Intro To Quantum Computing
Uploaded by
asdfasdfasdf786Intro To Quantum Computing
Uploaded by
asdfasdfasdf786hc | Q | ci
A Course in Quantum Computing
for the
Community College
Volume 1
Michael Loceff
Foothill College
mailto:[email protected]
c 2015 Michael Loceff
All Rights Reserved
This work is licensed under the Creative Commons
Attribution-NonCommercial-NoDerivatives 4.0 International License.
To view a copy of this license, visit
http://creativecommons.org/licenses/by-nc-nd/4.0/.
Contents
0 Introduction
0.1
21
Welcome to Volume One . . . . . . . . . . . . . . . . . . . . . . . . .
21
0.1.1
About this Volume . . . . . . . . . . . . . . . . . . . . . . . .
21
0.1.2
About this Introduction . . . . . . . . . . . . . . . . . . . . .
21
Bits and Qubits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
0.2.1
More Information than Zero and One . . . . . . . . . . . . . .
22
0.2.2
The Probabilistic Nature of Qubits . . . . . . . . . . . . . . .
22
0.2.3
Quantum Mechanics The Tool that Tames the Beast . . . .
24
0.2.4
Sneak Peek at the Coefficients and . . . . . . . . . . . . .
24
The Promise of Quantum Computing . . . . . . . . . . . . . . . . . .
25
0.3.1
Early Results . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
0.3.2
The Role of Computer Scientists
. . . . . . . . . . . . . . . .
26
The Two Sides of Quantum Computer Science . . . . . . . . . . . . .
26
0.4.1
Circuit Design . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
0.4.2
Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
0.5
Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
0.6
Navigating the Topics
29
0.2
0.3
0.4
. . . . . . . . . . . . . . . . . . . . . . . . . .
1 Complex Arithmetic
30
1.1
Complex Numbers for Quantum Computing . . . . . . . . . . . . . .
30
1.2
The Field of Complex Numbers . . . . . . . . . . . . . . . . . . . . .
31
1.2.1
The Real Numbers Just Dont Cut It . . . . . . . . . . . . . .
31
1.2.2
The Definition of C . . . . . . . . . . . . . . . . . . . . . . . .
31
1.2.3
The Complex Plane . . . . . . . . . . . . . . . . . . . . . . . .
32
1.2.4
Operations on Complex Numbers . . . . . . . . . . . . . . . .
33
1.2.5
C is a Field . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
Exploring the Complex Plane . . . . . . . . . . . . . . . . . . . . . .
35
1.3
1.4
1.5
1.3.1
Complex Numbers as Ordered Pairs of Real Numbers . . . . .
35
1.3.2
Real and Imaginary Axes and Polar Representation . . . . . .
35
1.3.3
Complex Conjugate and Modulus . . . . . . . . . . . . . . . .
36
Transcendental Functions and Their Identities . . . . . . . . . . . . .
38
1.4.1
The Complex Exponential Function Part 1: Pure Imaginary Case 38
1.4.2
Real sin() and cos() in Terms of the Complex Exponential . .
41
1.4.3
Complex Exponential Part 2: Any Complex Number . . . . .
41
1.4.4
The Complex Trigonometric Functions . . . . . . . . . . . . .
42
1.4.5
Polar Relations Expressed Using the Exponential . . . . . . .
42
Roots of Unity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
1.5.1
N Distinct Solutions to z N = 1 . . . . . . . . . . . . . . . .
44
1.5.2
Eulers Identity . . . . . . . . . . . . . . . . . . . . . . . . . .
46
1.5.3
Summation Notation . . . . . . . . . . . . . . . . . . . . . . .
47
1.5.4
Summing Roots-of-Unity and the Kronecker Delta . . . . . . .
48
2 Real Vector Spaces
51
2.1
Vector Spaces for Quantum Computing . . . . . . . . . . . . . . . . .
51
2.2
Vectors and Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . .
52
2.2.1
Standard Equipment: The Axioms . . . . . . . . . . . . . . .
52
2.2.2
Optional Equipment: Inner Products, Modulus and Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
2.2.3
A Bigger Vector Space: R . . . . . . . . . . . . . . . . . . . .
58
2.2.4
Preview of a Complex Vector Space: C2 . . . . . . . . . . . .
59
Bases for a Vector Space . . . . . . . . . . . . . . . . . . . . . . . . .
59
2.3.1
Linear Combination (or Superposition) . . . . . . . . . . . . .
59
2.3.2
The Notion of Basis
. . . . . . . . . . . . . . . . . . . . . . .
60
2.3.3
Coordinates of Vectors . . . . . . . . . . . . . . . . . . . . . .
65
2.3.4
Independence of Basis (or Not?) . . . . . . . . . . . . . . . . .
67
2.4
Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
68
2.5
Higher Dimensional Vector Spaces . . . . . . . . . . . . . . . . . . . .
68
2.6
More Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
2.3
3 Matrices
72
3.1
Matrices in Quantum Computing . . . . . . . . . . . . . . . . . . . .
72
3.2
Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
3.2.1
73
Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3
Matrix Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
3.3.1
Row Column . . . . . . . . . . . . . . . . . . . . . . . . . .
74
3.3.2
Definition of Matrix Multiplication . . . . . . . . . . . . . . .
74
3.3.3
Product of a Vector by a Matrix . . . . . . . . . . . . . . . . .
76
3.4
Matrix Transpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
3.5
Matrix Addition and Scalar Multiplication . . . . . . . . . . . . . . .
78
3.6
Identity Matrix and Zero Matrix . . . . . . . . . . . . . . . . . . . . .
78
3.7
Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
3.7.1
Determinant of a 2 2 Matrix . . . . . . . . . . . . . . . . .
79
3.7.2
Determinant of a 3 3 Matrix . . . . . . . . . . . . . . . . .
80
3.7.3
Determinant of an n n Matrix . . . . . . . . . . . . . . . .
81
3.7.4
Determinants of Products . . . . . . . . . . . . . . . . . . . .
81
3.8
Matrix Inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
82
3.9
Matrix Equations and Cramers Rule . . . . . . . . . . . . . . . . . .
83
3.9.1
Systems of Linear Equations . . . . . . . . . . . . . . . . . . .
83
3.9.2
Cramers Rule . . . . . . . . . . . . . . . . . . . . . . . . . . .
84
4 Hilbert Space
4.1
89
Complex Vector Spaces for Quantum Computing . . . . . . . . . . .
89
4.1.1
The Vector Nature of a Qubit . . . . . . . . . . . . . . . . . .
89
4.1.2
The Complex Nature of a Qubit . . . . . . . . . . . . . . . . .
90
4.2
The Complex Vector Space, C
. . . . . . . . . . . . . . . . . . . . .
90
4.3
The Complex Inner Product . . . . . . . . . . . . . . . . . . . . . . .
91
4.3.1
Norm and Distance . . . . . . . . . . . . . . . . . . . . . . . .
93
4.3.2
Expansion Coefficients . . . . . . . . . . . . . . . . . . . . . .
94
Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
97
4.4.1
Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
97
4.4.2
Old Friends and New Acquaintances . . . . . . . . . . . . . .
98
4.4.3
Some Useful Properties of Hilbert Spaces . . . . . . . . . . . .
99
4.4
4.5
4.6
Rays, Not Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5.1
Modeling Quantum Systems . . . . . . . . . . . . . . . . . . . 100
4.5.2
0 is not a Quantum State . . . . . . . . . . . . . . . . . . . . 103
4.5.3
Why? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
Almost There . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5 Linear Transformations
106
5.1
5.2
5.3
Linear Transformations for Quantum Computing . . . . . . . . . . . 106
5.1.1
A Concept More Fundamental Than the Matrix . . . . . . . . 106
5.1.2
The Role of Linear Transformations in Quantum Computing
Definitions and Examples . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2.1
Actions as well as Name Changes . . . . . . . . . . . . . . . . 107
5.2.2
Formal Definition of Linear Transformation . . . . . . . . . . 108
The Special Role of Bases . . . . . . . . . . . . . . . . . . . . . . . . 111
5.3.1
5.4
5.5
5.6
107
Application: Rotations in Space . . . . . . . . . . . . . . . . . 112
The Matrix of a Linear Transformation . . . . . . . . . . . . . . . . . 113
5.4.1
From Matrix to Linear Transformation . . . . . . . . . . . . . 113
5.4.2
From Linear Transformation to Matrix . . . . . . . . . . . . . 114
5.4.3
Dependence of a Matrix on Basis . . . . . . . . . . . . . . . . 116
5.4.4
The Transformation in an Orthonormal Basis . . . . . . . . . 119
Some Special Linear Transformations for Quantum Mechanics . . . . 124
5.5.1
The Adjoint of a Matrix . . . . . . . . . . . . . . . . . . . . . 124
5.5.2
Unitary Operators . . . . . . . . . . . . . . . . . . . . . . . . 125
5.5.3
Hermitian Operators . . . . . . . . . . . . . . . . . . . . . . . 130
Enter the Quantum World . . . . . . . . . . . . . . . . . . . . . . . . 131
6 The Experimental Basis of Quantum Computing
132
6.1
The Physical Underpinning for Spin 1/2 Quantum Mechanics
6.2
Physical Systems and Measurements . . . . . . . . . . . . . . . . . . 132
6.3
6.4
. . . . 132
6.2.1
Quantum Mechanics as a Model of Reality . . . . . . . . . . . 132
6.2.2
The Physical System, S . . . . . . . . . . . . . . . . . . . . . 133
6.2.3
Electron Spin as a Testbed for Quantum Mechanics
. . . . . 133
A Classical Attempt at Spin 1/2 Physics . . . . . . . . . . . . . . . . 134
6.3.1
An Imperfect Picture of Spin . . . . . . . . . . . . . . . . . . 134
6.3.2
A Naive Quantitative Definition of Electron Spin . . . . . . . 135
6.3.3
Spherical Representation . . . . . . . . . . . . . . . . . . . . . 136
Refining Our Model: Experiment #1 . . . . . . . . . . . . . . . . . . 137
6.4.1
The Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.4.2
The Actual Results . . . . . . . . . . . . . . . . . . . . . . . . 139
6.4.3
The Quantum Reality . . . . . . . . . . . . . . . . . . . . . . 139
6.4.4
A Follow-Up to Experiment #1 . . . . . . . . . . . . . . . . . 140
6.4.5
Results of the Follow-Up . . . . . . . . . . . . . . . . . . . . . 140
6.5
6.6
6.7
6.4.6
Quantum Mechanics Lesson #1 . . . . . . . . . . . . . . . . . 141
6.4.7
First Adjustment to the Model . . . . . . . . . . . . . . . . . 142
Refining Our Model: Experiment #2 . . . . . . . . . . . . . . . . . . 142
6.5.1
The Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.5.2
The Actual Results . . . . . . . . . . . . . . . . . . . . . . . . 144
6.5.3
The Quantum Reality . . . . . . . . . . . . . . . . . . . . . . 144
6.5.4
A Follow-Up to Experiment #2 . . . . . . . . . . . . . . . . . 144
6.5.5
Results of the Follow-Up . . . . . . . . . . . . . . . . . . . . . 145
6.5.6
Quantum Mechanics Lesson #2 . . . . . . . . . . . . . . . . . 145
6.5.7
Second Adjustment to the Model . . . . . . . . . . . . . . . . 146
Refining Our Model: Experiment #3 . . . . . . . . . . . . . . . . . . 147
6.6.1
What we Know . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.6.2
The Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.6.3
The Actual Results . . . . . . . . . . . . . . . . . . . . . . . . 150
6.6.4
The Quantum Reality . . . . . . . . . . . . . . . . . . . . . . 150
6.6.5
Quantum Mechanics Lesson #3 . . . . . . . . . . . . . . . . . 151
6.6.6
Third Adjustment to the Model . . . . . . . . . . . . . . . . . 151
Onward to Formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7 Time Independent Quantum Mechanics
153
7.1
Quantum Mechanics for Quantum Computing . . . . . . . . . . . . . 153
7.2
The Properties of Time-Independent Quantum Mechanics . . . . . . . 153
7.2.1
7.3
7.4
7.5
7.6
The State Space of a Physical System . . . . . . . . . . . . . . 154
The First Postulate of Quantum Mechanics . . . . . . . . . . . . . . . 155
7.3.1
Trait #1 (The State Space) . . . . . . . . . . . . . . . . . . . 155
7.3.2
The Fundamental State Space for Quantum Computing . . . . 156
7.3.3
Why Does a Projective H model Spin-1/2? . . . . . . . . . . . 157
The Second Postulate of Quantum Mechanics . . . . . . . . . . . . . 158
7.4.1
Trait #2 (The Operator for an Observable) . . . . . . . . . . 158
7.4.2
The Observable Sz . . . . . . . . . . . . . . . . . . . . . . . . 159
The Third Postulate of Quantum Mechanics . . . . . . . . . . . . . . 160
7.5.1
Trait #3 (The Eigenvalues of an Observable) . . . . . . . . . . 160
7.5.2
Eigenvectors and Eigenvalues . . . . . . . . . . . . . . . . . . 161
7.5.3
The Eigenvalues and Eigenvectors of Sz . . . . . . . . . . . . 162
Computing Eigenvectors and Eigenvalues . . . . . . . . . . . . . . . . 163
7.7
7.8
7.9
7.6.1
The Eigenvalues and Eigenvectors of Sy . . . . . . . . . . . . 163
7.6.2
The Eigenvalues and Eigenvectors of Sx . . . . . . . . . . . . 166
7.6.3
Summary of Eigenvectors and Eigenvalues for Spin-1/2 Observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Observables and Orthonormal Bases
. . . . . . . . . . . . . . . . . . 167
7.7.1
Trait #4 (Real Eigenvalues and Orthonormal Eigenvectors) . . 168
7.7.2
Using the x- or y-Basis . . . . . . . . . . . . . . . . . . . . . . 169
7.7.3
General States Expressed in Alternate Bases . . . . . . . . . . 170
The Completeness (or Closure) Relation . . . . . . . . . . . . . . . . 172
7.8.1
Orthonormal Bases in Higher Dimensions . . . . . . . . . . . . 172
7.8.2
Trait #5 (Closure Relation) . . . . . . . . . . . . . . . . . . . 173
The Fourth Postulate of Quantum Mechanics . . . . . . . . . . . . . 173
7.9.1
Trait #6 (Probability of Outcomes) . . . . . . . . . . . . . . . 173
7.10 The Fifth Postulate of Quantum Mechanics . . . . . . . . . . . . . . 176
7.10.1 Trait #7 (Post-Measurement Collapse) . . . . . . . . . . . . . 177
7.11 Summary of What Quantum Mechanics Tells Us About a System . . 178
7.12 Diracs Bra-ket Notation . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.12.1 Kets and Bras . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.12.2 The Adjoint of an Operator . . . . . . . . . . . . . . . . . . . 182
7.12.3 The Adjoint Conversion Rules . . . . . . . . . . . . . . . . . . 183
7.12.4 Trait #8 (Adjoint Conversion Rules) . . . . . . . . . . . . . . 184
7.13 Expectation Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
7.13.1 The Mean of the Experiments . . . . . . . . . . . . . . . . . . 185
7.13.2 Defining Expectation Value . . . . . . . . . . . . . . . . . . . 185
7.13.3 Computing Expectation Value . . . . . . . . . . . . . . . . . . 187
7.13.4 Trait #9 (Computing an Expectation Value) . . . . . . . . . . 187
7.13.5 Expectation Values in Spin-1/2 Systems . . . . . . . . . . . . 188
7.14 Its About Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
8 Time Dependent Quantum Mechanics
191
8.1
Time Evolution in Quantum Computing . . . . . . . . . . . . . . . . 191
8.2
The Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.3
8.2.1
Total Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.2.2
From Classical Hamiltonian to Quantum Hamiltonian . . . . . 192
The Hamiltonian for a Spin-1/2 System . . . . . . . . . . . . . . . . . 192
8.4
8.5
8.6
8.7
8.3.1
A Classical Hamiltonian . . . . . . . . . . . . . . . . . . . . . 192
8.3.2
A Quantum Hamiltonian . . . . . . . . . . . . . . . . . . . . . 194
The Energy Eigenkets . . . . . . . . . . . . . . . . . . . . . . . . . . 194
8.4.1
Relationship Between H and Sz . . . . . . . . . . . . . . . . . 194
8.4.2
Allowable Energies . . . . . . . . . . . . . . . . . . . . . . . . 195
Sixth Postulate of Quantum Mechanics . . . . . . . . . . . . . . . . . 196
8.5.1
The Schrodinger Equation . . . . . . . . . . . . . . . . . . . . 196
8.5.2
The Evolution of Spin in a Constant Magnetic Field . . . . . . 197
8.5.3
Stationary States . . . . . . . . . . . . . . . . . . . . . . . . . 200
8.5.4
General (Non-Stationary) States . . . . . . . . . . . . . . . . . 200
8.5.5
General Technique for Computing Time-Evolved States . . . . 201
Larmor Precession . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.6.1
The Time-Evolved Spin State in a Uniform B-Field . . . . . . 204
8.6.2
Evolution of the Spin Expectation Values . . . . . . . . . . . . 206
8.6.3
Summary of Larmor Precession . . . . . . . . . . . . . . . . . 210
The End and the Beginning . . . . . . . . . . . . . . . . . . . . . . . 211
9 The Qubit
212
9.1
Bits and Qubits as Vector Spaces . . . . . . . . . . . . . . . . . . . . 212
9.2
Classical Computation Models Informal Approach . . . . . . . . . . 212
9.2.1
9.3
9.4
9.5
Informal Definition of Bits and Gates . . . . . . . . . . . . . . 212
Classical Computation Models Formal Approach . . . . . . . . . . . 215
9.3.1
A Miniature Vector Space . . . . . . . . . . . . . . . . . . . . 215
9.3.2
Formal Definition of a (Classical) Bit . . . . . . . . . . . . . . 217
9.3.3
Formal Definition of a (Classical) Logical Operator . . . . . . 218
The Qubit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
9.4.1
Quantum Bits . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
9.4.2
Quantum Bit Values . . . . . . . . . . . . . . . . . . . . . . . 224
9.4.3
Usual Definition of Qubit and its Value . . . . . . . . . . . . . 224
9.4.4
Key Difference Between Bits and Qubits . . . . . . . . . . . . 225
Quantum Operators (Unary) . . . . . . . . . . . . . . . . . . . . . . . 227
9.5.1
Definition and Notation . . . . . . . . . . . . . . . . . . . . . 227
9.5.2
Case Study Our First Quantum Gate, QNOT . . . . . . . . 227
9.5.3
The Phase Flip, Z . . . . . . . . . . . . . . . . . . . . . . . . 233
9.5.4
The Bit-and-Phase Flip Operator, Y
. . . . . . . . . . . . . . 235
9.6
9.7
9.5.5
The Hadamard Gate, H . . . . . . . . . . . . . . . . . . . . . 237
9.5.6
Phase-Shift Gates, S, T and R . . . . . . . . . . . . . . . . . 240
Putting Unary Gates to Use . . . . . . . . . . . . . . . . . . . . . . . 241
9.6.1
Basis Conversion . . . . . . . . . . . . . . . . . . . . . . . . . 241
9.6.2
Combining Gates . . . . . . . . . . . . . . . . . . . . . . . . . 242
The Bloch Sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
9.7.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
9.7.2
Rewriting |i . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
9.7.3
The Expectation Vector for |i . . . . . . . . . . . . . . . . . 246
9.7.4
Definition of the Bloch Sphere . . . . . . . . . . . . . . . . . . 247
10 Tensor Products
248
10.1 Tensor Product for Quantum Computing . . . . . . . . . . . . . . . . 248
10.2 The Tensor Product of Two Vector Spaces . . . . . . . . . . . . . . . 249
10.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
10.2.2 Tensor Coordinates from Component-Space Coordinates . . . 257
10.3 Linear Operators on the Tensor Product Space . . . . . . . . . . . . . 263
10.3.1 Separable Operators . . . . . . . . . . . . . . . . . . . . . . . 263
10.3.2 The Matrix of a Separable Operator . . . . . . . . . . . . . . 265
10.3.3 The Matrix of a General Operator . . . . . . . . . . . . . . . . 267
10.3.4 Food for Thought . . . . . . . . . . . . . . . . . . . . . . . . . 268
11 Two Qubits and Binary Quantum Gates
269
11.1 The Jump from One to Two . . . . . . . . . . . . . . . . . . . . . . . 269
11.2 The State Space for Two Qubits . . . . . . . . . . . . . . . . . . . . . 270
11.2.1 Definition of a Two Quantum Bit (Bipartite) System . . . . 270
11.2.2 The Preferred Bipartite CBS . . . . . . . . . . . . . . . . . . . 271
11.2.3 Separable Bipartite States . . . . . . . . . . . . . . . . . . . . 272
11.2.4 Alternate Bipartite Bases . . . . . . . . . . . . . . . . . . . . 272
11.2.5 Non-Separable Bipartite Tensors . . . . . . . . . . . . . . . . . 275
11.2.6 Usual Definition of Two Qubits and their Values . . . . . . . . 276
11.3 Fundamental Two-Qubit Logic Gates . . . . . . . . . . . . . . . . . . 276
11.3.1 Binary Quantum Operators . . . . . . . . . . . . . . . . . . . 276
11.3.2 General Learning Example . . . . . . . . . . . . . . . . . . . . 277
11.3.3 Quantum Entanglement . . . . . . . . . . . . . . . . . . . . . 282
11.3.4 The Controlled-NOT (CNOT) Gate . . . . . . . . . . . . . . . 284
11.3.5 The Second-Order Hadamard Gate . . . . . . . . . . . . . . . 294
11.4 Measuring Along Alternate Bases . . . . . . . . . . . . . . . . . . . . 302
11.4.1 Measuring Along the x-Basis . . . . . . . . . . . . . . . . . . 303
11.4.2 Measuring Along any Separable Basis . . . . . . . . . . . . . . 303
11.4.3 The Entire Circuit Viewed in Terms of an Alternate Basis . . 305
11.4.4 Non-Separable Basis Conversion Gates . . . . . . . . . . . . . 305
11.5 Variations on the Fundamental Binary Qubit Gates . . . . . . . . . . 306
11.5.1 Other Controlled Logic Gates . . . . . . . . . . . . . . . . . . 306
11.5.2 More Separable Logic Gates . . . . . . . . . . . . . . . . . . . 312
11.6 The Born Rule and Partial Collapse . . . . . . . . . . . . . . . . . . . 315
11.6.1 The Born Rule for a Two-Qubit System . . . . . . . . . . . . 316
11.7 Multi-Gate Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
11.7.1 A Circuit that Produces Bell States . . . . . . . . . . . . . . . 319
11.7.2 A Circuit that Creates an Upside-Down CNOT . . . . . . . . 323
11.8 More than Two Qubits . . . . . . . . . . . . . . . . . . . . . . . . . . 325
11.8.1 Order-3 Tensor Products . . . . . . . . . . . . . . . . . . . . . 325
11.8.2 Three Qubit Systems . . . . . . . . . . . . . . . . . . . . . . . 326
11.8.3 Tripartite Bases and Separable States . . . . . . . . . . . . . . 326
11.8.4 End of Lesson . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
12 First Quantum Algorithms
335
12.1 Three Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
12.2 Superdense Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
12.2.1 Sending Information by Qubit . . . . . . . . . . . . . . . . . . 335
12.2.2 The Superdense Coding Algorithm . . . . . . . . . . . . . . . 336
12.3 Quantum Teleportation . . . . . . . . . . . . . . . . . . . . . . . . . . 342
12.3.1 The Quantum Teleportation Algorithm . . . . . . . . . . . . . 345
12.4 Introduction to Quantum Oracles . . . . . . . . . . . . . . . . . . . . 354
12.4.1 Boolean Functions and Reversibility . . . . . . . . . . . . . . . 354
12.4.2 The Quantum Oracle of a Boolean Function . . . . . . . . . . 357
12.5 Deutschs Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
12.5.1 Definitions and Statement of the Problem . . . . . . . . . . . 360
12.5.2 Deutschs Algorithm . . . . . . . . . . . . . . . . . . . . . . . 362
12.6 n Qubits and More Algorithms . . . . . . . . . . . . . . . . . . . . . 367
13 Multi-Qubit Systems and Algorithms
368
13.1 Moving Up from 2 Qubits to n Qubits . . . . . . . . . . . . . . . . . 368
13.2 General Tensor Products . . . . . . . . . . . . . . . . . . . . . . . . . 368
13.2.1 Recap of Order-2 Tensor Products . . . . . . . . . . . . . . . . 368
13.2.2 Recap of Order-3 Tensor Products . . . . . . . . . . . . . . . . 369
13.2.3 Higher Order Tensor Products . . . . . . . . . . . . . . . . . . 370
13.3 n-Qubit Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
13.3.1 Recap of Three Qubits . . . . . . . . . . . . . . . . . . . . . . 373
13.3.2 Three Qubit Logic Gates; the Toffoli Gate . . . . . . . . . . . 374
13.3.3 n Qubits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
13.3.4 n Qubit Logic Gates . . . . . . . . . . . . . . . . . . . . . . . 379
13.3.5 Oracles for n Qubit Functions . . . . . . . . . . . . . . . . . . 385
13.4 Significant Deterministic Speed-Up: The Deutsch-Jozsa Problem . . . 386
13.4.1 Deutsch-Jozsa Algorithm . . . . . . . . . . . . . . . . . . . . . 387
13.4.2 Quantum vs. Classical Time Complexity . . . . . . . . . . . . 392
13.4.3 Alternate Proof of the Deutsch-Jozsa Algorithm . . . . . . . . 395
13.5 True Non-Deterministic Speed-Up: The Bernstein-Vazirani Problem . 398
13.5.1 The Bernstein-Vazirani Algorithm . . . . . . . . . . . . . . . . 399
13.6 Generalized Born Rule . . . . . . . . . . . . . . . . . . . . . . . . . . 402
13.6.1 Trait #15 (Generalized Born Rule for (n + m)th order States) 402
13.7 Towards Advanced Quantum Algorithms . . . . . . . . . . . . . . . . 404
14 Probability Theory
405
14.1 Probability in Quantum Computing . . . . . . . . . . . . . . . . . . . 405
14.1.1 Probability for Classical Algorithms . . . . . . . . . . . . . . . 405
14.1.2 Probability for Quantum Algorithms . . . . . . . . . . . . . . 405
14.2 The Essential Vocabulary: Events vs. Probabilities . . . . . . . . . . 406
14.2.1 Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
14.2.2 Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
14.3 The Quantum Coin Flip . . . . . . . . . . . . . . . . . . . . . . . . . 407
14.4 Experimental Outcomes and the Sample Space . . . . . . . . . . . . . 408
14.4.1 Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
14.4.2 Requirements when Defining Outcomes . . . . . . . . . . . . . 410
14.4.3 An Incorrect Attempt at Defining Outcomes . . . . . . . . . . 411
14.4.4 Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
14.4.5 The Sample Space . . . . . . . . . . . . . . . . . . . . . . . . 412
14.4.6 Set Operations . . . . . . . . . . . . . . . . . . . . . . . . . . 413
14.5 Alternate Views of the Ten Qubit Coin Flip . . . . . . . . . . . . . . 414
14.6 Mod-2 Vectors in Probability Computations for Quantum Algorithms
416
14.6.1 Ordinary Linear Independence and Spanning . . . . . . . . . . 416
14.6.2 Linear Independence and Spanning in the Mod-2 Sense . . . . 417
14.6.3 Probability Warm-Up: Counting
. . . . . . . . . . . . . . . . 420
14.7 Fundamental Probability Theory . . . . . . . . . . . . . . . . . . . . 422
14.7.1 The Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
14.7.2 Assigning Probabilities in Finite, Equiprobable, Sample Spaces 422
14.8 Big Theorems and Consequences of the Probability Axioms . . . . . . 425
14.8.1 Unions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
14.8.2 Conditional Probability and Bayes Law . . . . . . . . . . . . 426
14.8.3 Statistical Independence . . . . . . . . . . . . . . . . . . . . . 427
14.8.4 Other Forms and Examples . . . . . . . . . . . . . . . . . . . 428
14.9 Wedge and Vee Notation and Lecture Recap . . . . . . . . . . . . . . 430
14.9.1 Disjoint Events . . . . . . . . . . . . . . . . . . . . . . . . . . 430
14.9.2 Partition of the Sample Space by Complements.
. . . . . . . 430
14.9.3 Bayes Law. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
14.9.4 Statistical Independence . . . . . . . . . . . . . . . . . . . . . 431
14.10Application to Deutsch-Jozsa . . . . . . . . . . . . . . . . . . . . . . 431
14.10.1 Sampling with Replacement . . . . . . . . . . . . . . . . . . . 432
14.10.2 Analysis Given a Balanced f . . . . . . . . . . . . . . . . . . . 432
14.10.3 Completion of Sampling with Replacement for Unknown f . . 434
14.10.4 Sampling without Replacement . . . . . . . . . . . . . . . . . 434
14.11A Condition for Constant Time Complexity in Non-Deterministic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
14.11.1 Non-Deterministic Algorithms . . . . . . . . . . . . . . . . . . 438
14.11.2 Preview of Time Complexity An Algorithms Dependence on
Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
14.11.3 Looping Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 439
14.11.4 Probabilistic Algorithms with Constant Time Complexity . . . 439
14.11.5 A Constant Time Condition for Looping Algorithms . . . . . . 439
15 Computational Complexity
442
15.1 Computational Complexity in Quantum Computing . . . . . . . . . . 442
15.2 An Algorithms Sensitivity to Size . . . . . . . . . . . . . . . . . . . . 442
15.2.1 Some Examples of Time Complexity . . . . . . . . . . . . . . 442
15.2.2 Time Complexity vs. Space Complexity . . . . . . . . . . . . 444
15.2.3 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
15.3 Big-O Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
15.3.1 Conflicting Ways to Measure an Algorithm . . . . . . . . . . . 445
15.3.2 Definition of Big-0 . . . . . . . . . . . . . . . . . . . . . . . . 446
15.3.3 Common Terminology for Certain Big-O Growth Rates
15.3.4 Factors and Terms We Can Ignore
. . . 446
. . . . . . . . . . . . . . . 447
15.4 Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
15.4.1 Definition of . . . . . . . . . . . . . . . . . . . . . . . . . . 449
15.5 Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
15.6 Little-o Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
15.7 Easy vs. Hard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
15.8 Wrap-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
16 Computational Basis States and Modular Arithmetic
451
16.1 Different Notations Used in Quantum Computing . . . . . . . . . . . 451
16.2 Notation and Equivalence of Three Environments . . . . . . . . . . . 451
16.2.1 First Environment n-qubit Hilbert Space, H(n)
Z2 )
16.2.2 The Second Environment: The Finite Group (Z
. . . . . . . 451
n
. . . . . . 453
16.2.3 The Third Environment: The Finite Group Z 2n with Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Z2 )n and (Z
Z2n , ) . . . . 459
16.2.4 Interchangeable Notation of H(n) , (Z
17 Quantum Oracles
462
17.1 Higher Dimensional Oracles and their Time Complexity . . . . . . . . 462
17.2 Simplest Oracle: a Boolean Function of One Bit . . . . . . . . . . . . 463
17.2.1 Circuit and Initial Remarks . . . . . . . . . . . . . . . . . . . 463
17.2.2 A Two-Qubit Oracles Action on the CBS . . . . . . . . . . . 465
17.2.3 Case Study #1: f (x) 1 . . . . . . . . . . . . . . . . . . . . 465
17.2.4 Case Study #2: f (x) x . . . . . . . . . . . . . . . . . . . . 466
17.2.5 Remaining Cases #3 and #4 . . . . . . . . . . . . . . . . . . 467
17.3 Integers Mod-2 Review . . . . . . . . . . . . . . . . . . . . . . . . . . 468
17.3.1 The Classical f at the Heart of an Oracle . . . . . . . . . . . 468
17.3.2 Mod-2 Notation for { 0, 1 }n . . . . . . . . . . . . . . . . . . . 468
17.3.3 Inside the Ket . . . . . . . . . . . . . . . . . . . . . . . . . 469
17.4 Intermediate Oracle: f is a Boolean Function of a Multiple Input Bits 470
17.4.1 Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
17.4.2 An (n + 1)-Qubit Oracles Action on the CBS . . . . . . . . . 470
17.4.3 Analyzing Uf for x = 0 . . . . . . . . . . . . . . . . . . . . . 471
17.4.4 Analyzing Uf for General x . . . . . . . . . . . . . . . . . . . 472
17.5 Advanced Oracle: f is a Multi-Valued function of a Multiple Input Bits 475
17.5.1 Circuit and Vocabulary . . . . . . . . . . . . . . . . . . . . . . 475
17.5.2 Smallest Advanced Oracle: m = 2 (Range of f Z 22 ) . . . . 475
17.5.3 The 4 4 Sub-Matrix for a Fixed x
. . . . . . . . . . . . . . 475
17.5.4 Easy Way to See Unitarity . . . . . . . . . . . . . . . . . . 481
17.5.5 The General Advanced Oracle: Any m (Range of f Z 2m ) . 481
17.6 The Complexity of a Quantum Algorithm Relative to the Oracle . . . 483
18 Simons Algorithm for Period-Finding
484
18.1 The Importance of Simons Algorithm . . . . . . . . . . . . . . . . . 484
18.2 Periodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
18.2.1 Ordinary Periodicity . . . . . . . . . . . . . . . . . . . . . . . 485
Z2 )n Periodicity . . . . . . . . . . . . . . . . . . . . . . . . . 485
18.2.2 (Z
18.2.3 Z2n Periodicity . . . . . . . . . . . . . . . . . . . . . . . . . . 486
18.2.4 1-to-1, 2-to-1 and n-to-1
. . . . . . . . . . . . . . . . . . . . 489
18.3 Simons Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
18.4 Simons Quantum Circuit Overview and the Master Plan . . . . . . . 492
18.4.1 The Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
18.4.2 The Plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
18.5 The Circuit Breakdown . . . . . . . . . . . . . . . . . . . . . . . . . . 494
18.6 Circuit Analysis Prior to Conceptual Measurement: Point B . . . . . 494
18.6.1 Hadamard Preparation of the A Register
. . . . . . . . . . . 494
18.6.2 The Quantum Oracle on CBS Inputs . . . . . . . . . . . . . . 495
18.6.3 The Quantum Oracle on Hadamard Superposition Inputs . . . 497
18.6.4 Partitioning the Domain into Cosets . . . . . . . . . . . . . . 498
18.6.5 Rewriting the Output of the Oracles B Register . . . . . . . 499
18.7 Analysis of the Remainder of the Circuit: Measurements . . . . . . . 500
18.7.1 Hypothetical Measurement of the B Register . . . . . . . . . 500
18.7.2 Effect of a Final Hadamard on A Register
. . . . . . . . . . 501
18.7.3 The Orthogonality of A Register Output Relative to the Unknown Period a . . . . . . . . . . . . . . . . . . . . . . . . . 503
18.7.4 Foregoing the Conceptual Measurement . . . . . . . . . . . . . 505
18.8 Circuit Analysis Conclusion . . . . . . . . . . . . . . . . . . . . . . . 506
18.9 Simons Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
18.9.1 Producing n 1 Linearly Independent Vectors . . . . . . . . 507
18.9.2 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 508
18.9.3 Strange Behavior . . . . . . . . . . . . . . . . . . . . . . . . . 510
18.10Time Complexity of the Quantum Algorithm . . . . . . . . . . . . . . 512
18.10.1 Producing n1 Linearly-Independent wk in Polynomial Time
Argument 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
18.10.2 Proof of Theorem Used by Argument 1 . . . . . . . . . . . . . 512
18.10.3 Summary of Argument 1 . . . . . . . . . . . . . . . . . . . . . 517
18.10.4 Producing n1 Linearly-Independent wk in Polynomial Time
Argument 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
18.10.5 Proof of Theorem Used by Argument 2 . . . . . . . . . . . . . 518
18.10.6 Summary of Argument 2 . . . . . . . . . . . . . . . . . . . . . 521
18.10.7 Discussion of the Two Proofs Complexity Estimates . . . . . 521
18.11The Hidden Classical Algorithms and Their Cost . . . . . . . . . . . 522
18.11.1 Unaccounted for Steps . . . . . . . . . . . . . . . . . . . . . . 522
18.12Solving Systems of Mod-2 Equations . . . . . . . . . . . . . . . . . . 523
18.12.1 Gaussian Elimination and Back Substitution . . . . . . . . . . 523
18.12.2 Gaussian Elimination . . . . . . . . . . . . . . . . . . . . . . . 524
18.12.3 Back Substitution . . . . . . . . . . . . . . . . . . . . . . . . . 528
18.12.4 The Total Cost of the Classical Techniques for Solving Mod-2
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
18.13Applying GE and Back Substitution to Simons Problem . . . . . . . 529
18.13.1 Linear Independence . . . . . . . . . . . . . . . . . . . . . . . 530
18.13.2 Completing the Basis with an nth Vector Not Orthogonal to a 531
18.13.3 Using Back-Substitution to Close the Deal . . . . . . . . . . . 534
18.13.4 The Full Cost of the Hidden Classical Algorithms . . . . . . . 534
18.14Adjusted Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
18.14.1 New Linear Independence Step . . . . . . . . . . . . . . . . . 534
18.14.2 New Solution of System Step . . . . . . . . . . . . . . . . . . 537
18.14.3 Cost of Adjusted Implementation of Simons Algorithm . . . . 537
18.15Classical Complexity of Simons Problem . . . . . . . . . . . . . . . . 538
18.15.1 Classical Deterministic Cost . . . . . . . . . . . . . . . . . . . 538
18.15.2 Classical Probabilistic Cost . . . . . . . . . . . . . . . . . . . 539
19 Real and Complex Fourier Series
542
19.1 The Classical Path to Quantum Fourier Transforms . . . . . . . . . . 542
19.2 Periodic Functions and Their Friends . . . . . . . . . . . . . . . . . . 543
19.2.1 Periodic Functions over R . . . . . . . . . . . . . . . . . . . . 543
19.2.2 Functions with Bounded Domain or Compact Support . . . . 544
19.2.3 The Connection Between Periodicity and Bounded Domain . . 545
19.3 The Real Fourier Series of a Real Period Function . . . . . . . . . . . 546
19.3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
19.3.2 Interpretation of the Fourier Series . . . . . . . . . . . . . . . 547
19.3.3 Example of a Fourier Series . . . . . . . . . . . . . . . . . . . 549
19.4 The Complex Fourier Series of a Periodic Function . . . . . . . . . . 554
19.4.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
19.4.2 Computing The Complex Fourier Coefficients . . . . . . . . . 555
19.5 Periods and Frequencies . . . . . . . . . . . . . . . . . . . . . . . . . 557
19.5.1 The Frequency of any Periodic Function . . . . . . . . . . . . 557
19.5.2 Ordinary Frequency . . . . . . . . . . . . . . . . . . . . . . . . 557
19.5.3 Angular Frequency . . . . . . . . . . . . . . . . . . . . . . . . 558
20 The Continuous Fourier Transform
560
20.1 From Series Transform . . . . . . . . . . . . . . . . . . . . . . . . . . 560
20.2 Motivation and Definitions . . . . . . . . . . . . . . . . . . . . . . . . 560
20.2.1 Non-Periodic Functions . . . . . . . . . . . . . . . . . . . . . . 560
20.2.2 Main Result of Complex Fourier Series . . . . . . . . . . . . . 561
20.2.3 Tweaking the Complex Fourier Series . . . . . . . . . . . . . . 561
20.2.4 Definition of the Fourier Transform . . . . . . . . . . . . . . . 562
20.2.5 The Inverse Fourier Transform . . . . . . . . . . . . . . . . . . 562
20.2.6 Real vs. Complex, Even vs. Odd . . . . . . . . . . . . . . . . 563
20.2.7 Conditions for a function to Possess a Fourier Transform . . . 564
20.3 Learning to Compute . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
20.3.1 Example 1: Rectangular Pulse . . . . . . . . . . . . . . . . . . 565
20.3.2 Example 2: Gaussian
20.4 Interlude: The Delta Function
. . . . . . . . . . . . . . . . . . . . . . 566
. . . . . . . . . . . . . . . . . . . . . 567
20.4.1 Characterization of the Delta Function . . . . . . . . . . . . . 567
20.4.2 The Delta Function as a Limit of Rectangles . . . . . . . . . . 568
20.4.3 The Delta Function as a Limit of Exponentials . . . . . . . . . 569
20.4.4 Sifting Property of the Delta Function . . . . . . . . . . . . . 569
20.5 Fourier Transforms Involving (x)
. . . . . . . . . . . . . . . . . . . 570
20.5.1 Example 3: A Constant . . . . . . . . . . . . . . . . . . . . . 570
20.5.2 Example 4: A Cosine . . . . . . . . . . . . . . . . . . . . . . . 572
20.5.3 Example 5: A Sine . . . . . . . . . . . . . . . . . . . . . . . . 572
20.6 Properties of the Fourier Transform . . . . . . . . . . . . . . . . . . . 573
20.6.1 Translation Invariance . . . . . . . . . . . . . . . . . . . . . . 573
20.6.2 Plancherels Theorem . . . . . . . . . . . . . . . . . . . . . . . 574
20.6.3 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
20.6.4 The Convolution Theorem . . . . . . . . . . . . . . . . . . . . 575
20.7 Period and Frequency in Fourier Transforms . . . . . . . . . . . . . . 575
20.8 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
20.8.1 Whats the FT Used For? . . . . . . . . . . . . . . . . . . . . 576
20.8.2 The Uncertainty Principle . . . . . . . . . . . . . . . . . . . . 578
20.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
21 The Discrete and Fast Fourier Transforms
580
21.1 From Continuous to Discrete . . . . . . . . . . . . . . . . . . . . . . . 580
21.2 Motivation and Definitions . . . . . . . . . . . . . . . . . . . . . . . . 581
21.2.1 Functions Mapping ZN C . . . . . . . . . . . . . . . . . 581
21.2.2 Defining the DFT
. . . . . . . . . . . . . . . . . . . . . . . . 582
21.3 Matrix Representation of the DFT . . . . . . . . . . . . . . . . . . . 587
21.4 Properties of DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587
21.4.1 Convolution of Two Vectors . . . . . . . . . . . . . . . . . . . 587
21.4.2 Translation Invariance (Shift Property) for Vectors . . . . . . 588
21.4.3 Computational Complexity of DFT . . . . . . . . . . . . . . . 588
21.5 Period and Frequency in Discrete Fourier Transforms . . . . . . . . . 588
21.6 A Cost-Benefit Preview of the FFT
. . . . . . . . . . . . . . . . . . 591
21.6.1 Benefit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
21.6.2 Cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
21.7 Recursive Equation for FFT
. . . . . . . . . . . . . . . . . . . . . . 592
21.7.1 Splitting f into f even and f odd . . . . . . . . . . . . . . . . . . 592
21.7.2 N th Order DFT in Terms of Two (N/2)th Order DFT s . . . 593
21.7.3 Danielson-Lanczos Recursion Relation . . . . . . . . . . . . . 594
21.7.4 Code Samples: Recursive Algorithm . . . . . . . . . . . . . . . 595
21.8 A Non-Recursive, N log N Solution that Defines the FFT
. . . . . . 597
21.8.1 The High-Level FFT Method . . . . . . . . . . . . . . . . . . 598
21.8.2 Bit-Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
21.8.3 Rebuilding from the Bit-Reversed Array . . . . . . . . . . . . 604
21.8.4 Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
21.8.5 Overall Complexity . . . . . . . . . . . . . . . . . . . . . . . . 609
21.8.6 Software Testing . . . . . . . . . . . . . . . . . . . . . . . . . 609
22 The Quantum Fourier Transform
610
22.1 From Classical Fourier Theory to the QFT
. . . . . . . . . . . . . . 610
22.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
n
22.2.1 From C2 to H(n) . . . . . . . . . . . . . . . . . . . . . . . . . 610
22.2.2 Approaches to Operator Definition . . . . . . . . . . . . . . . 611
22.2.3 Review of Hadamard Operator
22.2.4 Defining the QFT
. . . . . . . . . . . . . . . . . 612
. . . . . . . . . . . . . . . . . . . . . . . . 613
22.3 Features of the QFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
22.3.1 Shift Property . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
22.3.2 A Comparison between QFT and H . . . . . . . . . . . . . . 618
22.3.3 The Quantum Fourier Basis . . . . . . . . . . . . . . . . . . . 619
22.4 The QFT Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620
22.4.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620
22.4.2 The Math that Leads to the Circuit: Factoring the QFT . . . 621
22.4.3 The QFT Circuit from the Math: n = 3 Case Study . . . . . 627
22.4.4 The QFT Circuit from the Math: General Case . . . . . . . . 634
22.5 Computational Complexity of QFT . . . . . . . . . . . . . . . . . . . 635
23 Shors Algorithm
638
23.1 The Role of Shors Algorithms in Computing . . . . . . . . . . . . . . 638
23.1.1 Context for The Algorithms . . . . . . . . . . . . . . . . . . . 638
23.1.2 Period Finding and Factoring . . . . . . . . . . . . . . . . . . 638
23.1.3 The Period Finding Problem and its Key Idea . . . . . . . . . 639
23.2 Injective Periodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642
23.2.1 Functions of the Integers . . . . . . . . . . . . . . . . . . . . . 642
23.2.2 Functions of the Group Z M . . . . . . . . . . . . . . . . . . . 643
23.2.3 Discussion of Injective Periodicity . . . . . . . . . . . . . . . . 643
23.3 Shors Periodicity Problem . . . . . . . . . . . . . . . . . . . . . . . . 644
23.3.1 Definitions and Recasting the Problem . . . . . . . . . . . . . 646
Z2 )n CBS Connection
23.3.2 The Z N (Z
. . . . . . . . . . . . . . 648
23.4 Shors Quantum Circuit Overview and the Master Plan . . . . . . . . 649
23.4.1 The Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
23.4.2 The Plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
23.5 The Circuit Breakdown . . . . . . . . . . . . . . . . . . . . . . . . . . 651
23.6 Circuit Analysis Prior to Conceptual Measurement: Point B . . . . . 651
23.6.1 The Hadamard Preparation of the A register . . . . . . . . . . 651
23.6.2 The Quantum Oracle . . . . . . . . . . . . . . . . . . . . . . . 652
23.6.3 The Quantum Oracle on Hadamard Superposition Inputs . . . 653
23.7 Fork-in-the Road: An Instructional Case Followed by the General Case 653
23.8 Intermezzo Notation for GCD and Coprime . . . . . . . . . . . . . 654
23.8.1 Greatest Common Divisor . . . . . . . . . . . . . . . . . . . . 654
23.8.2 Coprime (Relatively Prime) . . . . . . . . . . . . . . . . . . . 654
23.9 First Fork: Easy Case (a N ) . . . . . . . . . . . . . . . . . . . . . . 655
23.9.1 Partitioning the Domain into Cosets . . . . . . . . . . . . . . 655
23.9.2 Rewriting the Output of the Oracle . . . . . . . . . . . . . . . 656
23.9.3 Implication of a Hypothetical Measurement of the B register
Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657
23.9.4 Effect of a Final QFT on the A Register . . . . . . . . . . . . 659
23.9.5 Computation of Final Measurement Probabilities (Easy Case)
661
{yc }a1
c=0
23.9.6 STEP I: Identify a Special Set of a Elements, C =
of
Certain Measurement Likelihood . . . . . . . . . . . . . . . . 661
23.9.7 Step II: Observe that Each cm Will be Measured with Equal
Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662
23.9.8 Step III: Prove that a Random Selection from [0, a 1] will be
Coprime-to-a 50% of the Time . . . . . . . . . . . . . . . . . . 664
23.9.9 Step IV: Observe that a y = cm Associated with c Coprime-to-a
Will be Measured with Probability 1/2 . . . . . . . . . . . . . 664
23.9.10 Step V: Observe that yc Associated with c Coprime to a Will
be Measured in Constant Time . . . . . . . . . . . . . . . . . 666
23.9.11 Algorithm and Complexity Analysis (Easy Case) . . . . . . . 667
23.10Second Fork: General Case (We do not Assume a N ) . . . . . . . . . 669
23.10.1 Partitioning the Domain into Cosets . . . . . . . . . . . . . . 670
23.10.2 Rewriting the Output of the Oracles output . . . . . . . . . . 672
23.10.3 Implication of a Hypothetical Measurement of the B register
Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673
23.10.4 Effect of a Final QFT on the A Register . . . . . . . . . . . . 675
23.10.5 Computation of Final Measurement Probabilities (General Case)677
23.10.6 STEP I: Identify (Without Proof ) a Special Set of a Elements,
a1
C = {yc }c=0
of High Measurement Likelihood . . . . . . . . . . 678
23.10.7 STEP II: Prove that the Values in, C = {yc }a1
c=0 Have High
Measurement Likelihood . . . . . . . . . . . . . . . . . . . . . 683
a1
23.10.8 STEP III: Associate {yc }a1
c=0 with {c/a}c=0
. . . . . . . . . . 694
23.10.9 STEP IV: Describe an O(log3 N ) Algorithm that Will Produce
c/a from yc . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
23.10.10Step V: Measure yc Associated with c Coprime to a in Constant
Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 699
23.10.11Algorithm and Complexity Analysis (General Case) . . . . . . 708
23.10.12Epilogue on Shors Period-Finding Algorithm . . . . . . . . . 710
24 Euclidean Algorithm and Continued Fractions
711
24.1 Ancient Algorithms for Quantum Computing . . . . . . . . . . . . . . 711
24.2 The Euclidean Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 711
24.2.1 Greatest Common Divisor . . . . . . . . . . . . . . . . . . . . 711
24.2.2 Long Division . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
24.2.3 The Euclidean Algorithm, EA(P, Q) . . . . . . . . . . . . . . 712
24.2.4 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 713
24.2.5 Time Complexity of EA . . . . . . . . . . . . . . . . . . . . . 714
24.3 Convergents and the Continued Fraction Alorithm . . . . . . . . . . . 715
24.3.1 Continued Fractions . . . . . . . . . . . . . . . . . . . . . . . 715
24.3.2 Computing the CFA ak s Using the EA if x is Rational . . . . 716
24.3.3 A Computer Science Method for Computing the ak s of Continued Fractions . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
24.3.4 Convergents of a Continued Fraction . . . . . . . . . . . . . . 718
24.3.5 An Algorithm for Computing the Convergents {nk /dk } . . . . 722
24.3.6 Easy Properties of the Convergents . . . . . . . . . . . . . . . 722
24.3.7 CFA: Our Special Brand of Continued Fraction Algorithm . . 723
25 From Period-Finding to Factoring
725
25.1 Period Finding to Factoring to RSA Encryption . . . . . . . . . . . . 725
25.2 The Problem and Two Classically Easy Cases . . . . . . . . . . . . . 725
25.3 A Sufficient Condition . . . . . . . . . . . . . . . . . . . . . . . . . . 727
25.4 A Third Easy Case and Order-Finding in ZM
. . . . . . . . . . . . . 728
25.5 Using the Order of y to Find the x of our Sufficient Condition
. . 730
25.6 The Time Complexity of f (x) = y (mod M ) . . . . . . . . . . . . . 732
25.7 The Complexity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 733
25.7.1 Complexity of Step 1 . . . . . . . . . . . . . . . . . . . . . . . 733
25.7.2 Complexity of Step 2 . . . . . . . . . . . . . . . . . . . . . . . 734
25.7.3 Absolute Complexity of Shors Factoring . . . . . . . . . . . . 734
List of Figures
736
List of Tables
741
Chapter 0
Introduction
0.1
0.1.1
Welcome to Volume One
About this Volume
This book accompanies CS 83A, the first of a three quarter quantum computing
sequence offered to sophomores during their second or third years at Foothill Community College in Los Altos Hills, California. This first course focuses on quantum
computing basics under the assumption that we have noise-free quantum components
with which to build our circuits. The subsequent courses deal with advanced algorithms and quantum computing in the presence of noise, specifically, error correction
and quantum encryption.
This is not a survey course; it skips many interesting aspects of quantum computing. On the other hand, it is in-depth. The focus is on doing. My hope is that some
of you will apply the hard skills you learn here to discover quantum algorithms
of your own. However, even if this is the only course you take on the subject, the
computational tools you learn will be applicable far beyond the confines of quantum
information theory.
0.1.2
About this Introduction
This short introduction contains samples of the math symbolism we will learn later
in the course. These equations are intended only to serve as a taste of whats ahead.
You are not expected to know what the expressions and symbols mean yet, so dont
panic. All will be revealed in the next few weeks. Consider this introduction to be a
no-obligation sneak preview of things to come.
21
0.2
0.2.1
Bits and Qubits
More Information than Zero and One
Classical computing is done with bits, which can be either 0 or 1, period.
Quantum computing happens in the world of quantum bits, called qubits. Until a
qubit, call it |i, is directly or indirectly measured, it is in a state that is neither 0
nor 1, but rather a superposition of the two, expressed formally as
|i
|0i + |1i .
The symbol |0i corresponds to classical 0 value, while |1i is associated with a
classical value of 1. Meanwhile, the symbols and stand for numbers that express
how much 0 and how much 1 are present in the qubit.
Well make this precise shortly, but the idea that you can have a teaspoon of 0
and a tablespoon of 1 contained in a single qubit immediately puts us on alert that
we are no longer in the world of classical computing. This eerie concept becomes all
the more magical when you consider that a qubit exists on a sub-atomic level (as
photon or the spin state of an electron, for example), orders of magnitude smaller
than the physical embodiment of a single classical bit which requires about a million
atoms (or in research labs as few as 12).
That an infinitely small entity such as a qubit can store so much more information
than a bulky classical bit comes at a price, however.
0.2.2
The Probabilistic Nature of Qubits
A Classical Experiment
If 100 classical one-bit memory locations are known to all hold the same value call
it x until we know what that value is then they all hold x = 0 or all hold x = 1. If
we measure the first location and find it to be 1, then we will have determined that
all 100 must hold a 1 (because of the assumption that all 100 locations are storing
the exact same value). Likewise, if we measure a 0, wed know that all 100 locations
Figure 1: After measuring one location, we know them all
22
contain the value 0. Measuring the other 99 locations would confirm our conclusion.
Everything is logical.
A Quantum Experiment
Qubits are a lot more slippery. Imagine a quantum computer capable of storing
qubits. In this hypothetical we can inspect the contents of any memory location in
our computer by attaching an output meter to that location and read a result off the
meter.
Lets try that last experiment in our new quantum computer. We load 100 qubit
memory locations with 100 identically prepared qubits. Identically prepared means
that each qubit has the exact same value, call it |i. (Never mind that I havent
explained what the value of a qubit means; it has some meaning, and Im asking you
imagine that all 100 have the same value.)
Next, we use our meter to measure the first location. As in the classical case
we discover that the output meter registers either a 0 or a 1. Thats already a
disappointment. We were hoping to get some science-fictiony-looking measurement
from a qubit, especially one with a name like |i. Never mind; we carry on. Say
the location gives us a measurement of 1.
Summary to this point. We loaded up all 100 locations with the same qubit,
peered into the first location, and saw that it contained an ordinary 1.
What should we expect if we measure the other 99 locations?
have no idea what to expect.
Answer: We
Figure 2: After measuring one location, we dont know much
Some of the remaining qubits may show us a 1, others a 0, all despite the fact
that the 100 locations initially stored identical qubit values.
This is disquieting. To our further annoyance,
the measurement of each qubit always produces either a 0 or a 1 on our
meter, despite our having prepared a state |i = |0i + |1i , falsely claiming
to be some exciting combination of the two classical values,
23
the measurement will have permanently destroyed the original state we prepared, leaving it in a classical condition of either 0 or 1, no more magical
superposition left in there,
as already stated we know nothing (well, almost nothing, but thats for another
day) about the measurement outcomes of the other 99 supposedly identically
prepared locations, and
most bizarre of all, in certain situations, measuring the state of any one of these
qubits will cause another qubit in a different computer, room, planet or galaxy
to be modified without the benefit of wires, radio waves or time.
0.2.3
Quantum Mechanics The Tool that Tames the Beast
Such wild behavior is actually well managed using quantum mechanics, the mathematical symbol-manipulation game that was invented in the early 20th century to
help explain and predict the behavior of very, very small things. I cited the truculent nature of qubits in this introduction as a bit of a sensationalism to both scare
and stimulate you. We can work with these things very easily despite their unusual
nature.
The challenge for us is that quantum mechanics and its application to information and algorithms is not something one can learn in a week or two. But one can
learn it in a few months, and thats what were going to do in this course. Ive prepared a sequence of lessons which will walk you through the fascinating mathematics
and quantum mechanics needed to understand the new algorithms. Because it takes
hard, analytical work, quantum computing isnt for everyone. But my hope is that
some among you will find this volume an accessible first step from which you can go
on to further study and eventually invent quantum algorithms of your own.
0.2.4
Sneak Peek at the Coefficients and
So as not to appear too secretive, Ill give you taste of what and roughly mean
for the state |i. They tell us the respective probabilities that we would obtain a
reading of either a 0 or a 1 were we to look into the memory location where |i
is stored. (In our quantum jargon, this is called measuring the state |i.) If, for
example, the values happened to be
1
and
= ,
2
3
=
,
2
then the
(
0 would be (1/2)2 = 14 = 25%
probability of measuring
.
1 would be ( 3/2)2 = 34 = 75%
24
In other words, if a qubit with the precise and definitive value
|i
|0i + |1i
is sitting in a one-(qu)bit memory location, and we look at that location, we will
actually see a
0
with probability
||2
with probability
||2 .
and
This is far from the whole story as well learn in our very first lecture, but it gives
you a feel for how the probabilistic nature of the quantum world can be both slippery
and quantitative at the same time. We dont know what well get when we query a
quantum memory register, but we do know what the probabilities will be.
0.3
0.3.1
The Promise of Quantum Computing
Early Results
If were going to expend time studying math and quantum mechanics, we should
expect something in return. The field is evolving rapidly, with new successes and
failures being reported weekly. However, there are a few established results which are
incontrovertible, and they are the reason so much effort is being brought to bear on
quantum computer design.
Of the early results, perhaps the most dramatic is Shors period-finding algorithm,
and it is this that I have selected as the endpoint of our first volume. It provides
a basis for factoring extremely large numbers in a reasonable time when such feats
would take classical computers billions of years. The applications, once implemented
are profound. However the consequences may give one pause; network security as we
know it would become obsolete.
Fortunately, or so we believe, there are different quantum techniques that offer
alternatives to current network security and which could render it far more secure
than it is today. (These require additional theory, beyond the basics that we learn in
volume 1 and will be covered in the sequel.)
There are also less sensational, but nevertheless real, improvements that have
been discovered. Grovers search algorithm for unsorted linear lists, while offering
a modest speed-up over classical searches, is attractive merely by the ubiquity of
search in computing. Related search techniques that look for items over a network
are being discovered now and promise to replicate such results for graphs rather than
list structures.
Quantum teleportation and super dense coding are among the simplest applications of quantum computing, and they provide a glimpse into possible new approaches
to more efficient communication. Well get to these in this volume.
25
0.3.2
The Role of Computer Scientists
Quantum computers dont exist yet. There is production grade hardware that appears
to leverage quantum behavior but does not exhibit the simple qubit processing needs
of the early or indeed most of the current quantum algorithms in computer
science. On the other hand, many university rigs possess the right stuff for quantum
algorithms, but they are years away from having the stability and/or size to appear
in manufactured form.
The engineers and physicists are doing their part.
The wonderful news for us computer scientists is that we dont have to wait.
Regardless of what the hardware ultimately looks like, we already know what it
will do. Thats because it is based on the most fundamental, firmly established and
despite my scary sounding lead-in surprisingly simple quantum mechanics. We know
what a qubit is, how a quantum logic gate will affect it, and what the consequences of
reading qubit registers are. There is nothing preventing us from designing algorithms
right now.
0.4
The Two Sides of Quantum Computer Science
Given that we can strap ourselves in and start work immediately, we should be clear
on the tasks at hand. There are two.
Circuit Design. We know what the individual components will be, even if
they dont exist yet. So we must gain some understanding and proficiency in
the assembly of these parts to produce full circuits.
Algorithm Design. Because quantum mechanics is probabilistic by nature,
well have to get used to the idea that the circuits dont always give us the
answer right away. In some algorithms they do, but in others, we have to send
the same inputs into the same circuits many times and let the laws of probability
play out. This requires us to analyze the math so we can know whether we have
a fighting chance of our algorithm converging to an answer with adequate error
tolerance.
0.4.1
Circuit Design
Classical
Classical logic gates are relatively easy to understand. An AND gate, for example,
has a common symbol and straightforward truth table that defines it:
26
x
y
xy
xy
You were introduced to logic like this in your first computer science class. After about
20 minutes of practice with various input combinations, you likely absorbed the full
meaning of the AND gate without serious incident.
Quantum
A quantum logic gate requires significant vocabulary and symbolism to even define,
never mind apply. If you promise not to panic, Ill give you a peek. Of course, youll
be trained in all the math and quantum mechanics in this course before we define
such a circuit officially. By then, youll be eating quantum logic for breakfast.
Well take the example of something called a second order Hadamard gate. We
would start by first considering the second order qubit on which the gate operates.
Such a thing is symbolized using the mysterious notation and a column of numbers,
|i2 =
.
Next, we would send this qubit through the Hadamard gate using the symbolism
(
|i2
H 2
H 2 |i2
Although it means little to us at this stage, the diagram shows the qubit |i2 entering
the Hadamard gate, and another, H 2 |i2 , coming out.
Finally, rather than a truth table, we will need a matrix to describe the behavior of
the gate. Its action on our qubit would be the result of matrix multiplication (another
27
topic well cover if you havent had it),
1
1
1
1
1
1
1
1
1
H 2 |i2 =
1 1 1
21
1 1 1
1
+++
1
+ .
2 +
+
Again, we see that there is a lot of unlearned symbolism, and not the kind that can be
explained in a few minutes or hours. Well need weeks. But the weeks will be packed
with exciting and useful information that you can apply to all areas of engineering
and science, not just quantum computing.
0.4.2
Algorithms
In quantum computing, we first design a small circuit using the components that are
(or will one day become) available to us. An example of such a circuit in diagram
form (with no explanation offered today) is
|0in
H n
H n
Uf
| {z }
(actual)
|0in
| {z }
(conceptual)
There are access points, A, B and C, to assist with the analysis of the circuit. When
we study these circuits in a few weeks, well be following the state of a qubit as it
makes its way through each access point.
Deterministic
The algorithm may be deterministic, in which case we get an answer immediately.
The final steps in the algorithm might read:
We run the circuit one time only and measure the output.
If we read a zero the function is constant.
If we read any non-zero value, the function is balanced.
This will differ from a corresponding classical algorithm that requires, typically, many
evaluations of the circuit (or in computer language, many loop passes).
28
Probabilistic
Or perhaps our algorithm will be probabilistic, which means that once in a blue moon
it will yield an incorrect answer. The final steps, then, might be:
If the above loop ended after n + T full passes, we failed.
Otherwise, we succeeded and have solved the problem with error probability < 1/2T and with a big-O time complexity of O (n3 ), i.e., in
polynomial time.
Once Again: I dont expect you to know about time complexity or probability
yet. Youll learn it all here.
Whether deterministic or probabilistic, we will be designing circuits and their
algorithms that can do things faster than their classical cousins.
0.5
Perspective
Quantum computing does not promise to do everything better than classical computing. In fact, the majority of our processing needs will almost certainly continue to
be met more efficiently with todays bit-based logic. We are designing new tools for
currently unsolvable problems, not to fix things that are currently unbroken.
0.6
Navigating the Topics
Most students will find that a cover-to-cover reading of this book does not match
their individual preparation or goals. One person may skip the chapters on complex
arithmetic and linear algebr a, while another may devote considerable and I hope
pleasurable time luxuriating in those subjects. You will find your path in one of
three ways:
1. Self-Selection. The titles of chapters and sections are visible in the click-able
table of contents. You can use them to evaluate whether a set of topics is likely
to be worth your time.
2. Chapter Introductions. The first sentence of some chapters or sections may
qualify them as optional, intended for those who want more coverage of a specific topic. If any such optional component is needed in the later volumes
accompanying CS 83B or CS 83C, the student will be referred back to it.
3. Tips Found at the Course Site. Students enrolled in CS 83A at Foothill
College will have access to the course web site where weekly modules, discussion
forums and private messages will contain individualized navigation advice.
Lets begin by learning some math.
29
Chapter 1
Complex Arithmetic
1.1
Complex Numbers for Quantum Computing
I presented the briefest of introductions to the quantum bit, or qubit, defining it as
an expression that combines the usual binary values 0 and 1 (in the guise of
the symbols |0i and |1i). The qubit was actually a combination, or superposition, of
those two binary values, each weighted by numbers and , respectively,
|i
|0i + |1i .
But what are these numbers and ? If you are careful not to take it too seriously,
you can imagine them to be numbers between 0 and 1, where small values mean less
probable or a small dose and larger values mean more probable or a high dose.
So a particular qubit value, call it |0 i, defined to be
|0 i
(.9798) |0i + (.200) |1i ,
would mean a large amount of the classical bit 0 and a small amount of the classical bit 1. Im intentionally using pedestrian terminology because its going to take
us a few weeks to rigorously define all this. However, I can reveal something immediately: real numbers will not work for or . The vagaries of quantum mechanics
require that these numbers be taken from the richer pool of complex numbers.
Our quest to learn quantum computing takes us through the field of quantum
mechanics, and the first step in that effort must always be a mastery of complex
arithmetic. Today we check off that box. And even if youve studied it in the past,
our treatment today might include a few surprises results well be using repeatedly
like Eulers formula, how to sum complex roots-of-unity, the complex exponential
function and polar forms. So without further ado, lets get started.
30
1.2
The Field of Complex Numbers
1.2.1
The Real Numbers Just Dont Cut It
We can only go so far in studying quantum information theory if we bind ourselves to
using only the real numbers, R. The hint that R is not adequate comes from simple
algebra. We know the solution of
x2 1
0,
is x = 1. But make the slightest revision,
x2 + 1
and we no longer have any solutions in the real numbers. Yet we need such solutions
in physics, engineering and indeed, in every quantitative field from economics to
neurobiology.
The problem is that the real numbers do not constitute a complete field, a term
that expresses the fact that there are equations that have no solutions in that number
system. We can force the last equation to have a solution by royal decree: we declare
the number
i
1
to be added to R. It is called an imaginary number. Make sure you understand the
meaning here. We are not computing a square root. We are defining a new number
whose name is i, and proclaiming that it have the property that
i2
1 .
i is merely shorthand for the black-and-white pattern of scribbles, 1. But those
scribbles tells us the essential property of this new number, i.
This gives us a solution to the equation x2 + 1 = 0, but there are infinitely
many other equations that still dont have solutions. It seems like we would have to
manufacture a new number for every equation that lacks a real solution (a.k.a. a real
zero in math jargon).
1.2.2
The Definition of C
However, we get lucky. The number i cant just be thrown in without also specifying
how we will respond when someone wants to add or multiply it by a number like
3 or -71.6. And once we do that, we start proliferating new combinations called
complex numbers. Each such non-trivial combination (i.e., one with an i in it) will be
a solutions to some equation that doesnt have a real zero. Here are a few examples
of the kinds of new numbers we will get:
2 + 3i, 1 i, .05 + .002i, + 2i, 52, 100i
31
All of these numbers are of the form
some real number + some real multiple of i,
Terminology
When there is no real part (the first term in the above sum) to the complex number,
it is called purely imaginary, or just imaginary. The last example in the list above is
purely imaginary.
If we take all these combinations, we have a new supersized number system, the
complex numbers, defined by
C a + bi | a R, b R and i = 1 .
1.2.3
The Complex Plane
Since every complex number is defined by an ordered pair of real numbers, (a, b),
where it is understood that
(a, b)
a + bi,
we have a natural way to represent each such number on a plane, whose x-axis is
the real axis (which expresses the value a), and y-axis is the imaginary axis (which
expresses the value b) (figure 1.1). This looks a lot like the real Cartesian plane, R2 ,
Figure 1.1: a few numbers plotted in the complex plane
but it isnt. Both pictures are a collection of ordered pairs of real numbers, but C
is richer than R2 when we dig deeper: it has product and quotient operations, both
missing in R2 . (Youll see.) For now, just be careful to not confuse them.
32
One usually uses z, c or w to denote complex numbers. Often we write the i
before the real coefficient:
z = x + iy,
c = a + ib,
w = u + iv
In quantum computing, our complex numbers are usually coefficients of the computational basis states a new term for those special symbols |0i and |1i we have been
toying with in which case we may use use Greek letters or , for the complex
numbers,
|i
|0i + |1i .
This notation emphasizes the fact that the complex numbers are scalars of the complex vector space under consideration.
[Note. If terms like vector space or scalar are new to you, fear not. Im not
officially defining them yet, and well have a full lecture on them. I just want to start
exposing you to some vocabulary early.]
Equality of Two Complex Numbers
The criteria for two complex numbers to be equal follows the template set by two
points in R2 being equal: both coordinates must be equal. If
z = x + iy
and
w = u + iv,
then
z=w
1.2.4
x=u
and
y = v.
Operations on Complex Numbers
We define what we mean by addition, multiplication, etc. next. We already know
from the definition of i, that
i2 = 1 ,
which tells us the answer to i i. From there, we build-up by assuming that the operations and + obey the same kinds of laws (commutative, associative, distributive)
as they do for the reals. With these ground rules, we can quickly demonstrate that
the only way to define addition (subtraction) and multiplication is
(a + ib) (c + id)
(a c) + i(b d)
and
(a + ib) (c + id)
(ac bd) + i(ad + bc) .
33
The rule for division can actually be derived from these,
a + ib
a + ib
c id
=
c + id
c + id
c id
(ac + bd) + i(bc ad)
c2 + d 2
bc ad
ac + bd
+ i 2
,
where c2 + d2 6= 0 .
= 2
2
c +d
c + d2
A special consequence of this is the oft cited identity
1
= i .
i
[Exercise. Prove it.]
=
Addition (or subtraction) can be pictured as the vectorial sum of the two complex
numbers (see figure 1.2). However, multiplication is more easily visualized when we
Figure 1.2: visualization of complex addition
get to polar coordinates, so well wait a few minutes before showing that picture.
[Exercise. For the following complex numbers,
2 + 3i, 1 i, .05 + .002i, + 2i, 52, 100i,
(a) square them,
(b) subtract 1 i from each of them,
(c) multiply each by 1 i,
(d) divide each by 5,
(e) divide each by i, and
(f) divide any three of them by 1 i. ]
[Exercise. Explain why it is not necessarily true that the square of a complex
number, z 2 , is positive or zero or even real, for that matter. If z 2 is real, does that
mean it must be non-negative?]
34
1.2.5
C is a Field
A number system that has addition and multiplication replete with the usual properties is called a field. What we have outlined above is the fact that C is, like R, a
field. When you have a field, you can then create a vector space over that field by
taking n-tuples of numbers from that field. Just as we have real n-dimensional vector
spaces, Rn , we can as easily create n-dimensional vector spaces over C which we call
Cn . (We have a whole lesson devoted to defining real and complex vector spaces.)
1.3
1.3.1
Exploring the Complex Plane
Complex Numbers as Ordered Pairs of Real Numbers
We already saw that each complex number has two aspects to it: the real term and
the term that has the i in it. This creates a natural correspondence between C and
R2 ,
x + iy
(x, y) .
As a consequence, a special name is given to Cartesian coordinates when applied to
complex numbers: the complex plane.
[Advanced Readers. For those of you who already know about vector spaces,
real and complex, Ill add a word of caution. This is not the same as a complex
vector space consisting of ordered pairs of complex numbers (z, w). The complex
plane consists of one point for every complex number, not a point for every ordered
pair of complex numbers.]
1.3.2
Real and Imaginary Axes and Polar Representation
The axes can still be referred to as the x-axis and y-axis, but they are more commonly
called the real -axis and imaginary axis. The number i sits on the imaginary axis, one
unit above the real axis. The number -3 is three units to the left of the imaginary
axis on the real axis. The number 1 + i is in the first quadrant. [Exercise. Look
up that term and describe what the second, third and fourth quadrants are.]
Besides using (x, y) to describe z, we can use the polar representation suggested
by polar coordinates (r, ) of the complex plane (figure 1.3).
x + iy r (cos + i sin )
35
Figure 1.3: the connection between cartesian and polar coordinates of a complex
number
Terminology
x = Re(z) ,
y = Im(z) ,
r = |z| ,
= arg z ,
3i, i, 900i, etc.
the real part of z
the imaginary part of z
the modulus (magnitude) of z
the argument (polar angle) of z
purely imaginary numbers
Note. Modulus will be discussed more fully in a moment. For now, |z| = r can be
taken as a definition or just terminology with the definition to follow.
1.3.3
Complex Conjugate and Modulus
The Conjugate of a Complex Number
We can apply an operation, complex conjugation, to any complex number to produce
another complex number by negating the imaginary part. If
z
x + iy
then its complex conjugate (or just conjugate) is designated and defined as
z
x iy ,
the complex conjugate of z .
A common alternate notation places a horizontal bar above z (instead of an asterisk
to the upper right),
z
z .
Geometrically, this is like reflecting z across the x (real) axis (figure 1.4).
36
Figure 1.4: conjugation as reflection
Examples
(35 + 8i)
2 i)
(2
(3 2 i)
(1.5)
=
=
=
=
35 8i
2 +
2i
(3 2 i
1.5
It is easy to show that conjugation distributes across sums and products, i.e.,
(w z)
(w + z)
=
=
w z
and
w + z .
These little factoids will come in handy when we study kets, bras and Hermitian
conjugates in a couple weeks.
[Exercise. Prove both assertions. What about quotients?]
The Modulus of a Complex Number
Just as in the case of R2 , the modulus of the complex z is the distance of the line
segment (in the complex plane) 0 z, that is,
p
|z|
Re(z)2 + Im(z)2
p
=
x2 + y 2 .
A short computation shows that multiplying z by its conjugate, z , results in a
37
Figure 1.5: modulus of a complex number
non-negative real number which is the square of the modulus of z.
|z|2
|z|
zz =
zz
zz
x2 + y 2
[Exercise. Fill in any details that are not immediately apparent.]
In these lectures you may have seen (and will continue to see) me square the
absolute value of a number or function. You might have wondered why I bothered to
use absolute value signs around something I was about to square. Now you know: the
value of the number might be complex and simply squaring it would not necessarily
result in a real much less positive number. The square of complex numbers are
not normally real. (If you didnt do the complex arithmetic exercise earlier, do it now
to get a few concrete examples of this phenomenon.)
1.4
Transcendental Functions and Their Identities
The term transcendental function has a formal definition, but for our purposes it
means functions like sin x, cos x, ex , sinh x, etc. Its time to talk about how they are
defined and relate to complex arithmetic.
1.4.1
The Complex Exponential Function Part 1: Pure Imaginary Case
From calculus, you may have learned that the real exponential function, exp(x) = ex ,
can be expressed by some authors, defined in terms of an infinite sum, the Taylor
38
expansion or Taylor series,
exp(x)
ex
X
xn
n=0
n!
This suggests that we can define a complex exponential function that has a similar
expansion, only for a complex z rather than a real x.
Complex Exponential of a Pure Imaginary Number and Eulers Formula
We start by defining a new function of a purely imaginary number, i, where is the
real angle (or arg) of the the number,
exp(i)
X
(i)n
n=0
n!
(A detail that I am skipping is the proof that this series converges to a complex
number for all real . But believe me, it does.)
Lets expand the sum, but first, an observation about increasing powers of i.
i0
i1
i2
i3
i4
i5
=1
=i
= 1
= i
=1
=i
...
n+4
i
= in
Apply these powers of i to the infinite sum.
i
2
i3
4
i
+
1 +
1!
2!
3!
4!
i5
6
i7
i8
+
+
5!
6!
7!
8!
+ ... .
Rearrange the terms so that all the real terms are together and all the imaginary
terms are together,
2
4
6
8
i
e
=
1
+
+
+ ...
2!
4!
6!
8!
3
5
7
+ i
+ ... .
1!
3!
5!
7!
You may recognize the two parenthetical expressions as the Taylor series for cos and
sin , and we pause to summarize this result of profound and universal importance.
39
Eulers Formula
ei
cos + i sin .
Notice the implication,
i 2
e
cos2 + sin2
1,
which leads to one of the most necessary and widely used facts in all of physics and
engineering,
i
e = 1 ,
for real .
[Exercise. Prove this last equality without recourse to Eulers formula, using
exponential identities alone.]
Eulers formula tells us how to visualize the exponential of a pure imaginary. If we
think of as time, then ei is a spec (if you graph it in the complex plane) traveling
around the unit-circle counter-clockwise at 1 radian-per-second (see Figure 1.6). I
Figure 1.6: ei after seconds, @ 1 rad/sec
cannot overstate the importance of Eulers formula and the accompanying picture.
Youll need it in this course and almost in any other engineering study.
40
1.4.2
Real sin() and cos() in Terms of the Complex Exponential
Now try this: Plug (minus theta) into Eulers formula, then add (or subtract) the
resulting equation to the original formula. Because of the trigonometric identities
sin ()
cos ()
sin (),
cos () ,
=
=
and
the so-called oddness and evenness of sin and cos, respectively, you would quickly
discover the first (or second) equality
cos
sin
ei + ei
,
2
ei ei
.
2i
These appear often in physics and engineering, and well be relying on them later in
the course.
1.4.3
Complex Exponential Part 2: Any Complex Number
We have defined exp() for pure imaginaries as an infinite sum, and we already knew
that exp() for reals was an infinite sum, so we combine the two to define exp() for
any complex number. If z = x + iy,
exp(z)
ez
ex+iy
ex eiy .
We can do this because each factor on the far right is a complex number (the first,
of course, happens to also be real), so we can take their product.
For completeness, we should note (this requires proof, not supplied) that everything we have done leads to the promised Taylor expansion for ez as a function of a
complex z, namely,
exp(z)
ez
X
zn
n=0
n!
This definition implies, among other things, the correct behavior of exp(z) with regard
to addition of exponents, that is
exp(z + w)
exp(z) exp(w) ,
or, more familiarly,
ez+w
41
ez ew .
Easy Proof of Addition Law of Sines and Cosines
A consequence of all this hard work is an easy way to remember the trigonometric
addition laws. I hope youre sitting down. Take A, B two real numbers they can
be angles. From Eulers formula, using = A + B, we get:
ei(A+B)
cos(A + B) + i sin(A + B)
On the other hand, we can apply Eulers formula to = A and = B, individually,
in this product:
eiA eiB
=
=
(cos A + i sin A) (cos B + i sin B)
(cos A cos B sin A sin B)
+ i (sin A cos B + cos A sin B)
Because of the law of exponents we know that the LHS of the last two equations are
equal, so their RHSs must also be equal. Finally, equate the real and imaginary parts:
cos(A + B)
sin(A + B)
=
=
cos A cos B sin A sin B
sin A cos B + cos A sin B
QED
1.4.4
The Complex Trigonometric Functions
We wont really need these, but lets record them for posterity. For a complex z, we
have,
cos z
z2
z4
z6
z8
+
+
+ ...
2!
4!
6!
8!
and
sin z
z
z3
z5
z7
+ ... .
1!
3!
5!
7!
We also get an Euler-like formula for general complex z, namely,
eiz
1.4.5
cos z + i sin z .
Polar Relations Expressed Using the Exponential
We are comfortable expressing a complex number as the sum of its real and imaginary
parts,
z
x + iy .
42
But from the equivalence of the Cartesian and polar coordinates of a complex number
seen in figure 1.3 and expressed by
x + iy r (cos + i sin ) ,
we can use the Euler formula on the RHS to obtain the very useful
z
rei .
The latter version gives us a variety of important identities. If z and w are two
complex numbers expressed in polar form,
z
rei
sei ,
and
then we have
zw = rs ei(+)
r i()
e
.
z/w =
s
and
Notice how the moduli multiply or divide and the args add or subtract (figure 1.7 ).
Figure 1.7: multiplication moduli multiply and args add
Also, we can express reciprocals and conjugates nicely, using
1/z
z
1 i
e
r
= rei .
=
and
In fact, that last equation is so useful, Ill restate it slightly. For any real number, ,
rei
= rei .
[Exercise. Using polar notation, find a short proof that the conjugate of a product
(quotient) is the product (quotient) of the conjugates. (This is an exercise you may
have done above using more ink.)]
43
A common way to use these relationships is through equivalent identities that put
the emphasis on the modulus and arg, separately,
|zw|
z
w
|z |
|z| |w| ,
|z|
,
|w|
|z| ,
arg(zw)
z
arg
w
arg z + arg w ,
arg z arg w ,
arg(z )
arg z .
and
[Exercise. Verify all of the last dozen or so polar identities using the results of the
earlier sections.]
1.5
Roots of Unity
An application of complex arithmetic that will be used extensively in Shors algorithm
and the quantum Fourier transform involves the roots of unity.
1.5.1
N Distinct Solutions to z N = 1
Here is something you dont see much in the real numbers. The equation
zN
1,
with a positive integer, N , has either one or two real roots (zeros), depending on
the parity of N . That is, if N is odd, the only zero is 1. If N is even, there are two
zeros, -1 and 1.
In complex algebra, things are different. Were going to see that there are N
distinct solutions to this equation. But first, we take a step back.
Consider the complex number (in polar form),
ei(2/N ) .
N is usually written
e2i/N ,
but I grouped the non-i factor in the exponent so you could clearly see that N was
of the form ei that we just finished studying. Indeed, knowing that = 2/N is a
real number allows us to use Eulers formula,
ei
cos + i sin ,
to immediately conclude the following:
44
N lies somewhere on the unit circle,
1 = e2i = 1,
2 = ei = 1,
3 = e2i/3 = 12 +
4 = ei/2 = i,
3
i,
2
and
for N > 4, as N increases (5, 6, 7, etc.), N marches clockwise along the upper
half of the unit circle approaching 1 (but never reaching it). For example 1000
is almost indistinguishable from 1, just above it.
For any N > 0, we can see that
(N )N
=
=
eiN (2/N ) = e2i
cos 2 + i sin 2 =
1,
earning it the name N th root of unity (sometimes hyphenated N th root-of-unity) as
wells as the symbolic formulation
N
1.
N =
In fact, we should call this the primitive N th root-of-unity to distinguish it from its
siblings which well meet in a moment. Finally, we can see that N is a non-real
(when N > 2) solution to the equation
zN
1.
Often, when we are using the same N and N for many pages, we omit the subscript
and use the simpler,
e2i/N ,
with N understood by context, especially helpful when used in large formulas.
If we fix a primitive N th root-of-unity for some N (think N = 5 or larger), and we
take powers of N , like N , (N )2 , (N )3 , etc., each power is just the rotation on the
unit circle of the previous power, counter-clockwise, by the angle 2/N . [Exercise.
Explain why.] Eventually well wrap around back to 1, producing N distinct complex
numbers,
N
1 = ei(2/N ) , ei2(2/N ) , ei3(2/N ) , . . . , ei2 = 1 .
(See Figure 1.8.) These are also N th roots-of-unity, generated by taking powers of
the primitive N th root.
[Exercise. Why are they called N th roots-of-unity? Hint: Raise any one of them
to the N th power.]
45
Figure 1.8: the fifth roots-of-unity
For our purposes, well consider the primitive N th root-of-unity to be the one
that has the smallest arg (angle), namely, e2i/N , and well call all the other powers,
ordinary, non-primitive roots. However, this is not how mathematicians would make
the distinction, and does not match the actual definition which includes some of the
other roots, a subtlety we can safely overlook.
As you can see, all N of the N th roots are the vertices of a regular polygon
inscribed in the unit circle, with one vertex anchored at the real number 1. This is a
fact about roots-of-unity that youll want to remember. And now we have met all N
of the distinct solutions to the equation z N = 1.
1.5.2
Eulers Identity
By looking at the fourth root-of-unity, you can get some interesting relationships.
ei/2
i,
e2i
1,
1 .
and
(See Figure 1.9.) That last equation is also known as Eulers identity (distinct from
Eulers formula). Sometimes it is written in the form
ei + 1
46
0.
Figure 1.9: Three of the fourth roots-of-unity (find the fourth)
Take a moment to ponder this amazing relationship between the constants, i, e and
.
1.5.3
Summation Notation
Theres a common notation used to write out long sums called the summation notation. Well use it throughout the course, beginning with the next subsection, so lets
formally introduce it here. Instead of using the ellipsis (. . .) to write a long sum, as
in
a1 + a2 + a3 + . . . + an ,
we symbolize like this
n
X
ak .
k=1
The index starting the sum is indicated below the large Greek Sigma, (), and the
final index in the sum is placed above it. For example, if we wanted to start the sum
from a0 and end at an1 , we would write
n1
X
ak .
k=0
A common variation is the infinite sum, written as
ak ,
k=0
47
where the start of the sum can be anything we want it to be: 0, 1, 5 or even .
[Exercise. Write out the sums
1 + 2 + 3 + . . . + 1999 ,
and
0 + 2 + 4 + . . . + 2N ,
using summation notation. ]
1.5.4
Summing Roots-of-Unity and the Kronecker Delta
We finish this tutorial on complex arithmetic by presenting some facts about roots-ofunity that will come in handy in a few weeks. You can take some of these as exercises
to confirm that you have mastered the ability to calculate with complex numbers.
For the remainder of this section, lets use the shorthand that I advertised earlier
and call the primitive N th root-of-unity , omitting the subscript N , which will be
implied.
We have seen that , as well as all of its integral powers,
,
2,
...,
N 1 ,
N = 0 = 1,
are solutions of the N th degree polynomial equation
zN 1
0.
But these N complex numbers are also solutions of another polynomial
(z 1) (z ) z 2 z N 1
= 0.
To see this, just plug any of the N th roots into this equation and see what happens.
Any two polynomials that have exactly the same roots are equal, so
z N 1 = (z 1) (z ) z 2 z N 1 .
By high school algebra, you can easily verify that the LHS can also be factored to
z N 1 = (z 1) z N 1 + z N 2 + . . . + z 2 + z + 1 .
We can therefore equate the RHS of the last two equations (and divide out the
common (z 1)) to yield
z N 1 + . . . + z + 1 = (z ) z N 1 .
This equation has a number of easily confirmed consequences that we will use going
forward. We list each one as an exercise.
48
Exercises
(a) Show that when 0 l < N ,
l(N 1) + l(N 2) + . . . + l + 1
(
N, l = 0
0, 0 < l < N
Hint: Plug l in for z in the last equation.
(b) Show that when N l < N ,
l(N 1) + l(N 2) + . . . + l + 1
(
N, l = 0, N
0, N < l < N, l 6= 0
Hint: Prove that, for all l, l+N = l , and apply the last result.
(c) Show that for any integer l,
l(N 1)
l(N 2)
+ ... + + 1
(
N, l = 0 (mod N )
0, l 6= 0 (mod N )
Hint: Add (or subtract) an integral multiple of N to (or from) l to bring it into
0
the interval [N, N ) and call l0 the new value of l. Argue that l (N 2) = l(N 2) ,
so this doesnt change the value of the sum. Finally, apply the last result.
(d) Show that for 0 j < N and 0 m < N ,
(jm)(N 1) + (jm)(N 2) + . . . + (jm) + 1
(
N, j = m
0, j 6= m
Hint: Set l = j m and apply the last result.
Kronecker Delta
All of these exercises have a common theme: When we add all of the N th rootsof-unity raised to certain powers, there is a massive cancellation causing the sum
to be zero except for special cases. Ill repeat the last result using Kronecker delta
symbol, as this will be very important in some upcoming algorithms. First, what is
the Kronecker delta?
Kronecker Delta. kj , the Kronecker delta, is the mathematical way
to express anything that is to be 0 unless the index k = j, in which case
it is 1,
(
1, if k = j
kj =
.
0, otherwise
49
You will see Kronecker delta throughout the course (and beyond), starting immediately, as I rewrite result (d) using it.
N
1
X
(jm)k
N jm .
k=0
We Are In the Air
Our flight is officially off the ground. Weve learned or reviewed everything needed
about complex arithmetic to sustain ourselves through the unusual terrain of quantum
computing. Theres a second subject that we will want to master, and thatll be our
first fly-over. Look out the window, as it should be coming into view now: linear
algebra.
50
Chapter 2
Real Vector Spaces
2.1
Vector Spaces for Quantum Computing
In each of these lessons we will inch a little closer to a complete understanding of
the qubit. At the moment, all we know about a qubit is that it can be written as a
collection of abstract symbols,
|i
|0i + |1i ,
where last lecture revealed two of the symbols, and , to be complex numbers.
Today, we add a little more information about it.
The qubit, |i, is a vector quantity.
Every vector lives in a world of other vectors, all of which bear certain similarities.
That world is called a vector space, and two of the similarities that all vectors in any
given vector space share are
its dimension (i.e., is it two dimensional? three dimensional? 10 dimensional?),
and
the kinds of ordinary numbers or scalars that support the vector space
operations (i.e., does it use real numbers? complex numbers? the tiny set of
integers {0, 1}?).
In this lecture well restrict our study to real vector spaces those whose scalars
are the real numbers. Well get to complex vector spaces in a couple days. As for
dimension, well start by focusing on two-dimensional vector spaces, and follow that
by meeting some higher dimensional spaces.
51
2.2
Vectors and Vector Spaces
R2 , sometimes referred to as Euclidean 2-space, will be our poster child for vector
spaces, and well see that everything we learn about R2 applies equally well to higher
dimensional vector spaces like R3 (three-dimensional), R4 (four-dimensional) or Rn
(n-dimensional, for any positive integer, n).
With that overview, lets get back down to earth and define the two-dimensional
real vector space R2 .
2.2.1
Standard Equipment: The Axioms
Every vector space has some rules, called the axioms, that define its objects and the
way in which those objects can be combined. When you learned about the integers,
you were introduced to the objects, (. . . 3, 2, 1, 0, 1, 2, 3, . . .) and the rules
(2 + 2 = 4, (3)(2) = 6, etc.). We now define the axioms the objects and rules
for the special vector space R2 .
The Objects
A vector space requires two sets of things to make sense: the scalars and the vectors.
Scalars
A vector space is based on some number system. For example, R2 is built on the
real numbers, R. These are the scalars of the vector space. In math lingo the scalars
are referred to as the underlying field.
Vectors
The other set of objects that constitute a vector space are the vectors. In the case
of R2 they are ordered pairs,
3
500
1
0
r=
, a=
, x
=
, y
=
.
7
1+
0
1
Youll note that I use boldface to name the vectors. Thats to help distinguish a
vector variable name like r, x or v, from a scalar name, likea, x or . Also, we will
3
usually consider vectors to be written as columns like
, not rows like (3, 7),
7
although this varies by author and context.
A more formal and complete description of the vectors in R2 is provided using set
notation,
(
)
x
R2
x, y R .
y
(See figure 2.1.) This is somewhat incomplete, though, because it only tells what the
52
Figure 2.1: A vector in R2
set of vectors is and fails to mention the scalars or rules.
Notation: Vector Transpose. When writing vectors in a paragraph, I may
write (3, 7) when Im being lazy, or (3, 7)t , when Im being good. The latter is
pronounced (3, 7) transpose, and it means turn this row vector into a column
vector. It goes both ways. If I had a column vector and wanted to turn it into a
row-vector, I would tag a superscript ()t onto the column vector.
Vocabulary. When we want to emphasize which set of scalars we are using for a
given vector space, we would say that the vectors space is defined over [insert name
of scalars here]. For example, the vector space R2 is a vector space over the reals (or
over R).
The Rules
Two operations and their properties are required for the above objects.
Vector Addition
We must be able to combine two vectors to produce a third vector,
v + w 7 u .
In R2 this is defined by adding vectors component-wise,
x1
x2
x1 + x2
+
=
.
y1
y2
y1 + y2
(See figure 2.2.)
Vector addition must obey the required rules. They are:
Zero Vector. Every vector space must have a unique zero vector, denoted by
boldface 0, which has the property that v + 0 = 0 + v = v, for all v in
the space. For R2 the zero vector is (0, 0)t .
53
Figure 2.2: Vector addition in R2
Vector Opposites. For each vector, v, there must be an additive inverse,
i.e., a unique v, such that v + (v) = 0. In R2 the additive inverse of (x, y)t
is (x, y)t .
Commutativity and Associativity. The vectors must satisfy the reasonable
conditions v + w = w + v and (v + w) + u = v + (w + u), for all v, w
and u.
Scalar Multiplication
We can scale a vector by a scalar, that is, we can multiply a vector by a scalar
to produce another vector,
cv 7 w .
In R2 this is defined in the expected way,
3
15
5
=
,
7
35
c
x1
y1
=
cx1
cy1
(See figure 2.3.)
Scalar multiplication must obey the required rules. They are:
Scalar Identity. 1v = v, for any vector v.
Associativity. (ab)v = a(bv), for any vector v and scalars a, b.
Distributivity. c(v + w) = cv + cw and (c + d)(v) = cv + dv, for any
vectors v, w and scalars c, d.
[Exercise. Assuming
3
500
r=
, a=
,
7
1+
x
=
1
0
,
(a) verify the associativity of vector addition using these three vectors,
(b) multiply each by the scalar 1/, and
(c) verify the distributivity using the first two vectors and the scalar 1/.]
54
Figure 2.3: Scalar multiplication in R2
2.2.2
Optional Equipment: Inner Products, Modulus and
Orthogonality
The vector spaces we encounter will have a metric a way to measure distances. The
metric is a side effect of a dot product, and we define this optional feature now.
[Caution. The phrase optional equipment means that not every vector space
has an inner product, not that this is optional reading. It is crucial that you master
this material for quantum mechanics and quantum computing since all of our vector
spaces will have inner products and we will use them constantly.]
Dot (or Inner) Product
When a vector space has this feature, it provides a way to multiply two vectors in
order to produce a scalar,
v w 7 c .
In R2 this is called the dot product, but in other contexts it may be referred to as an
inner product. There can be a difference between a dot product and an inner product
(in complex vector spaces, for example), so dont assume the terms are synonymous.
However, for R2 they are, with both defined by
x1
x2
= x1 x 2 + y 1 y 2 .
y1
y2
Inner products can be defined differently in different vector spaces,. However they
are defined, they must obey certain properties to get the title inner or dot product.
I wont burden you with them all, but one that is very common is a distributive
property.
v ( w1 + w2 )
=
55
v w1 + v w2 .
[Exercise. Look-up and list another property that an inner product must obey.]
[Exercise. Prove that the dot product in R2 , as defined above, obeys the distributive property.]
When we get to Hilbert spaces, there will be more to say about this.
Length (or Modulus or Norm)
An inner product, when present in a vector space, confers each vector with a length
(a.k.a. modulus or norm), denoted by either |v| or ||v||. The length (in most situations, including ours) of each vector must be a non-negative real number, even for
complex vector spaces coming later. So,
|v|
0.
For our friend R2 , if
v
x
,
y
then the length of v is defined to be
kvk
vv
s
x
x
y
y
p
x2 + y 2 ,
positive square root, assumed.
Orthogonality
Figure 2.4: Orthogonal vectors
If two vectors, v and w, have a dot product which is zero,
v w
0,
we say that they are orthogonal or mutually perpendicular. In the relatively visualizable spaces R2 and R3 , we can imagine the line segments 0 v and 0 w forming
right-angles with one another. (See figure 2.4.)
56
Examples of Orthogonal Vectors. A well known orthogonal pair is
1
0
,
.
0
1
Another is
1
1
,
.
1
1
[Exercise. Use the definition of dot product to show that each of the two sets of
vectors, listed above, is orthogonal.]
Examples of Non-Orthogonal Vectors. An example of a set that is not
orthogonal is
1
1
,
.
0
1
More Vocabulary. If a set of vectors is both orthogonal and each vector in the
set has unit length (norm = 1), it is called orthonormal.
The set
1
0
,
0
1
is orthonormal, while the set
1.1
0
,
0
1.1
is orthogonal (check it) but not orthonormal.
Some Unexpected Facts
A length seems like it should always be non-negative, but this wont be true in
some common vector spaces of first-year physics (hint: special relativity).
If v 6= 0, we would expect v v > 0, strictly greater than zero. However, not so
when we deal with the mod-2 vector spaces of quantum computing.
These two unexpected situations suggest that we need a term to distinguish the
typical, well-behaved inner products from those that are off-beat or oddball. That
term is positive definiteness.
Positive Definite Property. An inner-product is usually required to be positive
definite, meaning
v v 0, and
57
v 6= 0 kvk > 0 .
When a proposed inner product fails to meet these conditions, it is often not granted
the status inner (or dot) product but is instead called a pairing. When we come
across a pairing, Ill call it to your attention, and we can take appropriate action.
[Exercise. Assume
3
500
r=
, a=
7
1+
and x
=
1
0
.
(a) Compute the dot product of every pair in the group.
(b) Compute the norm of every vector in the group.
(c) Prove that the only vector in R2 which has zero norm is the zero vector, 0.
(d) For each vector, find an (i ) orthogonal and (ii ) non-orthogonal companion.]
2.2.3
A Bigger Vector Space: R3
Even though we have not learned what dimension means, we have an intuition that R2
is somehow two-dimensional: we graph its vectors on paper and it seems flat. Lets
take a leap now and see what its like to dip our toes into the higher dimensional
vector space R3 (over R). R3 is the set of all triples (or 3-tuples) of real numbers,
Figure 2.5: A vector in R3
R3
x
y
x, y, z R .
(This only tells us what the vectors are, but Im telling you now that the scalars
continue to be R.)
58
Examples of vectors in R3 are
2
0
/2
1 , 1 , 100 .
3.5
0
9
Its harder to graph these objects, but it can be done using three-D sketches (See
figure 2.5.).
[Exercise. Repeat some of the examples and exercises we did for R2 for this
richer vector space. In particular, define vector addition, scalar multiplication, dot
products, etc.]
2.2.4
Preview of a Complex Vector Space: C2
Well roll out complex vector spaces in their full glory when we come to the Hilbert
space lesson, but it wont hurt to put our cards on the table now. The most useful
vector spaces in this course will be ones in which the scalars are the complex numbers.
The simplest example is C2 .
Definition. C2 is the set of all ordered
(
x
C2
y
pairs of complex numbers,
)
x, y C .
You can verify that this is a vector space and guess its dimension and how the inner
product is defined. Then check your guesses online or look ahead in these lectures.
All I want to do here is introduce C2 so youll be ready to see vectors that have
complex components.
2.3
Bases for a Vector Space
If someone were to ask,What is the single most widely used mathematical concept is
in quantum mechanics and quantum computing, I would answer, the vector basis.
(At least thats my answer today.) It is used or implied at every turn, so much so,
that one forgets its there. But it is there, always. Without a solid understanding of
what a basis is, how it affects our window into a problem, and how different bases
relate to one another, one cannot participate in the conversation. Its time to check
off the what is a basis box.
Incidentally, we are transitioning now to properties that are not axioms. They are
consequences of the axioms, i.e., we can prove them.
2.3.1
Linear Combination (or Superposition)
Whenever you have a finite set of two or more vectors, you can combine them using
both scalar multiplication and vector addition. For example, with two vectors, v, w,
59
and two scalars, a, b, we can form the vector u by taking
u
av + bw .
Mathematicians call it a linear combination of v and w. Physicists call it a superposition of the two vectors. Superposition or linear combination, the idea is the same.
We are weighting each of the vectors, v and w, by scalar weights, a and b, respectively, then adding the results. In a sense, the two scalars tell the relative amounts
of each vector that we want in our result. (However, if you lean too heavily on that
metaphor, you will find yourself doing some fast talking when your students ask you
about negative numbers and complex scalars, so dont take it too far.)
The concept extends to sets containing more than two vectors. Say we have a
finite set of n vectors {vk } and corresponding scalars {ck }. Now a linear combination
would be expressed either long-hand or using summation notation,
u
=
=
c0 v0 + c1 v1 + . . . + cn1 vn1
n1
X
ck v k .
k=0
Linear combinations are fundamental to understanding vector bases.
[Exercise. Form three different linear combinations of the vectors
3
500
1
r=
, a=
, x
=
,
7
1+
0
showing the original scalars you chose for each one. Reduce each linear combination
into the simplest form you can. ]
[Exercise. Play around with those three to see if you can express any one of
them as a linear combination of the other two. Hint: You can, no matter how you
split them up.]
2.3.2
The Notion of Basis
One can find a subset of the vectors (usually a very tiny fraction of them compared
to the infinity of vectors in the space) which can be used to generate all the other
vectors through linear combinations. When we have such a subset that is, in a sense,
minimal (to be clarified), we call it a basis for the space.
The Natural Basis
In R2 , only two vectors are needed to produce through linear combination all the
rest. The most famous basis for this space is the standard (or natural or preferred )
basis, which Ill call A for now,
1
0
A = {
x, y
} =
,
.
0
1
60
For example, the vector (15, 3)t can be expressed as the linear combination
15
= 15 x
+ 3y
.
3
Note. In the diagram that follows, the vector pictured is not intended to be (15, 3)t .
(See figure 2.6.)
Figure 2.6: A vector expressed as linear combination of x
and y
Notation. It is common to use the small-hat, , on top of a vector, as in x
,
to denote a vector that has unit length. For the time being, well reserve the two
symbols, x
and y
for the vectors in the natural basis of R2 .
Properties of a Basis
Hang on, though, because we havent actually defined a basis yet; we only saw an
example of one. A basis has to have two properties called (i ) linear independence and
(ii ) completeness (or spanning).
(i) Linear Independence. A set of vectors is linearly independent if we cannot
express any one of them as a linear combination of the others. If we can express
one as a linear combination of the others, then it is called a linear dependent
set. A basis must be linearly independent.
We ask whether the following three vectors making up the set A0 is a basis. If
so, it must be linearly independent. Is it?
1
0
3
0
A
,
,
0
1
2
61
Since we can express (3, 2)t as
3
2
3x
+ 2
y,
A0 is not a linearly independent set of vectors.
Restating, to exercise our new vocabulary, (3, 2)t is weighted sum of x
and y
,
where 3 is the weight of the x
, and 2 is the weight of the y
. Thus, it is a
linear combination of those two and therefore linearly dependent on them. The
requirement that every basis be linear independent means that there cannot be
even one vector in the basis which can be expressed as a linear combination of
the others, so A0 cannot be a basis for R2 .
[Exercise. Is it also true that (1, 0)t can be expressed as a linear combination
of the other two, (0, 1)t and (3, 2)t ? Hint: Yes. See if you can fiddle with the
equation to find out what scalars are needed to get (1, 0)t as a linear combination
of the other two.]
[Exercise. Show that the zero vector (0, 0)t is never part of a linearly independent set. Hint: Find the coefficients of the remaining vectors that easily
produce (0, 0)t as a linear combination.]
(ii) Completeness (or Spanning). Completeness of a set of vectors means that
the set spans the entire vector space, which in turn means that any vector in
the space can be expressed as a linear combination of vectors in that set. A
basis must span the space.
Does the following set of two vectors, A00 , span R3 ?
0
1
00
0 ,
1
A
0
0
Since we cannot express (3, 1, 2)t as a linear combination of these two, i.e.,
3
1
0
1, 6= x 0 + y 1
2
0
0
for any x, y R, A00 does not span the set of vectors in R3 . In other words, A00
is not complete.
[Exercise. Find two vectors, such that if either one (individually) were added to
A , the augmented set would span the space.]
00
[Exercise. Find two vectors, such that, if either one were added to A00 , the
augmented set would still fail to be a spanning set.]
62
Definition of Basis
We now know enough to formally define a vector space basis.
Basis. A basis is a set of vectors that is linearly independent and
complete (spans the space).
Another way to phrase it is that a basis is a minimal spanning set, meaning we
cant remove any vectors from it without losing the spanning property.
Theorem. All bases for a given vector space have the same number of elements.
This is easy to prove (you can do it as an [exercise] if you wish). One consequence
is that all bases for R2 must have two vectors, since we know that the natural basis
has two elements. Similarly, all bases for R3 must have three elements.
Definition. The dimension of a vector space is the number of elements in any
basis.
[Exercise. Describe some vector space (over R) that is 10-dimensional. Hint:
The set of five-tuples, {(x0 , x1 , x2 , x3 , x4 )t xk R } forms a five-dimensional vector
space over R.]
Here is an often used fact that we should prove.
Theorem. If a set of vectors {vk }, is orthonormal, it is necessarily linearly
independent.
Proof. Well assume the theorem is false and arrive at a contradiction. So, we
pretend that {vk } is an orthonormal collection, yet one of them, say v0 , is a linear
combination of the others,
v0
n1
X
ck v k ,
k=1
where not all the ck can be 0, since that would imply that v0 = 0, which cannot
be a member of any orthonormal set (remember from earlier?). By orthonormality
v0 vk = 0 for all k 6= 0, but of course v0 v0 = 1, so we get the following chain of
equalities:
!
n1
n1
X
X
1 = v0 v0 = v0
ck v k
=
ck (v0 vk ) = 0 ,
k=1
a contradiction.
k=1
QED
Notice that even if the vectors were orthogonal, weaker than their being orthonormal, they would still have to be linearly independent.
63
Alternate Bases
There are many different pairs of vectors in R2 which can be used as a basis. Every
basis has exactly two vectors (by our theorem). Here is an alternate basis for R2 :
1
4
B = {b0 , b1 } =
,
.
1
1
For example, the vector (15, 3)t can be expressed as the linear combination
15
= (27/5) b0 + (12/5) b1 .
3
[Exercise. Multiply this out to verify that the coefficients, 27/5 and 12/5 work for
that vector and basis.]
And here is yet a third basis for R2 :
C
{c0 , c1 }
2/2 , 2/2
,
2/2
2/2
with the vector (15, 3)t expressible as the linear combination
15
=
9 2 c0 + 6 2 c1 .
3
[Exercise. Multiply this out to verify that the coefficients, 9 2 and 6 2 work for
that vector and basis.]
Note. In the diagrams, the vector pictured is not intended to be (15, 3)t . (See
figures 2.7 and 2.8.)
Figure 2.7: A vector expressed as linear combination of b0 and b1
64
Figure 2.8: A vector expressed as linear combination of c0 and c1
Expanding Along a Basis. When we want to call attention to the specific basis
we are using to express (i.e., to produce using the wrench of linear combination) a
vector, v, we would say we are expanding v along the so-and-so basis, e.g., we are
expanding v along the natural basis, or lets expand v along the B basis.
Orthonormal Bases
We are particularly interested in bases whose vectors have unit length and are mutually perpendicular. Not surprisingly, because of the definition of orthonormal vectors,
above, such bases are called orthonormal bases.
A is orthonormal:
B is not orthonormal:
C is orthonormal:
x
x
=y
y
= 1,
b0 b0 = 2 6= 1,
([Exercise])
x
y
=0
b0 b1 = 3 6= 0
If they are mutually perpendicular but do not have unit length, they are almost as
useful. Such a basis is called an orthogonal basis. If you get a basis to be orthogonal,
your hard work is done; you simply divide each basis vector by its norm in order to
make it orthonormal.
2.3.3
Coordinates of Vectors
As a set of vectors, R2 consists literally of ordered pairs of real numbers, { (x, y) } or
if you like, { (x, y)t }. Those are the vector objects of the vector space. However, each
vector also has coordinates relative to some basis, namely the specific scalars needed
to express that vector as a linear combination of the basis vectors. For example, if v
is expanded along the natural basis,
v
vx x
+ vy y
,
65
its coordinates relative to the natural basis are vx and vy . In other words, the coordinates are just weighting factors needed to expand that vector in that given basis.
Vocabulary. Sometimes the term coefficient is used instead of coordinate.
If we have a different basis, like B = {b0 , b1 }, and we expand v along that basis,
v
v0 b0 + v1 b1 ,
then its coordinates (coefficients) relative to the B basis are v0 and v1 .
As you see, the same vector, v, will have different coordinates relative to different
bases. However, the ordered pair that describes the pure object the vector, (x, y)t ,
itself does not change.
For some vector spaces, R2 being a prime example, it is easy to confuse the
coefficients with the actual vector. Relative to the natural (standard ) basis, the
coordinates happen to be the same as the numbers in the vector itself, so the vector
and the coordinates, when written in parentheses, are indistinguishable. If we need
to clarify that we are expressing coordinates of a vector, and not the raw vector itself,
we can label the ordered pair appropriately. Such a label would be the name of the
basis being used. Its easier to do than it is to say. Here is the vector (15, 3)t , first
expanded along the natural basis,
15
= 15 x
+ 3y
,
3
and now shown along with its coordinates in the standard basis,
15
15
=
.
3
3 A
For the non-preferred bases of the previous section, the coordinates for (15, 3), similarly labeled, are written as
15
27/5
9 2
=
=
3
12/5 B
6 2 C
Computing Expansion Coefficients along a Basis
In physics, it is common to use the term expansion coefficients to mean the coordinates
relative to some basis. This terminology seems to be used especially if the basis
happens to be orthonormal, although you may see the term used even if it isnt.
15
=
9 2 c0 + 6 2 c1 ,
3
so 9 2 and 6 2 are the expansion coefficients of
15
3
66
along the C basis.
If our basis happens to be orthonormal, there is a special trick we can use to find
the coordinates or expansion coefficients relative to that orthonormal basis.
Lets say we would like to expand v along the C basis, but we only know it in the
natural basis, A . We also know the C basis vectors, c0 and c1 , in the natural basis
A. In other words, we would like, but dont yet have,
v = 0 c0 + 1 c1 ,
i.e., we dont know the scalar weights 0 , 1 . An easy way to find the k is to dot
v with the two C basis vectors. We demonstrate why this works by computing 0 :
c0 v =
=
=
=
c0
c0
0
0
(0 c0 + 1 c1 )
0 c 0 + c 0 1 c 1
(c0 c0 ) + 1 (c0 c1 )
(1) + 1 (0) = 0
[Exercise. Justify each equality in this last derivation using the axioms of vector
space and assumption of orthonormality of C.]
[Exercise. This trick works almost as well with an orthogonal basis which does
not happen to be orthonormal. We just have to add a step; when computing the
expansion coefficient for the basis vector, ck , we must divide the dot product by |ck |2 .
Prove this and give an example.]
Thus, dotting by c0 produced the expansion coefficient 0 . Likewise, to find 1
just dot the v with c1 .
For the specific vector v = (15, 3)t and the basis C, lets verify that this actually
works for, say, the 0th expansion coefficient:
2/2 15
= 15 2/2 + 3 2/2 = 9 2 X
c0 v =
3
2/2
The reason we could add the check-mark is that this agrees with the expression we
had earlier for the vector v expanded along the C basis.
Remark. We actually did not have to know things in terms of the natural basis
A in order for this to work. If we had known the coordinates of v in some other basis
(it doesnt even have to be orthonormal), say B, and we also knew coordinates of the
C basis vectors with respect to B, then we could have done the same thing.
[Exercise. If youre up for it, prove this.]
2.3.4
Independence of Basis (or Not?)
The definition of inner product assumed that the n-tuples were, themselves, the
vectors and not some coordinate representation expanded along a specific basis. Now
67
that we know a vector can have different coordinates relative to different bases, we
ask is the inner product formula that I gave independent of basis?, i.e., can we
use the coordinates, (dx , dy ) and (ex , ey ) relative to some basis rather than the
numbers in the raw vector ordered pair to compute the inner product using the
simple dx ex + dy ey ? In general the answer is no.
[Exercise. Compute the length of the vector (15, 3)t by dotting it with itself. Now
do the same thing, but this time compute using that vectors coordinates relative to
the three bases A, B and C through use of the imputed formula given above. Do you
get the same answers? Which bases coordinates give the right inner-product answer?]
However, when working with orthonormal bases, the answer is yes, one can use
coordinates relative to that basis, instead of the pure vector coordinates, and apply
the simple formula to the coordinates.
[Exercise. Explain the results of the last exercise in light of this new assertion.]
[Exercise. See if you can prove the last assertion.]
Note. There is a way to use non-orthonormal basis coordinates to compute dot
products, but one must resort to a more elaborate matrix multiplication, the details
of which we shall skip (but its a nice [Exercise] should you wish to attempt it).
2.4
Subspaces
The set of vectors that are scalar multiples of a single vector, such as
2
a
=
a
aR ,
aR
1
.5a
is, itself, a vector space. It can be called a subspace of the larger space R2 . As
an exercise, you can confirm that any two vectors in this set, when added together,
produce a third vector which is also in the set. Same with scalar multiplication. So
the subspace is said to be closed under vector addition and scalar multiplication. In
fact, thats what we mean by a subspace.
Vector Subspace. A subspace of a (parent) vector space is a subset of
the parent vectors that is closed under the vector/scalar operations.
2.5
Higher Dimensional Vector Spaces
What we just covered lays the groundwork for more exotic yet commonly used
vector spaces in physics and engineering. Thats why I first listed the ideas and facts
in the more familiar setting of R2 (and R3 ). If you can abstract these ideas, rather
than just memorize their use in R2 and R3 , it will serve you well, even in this short
quantum computing course.
68
Axioms
We can extend directly from ordered pairs or triples to ordered n-tuples for any
positive integer n:
x0
x1
n
R
.. , xk R, k = 0 to n 1
n1
Set n = 10 and you are thinking about a 10-dimensional space.
The underlying field for Rn is the same real number system, R, that we used
for R2 . The components are named x0 , x1 , x2 , x3 , . . ., instead of x, y, z, . . .
(although, when you deal with relativity or most engineering problems, youll use the
four-dimensional sybols x, y, z and t, the last meaning time).
[Exercise. Define vector addition and scalar multiplication for Rn following the
lead set by R2 and R3 .]
Inner Product
The dot product is defined as you would expect. If
a0
a1
and
b =
a = ..
.
an1
b0
b1
..
.
bn1
then
ab
n
X
ak b k .
k=1
[Exercise. Compute the length of the vector (1, 1, 1, . . . , 1)t in Rn .]
Basis
The standard (or natural or preferred ) basis in Rn is
1
0
0 1
A =
.. , .. , . . . ,
. .
0
0
= {x
0 , x
1 , . . . , x
n1 } .
69
0
..
.
It is clearly orthonormal (why clearly?), and any other basis,
B
{ b0 , b1 , . . . , bn1 } ,
for Rn will therefore have n vectors. The orthonormal property would be satisfied by
the alternate basis B if and only if
bk bj = kj .
We defined the last symbol, kj , in our complex arithmetic lesson, but since many of
our readers will be skipping one or more of the early chapters, Ill reprise the definition
here.
Kronecker Delta. kj , the Kronecker delta, is the mathematical way
to express anything that is to be 0 unless the index k = j, in which case
it is 1,
(
1, if k = j
.
kj =
0, otherwise
Expressing any vector, v, in terms of a basis, say B, looks like
v =
n
X
k bk ,
k=1
and all the remaining properties and definitions follow exactly as in the case of the
smaller dimensions.
Computing Expansion Coefficients
I explained how to compute the expansion coefficient of an arbitrary vector along
an orthonormal basis. This move is done so frequently in a variety of contexts in
quantum mechanics and electromagnetism that it warrants being restated here in the
more general cases.
When the basis is orthonormal, we can find the expansion coefficients for a vector
v =
n
X
k bk ,
k=1
by dotting v with the basis vectors one-at-a-time. In practical terms, this means we
dot with bj to get j ,
bj v = bj
n
X
(k bk ) =
k=1
n
X
k (bj bk ) =
k=1
n
X
k=1
n
X
k=1
70
bj (k bk )
k jk = j .
2.6
More Exercises
Prove any of following that interest you:
1. If you have a spanning set of vectors that is not linearly independent, you can
find a subset that is a basis. (Argue why/how.) Give an example that demonstrates you cannot arbitrarily keep the correct number of vectors from the
original spanning set and expect those to be your basis; you have to select your
subset, carefully.
2. Demonstrate that the basis C, in the examples above, is orthonormal.
3. The Gram-Schmidt Process (Tricky). Starting with an arbitrary basis of a
subspace, you can construct an orthonormal basis which spans the same set as
your arbitrary basis. This is not the same as selecting a subset as in exercise 1,
but involves creating new vectors systematically from the existing basis. Hint:
You can keep any of the original vectors, as is, but from there, you might not
be able to keep any of the other vectors in the final, orthonormal, set. Use what
you know about dot products, perpendicular vectors and projection of vectors.
Draw pictures.
4. Show that the set of vectors in R2 which have length < 1 does not constitute a
vector subspace R2 .
5. Prove that the set of points on the line y = 3x 1 does not constitute a vector
subspace of R2 .
6. Argue that all one-dimensional vector subspaces of R2 consist of lines through
the origin.
7. Argue that all two-dimensional subspaces of R3 consist of planes through the
origin.
And that does it for our overview of vector spaces. The extension of these concepts
to the more exotic sounding vector spaces head complex vector spaces, Hilbert spaces
and spaces over the finite field Z2 will be very easy if you understand whats in this
section and have done several of the exercises.
71
Chapter 3
Matrices
3.1
Matrices in Quantum Computing
If a qubit is the replacement for the classical bit, what replaces the classical logic
gate? I gave you a sneak peek at the answer in the introductory lesson. There, I
first showed you the truth table of the conventional logic gate known to all computer
science freshmen as the AND gate (symbol ).
x
xy
Then, I mentioned that in the quantum world these truth tables get replaced by
something more abstract, called matrices. Like the truth table, a matrix contains
the rules of engagement when a qubit steps into its foyer. The matrix for a quantum
operator that well study later in the quarter is
1 1
1
1
1
1 1 1 1 ,
1 1 1
21
1 1 1
1
which represents a special gate called the second order Hadamard operator. Well
meet that officially in a few weeks.
Our job today is to define matrices formally and learn the specific ways in which
they can be manipulated and combined with the vectors we met the previous chapter.
72
3.2
Definitions
Definition of a Matrix. A matrix is a rectangular array of numbers,
variables or pretty much anything. It has rows and columns. Each
matrix has a particular size expressed as [# rows] [# columns],
for example, 2 2, 3 4, 7 1, 10 10, etc.
A 3 4 matrix (call it A) could be:
1 2 3 4
5 6 7 8
9 10 11 12
A 4 2 matrix (call it B) might be:
1
3
5
7
3.2.1
2
4
6
8
Notation
For this lesson I will number starting with 1, not 0, so
row 1
column 1
the first row
=
the top row of the matrix, and
the first column
the left column of the matrix.
I usually name matrices with non-boldface capital letters, A, B, etc. Occasionally,
I will use boldface, A, B. The context will always be clear.
Sometimes youll see the matrix represented abstractly as the (kl)th component
surrounded by parentheses,
Akl ,
which represents the entire matrix A, whose value in row k and column l is Akl . This
doesnt tell you the size, but youll always know that from the context.
Ill typically use m n or p q to describe the size of some general matrix and
n n or p p when I want to specify that we are dealing with a square matrices, i.e.,
one that has same number of rows as columns.
3.3
Matrix Multiplication
Matrix multiplication will turn out to be sensitive to order. In math lingo, it is not
a commutative operation,
AB 6= BA .
73
Therefore, its important that we note the order of the product in the definition.
First, we can only multiply the matrices A (n p) and B (q m) if p = q. So we will
only define the product AB for two matrices of sizes, A (n p) and B (p m). The
size of the product will be n m. Symbolically,
(n p) (p m)
(n m) .
Note that the inner dimension gets annihilated, leaving the outer dimensions to
determine the size of the product.
3.3.1
Row Column
We start by defining the product of a row by a column, which is just a short way of
saying a (1 l) matrix times an (l 1) matrix, i.e., two 1-dimensional matrices. It
is the simple dot product of the two entitles as if they were vectors,
5
6
1, 2, 3, 4
7 = (1)(5) + (2)(6) + (3)(7) + (4)(8)
8
= 18 .
This is the definition of matrix multiplication in the special case when the first is a
column vector and the second is a row vector. But that definition is used repeatedly
to generate the product of two general matrices, coming up next. Heres an example
when the vectors happen to have complex numbers in them.
5
6
1, i, 3, 2 3i
4i = (1)(5) + (i)(6) + (3)(4i) + (2 3i)(8i)
8
= 19 + 10i .
[For Advanced Readers. As you see, this is just the sum of the simple products
of the coordinates, even when the numbers are complex. For those of you already
familiar with complex inner products (a topic we will cover next week), please note
that this is not a complex inner product; we do not take the complex conjugate of either
vector. Even if the matrices are complex numbers, we take the ordinary complex
product of the corresponding elements and add them.]
3.3.2
Definition of Matrix Multiplication
The full definition of two matrices AB is more-or-less forced on us by everything we
have said so far. Based on the required size of the product and the definition of two
74
1-dimensional matrices, we must define the (kl)th element of the answer matrix to
be the dot product of the kth row of A with the lth column of B. Lets look at it
graphically before we see the formal definition.
We illustrate the computation of the 1-1 element of an answer matrix in Figure 3.1
and the computation of the 2-2 element of the same answer matrix in Figure 3.2.
Figure 3.1: Dot-product of the first row and first column yields element 1-1
Figure 3.2: Dot-product of the second row and second column yields element 2-2
The Formal Definition of Matrix Multiplication. If A is an n p matrix,
and B is a p m matrix, then C = AB is an n m matrix, whose (kl)th element is
given by
Ckl = (AB)kl
p
X
j=1
where k = 1, . . . , n and l = 1, . . . , m.
75
Akj Bjl ,
(3.1)
[Exercise. Fill in the rest of the elements in the product matrix, C, above.]
[Exercise. Compute the products
1 2 1
1 2 3
2 0 4 3 1 1
5 5 5
0 0 1
and
1 2
1
2 0 1
2 1 1 4 3 1
0 0
5
0 5 0
1 1
3
1
.
1
1
[Exercise. Compute the first product above in the opposite order. Did you get
the same answer?]
While matrix multiplication may not be commutative, it is associative,
A(BC) = (AB)C,
a property that is used frequently.
3.3.3
Product of a Vector by a Matrix
When a product happens to have an n 1 matrix in the second position, it is usually
called the product of a vector by a matrix because an n 1 matrix looks just like an
n- dimensional vector. So, if v is an n-dimensional vector and A is an n n matrix,
we already know what to do if we see Av. For example,
1 2i 3
4
4
3 1 1 2.5i = 11 + 2.5i .
0 0 1
1
1
Some Observations
Position. You cannot put the vector on the left of the matrix; the dimensions
would no longer be compatible. You can, however, put the transpose of the
vector on the left of the matrix, vt A. That does make sense and it has an
answer (see exercise, below).
Linear Transformation. Multiplying a vector by a matrix produces another
vector. It is our first example of something called a linear transformation,
which is a special kind of mapping that sends vectors to vectors. Our next
lecture covers linear transformations in depth.
76
[Exercise. Using the v and A from the last exercise, compute the product vt A.
[Exercise. Using the same A as above, let the vector w (1, 2.5i, 1)t and
compute the product Aw, then
compute A(v + w), and finally
compare A(v + w) with Av + Aw. Is this a coincidence? ]
3.4
Matrix Transpose
We briefly covered the transpose, vt , of a vector v. This is an action that converts
column vectors to row vectors and vice versa,
2
t
i
2, i, 3
=
3
t
1
2 = (1, 2, 3)
3
That was a special case of the more general operation, namely, taking the transpose
of an entire matrix. The transpose operation creates a new matrix whose rows are
the columns of the original matrix (and whose columns are the rows of the original).
More concisely, if A is the name of our original n m matrix, its transpose, At , is the
m n matrix defined by
( Akl )t
( Alk ) .
In other words, the element in position kl of the original is moved to position lk of
the transpose.
Examples:
1 2
5 6i
9 10
1
2
2 1
5
0
1
3 4
2
7 8 =
3
11 12
4
t
1
0 i
2
1 4 =
0
5 0
i
t
5 9
6i 10
7 11
8 12
2 5
1 0
1 5
4 0
[Exercise. Make up two matrices, one square and one not square, and show the
transpose of each.]
77
3.5
Matrix Addition and Scalar Multiplication
Matrices can be added (component-wise) and multiplied by a scalar (apply the scalar
to all nm elements in the matrix). Im going to let you be the authors of this section
in two short exercises.
[Exercise. Make these definitions precise using a formula and give an example of
each in a 3 3 case.]
[Exercise. Show that matrix multiplication is distributive over addition and both
associative and commutative in combination with scalar multiplication, i.e.,
A (c B1 + B2 )
3.6
=
=
c (AB1 ) + (AB2 )
(AB2 ) + c (AB1 ) . ]
Identity Matrix and Zero Matrix
Zero. A matrix whose elements are all 0 is called a zero matrix, and can be written
as 0 or ( 0 ), e.g.,
0 0
0 0 0
0 0
0 = 0 0 0
or
(0) =
0 0
0 0 0
0 0
Clearly, when you add the zero matrix to another matrix, it does not change anything
in the other matrix. When you multiply the zero matrix by another matrix, including
a vector, it will squash it to all 0s.
( 0 )A = ( 0 )
and
( 0 )v = 0
In words, the zero matrix is the additive identity for matrices.
Identity (or Unit). A square matrix whose diagonal elements (upper-left to
lower-right) are all 1 and whose off-diagonals are all 0 is called an identity matrix
or unit matrix, and can be written variously as I, Id, 1, 1 or 1, e.g.,
1 0 0 0
1 0 0
0 1 0 0
1 = 0 1 0
or
1 =
0 0 1 0
0 0 1
0 0 0 1
The identity matrix has the property that when you apply it to (multiply it with)
78
another matrix or vector, it leaves that matrix or vector unchanged, e.g.,
1 0 0 0
1
2i 0 1
1
2i 0 1
0 1 0 0 2 1 1
4
4
= 2 1 1
0 0 1 0 5
5
0
5
0
0
5
0
0 0 0 1
3 i 2 2 2
3 i 2 2 2
1 0 0 0
1
2i 0 1
1
2i 0 1
2 1 1
4
4
0 1 0 0 = 2 1 1
5
5
0
5
0 0 0 1 0
0
5
0
3 i 2 2 2
0 0 0 1
3 i 2 2 2
Notice that multiplication by a unit matrix has the same (non) effect whether it
appears on either side of its co-multiplicand. The rule for both matrices and vectors
is
1M = M 1 = M,
1 v = v and
vt 1 = vt .
In words, the unit matrix is the multiplicative identity for matrices.
3.7
Determinants
Associated with every square matrix is a scalar called its determinant. There is very
little physics, math, statistics or any other science that we can do without a working
knowledge of the determinant. Lets check that box now.
3.7.1
Determinant of a 2 2 Matrix
Consider any (real or complex) 2 2 matrix,
a b
A =
c d
The determinant of A is defined (and written) as follows,
a b
ad bc .
det(A) =
c d
For example,
i 1+i
6 5i
=
=
(i)(5i) (1 + i)( 6)
5
6 i 6
(5 6) i 6 .
79
[Exercise. Compute the following determinants:
1 2
3 4 = ?
4
1
+
i
= ?
1i
2
3.7.2
Determinant of a 3 3 Matrix
Well give an explicit definition of a 3 3 matrix and this will suggest how to proceed
to the n n case.
a b c
d e f a e f b d f + c d e
g h
h i
g i
g h i
= a minor of a
b minor of b
+ c minor of c
(Sorry, I had to use the variable name i, not for the 1, but to mean the 3-3 element
of the matrix, since I ran out of reasonable letters.) The latter defines the minors of
a matrix element to be the determinant of the smaller matrix constructed by crossing
out that elements row and column (See Figure 3.3.)
Figure 3.3: Minor of a matrix element
[Exercise. Compute the following determinants:
1 2 3
2 1 2 = ?
1 3 4
1
0
0
0
4
1 + i = ?
0 1i
2
Compare the last answer with the last 2 2 determinant exercise answer. Any
thoughts?]
80
3.7.3
Determinant of an n n Matrix
The 33 definition tells us to proceed recursively for any square matrix of size n.
We define its determinant as an alternating sum of the first row elements times their
minors,
det(A) = A
A11 minor of A11
A12 minor of A12
+ A13 minor of A13
+
n
X
k+1
=
(1) A1k minor of A1k .
k=1
I think you know whats coming. Why row 1 ? No reason at all (except that every
square matrix has a first row). In the definition above we would say that we expanded
the determinant along the first row, but we could have expanded it along any row
or any column, for that matter and gotten the same answer.
However, there is one detail that has to be adjusted if we expand along some other
row (or column). The expression
(1)k+1
has to be changed if we expand along the jth row (column) rather than the first row
(column). The 1 above becomes j,
(1)k+j ,
giving the formula
det(A)
n
X
k+j
(1) Ajk minor of Ajk ,
k=1
good for any row (column) j.
[Exercise. Rewrite the definition of an n n determinant expanded along the
3rd column; along the kth column.]
[Exercise. Expand one of the 22 determinants and one of the 33 determinants
in the exercises above along a different row or column and confirm that you get the
same answer as before.]
[Exercise. If you dare, prove that the determinant is independent of the choice
of row/column used to expand it.]
3.7.4
Determinants of Products
det(AB)
det(A) det(B) .
We dont need to prove this or do exercises, but make a mental note as we need it in
the following sections.
81
3.8
Matrix Inverses
Since we have the multiplicative identity (i.e., 1) for matrices we can ask whether,
given an arbitrary square matrix A, the inverse of A can be found. That is, can we
find a B such that AB = BA = 1? The answer is sometimes. Not all matrices have
inverses. If A does have an inverse, we say that A is invertible or non- singular, and
write its inverse as A1 . Shown in a couple different notations, just to encourage
flexibility, a matrix inverse must satisfy (and is defined by)
M 1 M
A1 A
=
=
M M 1 = 1 or
A A1 = I .
Here is one of two little theorems that well need when we introduce quantum
mechanics.
Little Inverse Theorem A. If M v = 0 for some non-zero vector
v, then M has no inverse.
Proof by Contradiction. Well assume that M 1 exists and reach a contradiction. Let v be some non-zero vector that M sends to 0. Then,
1
v = Iv =
M M v = M 1 (M v) = M 1 0 = 0,
contradicting the choice of v 6= 0.
QED
[Exercise. Prove that can be strengthened to in the above theorem. That
is, prove the somewhat trickier direction: if M does not have an inverse, there
exists a non-zero v satisfying M v = 0.]
The condition that a matrix be invertible is closely related to its determinant.
Big Inverse Theorem. A matrix is non-singular (invertible) its determinant 6= 0.
While we dont have to prove this, its actually fun and easy. In fact, we can get
half way there in one line:
[Exercise. Prove that M non-singular det(M ) 6= 0 (in a single line). Hint:
Use the fact about determinants of products.]
The other direction is a consequence of a popular result in matrix equations. Since
one or more of our quantum algorithms will refer to this result, well devote the next
section to it and do an example. Before we leave this section, though, I need to place
into evidence, our second little theorem.
Little Inverse Theorem B. If M v = 0 for some non- zero vector v, then
det(M ) = 0.
Proof. Little Theorem A tells us that the hypothesis implies M has no inverse,
and the subsequent Big Theorem tells us that this forces det(M ) 6= 0.
QED
82
3.9
3.9.1
Matrix Equations and Cramers Rule
Systems of Linear Equations
A system of simultaneous linear equations with n unknowns is a set of equations in
variables x1 , x2 , . . . , xn , which only involves sums first order (linear ) terms of each
variable, e.g.,
x1
4x2 x5 + 3.2x1
+ x2 + x3 + x4 + x5
10x4 22x1
=
=
=
19 + 5x3
1
85 x2 + x1
To solve the system uniquely, one needs to have exactly at least as many equations as
there are unknowns. So the above system does not have a unique solution (although
you can find some simpler relationships between the variables if you try). Even if you
do have exactly the same number of equations as unknowns, there still may not be a
unique solution since one of the equations might to cite one possibility be a mere
multiple of one of the others, adding no new information. Each equation has to add
new information it must be independent of all the others.
Matrix Equations
We can express this system of equations concisely using the language of matrix multiplication,
x1
19
3.2 4 5 0 1
x
2
1 1 1 1 1 x3 = 1 .
23 1 0 10 0 x4
85
x5
If we were able to get two more relationships between the variables, independent of
these three, we would have a complete system represented by a square 5 5 matrix
on the LHS. For example,
3.2 4 5 0 1
x1
19
1
1 1
1
1 x2
1
23 1 0 10
0
x3 = 85 .
2.5 .09 50
0
1 x4
2 0 .83 1 17
x5
4
Setting M = the 5 5 matrix on the left, v = the vector of unknowns, and c = the
vector of constants on the right, this becomes
Mv
=
83
c.
How can we leverage the language of matrices to get a solution? We want to know
what all the xk are. Thats the same as having a vector equation in which v is all
alone on the LHS,
v
d,
and the d on the RHS is a vector of constants. The scent of a solution should be
wafting in the breeze. If M v = c were a scalar equation we would divide by M .
Since its a matrix equation there are two differences:
1. Instead of dividing by M , we multiply each side of the equation (on the left) by
M 1 .
2. M 1 may not even exist, so this only works if M is non-singular.
If M is non-singular, and we can calculate M 1 , we apply bullet 1 above,
M 1 M v = M 1 c
1 v = M 1 c
v = M 1 c
This leaves us with two action items.
1. Determine whether or not M is invertible (non-singular).
2. If it is, compute the inverse, M 1 .
We know how to do item 1; if det(M ) is non-zero, it can be inverted. For item 2,
we have to learn how to solve a system of linear equations, something we are perfectly
situated to do given the work we have already done in this lecture.
3.9.2
Cramers Rule
Todays technique, called Cramers rule, is a clean and easy way to invert a matrix,
but not used much in practice due to its proclivity toward round-off error and poor
computer performance. (Well learn alternatives based on so-called normal forms and
Gaussian elimination when we get into Simons and Shors quantum algorithms). But
Cramers rule is very cute, so lets get enlightened.
Cramers Rule. A system of linear equations,
Mv
c,
(like the 5 5 system above) can be solved uniquely for the unknowns
xk det(M ) 6= 0. In that case each xk is given by
det Mk
,
det M
where Mk is the matrix M with its kth column replaced by the constant
vector c (see Figure 3.4).
xk =
84
Figure 3.4: The numerator of Cramers fraction
We will not prove Cramers rule. You can take it on faith, look up a proof, or
have a whack at it yourself.
Example Application of Cramers Rule
We can compute the relevant determinants of the 5 5 system above with the help
of Mathematica or similar software which computes determinants for us. Or, we can
write our own determinant calculator using very few instructions and recursion, and
do it all ourselves. First, we check the main determinant,
det(M )
167077 6= 0 .
So M is non-singular system is solvable. Now for Cramers numerators:
det(M1 )
det(M2 )
det(M3 )
det(M4 )
det(M5 )
=
=
=
=
=
635709
199789
423127
21996.3
176281
Using this, we solve for the first unknown,
x1
635709
167077
3.80488 .
[Exercise. Compute the other four xk and confirm that any one of the equations
in the original system holds for these five values.]
Computing Matrix Inverses Using Cramers Rule
As noted, Cramers rule is not very useful for writing software, but we can apply it
to the problem of finding a matrix inverse, especially when dealing with small, 2 2,
85
matrices by hand. Say we are given such a matrix
a b
M =
,
c d
that we have confirmed to be non-singular by computing ad bc. So we know it has
an inverse, which we will temporarily write as
e f
1
M
=
.
g h
The only thing we can say with certainty,
a b
e
1
MM
=
c d
g
at this point, is that
1 0
f
= 1 =
.
h
0 1
Our goal is to solve this matrix equation. We do so in two parts. First, we break this
into an equation which only involves the first column of the purported inverse,
a b
e
1
=
.
c d
g
0
This is exactly the kind of 2-equation linear system we have already conquered.
Cramers rule tells us that it has a solution (since det(M ) 6= 0) and the solution
is given by
1 b
det
0 d
e =
det M
and
a 1
det
c 0
g =
.
det M
The same moves can be used on the second column of M 1 , to solve
a b
f
0
=
.
c d
h
1
[Exercise. Write down the corresponding quotients that give f and h.]
Example. We determine the invertibility of M and, if invertible, compute its
inverse, where
12 1
M =
.
15 1
The determinant is
12 1
15 1
12 15
86
27 6= 0,
which tells us that the inverse does exist. Setting
e f
1
M
=
,
g h
e
we solve first for the left column
. Following the procedure above,
g
1 1
det
0 1
1
1
e =
=
=
det M
27
27
and
12 1
det
15 0
det M
=
15
,
27
which we dont simplify ... yet. Continuing on to solve for f and h results in the final
inverse matrix
1 1 1
1/27 1/27
1
,
M
=
=
15/27 12/27
27 15 12
and we can see why we did not simply the expression of g.
[Exercise. For the above example, confirm that M M 1 = Id.]
[Exercise. Compute the inverse of
M
1 2
3 4
using Cramers rule. Check your work.]
[Exercise. Show that
1 0 2
3 4 0
2 0 0
is non-singular, then compute its inverse using Cramers rule. Check your work.]
Completing the Proof of the Big Inverse Theorem
Remember this
Big Inverse Theorem. A matrix is non-singular (invertible) its determinant 6= 0.
87
?
Your proved in an exercise. Now you can do in another exercise.
[Exercise. Using Cramers rule as a starting point, prove that
det(M ) 6= 0 M is non-singular .
Hint. We just did it in our little 2 2 and 3 3 matrices. ]
88
Chapter 4
Hilbert Space
4.1
Complex Vector Spaces for Quantum Computing
4.1.1
The Vector Nature of a Qubit
We are gradually constructing a mathematical language that can be used to accurately
model the physical qubit and hardware that processes it. A typical qubit is expressed
using the symbolism
|i
|0i + |1i .
In the sneak peek given in the introduction I leaked the meaning of the two numbers
and . They embody the probabilistic nature of quantum bits. While someone might
have prior knowledge that a qubit with the precise value |0i + |1i (whatever
that means) was sitting in a memory location, if we tried to read the value of that
location that is, if we measured it we would see evidence of neither nor , but
instead observe only a classical bit; our measurement device would register one of two
possible outcomes: 0 or 1. Yet and do play a role here. They tell us our
measurement would be
0
with probability
||2
with probability
||2 .
and
Obviously, theres a lot yet to understand about how this unpredictable behavior can
be put to good use, but for today we move a step closer by adding the following clue:
The qubit shown above is a vector having unit length, and the , are
its coordinates in what we shall see is the vector spaces natural basis
{ |0i , |1i }.
In that sense, a qubit seems to correspond, if not physically at least mathematically,
to a vector in some vector space where and are two scalars of the system.
89
4.1.2
The Complex Nature of a Qubit
We need to know whatever is possible to know about a hypothetical quantum circuit
using paper and pencil so we can predict what the hardware will do before we build
it. Im happy to report that the mathematical language that works in this situation is
one of the most accurate and well tested in the history of science: quantum mechanics.
It behooves me to report that a real vector space like R2 or R3 doesnt work. To make
accurate predictions about qubits and the logic circuitry that entangles them, well
need to allow and to be complex. Thus, our vector space must be defined over a
complex scalar field.
Besides the straightforward consequences resulting from the use of complex arithmetic, well have to define an inner product that differs from the simple dot product
of real vector spaces. At that point well be working in what mathematicians call
complex Hilbert space or simply Hilbert space. To make the correspondence between
our math and the real world complete, well have to add one last tweak: we will consider only vectors of unit length. That is, we will need to work on something called
the projective sphere a.k.a projective Hilbert space.
Today we learn how to manipulate vectors in a complex Hilbert space and the
projective sphere within it.
4.2
The Complex Vector Space, Cn
Since we are secure enough in the vocabulary of real vector spaces, theres no need
to dilly-dally around with 2- or 3-dimensional spaces. We can go directly to the fully
general n-dimensional complex space, which we call Cn .
Its scalars are the complex numbers, C, and its vectors are n-tuples of complex
numbers,
c
0
c
1
n
C
.. , ck C, k = 0 to n 1
n1
Except for the inner product, everything else works like the real space Rn with
the plot twist that the components are complex. In fact, the natural basis is actually
identical to Rn s basis.
[Exercise. Prove that the same natural basis works for both Rn and Cn .]
There are, however, other bases for Cn which have no counterpart in Rn , since
their components can be complex.
[Exercise. Drum up a basis for Cn that has no real counterpart. Then find a
vector in Cn which is not in Rn but whose coordinates relative to this basis are all
real.]
90
[Exercise. Is Rn Cn as a set? As a subspace? Justify your answers.]
4.3
The Complex Inner Product
So, whats wrong with Rn s dot product in the complex case? If we were to define it
as we do in the real case, for example as in R2 ,
x1
x2
= x1 x 2 + y 1 y 2 ,
y1
y2
wed have a problem with lengths of vectors which we want to be 0. Recall that
lengths are defined by dotting a vector with itself (Ill remind you about the exact
details in a moment). To wit, for a complex a,
aa
?
=
n1
X
ak 2
k=0
is not necessarily real, never mind non-negative (try a vector whose components are
all 1 + i). When it is real, it could still be negative:
5i
5i
= 25 + 9 = 16 .
3
3
We dont want to have complex, imaginary or negative lengths typically, so we need
a different definition of dot product for complex vector spaces. We define the product
ab
n1
X
ak b k
n1
X
(ak ) bk .
k=0
k=0
When defined in this way, some authors prefer the term inner product, reserving
dot product for the real vector space analog (although some authors say complex dot
product, so you have to adapt.)
Notation. An alternative notation for the (complex) inner product is sometimes
used,
ha, bi ,
and in quantum mechanics, youll always see
ha | bi .
Caution #1. The complex inner product is not commutative, i.e.,
ha | bi
6=
91
hb | ai .
There is an important relationship between the two, however, namely
ha | bi
hb | ai .
[Exercise. Show that the definition of complex inner product implies the above
result.]
Caution #2. The complex inner product can be defined by conjugating the bk s,
rather than the ak s, and this would produce a different result, one which is the complex
conjugate of our defined inner product. However, physicists and we conjugate the
left-hand vectors coordinates because it produces nicer looking formulas.
Example. Let
1+i
a =
3
1 2i
b =
.
5i
and
Then
ha | bi
=
=
=
=
(1 + i) (1 2i) + (3 ) 5i
(1 i) (1 2i) + (3) 5i
(1 i 2i 2) + 15i
1 + 12i .
[Exercise. Compute hb | ai and verify that it is the complex conjugate of ha | bi.]
More Properties of the Complex Inner Product
In our linear algebra lesson we learned that inner products are distributive. In the
second position, this means
ha | b + b0 i
ha | bi + ha | b0 i ,
and the same is true in the first position. For the record, we should collect two more
properties that apply to the all-important complex inner product.
It is linear in the second position. If c is a complex scalar,
c ha | bi
ha | cbi .
It is anti-linear in the first position. If c is a complex scalar,
c ha | bi
hc a | bi .
Again, this asymmetry is reversed in much of mathematics and engineering. Quantum
mechanics naturally leads to anti-linearity in the left vector position, while you will
normally see it in the right vector position in other fields.
92
4.3.1
Norm and Distance
The inner product on Cn confers it with a metric, that is, a way to measure things.
There are two important concepts that emerge:
1. Norm. The norm of vector, now repurposed to our current complex vector
space, is defined by
v
v
u n1
u n1
uX
uX
2
t
kak
|ak |
= t (ak ) ak .
k=0
k=0
2. Distance. The distance between vectors a and b is
v
u n1
uX
|bk ak |2 .
dist(a, b) kb ak = t
k=0
With the now correct definition of inner product we get the desired behavior when
we compute a norm,
kak
ha | ai
n1
X
(ak ) ak
k=0
n1
X
|ak |2
k=0
0.
Since the length (norm or modulus) of a, kak, is the non-negative square root of this
value, once again all is well: lengths are real and non-negative.
More Notation. You may see the modulus for a vector written in normal (not
bold) face, as in
a
kak .
Another common syntax uses single, rather than double, bars,
|a|
kak .
Its all acceptable.
Example (Continued). The norm (a.k.a. modulus) of a in the earlier example
is
p
p
ha | ai =
(1 + i) (1 + i) + (3 ) 3
a = kak
p
=
(1 i) (1 + i) + (3) 3
p
=
(1 i + i + 1) + 9
=
11 .
[Exercise. Compute the norm of b from that same example, above.]
93
Complex Orthonormality and Linear Independence
We had a little theorem about orthonormal sets in real vector spaces. Its just as true
for complex vector spaces with the complex inner product.
Theorem. If a set of vectors {ak } is orthonormal, it is necessarily linearly independent.
Proof. Well assume the theorem is false and arrive at a contradiction. Say {ak }
is an orthonormal collection, yet one of them, a0 , is a linear combination of the others,
i.e.,
a0
n1
X
ck ak .
k=1
By orthonormality ha0 | ak i = 0 for all k 6= 0, and ha0 | a0 i = 1, so
! +
*
n1
n1
X
X
ck ak
=
ck ha0 | ak i
1 = ha0 | a0 i =
a0
k=1
a contradiction.
4.3.2
0,
k=1
QED
Expansion Coefficients
The dot-product trick for computing expansion coefficients along an orthonormal
basis works in this regime, but we have to be careful due to this slight asymmetry.
If we want to expand v along an orthonormal B = {bk }, we still dot it with the
individual basis vectors. Say the (as yet unknown) coefficients of v in this basis are
( . . . , k , . . . ). We compute each one using
* n1
+
X
hbk | vi =
bk
j bj
j=0
n1
X
j hbk | bj i ,
j=0
and since orthonormality means
hbk | bj i
kj ,
the last sum collapses to the desired k . However, we had to be careful to place our v
on the right side of the inner product. Otherwise, we would not get the kth expansion
coefficient, but its .
[Exercise. Fill in the blank].
Example. In C we expand v =
A
{
e0 ,
e1 }
1+i
1i
=
94
along the natural basis A,
1
0
,
.
0
1
(We do this easy example because the answer is obvious: the coordinates along A
should match the pure vector components, otherwise we have a problem with our
technique. Lets see ...)
We seek
v0
.
v1 A
The dotting trick says
v0
=
=
h
e0 | vi = (e00 ) v0 + (e01 ) v1
1 (1 + i) + 0 (1 i) = 1 + i ,
v1
=
=
h
e1 | vi = (e10 ) v0 + (e11 ) v1
0 (1 + i) + 1 (1 i) = 1 i ,
and
so
1+i
1i
=
1+i
1i
,
A
as expected (X). Of course that wasnt much fun since natural basis vector components were real (0 and 1), thus there was nothing to conjugate. Lets do one with a
little crunch.
1+i
2
Example. In C we expand the same v =
along the basis B,
1i
B
n
o
0 , b
1
b
2/2
i
2/2
,
.
2/2
i 2/2
First, we confirm that this basis is orthonormal (because the dot-product trick only
works for orthonormal bases).
D E
0 b
1
b
= (b00 ) b10 + (b01 ) b11
= ( 2/2) (i 2/2) + ( 2/2) (i 2/2)
= i/2 i/2 = 0.X
Also,
D
b0 b
0
=
=
=
(b00 ) b00 + (b01 ) b01
( 2/2) ( 2/2) + ( 2/2) ( 2/2)
1/2 + 1/2 = 1 X,
95
and
D
E
1 b
1
b
=
=
=
(b10 ) b10 + (b11 ) b11
(i 2/2) (i 2/2) + (i 2/2) (i 2/2)
1/2 + 1/2 = 1 X,
which establishes orthonormality.
We seek
v0
.
v1 B
The dotting trick says
v0
E
0 v
b
= (b00 ) v0 + (b01 ) v1
( 2/2) (1 + i) + ( 2/2) (1 i)
2,
D
=
=
=
and
v1
b1 v
= (b10 ) v0 + (b11 ) v1
(i 2/2) (1 + i) + (i 2/2) (1 i)
2,
D
=
=
=
so
1+i
1i
=
2, .
2, B
Finally, we check our work.
2 b0 +
2 b1
=
=
=
2/2
i
2/2
2
+
2
2/2
i 2/2
1
i
+
1
i
1+i
X
1i
[Exercise. Work in C2 and use the same v as above to get its coordinates relative
to the basis C,
1 1+i
1
2+i
C {
c0 ,
c1 } =
,
.
1
3
15 3 + i
Before starting, demonstrate that this basis is orthonormal.]
96
4.4
4.4.1
Hilbert Space
Definitions
A Hilbert Space is a real or complex vector space that
1. has an inner product and
2. is complete.
You already know what an inner product is. A vector space that has an inner product
is usually called an inner-product space.
Completeness is different from the property that goes by the same name (unfortunately) involving vectors spanning a space. This is not that. Rather, completeness
here means (please grab something stable and try not to hit your head) that any
Cauchy sequence of vectors in the space converges to some vector also in the space.
Unless you remember your advanced calculus, completeness wont mean much, and
its not a criterion that we will ever use, explicitly. However, if you want one more
degree of explanation, read the next paragraph.
Completeness of an Inner Product (Optional Reading)
The inner product, as we saw, imbues the vector space with a distance function,
dist( , ). Once we have a distance function, we can inspect any infinite sequence
of vectors, like { a0 , a1 , a2 , . . . }, and test whether the distance between consecutive
vectors in that sequence, dist(ak+1 , ak ), get small fast (i.e., obeys the Cauchy criterion
please research the meaning of this phrase elsewhere, if interested). If it does,
{ a0 , a1 , a2 , . . . } is called a Cauchy sequence. The vector space is complete if every
Cauchy sequence of vectors in the space approaches a unique vector, also in the space,
called the limit of the sequence.
Illustration of Completeness. The completeness criterion does not require a
vector space in order that it be satisfied. It requires only a set and a norm (or metric)
which allows us to measure distances. So we can ask about completeness of many sets
that are not even vector spaces. The simplest example happens to not be a vector
space.
Consider the interval [0, 1] in R which includes both endpoints. This
set
is
1
complete with respect to the usual norm in R. For example, the sequence 1 k k=2
is Cauchy and converges to 1, which is, indeed, in [0, 1]. This does not establish that
[0, 1] is complete, but it indicates the kind of challenging sequence that might prove
a set is not Cauchy.(See Figure 4.1)
For a counter-example, consider the interval (0, 1) in R which does not include
either endpoint. The same sequence, just discussed, is still in the set, but its limit,
1, is not. So this set is not complete. (See Figure 4.2)
97
Figure 4.1: The Cauchy sequence 1 k1 k=2 has its limit in [0, 1]
Figure 4.2: The Cauchy sequence 1 k1 k=2 does not have its limit in (0, 1)
Notation
When I want to emphasize that were working in a Hilbert space, Ill use the letter
H just as I use terms like R3 or Cn to denote real or complex vector spaces. H could
take the form of a C2 or a Cn , and I normally wont specify the dimension of H until
we get into tensor products.
4.4.2
Old Friends and New Acquaintances
Finite Dimensional Hilbert Spaces
Before we go further, lets give recognition to an old friend which happens to be
a Hilbert space: Rn with the usual dot-product. This Euclidean space (actually
these Euclidean spaces, since there is one for each positive n) does(do) meet the
completeness criterion, so it(they) are Hilbert space(s).
As we have been discussing on this page so, too, are the complex inner-product
spaces, Cn , for each positive n.
For us, the most important Hilbert space will be C2 .
Thus, we already know two classes of inner-product spaces that are Hilbert.
98
Infinite Dimensional Hilbert Spaces
When physicists use the term Hilbert space, they often mean something slightly more
exotic: function spaces. These are vector spaces whose vectors consist of sufficiently
well-behaved functions, and whose inner product is usually defined in terms of an
integral.
The Space L2 [a, b]. For example, all complex-valued functions, f (x), defined
over the real interval [a, b] and which are square-integrable, i.e.,
Z b
|f (x)|2 dx < ,
a
form a vector space. We can define an inner product for any two such functions, f
and g, using
Z b
hf | gi
f (x) g(x) dx ,
a
and this inner-product will give a distance and norm that satisfy the completeness
criterion. Hilbert spaces very much like these are used to model the momentum and
position of sub-atomic particles.
4.4.3
Some Useful Properties of Hilbert Spaces
We will be making implicit use of the following consequences of real and complex
inner-product spaces as defined above, so its good to take a moment and meditate
on each one.
Triangle Inequality
Both the real dot product of Rn and the complex inner product of Cn satisfy the
triangle inequality condition: For any vectors, x, y and z,
dist(x, z)
dist(x, y) + dist(y, z) .
Pictured in R2 , we can see why this is called the triangle inequality. (See Figure 4.3).
[Exercise. Pick three vectors in C3 and verify that the triangle inequality is
satisfied. Do this at least twice, once when the three vectors do not all lie on the
same complex line, { x C } and once when all three do lie on the same line
and y is between x and z.]
Cauchy-Schwarz Inequality
A more fundamental property of inner-product spaces is the Cauchy-Schwarz inequality which says that any two vectors, x and y, of an inner-product space satisfy
|hx | yi|2
99
kxk2 kyk2 .
Figure 4.3: Triangle inequality in a metric space.svg from Wikipedia
The LHS is the absolute-value-squared of a complex scalar, while the RHS is the
product of two vector norms (squared), each of which is necessarily a non-negative
real. Therefore, it is an inequality between two non-negative values. In words, it says
that the magnitude of an inner product is never more than the product of the two
component vector magnitudes.
The Cauchy-Schwarz inequality becomes an exact equality if and only if the two
vectors form a linearly dependent set, i.e., one is a scalar multiple of the other.
[Exercise. Pick two vectors in C3 and verify that the Cauchy-Schwarz inequality
is satisfied. Do this at least twice, once when the two vectors are linearly independent
and once when the are linearly dependent.]
4.5
4.5.1
Rays, Not Points
Modeling Quantum Systems
There is one final adjustment to be made if we are to accurately model quantum
systems with these Hilbert spaces. It will simultaneously simplify your computations
and confuse you. My hope is that there will be more simplify and less confuse.
Everything we do computationally will indeed take place in some Hilbert space
H = Cn . In that regard, we have covered all the basic moves in this short lesson
(minus linear transformations, which we present next time). However, in quantum
mechanics, it is not true that every vector in H corresponds to a distinct physical
state of our system. Rather, the situation is actually a little easier (if we dont resist).
First a definition:
Definition of Ray. A ray (through the origin, 0) in a complex vector space is
the set of all the scalar multiples of some non-zero vector,
Ray belonging to a 6= 0
a C
Each possible physical state of our quantum system a concept we will cover soon
will be represented by a unit vector (vector of length one) in H. Stated differently,
any two non-zero H-vectors that differ by a complex scalar represent the same state,
100
so we may as well choose a vector of length one (kvk = 1) that is on the same ray as
either (or both) of those two; we dont have to distinguish between any of the infinite
number of vectors on that ray.
This equivalence of all vectors on a given ray makes the objects in our mathematical model not points in H, but rays through the origin of H.
Figure 4.4: A 3-D quantum state is a ray in its underlying H = C3
If two complex n-tuples differ by a mere scalar multiple, , they are to be considered the same state in our state space, i.e.,
b
= a
b a
(as H-space vectors)
(as a physical state),
to be read a is equivalent to b, or a and b represent the same quantum state.
This identifies
[the quantum states modeled by H ]
with
[rays through the origin of that H = Cn ].
(See Figure 4.4 for an example of an H-state represented by its C3 ray.)
If we were to use brackets, [a], to mean the ray represented by the n-tuple a,
then we could express a quantum state in H, as
An H-state [a] =
a C
Example. All three of these complex ordered-pairs lie on the same ray in C2 , so
they all represent the same quantum state modeled by this two-dimensionbal H:
!
3i
2
10
5
a=
, b=
and c = 1+3i
i
5i
10
101
[Exercise. Elaborate.]
Dividing any one of them by its norm will produce a unit vector (vector with
modulus one) which also represents the same ray. Using the simplest representative,
2
2
!
2
i
i
a
5
,
=
= p
=
i
kak
5
(2)(2) + (i)(i)
5
often written as
1
2
.
i
For example, if we ended up with a as the answer to a problem, we could replace
it by a unit vector along its ray,
a
.
kak
(See Figure 4.5.)
Figure 4.5: Dividing a vector by its norm yields a unit vector on the same ray
Computing a unit length alternative to a given vector in H turns out to be of
universal applicability in quantum mechanics, because it makes possible the computation of probabilities for each outcome of a measurement. The fact that a vector has
norm = 1 corresponds to the various possible measurement probabilities adding to 1
(100% chance of getting some measurement). Well see all this soon enough.
Caution. Figures 4.4 and 4.5 suggest that once you know the magnitude of a,
that will nail-it-down as a unique C3 representative at that distance along the ray.
This is far from true. Each point pictured on the ray, itself, constitutes infinitely
many different n-tuples in Cn , all differing by a factor of ei , for some real .
102
Example (continued). We have seen that a= (2, i)t has norm 5. A different
i/6
i/6
Cn representative for this H-space vector
is
e
a.
We
can
easily
see
that
e
a
i/6
has the same length as a. In words, e
= 1 (prove it or go back and review your
complex arithmetic module), so multiplying by ei/6 , while changing the Cn vector,
will not change its modulus (norm). Thus, that adjustment not only produces a
different representative, it does so without changing the norm. Still, it doesnt hurt
to calculate norm of the product the long way, just for exercise:
!
i/6
2 (cos /6 + i sin /6)
2e
i/6
e
a =
=
i ei/6
i (cos /6 + i sin /6)
2 cos /6 + 2i sin /6
=
1 sin /6 + i cos /6
!
2 23 + 2i 12
=
1 21 + i 23
!
3 + i
=
21 + i 2 3
!
1 2 3 + 2i
=
2 1 + i 3
I did the hard part: I simplified the rotated vector (we call the act of multiplying by
ei for real a rotation because of the geometric implication which you can imagine,
look-up, or simply accept).
All thats left to do in this computation is calculate the
norm and see that it is 5.
[Exercise. Close the deal.]
4.5.2
0 is not a Quantum State
The states we are to model correspond to unit vectors in H. We all get that, now.
But what are the implications?
A state has to be a normalizable vector, which 0 is not. 0 will be the only
vector in H that does not correspond to a physical quantum state.
As we already hammered home, every other vector does correspond to some
state, but it isnt the only vector for that state. Any other vector that is a
scalar multiple of it represents the same state.
The mathematical entity that we get if we take the collection of all rays, {[a]}, as
its points, is a new construct with the name: complex projective sphere. This
new entity is, indeed, in one-to-one correspondence with the quantum states,
but ...
103
... the complex projective sphere is not a vector space, so we dont want to go
too far in attempting to define it; any attempt to make a formal mathematical
entity just so that it corresponds, one-to-one, with the quantum states it models
results in a non-vector space. Among other things, there is no 0-vector in such
a projective sphere, thus, no vector addition.
The Drill. With this in mind, we satisfy ourselves with the following process. It
may lack concreteness until we start working on specific problems, but it should give
you the feel for whats in store.
1. Identify a unit vector, v
H, corresponding to the quantum state of our
problem.
2. Sometimes we will work with a scalar multiple, v =
v H = Cn , because it
simplifies our computations and we know that this vector lies on the same ray
as the original. Eventually, well re-normalize by dividing by to bring it back
to the projective sphere.
3. Often we will apply a unitary transformation, U , directly to v
. U v
will be
a unit vector because, by definition (upcoming lecture on linear transformations), unitary transformations preserve distances. Thus, U keeps vectors on
the projective sphere.
4. In all cases, we will take care to apply valid operations to our unit state vector,
making sure that we end up with an answer which is also a unit vector on the
projective sphere.
4.5.3
Why?
There one question (at least) that you should ask and demand be answered.
Why is this called a projective sphere? Good question. Since the states
of our quantum system are rays in H, and we would prefer to visualize vectors as
points, not rays, we go back to the underlying Cn and project the entire ray (maybe
collapse would be a better word) onto the surface of an n-dimensional sphere (whose
real dimension is actually 2(n 1), but never mind that). We are projecting all those
representatives onto a single point on the complex n-sphere. (See Figure 4.5.) Caution: Each point on that sphere still has infinitely many representatives impossible
to picture due to a potential scalar factor ei , for real .]
None of this is to say that scalar multiples, a.k.a. phase changes, never matter.
When we start combining vectors in H, their relative phase will become important,
and so we shall need to retain individual scalars associated with each component
n-tuple. Dont be intimidated; well get to that in cautious, deliberate steps.
104
4.6
Almost There
We have one last math lesson to dance through after which we will be ready to
learn graduate level quantum mechanics (and do so without any prior knowledge of
undergraduate quantum mechanics). This final topic is linear transformations. Rest
up. Then attack it.
105
Chapter 5
Linear Transformations
5.1
5.1.1
Linear Transformations for Quantum Computing
A Concept More Fundamental Than the Matrix
In the last lecture we completed the mathematical preliminaries needed to formally
define a qubit, the vector object that will model a physical memory location in a
quantum computer. Of course, we havent officially defined the qubit yet, but at
least we have a pretty good idea that it will be associated with a vector on the
projective sphere of some Hilbert space. Were fully armed to deal with qubits when
the day comes (very soon, I promise).
This lesson will fill the remaining gap in your math: it provides the theory and
computational skills needed to work with quantum logic gates, the hardware that
processes qubits.
The introductory remarks of some prior lessons suggested that a quantum logic
gate will be described by a matrix. Consider the Y operator, a quantum gate that
well learn about later in the course. Well see that Y takes one qubit in and produces
one qubit out,
Y
and its matrix will be described by
Y
0 i
.
i 0
Fair enough, except for one small problem.
A matrix for a logic gate implicitly assumes that there is an underlying basis that
is being used in its construction. Since we know that every vector space can have
many bases, it will turn out that the matrix for any logic gate like Y will mutate
106
depending which basis we are using. For example, there is a basis in which the same
logic gate has a different matrix, i.e.,
1 0
Y =
.
0 1
This should disturb you. If a matrix defines a quantum logic gate, and there can be
more than one matrix describing that gate, how can we ever know anything?
There is a more fundamental concept than the matrix of linear algebra, that of
the linear transformation. As youll learn today, a linear transformation is the basisindependent entity that describes a logic gate. While a linear transformations matrix
will change depending on which underlying basis we use to construct it, its life-giving
linear transformation remains fixed. You can wear different clothes, but underneath
youre still you.
5.1.2
The Role of Linear Transformations in Quantum Computing
There are many reasons we care about linear transformations. Ill list a few to give
you some context, and well discover more as we go.
Every algorithm in quantum computing will require taking a measurement. In
quantum mechanics, a measurement is associated with a certain type of linear
transformation. Physicists call it by the name Hermitian operator, which well
define shortly. Such beasts are, at their core, linear transformations. [Beware:
I did not say that taking a measurement was a linear transformation; it is not.
I said that every measurement is associated with a linear transformation.]
In quantum computing well be replacing our old logic gates (AND, XOR, etc.)
with a special kind of quantum gate whose basis independent entity is called a
unitary operator. But a unitary operator is nothing more than a linear transformation which has some additional properties.
We often want to express the same quantum state or qubit in different bases.
One way to convert the coordinates of the qubit from one basis to another is to
subject the coordinate vector to a linear transformation.
5.2
5.2.1
Definitions and Examples
Actions as well as Name Changes
Linear Transformations are the verbs of vector spaces. They can map vectors of one
vector space, V, into a different space, W,
T
V W ,
107
or they can move vectors around, keeping them in the same space,
T
V V .
They describe actions that we take on our vectors. They can move a vector by
mapping it onto another vector. They can also be applied to vectors that dont move
at all. For example, we often want to expand a vector along a basis that is different
from the one originally provided, and linear transformations help us there as well.
5.2.2
Formal Definition of Linear Transformation
A linear transformation, T , is a map from a vector space (the domain) into itself or
into another vector space (the range),
T
V W
(W could be V) ,
that sends input vectors to output vectors,
T (v) 7 w .
That describes the mapping aspect (Figure 5.1). However, a linear transformation
Figure 5.1: T mapping a vector v in R3 to a w in R3
has the following additional properties that allow us to call the mapping linear :
T (c v) = c T (v) and
T (v1 + v2 ) = T (v1 ) + T (v2 ) .
Here, c is a scalar of the domain vector space and v1 , v2 are domain vectors. These
conditions have to be satisfied for all vectors and scalars.
Besides possibly being different spaces, the domain and range can also have different dimensions. However, today, well just work with linear transformation of a
space into itself.
A linear transformation can be interpreted many different ways. When learning
about them, its best to think of them as mappings or positional changes which
108
convert a vector (having a certain direction and length) into a different vector (having
a different direction and length).
Notation. Sometimes the parentheses are omitted when applying a linear transformation to a vector:
Tv
T (v) .
Famous Linear Transformations
Here are a few useful linear transformations, some of which refer to Rn or Cn , some
to function spaces not studied much in this course. Ive included figures for a few
that well meet later today.
v is any vector in an n dimensional vector space on which the linear transformation
acts. x
k is the kth natural basis vector, vk is the kth coordinate of v in the natural
is any unit vector.
basis, c is a scalar, and n
1(v) v (Identity)
0(v)
Sc (v)
Pk (v)
Pn (v)
D()
Z x
(f )
0 (Zero)
c v (Scale) (F igure 5.2)
vk x
k (Projection onto x
k ) (F igure 5.3)
(v n
) n
(Projection onto n
) (F igure 5.4)
0
(Differentiation)
Z x
f (x0 ) dx0 (Anti-differentiation)
TA (v) Av
(Multiplication by matrix of constants, A)
Figure 5.2: A scaling transformation
Lets pick one and demonstrate that it is, indeed, linear. We can prove both
conditions at the same time using the vector cv1 + v2 as a starting point. Ive
109
3
Figure 5.3: Projection onto the direction z, a.k.a. x
Figure 5.4: Projection onto an arbitrary direction n
selected the projection onto an arbitrary direction, n
, as the example.
Pn (c v1 + v2 ) =
=
=
=
[(c v1 + v2 ) n
] n
= [(c v1 ) n
+ v2 n
] n
[c (v1 n
) + v2 n
] n
= [c (v1 n
)] n
+ [v2 n
] n
c [(v1 n
)] n
+ [v2 n
] n
c Pn (v1 ) + Pn (v2 ) . QED
We also listed an example in which TA was defined by the multiplicative action
of a matrix, A, consisting of scalar constants (i.e., the matrix cant have variables or
formulas in it). To make this precise, let A have size mn and v be any n-dimensional
vector. (They both have to use the same field of scalars, say R or C.) From the rules
of matrix multiplication we can easily see that
A(v + w)
A(cv)
=
=
A(v) + A(w)
cA(v).
and
Therefore,
TA (v) Av
is linear. This is the linear transformation, TA , induced by the matrix A. Well look
at it more closely in a moment.
[Exercise. Prove these two claims about matrix-multiplication.]
110
[Exercise. Look at these mapping of C3 into itself.
x
2x
2y
T1 : y
7
z
2z
x
2ix
T2 : y 7
3y
z
(4 1i)z
x+2
x
2i
T3 : y 7 y +
z
z+ 2
x
y
x
T4 : y
7
z
z
x
xy
y
T5 : y
7
z
z2
x
0
T6 : y 7 xyz
z
0
Which are linear, which are not? Support each claim with a proof or counter example.]
5.3
The Special Role of Bases
Basis vectors play a powerful role in the study of linear transformations as a consequence of linearity.
Lets pick any basis for our domain vector space (it doesnt have to be the natural
basis or even be orthonormal),
B = {b1 , b2 , ...} .
The definition of basis means we can express any vector as a linear combination
of the bk s (uniquely) using the appropriate scalars, k , from the underlying field.
Now, choose any random vector from the space and expand it along this basis (and
remember there is only one way to do this for every basis),
v
n
X
k bk .
k=1
We now apply T to v and make use of the linearity (the sum could be infinite) to get
!
n
n
X
X
k bk
=
k T (bk )
Tv = T
k=1
k=1
111
What does this say? It tells us that if we know what T does to the basis vectors,
we know what it does to all vectors in the space. Lets say T is some hard-todetermine function which is actually not known analytically (by formula), but by
experimentation we are able to determine its action on the basis. We can extend that
knowledge to any vector because the last result tells us that the coordinates of the
vector combined with the known values of T on the basis are enough. In short, the
small set of vectors
{T (b1 ) , T (b2 ) , ...}
completely determines T .
5.3.1
Application: Rotations in Space
We now apply the theory to a linear transformation which seems to rotate points by
90 counter-clockwise in R2 . Lets call the rotation R/2 , since /2 radians is 90 .
The plan
We must first believe that a rotation is linear. I leave this to you.
[Exercise. Argue, heuristically, that R/2 is linear. Hint: Use the equations that
define linearity, but proceed intuitively, since you dont yet have a formula for R/2 .]
Next, we will use geometry to easily find the action of T on the two standard basis
vectors, {
x, y
}. Finally, well extend that to all of R2 , using linearity. Now for the
details.
The Details
Figures 5.5 and 5.6 show the result of the rotation on the two basis vectors.
Figure 5.5: Rotation of x
counter-clockwise by /2
These figures suggest that, in R2 ,
R/2 (
x)
R/2 (
y)
=
=
112
and
x,
Figure 5.6: Rotation of y
counter-clockwise by /2
therefore, for any v, with natural basis coordinates (vx , vy )t , linearity allows us to
write
R/2 (v)
=
=
=
R/2 (vx x
+ vy y
)
vx R/2 (
x) + vy R/2 (
y)
vx y
vy x
.
From knowledge of the linear transformation on the basis alone, we were able to derive
a formula applicable to the entire space. Stated in column vector form,
y
x
R/2
=
,
y
x
and we have our formula for all space (assuming the natural basis coordinates). Its
that easy.
[Exercise. Develop the formula for a counter-clockwise rotation (again in R2 )
through an arbitrary angle, . Show your derivation based on its effect on the natural
basis vectors.]
[Exercise. What is the formula for a rotation, Rz, /2 , about the z-axis in R3
through a 90 angle, counter-clockwise when looking down from the positive z-axis.
Show your derivation based on its effect on the natural basis vectors, {
x, y
,
z}.]
5.4
5.4.1
The Matrix of a Linear Transformation
From Matrix to Linear Transformation
You showed in one of your exercises above that any matrix A (containing scalar
constants) induces a linear transformation TA . Specifically, say A is a complex matrix
of size m n and v is a vector in Cn . Then the formula
TA (v)
Av
defines a mapping
TA
Cn Cm ,
113
which turns out to be linear. As you see, both A, and therefore, TA , might map
vectors into a different-sized vector space; sometimes m > n, sometimes m < n and
sometimes m = n.
5.4.2
From Linear Transformation to Matrix
We can go the other way. Starting with a linear transformation, T , we can construct
a matrix MT , that represents it. There is a little fine print here which well get to in
a moment (and you may sense it even before we get there).
For simplicity, assume we are working in a vector space and using some natural
basis A = {ak }. (Think {
x, y
,
z} of R3 .) So every vector can be expanded along A
by
v
n
X
k ak .
k=1
We showed that the action of T on the few vectors in basis A completely determines
its definition on all vectors v. This happened because of linearity,
Tv
n
X
k T (ak ) .
k=1
Lets write the sum in a more instructive way as the formal (but not quite legal) dot
product of a (row of vectors) with a (column of scalars). Shorthand for the above
sum then becomes
1
!
2
Tv =
T (a1 ) , T (a2 ) , . . . , T (an ) .. .
.
n
Now we expand each vector T (ak ) vertically into its coordinates relative to the same
basis, A, and we will have a legitimate product,
(T a1 )1 (T a2 )1 (T an )1
1
(T a1 )2 (T a2 )2 (T an )2 2
..
..
.. .. .
.
.
.
.
.
. .
(T a1 )n (T a2 )n . . . (T an )n
n
Writing ajk in place of (T ak )j , we have the
a11 a12
a21 22
T v = ..
..
.
.
an1 an2
simpler statement,
a1n
1
a2n 2
.. ,
. . . ..
.
.
ann
n
114
which reveals that T is nothing more than multiplication by a matrix made up of the
constants ajk .
Executive Summary. To get a matrix, MT , for any linear transformation, T ,
form a matrix whose columns are T applied to each basis vector.
We can then multiply any vector v by the matrix MT = ajk to get T (v).
Notation
Because this duality between matrices and linear transformations is so tight, we rarely
bother to distinguish one from the other. If we start with a linear transformation, T ,
we just use T as its matrix (and do away with the notation MT ). If we start with a
matrix, A, we just use A as its induced linear transformation, (and do away with the
notation TA ).
Example
Weve got the formula for a linear transformation that rotates vectors counter-clockwise
in R2 , namely,
x
y
R/2
=
.
y
x
To compute its matrix relative to the natural basis, {
x, y
}, we form
1
0
MR/2 =
R/2 (
x) , R/2 (
y)
=
R/2
, R/2
0
1
0 1
=
.
1 0
We can verify that it works by multiplying this matrix by an arbitrary vector,
0 1
x
0 x + (1) y
y
=
=
,
1 0
y
1x + 0y
x
as required.
[Exercise. Show that the matrix for the scaling transformation ,
S3i (v)
3i v ,
is
MS3i
3i 0
,
0 3i
and verify that it works by multiplying this matrix by an arbitrary vector to recover
the definition of S3i .]
115
5.4.3
Dependence of a Matrix on Basis
When we are using an obvious and natural orthonormal basis (like {
x, y
,
z} for C3 ),
everything is fairly straightforward. However, there will be times when we want a
different basis than the one we thought, originally, was the preferred or natural. The
rule is very strict, but simple.
In order to enable matrix multiplication to give us the result of some linear
T applied to a vector v, we must make sure T and v are both expressed
in a common basis. In that case, w = MT v only makes sense if both
vectors and all the columns of MT are expanded along that common basis.
In short, dont mix bases when you are expressing vectors and linear transformations.
And dont forget what your underlying common basis is, especially if its not the
preferred basis.
Remember, there is always some innate definition of our vectors, independent of
basis. Distinct from this, we have the expression of those vectors in a basis.
Say v is a vector. If we want to express it using coordinates along the A basis on
Sunday, then cabaret along the B basis all day Monday, thats perfectly fine. They
are all the same object, and this notation could be employed to express that fact:
v = v|A = v|B
2
This is confusing for beginners,
because they usually start with vectors in R , which
x
have innate coordinates,
, that really dont refer to a basis. As weve seen,
y
though, the vector happens to look the same as its expression in the natural basis,
{
x, y
}, so without the extra notation, above, theres no way to know (nor is there
usually the need to know) whether the author is referring to the vector, or its coordinates in the natural basis. If we ever want to disambiguate the conversation, we will
use the notation, v|A , where A is some basis we have previously described.
The same goes for T s matrix. If I want to describe T in a particular basis, I will
say something like
T = T |A = T |B ,
and now I dont even need to use MT , since the |A or |B implies we are talking about
a matrix.
So the basis-free statement,
w
T (v) ,
can be viewed in one basis as
w|A
T |A v|A ,
w|B
T |B v|B .
and in another basis as
116
Matrix of MT in a Non-Standard Basis
The development of our formula of a matrix MT of a transformation, T , led to
!
MT
T (a1 ) , T (a2 ) , . . . , T (an )
when we were assuming the preferred basis,
A
{a1 , a2 , . . .}
1
0
0 1
, , ... .
..
...
There was nothing special about the preferred basis in this formula; if we had any
basis even one that was non-orthonormal the formula would still be
!
T
=
T (b1 ) , T (b2 ) , . . . , T (bn )
.
B
The way to see this most easily is to first note that in any basis B, each basis vector
bk , when expressed in its own B-coordinates, looks exactly like the kth preferred basis
element, i.e.,
0
..
.
bk
= 1 kth element .
.
B
..
0
B
[Exercise. prove it.]
Now we can see that the proposed T applied to bk B gives the correct result. Writing
B
everything in the B basis (some labels omitted) and using matrix multiplication we
have
0
! ...
1
. . . , T (bk ) , . . .
=
T
(b
)
X
k .
B
B
.
..
0
Example
The transformation
x
x
T :
7
y
(3i) y
117
in the context of the vector space C2 is going to have the preferred-basis matrix
1 0
MT =
.
0 3i
[Exercise. Prove it.]
Lets see what T looks like when expressed in the non-orthogonal basis
2
1
C =
,
.
0
1
[Exercise. Prove that this C is a basis for C2 .]
Using our formula, we get
T
T (c1 ) , T (c2 )
!
C
We compute each T (ck ) , k = 1, 2.
C
First up, T (c1 ) :
C
T (c1 )
2
T
0
2
.
0
Everything needs to be expressed in the C basis, so we show that for this vector:
2
2
1
1
= 1
,
+ 0
=
1
0 C
0
0
so this last column vector will be the first column of our matrix.
Next, T (c2 ) :
C
T (c2 )
1
T
1
1
.
3i
We have to express this, too, in the C basis, a task that requires a modicum of algebra:
1
2
1
=
+
.
3i
0
1
Solving this system should be no problem for you (exercise), giving,
1 3i
2
3i ,
118
so
T (c2 )
1
3i
!
3i
13i
2
Plugging these in, we get,
T
5.4.4
1
0
!
.
3i
13i
2
The Transformation in an Orthonormal Basis
When we have an orthonormal basis, things get easy. We start our analysis by working
in a natural, orthonormal basis, where all vectors and matrices are simply written
down without even thinking about the basis, even though its always there, lurking
behind the page. Today, were using A to designate the natural basis, so
v = v|A
would have the same coordinates on both sides of the equations, perhaps
1+i 2
.
3.2 i
Formula for Matrix Elements Using the Inner Product Trick
Even though we already know how to get the matrix in the preferred basis, A, lets
analyze it, briefly. Take the scaling transformation S2/2 in R2 whose matrix is
2/2 0
.
0
2/2 A
(You constructed the matrix for an Sc in an earlier exercise today using the scaling
factor c = 3i, but just to summarize, we apply S2/2 to the two vectors (1, 0)t and
(0, 1)t , then make the output vectors the columns of the required matrix.) Remember,
each column is computed by listing the coordinates of T (
x) and T (
y) in the natural
basis. But for either vector, its coordinates can be expressed using the dotting trick
because we have an orthonormal basis. So,
0
2/2
x
x
0
2/2
2/2 0
.
=
0
2/2 A
0
2/2
y
y
0
2/2
This is true for any T and the preferred basis in R2 . To illustrate, lets take the upper
left component. Its given by the dot product
T11
h
x | T (
x)i .
119
Similarly, the other three are given by
T21
T12
T22
=
=
=
h
y | T (
x)i ,
h
x | T (
y)i
h
y | T (
y)i .
and
In other words,
T
h
x | T (
x)i
h
x | T (
y)i
h
y | T (
x)i
h
y | T (
y)i
!
.
To make things crystal clear, lets rename the natural basis vectors
A
{
e1 ,
e2 },
producing the very illuminating
T
T11 T12
!
=
T21 T22
h
e1 | T (
e1 )i
h
e1 | T (
e2 )i
h
e2 | T (
e1 )i
h
e2 | T (
e2 )i
!
.
But this formula, and the logic that led to it, would work for any orthonormal basis,
not just A, and in any vector space, not just R2 .
Summary. The jkth matrix element for the transformation, T , in an orthonormal
basis,
B
{ b1 , b2 , . . . . bn }
is given by
Tjk
E
j T (b
k ) ,
b
so
E
b1 T (b1 )
D
E
2 T (b
1 )
b
D
T
D
E
b1 T (b2 )
D
E.
2 T (b
2 )
b
Not only that, but we dont even have to start with a preferred basis to express our
T and B that are used in the formula. As long as T and B are both expressed in
the same basis say someD third
we can use the coordinates and matrix elements
C E
relative to C to compute bj T (bk ) and thus give us the matrix of TB .
Happy Note. That third basis, C, doesnt even have to be orthonormal. As long
as B, the desired basis in which we are seeking a new representation, is orthonormal,
it all works.
120
Example 1
Lets represent the scaling transformation
2/2
0
.
S2/2 =
0
2/2 A
in a basis that we encountered in a previous lesson (which was named C then, but
well call B today, to make clear the application of the above formulas),
2/2 , 2/2
B {b1 , b2 } =
.
2/2
2/2
We just plug in (intermediate B labels omitted):
D
E D
E
b
T
(
b
)
b
T
(
b
)
1
1
1
2
= D
S2/2
E D
E
B
2 T (b
1 )
2 T (b
2 )
b
b
b1 22
D
2 2
b
2
D
=
b1
*
2
b
*
=
1
b
1
b
1
2
1
2
!+
1
2
1
2
!+
2
2
2
2
D
1 2
b
2
D
2 2
b
2
2
b
2
b
!+
1
2
1
b
12
*
!+
b2 1
2
*
.
B
Thats surprising (or not). The matrix is the same in the B basis as it is in the A
basis.
1) Does it make sense?
2) Is this going to be true for all orthonormal bases and all transformations?
These are the kinds of questions you have to ask yourself when you are manipulating
mathematical symbols in new and unfamiliar territory. Ill walk you through it.
1) Does it make sense?
121
Lets look at the result of
S
3
2/2
10
in both bases.
First well compute it in the A-basis and transform the resulting output vector
to the B-basis.
When weve done that, well start over, but this time first convert the starting
vector and matrix into B-basis coordinates and use those to compute the result,
giving us an answer in terms of the B-basis.
Well compare the two answers to see if they are equal.
By picking a somewhat random-looking vector, (3, 10)t , and discovering that
both T |A and T |B turn it into equivalent output vectors, we will be satisfied that the
result makes sense; apparently it is true that the matrix for S2/2 looks the same
when viewed in both bases. But we havent tested that, so lets do it.
A-Basis. In the A basis we already know that the output vector must be
3 2/2
,
5 2
A
because thats just the application of the transformation to the innate vector, which
is the same as applying the matrix to the preferred coordinates. We transform the
output vector into B-basis:
3 2/2
b1
5 2
3 2/2
=
5 2
3 2/2
A
b2
5 2
B
!
!
7
32 + 5
2
=
=
3
13
+5
2
2
B
B
3
B-Basis. Now, do it again, but this time convert the input vector,
, to
10
B-coordinates, and run it through S2/2 B . First the input vector:
3
b
1
10
=
10 A
3
b2
10
B
3 2 2 + 102 2
3 2
10 2
+
2
2
122
=
B
!
7 2
2
13 2
2
B
Finally, we apply S2/2 B to these B coordinates:
S2/2
!
7 2
2
13 2
2
B
2
2
0
!
7
2
.
13
2
2
2
!
7 2
2
13 2
2
And ... we get the same answer. Apparently there is no disagreement, and other tests
would give the same
results, so it seems we made no mistake when we derived the
matrix for S2/2 B and got the same answer as S2/2 A .
This is not a proof, but it easy enough to do (next exercise). A mathematicallyminded student might prefer to just do the proof and be done with it, while an
applications-oriented person may prefer just to test it on a random vector to see if
s/he is on the right track.
[Exercise. Prove that for any v, you get the same result for S2/2 (v) whether
you use coordinates in the A basis or in the B basis, thus confirming once-and-for-all
that the matrix of this scaling transformation is the same in both bases.]
2) Do we always get the same matrix?
Certainly not otherwise would I have burdened you with a formula for computing
the matrix of a linear transformation in an arbitrary basis?
Example 2
We compute the matrices for a projection transformation,
2/2
Pn (v) (v n
) n
,
where n
=
.
2/2
in the two bases, above.
The A Matrix. Ill use the common notation introduced earlier,
ek , for the kth
natural basis vector.
Pn
(Pn )11 (Pn )12
!
=
(Pn )21 (Pn )22
h
e1 | Pn (
e1 )i
h
e1 | Pn (
e2 )i
h
e2 | Pn (
e1 )i
h
e2 | Pn (
e2 )i
Since
Pn (
e1 )
Pn (
e2 )
1/2
1/2
1/2
1/2
123
and
!
.
(exercise: prove it), then
Pn
1/2 1/2
1/2 1/2
(exercise: prove it).
The B Matrix. Now use the B basis.
E
D
1 Pn (b
1 )
b
E
Pn
=
= D
B
b2 Pn (b1 )
D
E
1 Pn (b
2 )
b
D
E .
2 Pn (b
2 )
b
B
1 and Pn (b
1 ) have to be expressed in this same
Remember, for this to work, both b
basis, and we almost always express them in the B basis, since thats how we know
everything. Now,
2/2
and
Pn (b1 ) =
2/2
0
2 ) =
Pn (b
0
(exercise: prove it), so
Pn
1 0
0 0 B
(exercise: prove it), a very different matrix than Pn .
A
[Exercise. Using (3, 10) (or a general v) confirm (prove) that the matrices
Pn A and Pn B produce output coordinates of the identical intrinsic output vector,
Pn (3, 10)t . Hint: use an argument similar to the one above.]
5.5
Some Special Linear Transformations for Quantum Mechanics
We will be using two special types of linear transformations frequently in quantum
computing, unitary and Hermitian operators. Each has its special properties that
are sometimes easier to state in terms of their matrices. For that, we need to define
matrix adjoints.
5.5.1
The Adjoint of a Matrix
Given complex n m matrix, M , we can form an m n matrix, M , by taking M s
conjugate transpose,
M
Mt
or, in terms of components,
M jk = Mkj
.
124
In other words, we form the transpose of M , then take the complex conjugate of every
element in that matrix.
Examples.
1i
3
0
2.5
7 + 7i
6
7i
88
1+i
0
6
3
7i
2.5
7 7i 88
1+i
5
3 2i
2.5
7 7i
99
ei
1i
5
3 + 2i
99
2.5
7 + 7i ei
When either m = 1 or n = 1, the matrix is usually viewed as a vector. The adjoint
operation turns column vectors into row vectors (while also transposing them), and
vice versa:
!
1 + i
=
1 i , 2 + 2i
2 2i
!
3 7i
3 + 7i , 2
=
2
The Adjoint of a Linear Transformation
Because linear transformations always have a matrix associated with them (once we
have established a basis), we can carry the definition of adjoint easily over to linear
transformations.
Given a linear transformation, T with matrix MT (in some agreed-upon basis), its
adjoint, T , is the linear transformation defined by the matrix (multiplication) MT .
It can be stated in terms of its action on an arbitrary vector,
T (v)
(MT ) v ,
for all v V .
[Food for Thought. This definition requires that we have a basis established,
otherwise we cant get a matrix. But is the adjoint of a linear transformation, in this
definition, going to be different for different bases? Hit the CS 83A discussion forums,
please.]
5.5.2
Unitary Operators
Informal Definition. A unitary operator is a linear transformation that preserves
distances.
In a moment, we will make this precise, but first the big news.
125
Quantum Logic Gates Are Unitary Operators
v
Uv
Quantum computing uses a much richer class of logical operators (or gates) than
classical computing. Rather than the relatively small set of gates: XOR, AND,
NAND, etc., in classical computing, quantum computers have infinitely many different
logic gates that can be applied. On the other hand, the logical operators of quantum
computing are of a special form not required of classical computing; they must be
unitary, i.e., reversible.
Definition and Examples
Theoretical Definition. A linear transformation, U , is unitary (a.k.a.
a unitary operator) if it preserves inner products, i.e.,
h Uv | Uw i
hv|wi ,
for all vectors v, w.
While this is a statement about the preservation of inner products, it has many
implications, one of which is that that distances are preserved, i.e., for all v, kU vk =
kvk .
Theorem. The following characterizations of unitary transformations
are equivalent:
i) A linear transformation is unitary if it preserves the lengths of vectors: kU vk = kvk.
ii) A linear transformation is unitary if it preserves inner products,
h U v | U w i = hv | wi.
iii) A linear transformation is unitary if its matrix (in an orthonormal
basis) has orthonormal rows (and columns).
Caution. This theorem is true, in part, because inner products are positivedefinite, i.e., v 6= 0 kvk > 0 (lengths of non-zero vectors are strictly positive).
However, well encounter some very important pseudo inner products that are not
positive definite, and in those cases the three conditions will not be interchangeable.
More on that when we introduce the classical bit and quantum bit, a fews lectures
hence.
A concrete example that serves as a classic specimen is a rotation, R , in any
Euclidean space, real or complex.
126
Notation. It is common to use the letter, U , rather than T , for a unitary operator
(in the absence of a more specific designation, like R ).
Because of this theorem, we can use any of the three conditions as the definition
of unitarity, and for practical reasons, we choose the third.
Practical Definition #1. A linear transformation, U , is unitary (a.k.a. a
unitary operator ) if its matrix (in any orthonormal basis) has orthonormal columns
(or equivalently orthonormal rows), i.e., U is unitary for any orthonormal basis,
B
{bk }nk=1 ,
the column vectors of U B satisfy
h U (bj ) | U (bk ) i
jk .
Note that we only need to verify this condition for a single orthonormal basis (exercise,
below).
The Matrix of a Unitary Operator. A Matrix, M , is called unitary if its
adjoint is also its inverse, i.e.,
M M
MM
1.
It is easily seen by looking at our practical definition (condition 3 ) of a unitary
operator that an operator is unitary its matrix is unitary. This leads to yet
another commonly stated, and equivalent, condition of a unitary operator.
Practical Definition #2. A unitary operator, U , is one in which
U U
UU
1.
Some Unitary Operators
The 90 Rotations. The /2 counter-clockwise rotation in R2 ,
x
y
R/2
=
,
y
x
is unitary, since its matrix in the preferred basis, {
x, y
}, is
0 1
MR/2 =
,
1 0
and it is easily seen that the columns are orthonormal.
General Rotations. In a previous exercise, you showed that counterclockwise rotation in R2 was, in the standard basis,
cos sin
MR =
.
sin cos
127
Dotting the two columns, we get
cos
sin
sin
cos
cos sin sin cos
0,
and dotting a column with itself (well just do the first column to demonstrate) gives
cos
cos
= cos2 + sin2 = 1 ,
sin
sin
so, again, we get orthonormality.
Phase Changes. A transformation which only modifies the arg of each coordinate,
i
x
e x
,
=
,
y
ei y
has the matrix
,
i
e
0
.
0 ei
Because this is a complex matrix, we have to apply the full inner-product machinery,
which requires that we not forget to take conjugates. The inner product of the two
columns is easy enough,
i
0
e
= ei 0 + 0 ei = 0 ,
0 ei
but do notice that we had to take the complex conjugate of the first vector even
though failing to have done so would have been an error that still gave us the right
answer. The inner product of a column with itself (well just show the first column)
gives
i i
e
e
= ei ei + 0 0
0 0
=
e0 + 0
1,
and we have orthonormality. Once again, the complex conjugate was essential in the
computation.
Some Non-Unitary Operators
Scaling by a Non-Unit Vector. Scaling by
the matrix
c
Sc =
0
128
a complex (or real) c, with |c| =
6 1 has
0
,
c
whose columns are orthogonal (do it), but whose column vectors are not unit length,
since
c c
= c c + 0 0 = |c|2 6= 1 ,
0 0
by construction.
Note. In our projective Hilbert spaces, such transformations dont really exist,
since we consider all vectors which differ by a scalar multiple to be the same entity
(state). While this example is fine for learning, and true for non-projective Hilbert
spaces, it doesnt represent a real operator in quantum mechanics.
A Projection Operator. Another example we saw a moment ago is the projection onto a vectors 1-dimensional subspace in R2 (although this will be true of such
a projection in Rn , n 2). That was
2/2
Pn (v) (v n
) n
,
where n
=
.
2/2
We only need look at Pn (v) in either of the two bases for which we computed its
matrices to see that this is not unitary. Those are done above, but you can fill in a
small detail:
[Exercise. Show that both matrices for this projection operator fail to have
orthonormal columns.]
A Couple Loose Ends
has orthonormal
[Exercise. Prove
that,
for
orthonormal
bases
A
and
B,
if
U
A
columns, then U B does, too. Thus, the matrix condition for unitarity is independent
of orthonormal basis.]
[Exercise. Prove that is not true that a unitary U will have a matrix with
orthonormal columns for all bases. Do this by providing a counter-example as follows.
1. Use the rotation in R/2 in R2 and the non-orthogonal basis
1
2
,
= { d1 , d2 }
D =
0
1
.
2. Write down the matrix for R/2 D using the previously developed formula, true
for all bases,
#
"
R/2 D =
R/2 (d1 ) ,
R/2 (d2 )
D
whose four components are, of course, unknown to us. Well express them
temporarily as
.
D
129
3. Computing the matrix in the previous step requires a little calculation: you
cannot use the earlier trick
D
E
j T (b
k ) ,
Tjk
=
b
B
since that formula only applies to an orthonormal basis B, which D is not.
,
Instead, you must solve some simple equations. To get the matrix
for
R
/2
D
t
find the column vectors. The first column is R/2 (2, 0) (D-coordinates
D
taken after we have applied R/2 ). You can easily confirm that (in natural
coordinates),
2
0
R/2
=
,
2 A
0 A
by drawing a picture. Now, get the D-coordinates of this vector by solving the
A-coordinate equation,
0
2
1
=
+
2
0
1
for and .
4. Even though this will immediately tell you that R/2 D is not orthonormal, go
on to get the full matrix by solving for the second column. (, )t , and showing
that neither column is normalized and the two columns are not orthogonal.]
5.5.3
Hermitian Operators
While unitary operators are the linear transformations that can be used to represent
quantum logic gates, there are types of operators associated with the measurements
of a quantum state. These are the called Hermitian operators.
Preview Characterization. In quantum mechanics, a Hermitian operator will
be an operator that is associated with some observable, that is, something about
the system, such as velocity, momentum, spin that we can imagine measuring. For
quantum computer scientists, the observable will be state of a quantum bit.
We will fully explore the connection between Hermitian operators and observables in the very next lesson (quantum mechanics), but right now lets least see the
mathematical definition of a Hermitian operator.
130
Definitions and Examples
Hermitian Matrices. A matrix, M , which is equal to its adjoint is
called Hermitian, i.e.,
M is Hermitian
M.
[Exercise. Show that the matrices
1 0
3 0
0 2 1 0
3 1 3 0
0 0
0
1
1i
1 + i
7 + 7i
2.5
7 7i
88
and
are Hermitian.]
[Exercise. Explain why the
1+i
0
3
2.5
7 7i
matrices
6
7i
88
and
1 0
3 0
0 2 1 0
3 1 3 0
are not Hermitian.]
[Exercise. Explain why a Hermitian matrix has real elements along its diagonal.]
Hermitian Operators. A Hermitian operator is a linear transformation, T , whose matrix, MT (in any basis) is Hermitian.
The definition of Hermitian operator implies that the basis chosen does not matter:
either all of T s matrices in all bases will be Hermitian or none will be. This is a fact,
but we wont bother proving it.
Hermitian operators play a starring role in quantum mechanics, as well see shortly.
5.6
Enter the Quantum World
Well you did it. After a few weeks of intense but rewarding math, you are ready to
learn some fully caffeinated quantum mechanics. Ill walk you through it in a single
chapter that will occupy us for about a week at which time youll be fully a certified
quantum mechanic.
131
Chapter 6
The Experimental Basis of
Quantum Computing
6.1
The Physical Underpinning for Spin 1/2 Quantum Mechanics
This is the first of a three-chapter lesson on quantum mechanics. For our purposes
in CS 83A, only the second of the three the next chapter contains the essential
formalism that well be using in our algorithms. However, a light reading of this first
chapter will help frame the theory that comes next.
6.2
Physical Systems and Measurements
6.2.1
Quantum Mechanics as a Model of Reality
Quantum mechanics is not quantum physics. Rather, it is the collection of mathematical tools used to analyze physical systems which are, to the best of anyones
ability to test, known to behave according to the laws of quantum physics.
As computer scientists, we will not concern ourselves with how the engineers
implement the physical hardware that exhibits predictable quantum behavior (any
more than we needed to know how they constructed a single classical bit capable of
holding a 1 or a 0 out of beach sand in order to write classical software or algorithms).
Certainly this is of great interest, but it wont interfere with our ability to play our
part in moving the field of quantum information forward. As of this writing, we dont
know which engineering efforts will result in the first generation of true quantum
hardware, but our algorithms should work regardless of the specific solution they end
up dropping at our doorstep.
That said, a brief overview of the physics will provide some stabilizing terra firma
to which the math can be moored.
132
6.2.2
The Physical System, S
Heres the set-up. We have a physical system, call it S (thats a script S), and an
apparatus that can measure some property of S . Also, we have it on good authority
that S behaves according to quantum weirdness; 100 years of experimentation has
confirmed certain things about the behavior of S and its measurement outcomes.
Here are some examples the last few may be unfamiliar to you, but I will elaborate,
shortly.
The system is a proton and the measurement is the velocity ( momentum) of
the proton.
The system is a proton and the measurement is the position of the proton.
The system is a hydrogen atom (one proton + one electron) and the measurement is the potential energy state of the atom.
The system is an electron and the measurement is the z-component magnitude
of the electrons spin.
The system is an electron and the measurement is the x-component magnitude
of the electrons spin.
The system is an electron and the measurement is the magnitude of the elec.
trons spin projected onto the direction n
In each case, we are measuring a real number that our apparatus somehow is
capable of detecting. In practice, the apparatus is usually measuring something related
to our desired quantity, and we follow that with a computation to get the value of
interest (velocity, momentum, energy, z-component of spin, etc.).
6.2.3
Electron Spin as a Testbed for Quantum Mechanics
One usually studies momentum and position as the relevant measurable quantities,
especially in a first course in quantum mechanics. For us, an electrons spin will be
a better choice.
A Reason to Learn Quantum Mechanics Through Spin
The Hilbert spaces that we get if we measure momentum or position are infinite dimensional vector spaces, and the corresponding linear combinations become integrals
rather than sums. We prefer to avoid calculus in this course. For spin, however, our
vector spaces are two dimensional about as simple as they come. Sums work great.
133
A Reason Computer Scientists Need to Understand Spin
The spin of an electron, which is known as a spin 1/2 particle, has exactly the kind
of property that we can incorporate into our algorithms. It will have classical aspects
that allow it to be viewed as a classical bit (0 or 1), yet it still has quantum-mechanical
aspects that enable us process it while it is in an in-between state a mixture of 0
and 1.
Spin Offers a Minor Challenge
Unlike momentum or position, spin is not part of common vernacular, so we have
to spend a moment defining it (in a way that will necessarily be incomplete and
inaccurate since we dont have a full 12 weeks to devote to the electromagnetism,
classical mechanics and quantum physics required to do it justice).
6.3
6.3.1
A Classical Attempt at Spin 1/2 Physics
An Imperfect Picture of Spin
The Scalar and Vector Sides of Spin
Spin is a property that every electron possesses. Some properties like charge and
mass are the same for all electrons, while others like position and momentum vary
depending on the electron in question and the exact moment at which we measure.
The spin or more accurate term spin state of an electron has aspects of both.
There is an overall magnitude associated with an electrons spin state that does not
change. It is represented by the number 1/2, a value shared by all electrons at all
times. But then each election can have its own unique vector orientation that varies
from electron-to-electron or moment-to-moment. Well sneak up on the true quantum
definition of these two aspects of quantum spin in steps by first by thinking of an
electron using inaccurate but intuitive imagery, and well make adjustments as we
perform experiments that progressively force us to change our attitude.
Electromagnetic Fields Use in Learning About Spin States
Electrons appear to interact with external electromagnetic fields as if they were
tiny, negatively charged masses rotating at a constant speed about an internal axis.
While they truly are negatively charged and massive, no actual rotation takes place.
Nevertheless, the math resulting from this imagery can be used as a starting point to
make predictions about how they will behave in such a field. The predictions wont
be right initially, but theyll lead us to better models.
134
A Useful Mental Image of Spin
We indulge our desire to apply classical physics and with an understanding that
its not necessarily true consider the electron to be a rotating, charged body. Such
an assumption would imbue every electron with an intrinsic angular momentum that
we call spin. If you have not studied basic physics, you can imagine a spinning top; it
has a certain mass distribution, spins at a certain rate (frequency) and its rotational
axis has a certain direction. Combining these three things into a single vector, we
end up defining the angular momentum of the top or, in our case, the spin of the
electron. (See Figure 6.1.)
Figure 6.1: Classical angular momentum
6.3.2
A Naive Quantitative Definition of Electron Spin
Translating our imperfect model into the language of math, we initially define an
electrons spin (state) to be the vector S which embodies two ideas,
1. first, its quantity of angular momentum (how heavy it is combined with how
fast its rotating), which I will call S, and
2. second, the orientation (or direction) of its imagined rotational axis, which I
S or sometimes just S.
will call n
S , can be represented by a unit vector
The first entity, S, is a scalar. The second, n
that points in the direction of the rotational axis (where we adjudicate up vs. down
by a right hand rule, which I will let you recall from any one of your early math
classes). (See Figure 6.2.)
So, the total spin vector will be written
Sx
Sy ,
Sz
135
Figure 6.2: A classical idea for spin: A 3-D direction and a scalar magnitude
and we can break it into the two aspects, its scalar magnitude,
q
Sx 2 + Sy 2 + Sz 2 ,
S |S| =
and a unit vector that embodies only its orientation (direction)
S = S
n
S
.
|S|
S, the spin magnitude,
is the same for all electrons under all conditions. For the
3
record, its value is 2 ~, where ~ is a tiny number known as Planks constant. (In
street terminology thats spin 1/2.) After today, we wont rely on the exact expression, but for now we will keep it on the books by making explicit the relationship
!
3
S .
~ n
S =
2
as the only
The constancy of its magnitude leaves the electrons spin orientation, S,
spin-related entity that can change from moment-to-moment or electron-to-electron.
6.3.3
Spherical Representation
Figure 6.3: Polar and azimuthal angles for the (unit) spin direction
You may have noticed that we dont need all three components nx , ny and nz .
S is unit length, the third can be derived from the other two. [Exercise.
Since n
How?] A common way to express spin direction using only two real numbers is
through the so-called polar and azimuthal angles, and (See Figure 6.3).
136
In fact, Spherical coordinates provide an alternate means of expressing any vector
using these two angles plus the vectors length, r. In this language a vector can
be written ( r, , )Sph , rather than the usual (x, y, z). (The subscript Sph is
usually not shown if the context makes clear we are using spherical coordinates.) For
example,
3
1
1 =
.
.615
1
/4 Sph
S is written ( 1, , )Sph , where
In the language of spherical coordinates the vector n
S has unit length. The two remaining
the first coordinate is always 1 because n
coordinates are the ones we just defined, the angles depicted in Figure 6.3.
and will be important alternatives to the Euclidean coordinates (nx , ny , nz ),
especially as we study the Bloch sphere, density matrices and mixed states, topics in
the next quantum computing course, CS 83B.
6.4
Refining Our Model: Experiment #1
We proceed along classical grounds and, with the aid of some experimental physicists,
design an experiment.
6.4.1
The Experiment
We prepare a bunch of electrons in equiprobable random states, take some measurements of their spin states and compare the result with what we would expect
classically. Here are the details:
1. The States. Lets instruct our experimental physicists to prepare an electron
soup: billions upon billions of electrons in completely random spin orientations.
No one direction (or range of directions) should be more represented than any
other.
Figure 6.4: A soup of electrons with randomly oriented spins
2. The Measurement. Even classically there is no obvious way to measure the
entire 3-D vector S at once; we have to detect the three scalar components,
137
Figure 6.5: The z-projection of one electrons spin
individually. Therefore, well ask the physicists to measure only the real valued
component Sz , the projection of S onto the z-axis. They assure us they can do
this without subjecting the electrons to any net forces that would (classically)
modify the z-component of S.
To aid the visualization, we imagine measuring each electron one-at-a-time and
noting the z-component of spin after each trial. (See Figure 6.5.)
3. The Classical Expectation. The
results are easy to predict classically since
electrons are oriented randomly;
the length of the spin is fixed at 23 ~, and the
3
3
we expect Sz to vary between + 2 ~ and 2 ~. For example, in one extreme
case, we could find
an electron whose spin
is oriented straight up, i.e., in the positive zdirection with Sz = + 23 ~ possessing 100% of the vectors length.
Similar logic implies the detection of other electrons
pointing straight down (Sz =
3
2
~),
lying entirely in the x-y plane (Sz = 0),
% pointing mostly up (Sz = +.8~),
& slightly down (Sz = .1~),
etc.
Figure 6.6: The Classical range of z-projection of spin
138
6.4.2
The Actual Results
We get no such cooperation from nature. In fact, we only and always get one of two
z-values of spin: Sz = +~/2 and Sz = ~/2. Furthermore, the two readings appear
to be somewhat random, occurring with about equal likelihood and no pattern:
~
~
~
~
~
~
+
+
+
+
2
2
2
2
2
2
Figure 6.7: The measurements force Sz to snap into one of two values.
6.4.3
The Quantum Reality
There are two surprises here, each revealing its own quantum truth that we will be
forced to accept.
Surprise #1. There are infinitely many quantum spin states available for electrons to secretly experience. Yet when we measure the z-component of
spin, the
~
uncooperative particle always reports that this value is either + 2 or ~2 ,
each choice occurring with about equal likelihood. We will call the z-component
of spin the observable Sz (observable indicating that we can measure it) and
accept from the vast body of experimental evidence that measuring the observable Sz forces the spin state to collapse such that its Sz snaps to either one
of the two
allowable values called eigenvalues of the observable Sz . Well call
the + ~2 outcome the +z outcome (or just the (+) outcome) and the ~2
outcome the z (or the ()) outcome.
Surprise #2. Even if we can somehow accept the collapse of the infinity of
random states into the two measurable ones, we cannot help but wonder why the
electrons projection onto the
z-axis is not the entire length
of
the vector, that
3
3
is, either straight up at + 2 ~ or straight down at 2 ~. The electron
stubbornly wants to give us only a fraction of that amount, 58%. This
corresponds to two groups. The up group which forms the angle 55
139
(0.955 rad) with the positive z-axis and down group which forms that same
55 angle with the negative z-axis. The explanation
for not being able to get
3
a measurement that has the full length 2 ~ is hard to describe without a
more complete study of quantum mechanics. Briefly it is due to the Heisenberg
uncertainty principle. If the spin were to collapse to a state that was any closer
to the vertical z-axis, we would have too much simultaneous knowledge about
its
and y- components (too close to 0) and its z-component (too close to
x
3
2 ~). (See Figure 6.8.) This would violate Heisenberg, which requires
the combined variation of these observables be larger than a fixed constant.
Therefore, Sz must give up some of its claim on the full spin magnitude, 23 ~,
Figure 6.8: Near vertical spin measurements give illegally accurate knowledge of
Sx , Sy and Sz .
resulting in a shorter Sz , specifically 12 ~.
6.4.4
A Follow-Up to Experiment #1
Before we move on to a new experimental set-up, we do one more test. The physicists
tell us that a side-effect of the experimental design was that the electrons which
measured a +z are now separated from those that measured z, and without building
any new equipment, we can subject each group to a follow-up Sz measurement. The
follow-up test, like the original, will not add any forces or torques that could change
the Sz (although we are now beginning to wonder, given the results of the first set of
trials). (See Figure 6.9.)
6.4.5
Results of the Follow-Up
We are reassured by the results. All of the electrons in the (+) group faithfully
measure +z, whether done immediately after the first test or after a period of time.
Similarly, the Sz value of the () group is unchanged; those rascals continue in their
down-z orientation. We have apparently created two special states for the electrons
which are distinguishable and verifiable. (See Figure 6.10.)
140
Figure 6.9: Plans for a follow-up to experiment # 1
Figure 6.10: Results of follow-up measurements of Sz on the +z (|+i) and z (|i)
groups.
|+i and |i States. We give a name to the state of the electrons in the (+)
group: we call it the |+iz state (or simply the |+i state, since we consider the z-axis
to be the preferred axis in which to project the spin). We say that the () group is
in the |iz (or just the |i) state. Verbally, these two states are pronounced plus
ket and minus ket.
6.4.6
Quantum Mechanics Lesson #1
Either by intuition or a careful study of the experimental record, we believe there
are infinitely many spin values represented by the electrons in our random electron
soup prepared by the physicists, yet when we measure the Sz component, each state
collapses into one of two possible outcomes. Repeated testing of Sz after the first
collapse, confirms that the collapsed state is stable: we have prepared the electrons
in a special spin-up or spin-down state, |+i or |i.
|+ix and |ix States. Certainly this would also be true had we projected
the spin vector onto the x-axis (measuring Sx ) or the y-axis (measuring Sy ). For
example, if we had tested Sx , we would have collapsed the electrons into a different
pair of states, |+ix and |ix (this time, we need the subscript x to distinguish it from
our preferred z-axis states {|+i , |i}).
Whats the relationship between the two z-states, {|+i , |i} and the x-states,
141
{|+ix , |ix }? Well need another experiment.
6.4.7
First Adjustment to the Model
The imagery of a spinning charged massive object has to be abandoned. Either our
classical physics doesnt work at this subatomic level or the mechanism of electron
spin is not what we thought it was or both. Otherwise, we would have observed
a continuum of values Sz values in our measurements. Yet the experiments that we
have been doing are designed to measure spin as if it were a 3-dimensional vector, and
the electrons are responding to something in our apparatus during the experiments,
so while the model isnt completely accurate, the representation of spin by a vector,
S (or equivalently, two angular parameters, and , and one scalar S) may still have
some merit. Were going to stop pretending that anything is really spinning, but well
stick with the vector quantity S and its magnitude S, just in case theres some useful
information in there.
6.5
Refining Our Model: Experiment #2
The output of our first experiment will serve as the input of our second. (See Figure 6.11.)
Figure 6.11: Experiment #2: |+iz electrons enter an Sx measurement apparatus.
6.5.1
The Experiment
We will take only the electrons that collapsed to |+i as input to a new apparatus. We
observed that this was essentially half of the original group, the other half consisting
of the electrons that collapsed into the state |i (which we now throw away).
Our second apparatus is going to measure the x-axis projection. Lets repeat this
using the same bullet-points as we did for the first experiment.
1. The States. The input electrons are in a specific state, |+i, whose z-spins always point (as close as possible to) up. This is in contrast to the first experiment
where the electrons were randomly oriented.
142
Figure 6.12: The input state for experiment # 2
2. The Measurement. Instead of measuring Sz , we measure Sx . We are curious
Figure 6.13: The x-projection of spin
to know whether electrons that start in the z-up state have any x-projection
preference. We direct our physicists to measure only the real valued component
Sx , the projection of S onto the x-axis.
3. The Classical Expectation.
Figure 6.14: Viewed from top left, the classical range of x-projection of spin.
Clinging desperately to those classical ideas which have not been ruled out by
the first experiment, we imagine Sx and Sy to be in any relative amounts that
complement |Sz |, now firmly fixed at ~2 . This would allow values for those two
q
such that |Sx |2 + |Sy |2 + 14 ~2 = 23 ~. If true, prior to this second measurement
Sx would be smeared over a range of values. (See Figure 6.14.)
143
6.5.2
The Actual Results
As before, our classical expectations are dashed. We get one of two x-values of spin:
Sx = +~/2 or Sx = ~/2. And again the two readings occur randomly with near
equal probability:
~
~
~
~
~
~
+
+
+
.
2
2
2
2
2
2
Also, when we subject each output group to further Sx tests, we find that after the
first Sx collapse each group is locked in its own state as long as we only test Sx .
6.5.3
The Quantum Reality
Our first experiment prepared us to expect the non-classical. Apparently, when an
electron starts out in a |+iz spin state, an Sx measurement causes it to collapse into
a state such that we only see one of two allowable eigenvalues of the observable Sx ,
namely + ~2 (the (+) outcome) and ~2 (the () outcome).
Despite the quantum results, we believe we have at least isolated two groups of
electrons, both in the |+i (z-up) state, but one in the |ix (x-down) state and the
other in the |+ix (x-up). (See Figure 6.15.) However, we had better test this belief.
Figure 6.15: A guess about the states of two groups after experiment #2
6.5.4
A Follow-Up to Experiment #2
Using the two detectors the physicists built, we hone in on one of the two output
groups of experiment, #2, choosing (for variety) the |ix group. We believe it to
contain only electrons with Sz = (+) and Sx = (). In case you are reading too
fast, thats z-up and x-down. To confirm, we run this group through a third test that
measures Sz again, fully anticipating a boring run of (+) readings.
144
6.5.5
Results of the Follow-Up
Here is what we find:
~
~
~
~
~
~
+
+
+
2
2
2
2
2
2
We are speechless. There are now equal numbers of z-up and z-down spins in a group
of electrons that we initially selected from a purely z-up pool. (See Figure 6.16.)
Furthermore, the physicists assure us that nothing in the second apparatus could
Figure 6.16: |i electrons emerging from a group of |+i electrons
have turned the electrons viz-a-viz the z-component there were no net z-direction
forces.
6.5.6
Quantum Mechanics Lesson #2
Even if we accept the crudeness of only two measurable states per observable, we
still cant force both Sz and Sx into specific states at the same time. Measuring Sx
destroys all information about Sz . (See Figure 6.17.)
After more experimentation, we find this to be true of any pair of distinct directions, Sx , Sy , Sz , or even Sn , where n
is some arbitrarily direction onto which
we project S. There is no such thing as knowledge of three independent directional
components such as Sx , Sy , and Sz . The theorists will say that Sz and Sx (and any
such pair of different directions) are incompatible observables. Even if we lower our
expectations to accept only discrete (in fact binary) outcomes |+i, |+ix , |iy , etc.
we still cant know or prepare two at once.
[Exception. If we select for |+ix (or any |+in ) followed by a selection for its
polar opposite, |ix (or |in ), there will be predictably zero electrons in the final
output.]
145
Figure 6.17: The destruction of |+i Sz information after measuring Sx
Preview
As we begin to construct a two-dimensional Hilbert space to represent spin, we already
have a clue about what all this might mean. Since, the |ix output of the second
apparatus was found (by the third apparatus) to contain portions of both |+i and
|i, it might be expressible by the equation
|ix
c+ |+i
c |i ,
where c+ and c are scalar weights which express how much |+i and |i constitute
|ix . Furthermore, it would make sense that the scalars c+ and c have equal magnitude if they are to reflect the observed (roughly) equal number of z-up and z-down
spins detected by the third apparatus when testing the |ix group.
We can push this even further. Because we are going to be working with normalized vectors (recall the projective
sphere in Hilbert space?), it will turn out that their
common magnitude will be 1/ 2.
For this particular combination of vector states,
|ix
1
1
|+i |i
2
2
|+i |i
.
2
In words (that we shall make precise in a few moments), the |ix vector can be
expressed as a linear combination of the Sz vectors |+i and |i. This hints at the
idea that |+i and |i form a basis for a very simple 2-dimensional Hilbert space, the
foundation of all quantum computing.
6.5.7
Second Adjustment to the Model
Every time we measure one of the three scalar coordinates Sx , Sy , or Sz , information
we thought we had about the other two seems to vanish. The three dimensional vector
146
S is crumbling before our eyes. In its place, a new model is emerging that of a two
dimensional vector space whose two basis vectors appear to be the two z-spin states
|+i and |i which represent a quantum z-up and quantum z-down, respectively. This
is a difficult transition to make, and Im asking you to accept the concept without
trying too hard to visualize it in your normal three-dimensional world view. Here are
three counter-intuitive ideas, the seeds of which are present in the recent outcomes
of experiment #2 and its follow-up:
1. Rather than electron spin being modeled by classical three dimensional unit
vectors
Sx
1
Sy =
Sz
Sph
in a real vector space with basis {
x, y
,
z}, we are heading toward a model
where spin states are represented by two dimensional unit vectors
c+
c
in a complex vector space with basis { |+iz , |iz }.
2. In contrast to classical spin, where the unit vectors with z-components +1 and
-1 are merely scalar multiples of the same basis vector
z,
0
0
0 = (+1)
0 = (1)
z
and
z
+1
1
we are positing a model in which the two polar opposite z-spin states, |+i = |+iz
and |i = |iz , are linearly independent of one another.
3. In even starker contrast to classical spin, where the unit vector x
is linearly
independent of
z, the experiments seem to suggest that unit vector |ix can be
formed by taking a linear combination of |+i and |i, specifically
|ix
|+i |i
.
2
But dont give up on the spherical coordinates and just yet. They have a role
to play, and when we study expectation values and the Bloch sphere, youll see what
that role is. Meanwhile, we have one more experiment to perform.
6.6
6.6.1
Refining Our Model: Experiment #3
What we Know
We have learned that measuring one of the scalars related to spin, Sz , Sx or any Sn ,
has two effects:
147
1. It collapses the electrons into one of two
spin states for that observable, one that
produces a reading of up (+) = + ~2 , and the other that produced a reading
of down () = ~2 .
2. It destroys any information about the other spin axes, or in quantum-ese, about
the other spin observables.
(Up/down might be better stated as left/right when referring to x-coordinates,
and forward-backward when referring to y-coordinates, but its safer to use the
consistent terminology of spin-up/spin-down for any projection axis, not just the
z-axis.)
There is something even more subtle we can learn, and this final experiment will
tease that extra information out of the stubborn states.
6.6.2
The Experiment
We can prepare a pure state that is between |+i and |i if we are clever. We direct
our physicists to rotate the first apparatus relative to the second apparatus by an
angle counter-clockwise from the z-axis (axis of rotation being the y-axis, which
ensures that the x-z plane is rotated into itself (see Figure 6.18).
Figure 6.18: A spin direction with polar angle from +z, represented by |i
Lets call this rotated state |i, just so it has a name that is distinct from |+i
and |i.
If we only rotate by a tiny , we have a high dose of |+i and a small dose of |i in
our rotated state, |i. On the other hand, if we rotate by nearly 180 ( radians), |i
would have mostly |i and very little |+i in it. Before this lesson ends, well prove
that the right way to express the relationship between and the relative amounts of
|+i and |i contained in |i is
|i = cos
|+i + sin
|i .
2
2
148
By selecting the same (+) group coming out of the first apparatus (but now tilted at
an angle ) as input into the second apparatus, we have effectively changed our input
states going into the second apparatus from purely |+i to purely |i.
We now measure Sz , the spin projected onto the z-axis. The exact features of
what I just described can be stated using the earlier three-bullet format.
1. The States. This time the input electrons are in a specific state, |i, whose
z-spin direction forms an angle from the z-axis (and for specificity, whose
spherical coordinate for the azimuthal angle, , is 0). (See Figure 6.19.)
Figure 6.19: The prepared state for experiment #3, prior to measurement
2. The Measurement. We follow the preparation of this rotated state with an
Sz measurement.
3. The Classical Expectation. Weve been around the block enough to realize
that we shouldnt expect aclassical result. If this were
a purely classical situa
3
3
tion, the spin magnitude, 2 ~, would lead to Sz = 2 ~ cos (). But we already
know the largest Sz ever reads is 21 ~, so maybe we attenuate that number
by cos , and predict ~2 cos (). Those are the only two ideas we have at the
moment. (See Figure 6.20.)
Figure 6.20: Classical, or semi-classical expectation of experiment #3s measurement
149
6.6.3
The Actual Results
It is perhaps not surprising that we always read one of two z-values of spin: Sz = +~/2
and Sz = ~/2. The two readings occur somewhat randomly:
~
~
~
~
~
~
+
+
+
+
2
2
2
2
2
2
However, closer analysis reveals that they are not equally likely. As we try different
s and tally the results, we get the summary shown Figure 6.21.
Figure 6.21: Probabilities of measuring |i from starting state |i, from +z
In other words,
~
% outcomes which = +
2
~
% outcomes which =
2
cos
100%
2
2
sin
100% .
2
2
and
Notice how nicely this agrees with our discussion of experiment #2. There, we
prepared a |ix state to go into the final Sz tester. |+ix is intuitively 90 from the
z-axis, so in that experiment
our was 90 . That would make 2 = 45 , whose cosine
and sine are both = 22 . For these values, the formula above gives a predicted 50%
(i.e., equal) frequency to each outcome, (+) and (), and thats exactly what we
found when we measured Sz starting with |ix electrons.
6.6.4
The Quantum Reality
Apparently, no matter what quantum state our electrons spin is in, we always measure the magnitude projected onto an axis as ~/2. We suspected this after the
first two experiments, but now firmly believe it. The only thing we havent tried is
to rotate the first apparatus into its most general orientation, one that includes a
non-zero azimuthal , but that would not change the results.
150
This also settles a debate that you might have been waging, mentally. Is spin, prior
to an Sz measurement, actually in some combination of the states |+i and |i? Yes.
Rotating the first apparatus relative to the second apparatus by a particular has a
physical impact on the outcomes. Even though the electrons collapse into one of |+i
and |i after the measurement, there is a difference between |i = .45 |+i + .89 |i
and | 0 i = .71 |+i + .71 |i: The first produces 20% (+) measurements, and the
second produces 50% (+) measurements.
6.6.5
Quantum Mechanics Lesson #3
If we prepare a state to be a normalized (length one) linear combination of the two
Sz states, |+i and |i, the two coefficients predict the probability of obtaining each
of the two possible eigenvalues ~2 for an Sz measurement. This is the meaning of
the coefficients of a state when expressed relative to the preferred z-basis.
This holds for any direction. If we wish to measure the projection of S onto
an arbitrary direction n
(in quotes because Im using slightly inaccurate classical
terminology), then we would expand our input state along the n
state vectors,
|i
|+in + |in ,
and ||2 and ||2 would give the desired probabilities.
The quantum physicist will tell us that spin of an electron relative to a predetermined direction can be in any of infinitely many states. However, when we
measure, it will collapse into one of two possible states, one up and one down. Furthermore the relative number of the up measurements vs. down measurements is
predicted precisely by the magnitude-squared of the coefficients of the prepared state.
Finally, we can use any two polar opposite directions in R3 (which we shall learn
correspond to two orthonormal vectors in C2 ) as a basis for defining the spin state.
{ |+i , |i }, the z-component measurements, are the most common, and have come
to be regarded as the preferred basis, but we can use { |+ix , |ix }, or any two oppositely directed unit vectors (i.e., orientations). Such pairs are considered alternate
bases for the space of spin states. In fact, as quantum computer scientists, well use
these alternate bases frequently.
6.6.6
Third Adjustment to the Model
Were now close enough to an actual model of quantum spin 1/2 particles to skip
the game of making a minor adjustment to our evolving picture. In the next lesson
well add the new information about the meaning of the scalar coefficients and the
probabilities they reflect and give the full Hilbert space description of spin 1/2 physics
motivated by our three experiments.
Also, were ready to leave the physical world, at least until we study time evolution.
We can thank our experimental physicists and let them get back to their job of
designing a quantum computer. We have a model to build and algorithms to design.
151
6.7
Onward to Formalism
In physics the word formalism refers to the abstract mathematical notation necessary to scribble predictions about what a physical system such as a quantum computer
will do if we build it. For our purposes, the formalism is what we need to know, and
thats the content of the next two chapters. They will provide a strict but limited set
of rules that we can use to accurately understand and design quantum algorithms.
Of those chapters, only the next (the second of these three) is necessary for CS
83A, but what you have just read has prepared you with a sound intuition about
the properties and techniques that constitute the formalism of quantum mechanics,
especially in the case of the spin 1/2 model used in quantum computing.
152
Chapter 7
Time Independent Quantum
Mechanics
7.1
Quantum Mechanics for Quantum Computing
This is the second in a three-chapter introduction to quantum mechanics and the only
chapter that is required for the first course, CS 83A. A brief and light reading of the
prior lesson in which we introduce about a half dozen conceptual experiments will
help give this chapter context. However, even if you skipped that on first reading,
you should still find the following to be self-contained.
Our goal today is twofold.
1. We want you to master the notation used by physicists and computer scientists
to scribble, calculate and analyze quantum algorithms and their associated logic
circuits.
2. We want you to be able to recognize and make practical use of the direct correspondence between the math and the physical quantum circuitry.
By developing this knowledge, you will learn how manipulating symbols on paper
affects the design and analysis of actual algorithms, hardware logic gates and measurements of output registers.
7.2
The Properties of Time-Independent Quantum
Mechanics
Lets do some quantum mechanics. Well be stating a series of properties Ill call
them traits that will either be unprovable but experimentally verified postulates, or
provable consequences of those postulates. Well be citing them all in our study of
quantum computing.
153
In this lesson, time will not be a variable; the physics and mathematics pertain
to a single instant.
7.2.1
The State Space of a Physical System
Physical System (Again)
Lets define (or re-define if you read the last chapter) a physical system S to be some
conceptual or actual apparatus that has a carefully controlled and limited number of
measurable states. Its implied that there are very few aspects of S that can change
or be measured, since otherwise it would be too chaotic to lend itself to analysis.
Consider two examples. The first is studied early in an undergraduate quantum
mechanics course; we wont dwell on it. The second forms the basis of quantum
computing, so well be giving it considerable play time.
A Particles Energy and Position in One-Dimension
We build hardware that allows a particle (mass = m) to move freely in one dimension
between two boundaries, say from 0 cm to 5 cm. The particle cant get out of that 1dimensional box but is otherwise free to roam around inside (no forces acting). We
build an apparatus to test the particles energy. Using elementary quantum mechanics
we discover that the particles energy measurement can only attain certain discrete
values, E0 , E1 , E2 , . . . . Furthermore, once we know which energy, Ek , the particle
has, we can form a probability curve that predicts the likelihood of finding the particle
at various locations within the interval [0, 5].
Spin: A Stern-Gerlach Apparatus
We shoot a single silver atom through a properly oriented inhomogeneous magnetic
field. The atoms deflection, up or down, will depend on the atoms overall angular
momentum. By placing two or more of these systems in series with various angular
orientations, we can study the spin of the outermost electron in directions Sz , Sx ,
Sy or general Sn . While not practical for quantum computation, this apparatus is
exactly how the earlier spin-1/2 results were obtained.
[Optional.] If you are curious, here are a few details about the Stern-Gerlach
apparatus.
Silver Atoms Work Well. You might think this is too complex a system for
measuring a single electrons spin. After all, there are 47 electrons and each has
spin not to mention a confounding orbital angular momentum from its motion
about the nucleus. However, orbital angular momentum is net-zero due to
statistically random paths, and the spin of the inner 46 electrons cancel (they
are paired, one up, one down). This leaves only the spin of the outermost
154
electron #47 to account for the spin of the atom as a whole. (Two other facts
recommend the use of silver atoms. First, we cant use charged particles since
the so-called Lorentz force would overshadow the subtle spin effects. Second,
silver atoms are heavy enough that their deflection can be calculated based
solely on classical equations.)
We Prepare a Fixed Initial State. An atom can be prepared in a spin state
associated with any pre-determined direction, n
, prior to subjecting it to a final
Stern-Gerlach tester. We do this by selecting a |+i electron from a preliminary
Stern-Gerlach Sz tester then orient a second tester in an n
direction relative
the original.
The Measurements and Outcomes. The deflection of the silver atom is detected
as it hits a collector plate at the far end of the last apparatus giving us the
measurements ~2 , and therefore the collapsed states |+ix , |in , etc. The
results correspond precisely with our experiments #1, #2 and #3 discussed
earlier.
Stern-Gerlach is the physical system to keep in mind as you study the math that
follows.
7.3
7.3.1
The First Postulate of Quantum Mechanics
Trait #1 (The State Space)
For any physical system, S , we associate a Hilbert space, H. Each physical state in
S corresponds to some ray in H, or stated another way, a point on the projective
sphere (all unit vectors) of H.
physical state S
v H,
|v| = 1 .
The Hilbert space H is often called the state space of the system S , or just the state
space.
Vocabulary and Notation. Physicists and quantum computer scientists alike
express the vectors in the state space using ket notation. This means a vector in
state space is written
| i,
and is usually referred to as a ket. The Greek letter is typically used to label any
old state. As needed we will be replacing it with specific and individual labels when
we want to differentiate two state vectors, express something known about the vector,
or discuss a famous vector that is universally labeled. Examples we will encounter
include
| a i , | uk i , | + i ,
155
and
| + iy .
When studying Hilbert spaces, I mentioned that a single physical state corresponds to an infinite number of vectors, all on the same ray, so we typically choose
a normalized representative having unit length. Its the job of quantum physicists
to describe how to match the physical states with normalized vectors and ours as
quantum computer scientists to understand and respect the correspondence.
7.3.2
The Fundamental State Space for Quantum Computing
Based on our experiments involving spin-1/2 systems from the (optional) previous
chapter, we were led to the conclusion that any spin state can be described by a
linear combination of two special states |+i and |i. If you skipped that chapter,
this will serve to officially define those two states. They are the natural basis kets of a
2-dimensional complex Hilbert space, H. In other words, we construct a simple vector
space of complex ordered pairs with the usual complex inner product and decree the
natural basis to be the two measurable states |+i and |i in S . Symbolically,
, C
with
H
1
|+i
and
0
0
|i
.
1
In this regime, any physical spin state |i S can be expressed as a normalized
vector expanded along this natural basis using
|i =
= |+i + |i ,
where
||2 + ||2
1.
The length requirement reflects that physical states reside on the projective sphere of
H.
[Exercise. Demonstrate that { |+i , |i } is an orthonormal pair. Caution:
Even though this may seem trivial, be sure you are using the complex, not the real,
inner product.]
The Orthonormality Expressions
In the heat of a big quantum computation, basis kets will kill each other off, turning
themselves into the scalars 0 and 1 because the last exercise says that
h+ | +i
h+ | i
=
=
h | i
h | +i
156
=
=
1,
0.
and
While this doesnt rise to the level of trait, memorize it. Every quantum mechanic
relies on it.
[Exercise. Demonstrate that the set { |+i , |i } forms a basis (the z-basis) for
H. Hint: Even though only the projective sphere models S , we still have to account
for the entire expanse of H including all the vectors off the unit-sphere if we are going
to make claims about spanning the space.]
The x-Basis for H
Lets complete a thought we started in our discussion of experiment #2 from last
time. We did a little hand-waving to suggest
|ix
|+i |i
,
2
but now we can make this official (or, if you skipped the last chapter, let this serve
as the definition of two new kets).
1
1
|ix
,
2 1
that is, it is the vector in H whose coordinates along the z-bases are as shown. We
may as well define the |+ix vector. It is
1
|+i + |i
1
=
.
|+ix =
2
2 1
[Exercise. Demonstrate that { |+ix , |ix } is an orthonormal pair.]
[Exercise. Demonstrate that the set { |+ix , |ix } forms a basis (the x-basis)
for H.]
7.3.3
Why Does a Projective H model Spin-1/2?
There may be two questions (at least) that you are inclined to ask.
1. Why do we need a complex vector space as opposed to a real one to represent
spin?
2. Why do spin states have to live on the projective sphere? Why not any point
in H or perhaps the sphere of radius 94022?
I can answer item 1 now (and item 2 further down the page). Obviously, there was
nothing magical about the z-axis or the x-axis. I could have selected any direction
in which to start my experiments at the beginning of the last lesson and then picked
any other axis for the second apparatus. In particular, I might have selected the
157
same z-axis for the first measurement, but used the y-axis for the second one. Our
interpretation of the results would then have suggested that |iy contain equal parts
|+i and |i,
|iy
c+ |+i
c |i ,
and similarly for |+iy . If we were forced to use real scalars, we would have to pick the
same two scalars 12 for c (although we could choose which got the + and which
got the sign, a meaningless difference). Wed end up with
1
1
(warning: not true) ,
|iy
1
2
which would force them to be identical to the vectors |+ix and |ix , perhaps the
order of the vectors, reversed. But this cant be true, since the x-kets and the y-kets
can no more be identical to each other than either to the z-kets, and certainly neither
are identical to the z-kets. (If they were, then repeated testing would never have split
the original |+i into two equal groups, |+ix and|ix .) So there are just not enough
real numbers to form a third pair of basis vectors in the y-direction, distinct from the
x-basis and the z-basis.
If we allow complex scalars, the problem goes away. We can define
1
1
|iy
,
2 i
Now all three pairs are totally different orthonormal bases for H, yet each one contains
equal amounts of |+i and |i.
7.4
The Second Postulate of Quantum Mechanics
Trait #1 seemed natural after understanding the experimental outcomes of the spin1/2 apparatus. Trait #2 is going to require that we cross a small abstract bridge,
but I promise, you have all the math necessary to get to the other side.
7.4.1
Trait #2 (The Operator for an Observable)
An observable quantity A (in S ) corresponds to an operator a linear transformation
in H. The operator for an observable is always Hermitian.
[Note. There is a Trait #2a that goes along with this, but you dont have the
vocabulary yet, so we defer for the time being.]
In case you forgot what Hermitian means, the following symbolic version of Trait
#2, which Ill call Trait #2, should refresh your memory.
158
Trait #2 (Mathematical Version of Operator for an Observable):
Observable A S
TA : H H linear and TA = TA .
[Review. TA = TA is the Hermitian condition. The Hermitian condition on a
matrix M means that M s adjoint (conjugate transpose), M , is the same as M . The
Hermitian condition on a linear transformation means that its matrix (in any basis)
is its own adjoint. Thus, Trait #2 says that the matrix, TA , which purports to
correspond to the observable A, must be self-adjoint. Well reinforce all this in the
first example.]
Dont feel bad if you didnt expect this kind of definition; nothing we did above
suggested that there were linear transformations behind the observables Sz , Sx , etc.
never mind that they had to be Hermitian. Im telling you now, there are and they
do. This is an experimentally verified observation, not something we can prove, so it
is up to the theorists to guess the operator that corresponds to a specific observable.
Also, it is not obvious what we can do with the linear transformation of an observable,
even if we discover what it is. All in due time.
7.4.2
The Observable Sz
The linear transformation for Sz , associated with the z-component of electron spin
(a concept from the optional previous chapter), is represented by the matrix
~
0
~ 1 0
2
.
=
2 0 1
0 ~2
Blank stares from the audience.
Never fear. Youll soon understand why multiplication by this matrix is the operator (in H) chosen to represent the observable Sz (in the physical system S ). And
if you did not read that chapter, just take this matrix as the definition of Sz .
Notation. Sometimes we abandon the constant ~/2 and use a simpler matrix to
which we give the name z a.k.a. the Pauli spin matrix (in the z-direction),
1 0
z
.
0 1
Meanwhile, the least we can do is to confirm that the matrix for Sz is Hermitian.
[Exercise. Prove the matrix for Sz is Hermitian.]
We will now start referring to
i) the observable spin projected onto the z-axis,
159
ii) its associated linear operator, and
iii) the matrix for that operator
all using the one name, Sz .
Which Basis? Everything we do in this chapter and in the entire the volume
assumes that vectors and matrices are represented in the z-basis, which is the
preferred basis that emerges from the Sz observable unless explicitly stated. There
are other bases, but well label things if and when we expand along them.
As for the linear operators associated with the measurements of the other two
major spin directions Sx and Sy , they turn out to be
~ 0 1
Sx
2 1 0
~ 0 i
Sy
,
2 i 0
both represented in terms of the z-basis (even though these operators do define their
own bases).
Without the factor of ~/2 these, too, get the title of Pauli spin matrices (in the xand y-directions, respectively). Collecting the three Pauli spin matrices for reference,
we have
1 0
0 1
0 i
z =
x =
y =
.
0 1
1 0
i 0
Tip. This also demonstrates a universal truth. Any observable expressed it its
own basis is always a diagonal matrix with the eigenvalues appearing along that
diagonal. Because the actual observables matrices are the Pauli operators with the
factor of ~/2 out front, we can see that the eigenvalues of Sz do, in fact, appear along
that matrixs diagonal.
7.5
The Third Postulate of Quantum Mechanics
This trait will add some math vocabulary and a bit of quantum computational skill
to our arsenal.
7.5.1
Trait #3 (The Eigenvalues of an Observable)
The only possible measurement outcomes of an observable quantity A are special real
numbers, a1 , a2 , a3 , . . ., associated with the operators matrix. The special numbers
a1 , a2 , a3 , . . . are known as eigenvalues of the matrix.
[Note to Physicists. Im only considering finite or countable eigenvalues because
our study of quantum computing always has only finitely many special ak . In physics,
160
when measuring position or momentum, the set of eigenvalues is continuous (nonenumerable).]
Obviously, we have to learn what an eigenvalue is. Whatever it is, when we
compute the eigenvalues of, say, the matrix Sz , this trait tells us that they must be
~/2, since those were the values we measured when we tested the observable Sz . If
we had not already done the experiment but knew that the matrix for the observable
was Sz above, Trait #3 would allow us to predict the possible outcomes, something
we will do now.
7.5.2
Eigenvectors and Eigenvalues
Given any complex or real matrix M , we can often find certain special non-zero
vectors called eigenvector s for M . An eigenvector u (for M ) has the property that
u
Mu
6
=
=
0
and
a u , for some (possibly complex) scalar a .
In words, applying M to the non-zero u results in a scalar multiple of u. The scalar
multiple, a, associated with u is called an eigenvalue of M , while u is the (an)
eigenvector belonging to a. This creates a possible collection of special eigenvectoreigenvalue pairs,
on or
n
uak ak
k=1
for any matrix M .
Note. There may be more than one eigenvector, u0a , u00a , u000
a , . . ., for a given
eigenvalue, a, but in this course, well typically see the eigenvalues and eigenvectors
uniquely paired. More on this, shortly.
[Exercise. Give two reasons we require eigenvectors to be non-zero. One should
be based purely on the arithmetic of any vector space, and a second should spring
from the physical need to use projective Hilbert spaces as our quantum system models.
Hint: Both reasons are very simple.]
[Exercise. Show that if u is an eigenvector for M , then any scalar multiple, u,
is also an eigenvector.]
In light of the last exercise, when we find an eigenvector for a matrix, we consider
all of its scalar multiples to be the same eigenvector; we only consider eigenvectors
to be distinct if they are not scalar multiples of one another. Since we will be working with projective Hilbert spaces, whenever we solve for an eigenvector, we always
normalize it (divide by its magnitude) to force it onto the projective sphere. From
there, we can also multiply by a unit-length complex number, ei , in order to put it
into a particularly neat form.
Vocabulary. If an eigenvalue, a, corresponds to only one eigenvector, ua , it is
called a non-degenerate eigenvalue. However, if a works for two or more (linearly independent, understood) eigenvectors, u0a , u00a , . . ., it is called a degenerate eigenvalue.
161
There are two facts that I will state without proof. (They are easy enough to be
exercises.)
Uniqueness. Non-Degenerate eigenvalues have unique (up to scalar multiple)
eigenvectors. Degenerate eigenvalues do not have unique eigenvectors.
Diagonality. When the eigenvectors of a matrix, M , form a basis for the vector
space, we call it an eigenbasis for the space. M , expressed as a matrix in its
own eigenbasis, is a diagonal matrix (0 everywhere except for diagonal from
position 1-1 to n-n).
[Exercise. Prove one or both of these facts.]
7.5.3
The Eigenvalues and Eigenvectors of Sz
Lets examine M = Sz . Most vectors (say, (1, 2)t , for example) are not eigenvectors
of Sz . For example,
~ 1 0
~ 1
1
1
=
6= a
2
2
2 0 1
2 2
for any complex scalar, a. However, the vector (1, 0)t is an eigenvector, as
~ 1 0
~ 1
1
=
0
2 0 1
2 0
demonstrates. It also tells us that ~2 is the eigenvalue associated with the vector
(1, 0)t , which is exactly what Trait #3 requires.
[Exercise. Show that (0, 1)t and ~2 form another eigenvector-eigenvalue pair for
Sz .]
This confirms that Trait #3 works for Sz ; we have identified the eigenvalues for
Sz and they do, indeed, represent the only measurable values of the observable Sz in
our experiments.
All of this results in a more informative variant of Trait #3, which Ill call Trait
#3.
Trait #3: The only possible measurement outcomes of an observable, A, are the
solutions {ak } to the eigenvector-eigenvalue equation
TA |uk i
ak |uk i .
The values {ak } are always real, and are called the eigenvalues of the observable,
while their corresponding kets, {|uk i} are called the eigenkets. If each eigenvalue has
a unique eigenket associated with it, the observable is called non-degenerate. On the
other hand, if there are two or more different eigenkets that make the equation true
162
for the same eigenvalue, that eigenvalue is called a degenerate eigenvalue, and the
observable is called degenerate.
You may be wondering why we can say that the eigenvalues of an observable are
always real when we have mentioned that, for a general matrix operator, we can get
complex eigenvalues. This is related to the theoretical definition of an observable
which requires it to be of a special form that always has real eigenvalues.
7.6
Computing Eigenvectors and Eigenvalues
We take a short detour to acquire the skills needed to compute eigenvalues for any
observable.
The Eigenvalue Theorem. Given a matrix M , its eigenvalues are solutions to
the system of simultaneous equations in the unknown
det (M I)
0.
The equation in the Eigenvalue Theorem is known as the characteristic equation
for M .
Proof. Assume that a is an eigenvalue for M whose corresponding eigenvector is
u. Then
Mu
Mu
(M aI) u
=
=
=
au
a Iu
0.
Keeping in mind that eigenvectors are always non-zero, we have shown that the matrix
M aI maps a non-zero u into 0. But thats the hypothesis of the Little Inverse
Theorem B of our matrix lesson, so we get
det (M aI)
0.
In words, a is a solution to the characteristic equation.
QED
This tells us to solve the characteristic equation for an observable in order to
obtain its eigenvalues. Getting the eigenvectors from the eigenvalues is more of an
art and is best done using the special properties that your physical system S imbues
to the state space H. For spin-1/2 state spaces, it is usually very easy and can be
done by hand.
7.6.1
The Eigenvalues and Eigenvectors of Sy
The characteristic equation is
det (Sy I)
163
0,
which is solved like so:
~ 0 i
2 i 0
0
~2 i
~i
2
2
~2
4
~
2
Of course, we knew the answer, because we did the experiments (and in fact, the
theoreticians crafted the Sy matrix based on the results of the experimentalists). Now
comes the fun part. We want to figure out the eigenvectors for these eigenvalues. Get
ready to do your first actual quantum mechanical calculation.
Eigenvector for (+ ~ /2). The eigenvector has to satisfy
Sy u
~
+ u,
2
so we view this as an equation in the unknown coordinates (expressed in the preferred
z-basis where we have been working all along),
~ v1
~ 0 i
v1
=
.
v2
2 i 0
2 v2
This reduces to
i v2
i v1
=
=
v1
v2 .
and
There are two equations in two unknowns. Wrong. There are four unknowns (each
coordinate, vk , is a complex number, defined by two real numbers). This is somewhat
expected since we know that the solution will be a ray of vectors all differing by a
complex scalar factor. We can solve for any one of the vectors on this ray as a first
step. We do this by guessing that this ray has some non-zero first coordinate (and if
we guess wrong, we would try again, the second time knowing that it must therefore
have a non-zero second coordinate [Exercise. Why?]. Using this guess, we can pick
v1 =1, since any non-zero first coordinate can be made to = 1 by a scalar multiple of
the entire vector. With this we get the complex equations
i v2
i1
=
=
1
v2 ,
revealing that v2 = i, so
u
1
,
i
164
which we must (always) normalize by projecting onto the unit (projective) sphere.
u
|+i + i |i
1 1
=
=
.
|u|
2 i
2
The last equality is the expression of u explicitly in terms of the z-basis {|+i , |i}.
Alternate Method. We got lucky in that once we substituted 1 for v1 , we were
able to read off v2 immediately. Sometimes, the equation is messier, and we need to
do a little work. In that case, naming the real and imaginary parts of v2 helps.
v2
a + bi,
and substituting this into the original equations containing v2 , above, gives
i (a + b i)
i
=
=
1
(a + b i) ,
or
b ai =
i =
1
a + bi.
Lets solve the second equation for a, then substitute into the first, as in
a
b (i b i) i
1
=
=
=
i bi
1
1.
What does this mean? It means that we get a very agreeable second equation; the b
disappears resulting in a true identity (a tautology to the logician). We can, therefore
let b be anything. Again, when given a choice, choose 1. So b = 1 and substituting
that into any of the earlier equations gives a = 0. Thus,
v2
a + bi
0 + 1i
i,
the same result we got instantly the first time. We would then go on to normalize u
as before.
Wrong Guess for v1 ? If, however, after substituting for a and solving the first
equation, b disappeared and produced a falsehood (like 0 = 1), then no b would be
suitable. That would mean our original choice of v1 = 1 was not defensible; v1 could
not have been a non-zero value. We would simply change that assumption, set v1 = 0
and go on to solve for v2 (either directly or by solving for a and b to get it). This
time, we would be certain to get a solution. In fact, any time you end up facing a
contradiction (3 = 4) instead of a tautology (7 = 7), then your original guess for v1
has to be changed. Just redefine v1 (if you chose 1, change it to 0) and everything
will work out.
165
Too Many Solutions? (Optional Reading) In our well-behaved spin-1/2 state
space, each eigenvalue determines a single ray in the state-space, so it only takes one
unit eigenvector to describe it; you might say that the subspace of H spanned by the
eigenvectors corresponding to each eigenvalue is 1-dimensional. All of its eigenvectors
differ by scalar factor. But in other physical systems the eigenvector equation related
to a single eigenvalue may yield too many solutions (even after accounting for the
scalar multiples on the same ray we already know about). In other words, there
may be multiple linearly-independent solutions to the equation for one eigenvalue. If
so, we select an orthonormal set of eigenvectors that correspond to the degenerate
eigenvalue, as follows.
1. First observe (you can prove this as an easy [exercise]) that the set of all
eigenvectors belonging to that eigenvalue form a vector subspace of the state
space.
2. Use basic linear algebra to find any basis for this subspace.
3. Apply a process named Gram-Schmidt to replace that basis by an orthonormal basis.
Repeat this for any eigenvalue that is degenerate. You will get an optional exercise
that explains why we want an orthonormal basis for the eigenspace of a degenerate
eigenvalue.
Eigenvector for (- ~ /2). Now its your turn.
[Exercise. Find the eigenvector for the negative eigenvalue.]
7.6.2
The Eigenvalues and Eigenvectors of Sx
Roll up your sleeves and do it all for the third primary direction.
[Exercise. Compute the eigenvalues and eigenvectors for the observable Sx .]
Congratulations. You are now doing full-strength quantum mechanics (not
quantum physics, but, yes, quantum mechanics).
7.6.3
Summary of Eigenvectors and Eigenvalues for Spin-1/2
Observables
Spoiler Alert. These are the results of the two last exercises.
166
The eigenvalues and eigenvectors for Sz , Sx , and Sy are:
~
~
1
0
Sz :
+
0
1
2
2
1 1
1
~
~
1
Sx :
+
,
2
2
2 1
2 1
~
~
1 1
1
1
,
Sy :
+
2
2
2 i
2 i
Expressed explicitly in terms of the z-basis vectors we find
Sz :
Sx :
Sy :
~
~
|+i ,
|i
2
2
~
|+i + |i
|+ix =
,
2
2
~
|+i + i |i
|+iy =
,
2
2
~
|+i |i
|ix =
2
2
~
|+i i |i
|iy =
2
2
We saw the x-kets and y-kets before when we were trying to make sense out of the
50-50 split of a |ix state into the two states |+iz and |iz . Now, the expressions reemerge as the result of a rigorous calculation of the eigenvectors for the observables Sx
and Sy . Evidently, the eigenvectors of Sx are the same two vectors that you showed
(in an exercise) were an alternative orthonormal basis for H, and likewise for the
eigenvectors of Sy .
Using these expressions along with the distributive property of inner products, it
is easy to show orthonormality relations like
xh+
| + ix
x h + | ix
=
=
1,
0.
or
[Exercise. Prove the above two equalities as well as the remaining combinations
that demonstrate that both the x-bases and y-basis are each (individually) orthonormal.]
7.7
Observables and Orthonormal Bases
Thus far, we have met a simple 2-dimension Hilbert space, H, whose vectors correspond to states of a spin-1/2 physical system, S . The correspondence is not 1-to-1,
since an entire ray of vectors in H correspond to the same physical state, but we
are learning to live with that by remembering that we can and will normalize all
vectors whenever we want to see a proper representative of that ray on the unit (projective) sphere. In addition, we discovered that our three most common observables
Sz , Sx and Sy correspond to operators whose eigenvalues are all ~2 and whose
167
eigenvectors form three different 2-vector bases for the 2-dimensional H. Each of the
bases is an orthonormal basis.
Id like to distill two observations and award them collectively the title of trait.
7.7.1
Trait #4 (Real Eigenvalues and Orthonormal Eigenvectors)
An observable A in S will always correspond to an operator
1. whose eigenvalues, a1 , a2 , . . ., are real, and
2. whose eigenvectors ua1 , ua2 , . . ., form an orthonormal basis for the state space
H.
I did not call this trait a postulate, because it isnt; you can prove these two
properties based on Trait #2, which connects observables to Hermitian operators.
Although we wont spend our limited time proving it, if you are interested, try the
next few exercises.
Note. I wrote the trait as if all the eigenvalues were non-degenerate. It is still true,
even for degenerate eigenvalues, although then we would have to label the eigenvectors
more carefully.
[Exercise. Show that a Hermitian operators eigenvalues are real.]
[Exercise. Show that eigenvectors corresponding to distinct eigenvalues are orthogonal.]
[Exercise. In an optional passage, above, I mentioned that a degenerate eigenvalue determines not a single eigenvector, but a vector subspace of eigenvectors, from
which we can always select an orthonormal basis. Use this fact, combined with the
last exercise to construct a complete orthonormal set of eigenvectors, including those
that are non-degenerate (whose eigenspace is one-dimensional) and those that are
degenerate (whose eigenspace requires multiple vectors to span it).]
I told a small white lie a moment ago when I said that this is totally provable
from the second postulate (Trait #2). There is one detail that I left out of the
second postulate which is needed to prove these observations. You did not possess
the vocabulary to understand it at the time, but now you do. In Trait #2 I said
that the observable had to correspond to a Hilbert-space operator. What I left out
was the following mini-trait, which Ill call
Trait #2a (Completeness of the Eigenbasis): The eigenvectors of an observable span and since they are orthogonal, constitute a basis for the state space.
Furthermore, every measurable quantity of the physical system S corresponds to an
observable, thus its eigenvectors can be chosen as a basis whenever convenient.
If we had an observable whose eigenvectors turned out not to span the state
space, we did a bad job of defining the state space and would have to go back and
168
figure out a better mathematical Hilbert space to model S . Similarly, if there were
a measurable quantity for which we could not identify a linear operator, we have not
properly modeled S .
7.7.2
Using the x- or y-Basis
While weve been doggedly using the z-axis kets to act as a basis for our 2-dimensional
state space, we know we can use the x-kets or y-kets. Lets take the x-kets as an
example.
First and most easily, when we express any basis vector in its own coordinates,
the results always look like natural basis coordinates, i.e., (0, 1)t or (1, 0)t . This was
an exercise back in our linear algebra lecture, but you can take a moment to digest
it again. So, if we were to switch to using the x-basis for our state space vectors we
would surely see that
1
|+ix =
and
0 x
0
|ix =
.
1 x
How would our familiar z-kets now look in this new basis? You can do this by starting
with the expressions for the x-kets in terms of |+i and |i that we already have,
|+ix
|ix
|+i + |i
2
|+i |i
,
2
and
and solve for the z-kets in terms of the x-kets. It turns out that doing so results in
dej`a vu,
|+i
|i
|+ix + |ix
2
|+ix |ix
.
2
and
Its a bit of a coincidence, and this symmetry is not quite duplicated when we express
the y-kets in terms of |+ix and |ix . Ill do one, and you can do the other as an
exercise.
|+iy in the x-Basis. The approach is to first write down |+iy in terms of |+i and
|i (already known), then replace those two z-basis kets with their x-representation,
169
shown above. Here we go.
|+iy
|+i + i |i
|+ix + |ix
(1 + i) |+ix + (1 i) |ix
2
(1i)
2 |+ix 2i |ix
=
=
2
+ i
|+ix |ix
normalize
|+ix i |ix
|+ix i |ix
[Exercise. Show that
|iy
|+ix + i |ix i
[Exercise. Express |i and |ix in the y-basis.]
7.7.3
General States Expressed in Alternate Bases
Expressing |+i or |i in a different basis, like the x-basis, is just a special case of
something you already know about. Any vector, whose coordinates are given in the
preferred basis, like
|i = |+i + |i =
,
has an alternate representation when its coefficients are expanded along a different
basis, say the x-basis,
0
0
0
.
|i = |+ix + |ix =
0 x
If we have everything expressed in the preferred basis, then one day we find ourselves
in need of all our vectors names in the alternate (say x) basis, how do we do it?
Well there is one way I didnt develop (yet) and it involves finding a simple matrix
by which you could multiply all the preferred vectors to produce their coefficients
expanded along the alternate basis. But we usually dont need this matrix. Since
we are using orthonormal bases, we can convert to the alternate coordinates directly
using the dot-with-the-basis vector trick. Typically, we only have one or two vectors
whose coordinates we need in this alternate basis, so we just apply that trick.
For example, to get a state vector, |i, expressed in the x-basis, we just form the
two inner-products,
!
x h+ | i
|i =
,
x h | i
170
where the syntax
x h+ | i
means we are taking the complex inner product of |i on the right with the x-basis
ket |+ix on the left. (Dont forget that we have to take the Hermitian conjugate of
the left vector for complex inner products.)
We are implying by context that the column vector on the RHS is expressed in
x-coordinates since thats the whole point of the paragraph. But if we want to be
super explicit, we could write it as
!
!
x h+ | i
x h+ | i
|i =
,
, or
x h | i x
x h | i x
with or without the long vertical line, depending on author and time-of-day.
Showing the same thing in terms of the x-kets explicitly, we get
|i
x h+ | i
*
=
1
2
1
2
|+ix + x h | i |ix
+
* 1 +
2
|+ix +
|ix
1
2
Notice that the coordinates of the three vectors, |ix and |i are expressed in the
preferred z-basis. We can compute inner products in any orthonormal bases, and since
we happen to know everything in the z-basis, why not? Try not to get confused. We
are looking for the coordinates in the x-basis, so we need to dot with the x-basis
vectors, but we use z-coordinates of those vectors (and the z-coordinates of state |i)
to compute those two scalars.
Example. The (implied z-spin) state vector
!
1+i
23
has the x-spin coordinates,
*
1+i
1
6
2
1 2
2
3
1+i
1
6
2
2
1
2
3
[Exercise. Compute and simplify.]
171
7.8
7.8.1
The Completeness (or Closure) Relation
Orthonormal Bases in Higher Dimensions
Our casual lesson about conversion from the z-basis to the x-basis has brought us to
one of the most computationally useful tools in quantum mechanics. Its something
that we use when doodling with pen and paper to work out problems and construct
algorithms, so we dont want to miss the opportunity to establish it formally, right
now. We just saw that
|i
xh+
| ix |+ix
xh
| ix |ix ,
which was true because {|+ix , |ix } formed an orthonormal basis for our state space.
In systems with higher dimensions we have a more general formula. Where do we
get state spaces that have dimensions higher than two? A spin 1 system photons
has 3 dimensions; a spin-3/2 system delta particle has 4-dimensions and later
when we get into multi-qubit systems, well be taking tensor products of our humble
2-dimensional H to form 8-dimensional or larger state spaces. And the state space
that models position and momentum are infinite dimensional (but dont let that scare
you they are actually just as easy to work with we use use integrals instead of
sums).
If we have an n-dimensional state space then we would have an orthonormal basis
for that space, say,
n
on
|uk i
.
k=1
The |uk i basis may or may not be a preferred basis doesnt matter. Using the
dot-product trick we can always expand any state in that space, say |i, along the
basis just like we did for the x-basis in 2-dimensions. Only, now, we have a larger
sum
|i
n
X
huk | i |uk i .
k=1
This is a weighted sum of the uk -kets by the scalars huk | i. There is no law that
prevents us from placing the scalars on the right side of the vectors, as in
|i
n
X
|uk i huk | i .
k=1
Now, we can take the |i out of the sum
|i
n
X
!
|uk i huk | |i .
k=1
172
Look at what we have. We are subjecting any state vector |i to something that
looks like an operator and getting that same state vector back again. In other words,
that fancy looking operator-sum is nothing but an identity operator 1.
!
n
X
|uk i huk |
= 1.
k=1
This simple relation, called the completeness or closure relation, can be applied by
inserting the sum into any state equation without changing it, since it is the same as
inserting an identity operator (an identity matrix) into an equation involving vectors.
Well use it a little in this course, CS 83A, and a lot in the next courses CS 83B and
CS 83C, here at Foothill College.
P
[Exercise. Explain how the sum, k |uk i huk | is, in fact, a linear transformation
that can act on a vector |i. Hint: After applying it to |i and distributing, each
term is just an inner-product (resulting in a scalar) times a vector. Thus, you can
analyze a simple inner product first and later take the sum, invoking the properties
of linearity.]
This is worthy of its own trait.
7.8.2
Trait #5 (Closure Relation)
Any orthonormal basis { |uk i } for our Hilbert space H satisfies the closure relation
!
n
X
|uk i huk |
= 1.
k=1
In particular, the eigenvectors of an observable will always satisfy the closure relation.
We wont work any examples as this will take us too far afield, and in our simple
2-dimensional H theres not much to see, but we now have it on-the-record.
7.9
The Fourth Postulate of Quantum Mechanics
There are two versions of this one, and we only need the simpler of the two which
holds in the case of non-degenerate eigenvalues, the situation that we will be using for
our spin-1/2 state spaces. (Also, youll note our traits are no longer in-sync with the
postulates, an inevitable circumstance since there are more traits than postulates.)
7.9.1
Trait #6 (Probability of Outcomes)
If a system is in the (normalized) state |i, and this state is expanded along the
eigenbasis { |uk i } of some observable, A, i.e.,
n
X
|i =
ck |uk i ,
k=1
173
then the probability that a measurement of A will yield a non-degenerate eigenvalue
ak (associated with the eigenvector |uk i) is |ck |2 .
Vocabulary. The expansion coefficients, ck , for state |i, are often referred to as
amplitudes by physicists.
In this language, the probability of obtaining a measurement outcome ak for observable A is the magnitude-squared of the amplitude ck standing next to the eigenvector |uk i associated with the outcome ak .
The complex coordinates of the state determine the statistical outcome of repeated
experimentation. This is about as quantum mechanical a concept as there is. It tells
us the following.
If a state is a superposition (non-trivial linear combination) of two or more
eigenkets, we cannot know an outcome of a quantum measurement with certainty.
This is not a lack of knowledge about the system, but a statement about what
it means to know everything knowable about the system, namely the full description of the state.
The coefficient (or amplitude) ck gives you the probability of the outcome ak ,
namely
P ak |i = ck ck = |ck |2 ,
to be read, when a system in state |i is measured, its probability of outcome,
ak , is the square magnitude of ck . The observable in question is understood to
be the one whose eigenvalues-eigenvectors are { ak |uk i }.
Probability Example 1
Lets analyze our original Experiment #2 (last optional chapter), in which we
started with a group of electrons in the +z state, i.e., |+i, and tested each of their
spin projections along the x-axis, i.e., we measured the observable Sx . In order to
predict the probabilities of the outcome of the Sx measurement using the fourth
QM postulate our Trait #6 we expand the pre-measurement state along the
eigenvector basis for Sx and then examine each coefficient. The pre-measurements
state is |+i. We have previously expressed that state in terms of the Sx eigenvectors,
and the result was
|+ix + |ix
.
|+i =
2
The amplitude of each of the two possible outcome states (the two eigenstates |+ix
and |ix ) is 12 . This tell us that each eigenstate outcome is equally likely and, in
fact, determined by
1
1
1
1
P Sx = + ~
=
=
,
2
2 |+i
2
2
174
while
1
P Sx = ~
2 |+i
=
1
.
2
Notice that the two probabilities add to 1. Is this a happy coincidence? I think not.
The first postulate of QM (our Trait #1) guarantees that we are using unit vectors
to correspond to system states. If we had a non-unit vector that was supposed to
represent that state, wed need to normalize it first before attempting to compute the
probabilities.
Probability Example 2
The follow-up to Experiment #2 was the most shocking to us, so we should see
how it is predicted by Trait #6. The starting point for this measurement was the
output of the second apparatus, specifically, the x group: |ix . We then measured
Sz , so we need to know the coefficients of the state |ix along the Sz eigenbasis. We
computed this already, and they are contained in
|ix
|+i |i
.
2
The arithmetic we just did works exactly the same here, and our predictions are the
same:
1
1
1
1
P Sz = + ~
=
=
2
2 |ix
2
2
and
1
P Sz = ~
2 |ix
1
1
2
2
1
.
2
Notice that, despite the amplitudes negative signs, the probabilities still come out
non-negative.
[Exercise. Analyze the Sy measurement probabilities of an electron in the state
|i. Be careful. This time we have a complex number to conjugate.]
Probability Example 3
The z-spin coordinates of a state, |i are given by
!
1+i
23
175
The probability of detecting a z-UP spin is given by the coefficient (amplitude) of
the |+i, and is
1+i
1
1i
P Sz = + ~
=
2 |i
6
6
2
1
=
=
.
6
3
The probability of detecting an x-DOWN spin starting with that same |i requires
that we project that state along the x-basis,
|i
c+ |+ix + c |ix .
However, since we only care about the x-down state, we can just compute the |ix
coefficient, which we do using the dot-product trick.
* 1 1+i +
1+i
1
6
2
c =
=
+
2
1
2
12
3
3
=
1+i+2
12
3+i
.
12
Now we take the magnitude squared,
3i
3+i
2
|c |
=
12
12
10
12
5
.
6
[Exercise. Compute the z and +x spin probabilities for this |i and confirm
that they complement their respective partners that we computed above. Explain
what I mean by complementing their partners.]
[Exercise. Compute the +y and y spin probabilities for this |i.]
7.10
The Fifth Postulate of Quantum Mechanics
After the first experiment of the last lesson, in which we measured the z-spin of
random electrons and noted that the measurement caused them to split into two
equal groups, one having all z-up states and the other having all z-down states,
we did a follow-up. We tested Sz again on each output group separately. In case
you missed it, heres what happened. When we re-tested the |+i group, the results
always gave us +z readings, and when we re-tested the |i group we always got z
readings. It was as if, after the first measurement had divided our electrons into two
equal groups of +z and z spins, any subsequent tests suggested that the two groups
were frozen into their two respective states, as long as we only tested Sz .
However, the moment we tested a different observable, like Sx , we disturbed that
z-axis predictability. So the collapse of the system into the Sz state was only valid as
long as we continued to test that one observable. This is our next trait.
176
7.10.1
Trait #7 (Post-Measurement Collapse)
If the measurement of an observable of system S results in the eigenvalue, ak , then
the system collapses into the (an) eigenvector |uk i, associated with ak . Further
measurements on this collapsed state yields the eigenvalue ak with 100% certainty.
Vocabulary Review. The eigenvectors of an observable are also known as eigenkets.
Weve been saying all along that the eigenvalue ak might be degenerate. The implication here is that there may be more than one possibility for the collapsed state,
specifically, any of the eigenvectors |u0k i , |u00k i , |u000
k i , . . . which correspond to ak . We
wont encounter this situation immediately, but it will arise later in the course. The
early easy cases will consist of non-degenerate eigenvalues whose probabilities are easily computed by the amplitudes of their respective unique eigenvectors. Later, when
we get to degenerate eigenvalues, we wont be sure which of the eigenvectors corresponding to that eigenvalue represents the state into which the system collapsed.
Yet, even knowing that it collapsed to the small subset of eigenvectors corresponding
to a single eigenvalue (in the degenerate case) is invaluable information that plays
directly into our algorithms.
Preparing Special System States
The impact of Trait #7 is that engineers can prepare special states to act as input
into our quantum hardware logic. This is akin to setting a registers value in a classical
computer using an assignment statement, prior to beginning further logic.
Example: Preparing a Basis (Eigenvector) State. We did this in Experiment #1. After measuring the observable Sz , we ended up with two groups. By
selecting either one of those groups, we will be getting either the |+i state or the |i
state, as we wish. Any future testing of Sz will confirm that we stay in those states,
as long as we dont subject the system to forces that modify it.
Example: Preparing a Superposition State. We did this, too. In our late
stages of tinkering with Experiment #2, we focused on the output of the second
apparatus by choosing the |ix group for further investigation. Using our state space
vocabulary, we expand the state |ix in terms of the z-eigenkets,
|+i + |i
,
2
and realize that we have prepared a state which, with respect to the z-spin observable,
is not a basis state but a linear combination a superposition of the two z-basis
states (with equally weighted components). This kind of state preparation will be
very important for quantum algorithms, because it represents starting out in a state
which is neither 0 nor 1, but a combination of the the two. This allows us to work
with a single state in our quantum processor and get two results for the price of one.
A single qubit and a single machine cycle will simultaneously produce answers for
both 0 and 1.
|+ix
177
But, after we have prepared one of these states, how do we go about giving it to
a quantum processor, and what is a quantum processor. Thats answered in the first
quantum computing lesson coming up any day now. Today, we carry on with pure
quantum mechanics to acquire the full quiver of q-darts.
7.11
Summary of What Quantum Mechanics Tells
Us About a System
We now have a complete set of numbers and mathematical entities that give us all
there is to know about any quantum system. We know
1. the possible states |i of the system (vectors in our state space),
2. the possible outcomes of a measurement of an observable of that system (the
eigenvalues, ak , of the observable),
3. the eigenvectors states (a.k.a. eigenbasis), |uk i, into which the system collapses
after we detect a value, ak , of that observable, and
4. the probabilities associated with each eigenvalue outcome (specifically, |ck |2 ,
which are derived from the amplitudes, ck , of |i expanded along the eigenbasis
|uk i).
In words, the system is in a state of probabilities. We can only get certain special
outcomes of measurement. Once we measure, the system collapses into a special state
associated with that special outcome. The probability that this measurement occurs
is predicted by the coefficients of the states eigenvector expansion.
7.12
Diracs Bra-ket Notation
We take a short break from physics to talk about the universal notation made popular
by the famous physicist Paul Dirac.
7.12.1
Kets and Bras
Its time to formalize something weve been using loosely up to this point: the bracket,
or bra-ket, notation. The physicists expression for a complex inner product,
hv | wi ,
h | i ,
huk | i ,
h+ | i ,
can be viewed, not as a combination of two vectors in the same vector space, but
rather as two vectors, each from different vector space. Take, for example,
h | i,
178
The RHS of the inner product, |i, is the familiar vector in our state space, or ket
space. Nothing new there. But the LHS, h|, is to be thought of as a vector from a
new vector space, called the bra-space (mathematicians call it the dual space). The
bra space is constructed by taking conjugate transpose of the vectors in the ket space,
that is,
|i =
h| = ( , ) .
Meanwhile, the scalars for the bra space are the same: the complex numbers, C.
Examples
Here are some kets (not necessarily
!
1 + i
|i =
2 2i
!
3/2
|i =
i
!
5i
1
|i =
2 0
!
1 1
|i =
2 1
normalized) and their associated bras.
1 i,
2 + 2i
h| =
h| =
1
h| = ( 5i , 0 )
2
1
h| = ( 1 , 1 )
2
|+i
h+|
|iy
y h|
3/2 , i
[Exercise. Show that
|+i + i |i
h+| i h| .
Hint: Its probably easier to do this without reading a hint, but if youre stuck
... write out the LHS as a single column vector and take the conjugate transpose.
Meanwhile the RHS can be constructed by constructing the bras for the two z-basis
vectors (again using coordinates) and combining them. The two efforts should result
in the same vector. ]
Vocabulary and Notation
Notice that bras are written as row-vectors, which is why we call them conjugate
transposes of kets. The dagger () is used to express the fact that a ket and bra bear
179
this conjugate transpose relationship,
h|
|i
|i
h| .
and
This should sound very familiar. Where have we seen conjugate transpose before?
Answer: When we defined the adjoint of a matrix. We even used the same dagger
() notation. In fact, you saw an example in which the matrix had only one row (or
one column) i.e., a vector. (See the lesson on linear transformations.) This is the
same operation: conjugate transpose.
Be careful not say that a ket is the complex conjugate of a bra. Complex conjugation is used for scalars, only. Again, we say that a bra is the adjoint of the ket (and
vice versa). If you want to be literal, you can always say conjugate transpose. An
alternate term physicists like to use is Hermitian conjugate, the bra is the Hermitian
conjugate of the ket.
Example
Lets demonstrate that the sum of two bras is also a bra. What does that even mean?
If h| and h| are bras, they must be the adjoints of two kets,
|i =
h| = ( , ) .
|i =
h| = ( , ) .
We add the bras component-wise,
h| + h|
( + ,
+ ) ,
which you can see is the Hermitian conjugate of the ket
+
|i + |i =
,
+
i.e.,
h| + h|
( |i + |i ) .
Thats all a bra needs to be: the Hermitian conjugate of some ket. So the sum is a
bra.
There is (at least) one thing we must confirm. As always, when we define anything
in terms of coordinates, we need to be sure that the definition is independent of our
choice of basis (since coordinates arise from some basis). I wont prove this, but you
may choose to do so as an exercise.
180
[Exercise. Pick any three axioms of a vector space and prove that the bras in
the bra space obey them.]
[Exercise. Show that the definition of bra space is independent of basis.]
Remain Calm. There is no cause for alarm. Bra space is simply a device that
allows us manipulate the equations without making mistakes. It gives us the ability to
talk about the LHS and the RHS of an inner product individually and symmetrically,
unattached to the inner product.
Elaborate Example
We will use the bra notation to compute the inner product of the two somewhat
complicated kets,
c |i + |i
on the left, and
d |i f |i
g
on the right ,
where c, d, f and g are some complex scalars. The idea is very simple. We first take
the Hermitian conjugate of the intended left vector by
1. turning all of the kets (in that left vector) into bras,
2. taking the complex conjugate of any scalars (in that left vector) which are
outside a ket,
3. forming the desired inner product, and
4. using the distributive property to combine the component kets and bras.
Applying steps 1 and 2 on the left ket produces the bra
c h| + h| .
Step 3 gives us
d |i f |i
c h| + h|
,
g
and applying step 4 we get
c d h | i + d h | i c f h | i + f h | i
.
g
It seems overly complicated, but usually we apply it to simpler combinations and it
is far less cumbersome than turning all the constituent kets into their coordinates,
combining them, taking the conjugates of the left ket and doing the final dot.
181
Simple Example
y h+ | +i
=
=
=
|+i + i |i
h
| + i
2
h+| i h|
| + i
2
1
h+ | +i i h | +i
= .
2
2
The first thing we did was to express |+iy in the z-basis without converting it to a
bra. Then, we used the techniques just presented to convert that larger expression
into a bra. From there, it was a matter of distributing the individual kets and bras
and letting them neutralize each other. Normally, we would perform the first two
steps at once, as the next example demonstrates.
|+i + |i
h+| + i h|
h | +ix =
y
2
2
h+ | +i + i h | +i + h+ | i + i h | i
=
2
1 + i
=
.
2
[Exercise. Compute
h+ | +ix
and
h | +iy . ]
Summary
The bra space is a different vector space from the ket (our state) space. It is, however,
an exact copy (an isomorphism) of the state space in the case of finite dimensions
- all we ever use in quantum computing. You now have enough chops to prove this
easily, so I leave it as an ...
[Exercise. Prove that the adjoint of the ket basis is a basis for bra space. Hint:
Start with any bra. Find the ket from which it came (this step is not always possible in
infinite dimensional Hilbert space, as your physics instructors will tell you). Expand
that ket in any state space basis, then . . . .]
7.12.2
The Adjoint of an Operator
We now leverage the earlier definition of a matrix adjoint and extend our ability to
translate expressions from the ket space to the bra space (and back). If A is an
operator (or matrix) in our state space (not necessarily Hermitian it could be any
operator), then its adjoint A can be viewed as an operator (or matrix) in the bra
space.
A : h|
7
182
h| .
A operates on bras, but since bras are row-vectors, it has to operate on the right,
not left:
h| A
h| .
And the symmetry we would like to see is that the output h| to which A maps
h| is the bra corresponding to the ket A |i. That dizzying sentence translated into
symbols is
h|
A h| .
Example. Start with our familiar 2-dimensional state space and consider the
operator,
i i
A =
.
1 0
Its adjoint is
A
i 1
.
i 0
A maps the bra, h| = ( 1 + i , 3 ) into
i 1
h| A = ( 1 + i , 3 )
i 0
( 1 + 2i , 1 + i ) .
Meanwhile, back in ket space, A maps |i = h| into
i i
1i
1 2i
A |i =
=
.
1 0
3
1i
As you can see by comparing the RHS of both calculations, the adjoint of A |i is
h| A , in agreement with the claim.
7.12.3
The Adjoint Conversion Rules
The reason we added adjoint operators into the mix was to supply the final key to
the processes of converting any combination of state space kets into bras and any
combination of bras into kets. Say we desire to convert the ket expression
c A |i + d h+ | i |i
into its bra counterpart. The rules will guide us, and they work for expressions far
more complex with equal ease.
We state the rules, then we will try them out on this expression. This calls for a
new trait.
183
7.12.4
Trait #8 (Adjoint Conversion Rules)
The terms of a sum can be (but dont have to be) left in the same order.
The order of factors in a product are reversed.
Kets are converted to bras (i.e., take the adjoints).
Bras are converted to kets (i.e., take the adjoints).
Operators are converted to their adjoints.
Scalars are converted to their complex conjugates.
(Covered by the above, but stated separately anyway:) Inner products are reversed.
When done (for readability only), rearrange each product so that the scalars are
on the left of the vectors.
If we apply the adjoint conversion rules, except for the readability step, to the
above combination we get
c A |i + d h+ | i |i
h| A c + h| h | +i d ,
which we rearrange to
c h| A + d h | +i h| .
Youll get fast at this with practice. I dont want to spend any more real estate on
the topic, since we dont apply it very much in our first course, CS 83A, but here are
a couple exercises that will take care of any lingering urges.
[Exercise. Use the rules to convert the resulting bra of the last example back
into a ket. Confirm that you get the ket we started with.]
[Exercise. Create a wild ket expression consisting of actual literal scalars, matrices and column vectors. Use the rules to convert it to a bra. Then use the same rules
to convert the bra back to a ket. Confirm that you get the ket you started with.]
7.13
Expectation Values
Say we have modeled some physical quantum system, S , with a Hilbert space, H.
Imagine, further, that we want to study some observable, A, that has (all nondegenerate) eigenvalues {ak } with corresponding eigenkets { |uk i }. Most importantly,
we assume that we can prepare many identical copies of S , all in the same state |i.
184
(We did this very thing by selecting only z-up electrons in a Stern-Gerlach-like apparatus, for example.) We now look at our state expanded along the A eigenbasis,
X
|i =
ck |uk i .
k
How do the amplitudes, ck , and their corresponding probabilities, |ck |2 , make themselves felt, by us human experimenters?
The answer to this question starts by taking many repeated measurements of the
observable A on these many identical states |i and recording our results.
[Exercise. Explain why we cant get the same results by repeating the A measurements on a single system S in state |i.]
7.13.1
The Mean of the Experiments
Well take a large number, N , of measurements, record them, and start doing elementary statistics on the results. We label the measurements we get using
jth measurement of A
mj .
As a start, we compute average (or mean) of all N measurements,
m
N
1 X
mj .
N j=1
If we take a large enough N , what do we expect this average to be? This answer
comes from the statistical axiom called the law of large numbers, which says that this
value will approach the expectation value, , as N , that is,
lim m = .
This is good and wonderful, but I have not yet defined the expectation value . Better
do that, fast.
[Note. I should really have labeled m with N , as in mN , to indicate that each
average depends on the number of measurements taken, as we are imagining that we
can do the experiment with larger and larger N . But you understand this without
the extra notation.]
7.13.2
Defining Expectation Value
This conveniently brings us back around to Trait #6 (the Fourth Postulate of QM )
concerning probabilities of outcomes. It asserts that the probability of detecting the
185
measurement (eigenvalue) ak is given by its kth expansion coefficient along the A
basis (ck ), specifically by its magnitude-squared, |ck |2 ,
P ak |i = ck ck = |ck |2 .
This motivates the definition. Expectation value, , is defined by summing up each
possible outcome (the eigenvalues ak ), weighted by their respective probabilities, |ck |2 :
X
|ck |2 ak .
k
In case you dont see why this has the feeling of an expectation value (something we
might expect from a typical measurement, if we were forced to place a bet), read it
in English:
The first measurable value times the probability of getting that value
plus
the second measurable value times the probability of getting that value
plus
... .
In physics, rather than using the Greek letter , the notation for expectation value
focuses attention on the observable we are measuring, A,
X
hAi|i
|ck |2 ak .
k
Fair Warning. In quantum mechanics, you will often see the expectation value of
the observable, A, written without the subscript, |i,
hAi ,
but this doesnt technically make sense. There is no such thing as an expectation
value for an observable that applies without some assumed state; you must know
which |i has been prepared prior to doing the experiment. If this is not obvious to
you, look up at the definition one more time: We dont have any ck to use in the
formula unless there is a |i in the room, because
X
|i =
ck |uk i .
k
When authors suppress the subscript state on the expectation value, its usually
because the context strongly implies the state or the state is explicitly described
earlier and applies until further notice.
Calculating an expectation value tells us one way we can use the amplitudes.
This, in turn acts as an approximation of the average m, of a set of experimental
measurements on multiple systems in the identical state.
186
7.13.3
Computing Expectation Value
It seems like we should be done with this section. We have a formula for the expectation value, hAi|i , so what else is there to do? It turns out that computing that sum
isnt always as easy or efficient as computing the value a different way.
7.13.4
Trait #9 (Computing an Expectation Value)
The expectation value of an observable A of a system in state |i is usually computed
using the expression h | A | i.
To add formality, we will also call this the
Expectation Value Theorem.
hAi|i
h | A | i .
The A on the RHS can be thought of either as the operator representing A, or the
matrix for the operator. The expression can be organized in various ways, all of which
result in the same real number. For instance, we can first compute A |i and then
take the inner product
h| A |i ,
or we can first apply A to the bra h|, and then dot it with the ket,
h| A |i .
If you do it this way, be careful not take the adjoint of A. Just because we apply
an operator to a bra does not mean we have to take its Hermitian conjugate. The
formula says to use A, not A , regardless of which vector we feed it.
[Exercise. Prove that the two interpretations of h | A | i are equal by expressing
everything in component form with respect to any basis. Hint: Its just (a row vector)
(a matrix) (a column vector), so multiply it out both ways.]
Proof of the Expectation Value Theorem. This is actually one way to prove
the last exercise nicely. Express everything in the A-basis (i.e., the eigenkets of A
which Trait #4 tells us form a basis for the state space H). We already know that
c1
c2
X
|i =
ck |uk i = c3
,
..
k
.
cn A-basis
which means (by our adjoint conversion rules)
X
h| =
ck huk | = (c1 , c2 , c3 , . . . , cn )A-basis .
k
187
Finally, whats A in its own eigenbasis? We know that any basis vector expressed in
that basis coordinates has a preferred basis look, (1, 0, 0, 0, . . . )t , (0, 1, 0, 0, . . . )t ,
etc. To that, add the definition of an eigenvector and eigenvalue
Mu
au,
and you will conclude that, in its own eigenbasis, the matrix for A is 0 everywhere
except along its diagonal, which holds the eigenvalues,
a1 0 0 0
0 a2 0 0
A = 0 0 a3 0
.
.. .. .. . .
..
. . .
. .
0 0 0 an A-basis
[Oops. I just gave away the answer to one of todays exercises.
have all our players in coordinate form, so
a1 0 0
0 a2 0
0
0 a3
h | A | i = (c1 , c2 , c3 , . . . , cn )
.. .. .. . .
. . .
.
0 0 0
a1 c 1
a2 c 2
= (c1 , c2 , c3 , . . . , cn ) a3 c3
..
.
an c n
X
=
|ck |2 ak ,
Which one?] We now
0
c1
c2
0
0
c3
.. ..
. .
an
cn
which is the definition of hAi|i .
7.13.5
QED
Expectation Values in Spin-1/2 Systems
Ill do a few examples to demonstrate how this looks in our 2-dimensional world.
The Expectation Value of Sz Given the State |+i
This is a great sanity check, since we know from Trait #7 (the fifth postulate of
QM) that Sz will always report a + ~2 with certainty if we start in that state. Lets
confirm it.
~ 1 0
~
1
h+ | Sz | +i = (1, 0)
= + .
0
2 0 1
2
188
That was painless. Notice that this result is weaker than what we already knew
from Trait #7. This is telling us that the average reading will approach the (+)
eigenvalue in the long run, but in fact every measurement will be (+).
[Exercise. Show that the expectation value h | Sz | i = ~2 .]
The Expectation Value of Sx Given the State |ix
Once again, we know the answer should be ~2 , because were starting in an eigenstate
of Sx , but we will do the computation in the z-basis, which involves a wee-bit more
arithmetic and will serve to give us some extra practice.
1
~ 0 1
1
1
h | Sx | ix = (1, 1)
x
1
1
0
2
2
2
~
0 1
1
=
(1, 1)
1 0
1
4
~
~
1
(1, 1)
= .
=
1
4
2
[Exercise. Show that the expectation value h+ | Sx | +ix = + ~2 .]
x
The Expectation Value of Sz Given the State |iy
This time, its not so obvious. However, we can guess. Since, the state
|+i i |i
2
we see that the probability for each outome is 1/2. Over time half will result in an
Sz measurement of + ~2 , and half will give us ~2 , so the average should be close to 0.
Lets verify that.
1
~ 1 0
1
1
h | Sz | iy = (1, +i)
0
1
i
y
2
2
2
~
1 0
1
(1, +i)
=
0
1
i
4
~
1
=
(1, +i)
= 0.
i
4
|iy
[Exercise. Show that the expectation value h+ | Sz | +iy = 0 .]
y
7.14
Its About Time
We are done with all the math and physics necessary to do rigorous quantum computing at the basic level. Well be adding a few lessons on math as the course progresses,
189
but for now youre ready to dive into the lectures on single qubit systems and early
algorithms.
The next chapter is a completion of our quantum mechanics primer that covers
the basics of time evolution, that is, it describes the laws by which quantum systems
evolve over time. It isnt required for CS 83A, but youll need it for the later courses
in the sequence. You can skip it if you are so inclined or, if you opt in immediately,
itll provide the final postulates and traits that comprise a complete study of quantum
formalism including the all important Schrodinger equation.
Whether you choose to go directly to the chapter on qubits or first learn the
essentials of the time dependent Schrodinger equation youre in for a treat. Both
subjects provide a sense of purpose and completion to all the hard work weve done
up to this point.
Either way, its about time.
190
Chapter 8
Time Dependent Quantum
Mechanics
8.1
Time Evolution in Quantum Computing
Once a quantum system is put into a known state it will inevitably evolve over
time. This might be a result of the natural laws of physics taking their toll on the
undisturbed system or it may be that we are intentionally subjecting the system to
a modifying force such as a quantum logic gate. In either case, the transformation is
modeled by a linear operator that is unitary.
Weve seen unitary matrices already, and even if you skip this chapter, youre
equipped to go on to study qubits and quantum logic because the matrices and
vectors involved do not have a time variable, t, in their respective coefficients - they
are all complex constants.
But there are unitary transformation and quantum state vectors in which the
elements themselves are functions of time, and it is that kind of evolution we will
study today. It could represent the noise inherent in a system or it may just be a
predictable change that results from the kind of hardware we are using.
8.2
8.2.1
The Hamiltonian
Total Energy
It turns out that the time evolution of a quantum system is completely determined by
the total energy of the system (potential plus kinetic). There is a name for this quantity: The Hamiltonian. However the term is used differently in quantum mechanics
than in classical mechanics, as well see in a moment.
191
8.2.2
From Classical Hamiltonian to Quantum Hamiltonian
We have to figure out a way to express the energy of a system S using pen and paper
so we can manipulate symbols, work problems and make predictions about how the
system will look at 5 PM if we know how it started out at 8 AM. It sounds like a
daunting task, but the 20th century physicists gave us a conceptually simple recipe
for the process. The first step is to define a quantum operator a matrix for our
state space that corresponds to the total energy. Well call this recipe ...
Trait #10 (Constructing the Hamiltonian)
To construct an operator, H, that represents the energy of a quantum system:
1. Express the classical energy, H , formulaically in terms of basic classical concepts (e.g., position, momentum, angular momentum, etc.). You will have H
on the LHS of a formula and all the more basic terms on the RHS.
2. Replace the occurrence of the classical variables on the RHS by their (well
known) quantum operators, and replace the symbol for classical energy, H ,
on the LHS by its quantum symbol H.
Vocabulary. Although the total energy, whether classical or quantum, is a scalar,
the Hamiltonian in the quantum case is typically the operator associated with the
(measurement of) that scalar, while the classical Hamiltonian continues to be synonymous with the scalar, itself. As you can see from the reading Trait #10, to
distinguish the classical Hamiltonian, which works for macroscopic quantities, from
the quantum Hamiltonian, we use H for classical and H for quantum).
This is a bit hard to visualize until we do it. Also, it assumes we have already been
told what operators are associated with the basic physical concepts on the RHS. As
it happens, there are very few such basic concepts, and their quantum operators are
well known. For example, 3-dimensional positional coordinates (x, y, z) correspond to
three simple quantum operators X, Y and Z. Because I have not burdened you with
the Hilbert space that models position and momentum, I cant give you a meaningful
and short example using those somewhat familiar concepts, but in a moment youll
see every detail in our spin-1/2 Hilbert space, which is all we really care about.
8.3
8.3.1
The Hamiltonian for a Spin-1/2 System
A Classical Hamiltonian
Consider a stationary electron in a constant magnetic field. (Warning: this is not
the Stern-Gerlach experiment since that requires a non-homogeneous magnetic-field,
not to mention a moving electron.) Because the electron is stationary, all the energy
is potential and depends only on the spin (direction and magnitude) of the electron
192
in relation to the magnetic field. (You may challenge that I forgot to account for
the rotational kinetic energy, but an electron has no spatial extent, so there is no
classical moment of inertia, and therefore no rotational kinetic energy.) We want to
build a classical energy equation, so we treat spin as a classical 3-dimensional vector
representing the intrinsic angular momentum, (sx , sy , sz )t . A dot product between
this vector and the magnetic field vector, B, expresses this potential energy and yields
the following classical Hamiltonian
H
B S ,
where is a scalar known by the impressive name gyromagnetic ratio whose value is
not relevant at the moment. (We are temporarily viewing the system as if it were
classical in order to achieve step 1 in Trait #10, but please understand that it
already has one foot in the quantum world simply by the inclusion of the scalar .
Since scalars dont affect us, this apparent infraction doesnt disturb the process.)
Defining the z-Direction. The dot product only cares about the relationship
between two vectors,
vw
vw cos ,
where is the angle between them, so we can rotate the pair as a fixed assembly, that
is, preserving angle . Therefore, lets establish a magnetic field (with magnitude B)
pointing in the +z-direction,
0
0 ,
B = B z =
B
and let the spin vector, S, go along for the rotational ride. This does not produce
a unique direction for the spin, but we only care about the polar angle , which we
have constrained to remain unchanged. Equivalently, we can define the z-direction
to be wherever our B field points. Either way, we get a very neat simplification.
The classical spin has well-defined real-valued components sx , sy and sz (not
operators yet),
sx
sy ,
S =
sz
and we use this vector to evaluate the dot product, above. Substituting gives
0
sx
H = 0 sy
B
sz
=
B sz .
This completes step 1 in Trait #10, and I can finally show you how easy it is to do
step 2.
193
8.3.2
A Quantum Hamiltonian
We saw that (by a century of experimentation) the quantum operator corresponding
to the classical z-component of spin is Sz . So we simply replace the classical sz with
our now familiar quantum operator operator Sz to get the quantum Hamiltonian,
H=
B Sz .
Its that simple.
Note that while Sz may change from moment-to-moment (we arent sure yet), B
and are constants, so the Hamiltonian is not time-dependent; it always gives an
answer based solely on the relationship between S and B.
We now have a scalar relation between the Hamiltonian operator, H, and the
z-spin projection operator, Sz , giving us a short-cut in our quest for H; we need only
substitute the matrix we computed earlier for Sz into this equation and get
~ 1 0
H = B
.
2 0 1
8.4
8.4.1
The Energy Eigenkets
Relationship Between H and Sz
Because the operator H is a scalar multiple of our well studied Sz , we can easily find
its eigenvectors (also known as the energy eigenkets) and eigenvalues; the eigenvectors
are the same |i, and to get their respective eigenvalues we simply multiply those
of Sz by B. Wed better prove this, as it might not be obvious to everyone. We
know that eigenvalue-eigenvector pairs for Sz are given by
~
~
|+i ,
|+i
and
Sz |+i =
2
2
~
~
Sz |i =
|i ,
|i
,
2
2
but we now know that Sz and B bear the scalar relationship,
1
Sz =
H.
B
Substituting this into the Sz eigenvector expressions gives
1
~
H |+i =
|+i
and
B
2
1
~
H |i =
|i ,
B
2
194
or
H |+i
H |i
B~
|+i
2
B~
|i .
2
and
But this says that H has the same two eigenvectors as Sz , only they are associated
with different eigenvalues,
|+i
|i
B~
,
2
B~
+
.
2
if the electron collapses into
If we measure the energy of the system, we will get B~
2
the |+i state, pointing as close as quantum mechanics allows toward the (+z)-axis,
the direction of the B-field. Meanwhile, if it collapses into the |i state, we will
get + B~
, pointing as far as quantum mechanics allows from (+z)-axis, opposite the
2
direction of the B-field.
Does this make sense?
Yes. When a magnetic dipole (which is what electron spin represents, discounting
the gyromagnetic ratio ) is pointing in the direction of a magnetic field, energy is
minimum. Imagine a compass needle pointing magnetic north. The potential energy
of the compass-Earth system is at its minimum: It takes no energy to maintain that
configuration. However, if the dipole is pointing opposite the direction of the magnetic
field, energy is at its maximum: Imagine the compass needle pointing magnetic south.
Now it takes energy to hold the needle in place. The potential energy of the compassEarth system is maximum. The fact that the energy measurement is negative for the
|+i but positive for the |i faithfully represents physical reality.
Also, note that for a spin-1/2 system, any state vector, |i, has the same coordinate expression whether we expand it along the eigenkets of Sz or those of H, as
they are the same two kets, |+i and |i.
8.4.2
Allowable Energies
Knowledge of the energy eigenvectors is essential to predicting time evolution of any
state, as we are about to see. But first, we apply Trait #3 (the third postulate
of QM: allowable eigenvalues of an observable) to the Hamiltonian to make a traitworthy observation that is universal in quantum physics.
Trait #11 (Quantization of Energy)
The only allowable (i.e., measurable) energies of a quantum system are the eigenvalues
of the Hamiltonian.
195
Lets take a short side-trip to give the crucial postulate that expresses, in full generality, how any quantum state evolves based on the systems Hamiltonian operator,
H.
8.5
8.5.1
Sixth Postulate of Quantum Mechanics
The Schr
odinger Equation
The time-dependent Schrodinger Equation is the quantum-mechanical answer to the
question, How does a state evolve over time? To start, we replace a general (but
fixed) state with one which is a function of time,
|i
| (t) i .
As a consequence, its expansion coefficients will also be complex scalar functions of
time.
| (t) i
n
X
ck (t) |uk i .
k=1
Everything we did earlier still holds if we freeze time at some t0 . We would then
evaluate the system at that instant as if it were not time-dependent and we were
working with the fixed state
|i
|(t )i
n
X
c0k |uk i ,
k=1
c0k
ck (t ) .
To get to time t = t0 (or any future t > 0), though, we need to know the exact
formula for those coefficients, ck (t), so we can plug-in t = t0 and produce this fixed
state. Thats where the sixth postulate of quantum mechanics comes in.
Trait #12 (The Time-Dependent Schr
odinger Equation)
The time evolution of a state vector is governed by the Schrodinger Equation
i~
d
| (t) i
dt
H(t) | (t) i .
Notice that the Hamiltonian can change from moment-to-moment, although it does
not always do so. For our purposes, it is not time dependent, so we get a simpler
form of the Schrodinger Equation,
i~
d
| (t) i
dt
=
196
H | (t) i .
This is still a time-dependent equation; we merely have a simplification acknowledging
that t does not appear in the matrix for H.
The Other Schr
odinger Equation? In case youre wondering whether
theres a time-independent Schrodinger equation, the answer is, it depends on whom
you ask. Purists say, not really, but most of us consider the eigenket-eigenvalue
equation of Trait #3,
TA |uk i
ak |uk i ,
when applied to the operator A = H, to be the time-independent Schrodinger equation.
Solving the time-independent Schrodinger equation, which produces the energy
eigenkets and eigenvalues, is typically the first step in the process of solving the more
general time-dependent Schrodinger equation.
8.5.2
The Evolution of Spin in a Constant Magnetic Field
You are ready to solve your first Schrodinger equation. We consider the system of a
stationary electron in a uniform z-up directed B-field, whose Hamiltonian we have
already cracked (meaning we solved the time-independent Schrodinger equation).
Because its so exciting, Im going to summarize the set-up for you. The players
are
our state vector, |(t)i, with its two expansion coefficients, the unknown functions c1 (t) and c2 (t),
!
c1 (t)
| (t) i = c1 (t) |+i + c2 (t) |i =
,
c2 (t)
z
the Hamiltonian operator,
H
~
B
2
1 0
,
0 1
and the Schrodinger equation that relates the two,
i~
d
| (t) i
dt
H | (t) i .
Lets compute. Substitute the coordinate functions in for |i in the Schrodinger
equation,
!
!
c1 (t)
d c1 (t)
~ 1 0
i~
= B
.
dt c2 (t)
2 0 1
c2 (t)
197
This is equivalent to
i~
!
d
c
1
dt
i~
d
c
dt 2
B ~2 c1
!
,
B ~2 c2
or
d
c1
dt
d
c2
dt
Bi
c1
and
2
Bi
c2 .
2
=
=
From calculus we know that the solutions to
dx
dt
kx ,
with constant k, are the family of equations
x(t)
Cekt ,
one solution for each complex constant C. (If you didnt know that, you can verify it
now by differentiating the last equation.) The constant C is determined by the initial
condition, at time t = 0,
C
x(0) .
We have two such differential equations, one for each constant
k =
Bi
,
2
so
c1 (t)
C1 eit(B/2) ,
c2 (t)
C2 eit(B/2) .
The initial condition is that
|(0)i
|i ,
our starting state. In other words, the initial conditions for our two equations are
c1 (0)
c2 (0)
=
=
0
0 ,
and
where we are saying 0 and 0 are the two scalar coefficients of the state |i at time
t = 0. That gives us the constants
C1
C2
=
=
198
0 ,
0 ,
the complete formulas for the time-dependent coefficients,
c1 (t)
0 eit(B/2) ,
c2 (t)
0 eit(B/2) .
and, finally, our time-dependent state in its full glory,
|(t)i
0 eit(B/2) |+i
0 eit(B/2) |i .
We have solved the Schrodinger equation using first quarter calculus.
If you can manage to not be too confused, lets reduce our notation by using c1 and
c2 (without the parameter (t)), rather than 0 and 0 , to mean the initial coefficients
at time t = 0 giving us the slightly cleaner
|(t)i
c1 eit(B/2) |+i
c2 eit(B/2) |i .
We pause to consider how this can be generalized to any situation (in the case of
finite or enumerable eigenvalues). Ill introduce an odd notation that physicists have
universally adopted, namely that the eigenket associated with eigenvalue a, will be
that same a inside the ket symbol, i.e., |ai .
1. We first solved the systems Hamiltonian the time-independent Schrodinger
equation, if you like to get the allowable energies, { Ek } and their associated
eigenkets, { |Ek i }.
2. Next, we expanded the initial state, |i, along the energy basis,
X
ck |Ek i .
|i =
k
3. We solved the Schrodinger equation for each time-dependent amplitude, ck (t),
which yielded
ck (t)
ck eitEk /~ .
4. Finally, we attached the exponentials as factors to each of the terms in the
original sum for |i to get the full, time-dependent state,
X
|(t)i =
ck eitEk /~ |Ek i .
k
This will be packaged into a trait in a minute. But first ...
199
8.5.3
Stationary States
Notice what this implies if our initial state happens to be one of the energy eigenstates.
In our spin-1/2 system that would be either |+i or |i. Take the |i = |+i. The
result is
|(t)i
eit(B/2) |+i ,
Introducing the shorthand t Bt/2, the time evolution merely causes a phase
factor of eit to appear. But remember that our state-space does not differentiate
between scalar multiples of state vectors, so
|(t)i
eit |+i
|+i
|i .
The state does not change over time.
Looked at another way, if you start out in state |+i and measure total energy, H,
with certainly (which we already knew, since the coefficient of |+i is 1,
you get B~
2
2
and 1 = 1). This also means that the state into which |+i evolves over time t, also
has 100% probability of remaining |+i as the coefficient of the |+i reveals:
|c1 (t)|2
c1 (t) c1 (t)
eit(B/2) eit(B/2)
e0
1.
Its one and only expansion coefficient changes by a factor of eit whose square magnitude is 1 regardless of t. This is big enough to call a trait.
Trait #13 (Stationary States)
An eigenstate of the Hamiltonian operator evolves in such a way that its measurement
outcome does not change; it remains in the same eigenstate.
For this reason, eigenstates are often called stationary states of the system.
8.5.4
General (Non-Stationary) States
Does this mean that we can throw away the phase factors eit for times t 6= 0? It
may surprise you that the answer is a big fat no. Lets take the example of an initial
state that is not an energy eigenket. One of the y-eigenstates will suffice,
|i
|iy
|+i i |i
.
2
The two initial coefficients are
c1
c2
and
2
i
.
2
200
Allow this state to evolve for a time t according to the Schrodinger equation,
|(t)i
eit(B/2) |+i
i eit(B/2) |i
.
2
The state is unchanged if we multiply by an overall scalar multiple, so lets turn
the |+i coefficient real by multiplying the entire fraction by its complex conjugate
eit(B/2) , giving the equivalent state
|(t)i
|+i
i eit(B) |i
,
2
whose coefficients are
c1 (t)
c2 (t)
and
2
ieit(B)
.
2
Comparing this to the original state, we see that while the coefficient of |+i never
changes, the coefficient of |i changes depending on t. To dramatize this, consider
time t0 = /(B) (an impossibly tiny fraction of a second if you were to compute it
using real values for and a typical B-field).
|(t0 )i
=
=
|+i
|+i
i ei |i
2
+ i |i
=
2
|+i
i (1) |i
|+iy .
We started out in state |iy , and a fraction of a second later found ourselves in state
|+iy . Thats a pretty dramatic change. If we were to measure Sy initially, we would,
with certainty, receive a -~/2 reading. Wait a mere /(B) seconds later, and a
measurement would result, with equal certainty, in the opposite value +~/2.
8.5.5
General Technique for Computing Time-Evolved States
Weve set the stage for a technique that applies throughout quantum mechanics. Well
call it a trait.
Trait #14 (Evolution of Any Observable)
To determine the time-evolved probability of the outcome of any observable, A, starting in the initial state |i,
1. compute the Energy eigenvalues and eigenkets for the system, { Ek |Ek i} ,
by solving the time-independent Schrodinger equation, H |Ek i = Ek |Ek i,
201
2. expand |i along the energy basis: |i
k ck
|Ek i,
3. attach (as factors)
the time-dependent phase factors, eitEk /~ to each term,
P
itEk /~
|(t)i =
|Ek i,
k ck e
4. dot this expression with the desired eigenket, |uj i of A, to find its amplitude,
j (t) = huj | (t)i, and
5. the magnitude-squared of this amplitude, |j (t)|2 , will be the probability of an
A-measurement producing the eigenvalue ak at time t.
The short version is that for any observable of interest, A, you first solve the systems time-independent Schrodinger equation and use its energy eigenbasis to express
your state, |i. Incorporate the time dependence into that expression and dot that
with the desired eigenstate of A to get your amplitude and, ultimately, probability.
(I keep using quotes with the verb dot, because this is really an inner -product,
rather than a real dot-product, requiring the left vectors coordinates to be conjugated.)
Example
We continue to examine the evolution of an electron that starts in the y-down state,
|iy . Weve already done the first three steps of Trait #14 and found that after
time t the state |iy evolves to
i eit(B) |i
.
2
(I used | ()t iy , rather than the somewhat confusing |(t)iy , to designate the states
dependence on time.)
| ()t iy
|+i
Thats the official answer to the question how does |iy evolve?, but to see how
we would use this information, we have to pick an observable we are curious about
and apply Trait #14, steps 4 and 5.
Lets ask about Sy , the y-projection of spin specifically the probability of measuring a |+iy at time t. Step 4 says to dot the time-evolved state with the vector
|+iy , so the amplitude (step 4) is
c+y
yh +
| ()t iy .
Ill help you read it: the left vector of the inner product is the + ~2 eigenket of the
operator Sy , |+iy , independent of time. The right vector of the inner product is
our starting state, |iy , but evolved to a later time, t.
Because everything is expressed in terms of the z-basis, we have to be sure we
stay in that realm. The z-coordinates of |+iy are obtained from our familiar
!
1/ 2
|+i + i |i
|+iy =
=
.
2
i/ 2 z
202
I added the subscript z on the RHS to emphasize that we are displaying the vector
|+iy in the z-coordinates, as usual. If we are to use this on the left side of a complex
inner product we have to take the conjugate of all components. This is easy to see in
the coordinate form,
1/ 2 , i/ 2 z ,
y h+| =
but lets see how we can avoid looking inside the vector by applying our adjoint
conversion rules to the expression defining |+iy to create a bra for this vector. Ill
give you the result, and you can supply the (very few) details as an ...
[Exercise. Show that
y h+|
|+iy
h+| i h|
. ]
2
Getting back to the computation of the amplitude, c+y , substitute the computed
values into the inner product to get
|+i i eit(B) |i
h+| i h|
c+y =
2
2
iBt
h+ | +i e
h | i
1 eiBt
=
=
.
2
2
(The last two equalities made use of the orthonormality of any observable eigenbasis
(Trait #4).)
Finally, step 5 says that the probability of measuring Sy = + ~2 at any time t is
1 eiBt
1 eiBt
2
|c+y |
= c+y c+y =
,
2
2
where we used the fact (see the complex number lecture) that, for real ,
ei
= ei .
Lets simplify the notation by setting
Bt
and complete the computation with that substitution,
|c+y |2
=
=
=
2 ei ei
4
1
1 ei + ei
2
2
2
1
1
cos .
2
2
203
Undoing the substitution gives us the final result
~
1
1
P Sy (t) = +
=
cos (B t) .
2
2
2
As you can see, the probability of measuring an up-y state oscillates between 0 and
1 sinusoidally over time. Note that this is consistent with our initial state at time
t = 0: cos 0 = 1, so the probability of measuring + ~2 is zero; it had to be since we
started in state |iy , and when you are in a eigenstate (|iy ), the measurement of the
observable corresponding to that eigenstate (Sy ) is guaranteed to be the eigenstates
eigenvalue ( ~2 ). Likewise, if we test precisely at t = /(B), we get 21 12 (1) = 1,
a certainty that we will detect + ~2 , the |+iy eigenvalue.
We can stop the clock at times between those two extremes to get any probability
we like.
[Exercise. What is the probability of measuring Sy (t) = + ~2 at the (chronologically ordered) times
(a) t = /(6B) ,
(b) t = /(4B) ,
(c) t = /(3B) ,
(d) t = /(2B) .
[Exercise. Do the same analysis to get the probability that Sy measured at time
t will be ~2 . Confirm that at any time t, the two probabilities add to 1.]
8.6
Larmor Precession
We complete this lecture by combining expectation value with time evolution to get
a famous result which tells us how we can relate the 3-dimensional real vector of
a classical angular momentum vector to the quantum spin-1/2 electron, which is a
2-dimensional complex vector.
8.6.1
The Time-Evolved Spin State in a Uniform B-Field
We assume that we have many electrons in the same initial spin state, and we look
at that states expansion along the z-basis,
|i
c1 |+i + c2 |i ,
where, by normalization we know that
|c1 |2 +
|c2 |2
204
1.
We let the state (of any one of these systems, since they are all the same) evolve for a
time t. We have already solved the Schrodinger equation and found that the evolved
state at that time will be
|(t)i
=
=
c2 eit(B/2) |i
c1 eit(B/2) |+i +
!
c1 eit(B/2)
.
c2 eit(B/2)
This, then is the state at time t, prior to measurement.
Rewriting |(t)i
It will help to represent this state by an equivalent vector that is a mere unit scalar
multiple of itself. To do that, we first express c1 and c2 in polar form,
c1
c2
c ei1
i2
se
and
,
giving the equivalent state
|(t)i
c ei1 eit(B/2)
s ei2 eit(B/2)
!
.
1 +2
Then we multiply by the unit scalar ei( 2 ) to get a more balanced equivalent
state,
i( 1 2 2 ) it(B/2)
e
ce
.
|(t)i =
1 2
i( 2 ) it(B/2)
se
e
Now, we simplify by making the substitutions
=
=
B
and
1 2 ,
to get the simple and balanced Hilbert space representative of our state,
!
!
c ei (t + 0 )/2
c ei0 /2 eit/2
|(t)i =
=
.
s ei (t + 0 )/2
s ei0 /2 eit/2
We get a nice simplification by using the notation
t + 0
,
2
to express our evolving state very concisely as
!
c ei (t)
|(t)i =
.
s ei (t)
(t)
205
A Convenient Angle
There is one last observation before we start to compute. Since |i is normalized,
|c|2 +
|s|2
1,
the amplitudes c and s have moduli (absolute values) that are consistent with the
sine and cosine of some angle. Furthermore, we can name that angle anything we
like. Call it /2 for reasons that will become clear in about 60 seconds. [Start of
60 seconds.]
We have proclaimed the angle to be such that
c
sin ,
2
cos
and
which is why I named the amplitudes c and s.
Also, were going to run into the two expressions cs and c2 s2 a little later,
so lets see if we can write those in terms of our angle . The addition law of sines
implies that
cs
cos
sin
2
2
sin
,
2
while the addition law of cosines yields
c2 s 2
cos2
sin2
2
2
cos .
By letting /2 be the common angle that we used to represent c and s (instead of,
say, ) we ended up with plain old on the RHS of these formulas, which is the form
well need. [End of 60 seconds.]
Although well start out using c and s for the moduli of |is amplitudes, well
eventually want to make these substitutions when the time comes. The angle will
have a geometric significance.
8.6.2
Evolution of the Spin Expectation Values
Experiment #4: Measuring |(t)i at Time t
We ask the physicists to measure one of Sx , Sy or Sz at time t. Now they can
only measure one of those observables per electron, because once they do, |(t)i will
collapse into one of the six basis states, |i, |ix or |iy , after which time, it becomes
useless. But thats okay; were swimming in |(t)i electrons. We pick a very large
number, N , (say N is a million). We test 3N electrons, measuring Sz on the first
N , Sx on the second N , and Sy on the third N . We will have measured the state
206
|(t)i 3N or 3 million times. The physicists record the measurements producing a
certain number of +z values, a certain number of z, etc. We ask them to compute
the average of the Sz results a number between - ~2 and + ~2 , and the same with the
Sx and Sy results.
So theyll have three numbers in the end, mx , my and mz .
But hold on a second. We dont have to bother the physicists, because when N
is large, we know from the the law of large numbers that the average values of each
of the three spin projections are approximated very closely by the expectation values
of the operators. So, lets compute those instead of wasting a lot of time and money.
The Expectation Values at Time t
We do this for each observable Sz , Sx and Sy , individually, then find a way to combine
the answers. Well begin with hSz (t)i.
A. The Expectation Value for the z-spin Observable: hSz (t)i
Trait #9 (the expectation value theorem), tells us that we can compute this using
h(t) | Sz | (t)i .
With the help of Trait #8 (the adjoint conversion rules) and keeping in mind that
c and s are real, we find
h(t) | Sz | (t)i
=
=
=
~
c ei (t) , s ei (t)
2
c ei (t)
s ei (t)
0 1
~
c ei(t)
i (t)
i (t)
ce
, se
2 s ei(t)
~ 2
~
c s2
=
cos .
2
2
We can draw some quick conclusions from this (and subtler ones later).
1. The expectation value of Sz does not change with time.
2. If |i = |+i, i.e., c = 1 and s = 0, the expectation value is + ~2 , as it must, since
|+i is a stationary state and so always yields a (+) measurement.
3. If |i = |i, i.e., c = 0 and s = 1, the expectation value is ~2 , again consistent
with |i being the other stationary state, the one that always yields a ()
measurement.
207
B. The Expectation Value for the x-spin Observable: hSx (t)i
We compute
h(t) | Sx | (t)i .
Using our adjoint conversion rules again, we find
h(t) | Sx | (t)i
=
~
c ei (t) , s ei (t)
2
~
c ei (t) , s ei (t)
2
0 1
c ei (t)
s ei (t)
1 0
s ei (t)
c ei (t)
+ cs e2i (t) .
~
cs e2i (t)
2
This is nice, but a slight rearrangement should give you a brilliant idea,
=
e2i (t) + e2i (t)
.
2
Look back at our lesson on complex numbers, especially the consequences of the Euler
formula, and youll discover that the fraction simplifies to cos (2(t)). Now we have
h(t) | Sx | (t)i
h(t) | Sx | (t)i
cs ~
cs ~ cos (2(t)) ,
which, after undoing our substitutions for cs and (t) we set up in the convenient
angle section, looks like
h(t) | Sx | (t)i
cs ~ cos (t + 0 )
~
sin cos (t + 0 ) .
2
Observe these consequences.
1. The expectation value of Sx varies sinusoidally with time at a frequency = B,
not counting the situation in item 2.
2. If |i is either |+i or |i, then cs = 0 producing an expectation value of 0 at
all times. This is consistent with the following two facts.
These two z-eigenkets, when expressed in the x-basis, have equal doses
( 12 ) of |+ix and |ix , as you can tell from
|i
|+ix |ix
,
2
so we would expect a roughly equal collapse into the |+ix and |ix states,
averaging to 0.
Weve already established that the two kets |i are stationary states of H,
so whatever holds at time t = 0, holds for all time.
208
C. The Expectation Value for the y-spin Observable: hSy (t)i
We compute
h(t) | Sy | (t)i ,
The calculation proceeds much like that of hSx i.
h(t) | Sy | (t)i
=
~
c ei (t) , s ei (t)
2
c ei (t) , s ei (t)
~
cs e2i (t)
2i
~
2
!
0 i
i
c ei (t)
is ei (t)
s ei (t)
!
ic ei (t)
cs e2i (t)
Rearranging and applying one of our Euler formulas, we find
h(t) | Sy | (t)i
e2i (t) e2i (t)
2i
cs ~ sin (2(t))
cs ~ sin (t + 20 )
cs ~
~
sin sin (t + 0 ) .
2
Again, some observations.
1. The expectation value of Sy varies sinusoidally with time at a frequency = B,
not counting the situation in item 2.
2. If |i is either |+i or |i, then cs = 0 yielding an expectation value of 0 for
all time. Again, this is consistent with these two z-eigenkets, being stationary
states and having y-basis amplitudes of equal magnitude ( 12 ),
|+i
|+iy + |iy
|i
|+iy |iy
,
i 2
and
(from a prior exercise).
The Expectation Vector and Larmor Precession
In classical mechanics, Larmor precession describes the way in which a magnetic
dipoles axis a real 3-dimensional vector revolves about a magnetic field vector.
209
In contrast, the quantum mechanical spin state-vector lives in a 2-dimensional Hilbert
space, not 3-dimensional real space, so we dont have simultaneously measurable x, y,
and z-components which we can study. However, we can define a real (and evolving)
3-dimensional vector s(t) to be
s(t)
hSx i|(t)i
hS i
y |(t)i .
hSz i|(t)i
This s(t) is a true 3-dimensional (time-dependent) vector whose real coordinates are
the three expectation values, hSx i, hSy i and hSz i, at time, t.
In the previous section we showed that
~
sin cos (t + 0 )
2
s(t) = ~2 sin sin (t + 0 )
~
2
cos
sin cos (t + 0 )
sin sin (t + 0 ) .
2
cos
If this is not speaking to you, drop the factor of ~/2 and set (t) = t + 0 . What
we get is the 3-dimensional vector
sin cos (t)
s(t) sin sin (t) .
cos
It is a unit vector in R3 whose spherical coordinates are (1, , (t)), i.e., it has a polar
angle and azimuthal angle (t). (I dont use exclamation points, but if I did, I would
use one here.) We are looking at a vector that has a fixed z-coordinate, but whose
x and y-coordinates are in a clockwise circular orbit around the origin (clockwise
because of the y-coordinates minus sign [Exercise]). This is called precession.
Since our B-field was defined to point in the +z direction, we have discovered the
meaning of the vector s(t) = (hSx (t)i , hSy (t)i , hSz (t)i )t .
This is something that recurs with regularity in quantum physics. The quantum state vectors, themselves, do not behave like classical vectors. However, their
expectation values do.
8.6.3
Summary of Larmor Precession
When a motionless electron is initially in spin state
!
c ei1
|i =
.
s ei2
210
within a uniform magnetic field B pointing in the +z direction, the Schrodinger equation tells us that its expectation value vector, s(t), evolves according to the formula
s(t)
sin cos (t + 0 )
sin sin (t + 0 ) .
2
cos
, and 0 are parameters related to |i and B as follows:
is the polar angle that the real vector, s(t), makes with the z-axis in our R3 .
It relates to |i in that /2 is the angle that expresses the magnitudes c and s
of |is Hilbert-space coordinates,
c
sin .
2
cos
and
/2 ranges from 0 (when it defines the up-z state, |+i) to /2 (when it defines
the down-z state, |i state), allowing to range from 0 to in R3 .
is the Larmor frequency, defined by the magnitude of the B-field and the
constant, (the gyromagnetic ratio),
B .
0 is the relative phase of the two amplitudes,
0
8.7
1 2 .
The End and the Beginning
This completes our tutorial on both time-independent and time-evolving quantum
mechanics. If you studied this last lesson as part of the required reading for CS 83B,
you are ready to move on to the next phase in that course. If you chose to read it as
an optional section in CS 83A, good for you. Its time to start learning about qubits.
211
Chapter 9
The Qubit
|i
9.1
|0i + |1i
Bits and Qubits as Vector Spaces
All the hard work is done. You have mastered the math, spin-physics and computational quantum mechanics needed to begin doing quantum computer science.
While useful to our understanding of quantum mechanics, the spin-1/2 physical
system, S , is no longer useful to us as computer scientists. We will work entirely
inside its state space H from this point forward, applying the postulates and traits
of quantum mechanics directly to that abstract system, indifferent to the particular
S that the Hilbert space H is modeling.
Although in practical terms a quantum bit, or qubit, is a variable superposition of
two basis states in H of the form shown at the top of this chapter, formally qubits
as well as classical bits are actually vector spaces. You read that correctly. A single
qubit is not a vector, but an entire vector space. Well see how that works shortly.
To get oriented, well do a quick review of classical bits in the new formalism, then
well retrace our steps for qubits. Lets jump in.
9.2
9.2.1
Classical Computation Models Informal Approach
Informal Definition of Bits and Gates
In order to study the qubit, lets establish a linguistic foundation by defining the more
familiar classical bit.
212
Folksy Definition of a Bit
Heres one possible way to define a bit without going formal.
A Bit (Informal). A bit is an entity, x, capable of being in one of
two states which we label 0 and 1.
The main take-away is that 0 and 1 are not bits. They are the values or states that
the bit can attain.
We can use the notation
x = 0
to mean that x is in the state 0, an observation (or possibly a question) about the
state bit x is in. We also use the same notation to express the imperative, put x into
the state 0. The latter is the programmers assignment statement.
Folksy Definition of a Logical Operator a.k.a. Gate
What about the logical operators like AND or XOR which transform bits to other
bits? We can define those, too, using similarly loose language.
Classical Logical (Boolean) Operator. A logical operator is a function that takes one or more bits as input and produces a single bit as
output.
Logical operators are also called logic gates or just gates when implemented in
circuits diagrams.
Note: We are only considering functions that have a single output bit. If one
wanted to build a logic gate with multiple output bits it could be done by combining
several single-output logic gates, one for each output bit.
Examples of Logical Operators
The Negation Operator. NOT (symbol ) is a logical operator on a single bit
defined by the formula
(
0 , if x = 1 and
x
.
1 , otherwise
A common alternate notation for the NOT operator is the overline, x = x.
The Exclusive-Or Operator. XOR (symbol ) is an operator on two bits
defined by the formula
(
0 , if x = y and
xy
.
1 , otherwise
213
Mathematically, x y is called the mod-2 sum of x and y, language that is used
throughout classical and quantum logic.
Unary and Binary Operators
While logic gates can take any number of inputs, those that have one or two inputs
are given special names.
A unary operator is a gate that takes single input bit, and a binary
operator is one that takes two input bits.
As you can see, NOT is a unary operator while XOR is a binary operator.
Truth Tables
Informally, unary and binary operators are often described using truth tables. Here
are the traditional gate symbols and truth tables for NOT and XOR.
x
y
xy
x (or x)
xy
We could go on like this to define more operators and their corresponding logic gates.
However, its a little disorganized and will not help you jump to a qubit, so lets try
something a little more formal.
Our short formal development of classical logic in this lesson will be restricted to
the study of unary operators. We are learning about a single qubit today which is
analogous to one classical bit and operators that act on only one classical bit, i.e.,
unary operators. It sounds a little boring, I know, but thats because in classical logic,
unary operators are boring (there are only four). But as you are about to see, in the
quantum world there are infinitely many different unary operators, all useful.
214
9.3
Classical Computation Models Formal Approach
The definition of a quantum bit is necessarily abstract, and if I were to define it at
this point, you might not recognize the relationship between it and a classical bit.
To be fair to qubits, well give a formal definition of a classical bit first, using the
same language we will need in the quantum case. This will allow us to establish some
vocabulary in the classical context that will be re-usable in the quantum world and
give us a reference for comparing the two regimes on a level playing field.
9.3.1
A Miniature Vector Space
You are familiar with R2 , the vector space of ordered pairs of real numbers. Id like
to introduce you to an infinitely simpler vector space, B = B2 .
The Tiny Field B
In place of the field R of real numbers, we define a new field of numbers,
B
{0, 1} .
Thats right, just the set containing two numbers. But we allow the set to have
addition, , and multiplication, , defined by
is addition mod-2
0 1 = 1
1 0 = 1
0 0 = 0
1 1 = 0
is ordinary
0 1
1 0
0 0
1 1
multiplication
= 0
= 0
= 0
= 1
Of course is nothing other than the familiar XOR operation, although in this
context we also get negative mod-2 numbers (1 = 1) and subtraction mod-2 (0 1 =
0 (1) = 1), should we need them.
The Vector Space B
We define B = B2 to be the vector space whose scalars come from B and whose vectors
(objects) are ordered pairs of numbers from B. This vector space is so small I can list
215
its objects on one line,
2
B = B
0
0
1
1
,
,
,
.
0
1
0
1
Im not going to bother proving that B obeys all the properties of a vector space, and
you dont have to either. But if you are interested, its a fun ...
[Exercise. Show that B obeys the properties of a field (multiplicative inverses,
distributive properties, etc.) and that B obeys the properties of a vector space.]
The Mod-2 Inner Product
Not only that, but there is an inner product,
x1
x2
= x1 x2 y1 y2 .
y1
y2
Youll notice something curious about this inner product (which is why I put it in
quotes and used the operator rather than ). The vector (1, 1)t is a non-zero
vector which is orthogonal to itself. Dont let this bother you. In computer science,
such non-standard inner products exist. (Mathematicians call them pairings, since
they are not positive definite, i.e., it is not true that v 6= 0 kvk > 0. Well
just call a weird inner product.) This one will be the key to Simons algorithm,
presented later in the course. The inner product gives rise to a modulus or length
for vectors through the same mechanism we used in R2 ,
kxk = |x|
x x,
where I have shown two different notations for modulus. With this definition we find
0
=
1
= 1 and
0
1
0
=
1
= 0.
0
1
The strange but necessary equality on the lower right is a consequence of this
oddball inner product on B = B2 .
[Exercise. Verify the above moduli.]
[Notation. Sometimes Ill use an ordinary dot for the inner product, (x1 , y1 )t
(x2 , y2 )t instead of the circle dot, (x1 , y1 )t (x2 , y2 )t . When you see a vector on
each side of youll know that we really mean the mod-2 inner product, not mod-2
multiplication.]
Dimension of B
If B is a vector space, what is its dimension, and what is its natural basis? Thats
not hard to guess. The usual suspects will work. Its a short exercise.
216
[Exercise. Prove that B is 2-dimensional by showing that
1
0
,
0
1
form an orthonormal basis. Hint: There are only four vectors, so express each in
this basis. As for linear independence and orthonormality, I leave that to you.]
9.3.2
Formal Definition of a (Classical) Bit
Definition of Bit. A bit is (any copy of ) the entire vector space B.
Sounds strange, I know, making a bit equal to an entire vector space. We think of a
bit as capable of holding a 1 or a 0. This is expressed as follows.
Definition of Bit Value. A bits value or state is any normalized
(i.e., unit) vector in B.
A bit, itself, is not committed to any particular value until we say which unit-vector
in B we are assigning it. Since there are only two unit vectors in B, that narrows the
field down to one of two values, which Ill label as follows:
1
[0]
and
0
0
[1]
.
1
The other two vectors (0, 0)t and (1, 1)t , have length 0 so cannot be normalized (i.e.,
we cannot divide them by their length to form a unit vector).
How Programmers Can Think About Formal Bits
If the definition feels too abstract, try this out for size. As a programmer, youre
familiar with the idea of a variable (LVALUE ) which corresponds to a memory location. That variable is capable of holding one specific value (the RVALUE ) at any
point in time, although there are many possible values we can assign it. The vector
space B is like that memory location: it is capable of holding one of several different
values but is not committed to any until we make the assignment. Putting a value
into a memory location with an assignment statement like x = 51; corresponds to
choosing one of the unit vectors in B to be assigned to the bit. So the values allowed
to be stored in the formal bit (copy B) are any of the unit vectors in B.
217
Multiple Bits
If we have several bits we have several copies of B, and we can name each with a
variable like x, y or z. We can assign values to these variables with the familiar
syntax
x
[1] ,
[1] ,
[0] ,
etc.
A classical bit (uncommitted to a value) can also be viewed as a variable linear
combination of the two basis vectors in B.
Alternate Definition of Bit. A bit is a variable superposition of the
two natural basis vectors of B,
x
[0] + [1] ,
2 2
where
1.
Since and are scalars of B, they can only be 0 and 1, so the normalization
condition implies exactly one of them is 1 and the other is 0.
Two questions are undoubtedly irritating you.
1. Is there any benefit, outside of learning quantum computation, of having such
an abstract and convoluted definition of a classical bit?
2. How will this definition help us grasp the qubit of quantum computing?
Keep reading.
9.3.3
Formal Definition of a (Classical) Logical Operator
We will only consider unary (i.e., one bit) operators in this section. Binary operators
come later.
Unary Operators
Definition of Logical Unary Operator. A logical unary operator or (one bit operator) is a linear transformation of B that maps
normalized vectors to other normalized vectors.
This is pretty abstract, but we can see how it works by looking at the only four logical
unary operators in sight.
218
The constant-[0] operator A(x) [0]. This maps any bit into the 0-bit.
(Dont forget, in B, the 0-bit is not the 0-vector, it is the unit vector (1, 0)t .)
Using older, informal truth tables, we would describe this operator as follows:
x
[0]-op
In our new formal language, the constant-[0] operator corresponds to the linear
transformation whose matrix is
1 1
,
0 0
since, for any unit vector (bit value) (, )t , we have
1 1
1
=
=
0 0
0
0
[0] ,
the second-from-last equality is due to the fact that, by the normalization requirement on bits, exactly one of and must be 1.
The constant-[1] operator A(x) [1]. This maps any bit into the 1-bit.
Informally, it is described by:
x
[1]-op
Formally, it corresponds to the linear transformation
0 0
,
1 1
since, for any unit vector (bit value) (, )t , we have
0 0
0
0
=
=
1 1
[1] .
The negation (or NOT) operator A(x) x. This maps any bit into its
logical opposite. It corresponds to
0 1
,
1 0
219
since, for any x = (, )t , we get
0 1
0 1
x =
1 0
1 0
x.
[Exercise. Using the formula, verify that [0] = [1] and [1] = [0].]
The identity operator A(x) 1x = x. This maps any bit into itself and
corresponds to
1 0
.
0 1
[Exercise. Perform the matrix multiplication to confirm that 1[0] = [0] and
1[1] = [1].]
Linear Transformations that are Not Unary Operators
Apparently any linear transformation on B other than the four listed above will not
correspond to a logical unary operator. For example, the zero-operator
0 0
0 0
isnt listed. This makes sense since it does not map normalized vectors to normalized
vectors. For example,
0 0
0 0
1
0
[0] =
=
= 0,
0 0
0 0
0
0
not a unit vector. Another example is the matrix
1 0
,
1 0
since it maps the bit
1
1
[1] to the non-normalizable 0,
0
1 0
0
[1] =
=
0
1 0
1
0
0
0.
[Exercise. Demonstrate that
1 1
1 1
does not represent a logical operator by finding a unit vector that it maps to a nonunit vector, thus violating the definition of a logical unary operator.]
Defining logical operators as matrices provides a unified way to represent the
seemingly random and unrelated four logical operators that we usually define using
isolated truth tables. Theres a second advantage to this characterization.
220
Reversible Logic Gates
A reversible operator or logic gate is one that can be undone by applying another operator (which might be the same as the original). In other words, its associated matrix
has an inverse. For example, the constant-[1] operator is not reversible, because it
forces its input to the output state [1],
0 0
0
=
,
1 1
1
and theres no way to reconstruct (, )t reliably from the constant bit [1]; it has
erased information that can never be recovered. Thus, the operator, and its associated
logic gate, is called irreversible.
Of the four logical operators on one classical bit, only two of them are reversible:
the identity, 1, and negation, . In fact, to reverse them, they can each be reapplied
to the output to get the back original input bit value with 100% reliability.
Characterization of Reversible Operators.
versible its matrix is unitary.
An operator is re-
Example. We show that the matrix for is unitary by looking at its columns.
The first column is
0
,
1
which
has unit length, since (0, 1)t (0, 1)t = 1, and
is orthogonal to the to the second column vector, since (0, 1)t (1, 0)t = 0.
Reversibility is a very powerful property in quantum computing, so our ability to
characterize it with a simple criterion such as unitarity is of great value. This is easy
to check in the classical case, since there are only four unary operators: two have
unitary matrices and two do not.
[Exercise. Show that the matrix for 1 is unitary and that the two constant
operators matrices are not.]
An Anomaly with Unitary Matrices in B
We learned that unitary transformations have the following equivalent properties.
1. They preserve the lengths of vectors: kAvk = kvk ,
2. they preserve inner products, h Av | Aw i = hv | wi , and
221
3. their rows (or columns) are orthonormal.
While this is true of usual (i,e., positive definite) inner products, the pairing in B
causes these conditions to lose sync. All four possible logical operators on bits do
preserve lengths and so meet condition 1; weve already shown that they map unit
vectors to unit vectors, and we can also check that they map vectors of length zero
to other vectors of length zero:
[Exercise. Verify that the constant-[1] operator maps the zero-length vectors
(1, 1)t and (0, 0)t to the zero-length (0, 0)t . Do the same for the constant-[0] operator.
Thus both of these non-unitary operators preserve the (oddball) mod-2 length.]
However, the constant-[1] and [0] operators do not preserve inner products, nor do
they have unitary matrices.
[Exercise. Find two vectors in B that have inner product 0, yet whose two images
under constant-[1] operator have inner product 1.]
This unusual situation has the consequence that, of the four linear transformations
which qualify to be logical operators (preserving lengths) in B, only two are reversible
(have unitary matrices).
Although the preservation-of-lengths condition (1), is enough to provide B
with four logical operators, that condition is not adequate to assure that
they are all unitary (reversible). Only two of the operators satisfy the
stronger requirements 2 or 3.
This distinction between length-preserving operators and unitary operators is not
replicated in qubits, and the consequences will be profound.
9.4
The Qubit
We are finally ready to go quantum. At each turn there will be a classical concept
and vocabulary available for comparison because we invested a few minutes studying
the formal definitions in that familiar context.
9.4.1
Quantum Bits
Definition of Qubit. A quantum bit, a.k.a. qubit, is (any copy
of ) the entire vector space H.
Notice that we dont need to restrict ourselves, yet, to the projective sphere. That
will come when we describe the allowed values a bit can take.
222
Comparison with Classical Logic
The classical bit was a 2-D vector space, B, over the finite scalar field B, and now we
see that a quantum bit is a 2-D vector space, H, over the infinite scalar field C.
Besides the underlying scalar fields being different the tiny 2-element B vs. the
infinite and rich C the vectors themselves are worlds apart. There are only four
vectors in B, while H has infinitely many.
However, there is at least one similarity: they are both two dimensional. The
natural basis for B is [0] and[1], while the natural basis for H is |+i and |i.
Quantum Computing Vocabulary
The quantum computing world has adopted different terms for many of the established
quantum physics entities. This will be the first of several sections introducing that
new vocabulary.
Symbols for Basis Vectors. To reinforce the connection between the qubit
and the bit, we abandon the vector symbols |+i and |i and in their places use |0i
and |1i same vectors, different names. This is true whether we are referring to the
preferred z-basis, or any other orthonormal basis.
|+i |0i
|i |1i
|+ix |0ix
|ix |1ix
|+iy |0iy
|iy |1iy
Alternate x-Basis Notation. Many authors use the shorter notation for the
x-basis,
|0ix |+i
|1ix |i ,
but I will eschew that for the time being; |+i and |i already have a z-basis meaning
in ordinary quantum mechanics, and using them for the x-basis too soon will cause
confusion. However, be prepared for me to call |+i and |i into action as x-basis
CBS, particularly when we need the variable x for another purpose.
Computational Basis States. Instead of using the term eigenbasis, computer
scientists refer to computational basis states (or CBS when Im in a hurry). For
example, we dont talk about the eigenbasis of Sz , { |+i , |i }. Rather, we speak of
the preferred computational basis, { |0i , |1i }. You are welcome to imagine it as being
associated with the observable Sz , but we really dont care what physical observable
led to this basis. We only care about the Hilbert space H, not the physical system,
S , from which it arose. We dont even know what kind of physics will be used to
build quantum computers (yet). Whatever physical hardware is used, it will give us
the 2-D Hilbert space H.
223
Alternate bases like { |0ix , |1ix }, { |0iy , |1iy } or even { |0in , |1in } for some di , when needed, are also called computational bases, but we usually qualify
rection n
them using the term alternate computational basis. We still have the short-hand
terms z-basis, x-basis, etc., which avoid the naming conflict, altogether. These alternate computational bases are still defined by their expansions in the preferred, z-basis
( |1ix = |0i2|1i , e.g.), and all the old relationships remain.
9.4.2
Quantum Bit Values
Definition of Qubit Value. The value or state of a qubit is any
unit (or normalized) vector in H.
In other words, a qubit is an entity the Hilbert space H whose value can be any
vector on the projective sphere of that space.
A qubit, itself, is not committed to any particular value until we say which specific
unit-vector in H we are assigning to it.
Normalization
Why do we restrict qubit values to the projective sphere? This is a quantum system,
so states are always normalized vectors. Thats how we defined the state space in our
quantum mechanics lesson. A qubits state (or the state representing any quantum
system) has to reflect probabilities which sum to 1. That means the magnitudesquared of all the amplitudes must sum to 1. Well, thats the projective sphere. Any
vectors off that sphere cannot claim to reflect reality.
Certainly there will be times when we choose to work with un-normalized vectors, especially in CS 83B, but that will be a computational convenience that must
eventually be corrected by normalizing the answers.
Comparison with Classical Logic
Just as a classical bit was capable of storing a state [0] or [1] (the two unit vectors
in B), qubits can be placed in specific states which are normalized vectors in H.
This time, however, there are infinitely many unit vectors in H: we have the entire
projective sphere from which to choose.
9.4.3
Usual Definition of Qubit and its Value
A more realistic working definition of qubit parallels that of the alternative formal
definition of bit.
224
Alternative Definition of Qubit. A qubit is a variable superposition
of the two natural basis vectors of H,
|i
|0i + |1i ,
||2 + ||2
where the complex scalars satisfy
1.
I used the word variable to call attention to the fact that the qubit stores a value,
but is not the value, itself.
Global Phase Factors
We never forget that even if we have a normalized state,
|i
|0i + |1i ,
it is not a unique representation; any length-1 scalar multiple
ei |i
ei |0i + ei |1i ,
real ,
still resides on the projective sphere, and since it is a scalar multiple of |i, it is a
valid representative of same state or qubit value. We would say |i and ei |i differ
by an overall, or global, phase factor , a condition that does not change anything,
but can be used to put |i into a more workable form.
Comparison with Classical Logic
The alternate/working expressions for bits in the two regimes look similar.
Classical:
x = [0] + [1] ,
Quantum:
|i = |0i + |1i ,
2 2
||2 + ||2
= 1
= 1
In the classical case, and could only be 0 or 1, and they could not be the same
(normalizability), leading to only two possible states for x, namely, [0] or [1]. In the
quantum case, and can be any of an infinite combination of complex scalars,
leading to an infinite number of distinct values for the state |i.
9.4.4
Key Difference Between Bits and Qubits
Whats the big idea behind qubits? They are considerably more difficult to define and
study than classical bits, so we deserve to know what we are getting for our money.
225
Parallel Processing
The main motivation comes from our Trait #6, the fourth postulate of quantum
mechanics. It tells us that until we measure the state |i, it has a probability of
landing in either state, |0i or |1i, the exact details given by the magnitudes of the
complex amplitudes and .
As long as we dont measure |i, it is like the Centaur of Greek mythology: part |0i
and part |1i; when we process this beast with quantum logic gates, we will be sending
both alternative binary values through the hardware in a single pass. That will change
it to another normalized state vector (say |i), which has different amplitudes, and
that can be sent through further logic gates, again retaining the potential to be part
|0i and part |1i, but with different probabilities.
Classical Computing as a Subset
Classical logic can be emulated using quantum gates.
This is slightly less obvious than it looks, and well have to study how one constructs even a simple AND gate using quantum logic. Nevertheless, we can correctly
forecast that the computational basis states, |0i and |1i, will correspond to the classical bit values [0] and [1]. These two lonely souls, however, are now swimming in a
continuum of non-classical states.
Measurement Turns Qubits into Bits
Well, this is not such a great selling point. If qubits are so much better than bits,
why not leave them alone? Worse still, our Trait #7, the fifth postulate of quantum
mechanics, means that even if we dont want to collapse the qubit into a computational
basis state, once we measure it, we will have done just that. We will lose the exquisite
subtleties of and , turning the entire state into a |0i or |1i.
Once you go down this line of thought, you begin to question the entire enterprise.
We cant get any answers if we dont test the output states, and if they always collapse,
what good did the amplitudes do? This skepticism is reasonable. For now, I can only
give you some ideas, and ask you to wait to see the examples.
1. We can do a lot of processing below the surface of the quantum ocean, manipulating the quantum states without attempting an information-destroying
measurement that brings them up for air until the time is right.
2. We can use Trait #6, the fourth postulate, in reverse: Rather than looking at
amplitudes as a prediction of experimental outcomes, we can view the relative
distribution of several measurement outcomes to guess at the amplitudes of the
output state.
3. By preparing our states and quantum logic carefully, we can load the dice so
226
that the likelihood of getting an information-rich collapsed result will be greatly
enhanced.
9.5
Quantum Operators (Unary)
9.5.1
Definition and Notation
Definition of a Unary Quantum Logical Operator. A unary
quantum logical operator or unary quantum gate is a linear
transformation of H that maps normalized (unit) vectors to other normalized vectors.
The first important observation is that since H is 2-dimensional, a unary quantum
operator can be represented by a 2 2 matrix
a00 a01
.
a10 a11
Notation. I am now numbering starting from 0 rather than 1. This will continue
for the remainder of the course.
The big difference between these matrices and those of classical unary operators is
that here the coefficients come from the infinite pool of complex numbers. Of course,
not every 2 2 matrix qualifies as a quantum gate, a fact well get to shortly.
9.5.2
Case Study Our First Quantum Gate, QNOT
Well examine every angle of this first example. It is precisely because of its simplicity that we can easily see the important differences between classical and quantum
computational logic.
The Quantum NOT (or QNOT) Operator, X
The QNOT operator swaps the amplitudes of any state vector. It corresponds to
0 1
,
1 0
the same matrix that represents the NOT operator, , of classical computing. The
difference here is not in the operator but in the vast quantity of qubits to which we
can apply it. Using |i = (, )t , as we will for this entire lecture, we find
0 1
X |i =
=
.
1 0
In the special case of a CBS ket, we find that this does indeed change the state from
|0i to |1i and vice versa.
[Exercise. Using the formula, verify that X |0i = |1i and X |1i = |0i.]
227
Notation and Vocabulary
The X operator is sometimes called the bit flip operator, because it flips the CBS
coefficients, . In the special case of a pure CBS input, like |0i, it flips it to
the other CBS, |1i.
The reason QNOT is usually labeled using the letter X is that, other than the
factor of ~2 , the matrix is the same as the spin-1/2 observable Sx . In fact, youll recall
from the quantum mechanics lesson that QNOT is precisely the Pauli spin matrix in
the x-direction,
0 1
X = x
=
.
1 0
Its best not to read anything too deep into this. The matrix that models an observable is used differently than one that performs a reversible operation on a qubit.
Here, we are swapping amplitudes and therefore negating computational basis states.
Thats the important take-away.
Gate Symbol and Circuits
In a circuit diagram, we would use
X
when we want to express an X gate or, less frequently, if we want to be overly explicit,
QN OT
We might show the effect of the gate right on the circuit diagram,
|0i + |1i
|0i + |1i .
The input state is placed on the left of the gate symbol and the output state on the
right.
Computational Basis State Emphasis
Because any linear operator is completely determined by its action on a basis, you will
often see an operator like X defined only on the CBS states, and you are expected to
know that this should be extended to the entire H using linearity. In this case, the
letters x and y are usually used to label a CBS, so
|0i
|xi
or
|1i
228
to be distinguished from |i, which can take on infinitely many superpositions of
these two basis kets. With this convention, any quantum logic gate can be defined
by its action on the |xi (and sometimes |yi, |zi or |wi, if we need more input kets).
For the X gate, it might look like this
|xi
|xi .
This expresses the two possible input states and says that X |0i = |0i = |1i, while,
X |1i = |1i = |0i. Using alternative notation,
|xi
|xi .
In fact, youll often see the mod-2 operator for some CBS logic. If we used that to
define X, it would look like this:
|xi
|1 xi .
[Exercise. Why does the last expression result in a logical negation of the CBS?]
You Must Remember This. The operators (, , etc.) used inside the kets
on the variables x and/or y apply only to the binary values 0 or 1 that label the basis
states. They make no sense for general states. We must extend linearly to the rest
of our Hilbert space.
Sample Problem
Given the definition of the bit flip operator, X, in terms of the CBS |xi,
|xi
|xi ,
what is the action of X on an arbitrary state |i, and what is the matrix for X?
Expand |i along the computational basis and apply X:
X |i
=
=
=
=
=
X ( |0i + |1i)
X |0i + X |1i
|0i + |1i
|1i + |0i
|0i + |1i ,
in agreement with our original definition when viewed in coordinate form.
The matrix for X would be constructed, as with any linear transformation, by
applying X to each basis ket and setting the columns of the matrix to those results.
!
!
0 1
X |0i , X |1i
|1i , |0i
MX =
=
=
.
1 0
229
While the problem didnt ask for it, lets round out the study by viewing X in terms
of a kets CBS coordinates,(, )t . For that, we can read it directly off the final
derivation of X |i, above, or apply the matrix, which I will do now.
0 1
=
.
1 0
One Last Time: In the literature, most logic gates are defined in terms of |xi,
|yi, etc. This is only the action on the CBS, and it is up to us to fill in the blanks,
get its action on the general |i and produce the matrix for the gate.
Comparison with Classical Logic
Weve seen that there is at least one classical operator, NOT, that has an analog,
QNOT, in the quantum world. Their matrices look the same and they affect the
classical bits, [0] / [1], and corresponding CBS counterparts, |0i / |1i, identically, but
thats where the parallels end.
What about the other three classical unary operators? You can probably guess
that the quantum identity,
1 0
1
,
0 1
exhibits the same similarities and differences as the QNOT did with the NOT. The
matrices are identical and have the same effect on classical/CBS kets, but beyond
that, the operators work in different worlds.
[Exercise. Express the gate (i.e., circuit) definition of 1 in terms of a CBS |xi
and give its action on a general |i. Discuss other differences between the classical
and quantum identity.]
That leaves the two constant operators, the [0]-op and the [1]-op. The simple
answer is there are no quantum counterparts for these. The reason gets to the heart
of quantum computation.
Unitarity is Not Optional
The thing that distinguished and 1 from [0]-op and the [1]-op in the classical case
was unitarity. The first two were unitary and the last two were not. How did those
last two even sneak into the classical operator club? They had the property that they
preserved the lengths of vectors. Thats all an operator requires. But these constant
ops were not reversible which implied that their matrices were not unitary. The
quirkiness of certain operators being length-preserving yet non-unitary was a fluke
of nature caused by the strange mod-2 inner product on B. It allowed a distinction
between length-preservation and unitary.
In quantum computing, no such distinction exists. The self-same requirement
that operators map unit vectors to other unit vectors in H forces operators to be
230
unitary. This is because H has a well-behaved (positive definite) inner product. We
saw that with any positive definite inner product, unitarity and length-preservation
are equivalent.
But if you dont want to be that abstract, you need only try on either constant
op for size and see if it fits. Lets apply an attempted analog of the [0]-op to the unit
vector |0ix . For fun, well do it twice. First try out the (non-unitary) matrix for the
classical [0]-op in the quantum regime. We find
!
|0i + |1i
1
1 1
1 1
1
1 1
0
=
+
0 0
0
0
0
0
0
1
2
2
!
1
1
1
2
=
+
=
,
0
0
0
2
not a unit vector. To see a different approach we show it using a putative operator,
A, defined by its CBS action that ignores the CBS input, always answering with a
|0i,
|xi
|0i .
Apply A to the same unit vector |0ix .
|0i + |1i
A |0i + A |1i
A
=
2
2
2 |0i ,
=
=
|0i + |0i
again, not normalized.
Measurement
Finally, we ask what impact QNOT has on the measurement probabilities of a qubit.
Here is a picture of the circuit with two potential measurement access points, A
(before the gate) and B (after ):
|i
X
Of course, we know from Trait #7 (fifth postulate of QM) that once we measure the
state it collapses into a CBS, so we cannot measure both points on the same sample
of our system. If we measure at A, the system will collapse into either |0i or |1i and
we will no longer have |i going into X. So the way to interpret this diagram is to
231
visualize many different copies of the system in the same state. We measure some at
access point A and others at access point B. The math tells us that
(
|0i + |1i at Point A
|i =
|0i + |1i at Point B
So, if we measure 1000 identical states at point A,
|i
we will get
# of 0 -measurements
# of 1 -measurements
1000 ||2 and
1000 ||2 ,
by Trait #6, the fourth QM postulate. However, if we do not measure those states,
but instead send them through X and only then measure them (at B),
|i
the same trait applied to the output expression tells us that the probabilities will be
swapped,
# of 0 -measurements
# of 1 -measurements
1000 ||2 and
1000 ||2 ,
Composition of Gates
Recall from the lesson on linear transformation that any unitary matrix, U , satisfies
U U
UU
1,
where U is the conjugate transpose a.k.a, adjoint, of U . In other words its adjoint is
also its inverse.
Because unitarity is required of all quantum gates in particular QNOT we
know that
X X
But the matrix for X tells us that
0
X
=
1
XX
1,
it is self-adjoint:
1
0 1
=
0
1 0
Therefore, in this case, we get the even stronger identity
X2 =
232
1.
X.
X is its own inverse. If we apply X (or any self-adjoint operator) consecutively
without an intervening measurement, we should get our original state back. For
QNOT, this means
|i
|i ,
which can be verified either algebraically, by multiplying the vectors and matrices, or
experimentally, by taking lots of sample measurements on identically prepared states.
Algebraically, for example, the state of the qubit at points A, B and C,
|i
will be
This leads to the expectation that for all quantum gates as long as we dont measure
anything we can keep sending the output of one into the input of another, and while
they will be transformed, the exquisite detail of the qubits amplitudes remain intact.
They may be hidden in algebraic changes caused by the quantum gates, but they will
be retrievable due to the unitarity of our gates. However, make a measurement, and
we will have destroyed that information; measurements are not unitary operations.
Weve covered the only two classical gates that have quantum alter egos. Lets go
on to meet some one-bit quantum gates that have no classical counterpart.
9.5.3
The Phase Flip, Z
If the bit flip, X, is the operator that coincides with the physical observable Sx ,
then the phase flip, Z, turns out to be the Pauli matrix z associated with the Sz
observable.
The Z operator, whose matrix is defined to be
1 0
,
0 1
negates the second amplitude of a state vector, leaving the first unchanged,
1 0
Z |i =
=
.
0 1
233
Notation and Vocabulary
Z is called a phase flip, because it changes (maximally) the relative phase of the two
amplitudes of |i. Why is multiplication by -1 a maximal phase change? Because
1
ei ,
so multiplying by this scalar is a = 180 rotation of in the complex plane. You
cant get more maximal than rotating something in the complex plane by 180 .
[Exercise. Graph and ei . If you are stuck, express in polar form and use
the Euler formula.]
Furthermore, if we modify only one of the two amplitudes in a states CBS expansion, thats a relative phase change, something we know matters (unlike absolute
phase changes, which are inconsequential multiples of the entire ket by a complex
unit).
[Exercise. Write out a summary of why a relative phase change matters. Hint:
Review the sections about time evolution of a quantum state from the third (optional)
time-dependent quantum mechanics lesson, particularly Traits #13 and #14.]
The use of the symbol, Z, to represent this operator is, as already observed, due
to its matrix being identical to z of the Sz observable.
Gate Symbol and Circuits
In a circuit diagram we use
Z
to express a Z gate. Its full effect on a general state in diagram form is
|0i + |1i
|0i |1i .
Computational Basis State Emphasis
It is uncommon to try to express Z in compact computational basis form; it leaves
|0i unchanged, and negates |1i, an operation that has no classical counterpart (in B,
1 = 1, so [1] = [1], and the operation would look like an identity). However, if
pressed to do so, we could use
|xi
(1)x |xi .
[Exercise. Show that this agrees with Zs action on a general |i by expanding |i
along the computational basis and using this formula while applying linearity.]
234
Measurement
It is interesting to note that Z has no measurement consequences on a single qubit
since the magnitude of both amplitudes remain unchanged by the gate. The circuit
|i
Z
produces the A-to-B transition
yielding the same probabilities, ||2 and ||2 , at both access points. Therefore, both
|i and Z |i will have identical measurement likelihoods; if we have 1000 electrons
or photons in spin state |i and 1000 in Z |i, a measurement of all of them will
throw about ||2 1000 into state |0i and ||2 1000 into state |1i.
However, two states that have the same measurement probabilities are not
necessarily the same state.
The relative phase difference between |i and Z |i can be felt the moment we
try to combine (incorporate into a larger expression) either state using superposition.
Mathematically, we can see this by noticing that
|i + |i
2
|i ,
i.e., we cannot create a new state from a single state, which is inherently linearly
dependent with itself, while a distinct normalized state can be formed by
|i + Z |i
2
=
=
( + ) |0i + ( ) |1i
2
2 |0i
= |0i .
2
Unless it so happens that = 1 and = 0, we get different results when using |i
and Z |i in the second position of these two legal superpositions, demonstrating that
they do, indeed, have different physical consequences.
9.5.4
The Bit-and-Phase Flip Operator, Y
After seeing the X and Z operators, we are compelled to wonder about the operator
Y that corresponds to the Pauli matrix y associated with the Sy observable. It is
just as important as those, but has no formal name (well give it one in a moment).
235
The Y operator is defined by the matrix
0 i
,
i 0
which has the following effect on a general ket:
0 i
Y |i =
= i
i 0
,
that last equality accomplished by multiplying the result by the innocuous unit scalar,
i.
Notation and Vocabulary
Although Y has no official name, we will call it the bit-and-phase flip, because it flips
both the bits and the relative phase, simultaneously.
The symbol Y is used to represent this operator because it is identical to the Pauli
matrix y .
Gate Symbol and Circuits
In a circuit diagram, we use
Y
to express a Y gate. Its full effect on a general state in diagram form is
|0i + |1i
i |0i + i |1i .
Computational Basis State Emphasis
No one expresses Y in computational basis form, but it can be done.
[Exercise. Find a formula that expresses Y s action on a CBS.]
Whats with the i? [Optional Reading]. Clearly, we can pull the i out
of the matrix (or the output ket) with impunity since we can always multiply any
normalized state by a unit scalar without changing the state. That makes us wonder
what its doing there in the first place. Superficially, the i makes Y = y , a desirable
completion to the pattern started with X = x and Z = z . But more deeply, by
attaching a factor of i to this operator, we can, with notational ease, produce a linear
combination of the three matrices using real coefficients, nx , ny and nz ,
0 1
0 i
1 0
nz
nx ny i
nx
+ ny
+ nz
=
.
1 0
i 0
0 1
nx + ny i
nz
236
This may seem less than a compelling justification, but the expression is of the syntactic form
,
n
is a real 3-D unit vector (nx , ny , nz )t ,
where, n
is a vector of operators
x
y ,
z
, represents the matrix for the observable Sn
and their formal dot product,
n
).
(the measurement of spin in the most general direction defined by a unit vector n
This will be developed and used in the next course CS 83B.
Measurement
As a combination of both bit flip (which swaps probabilities) and phase flip (which
does not change probabilities), Y has the same measurement consequence as a simple
QNOT:
|i
Y
This gives an A-to-B transition
i
which causes the probabilities, ||2 and ||2 to get swapped at access point B.
9.5.5
The Hadamard Gate, H
While all these quantum gates are essential, the one having the most far-reaching
consequences and personifying the essence of quantum logic is the Hadamard gate.
The Hadamard operator, H, is defined by the matrix
1 1 1
.
2 1 1
Traditionally, we first address its effect on the CBS states,
1 1 1
|0i + |1i
1
H |0i =
=
0
2 1 1
2
1 1 1
|0i |1i
0
H |1i =
=
.
1
2 1 1
2
237
and
We immediately recognize that this rotates the z-basis kets onto the x-basis kets,
H |0i
|0ix
H |1i
|1ix .
and
[Exercise. Prove that H is unitary.]
H of a General State
This is to be extended to an arbitrary state, |i, using the obvious rules. Rather than
approach it by expanding |i along the z-basis then extending linearly, its perhaps
faster to view everything in terms of matrices and column vectors,
1 1 1
1 +
H |i =
=
,
2 1 1
2
which can be grouped
H |i
|0i +
|1i
Gate Symbol and Circuits
In a circuit diagram, we use
H
to express an H gate. Its full effect on a general state in diagram form is
+
|0i
+
|1i .
|0i + |1i
H
2
2
Computational Basis State Emphasis
It turns out to be very useful to express H in compact computational basis form. Ill
give you the answer, and let you prove it for for yourself.
|xi
|0i + (1)x |1i
.
2
[Exercise. Show that this formula gives the right result on each of the two CBS
kets.]
238
Measurement
Compared to the previous gates, the Hadamard gate has a more complex and subtle
effect on measurement probabilities. The circuit
|i
H
gives an A-to-B transition
!
,
causing the output probabilities to go from ||2 and ||2 at point A to
P H |i & |0i
P H |i & |1i
| + |2
2
| |2
,
2
at point B.
Notation. The diagonal arrow, & , is to be read when measured, collapses to.
To make this concrete, here are the transition probabilities of a few input states
(details left as an exercise).
|i
|0i =
0
1/
2
|1ix =
1/ 2
i
.3
|0 i =
.7
1/2
|1 i =
3/2
Probabilities Before
P & |0i = 1
P & |1i = 0
P & |0i = 1/2
P & |1i = 1/2
P & |0i = .3
P & |1i = .7
P & |0i = 1/4
P & |1i = 3/4
[Exercise. Verify these probabilities.]
239
Probabilities After
P & |0i = 1/2
P & |1i = 1/2
P & |0i = 0
P & |1i = 1
P & |0i = 1/2
P & |1i = 1/2
P & |0i = 2 +4 3
P & |1i = 2 4 3
9.5.6
Phase-Shift Gates, S, T and R
The phase-flip gate, Z, is a special case of a more general (relative) phase shift
operation in which the coefficient of |1i is shifted by = radians. There are
two other common shift amounts, /2 (the S operator) and /4 (the T operator).
Beyond that we use the most general amount, any (the R operator). Of course,
they can all be defined in terms of R , so well define that one first.
The phase shift operator, R , is defined by the matrix
1 0
,
0 ei
where is any real number. It leaves the coefficient of |0i unchanged and shifts (or
rotates) |1is coefficient by a relative angle , whose meaning we will discuss in a
moment. Heres the effect on a general state:
1 0
R |i =
=
.
0 ei
ei
The operators S and T are defined in terms of R ,
1 0
S R/2 =
,
0 i
and
T
R/4
1
0
.
0 ei/4
Vocabulary
S is called the phase gate, and in an apparent naming error T is referred to as the
/8 gate, but this is not actually an error as much as it is a change of notation.
Like a state vector, any unitary operator on state space can be multiplied by a unit
scalar with impunity. Weve seen the reason, but to remind you, state vectors are
rays in state space, all vectors on the ray considered to be the same state. Since
we are also working on the projective sphere, any unit scalar not only represents the
same state, but keeps the vector on the projective sphere.
With that in mind, if we want to see a more balanced version of T , we multiply
it by ei/8 and get the equivalent operator,
i/8
e
0
T
.
=
0
ei/8
Measurement
These phase shift gates leave the probabilities alone for single qubit systems, just as
the phase flip gate, Z, did. You can parallel the exposition presented for Z in the
current situation to verify this.
240
9.6
Putting Unary Gates to Use
9.6.1
Basis Conversion
Every quantum gate is unitary, a fact that has two important consequences, the first
of which we have already noted.
1. Quantum gates are reversible; their adjoints can be used to undo their action.
2. Quantum gates map any orthonormal CBS to another orthonormal CBS.
The second item is true using logic from our linear algebra lesson as follows. Unitary
operators preserve inner products. For example, since the natural CBS { |0i , |1i }
which we can also express using the more general notation { |xi }1x=0 satisfies the
orthonormality relation
hx | yi
xy
(kronecker delta) ,
then the unitarity of U means we also have
x U U y
xy .
(Ill remind you that the LHS of last equation is nothing but the inner product of
U |yi with U |xi expressed in terms of the adjoint conversion rules introduced in our
lesson on quantum mechanics).
That tells us that { U |0i , U |1i } are orthonormal, and since the dimension of
H = 2 they also span the space. In other words, they form an orthonormal basis,
as claimed.
Lets call this the basis conversion property of unitary transformations and make
it a theorem.
Theorem (Basis Conversion Property). If U is a unitary operator
and A is an orthonormal basis, then U (A), i.e., the image of vectors A
under U , is another orthonormal basis, B.
[Exercise. Prove the theorem for any dimension, N . Hint: Let bk = U (ak )
be the kth vector produced by subjecting ak A to U . Review the inner productpreserving property of a unitary operator and apply that to any two vectors bk
and bj in the image of A. What does that say about the full set of vectors B =
U (A)? Finally, what do you know about the number of vectors in the basis for an
N -dimensional vector space?]
241
QNOT and Bases
When applied to some gates, like QNOT, this is a somewhat trivial observation since
QNOT maps the z-basis to itself:
X : |0i
X : |1i
7 |1i
7 |0i .
Hadamard and Bases
In other situations, this is a very useful and interesting conversion. For example, if
you look back at its effect on the CBS, the Hadamard gate takes the z-basis to the
x-basis:
H : |0i
H : |1i
7 |0ix
7 |1ix ,
and, since every quantum gates adjoint is its inverse, and H is self-adjoint (easy
[Exercise]), it works in the reverse direction as well,
H : |0ix
H : |1ix
7 |0i
7 |1i .
[Exercise. Identify the unitary operator that has the effect of converting between
the z-basis and the y-basis.]
9.6.2
Combining Gates
We can place two or more one-qubit gates in series to create desired results as the
next two sections will demonstrate.
An x-Basis QNOT
The QNOT gate (a.k.a. X), swaps the two CBS states, but only relative to the z-basis,
because thats how we defined it. An easy experiment shows that this is not true of
another basis, such as the x-basis:
|0i |1i
X |0i X |1i
X |1ix = X
=
2
2
|1i |0i
= |1ix
=
= |1ix .
2
[Exercise. To what is the final equality due? What is X |0ix ?]
242
[Exercise. Why is this not a surprise? Hint: Revive your quantum mechanics
knowledge. X is proportional to the matrix for the observable, Sx , whose eigenvectors
are |0ix and |1ix , by definition. What is an eigenvector?]
If we wanted to construct a gate, QN OTx , that does have the desired swapping
effect on the x-basis we could approach it in a number of ways, two of which are
1. brute force using linear algebra, and
2. a gate-combination using the basis-transforming power of the H-gate.
Ill outline the first approach, leaving the details to you, then show you the second
approach in its full glory.
Brute Force. We assert the desired behavior by declaring it to be true (and
confirming that our guess results in a unitary transformation). Here, that means
stating that QN OTx |0ix = |1ix and QN OTx |1ix = |0ix . Express this as two
equations involving the matrix for QN OTx and the x-basis kets in coordinate form
(everything in z-basis coordinates, of course). Youll get four simultaneous equations
for the four unknown matrix elements of QN OTx . This creates a definition in terms
of the natural CBS. We confirm its unitary and weve got our gate.
[Exercise. Fill in the details for QN OTx .]
Gate Combination. We know that QNOT (i.e., X) swaps the z-CBS, and H
converts between z-CBS and x-CBS. So we use H to map the x-basis to the z-basis,
apply X to the z-basis and convert the results back to the x-basis:
H
Lets confirm that our plan worked.
Note. Gates operate from left-to-right, but operator algebra moves from rightto-left. When translating a circuit into a product of matrices, we must reverse the
order. In this case, we cant see that effect, since the gate is symmetric, but it
wont always be, so stay alert.
Substituting the matrices for their gates (and reversing order), we get
1 1 1
1 1 1
0 1
QN OTx (H) (X) (H) =
1 0
2 1 1
2 1 1
=
1
2
1
1
1 1
1 1
1
1
1
2
2 0
0 2
=
1 0
,
0 1
and we have our matrix. Its easy to confirm that the matrix swaps the x-CBS kets
and is identical to the matrix we would get using brute force (Ill let you check that).
243
Circuit Identities
The above example has a nice side-effect. By comparing the result with one of our
basic gates, we find that H X H is equivalent to Z,
H
There are more circuit identities that can be generated, some by looking at the
matrices, and others by thinking about the effects of the constituent gates and confirming your guess through matrix multiplication.
Here is one you can verify,
H
and heres another,
Z
The last pattern is true for any quantum logic gate, U , which is self-adjoint because
then U 2 = U U = 1, the first equality by self-adjoint-ness and the second by
unitarity.
Some operator equivalences are not shown in gate form, but rather using the
algebraic operators. For example
XZ
i Y
or
XY Z
ZY X
i1.
Thats because the algebra shows a global phase factor which may appear awkward
in gate form yet is still important if the combination is to be used in a larger circuit.
As you may recall, even though a phase factor may not have observable consequences
on the state alone, if that state is combined with other states prior to measurement,
the global phase factor can turn into a relative phase difference, which does have
observable consequences.
I will finish by reminding you that the algebra and the circuit are read in opposite
order. Thus
XY Z
i1.
corresponds to the circuit diagram
Z
i1
This completes the basics of Qubits and their unary operators. There is one final
topic that every quantum computer scientist should know. It is not going to be used
much in this course, but will appear in CS 83B and CS 83C. It belongs in this chapter,
so consider it recommended, but not required.
244
9.7
9.7.1
The Bloch Sphere
Introduction
Our goal is to find a visual 3-D representation for the qubits in H. To that end, we
will briefly allude to the lecture on quantum mechanics.
If you studied the optional time evolution of a general spin state corresponding
to a special physical system an electron in constant magnetic field B you learned
that the expectation value of all three observables formed a real 3-D time-evolving
vector,
hSx i|(t)i
s(t) hSy i|(t)i .
hSz i|(t)i
which precesses around the direction of B.
Do Not Panic. You dont have to remember or even re-study that section. This
is merely a reference in case you want to connect that material with the following,
which is otherwise self-contained.
The time evolution started out in an initial spin state at time t = 0, and we
followed its development at later times. However, today, we need only consider the
fixed state at its initial time. No evolution is involved.
With that scary introduction, lets calmly look at any general states expansion
along the preferred CBS,
|i
9.7.2
c1 |0i + c2 |1i .
Rewriting |i
We start by finding a more informative representation of |i. First, express c1 and
c2 in polar form, giving the equivalent state
!
c ei1
.
|i =
s ei2
Then multiply by the unit scalar ei(
|i
1 +2
2
) to get a more balanced equivalent state,
i( 1 2 2 )
ce
.
1 2
s ei( 2 )
Now, simplify by making the substitution
1 2
2
245
to get a balanced Hilbert space representative of our qubit,
!
c ei
|i =
.
s ei
As a qubit, |i is normalized,
c2 +
s2
1,
so c and s can be equated with the sine and cosine of some angle which we call 2 , i.e.,
c
sin .
2
cos
and
Well see why we pick 2 , not , next.
9.7.3
The Expectation Vector for |i
The three logic gates, X, Y and Z, are represented by unitary matrices, but they also
happen to be Hermitian. This authorizes us to consider them observables defined by
those matrices. In fact, we already related them to the observables Sx , Sy and Sz ,
the spin measurements along the principal axes, notwithstanding the removal of the
factor ~2 . Therefore, each of these observables has an expectation value a prediction
about the average of many experiments in which we repeatedly send lots of qubits
in the same state, |i, through one of these gates, measure the output (causing a
collapse), compute the average for many trials, and do so for all three gates X, Y
and Z. We define a real 3-D vector, s, that collects the expectation value into one
column, to be
hXi|i
s hY i|i .
hZi|i
At the end of the quantum mechanics lecture we essentially computed these expectation values. If you like, go back and plug t = 0 into the formula there. Aside from
the factor of ~2 (caused by the difference between X/Y /Z and Sx /Sy /Sz ), you will get
sin cos
s = sin sin .
cos
By defining c and s in terms of 2 , we ended up with expectation values that had the
whole angle, , in them. This is a unit vector in R3 whose spherical coordinates are
(1, , )t , i.e., it has a polar angle and azimuthal angle . It is a point on the
unit sphere.
246
9.7.4
Definition of the Bloch Sphere
The sphere in R3 defined by
n
o
n |
n| = 1
=
is called the Bloch sphere when the coordinates of each point on the sphere n
(x, y, z)t are interpreted as the three expectation values hXi, hY i and hZi for some
on the Bloch
qubit state, |i. Each qubit value, |i, in H corresponds to a point n
sphere.
=
If we use spherical coordinates to represent points on the sphere, then n
t
(1, , ) corresponds to the |i = |0i + |1i in our Hilbert space H according to
!
1,
cos 2 ei
i
= ,
H.
n
Bloch sphere |i =
sin
e
2
Sph
Now we see that a polar angle, , of a point on the Bloch sphere gives the magnitudes
of its corresponding qubit coordinates, but not directly; when is the polar angle,
/2 is used (through sine and cosine) for the qubit coordinate magnitudes.
247
Chapter 10
Tensor Products
V W
10.1
Tensor Product for Quantum Computing
While quantum unary gates are far richer in variety and applicability than classical
unary gates, there is only so much fun one can have with circuit elements, like
U
which have a single input. We need to combine qubits (a process called quantum
entanglement), and to do that well need gates that have, at a minimum, two inputs,
U
(You may notice that there are also two outputs, an inevitable consequence of unitarity that well discuss in this hour.)
In order to feed two qubits into a binary quantum gate, we need a new tool to
help us calculate, and that tool is the tensor product of the two single qubit state
spaces
H H.
The concepts of tensors are no harder to master if we define the general tensor product
of any two vector spaces, V and W of dimensions l and m, respectively,
V W ,
and this approach will serve us well later in the course. We will then apply what we
learn by setting V = W = H.
248
The tensor product of more than two component spaces like
V0 V1 V2 ,
or more relevant to us,
H H H ,
presents no difficulty once we have mastered the order 2 tensors (product of just
two spaces). When we need a product of more than two spaces, as we shall in a future
lecture, Ill guide you. For now, lets learn what it means to form the tensor product
of just two vector spaces.
10.2
The Tensor Product of Two Vector Spaces
10.2.1
Definitions
Whenever we construct a new vector space like V W , we need to handle the required
equipment. That means
1. specifying the scalars and vectors,
2. defining vector addition and scalar multiplication, and
3. confirming all the required properties.
Items 1 and 2 are easy, and we wont be overly compulsive about item 3, so it
should not be too painful. Well also want to cover the two normally optional but
for us required topics,
4. defining the inner product and
5. establishing the preferred basis.
If you find this sort of abstraction drudgery, think about the fact that tensors are the
requisite pillars of many fields including structural engineering, particle physics and
general relativity. Your attention here will not go unrewarded.
Overview
The new vector space is based on the two vector spaces V (dimension = l) and W
(dimension = m) and is called tensor product of V and W , written V W . The new
space will turn out to have dimension = lm, the product of the two component space
dimensions).
249
The Scalars of V W
Both V and W must have a common scalar set in order to form their inner product,
and that set will be the scalar set for V W . For real vector spaces like R2 and R3 ,
the scalars for
R2 R3
would then be R. For the Hilbert spaces of quantum physics (and quantum computing), the scalars are C.
The Vectors in V W
Vectors of the tensor product space are formed in two stages. I like to Compartmental -ize them as follows:
1. We start by populating V W with the formal symbols
v w
consisting of one vector v from V and another vector w from W . v w is called
the tensor product of the two vectors v and w. There is no way to further merge
these two vectors; v w is as concise as their tensor product gets.
For example, in the case of (5, 6)t R2 and (, 0, 3)t R3 , they provide the
tensor product vector
5
0 R2 R3 ,
6
3
with no further simplification possible (notwithstanding the natural basis coordinate representation that we will get to, below).
Caution. No one is saying that all these formal products vw are distinct from
one another. When we define the operations, well see there is much duplication.
2. The formal vector products constructed in step 1 produce only a small subset
of the tensor product space V W . The most general vector is a finite sum of
such symbols.
The full space V W consists of all finite sums of the form
X
vk wk , with
k
vk V and wk W.
250
For example, in the case of the complex vector spaces V = C3 and W = C4 ,
one such typical vector in the product space C3 C4 would be
i
1
1i
1+i
1i
2
2
i
6 + 6
2 .
3
0 +
7
3i
3i
2
4
0
i
Although this particular sum cant be further simplified as a combination of
separable tensors, we can always simplify long sums so that there are at most
lm terms in them. Thats because (as well learn) there are lm basis vectors for
the tensor product space.
Second Caution. Although these sums produce all the vectors in V W ,
they do so many times over. In other words, it will not be true that every sum
created in this step is a distinct tensor from every other sum.
Vocabulary
Product Space. The tensor product of two vector spaces is sometimes referred
to as the product space.
Tensors. Vectors in the product space are sometimes called tensors, emphasizing that they live in the tensor product space of two vector spaces. However,
they are still vectors.
Separable Tensors. Those vectors in V W which arise in the step 1 are
called separable tensors; they can be separated into two component vectors,
one from each space, whose product is the (separable) tensor. Step 2 presages
that most tensors in the product space are not separable. Separable tensors are
sometimes called pure or simple.
Tensor Product. We can use the term tensor product to mean either the
product space or the individual separable tensors. Thus V W is the tensor
product of two spaces, while v w is the (separable) tensor product of two
vectors.
Vector Addition
This operation is built-into the definition of a tensor; since the general tensor is the
sum of separable tensors, adding two of them merely produces another sum, which
is automatically a tensor. The twist, if we can call it that, is how we equate those
sums which actually represent the same tensor. This all expressed in the following
two bullets.
251
For any two tensors,
=
0 =
v k wk
vj0 wj0 ,
their vector sum is the combined sum, i.e.,
X
X
+ 0
v k wk +
vj0 wj0 ,
j
which simply expresses the fact that a sum of two finite sums is itself a finite
sum and therefore agrees with our original definition of a vector object in the
product space. The sum may need simplification, but it is a valid object in the
product space.
The tensor product distributes over sums in the component space,
(v + v0 ) w = v w + v0 w
v (w + w0 ) = v w + v w0 .
and
Practically, we only need to understand this as a distributive property, but
in theoretical terms it has the effect of producing countless sets of equivalent
vectors. That is, it tells how different formal sums in step 2 of The Vectors
of V W might represent the same actual tensor.
Commutativity. Vector addition commutes by definition; the formal sums in step
2 are declared to be order-independent.
Scalar Multiplication
Let c be a scalar. We define its product with a tensor in two steps.
c Times a Separable Tensor.
c(v w)
(cv) w
v (cw) .
This definition requires declares that it doesnt matter which component
of the separable tensor gets the c. Either way, the two tensors formed are the
same. It has the effect of establishing the equivalence of initially distinct formal
symbols v w in step 1 of The Vectors of V W .
c Times a Sum of (Separable) Tensors.
c(v0 w0 + v1 w1 )
c(v0 w0 ) + c(v1 w1 ) .
Because any tensor can be written as the finite sum of separable tensors, this
requirement covers the balance of the product space. You can distribute c over
as large a sum as you like.
Order. While we can place c on either the left or right of a vector, it is usually
placed on the left, as in ordinary vector spaces.
252
The Requisite Properties of V W
Properties like commutative addition, distributivity of scalar multiplication are automatic consequences of the above definitions. Ill leave it as an optional exercise.
[Exercise. Prove that V W satisfies the various operational requirements of a
vector space.]
Inner Products in V W
If V and W possess inner products, an inner product is conferred to the product
space V W by
hv w | v0 w0 i
hv | v0 i hw | w0 i ,
where on the RHS is scalar multiplication. This only defines the tensor product of
two separable tensors, however we extend this to all tensors by asserting a distributive
property (or if you prefer the terminology, by extending linearly). For example,
h v w | v0 w0 + v00 w00 i
hv w | v0 w0 i + hv w | v00 w00 i .
[Exercise. Compute the inner product
*
5
0
6
3
in R2 R3
+
1
0
2
1
3
[Exercise. Compute the inner product in C3 C4
1
i
1
* 1 + i
+
i
1
2
i
6 + 6 0
3
0
1
3i
3i
1
4
1
i
[Exercise. Prove that the definition of inner product satisfies all the usual requirements or a dot or inner product. Be sure to cover distributivity and positive
definiteness.]
The Natural Basis for V W
Of all the aspects of tensor products, the one that we will use most frequently is the
preferred basis.
Tensor Product Basis Theorem (Orthonormal Version). If V has
dimension l with orthonormal basis
n
ol1
vk
,
k=0
253
and W has dimension m, with orthonormal basis
n
om1
wj
,
j=0
then V W has dimension lm and inherits a natural (preferred) orthonormal basis
n
ol1, m1
vk w j
j, k = 0, 0
from V and W . For example, the natural basis for R2 R3 is
1
0
0
1
1
1
0 ,
1 ,
0 ,
0
0
0
0
0
1
1
0
0
0
0
0
0 ,
1 ,
0
1
1
1
0
0
1
Proof of Basis Theorem. Ill guide you through the proof, and you can fill in
the gaps as an exercise if you care to.
Spanning. A basis must span the space. We need to show that any tensor can
be expressed as a linear combination of the alleged basis vectors vk wj . This is an
easy two parter:
1. Any v V can be expanded along the V - basis as
X
v =
k v k ,
and any w W can be expanded along the W - basis as
X
w =
j wj ,
which implies that any separable tensor v w can be expressed
X
vw =
k j (vk wj ) .
[Exercise. Prove the last identity by applying linearity (distributive properties
of the tensor product space).]
2. Any tensor is a sum of separable tensors, so item 1 tells us that it, too, can
be expressed as a linear combination of vk wj . [Exercise. Demonstrate this
algebraically.]
254
Linear Independence and Orthonormality. We rely on a little theorem to
which I subjected you twice, first in the linear algebra lecture, then again in the
Hilbert space lecture. It said that an orthonormality linearly independence. We
now show that the set {vk wj } is orthonormal collection of vectors.
hvk wj | vk0 wj 0 i
=
=
hvk | vk0 i hwj | wj 0 i
(
1 , if k = k 0 and j = j 0
0 , otherwise
This is the definition of orthonormality, so by our little theorem the vectors in
{vk wj } are linearly independent.
QED
The basis theorem works even if the component bases are not orthonormal. Of
course, in that case, the inherited tensor basis is not orthonormal.
Tensor Product Basis Theorem (General Version). If V has dimension l with basis
n
ol1
vk
,
k=0
and W has dimension m, with basis
om1
n
wj
,
j=0
then V W has dimension lm and inherits a natural (preferred) basis
n
v k wj
ol1, m1
j, k = 0, 0
from V and W .
If V and W each have an inner product (true for any of our vector spaces) this
theorem follows immediately from the orthonormal basis theorem.
[Exercise. Give a one sentence proof of this theorem based on the orthonormal
product basis theorem.]
The theorem is still true, even if the two spaces dont have inner products, but
we wont bother with that version.
Practical Summary of the Tensor Product Space
While we have outlined a rigorous construction for a tensor product space, it is usually
good enough for computer scientists to characterize the produce space in terms of the
tensor basis.
255
The produce space, U = V W consists of tensors, u, expressible as sums of
the separable basis
n
o
vk w j
j = 0, . . . , (n 1), k = 0, . . . , (m 1) ,
weighted by scalar weights ckj ,
(n1)(m1)
ckj (vk wj ) .
k=0
j=0
The nm basis tensors, vk wj , are induced by the two component bases,
V
vk
n1
wj
m1
and
k=0
j=0
The sums, products and equivalence of tensor expressions are defined by the required
distributive and commutative properties listed earlier, but can often be taken as the
natural rules one would expect.
Conventional Order of Tensor Basis
While not universal, when we need to list the tensor basis linearly, the most common
convention is to let the left basis index increment slowly and the right increment
quickly. It is V -major / W -minor format if you will, an echo of the row-major
(column-minor ) ordering choice of arrays in computer science,
n
v0 w0 , v0 w1 , v0 w2 , . . . , v0 wm1 ,
v1 w0 , v1 w1 , v1 w2 , . . . , v1 wm1 ,
..
.
o
vl1 w0 , vl1 w1 , vl1 w2 , vl1 wm1 .
You might even see these basis tensors labeled using the shorthand like kj ,
n
00 , 01 , 02 , . . . , 0(m1) ,
10 , 11 , 12 , . . . , 1(m1) ,
..
.
o
(l1)0 , (l1)1 , (l1)2 , . . . , (l1)(m1) .
256
10.2.2
Tensor Coordinates from Component-Space Coordinates
We have defined everything relating to a tensor product space V W . Before we
apply that to qubits, we must make sure we can quickly write down coordinates of
a tensor in the preferred (natural) basis. This starts, as always, with (i) separable
tensors, from which we move to the (ii) basis tensors, and finally graduate to (iii)
general tensors.
Natural Coordinates of Separable Tensors
We are looking for the coordinates of a pure tensor, expressible as a product of two
component vectors whose preferred coordinates we already know,
c0
d0
c1 d1
v w = .. .. .
. .
cl1
dm1
[A Reminder. We are numbering staring with 0, rather than 1, now that we are in
computing lessons.]
Im going to give you the answer immediately, and allow you to skip the explanation if you are in a hurry.
c0 d 0
c0 d 1
d0
..
.
1
c0 ..
c0 dm1
.
cd
1 0
dm1
cd
1 1
..
.
d
0
d
c1 dm1
1
c0
d0
c1 .
c1 d 1
.
c
d
2
0
.. .. =
c2 d 1 .
dm1
. .
..
.
cl1
dm1
cd
..
2 m1
.
d
0
d1
c
c d
l1 ..
l1 0
.
c d
l1 1
dm1
..
.
cl1 dm1
257
Example. For (5, 6)t R2 and (, 0, 3)t R3 , their tensor product in R2 R3
has natural basis coordinates given by
5
0
15
5
0 =
6 .
6
3
0
18
[Exercise. Give the coordinate representation of the tensor
1
1+i
2
6
3
3i
4
in the natural tensor basis.]
Explanation. Ill demonstrate the validity of the formula in the special case
R R3 . The derivation will work for any V and W (as the upcoming exercise
shows).
2
Consider the tensor product of two general vectors,
d0
c0
d1 .
c1
d2
Lets zoom in on the meaning of each column vector which is, after all, shorthand
notation for the following,
1
0
0
1
0
c0
+ c1
d0 0 + d1 1 + d2 0 .
0
1
0
0
1
Now apply the linearity to equate the above with
1
0
1
1
c0 d0
0
+ c0 d1
1
0
0
0
0
1
0
0
0
+ c1 d0
0 + c1 d1
1
1
1
0
0
0
1
c0 d 2
0
0
1
0
0
c1 d 2
0 .
1
1
Next, identify each of the basis tensor products with their symbolic vk (for R2 ) and
wj (for R3 ) to see it more clearly,
c0 d0 (v0 w0 )
+ c1 d0 (v1 w0 )
+
+
c0 d1 (v0 w1 )
c1 d1 (v1 w1 )
258
+
+
c0 d2 (v0 w2 )
c1 d2 (v1 w2 ) .
The basis tensors are listed in the conventional (V -major / W -minor format) allowing us to write down the coordinates of the tensor,
c0 d 0
c0 d 1
c0 d 2
c1 d 0 ,
c1 d 1
c1 d 2
as claimed.
QED
[Exercise. Replicate the demonstration for the coordinates of a separable tensor
v w in any product space V W.]
Natural Coordinates of Basis Tensors
By definition and a previous exercise (see the lecture on linear transformations) we
know that each basis vector, bk , when expressed in its own B-coordinates looks exactly
like the kth preferred basis element, i.e.,
0
..
.
bk
= 1 kth element .
.
B
..
0
The is true for our tensor basis, which means that once
have lm basis vectors
1
0
0
..
..
..
.
.
.
0 , , 1 , , 0
.
.
.
..
..
..
0
0
1
we embrace that basis, we
lm ,
which dont make reference to the two component vector spaces or the inherited V major / W -minor ordering we decided to use. However, for this to be useful, we need
an implied correspondence between these vectors and the inherited basis
n
v0 w0 , v0 w1 , v0 w2 , . . . , v0 wm1 ,
v1 w0 , v1 w1 , v1 w2 , . . . , v1 wm1 ,
...
o
vl1 w0 , vl1 w1 , vl1 w2 , vl1 wm1 .
259
This is all fine, as long as we remember that those self-referential tensor bases assume some agreed-upon ordering system, and in this course that will be the V -major
ordering.
To illustrate this, say we are working with the basis vector
0
0
0
R2 R3 .
0
1
0
In the rare times when we need to relate this back to our original vector spaces we
would count: The 1 is in position 4 (counting from 0), and relative to R2 and R3 this
means
4
1 3 + 1,
yielding the individual basis vectors #1 in V and #1 in W . Therefore, the correspondence is
0
0
0
0
0
v1 w1 =
1
0
1
0
1
0
which we can confirm by multiplying out the RHS as we learned in the last section
0
0
1
1
0
00
0 1
0 0
1 0
1 1
10
0
0
0
.
0
1
0
This will be easy enough with our 2-dimensional component spaces H, but if you
arent prepared for it, you might find yourself drifting aimlessly when faced with a
long column basis tensor and dont know what to do with it.
260
Natural Coordinates of General Tensors
Because the tensor space has dimension lm, we know that any tensor in the natural
basis will look like
0
1
.. .
.
lm2
lm1
The only thing worth noting here is the correspondence between this and the component spaces V and W . If we are lucky enough to have a separable tensor in our
hands, this would have the special form
c0 d 0
c0 d 1
c0 d 2
..
.
c1 d 0
c1 d 1
c1 d 2 ,
.
..
c d
2 0
c d
2 1
c d
2 2
..
.
and we might be able to figure out the component vectors from this. However, in
general, we dont have separable tensors. All we can say is that this tensor is a linear
combination of the lm basis vectors, and just accept that it has the somewhat random
components, i which we might label simply
0
1
.. ,
.
lm2
lm1
261
or we might label with an eye on our component spaces
00
01
02
..
.
0(m1)
10
11
12
.
. ,
.
1(m1)
20
21
22
.
..
2(m1)
..
.
with the awareness that these components kj may not be products of two factors
ck dj originating in two vectors (c0 , . . . , cl1 )t and (d0 , . . . , dm1 )t .
One thing we do know: Any tensor can be written as a weighted-sum of, at most,
lm separable tensors. (If this is not immediately obvious, please review the tensor
product basis theorem.)
Example. Lets compute the coordinates of the non-separable tensor
1
0
3
1
=
2
+
1
6
0
1
in the natural basis. We
component and adding,
1
3 2
= +
6 2
do it by applying the above formula to each separable
0
1 1
0 1
1
3
0
6
1
+ 1
6
0
12
0
6
0
3
7
1 + 3
6 .
12
6
Tensors as Matrices
If the two component spaces have dimension l and m, we know that the product space
has dimension lm. More recently weve been talking about how these lm coordinates
262
might be organized. Separable or not, a tensor is completely determined by its lm
preferred coefficients, which suggests a rectangle of numbers,
00
01
02
0(m1)
10
11
12
1(m1)
..
,
..
..
..
.
.
.
.
.
.
.
(l1)0 (l1)1 (l1)2 (l1)(m1)
which happens to have a more organized structure in the separable case,
c0 d0
c0 d 1
c0 d2 c0 dm1
c1 d0
c1 d 1
c1 d2 c1 dm1
..
.
..
..
..
...
.
.
.
.
cl1 d0 cl1 d1 cl1 d2 cl1 dm1
This matrix is not to be interpreted as a linear transformation of either component
space it is just a vector in the product space. (It does have a meaning as scalarvalued function, but well leave that as a topic for courses in relativity, particle physics
or structural engineering.)
Sometimes the lm column vector model serves tensor imagery best, while other
times the l m matrix model works better. Its good to be ready to use either one
as the situation demands.
10.3
Linear Operators on the Tensor Product Space
The product space is a vector space, and as such supports linear transformations.
Everything we know about them applies here: they have matrix representations in
the natural basis, some are unitary, some not, some are Hermitian, some not, some
have inverses, some dont, etc. The only special and new topic that confronts us is
how the linear transformation of the two component spaces, V and W , inform those
of the product space, V W .
The situation will feel familiar. Just as tensors in V W fall into two main
classes, those that are separable (products, v w, of two vectors) and those that
arent (they are built from sums of separable tensors), the operators on V W have
a corresponding breakdown.
10.3.1
Separable Operators
If A is a linear transformation (a.k.a. an operator) on V
A : V V ,
and B is a linear transformation on W
B : W W ,
263
then their tensor product, A B is a linear transformation on the product space
V W,
A B : V W V W ,
defined by its action on the separable tensors
[A B](v w)
Av Bw ,
and extended to general tensors linearly.
Note 1: We could have defined A B first on just the lm basis vectors vk wj ,
since they span the space. However, its so useful to remember that we can use this
formula on any two component vectors v and w, that I prefer to make this the official
definition.
Note 2: A and/or B need not map their respective vector spaces into themselves.
For example, perhaps A : V 7 V 0 and B : W 7 W 0 . Then AB : V W 7 V 0 W 0 .
However, we will usually encounter the simpler case covered by the definition above.
One must verify that this results in a linear transformation (operator) on the
product space by proving that for an , V W and scalar c,
[A B] ( + )
[A B] + [A B]
[A B] (c)
c [A B] .
and
This is very easy to do as an ...
[Exercise. Prove this by first verifiying it on separable tensors then showing that
the extension to general tensors preserves the properties.]
Example. Let A be defined on R2 by
v
Av = A 0
v1
v0 + 2v1
v1
and B be defined on R3 by
Bw
w0
w1 .
w2
On separable tensors, then, A B has the effect expressed by
w0
v0 + 2v1
[A B](v w) =
w1 ,
v1
w2
and this is extended linearly to general tensors. To get specific, we apply A B to
the tensor
1
0
3
1
=
2
+
1
6
0
1
264
to get
[A B]
0
3 + (12)
1+0
2
+
6
0
2
0
9
1
6
0
2
with no simplification obvious. However, an equivalent expression can be formed by
extracting a factor of :
0
1
9
1
[A B] =
2
+
1 .
6
0
1
We can always forsake the separable components and instead express this as a column
vector in the product space by adding the two separable tensors,
0
9
1
[A B] =
2 +
6
0
2
9
0
9
18
17
2
9
9 2
=
+ 0 = 6 .
6
12
0
12
6 2
0
6 2
10.3.2
The Matrix of a Separable Operator
We can write down the matrix for any separable operator using a method very similar
to that used to produce the coordinate representation of a separable tensor product of
vectors. As you recall, we used a V -major format which multiplied the entire vector
w by each coordinate of v to build the result. The same will work here. We use an
A-major format, i.e., the left operators coordinates change more slowly than the
right operators. Stated differently, we multiply each element of the left matrix by
265
the entire matrix on
a00
a10
..
.
the right.
a01 a0(l1)
b00 b01
a11 a1(l1) b10 b11
.. . .
..
..
..
.
.
.
.
.
b00 b01
b00
a b10 b11 a b10
01
00
..
.. . .
..
.
.
.
b00 b01
b00
b10
a10 b10 b11
a11
..
.. . .
..
.
.
.
.
..
.
b0(m1)
b1(m1)
..
..
.
.
b01
b11
.. . .
b01
b11
.. . .
.
.
..
..
.
.
This works based on a V -major column format for the vectors in the product space.
If we had used a W -major column format, then we would have had to define the
product matrix using a B-major rule rather than the A-major rule given above.
Example. The matrices for the A and B of our last example are given by
0 0
1 2
A =
and B = 0 0 ,
0 1
0 0
so the tensor product transformation A B is immediately written down as
0 0
2 0 0
0 0
0 2 0
0 0
0 0 2
.
AB =
0 0 0
0 0
0 0 0
0 0
0 0 0
0 0
Example. Weve already applied the definition of a tensor operator directly to
compute the above operator applied to the tensor
3
7
1
0
1 + 3
3
1
=
2
+
1
=
6
0
6
1
12
6
266
and found that, after reducing the result to a column vector,
9
17
9 2
[A B] =
12
6 2
It now behooves us to do a sanity check. We must confirm that multiplication of
by the imputed matrix for A B will produce the same result.
0 0
2 0 0
3
0 0
0 2 0
0 0
7
0
0
2
1 + 3
[A B] =
6 = ?
0 0 0
0 0
12
0 0 0
0 0
6
0 0 0
0 0
[Exercise. Do the matrix multiplication, fill in the question mark, and see if it
agrees with the column tensor above.]
10.3.3
The Matrix of a General Operator
As with the vectors in the product space (i.e., tensors), we can represent any operator
on the product space as sum of no more than (lm)2 separable operators, since there
are that many elements in a tensor product matrix. The following exercise will make
this clear.
[Exercise. Let the dimensions of our two component spaces be l and m. Then
pick two integers p and q in the range 0 p, q < lm. Show that the matrix
Ppq
which has a 1 in position (p, q) and 0 in all other positions, is separable. Hint: You
need to find an l l matrix and an m m matrix whose tensor product has 1 in the
right position and 0s everywhere else. Start by partitioning Ppq into sub-matrices of
size m m. Which sub-matrix does the lonely 1 fall into? Where in that m m
sub-matrix does that lonely 1 fall? ]
[Exercise. Show that the set of all (lm)2 matrices
n
o
Ppq 0 p, q < lm ,
spans the set of linear transformations on V W .]
267
10.3.4
Food for Thought
Before we move on to multi-qubit systems, here are a few more things for you may
wish to ponder.
[Exercise. Is the product of unitary operators unitary in the product space?]
[Exercise. Is the product of Hermitian operators Hermitian in the product
space?]
[Exercise. Is the product of invertible operators invertible in the product space?]
268
Chapter 11
Two Qubits and Binary Quantum
Gates
|i2
|0iA|0iB + |0iA |1iB
+ |1iA |0iB + |1iA |1iB
11.1
The Jump from One to Two
Classical Binary Systems
Elevating a classical one-bit system to a classical two-bit system doesnt seem to
require any fancy math. We simply slap together two single bits and start defining
binary logic gates like AND and OR which take both bits as input. That works
because classical bits cannot become entangled the way in which quantum bits can.
Nevertheless, we could use tensor products to define the classical two-bit system if
we were inclined to see the formal definition. As with the one-bit systems, it would
be rather dull and its only purpose would be to allow a fair comparison between two
classical bits and two quantum bits (the latter requiring the tensor product). I wont
bother with the formal treatment of two classical bits, but leave that as an exercise
after you have seen how we do it in the quantum case.
Quantum Binary Systems
Tensor algebra allows us to make the leap from the 2-D Hilbert space of one qubit to
a 4-D Hilbert space of two qubits. Once mastered, advancing to n qubits for n > 2
is straightforward. Therefore, we move carefully through the n = 2 case, where the
concepts needed for higher dimensions are easiest to grasp.
269
11.2
The State Space for Two Qubits
11.2.1
Definition of a Two Quantum Bit (Bipartite) System
The reason that we go tensor for a two-qubit system is that the two bits may
become entangled (to be defined below). That forces us to treat two bits as if they
were a single state of a larger state space rather than keep them separate.
Definition of Two Qubits. A Two-qubit system is (any copy of ) the
entire product space H H.
Definition of a Two-Qubit Value. The value or state of a twoqubit system is any unit (or normalized) vector in H H.
In other words, two qubits form a single entity the tensor product space H H
whose value can be any vector (which happens also to be a tensor) on the projective
sphere of that product space.
The two-qubit entity itself is not committed to any particular value until we say
which specific unit-vector in H H we are assigning it.
Vocabulary and Notation
Two qubits are often referred to as a bipartite system. This term is inherited from
physics in which a composite system of two identical particles (thus bi-parti -te) can
be entangled.
To distinguish the two otherwise identical component Hilbert spaces I may use
subscripts, A for the left-space and B for the right space,
HA HB .
Another notation you might see emphasizes the order of the tensor product, that is,
the number of component spaces in our current case, two,
H(2) .
In this lesson, we are concerned with order-2 products, with a brief but important
section on order-3 products at the very end.
Finally, note that in the lesson on tensor products, we used the common abstract
names V and W for our two component spaces. In quantum computation the component spaces are usually called A and B. For example, whereas in that lecture I
talked about V -major ordering for the tensor coordinates, Ill now refer to A-major
ordering.
270
11.2.2
The Preferred Bipartite CBS
First and foremost, we need to establish symbolism for the computation basis states
(CBS ) of our product space. These states correspond to the two-bits of classical
computing, and they allow us to think of two ordinary bits as being embedded within
the rich continuum of a quantum bipartite state space.
Symbols for Basis Vectors
The tensor product of two 2-D vector spaces has dimension 2 2 = 4. Its inherited
preferred basis vectors are the separable products of the component space vectors,
{ |0i |0i , |0i |1i , |1i |0i , |1i |1i } .
These are the CBS of the bipartite system. There are some shorthand alternatives in
quantum computing.
|0i |0i |0i |0i |00i
|0i2
|0i |1i |0i |1i |01i
|1i2
|1i |0i |1i |0i |10i
|2i2
|1i |1i |1i |1i |11i
|3i2
All three of the alternatives that lack the symbol are seen frequently in computer
science, and we will switch between them freely based on the emphasis that the
context requires.
The notation of the first two columns admits the possibility of labeling each of
the component kets with the H from whence it came, A or B, as in
|0iA |0iB |0iA |0iB
|0iA |1iB |0iA |1iB
etc.
I will often omit the subscripts A and B when the context is clear and include them
when I want to emphasize which of the two component spaces the vectors comes from.
The labels are always expendable since the A-space ket is the one on the left and the
B-space ket is the one on the right. I will even include and/or omit them in the same
string of equalities, since it may be clear in certain expressions, but less so in others:
U
|0iA + |1iA |1iB
= U |0i |1i + |1i |1i = |0i |0i + |0i |1i
= |0iA |0iB + |1iB
(This is an equation we will develop later in this lesson.)
271
The densest of the notations in the stack a couple paragraphs back is
the encoded version which expresses the ket as an integer from 0 to 3. We should
reinforce this correspondence at the outset and add the coordinate representation of
each basis ket under the implied A-major ordering of the vectors suggested by their
presentation, above.
basis ket
encoded
coordinates
|0i |0i
|0i2
1
0
0
|0i |1i
|1i2
0
1
0
|1i |0i
|2i2
0
0
1
|1i |1i
|3i2
0
0
0
This table introduces an exponent-like notation, |xi2 , which is needed mainly in the
encoded form, since an integer representation for a CBS does not disclose its tensor
order (2 in this case) to the reader, while the other representations clearly imply that
we are looking at two-qubits.
11.2.3
Separable Bipartite States
The CBS are four special separable tensors. There is an infinity of other separable
tensors in H H of the form
|i |i |i |i .
Note that, unlike the CBS symbolism, there is no further alternative notation for a
general separable tensor. In particular, |i makes no sense.
11.2.4
Alternate Bipartite Bases
Inherited Second Order x-CBS
We can construct other computational bases like the inherited x-basis,
{ |0ix |0ix , |0ix |1ix , |1ix |0ix , |1ix |1ix } ,
or, using the common alternate notation the x-basis, as
{ |+i |+i , |+i |i , |i |+i , |i |i } .
You might see it condensed, as
{ |++i , |+i , |+i , |i } ,
272
or super condensed (my own notation),
2
|0i , |1i2 , |2i2 , |3i2 .
In this last version, Im using the subscript to indicate x basis and Im encoding
the ket labels into decimal integers, 0 through 3, for the four CBS states.
Inherited Second Order Mixed CBS
While rare, we could inherit from the z-basis for our A-space, and the x-basis for our
B-space, to create the hybrid
{ |0i |0ix , |0i |1ix , |1i |0ix , |1i |1ix } .
[Exercise. How do we know that this is an orthonormal basis for the product space?]
Notation
Whenever we choose to label a state as a CBS of a non-standard basis, we leave no
room to label the component spaces, |0iA |1iB , which then must be inferred from their
position (A left and B right).
[Exception. If we use the alternate notation |+i = |0ix and |i = |1ix , then we
once again have room for the labels A and B: |+iA |+iB , |iA |+iB , etc.]
No matter what basis we use, if pressed to show both the individual basis of interest
and label the component state spaces, we could express everything in the z-basis and
thus avoid the subscript conflict, entirely, as with
|0iB + |1iB
.
|0i |0ix = |0iA
2
The basis we use as input(s) to the quantum circuit and along which we measure
the output of the same circuit is what we mean when we speak of the computational
basis. By convention, this is the z-basis. It corresponds to the classical bits
|0i |0i [00]
|0i |1i [01]
|1i |0i [10]
|1i |1i [11]
Nevertheless, in principle there is nothing preventing us from having hardware that
measures output along a different orthonormal basis.
[Future Note. In fact, in the later courses, CS 83B and 83C, we will be considering measurements which are not only made with reference to a different observable
(than Sz ) but are not even bases: they are neither orthonormal nor linearly independent. These go by the names general measurement operators or positive operator
valued measures, and are the theoretical foundation for encryption and error correction.]
273
Example
We expand the separable state
|0ix |1iy
along the (usual) computational basis in H H.
|0i + |1i
|0i i |1i
|0ix |1iy =
2
2
|00i i |01i + |10i i |11i
.
2
As a sanity check we compute the modulus-squared of the product directly, i.e., by
summing the magnitude-squared of the four amplitudes,
2
1
1
i
i
1
1
i
i
=
+
+
+
|0ix |1iy
2
2
2
2
2
2
2
2
=
1,
as it darn well had better be.
This is the kind of test you can perform in the midst of a large computation to
be sure you havent made an arithmetic error.
We might have computed the modulus-squared of the tensor product using one
of two other techniques, and it wont hurt to confirm that we get the same 1 as an
answer. The techniques are
the definition of inner product in the tensor space (the product of component
inner-products),
2
= h |0ix |1iy |0ix |1iy i
|0ix |1iy
=
h0 | 0ix
h1 | 1iy
11
X,
or the adjoint conversion rules to form the left bra for the inner product,
2
=
h1|
h0|
|0i
|1i
=
h1|
h0
|
0i
|1iy
|0ix |1iy
y
x
x
y
y
x
x
=
h1 | 1iy
X .
However, neither of these would have been as thorough a check of our arithmetic as
the first approach.
[Exercise. Expand the separable state
.1 |0i + i .9 |1i
i .7 |0i + .3 |1i
along the computational basis in H H. Confirm that it is a unit tensor.]
274
How the Second Order x-Basis Looks when Expressed in the Natural Basis
If we combine the separable expression of the x-CBS kets,
|+i |+i , |+i |i , |i |+i ,
and
|i |i .
with the expansion of each component ket along the natural basis,
|+i
H |0i
|0i + |1i
|i
H |1i
|0i |1i
,
2
and
the four x-kets look like this, when expanded along the natural basis:
|+i |+i
|0i |0i + |0i |1i + |1i |0i + |1i |1i
,
2
|+i |i
|0i |0i |0i |1i + |1i |0i |1i |1i
,
2
|i |+i
|0i |0i + |0i |1i |1i |0i |1i |1i
,
2
|i |i
|0i |0i |0i |1i |1i |0i + |1i |1i
.
2
Notice something here that we will use in a future lecture:
When expanded along the z-basis, the x-basis kets have equal numbers of
+ and terms except for the zeroth CBS ket, |00ix , whose coordinates
are all +1.
I know it sounds silly when you say it out loud, but believe me, it will be very useful.
11.2.5
Non-Separable Bipartite Tensors
The typical tensor in a two-qubit system is a normalized finite sum of separable
tensors thats how we started defining product tensors. While such a tensor is not
necessarily separable, it can at least be expressed as a superposition of the four CBS
kets,
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i .
The exponent 2 on the LHS is, as mentioned earlier, a clue to the reader that |i
lives in a second-order tensor product space, a detail that might not be clear without
looking at the RHS. In particular, nothing is being squared.
275
11.2.6
Usual Definition of Two Qubits and their Values
This brings us to the common definition of a two-qubit system, which avoids the
above formalism.
Alternative Definition of Two Qubits. Two qubits are represented by a variable superposition of the four tensor basis vectors of HH,
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i
where the complex scalars satisfy
||2 + ||2 + ||2 + ||2
1.
Using our alternate notation, we can write this superposition either as
|i2
|00i + |01i + |10i + |11i
|i2
|0i2 + |1i2 + |2i2 + |3i2 .
or
We may also use alternate notation for scalars, especially when we prepare for higherorder product spaces:
|i2
c00 |00i + c01 |01i + c10 |10i + c11 |11i
or
|i2
c0 |0i2 + c1 |1i2 + c2 |2i2 + c3 |3i2 .
11.3
Fundamental Two-Qubit Logic Gates
11.3.1
Binary Quantum Operators
Definition and Terminology
A binary quantum operator is a unitary transformation, U , on the
two-qubit system H H.
As you can see, with all the vocabulary and skills we have mastered, definitions
can be very short now and still have significant content. For instance, we already
know that some binary quantum operators will be separable and others will not. The
simplest and most common gate, in fact, is not separable.
Binary quantum operators also go by the names two-qubit gates, binary qubit
operators, bipartite operators and various combinations of these.
276
Complete Description of Binary Quantum Operators
When we study specific binary qubit gates, we will do five things:
:
Show the symbol for that gate,
|xi |yi : define the gate on the computation basis states,
: construct the matrix for the gate,
|i2 : examine the behavior of the gate on an general state (i.e., one that is not
necessarily separable), and
& : discuss the measurement probabilities of the output registers.
11.3.2
General Learning Example
Well go through the checklist on a not-particularly-practical gate, but one that has
the generality needed to cover future gates.
:
The Symbol
Every binary qubit gate has two input lines, one for each input qubit, and two output
lines, one for each output qubit. The label for the unitary transformation associated
with the gate, say U , is placed inside a box connected to its inputs and outputs.
U
Although the data going into the two input lines can become entangled inside the
gate, we consider the top half of the gate to be a separate register from the lower
half. This can be confusing to new students, as we cant usually consider each output
line to be independent of its partner the way the picture suggests. More (a lot more)
about this shortly.
Vocabulary. The top input/output lines form an upper A register (or A channel )
while the bottom form a lower B register (or B channel ).
A-register out
A-register in
B-register in
B-register out
The labels A and B are usually implied, not written.
277
|xi |yi : Action on the CBS
Every operator is defined by its action on the basis, and in our case thats the computational basis. For binary gates, the symbolism for the general CBS is
|xi |yi
= |xi |yi ,
where x , y {0, 1} .
To demonstrate this on our learning gate, U , we define its action on the CBS, which
in turn defines the gate:
|xi
| yi
U
|yi
| x yi
It is very important to treat the LHS as a single two-qubit input state, not two
separate single qubits, and likewise with the output. In other words, it is really
saying
U |xi |yi
| yi | x yi
or, using shorter notation,
U |xi |yi
| yi | x yi
Furthermore, |xi |yi only represents the four CBS, so we have to extend this linearly
to the entire Hilbert space.
Lets make this concrete. Taking one of the four CBS, say |10i, the above definition
tells us to substitute 1 x and 0 y, to get the gates output,
U |1i |0i
= | 0i | 1 0i = |1i |0i .
[Exercise. Compute the effect of U on the other three CBS.]
: The Matrix
In our linear transformation lesson, we proved that the matrix MT that represents
an operator T can be written by applying T to each basis vector, ak , and placing the
answer vectors in the columns of the matrix,
!
MT
T (a1 ) , T (a2 ) , . . . , T (an )
T (v) is then just MT v. Applying the technique to U and the CBS {|xi |yi} we get
!
MU
U |00i , U |01i , U |10i , U |11i
278
Each of these columns must be turned into the coordinate representation in the
inherited tensor basis of the four U -values. Lets compute them. (Spoiler alert: this
was the last exercise):
0
0
U |00i = | 0i | 0 0i = |1i |1 0i = |1i |1i =
0 .
1
Similarly, we get
1
0
U |01i = | 1i | 0 1i = |0i |1 1i = |0i |0i =
0 ,
0
0
0
U |10i = | 0i | 1 0i = |1i |0 0i = |1i |0i =
1 ,
0
0
1
U |11i = | 1i | 1 1i = |0i |0 1i = |0i |1i =
0 ,
0
giving us the matrix
MU
0
0
0
1
1
0
0
0
0
0
1
0
0
1
,
0
0
which is, indeed, unitary. Incidentally, not every recipe you might conjure for the four
values U (|xi |yi) will produce a unitary matrix and therefore not yield a reversible
and thus valid quantum gate. (We learned last time that non-unitary matrices do
not keep state vectors on the projective sphere and therefore do not correspond to
physically sensible quantum operations.)
[Exercise. Go through the same steps on the putative operator defined by
U |xi |yi
|x yi | x yi .
Is this matrix unitary? Does U constitute a realizable quantum gate?]
|i2 : Behavior on General State
The general state is a superposition of the four basis kets,
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
279
whose coordinates are
|i2
so matrix multiplication gives
0 1
0 0
U |i2 =
0 0
1 0
=
0
0
1
0
0
0
|0i |0i + |0i |1i + |1i |0i + |1i |1i .
Evidently, U leaves the amplitude of the CBS ket, |1i |0i, alone and permutes the
other three amplitudes.
& : Measurement
[Note: Everything in this section, and in general when we speak of measurement,
assumes that we are measuring the states relative to the natural computational basis,
the z-basis, unless explicitly stated otherwise. This is { |0i , |1i } for a single register
and { |00i , |01i , |10i , |11i } for both registers. We can measure relative to other
CBSs, but then some of the results below cannot be applied exactly as stated. Of
course, I would not finish the day without giving you an example.]
Now that we know what U does to input states we can make some basic observations about how it changes the measurement probabilities. Consider a states
amplitudes both before and after the application of the gate (access points P and Q,
respectively):
(
|i2
U |i2
Easy Observations by Looking at Expansion Coefficients. Trait #6 (QMs
fourth postulate) tells us that a measurement of the input state (point P)
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
collapses it to |00i with probability ||2 . Meanwhile, a look at the U |i2 s amplitudes
(point Q),
U |i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
280
reveals that measuring the output there will land it on the state |00i with probability
||2 . This was the inputs probability of landing on |01i prior to the gate; U has
shifted the probability that a ket will register a 01 on our meter to the probability
that it will register a 00. In contrast, a glance at |is pre- and post-U amplitudes
of the CBS |10i tells us that the probability of this state being measured after U is
the same as before: ||2 .
Measurement of Separable Output States. By looking at the expansion
coefficients of the general output state, we can usually concoct a simple input state
that produces
a separable
output. For example, taking = = 0 gives a separable
input, |0iA + |1iA |1iB , as well as the following separable output:
U
|0iA + |1iA
|1iB
= U |0i |1i + |1i |1i = |0i |0i + |0i |1i
= |0iA |0iB + |1iB
We consider a post-gate measurement at access point Q:
|0i + |1i
|0i
U
|1i
|0i + |1i
Q
Measuring the A-register at the output (point Q) will yield a 0 with certainty
(the coefficient of the separable CBS component, |0i, is 1) yet will tell us nothing
about the B-register, which has a ||2 probability of yielding a 0 and ||2 chance of
yielding a 1, just as it did before we measured A. Similarly, measuring B at point
Q will collapse that output register into one of the two B-space CBS states (with
the probabilities ||2 and ||2 ) but will not change a subsequent measurement of the
A-register output, still certain to show us a 0.
(If this seems as though Im jumping to conclusions, it will be explained formally
when we get the Born rule, below.)
A slightly less trivial separable output state results from the input,
!
!
!
!
2
6
2
6
|00i +
|01i +
|10i +
|11i
|i2 =
4
4
4
4
!
|0iA + |1iA
|0iB + 3 |1iB
.
=
2
2
(As it happens, this input state is separable, but thats not required to produce a
separable output state, the topic of this example. I just made it so to add a little
symmetry.)
281
The output state can be written down instantly by permuting the amplitudes
according to U s formula,
!
!
!
!
6
6
2
2
U |i2 =
|00i +
|01i +
|10i +
|11i
4
4
4
4
!
3 |0iA + |1iA
|0iB + |1iB
,
2
2
and I have factored it for you, demonstrating the output states separability. Measuring either output register at access point Q,
|0i + |1i
|0i + 3 |1i
2
3 |0i + |1i
2
|0i + |1i
2
Q
has a non-zero probability of yielding one of the two CBS states for its respective H,
but it wont affect the measurement of the other output register. For example, measuring the B-qubit-out will land it in it |0iB or |1iB with equal probability. Regardless
of which result we get, it will not affect a future measurement of the A-qubit-out which
has a 3/4 chance of measuring 0 and 1/4 chance of showing us a 1.
Measuring One Register of a Separable State. This is characteristic of
separable states, whether they be input or output. Measuring either register does not
affect the probabilities of the other register. It only collapses the component vector of
the tensor, leaving the other vector un-collapsed.
11.3.3
Quantum Entanglement
The Measurement of Non-Separable Output States
As a final example before we get into real gates, we look at what happens when we try
to measure a non-separable output state of our general learning circuit just presented.
Consider the input,
1
1
2
|00i +
|01i
|i
=
2
2
|0iB + |1iB
= |0iA
,
2
282
a separable state that we also know under the alias,
|0i |0ix .
Applying U , this time using matrix notation (for fun and practice), yields
0 1 0 0
1
1
1 1
1 0
0 0 0 1
U |i2 =
=
0 0 1 0
2 0
2 0
1 0 0 0
0
1
|0iA |0iB + |1iA |1iB
,
2
clearly not factorable. Furthermore, unlike the separable output states we have studied, a measurement of either register forces its partner to collapse.
|0i
|0i + |1i
|0i |0i + |1i |1i
A
B
A
B
Q
For example, if we measure Bs output, and find it to be in state |1iB , since the
output ket has only one CBS tensor associated with that |1iB , namely |1iA |1iB , as
we can see from its form
|0i |0i + |1i |1i
,
2
we are forced to conclude that the A-register must have collapsed into its |1i state. If
this is not clear to you, imagine that the A-register had not collapsed to |1i. It would
then be possible to measure a 0 in the A-register. However, such a turn of events
would have landed a |1i in the B-register and |0i in the A-register, a combination
that is patently absent from the output kets CBS expansion, above.
Stated another way (if you are still unsure), there is only one bipartite state here,
and if, when expanded along the CBS basis, one of the four CBS kets is missing
from that expansion that CBS ket has a zero probability of being the result of a
measurement collapse. Since |0i |1i is not in the expansion, this state is not accessible
through a measurement. (And by the way, the same goes for |1i |0i.)
Definition of Quantum Entanglement
An entangled state in a product space is one that is not separable.
283
Non-Locality
Entangled states are also said to be non-local, meaning that if you are in a room
with only one of the two registers, you do not have full control over what happens
to the data there; an observer of the other register in a different room may measure
his qubit and affect your data even though you have done nothing. Furthermore, if
you measure the data in that register, your efforts are not confined to your room but
extend to the outside world where the other register is located. Likewise, separable
states are considered local, since they do allow full segregation of the actions on
separate registers. Each observer has total control of the destiny of his register, and
his actions dont affect the other observer.
The Entanglement Connection
Non-separable states are composed of entangled constituents. While each constituent
may be physically separated in space and time from its partner in the other register,
the two parts do not have independent world lines. Whatever happens to one affects
the other.
This is the single most important and widely used phenomenon in quantum computing, so be sure to digest it well.
Partial Collapse
In this last example a measurement and collapse of one register completely determined
the full and unique state of the output. However, often things are subtler. Measuring
one register may have the effect of only partially collapsing its partner. Well get to
that when we take up the Born rule.
Measurements Using a Different CBS
Everything weve done in the last two sections is true as long as we have a consistent
CBS from start to finish. The definition of U , its matrix and the measurements have
all used the same CBS. But funny things happen if we use a different measurement
basis than the one used to define the operator or express its matrix. Look for an
example in a few minutes.
Now lets do everything again, this time for a famous gate.
11.3.4
The Controlled-NOT (CNOT) Gate
This is the simplest and most commonly used binary gate in quantum computing.
284
:
The Symbol
The gate is usually drawn without an enclosing box as in
but I sometimes box it,
The A-register is often called the control bit, and the B-register the target bit,
control bit
target bit
Well explain that terminology in the next bullet.
|xi |yi : Action on the CBS
The CNOT gate has the following effect on the computational basis states:
|xi
|xi
|x yi
|yi
When viewed on the tiny set of four CBS tensors, it appears to leave the A-register
unchanged and to negate the B-register qubit or leave it alone, based on whether the
A-register is |1i or |0i:
(
|yi ,
if x = 0
|yi 7
| yi ,
if x = 1
Because of this, the gate is described as a controlled-NOT operator. The A-register is
called the control bit or control register, and the B-register is the target bit or target
register. We cannot use this simplistic description on a general state, however, as the
next sections will demonstrate.
285
: The Matrix
We compute the column vectors of the matrix by applying CNOT to the CBS tensors
to get
!
MCNOT
CNOT |00i , CNOT |01i , CNOT |10i , CNOT |11i
!
|00i , |01i , |11i , |10i
1
0
0
0
0
1
0
0
0
0
0
1
0
0
1
0
which, as always, we must recognize as unitary.
[Exercise. Prove that this is not a separable operator. Hint: What did we say
about the form of a separable operators matrix?]
|i2 : Behavior on General State
Applying CNOT to the general state,
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
we get
CNOT |i2
0
1
0
0
0
0
0
1
0
1
0
1
0
0
0
|0i |0i + |0i |1i + |1i |0i + |1i |1i .
A Meditation. If you are tempted to read on, feeling that you understand
everything we just covered, see how quickly you can answer this:
[Exercise. The CNOT is said to leave the source register unchanged and flip
the target register only if the source register input is |1i. Yet the matrix for CNOT
seems to always swap the last two amplitudes, , of any ket. Explain this.]
Caution. If you cannot do the last exercise, you should not continue reading, but
review the last few sections or ask a colleague for assistance until you see the light.
This is an important consequence of what we just covered. It is best that you apply
that knowledge to solve it rather than my blurting out the answer for you.
286
& : Measurement
First, well consider the amplitudes before and after the application of CNOT (access
points P and Q, respectively):
(
|i2
CNOT |i2
A measurement of the input state (point P),
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
will yield a 00 with probability ||2 and 01 with probability ||2 . A post-gate
measurement of CN OT |i2 (point Q),
CNOT |i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
will yield those first two readings with the same probabilities since their kets respective amplitudes are not changed by the gate. However, the probabilities of getting a
10 vs. a 11 reading are swapped. They go from ||2 and ||2 before the gate to
||2 and ||2 , after.
Separable Output States. Theres nothing new to say here, as we have covered
all such states in our learning example. Whenever we have a separable output state,
measuring one register has no affect on the other register. So while a measurement
of A causes it to collapse, B will continue to be in a superposition state until we
measure it (and vice versa).
Quantum Entanglement for CNOT
A separable bipartite state into CNOT gate does not usually result in a separable
state out of CNOT. To see this, consider the separable state
|0i + |1i
2
|0i
|i
= |0ix |0i =
2
going into CNOT:
|0i + |1i
|0i
?
?
287
When presented with a superposition state into either the A or B register, back
away very slowly from your circuit diagram. Turn, instead, to the linear algebra,
which never lies. The separable state should be resolved to its tensor basis form by
distributing the product over the sums,
1
|0i + |1i
1
|0i = |00i + |10i .
2
2
2
[Exception: If you have a separable operator as well as separable input state, we
dont need to expand the input state along the CBS, as the definition of separable
operator allows us to apply the component operators individually to the component
vectors. CNOT is not separable, so we have to expand.]
Now, apply CNOT using linearity,
1
1
CNOT |00i + |10i
=
2
2
=
=
1
1
CNOT (|00i) + CNOT (|10i)
2
2
1
1
|00i + |11i
2
2
|00i + |11i
.
2
This is the true output of the gate for the presented input. It is not separable as is
obvious by its simplicity; there are only two ways we might factor it: pulling out an
A-ket (a vector in the first H space) or pulling out a B-ket (a vector in the second H
space), and neither works.
(Repeat of ) Definition. An entangled state in a product space is
one that is not separable.
Getting back to the circuit diagram, we see there is nothing whatsoever we can
place in the question marks that would make that circuit sensible. Anything we
might try would make it appear as though we had a separable product on the RHS,
which we do not. The best we can do is consolidate the RHS of the gate into a single
bipartite qubit, indeed, an entangled state.
|0i + |1i
2
|0i
|00i + |11i
With an entangled output state such as this, measuring one output register causes the
collapse of both registers. We use this property frequently when designing quantum
algorithms.
Individual Measurement of Output Registers. Although we may have an
entangled state at the output of a gate, we are always allowed to measure each
288
register separately. No one can stop us from doing so; the two registers are distinct
physical entities at separate locations in the computer (or universe). Entanglement
and non-locality mean that the registers are connected to one another. Our intent to
measure one register must be accompanied by the awareness that, when dealing with
an entangled state, doing so will affect the other registers data.
When CNOTs Control Register gets a CBS ...
Now, consider the separable state
|i
=
|1i |0ix
|1i
|0i + |1i
going into CNOT:
|1i
|0i + |1i
?
?
I have chosen the A-register input to be |1i for variety. It could have been |0i with
the same (as of yet, undisclosed) outcome. The point is that this time our A-register
is a CBS while the B-register is a superposition. We know from experience to ignore
the circuit diagram and turn to the linear algebra.
1
|0i + |1i
1
|1i
= |10i + |11i ,
2
2
2
and we apply CNOT
1
1
CNOT |10i + |11i
2
2
=
=
=
1
1
CNOT (|10i) + CNOT (|11i)
2
2
1
1
|11i + |10i
2
2
|1i + |0i
|1i
.
2
Aha separable. Thats because the control-bit (the A-register) is a CBS; it does not
change during the linear application of CNOT so will be conveniently available for
factoring at the end. Therefore, for this input, we are authorized to label the output
registers, individually.
|1i
|0i + |1i
|1i
|0i + |1i
289
The two-qubit output state is unchanged. Not so fast. You have to do an ...
[Exercise. We are told that a |1i going into CNOTs control register means
we flip the B-register bit. Yet, the output state of this binary gate is the same as
the
Explain. Hint: Try the same example with a B-register input of
input state.
.3 |0i + .7 |1i. ]
[Exercise. Compute CNOT of an input tensor |1i |1ix . Does CNOT leave this
state unchanged?]
Summary. A CBS ket going into the control register (A) of a CNOT gate allows
us to preserve the two registers at the output: we do, indeed, get a separable state
out, with the control register output identical to the control register input. This is
true even if a superposition goes into the target register (B). If a superposition goes
into the control register, however, all bets are off (i.e., entanglement emerges at the
output).
What the Phrase The A-Register is Unchanged Really Means
The A, or control, register of the CNOT gate is said to be unaffected by the CNOT
gate, although this is overstating the case; it gives the false impression that a separable
bipartite state into CNOT results in a separable state out, which we see is not the
case. Yet, there are are at least two ways to interpret this characterization.
1. When a CBS state (of the preferred, z-basis) is presented to CNOTs A-register,
the output state is, indeed, separable, with the A-register unchanged.
2. If a non-trivial superposition state is presented to CNOTs A-register, the measurement probabilities (relative to the z-basis) of the A-register output are preserved.
We have already demonstrated item 1, so lets look now at item 2. The general
state, expressed along the natural basis is
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
and CNOTing this state gives
CNOT |i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i .
If we were to measure this state, we know that it would collapse to a CBS in H H.
With what probability would we find that the first CBS (register-A) in this state had
collapsed to |0i? |1i?
We have not had our formal lesson on probability yet, but you can no doubt
appreciate that, of the four possible CBS outcomes, two of them find the A-reg
collapsing to |0iA (even though they are independent events because the B-register of
290
one outcome is |0iB and of the other outcome is |1iB ). Therefore, to get the overall
probability that A collapses to |0iA , we simply add those two probabilities:
P (A-reg output & |0i) = P CNOT |i & |00i
+ P CNOT |i & |01i
=
||2 + ||2 ,
which is exactly the same probability of measuring a 0 on the input, |i, prior to
applying CNOT.
[Exercise. We did not compute the probability of measuring a 0 on the input,
|i. Do that that to confirm the claim.]
[Exercise. What trait of QM (and postulate) tells us that the individual probabilities are ||2 and ||2 ?]
[Exercise. Compute the probabilities of measuring a 1 in the A-register both
before and after the CNOT gate. Caution: This doesnt mean that we would measure
the same prepared state before and after CNOT. Due to the collapse of the state after
any measurement, we must prepare many identical states and measure some before
and others after then examine the outcome frequencies to see how the experimental
probabilities compare.]
Measuring Using a Different CBS
Now we come to the example that I promised: measuring in a basis different from the
one used to define and express the matrix for the gate. Lets present the following
four bipartite states
|00ix , |01ix , |10ix , and |11ix
to the input of CNOT and look at the output (do a measurement) in terms of the
x-basis (which consist of those four tensors).
|00ix : This one is easy because the z-coordinates are all the same.
|00ix
|0ix |0ix
1
|0i + |1i
1
|0i + |1i
1
2 1
2
2
1
CNOT applied to |00ix swaps the last two z-coordinates, which are identical,
so it is unchanged.
CNOT |00ix
|00ix
(Use any method you learned to confirm this rigorously.)
291
|01ix : Just repeat, but watch for signs.
|01ix
|0ix |1ix
1
|0i + |1i
|0i |1i
1
1
=
2 1
2
2
1
Lets use the matrix to get a quick result.
1
1 0 0 0
0 1 0 0 1 1
CNOT |01ix =
0 0 0 1 2 1
1
0 0 1 0
1
1
2 1
1
From here its easier to expand along the z-basis so we can factor,
1
CNOT |01ix =
|0i |0i |0i |1i |1i |0i + |1i |1i
2
1
=
|0i |0i |1i |0i |0i |1i + |1i |1i
2
1
|0i |1i |0i
|0i |1i |1i
=
2
1
=
|0i |1i |0i |1i
2
|0i |1i
|0i |1i
=
= |1ix |1ix = |11ix .
2
2
What is this? Looking at it in terms of the x-basis it left the B-register unchanged at |1ix but flipped the A-register from |0ix to |1ix .
Looking back at the |00ix case, we see that when the B-register held a qubit in
the Sx =0 state, the A-register was unaffected relative to the x-basis.
This looks suspiciously as though the B-register is now the control bit and A is
the target bit and, in fact, a computation of the remaining two cases, |10ix and
|11ix , would bear this out.
|10ix : CNOT |10ix
|10ix [Exercise.]
|11ix : CNOT |11ix
|01ix [Exercise.]
Summarizing, we see that
CNOT |00ix
CNOT |01ix
CNOT |10ix
CNOT |11ix
292
=
=
=
=
|00ix
|11ix
|10ix
|01ix ,
demonstrating that, in the x-basis, the B-register is the control and the A is the
target.
[Preview. We are going to revisit this in a circuit later today. For now, well call
it an upside-down action of the CNOT gate relative to the x-basis and later see how
to turn it into an actual upside-down CNOT gate for the natural CBS kets. When
we do that, well call it CNOT, because it will be controlled from bottom-up in
the z-basis. So for, however, weve only produced this bottom-up behavior in the
x-basis, so the gate name does not change.]
What About Measurement? I advertised this section as a study of measurement, yet all I did so far was make observations about the separable components of
the output which was an eye-opener in itself. Still, lets bring it back to the topic
of measurement.
Take any of the input states, say |10ix . Then the above results say that
CNOT |10ix
|10ix .
To turn this into a statement about measurement probabilities, we dot the output
state with the x-basis kets to get the four amplitudes. By orthogonality of CBS,
x h00 | 10ix
x h01 | 10ix
x h10 | 10ix
x h11 | 10ix
0,
producing the measurement probability of 100% for the state, |10ix . In other words,
for this state which has a B input of 0 (in x-coordinates) its output remains 0
with certainty, while As 1 (again, x-coordinates) is unchanged, also with certainty.
On the other hand, the input state |11ix gave us an output of |01ix , so the output
amplitudes become
x h00 | 01ix
x h01 | 01ix
x h10 | 01ix
x h11 | 01ix
0.
Here the input whose B x-basis component is 1 turns into an output with the B
x-basis component remaining 1 (with certainty) and an A x-basis input 1 becoming
flipped to 0 at the output (also with certainty).
This demonstrates that such statements like The A-register is left unchanged
or The A register is the control qubit, are loosey-goosey terms that must be taken
with a grain of salt. They are vague for non-separable states (as we saw, earlier) and
patently false for measurements in alternate CBSs.
293
11.3.5
The Second-Order Hadamard Gate
We construct the second order Hadamard gate by forming the tensor product of two
first order gates, so wed better first review that operator.
Review of the First Order Hadamard Gate
Recall that the first order Hadamard gate operates on the 2-dimensional Hilbert space,
H, of a single qubit according to its effect on the CBS states
H |0i
|0ix
|0i + |1i
H |1i
|1ix
|0i |1i
,
2
and
which, in an exercise, you showed was equivalent to the CBS formula
|xi
|0i + (1)x |1i
.
2
We learned that its matrix is
1
1 1
,
1 1
and that it affects a general qubit state, |i = |0i + |1i, according to
+
H |i =
|0i +
|1i .
2
2
:
Definition and Symbol
Definition. The second order Hadamard gate, also called the two-qubit or binary
Hadamard gate, is the tensor product of two single-qubit Hadamard gates,
H 2
H H.
Notation. You will see both H H and H 2 when referring to the second order
Hadamard gate. In a circuit diagram, it looks like this
H 2
when we want to show the separate A and B register, or like this
/
H 2
294
when we want to condense the input and output pipes into a multi-pipe. However, it
is often drawn as two individual H gates applied in parallel,
H
H
|xi |yi : Action on the CBS
When a two-qubit operator is defined as a tensor product of two single-qubit operators, we get its action on CBS kets free-of-charge compliments of the definition of
product operator. Recall that the tensor product of two operators T1 and T2 , requires
that its action on separable tensors be
[T1 T2 ](v w)
T1 v T2 w .
This forces the action of H H on the H(2) computational basis state |0i |0i (for
example) to be
|0i + |1i
|0i + |1i
[H H] |0i |0i
= H |0i H |0i =
.
2
2
Separability of CBS Output. Lets pause a moment to appreciate that when an
operator in H(2) is a pure product of individual operators, as this one is, CBS states
always map to separable states. We can see this in the last result, and we know it will
happen for the other three CBS states.
CBS Output in z-Basis Form. Separable or not, its always good to have
the basis-expansion of the gate output for the four CBS kets. Multiplying it out for
H 2 (|0i |0i), we find
1
1
|0i |0i + |0i |1i + |1i |0i + |1i |1i
1 .
=
[H H] |0i |0i
=
2
2 1
1
Doing the same thing for all four CBS kets, we get the identities
H 2 |0i |0i
|0i |0i + |0i |1i + |1i |0i + |1i |1i
2
H 2 |0i |1i
|0i |0i |0i |1i + |1i |0i |1i |1i
2
H 2 |1i |0i
|0i |0i + |0i |1i |1i |0i |1i |1i
2
H 2 |1i |1i
|0i |0i |0i |1i |1i |0i + |1i |1i
.
2
295
Condensed Form #1. There is a single-formula version of these four CBS results,
and it is needed when we move to three or more qubits, so we had better develop it
now. However, it takes a little explanation, so we allow a short side trip.
First, lets switch to encoded notation (0 00, 1 01, 2 10 and 3 11),
and view the above in the equivalent form,
2
|0i2 + |1i2 + |2i2 + |3i2
2
H 2 |1i2
|0i2 |1i2 + |2i2 |3i2
2
H 2 |2i2
|0i2 + |1i2 |2i2 |3i2
2
|0i2 |1i2 |2i2 + |3i2
.
2
|0i
|3i
Next, I will show (with your help) that all four of these can be summarized by the
single equation (note that x goes from the encoded value 0 to 3)
H 2 |xi2
=
(1)x 0 |0i2 + (1)x 1 |1i2 + (1)x 2 |2i2 + (1)x 3 |3i2
,
2
where the operator in the exponent of (1) is the mod-2 dot product that was
defined for the unusual vector space, B2 , during the lesson on single qubits. Ill
repeat it here in the current context. For two CBS states,
x = x1 x0
y = y1 y0 ,
and
where the RHS represents the two-bit string of the CBS (00, 01, etc.), we define
x y
x1 y 1 x0 y 0 ,
that is, we multiply corresponding bits and take their mod-2 sum. To get overly
explicit, here are a few computed mod-2 dot products:
x = x1 x0
3 = 11
1 = 01
0 = 00
1 = 01
1 = 01
y = y1 y0
3 = 11
1 = 01
2 = 10
3 = 11
2 = 10
296
x y
0
1
0
1
0
Ill now confirm that the last expression presented above for H 2 |xi2 , when applied
to the particular CBS |3i2
= |1i |1i, gives the right result:
H 2 |3i2
=
=
=
H 2 |11i
(1)11 00 |0i2 + (1)11 01 |1i2 + (1)11 10 |2i2 + (1)11 11 |3i2
2
2
2
0
1
1
(1) |0i + (1) |1i + (1) |2i2 + (1)0 |3i2
2
2
2
2
(|0i |1i |2i + |3i2
,
2
which matches the final of our four CBS equations for H 2 . Now, for the help I
warned you would be giving me.
[Exercise. Show that this formula produces the remaining three CBS expressions
for H 2 .]
We now present the condensed form using summation notation,
H
|xi
3
1 X
(1)x y |yi2 .
2 y=0
Most authors typically dont use the for the mod-2 dot product, but stick with a
simpler , and add some verbiage to the effect that this is a mod-2 dot product...,
in which case you would see it as
H
|xi
3
1 X
(1)x y |yi2 .
2 y=0
Putting this result into the circuit diagram gives us
(
2
|xi
1
2
3
X
(1)x y |yi2
y=0
or, using the multi-channel pipe notation,
2
|xi
1
2
3
X
(1)x y |yi2
y=0
Condensed Form #2. You may have noticed that the separability of the CBS
outputs, combined with the expression we already had for a single-qubit Hadamard,
H |xi
|0i + (1)x |1i
,
2
297
gives us another way to express H in terms of the CBS. Once again, invoking the
definition of a product operator, we get
[H H] |xi |yi
= H |xi H |yi
|0i + (1)x |1i
|0i + (1)y |1i
=
,
2
2
which, with much less fanfare than condensed form #1, produces a nice separable
circuit diagram definition for the binary Hadamard:
|0i + (1)x |1i
|xi
H 2
|0i + (1)y |1i
|yi
which I will call condensed form #2.
[Caution: The y in this circuit diagram labels the B-register CBS ket, a totally
different use from its indexing of the summation formula in condensed form #1.
However, in both condensed forms the use of y is standard.]
The reason for the hard work of condensed form #1 is that it is easier to generalize
to higher order qubit spaces, and proves more useful in those contexts.
[Exercise. Multiply-out form #2 to show it in terms of the four CBS tensor kets.]
[Exercise. The result of the last exercise will look very different from condensed
form #1. By naming your bits and variables carefully, demonstrate that both results
produce the same four scalar amplitudes standing next to the CBS tensor terms.]
: The Matrix
This time we have two different approaches available. As before we can compute
the column vectors of the matrix by applying H 2 to the CBS tensors. However,
the separability of the operator allows us to use the theory of tensor products to
write down the product matrix based on the two component matrices. The need for
frequent sanity checks wired into our collective computer science mindset induces us
to do both.
Method #1: Tensor Theory.
Using the standard A-major (B-minor) method, a separable operator A B can
298
be immediately written down using our formula
a00 a01 a0(l1)
b00 b01
a10 a11 a1(l1) b10 b11
..
.. . .
..
..
..
.
.
.
.
.
.
b00 b01
b00
a b10 b11 a b10
01
00
..
.. . .
..
.
.
.
b00 b01
b00
b10
a10 b10 b11
a11
..
.. . .
..
.
.
.
.
..
.
b0(m1)
b1(m1)
..
..
.
.
b01
b11
.. . .
b01
b11
.. . .
.
.
..
..
.
.
For us, this means
H H
1
1
1
1
2 1 1 1
2
1
1 1
1
1 1 1
1
1 1
2 1
1 1 1
1
2
1
1 1
1
2
1
1 1
1
1 1
1 12
1
1 1
1
1
,
1
1
a painlessly obtained result.
Method #2: Linear Algebra Theory.
Our earlier technique which works for any transformation, separable or not, instructs us to place the output images of the four CBS kets into columns of the desired
299
matrix. Using our four expressions for H 2 |xi |yi presented initially, we learn
1
1
1
1
H 2 |00i =
|00i + |01i + |10i + |11i
=
2
2 1
1
1
1
1
1
|00i |01i + |10i |11i
=
H 2 |01i =
1
2
2
1
1
1
1
1
|00i + |01i |10i |11i
=
H 2 |10i =
2
2 1
1
1
1 1
1
.
|00i |01i |10i + |11i
=
H 2 |11i =
2
2 1
1
Therefore,
!
MH 2
|00i ,
|01i ,
|10i ,
|11i
1
1
1
1
1
1 1 1 1 .
1 1 1
21
1 1 1
1
happily, the same result.
[Exercise. Complete the sanity check by confirming that this is a unitary matrix.]
|i2 : Behavior on General State
To a general
|i2
we apply our operator
H 2 |i2
1
1
1
1
1
1
1
1
1
1 1 1
21
1 1 1
1
300
+++
1
+ ,
2 +
+
which shows the result of applying the two-qubit Hadamard to any state.
& : Measurement
There is no concise phrase we can use to describe how the binary Hadamard affects
measurement probabilities of a general state. We must be content to describe it in
terms of the algebra. For example, testing at point P,
(
|i2
H 2
H 2 |i2
collapses |i2 to |00i with probability ||2 (as usual). Waiting, instead, to take the
measurement of H 2 |i2 (point Q), would produce a collapse to that same |00i with
the probability
+ + +
2
( + + + ) ( + + + )
=
,
2
4
an accurate but relatively uninformative fact.
Measurement of CBS Output States. However, there is something we can
say when we present any of the four CBS tensors to our Hadamard. Because its
separable, we can see its output clearly:
|0i + (1)x |1i
2
|0i + (1)y |1i
|xi
H 2
|yi
This tells us that if we measure either register at Q, we will read a 0 or 1 with
equal (50%) probability, and if we measure both registers at Q, well see a 00, 01,
10 or 11 also with equal (25%) probability. Compare this to the probabilities at
P, where we are going to get either of our two input qubit values, x and y with
100% probability (or, if we measure both, the bipartite input value xy with 100%
probability) since they are CBS states at the time of measurement.
301
Converting Between z-Basis and x-Basis using H 2
The induced z-basis for H(2) is
|0i |0i , |0i |1i , |1i |0i ,
and
|1i |1i ,
while the induced x-basis for H(2) is
|0ix |0ix , |0ix |1ix , |1ix |0ix ,
|1ix |1ix ,
and
or in alternate notation,
|+i |+i , |+i |i , |i |+i ,
and
|i |i .
Since H converts to-and-from the bases {|0i , |1i} and {|0ix , |1ix } in H(1) , it is a
short two-liner to confirm that the separable H 2 converts to-and-from their induced
second order counterparts in H(2) . Ill show this for one of the four CBS kets and you
can do the others. Lets take the third z-basis ket, |1i |0i,
H 2 |1i |0i
[H H] (|1i |0i)
|1ix |0ix
(H |1i) (H |0i)
|1ix |0ix
= |i |+i
=
[Exercise. Do the same for the other three CBS kets.]
This gives us a very useful way to view H 2 when we are thinking about bases:
H 2 is the transformation that takes the z-CBS to the x-CBS (and since
it is its own inverse, also takes x-CBS back to the z-CBS).
Algebraically, using the alternate |i notation for the x-basis, the forward direction
is
H 2
H 2
H 2
H 2
|00i
|01i
|10i
|11i
H 2
H 2
H 2
H 2
|++i
|+i
|+i
|i
=
=
=
=
|++i ,
|+i ,
|+i and
|i ,
and the inverse direction is
11.4
=
=
=
=
|00i ,
|01i ,
|10i and
|11i .
Measuring Along Alternate Bases
In the earlier study, when we were applying the naked CNOT gate directly to kets,
I described the final step as measuring the output relative to the x-basis. At the
time it was a bit of an informal description. How would it be done practically?
302
11.4.1
Measuring Along the x-Basis
We apply the observation of the last section. When we are ready to measure along
the x-basis, we insert the quantum gate that turns the x-basis into the z-basis, then
measure in our familiar z-basis. For a general circuit represented by a single operator
U we would use
H
,
U
H
to measure along the x-basis. (The meter symbols imply a natural z-basis measurement, unless otherwise stated.)
11.4.2
Measuring Along any Separable Basis
Say we have some non-preferred basis,
C
{ |0 i , |1 i } ,
of our first order Hilbert space, H(1) . Lets also assume we have the unary operator,
call it T , that converts from the natural z-CBS to the C-CBS,
T : z-basis C ,
T |0i = |0 i ,
T |1i = |1 i .
Because T is unitary, its inverse is T , meaning that T takes a C-CBS back to a
z-CBS.
T : C
z-basis ,
T |0 i = |0i ,
T |1 i = |1i .
Moving up to our second order Hilbert space, H(2) , the operator that converts from
the induced 4-element z-basis to the induced 4-element C-basis is T T , while to go
in the reverse direction we would use T T ,
T T |00 i = |00i ,
T T |01 i = |01i
T T |10 i = |10i ,
T T |11 i = |11i .
303
To measure a bipartite output state along the induced C-basis, we just apply T T
instead of the Hadamard gate H H at the end of the circuit:
T
U
Question to Ponder
How do we vocalize the measurement? Are we measuring the two qubits in the C basis
or in the z-basis? It is crucial to our understanding that we tidy up our language.
There are two ways we can say it and they are equivalent, but each is very carefully
worded, so please meditate on them.
1. If we apply the T gate first, then subsequent measurement will be in the z-basis.
2. If we are talking about the original output registers of U before applying the T
gates to them, we would say we are measuring that pair along the C basis. This
version has built-into it the implication that we are first sending the qubits
through T s, then measuring them the z-basis.
How We Interpret the Final Results
If we happen to know that, just after the main circuit but before the final basistransforming T T , we had a C-CBS state, then a final reading that showed the
binary number x (x = 0, 1, 2 or 3) on our meter would imply that the original
bipartite state before the T s was |x i. No collapse takes place when we have a CBS
ket and measure in that CBS basis.
However, if we had some superposition state of x-CBS kets, |i2 , at the end of our
main circuit but before the basis-transforming T T , we have to describe things
probabilistically. Lets express this superposition as
|i2
c00 |00 i + c01 |01 i + c10 |10 i + c11 |11 i .
By linearity,
T T |i2
c00 |00i + c01 |01i + c10 |10i + c11 |11i ,
a result that tells us the probabilities of detecting the four natural CBS kets on our
final meters are the same as the probabilities of our having detected the corresponding
x-CBS kets prior to the final T T gate. Those probabilities are, of course, |cx |2 ,
for x = 0, 1, 2 or 3.
304
11.4.3
The Entire Circuit Viewed in Terms of an Alternate
Basis
If we wanted to operate the circuit entirely using the alternate CBS, i.e., giving it
input states defined using alternate CBS coordinates as well as measuring along the
alternate basis, we would first create a circuit (represented here by a single operator
U ) that works in terms of the z-basis, then surround it with T and T gates,
T
T
U
11.4.4
Non-Separable Basis Conversion Gates
While on the topic, I should mention non-separable basis conversion transformations
(to/from the z-basis). A basis conversion operator doesnt have to be separable in
order for it to be useful in measuring along that basis. Like all quantum gates it is
unitary, so everything we said about separable conversion operators will work. To
measure along any basis, we find the binary operator, call it S, that takes the z-basis
to the other basis and use S prior to measurement.
.
An Entire Circuit in a Non-Separable Basis
If we wanted to design a circuit in terms of input coordinates and measurements
relative to a non-preferred (and non-separable) basis, we would
1. first define the circuit in terms of the natural z-basis, then
2. sandwich it between S and S gates.
The circuit diagram for this is
Well be using both separable and non-separable basis conversion operators today
and in future lessons.
305
11.5
Variations on the Fundamental Binary Qubit
Gates
Weve seen three examples of two-qubit unitary operators: a general learning example,
the CN OT and the H 2 . There is a tier of binary gates which are derivative of one
or more of those three, and we can place this tier in a kind of secondary category
and thereby leverage and/or parallel the good work weve already done.
11.5.1
Other Controlled Logic Gates
The Controlled-U Gate
The CNOT is a special case of a more general gate which places any unary (onequbit) operation in the B-register to be controlled by the qubit in the A-register.
If U is any unary gate, we can form a binary qubit gate called a controlled-U gate
(or U -operator ). Loosely stated, it applies the unary operator on one registers CBS
conditionally, based on the CBS qubit going into the other register.
:
The Symbol
The controlled-U gate is drawn like the CNOT gate with CNOTs operation replaced by the unary operator, U , that we wish to control:
U
For example, a controlled-bit flip (= controlled-QNOT) operator would be written
,
X
whereas a controlled-phase shift operator, with a shift angle of , would be
.
R
306
The A-register maintains its role of control bit/register, and the B-register the target
register or perhaps more appropriately, target (unary) operator.
control register
target operator
There is no accepted name for the operator, but Ill use the notation CU (i.e.,
CX, C(R ), etc.) in this course.
|xi |yi : Action on the CBS
It is easiest and most informative to give the effect on the CBS when we know which
specific operator, U , we are controlling. Yet, even for a general U we can give formal
expression using the power (exponent) of the matrix U . As with ordinary integer
exponents, U n simply means multiply U by itself n times with the usual convention
that U 0 = 1:
|xi
|xi
|yi
U x |yi
The abstract expression is equivalent to the more understandable
(
|yi ,
if x = 0
|yi 7
U |yi ,
if x = 1
It will help if we see the four CBS results as explicitly:
1
0
CU |00i = |0i |0i =
0
0
0
1
CU |01i = |0i |1i =
0
0
CU |10i
|1i U |0i
CU |11i
|1i U |1i
307
0
0
U00
U10
0
0
U01
U11
Here the Ujk are the four matrix elements of MU .
[Exercise. Verify this formulation.]
: The Matrix
The column vectors of the matrix are given by applying CU to the CBS tensors to
get
!
MCU
CU |00i , CU |01i , CU |10i , CU |11i
!
|00i , |01i , |1i U |0i , |1i U |1i
1
0
0
0
0 0
0
1 0
0
,
0 U00 U01
0 U10 U11
which can be more concisely expressed as
CU
or, the even cleaner
U
[Exercise. Prove that this is a separable operator for some U and not-separable
for others. Hint: What did we say about CNOT in this regard?]
[Exercise. Confirm unitarity.]
|i2 : Behavior on General State
Applying CU to the general state,
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
we get
CU |i2
1
0
0
0
0 0
0
1 0
0
0 U00 U01
0 U10 U11
U00 + U01
U10 + U11
|0i |0i + |0i |1i + (U00 + U01 ) |1i |0i + (U10 + U11 ) |1i |1i .
308
& : Measurement
Theres nothing new we have to add here, since measurement probabilities of the
target-register will depend on the specific U being controlled, and we have already
established that the measurement probabilities of the control register are unaffected
(see CNOT discussion).
The Controlled-Z Gate
Lets get specific by looking at CZ.
:
The Symbol
Its symbol in a circuit is
|xi |yi : Action on the CBS
Recall that we had what appeared to be an overly abstract expression for the unary
Z-gate operating on a CBS |xi,
Z |xi
(1)x |xi ,
but this actually helps us understand the CZ action using a similar expression:
CZ |xi |yi
(1)xy |xi |yi .
[Exercise. Prove this formula is correct. Hint: Apply the wordy definition of a
controlled Z-gate (leaves the B-reg alone if ... and applies Z to the B-reg if ...) to
the four CBS tensors and compare what you get with the formula (1)xy |xi |yi for
each of the four combinations of x and y.]
309
To see the four CBS results as explicitly,
CZ |00i
CZ |01i
CZ |10i
CZ |11i
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
: The Matrix
The column vectors of the matrix were produced above, so we can write it down
instantly:
1 0 0 0
0 1 0 0
CZ =
0 0 1 0
0 0 0 1
[Exercise. Prove that this (a) is not a separable operator and (b) is unitary.]
|i2 : Behavior on General State
Applying CU to the general state,
|i2
|0i |0i + |0i |1i + |1i |0i + |1i |1i ,
we get
CZ |i2
1
0
0
0
0
1
0
0
0 0
0 0
1 0
0 1
310
& : Measurement
As we noticed with the unary Z-operator, the probabilities of measurement (along the
preferred CBS) are not affected by the controlled-Z. However, the state is modified.
You can demonstrate this the same way we did for the unary Z: combine |i2 and
CZ |i2 with a second tensor and show that you (can) produce distinct results.
Swapping Roles in Controlled Gates
We could have (and still can) turn any of our controlled gates upside down:
U
Now the B-register takes on the role of control, and the A-register becomes the target:
target operator
control register
Lets refer to this version of a controlled gate using the notation (caution: not
seen outside of this course) (C) U
Everything has to be adjusted accordingly if we insist which we do on continuing to call the upper register the A-register, producing A-major (B-minor ) matrices
and coordinate representations. For example, the action on the CBS becomes
|xi
U y |xi
|yi
|yi
and the matrix we obtain (make sure you can compute this this) is
1
0
0
0
0 U
U01
00 0
(C) U =
,
0
0
1
0
0 U10 0 U11
where the Ujk are the four matrix elements of MU .
With this introduction, try the following exercise.
311
[Exercise. Prove that the binary quantum gates CZ and (C) Z are identical.
That is, show that
Z
=
Hint: If they are equal on the CBS, they are equal period.]
11.5.2
More Separable Logic Gates
The binary Hadamard gate, H H, is just one example of many possible separable
gates which, by definition, are constructed as the tensor product of two unary operators. Take any two single-qubit gates, say X and H. We form the product operator,
and use tensor theory to immediately write down its matrix,
0 MH
1 MH
X H =
1 MH
0 MH
0 0
1
0 0
2 1 1
1 1
1 1
1 1
.
0 0
0 0
[Exercise. Verify unitarity.]
Vocabulary. Separable operators are also called local operators.
Separable Operators on Separable States. All separable operators take separable states to other separable states, therefore, they preserve locality (which is why
they are also known as local operators). In particular, their action on the CBS result
in separable outputs.
One of the notational options we had for the binary Hadamard gate,
H
H
provided a visual representation of this locality. It correctly pictures the non-causal
separation between the two channels when separable inputs are presented to the gate.
This is true for all separable operators like our current X H,
|i
|i
X
H
X |i
H |i
demonstrating the complete isolation of each channel, but only when separable states
are sent to the gate. If an entangled state goes in, an entangled state will come out.
312
The channels are still separate, but the input and output must be displayed as unified
states,
(
)
X
2
|i
(X H) |i2
H
Separable Operators on Entangled States
A single-qubit operator, U , applied to one qubit of an entangled state is equivalent
to a separable binary operator like 1 U or U 1. Such local operations will
generally modify both qubits of the entangled state, so a separable operators effect
is not restricted to only one channel for such states. To reinforce this concept, let U
be any unary operator on the A-space and form the separable operator
U 1 : HA HB HA HB .
Lets apply this to a general |i2 , allowing the possibility that this is an entangled
state.
(U 1) |i2 = [U 1] |00i + |01i + |10i + |11i
= U |0i |0i + U |0i |1i + U |1i |0i + U |1i |1i .
Replacing U |0i and U |1i with their expansions along the CBS (and noting that the
matrix elements, Ujk , are the weights), we get
(U 1) |i2
=
U00 |0i + U10 |1i |0i + U00 |0i + U10 |1i |1i
+ U01 |0i + U11 |1i |0i + U01 |0i + U11 |1i |1i .
Now, distribute and collect terms for the four CBS tensor basis, to see that
(U 1) |i2 =
U00 + U01 |00i +
and without even finishing the regrouping we see that the amplitude of |00i can be
totally different from its original amplitude, proving that the new entangled state is
different, and potentially just as entangled.
[Exercise. Complete the unfinished expression I started for (U 1) |i2 , collect
terms for the four CBS kets, and combine the square-magnitudes of |0iB to get the
probability of B measuring a 0 . Do the same for the |1iB to get the probability
of B measuring a 1 . Might the probabilities of these measurements be affected by
applying U to the A-register?]
Vocabulary. This is called performing a local operation on one qubit of an
entangled pair.
313
The Matrix. It will be handy to have the matrix for the operator U 1 at
the ready for future algorithms. It can be written down instantly using the rules for
separable operator matrices (see the lesson on tensor products),
U00 0
U01 0
0 U00
0 U01
.
U10 0
U11 0
0 U10
0 U11
Example. Alice and Bob each share one qubit of the entangled pair
1
|00i + |11i
1
0
,
|00 i
=
2
2 0
1
which youll notice I have named |00 i for reasons that will be made clear later today
when we get to the Bell states. Alice will hold the A-register qubit (on the left of
each product) and Bob the B-register qubit, so you may wish to view the state using
the notation
|0iA |0iB + |1iA |1iB
.
2
Alice sends her qubit (i.e., the A-register) through a local QNOT operator,
0 1
X =
,
1 0
and Bob does nothing. This describes the full local operator
X 1
applied to the entire bipartite state. We want to know the effect this has on the
total entangled state, so we apply the matrix for X 1 to the state. Using our
pre-computed matrix for the general U 1 with U = X, we get
X 1 |00 i
0
0
1
0
0
0
0
1
1 0
1
0 1
1 0
0 0 2 0
1
0 0
|01i + |10i
.
2
314
1
1
2 1
0
We could have gotten the same result, perhaps more quickly, by distributing X 1
over the superposition and using the identity for separable operators,
[S T ] (v w)
S(v) T (w) .
Either way, the result will be used in our first quantum algorithm (superdense coding),
so its worth studying carefully. To help you, here is an exercise.
[Exercise. Using the entangled state |00 i as input, prove that the following local
operators applied on Alices end produce the entangled output states shown:
Z 1 |00 i = (|00i |11i) / 2 ,
iY 1
|00 i = (|01i |10i) / 2
and, for completeness, the somewhat trivial
1 1 |00 i = (|00i + |11i) / 2 .
11.6
The Born Rule and Partial Collapse
Weve had some experience with the famous entangled state
|00 i
|0i |0i + |1i |1i
.
2
If we measure the A-register of this state, it will of course collapse to either |0iA or
|1iA as all good single qubits must. That, in turn, forces the B-register into its version
of that same state, |0iB or |1iB , meaning if A collapsed to |0iA , B would collapse to
|0iB , and likewise for |1i; there simply are no terms present in |00 is CBS expansion
in which A and B are in different states. In more formal language, the CBS terms in
which the A qubit and B qubit differ have zero amplitude and therefore zero collapse
probability.
To make sure there is no doubt in your mind, consider a different state,
|i2
|0i |1i + |1i |0i
.
2
This time, if we measured the A-register and it collapsed to |0iA , that would force
the B-register to collapse to |1iB .
[Exercise. Explain this last statement. What happens to the B-register if we
measure the A-register and it collapses to |1iA ?]
All of this is an example of a more general phenomenon in which we measure one
qubit of an entangled bipartite state.
315
11.6.1
The Born Rule for a Two-Qubit System
Take the most general state
|i2
|00i + |01i + |10i + |11i
and rearrange it so that either the A-kets or B-kets are factored out of common terms.
Lets factor the A-kets for this illustration.
+ |1iA |0iB + |1iB .
|i2 = |0iA |0iB + |1iB
(I labeled the state spaces of each ket to reinforce which kets belong to which register,
but position implies this information even without the labels. I will often label a
particular step in a long computation when I feel it helps, leaving the other steps
unlabeled.)
What happens if we measure A and get a 0 ? Since there is only one term which
matches this state, namely,
|0i |0i + |1i ,
we are forced to conclude that the B-register is left in the non-normalized state
|0i + |1i .
There are a couple things that may be irritating you at this point.
1. We dont actually have a QM postulate that seems to suggest this claim.
2. We are not comfortable that Bs imputed state is not normalized.
We are on firm ground, however, because when the postulates of quantum mechanics
are presented in their full generality, both of these concerns are addressed. The
fifth postulate of QM (Trait #7), which addresses post-measurement collapse has
a generalization sometimes called the generalized Born rule. For the present, well
satisfy ourselves with a version that applies only to a bipartite states one-register
measurement. Well call it the . . .
Trait #15 (Born Rule for Bipartite States): If a bipartite state is factored
relative to the A-register,
|i2 = |0i |0i + |1i
+ |1i |0i + |1i ,
316
a measurement of the A-register will cause the collapse of the B-register according to
A & 0
|0i + |1i
B & p
||2 + ||2
A & 1
|0i + |1i
B & p
.
||2 + ||2
Note how thisphandles the non-normality of the state |0i + |1i: we divide though
by the norm ||2 + ||2 . (The same is seen for the alternative state |0i + |1i).
Vocabulary. Well call this simply the Born Rule.
Trait #15 has a partner which tells us what happens if we first factor out the
B-ket and measure the B-register.
[Exercise. State the Born Rule when we factor and measure the B -register.]
Checking the Born Rule in Simple Cases
You should always confirm your understanding of a general rule by trying it out on
simple cases to which you already have an answer.
Example. The state we encountered,
|i2
|0i |0i + |1i |1i
,
2
has
= = 1
= = 0,
and
so if we measure the A-register and find it to be in the state |0iA (by a measurement
of 0 on that register), the Born rule tells us that the state remaining the B-register
should be
|0i + |1i
p
||2 + ||2
1 |0i + 0 |1i
12 + 02
which is what we concluded on more elementary grounds.
|0i ,
X
[Exercise. Use this technique to prove that if we start in the same state and,
instead, measure the B-register with a result of 1, then the A-register is left in the
state |1iA with certainty.]
Example. Lets take a somewhat less trivial bipartite state. Consider
|i2
|0i |0i + |0i |1i + |1i |0i |1i |1i
,
2
317
and imagine that we test the B-register and find that it decided to collapse to |0iB .
To see what state this leaves the A-register in, factor out the B-kets of the original
to get an improved view of |i2 ,
|0iA + |1iA |0iB +
|0iA |1iA |1iB
|i2 =
.
2
Examination of this expression tells us that the A-register corresponding to |0iB is
some normalized representation of the vector |0i + |1i. Lets see if the Born rule
gives us that result. The expressions four scalars are
1
and
= = =
2
1
= ,
2
so a B-register collapse to |0iB will, according to the Born rule, leave the A-register
in the state
1
|0i + 21 |1i
|0i + |1i
|0i + |1i
p
= p2
=
,
2
||2 + ||2
(1/2)2 + (1/2)2
again, the expected normalized state.
[Exercise. Show that a measurement
of B & 1 for the same state results in an
A-register collapse to |0i |1i / 2 .]
Application of Born Rule to a Gate Output
The Born rule gets used in many important quantum algorithms, so theres no danger
of over-doing our practice. Lets take the separable gate 1 H, whose matrix you
should (by now) be able to write down blindfolded,
1 1
1 1 1
.
1H =
1 1
2
1 1
Hand it the most general |i2 = (, , , )t , which will produce gate output
1
2 +
We measure the B register and get a 1. To see whats left in the A-register, we
factor the B-kets at the output,
1 H |i2 =
( + ) |0i + ( + ) |1i |0i
+
( ) |0i + ( )) |1i |1i .
318
(At this point, we have to pause to avoid notational confusion. The of the Born
rule is actually our current ( ), with similar unfortunate name conflicts for the
other three Born variables, all of which Im sure you can handle.) The Born rule says
that the A-register will collapse to
( ) |0i + ( ) |1i
p
,
| |2 + | |2
which is as far as we need to go, although it wont hurt for you to do the ...
[Exercise. Simplify this and show that it is a normal vector.]
11.7
Multi-Gate Circuits
We have all the ingredients to make countless quantum circuits from the basic binary
gates introduced above. Well start with the famous Bell states.
11.7.1
A Circuit that Produces Bell States
There are four pairwise entangled states, known as Bell states or EPR pairs (for the
physicists Bell, Einstein, Podolsky and Rosen who discovered their special qualities).
Weve already met one,
|00 i
|00i + |11i
,
2
and the other three are
|01 i
|01i + |10i
,
2
|10 i
|00i |11i
|11 i
|01i |10i
.
2
and
The notation I have adopted above is after Nielsen and Chuang, but physicists also
use the alternative symbols
|00 i
|+ i ,
|01 i
|+ i ,
|10 i
| i
|11 i
| i .
and
In addition, I am not using the superscript to denote a bipartite state |00 i2 , since
the double index on (00 ) tells the story.
319
The circuit that produces these four states using the standard CBS basis for H(2)
as inputs is
)
|xi
H
|xy i ,
|yi
which can be seen as a combination of a unary Hadamard gate with a CNOT gate.
We could emphasize that this is a binary gate in its own right by calling it BELL and
boxing it,
|xi
)
|xy i .
|yi
In concrete terms, the algebraic expression,
|xi |yi
|0i |0i
|00 i ,
|0i |1i
|01 i ,
|1i |0i
|10 i
|1i |1i
|11 i .
BELL
|xy i
is telling us that
BELL
BELL
BELL
BELL
and
The Bell States Form an Orthonormal Basis
When studying unary quantum gates, we saw that they take orthonormal bases to
orthonormal bases. The same argument unitarity proves this to be true of any
dimension Hilbert space. Consequently, the Bell states form an orthonormal basis for
H(2) .
The Matrix for BELL
The matrix for this gate can be constructed using the various techniques we have
already studied. For example, you can use the standard linear algebra approach of
building columns for the matrix from the four outputs of the gate. For variety, lets
take a different path. The A-register Hadamard has a plain quantum wire below it,
meaning the B-register is implicitly performing an identity operator at that point.
So we could write the gate using the equivalent symbolism
,
1
320
a visual that demonstrates the application of
1
0
BELL = (CN OT ) (H 1) =
0
0
1
0
2 0
1
two known matrices
0 0 0
1
1
1 0 0
0
0 0 1 2 1
0 1 0
0
in series, that is,
0 1
0
1 0
1
0 1 0
1 0 1
0 1
0
1 0
1
.
1 0 1
0 1
0
At this point, you should do a few things.
[Exercise.
1. Explain why the order of the matrices is opposite the order of the gates in the
diagram.
2. Confirm the matrix multiplication, above.
3. Confirm that the matrix is unitary.
4. Confirm that the matrix gives the four Bell states when one presents the four
CBS states as inputs. (See below for a hint.)
5. Demonstrate that BELL is not separable. Hint: What do we know about the
matrix of a separable operator? ]
Ill do item 4 for the input state |10i, just to get the blood flowing.
1 0 1
0
1
0
1
1 0
0 1 0
1 0
BELL |10i =
=
2 0 1 0 1 1
2 0
1 0 1 0
1
0
=
|00i |11i
|10 i .
Four Bell States from One
We did an example earlier that demonstrated that
X 1 |00 i = (|01i + |10i) / 2 ,
321
and followed it with some exercises. We can now
the EPR pairs.
1 1 |00 i =
X 1 |00 i =
Z 1 |00 i =
iY 1 |00 i =
list those results in the language of
|00 i
|01 i
|10 i
|11 i
This might be a good time to appreciate one of todays earlier observations: a local
(read separable) operation on an entangled state changes the entire state, affecting
both qubits of the entangled pair.
BELL as a Basis Transforming Operator
The operator BELL takes natural CBS kets to the four Bell kets, the latter shown to
be an orthonormal basis for H(2) . But thats exactly what we call a basis transforming
operator. Viewed in this light BELL, like H 2 , can be used when we want to change
our basis. Unlike H 2 , however, BELL is not separable (a recent exercise) and not
its own inverse (to be proven in a few minutes).
Measuring Along the Bell Basis. We saw that to measure along any basis,
we find the binary operator, call it S, that takes the z-basis to the other basis and
use S prior to measurement.
Thus, to measure along the BELL basis (and we will, next lecture), we plug in BELL
for S,
BELL
And what is BELL ? Using the adjoint conversion rules, and remembering that the
order of operators in the circuit is opposite of that in the algebra, we find
BELL
=
=
i
(CN OT ) H 1
H 1 (CN OT ) ,
h
H 1
(CN OT )
the final equality a consequence of the fact that CNOT and H 1 are both self-adjoint.
322
[Exercise. Prove this last claim using the matrices for these two binary gates.]
In other words, we just reverse the order of the two sub-operators that comprise
BELL. This makes the circuit diagram for BELL come out to be
H
,
1
The matrix for BELL is easy to derive since we just take the transpose (everythings
real so no complex conjugation necessary):
0 1
0
1 0
1
1 0 1
0 1 0
BELL
11.7.2
1
0
2 0
1
1
0
2 1
0
0 0
1
1 1
0
.
0 0 1
1 1 0
A Circuit that Creates an Upside-Down CNOT
Earlier today we demonstrated that the CNOT gate, when presented and measured
relative to the x-basis, behaved upside down. Now, Im going to show you how to
turn this into a circuit that creates an upside down CNOT in the standard z-basis
CBS.
First, the answer. Here is your circuit, which well call CNOT, because it is
controlled from bottom-up:
|xi
|yi
|x yi
|yi
Notice that it has a right-side up (normal) CNOT gate at its core.
The first thing we can do is verify the claim indicated at the output registers.
As we just did with the BELL gate, we can do this most directly by multiplying the
matrices of the three binary sub-gates evident by grouping this gate as
H
H
.
323
That corresponds to the matrix product
1
1
1
1
1
1 1 1 1 1 0
C NOT =
1 1 1 0
21
1 1 1
1
0
0
1
0
0
0
0
0
1
0
1 1
1
1
0
1 1 1 1 1
1 2 1
1 1 1
0
1 1 1
1
1 1
1
1
1
1
1
1
1
1 1 1 1 1 1 1 1
1 1 1
1 1 1
1
4 1
1 1 1
1
1
1 1 1
1
0
0
0
0
0
0
1
0
0
1
0
4
1
0
4 0
0
0
0
0
4
0
0
4
0
0
4
0
0
0
1
.
0
0
This is an easy matrix to apply to the four CBS kets with the following results:
C NOT
C NOT
C NOT
C NOT
:
:
:
:
|00i
|01i
|10i
|11i
7
7
7
7
|00i
|11i
|10i
|01i
We can now plainly see that the B-register is controlling the A-registers QNOT
operation, as claimed.
Interpreting the Upside-Down CNOT Circuit
We now have two different studies of the upside-down CNOT. The first study concerned the naked CNOT and resulted in the observation that, relative to the x-CBS,
the B-register controlled the QNOT (X) operation on the A-register, thus it looks
upside-down if you are an x-basis ket. The second and current study concerns a new
circuit that had the CNOT surrounded by Hadamard gates and, taken as a whole is
a truly upside-down CNOT viewed in the ordinary z-CBS. How do these two studies
compare?
The key to understanding this comes from our recent observation that H 2 can
be viewed as a way to convert between the x-basis and the z-basis (in either direction
since its its own inverse).
Thus, we use the first third of the three-part circuit to let H 2 take z-CBS kets
to x-CBS kets. Next, we allow CNOT to act on the x-basis, which we saw from
our earlier study caused the B-register to be the control qubit because and only
because we are looking at x-basis kets. The output will be x-CBS kets (since we put
x-CBS kets into the central CNOT). Finally, in the last third of the circuit we let
H 2 convert the x-CBS kets back to z-CBS kets.
324
11.8
More than Two Qubits
We will officially introduce n-qubit systems for n > 2 next week, but we can find ways
to use binary qubit gates in circuits that have more than two inputs immediately as
long as we operate on no more than two qubits at at-a-time. This will lead to our
first quantum algorithms.
11.8.1
Order-3 Tensor Products
If a second order product space is the tensor product of two vector spaces,
W
AB,
then its easy to believe that a third order product space, would be constructed from
three component spaces,
W
ABC.
This can be formalized by relying on the second order construction, and applying it
twice, e.g.,
W
(A B) C .
Its actually less confusing to go through our order-2 tensor product development and
just extend all the definitions so that they work for three component spaces. For
example, taking a page from our tensor product lecture, we would start with the
formal vector symbols
abc
and produce all finite sums of these things. Then, as we did for the A B product
space, we could define tensor addition and scalar multiplication. The basic concepts
extend directly to third order product spaces. Ill cite a few highlights.
The dimension of the product space, W , is the product of the three dimensions,
dim (W )
dim (A) dim (B) dim (C)
and W has as its basis
wjkl
aj bk cl
where {aj }, {bk } and {cl } are the bases of the three component spaces.
The vectors (tensors) in the product space are uniquely expressed as superpositions of these basis tensors so that a typical tensor in W can be written
X
w =
cjkl (aj bk cl ) ,
j,k,l
where cjkl are the amplitudes of the CBS kets, scalars which we had been naming
, , , etc. in a simpler era.
325
A separable operator on the product space is one that arises from three component operators, TA , TB and TC , each defined on its respective component space,
A, B and C. This separable tensor operator is defined first by its action on
separable order-3 tensors
[TA TB TC ] (a b c)
TA (a) TB (b) TC (c)
and since the basis tensors are of this form, that establishes the action of TA
TB TC on the basis which in turn extends the action to the whole space.
If any of this seems hazy, I encourage you to refer back to the tensor produce
lecture and fill in details so that they extend to three component spaces.
[Exercise. Replicate the development of an order-2 tensor product space from
our past lecture to order-3 using the above definitions as a guide.]
11.8.2
Three Qubit Systems
Definition of Three Qubits. Three qubits are, collectively, the tensor product
space H H H, and the value of those qubits can be any tensor having unit length.
Vocabulary and Notation. Three qubits are often referred to as a tripartite
system.
To distinguish the three identical component spaces, we sometimes use subscripts,
HA H B H C .
The order of the tensor product, this time three, can used to label the state space:
H(3) .
11.8.3
Tripartite Bases and Separable States
The tensor product of three 2-D vector spaces has dimension 2 2 2 = 8. Its
inherited preferred basis tensors are the separable products of the component space
vectors,
n
|0i |0i |0i , |0i |0i |1i , |0i |1i |0i , |0i |1i |1i ,
o
|1i |0i |0i , |1i |0i |1i , |1i |1i |0i , |1i |1i |1i .
326
The shorthand alternatives are
|0i |0i |0i |0i |0i |0i |000i
|0i3
|0i |0i |1i |0i |0i |1i |001i
|1i3
|0i |1i |0i |0i |1i |0i |010i
|2i3
|0i |1i |1i |0i |1i |1i |011i
|3i3
|1i |0i |0i |1i |0i |0i |100i
|4i3
|1i |0i |1i |1i |0i |1i |101i
|5i3
|1i |1i |0i |1i |1i |0i |110i
|6i3
|1i |1i |1i |1i |1i |1i |111i
|7i3
The notation of the first two columns admits the possibility of labeling each of the
component kets with the H from which it came, A, B or C,
|0iA |0iB |0iC |0iA |0iB |0iC
|0iA |0iB |1iC |0iA |0iB |1iC
etc.
The densest of the notations expresses the CBS ket as an integer from 0 to 7. We
reinforce this correspondence and add the coordinate representation of each basis ket:
|000i
|0i3
1
0
0
0
0
0
0
0
|001i
|1i3
0
1
0
0
0
0
0
0
|010i
|2i3
0
0
1
0
0
0
0
0
|011i
|3i3
0
0
0
1
0
0
0
0
|100i
|4i3
0
0
1
0
1
0
0
0
|101i
|5i3
0
0
0
0
0
1
0
0
|110i
|6i3
0
0
1
0
0
0
1
0
|111i
|7i3
0
0
0
0
0
0
0
1
Note that that the exponent 3 is needed mainly in the encoded form, since an
integer representation for a CBS does not disclose its tensor order (3) to the reader,
while the other representations clearly reveal that the context is three-qubits.
The Channel Labels. We will use the same labeling scheme as before, but more
input lines means more labels. For three lines, we would name the registers A, B and
327
C, as in the hypothetical circuit
A-register in
A-register out
U
B-register in
B-register out
C-register in
C-register out
Working with Three Qubits
As I mentioned in the introduction, the current regime allows three or more inputs as
long as we only apply operators to two at-a-time. Lets look at a circuit that meets
that condition.
(
2
|i
H
|i
H
This circuit is receiving an order-3 tensor at its inputs. The first two registers, A and
B, get a (potentially) entangled bipartite state |i2 and the third, C, gets a single
qubit, |i. We analyze the circuit at the three access points, P , Q and R.
A Theoretical Approach. Well first do an example that is more general than
we normally need, but provides a surefire fallback technique if we are having a hard
time. Well give the input states the most general form,
|i =
and
|i
328
Access Point P. The initial tripartite tensor is
|i2 |i
Access Point Q. The first gate is a CNOT applied only to the entangled |i2 ,
so the overall effect is just
CNOT 1 |i2 |i
=
CNOT |i2 |i
by the rule for applying a separable operator to a separable state. (Although the first
two qubits are entangled, when the input is grouped as a tensor product of H(2) H,
|i2 |i is recognized as a separable second-order tensor.)
Applying the CNOT explicitly, we get
CNOT |i
|i
CNOT
We needed to multiply out the separable product in preparation for the next phase
of the circuit, which appears to operate on the last two registers, B and C.
Access Point R. The final operator is local to the last two registers, and takes
the form
1 H 2
Although it feels as though we might be able to take a short cut, the intermediate
tripartie state and final operator are sufficiently complicated to warrant treating it as
329
a fully entangled three-qubit state and just doing the big matrix multiplication.
1
1
1
1 0 0
0
0
1 1 1 1 0 0
0
0
1
1 1 1 0 0
0
0
h
i
1
1
1
1
1
0
0
0
0
2
1H
= 2 0 0
0
0
1
1
1
1
0 0
0
0 1 1 1 1
0 0
0
0 1
1 1 1
0 0
0
0 1 1 1 1
1
2
+
+
+
+
+
+
which could be written as the less compact basis expansion
1h
( + + + ) |000i + ( + ) |001i
2
+ ( + ) |010i + ( + ) |011i
+ ( + + + ) |100i + ( + ) |101i
i
+ ( + ) |110i + ( + ) |111i .
This isnt very enlightening because the coefficients are so general. But a concrete
example shows how the process can be streamlined.
A Practical Approach. Usually we have specific and nicely symmetric input
tensors. In this case, lets pretend we know that
|i2
|i
|00i + |11i
2
|1i .
330
and
For reference, Ill repeat the circuit for this specific input.
|00i + |11i
|1i
H
Access Point P. We begin in this state
|00i + |11i
|1i .
2
Access Point Q. Apply the two-qubit CNOT gate to registers A and B, which
has the overall effect of applying CNOT 1 to the full tripartite tensor.
CN OT 1
|i2 |i
=
CNOT |i2 1 |i
|00i + |11i
1 |1i
=
CNOT
2
CNOT |00i + CNOT |11i
=
|1i
2
|00i + |10i
=
|1i
2
=
|001i + |101i
,
2
a state that can be factored to our advantage:
|0i + |1i
|001i + |101i
=
|01i
2
2
Access Point R. Finally we apply the second two-qubit gate H 2 to the B and
C registers, which has the overall effect of applying 1 H 2 to the full tripartite
state. The factorization we found makes this an easy separable proposition,
h
i |0i + |1i
|0i + |1i
2
1H
|01i
=
H 2 |01i .
2
2
331
Referring back to the second order Hadamard on the two states in question, i.e.,
H 2 |01i
|00i |01i + |10i |11i
,
2
we find that
h
i |0i + |1i
1H
|01i
2
|00i |01i + |10i |11i
|0i + |1i
=
2
2
2
which can be factored into a fully separable
|0i + |1i
|0i |1i
|0i + |1i
.
2
2
2
On the other hand, if we had been keen enough to remember that H 2 is just the
separable H H, which has a special effect on a separable state like |01i = |0i |1i,
we could have gotten to the factored form faster using
h
i |0i + |1i
|0i + |1i
2
1H
|01i
=
H H |0i |1i
2
2
|0i + |1i
=
H |0i H |1i
2
|0i + |1i
|0i + |1i
|0i |1i
=
.
2
2
2
The moral is, dont worry about picking the wrong approach to a problem. If your
math is sound, youll get to the end zone either way.
Double Checking Our Work. Lets pause to see how this compares with the
general formula. Relative to the general |i2 and |i, the specific amplitudes we have
in this example are
1
= = ,
2
= 1
and
= = = 0
which causes the general expression
1h
( + + + ) |000i + ( + ) |001i
2
+ ( + ) |010i + ( + ) |011i
+ ( + + + ) |100i + ( + ) |101i
+ ( + ) |110i + ( + ) |111i
332
to reduce to
1h
2
() |000i + () |001i + () |010i + () |011i
+ () |100i + () |101i + () |110i + () |111i
1h 1
1
1
1
( |000i |001i + |010i |011i
=
2
2
2
2
2
i
1
1
1
1
+ |100i |101i + |110i |111i .
2
2
2
2
A short multiplication reveals this to be the answer we got for the input |00i+2|11i |1i,
without the general formula.
[Exercise. Verify that this is equal to the directly computed output state. Hint:
If you start with the first state we got, prior to the factorization, there is less to do.]
[Exercise. This had better be a normalized state as we started with unit vectors
and applied unitary gates. Confirm it.]
The Born Rule for Tripartite Systems
The Born rule can be generalized to any dimension and stated in many ways. For
now, lets state the rule for an order-three Hilbert space with registers A, B and C,
and in a way that favors factoring out the AB-registers.
Trait #15 (Born Rule for Tripartite States): If we have a tripartite state
that can be expressed as the sum of four terms
|i3
|0i2AB |0 iC + |1i2AB |1 iC + |2i2AB |2 iC + |3i2AB |3 iC ,
each of which is the product of a distinct CBS ket for HA HB and some general
first order (typically un-normalized) ket in the space HC ,
|ki2AB |k iC ,
then if we measure the first two registers, thus forcing their collapse into one of the
four basis states,
2
|0iAB , |1i2AB , |2i2AB , |3i2AB ,
the C register will be left in a normalized state associated with the measured CBS
333
ket. In other words,
|0 i
A B & |0i2
&
A B & |1i2
&
A B & |2i2
&
A B & |3i2
&
h0 | 0 i
|1 i
h1 | 1 i
,
,
|2 i
and
h2 | 2 i
|3 i
h3 | 3 i
Note the prime () in the tripartite Trait #15, to distinguish this from the unprimed Trait #15 for bipartite systems. Also, note that I suppressed the state-space
subscript labels A, B and C which are understood by context.
Well use this form of the Born rule for quantum teleportation in our next lecture.
11.8.4
End of Lesson
This was a landmark week, incorporating all the basic ideas of quantum computing.
We are now ready to study our first quantum algorithms.
334
Chapter 12
First Quantum Algorithms
12.1
Three Algorithms
This week we will see how quantum circuits and their associated algorithms can be
used to achieve results impossible in the world of classical digital computing. Well
cover three topics:
Superdense Coding
Quantum Teleportation
Deutschs Algorithm
The first two demonstrate quantum communication possibilities, and the third provides a learning framework for many quantum algorithms which execute faster (in
a sense) than their classical counterparts.
12.2
Superdense Coding
12.2.1
Sending Information by Qubit
|i
= |0i + |1i ,
seems like it holds an infinite amount of information. After all, and are complex
numbers, and even though you cant choose them arbitrarily (||2 + ||2 must be 1),
the mere fact that can be any complex number whose magnitude is 1 means it
could be an unending sequence of never-repeating digits, like 0.4193980022903 . . . .
If a sender A (an assistant quantum researcher named Alice) could pack |i with
that (and compatible ) and send it off in the form of a single photon to a receiver
B (another helper whose name is Bob) a few time zones away, A would be sending
an infinite string of digits to B encoded in that one sub-atomic particle.
335
The problem, of course, arises when the when B tries to look inside the received
state. All he can do is measure it once and only once (Trait #7, the fifth postulate
of QM), at which point he gets a 0 or 1 and both and are wiped off the face
of the Earth. That one measurement tells B very little.
[Exercise. But it does tell him something. What?]
In short, to communicate ||, A would have to prepare and send an infinite
number of identical states, then B would have to receive, test and record them. Only
then would B know || and || (although neither nor ). This is no better than
classical communication.
We have to lower our sights.
A Most Modest Wish
We are wondering what information, exactly, A (Alice) can send B (Bob) in the
form of a single qubit. We know its not infinite. At the other extreme is the most
modest super-classical capability we could hope for: two classical bits for the price
of one. I think that we need no lecture to affirm the claim that, in order to send a
two-digit binary message, i.e., one of
0
1
2
3
=
=
=
=
00 ,
01 ,
10 ,
11 ,
we would have to send more than one classical bit wed need two. Can we pack at
least this meager amount of classical information into one qubit with the confidence
that B would be able read the message?
12.2.2
The Superdense Coding Algorithm
We can, and to do so, we use the four Bell states (EPR pairs) from the two qubit
lecture,
|00 i
|00i + |11i
,
2
|01 i
|01i + |10i
,
2
|10 i
|00i |11i
|11 i
|01i |10i
.
2
336
and
Building the Communication Equipment
A and B prepare the state |00 i. (This can be done, for example, by sending a |00i
through the BELL gate,
|0i
)
|00 i ,
|0i
as we learned.) A takes the A register of the entangled state |00 i and B takes the
B register. B gets on a plane, placing his qubit in the overhead bin and travels a
few time zones away. This can all be done long before the classical two-bit message
is selected by A , but it has to be done. It can even be done by a third party who
sends the first qubit of this EPR pair to A and the second to B.
Defense of Your Objection. The sharing of this qubit does not constitute
sending more than one qubit of information (the phase yet to come), since it is
analogous to establishing a radio transmission protocol or message envelope, which
would have to be done even with classical bits. It is part of the equipment that B
and A use to communicate data, not the data itself.
Notation. In the few cases where we need it (and one is coming up), lets build
some notation. When a potentially entangled two-qubit state is separated physically
into two registers or by two observers, we need a way to talk about each individual
qubit. Well use
2
|i
for the A register (or A s) qubit, and
A
|i2
for the B register (or Bs) qubit.
Note that, unless |i2 happens to be separable and |00 i is clearly not we will be
faced with the reality that
|i2 6=
|i2 |i2 .
A
[Note. This does not mean that the A register and B register cant exist in physically
independent locations and be measured or processed independently by different observers. As we learned, one observer can modify or measure either qubit individually.
What it does mean is that the two registers are entangled so modifying or measuring
one will affect the other. Together they form a single state.]
With this language, the construction and distribution of each half of the entangled
|00 i to A and B can be symbolized by
|00 i A goes to A
|00 i
|00 i goes to B
B
337
A Encodes the Message
When A is ready to ship one of the four bit strings to B, she decides or is informed
which it is to be and takes the following action. She submits her half of the bipartite
state to one of four local gates according to the table
A to Send
A Applies
Equivalent Binary Gate
New Bipartite State
00
(nothing)
11
|00 i
01
X 1
|01 i
10
Z 1
|10 i
11
iY
iY 1
|11 i
(The 1s in the equivalent binary gate column reflect the fact that B is not touching his half of |00 i, which is effectively the identity operation as far as the B register
is concerned.)
And how do we know that the far right column is the result of A s local operation?
We apply the relevant matrix to |00 i and read off the answer (see section Four Bell
States from One in the two qubit lecture).
Compatibility Note. Most authors ask A to apply Z if she wants to encode
01 and X if she wants to encode 10 , but doing so results in the state |10 i for
01 and |01 i for 10 , not a very nice match-up and is why I chose to present the algorithm with those two gates swapped. Of course, it really doest matter which of the
four operators A uses for each encoding, as long as B uses the same correspondence
to decode.
If we encapsulate the four possible operators into one symbol, SD (for Super
Dense), which takes on the proper operation based on the message to be encoded,
A s job is to apply the local circuit,
h
i
( SD 1 ) |00 i
|00 i
SD
A
A
Notice that to describe A s half of the output state, we need to first show the full
effect of the bipartite operator and only then restrict
attention to A s qubit. We
cannot express it as a function of A s input, |00 i , alone.
A
A Sends Her Qubit
The message is now encoded in the bipartite state, but for B to decode it, he needs
both qubits. A now sends her qubit to B.
338
i
( SD1 )|00 i
A
B Measures Along the Bell Basis to Read the Message
When he gets A s qubit, B has the complete entangled bipartite state
h
i
( SD 1 ) |00 i
( SD 1 ) |00 i ,
h
i
( SD 1 ) |00 i
so he can now measure both qubits to determine which of the four Bell states he has.
Once thats done he reads the earlier table from right-to-left to recover the classical
two-bit message.
Refresher: Measuring Along the Bell Basis. Since this is the first time we
will have applied it in an algorithm, Ill summarize one way that B can measure
his entangled state along the Bell basis. When studying two qubit logic we learned
that to measure a bipartite state along a non-standard basis (call it C), we find the
binary operator that takes the z-basis to the other basis, call it S, and use S prior
to measurement:
)
(
S
(some C-basis state)
measure along z-basis
In this situation, S is BELL, whose adjoint we computed in that same lecture,
BELL
H
.
Adding the measurement symbols (the meters) along the z-basis, circuit becomes
(
(one of the four BELL states)
339
BELL
In terms of matrices, B subjects his two-qubit state to the matrix for BELL (also
computed last time),
1 0 0
1
1 0 1 1
0
.
BELL =
2 1 0 0 1
0 1 1 0
Bobs Action and Conclusion. Post-processing with the BELL gate turns
the four Bell states into four z-CBS kets; if B follows that gate with a z-basis measurement and sees a 01 , he will conclude that he had received the Bell state |01 i
from A , and likewise for the other states. So his role, after receiving the qubit sent
by A , is to
1. apply BELL to his two qubits, and
2. read the encoded message according to his results using the table
B measures
B Concludes A s Message to him is
00
00
01
01
10
10
11
11
In other words, the application of the BELL gate allowed B to interpret his z-basis
measurement reading xy as the message, itself.
The following exercise should help crystallize the algorithm.
[Exercise. Assume A wants to send the message 11.
i) Apply iY 1 to |00 i and confirm that you get |11 i out.
ii) Multiply the 4 4 matrix for BELL by the 4 1 state vector for |11 i to show
that B recovers the message 11.
A Circuit Representation of Superdense Coding
We can get a circuit for the overall superdense coding algorithm by adding some new
notation.
Classical Wires. Double lines (=) indicate the transfer of classical bits. We use
them to move one or more ordinary digits within a circuit.
340
Decisions Based on Classical Bits. We insert a dot symbol, , into a classical
line to indicate a general controlled operation, based on the content of that classical
data ([1] means apply an operator, [0] means dont).
Noiseless Transmission Between Communicators. To indicate the (typically radio) transmission of either quantum or classical data between sender and
receiver, we use the wavy line,
.
With this notation, the superdense coding circuit can be expressed as:
A-reg: |00 i
|xi
SD
B-reg: |00 i
BELL
|yi
B
The notation tells the story. A uses her two-bit classical message xy (traveling on
the double lines) to control (filled circles) which of the four operations (SD = 1, X,
Z or iY ) she will apply to her qubit. After sending her qubit to B, B measures both
qubits along the Bell basis to recover the message xy now sitting in the output
registers in natural z-basis form |xi |yi.
[Exercise. Measurement involves collapse and uncertainty. Why is B so certain
that his two measurements will always result in a true reproduction of the message
xy sent by A ? Hint: For each of the four possible messages, what bipartite state
is he holding at the moment of measurement?]
This can actually be tightened up. Youve seen several unary operator identities
in the single qubit lecture, one of which was XZ = iY . A slight revision of this
(verify as an exercise) is
ZX
iY ,
which enables us to define the elusive SD operation: we place a controlled-X gate and
controlled-Z gate in the A-channel under A s supervision. Each gate is controlled
by one of the two classical bits in her message. They work just like a quantum
Controlled-U gate, only simpler: if the classical control bit is 1, the target operation
is applied, if the bit is 0, it is not.
A-reg: |00 i
|xi
X
Z
B
BELL
B-reg: |00 i
|yi
341
For example, if both bits are 1, both gates get applied and result in in the desired
behavior: 11 ZX
= iY .
[Exercise. Remind us why the gates X and Z appear reversed in the circuit
relative to the algebraic identity iY = ZX.]
The Significance of Superdense Coding
This technique may not seem tremendously applicable considering its unimpressive
2-bit to 1-bit compression, but consider sending a large classical message, even one
that is already as densely compressed as classical logic will allow. This is a 2-to-1
improvement over the best classical technique when applied to the output of classical
compression. The fact that we have to send lots of entangled Bell states before our
message takes nothing away from our ability to send information in half the time (or
space) as before.
12.3
Quantum Teleportation
We re-enlist the help of our two most excellent researchers, Alice and Bob, and continue to refer to them by their code names A and B.
In superdense coding A sent B one qubit, |i = |0i + |1i, in order to reconstruct two classical bits. Quantum teleportation is the mirror image of this process. A
wants to send B the qubit, |i, by sending him just two classical bits of information.
Giving Teleportation Context
You might ask why she doesnt simply send B the one qubit and be done with it.
Why be so indirect and translate the quantum information into classical bits? There
are many answers, two of which I think are important.
1. As a practical matter, it may be impossible, or at least difficult, for A to
send B qubit information due to its instability and/or expense. By contrast,
humans have engineered highly reliable and economical classical communication
channels. Sending two bits is childs play.
2. Sending the original qubit rather than two classical bits is somewhat beside the
point. The very fact A can get the infinitely precise data embedded in the
continuous scalars and by sending something as crude as an integer from 0
to 3 should come as unexpectedly marvelous news, and we want to know why
and how this can be done.
342
Caveats
There is the usual caveat. Just because B gets the qubit doesnt mean he can know
what it is. He can no more examine its basis coefficients than A (or anyone in her
local lab who didnt already know their values) could. What we are doing here is
getting the qubit over to Bs lab so he can use it on his end for any purpose that A
could have (before the teleportation).
And then theres the unusual caveat. In the process of executing the teleportation,
A loses her copy of |i. Well see why as we describe the algorithm.
An Application of the Born Rule
I like to take any reasonable opportunity to restate important tools so as to establish
them in your mind. The Born rule is so frequently used that it warrants such a review
before we apply it to teleportation.
The Born rule for 3-qubit systems (Trait #15) tells us (and I am paraphrasing
with equally precise expressions) that if we have a tripartite state which can be
expressed as the sum of four terms (where the first factors of each term are AB-basis
kets):
|i3
|0i2AB |0 iC + |1i2AB |1 iC + |2i2AB |2 iC + |3i2AB |3 iC ,
then an AB-register measurement along the natural basis will force the corresponding
C-register collapse according to
A B & |0i2AB
&
|0 iC
,
k |0 iC k
A B & |1i2AB
&
|1 iC
,
k |1 iC k
etc.
There are two consequences that will prepare us for understanding quantum teleportation as well as anticipating other algorithms that might employ this special
technique.
Consequence #1. The rule works for any orthonormal basis in channels A and
B, not just the natural basis. Whichever basis we choose for the first two registers A
and B, it is along thatbasis that
we must make our two-qubit measurements. So, if
we use the Bell basis, |jk i , then a state in the form
|i3
|00 iAB |00 iC + |01 iAB |01 iC
+ |10 iAB |10 iC + |11 iAB |11 iC ,
343
when measured along that basis will force the corresponding C-register collapse according to
|00 iC
A B & |00 iAB
=
C &
,
k |00 iC k
A B & |01 iAB
&
|01 iC
,
k |01 iC k
etc.
This follows from the Trait #7, Post-Measurement Collapse, which tells us that AB
will collapse to one of the four CBS states regardless of which CBS we use forcing
C into the state that is glued to its partner in the above expansion.
p
The division by each k |jk i k (or, if you prefer, hjk | jk i ) is necessary because
the overall tripartite state, |i3 can only be normalized when the |jk i have non-unit
(in fact < 1) lengths.
[Exercise. We already know that |jk i are four normalized CBS kets. Show that
if the |jk i were normal vectors in HC , then |i3 would not be a normal vector. Hint:
Write down 3 h | i3 and apply orthonormality of the Bell states.]
Consequence #2. If we know that the four general states, |jk i, are just four
variations of a single known state, we may be able to glean even more specific information about the collapsed C-register. To cite the example needed today, say we
know that all four |jk i use the same two scalar coordinates, and , only in slightly
different combinations,
|0iC + |1iC
|0iC + |1iC
3
+ |01 iAB
|i
= |00 iAB
2
2
|0iC |1iC
|0iC + |1iC
+ |10 iAB
+ |11 iAB
.
2
2
(Each denominator 2 is needed to produce a normal state |i3 ; we cannot absorb
it into and , as those scalars are fixed by the normalized |i to be teleported.
However, the Born rule tells us that the collapse of the C-register will get rid of this
factor, leaving only one of the four numerators in the C-register.) Such a happy stateof-affairs will allow us to convert any of the four collapsed states in the C-register to
the one state,
|0i + |1i
by mere application of a simple unary operator. For example, if we find that AB
collapses to |00 i (by reading a 00 on our measuring apparatus), then C will have
already collapsed to the state |0i + |1i. Or, if AB collapses to |11 i (meter
reads11 ), then we apply the operator iY to C to recover |0i + |1i, because
0 1
iY |0i + |1i
=
=
.
1 0
Youll refer back to these two facts as we unroll the quantum teleportation algorithm.
344
12.3.1
The Quantum Teleportation Algorithm
We continue to exploit the EPR pairs which I list again for quick reference:
|00 i
|00i + |11i
,
2
|01 i
|01i + |10i
,
2
|10 i
|00i |11i
|11 i
|01i |10i
.
2
and
Building the Communication Equipment
A and B prepare and each get one qubit of the bipartite state |00 i.
A Produces a General Qubit for Teleportation
When A is ready to teleport a qubit to B, she manually produces or is given any
H-space qubit
|i
|iC
|0iC + |1iC .
The subscript C indicates that we have a qubit separate from the two entangled
qubits already created and distributed to our two messagers, a qubit which lives in
its own space with its own (natural) CBS basis { |0iC , |1iC }.
By tradition for this algorithm, we place the C-channel above the A/B-Channels:
|iC
register C
(
|00 iAB
register A
register B
The algebraic state of the entire system, then, starts out to be
|0i |0i + |1i |1i
3
A
B
A
B
.
|i
= |iC |00 iAB =
|0iC + |1iC
2
The Plan
A starts by teleporting a qubit to B. No information is actually sent. Rather, A does
something
to her entangled qubit, |00 i A , which instantaneously modifies Bs qubit,
|00 i B , faster than the speed of light. This is the meaning of the word teleportation.
345
She then follows that up by taking a measurement of her two qubits, getting two
classical bits of information the outcomes of the two register readings. Finally, she
sends the result of that measurement as a classical two-bit message to B (sorry, we
have to obey the Einsteins speed limit for this part). B will use the two classical bits
he receives from Alice to tweak his qubit (already modified by Alices teleportation)
into the desired state, |i.
A Expresses the System State in the Bell Basis (No Action Yet)
In the z-basis, all the information about |i is contained in A s C-register. She
wants to move that information over to Bs B-register. Before she even does anything
physical, she can accomplish most of the hard work by just rearranging the tripartite
state |i3 in a factored form expanded along a CA Bell -basis rather than a CA
z-basis. In other words, wed like to see
|i3
?
=
|00 iCA |00 iB + |01 iCA |01 iB
+ |10 iCA |10 iB + |11 iCA |11 iB ,
where the |jk iB are (for the moment) four unknown B-channel states. We can
only arrive at such a CA Bell basis expression if the two channels A and C become
entangled, which they are not, initially. Well get to that.
In our short review of the Born rule, above, I gave you a preview of the actual
expression well need. This is what we would like/wish/hope for:
|0iB + |1iB
|0iB + |1iB
?
3
+ |01 iCA
|i
= |00 iCA
2
2
|0iB |1iB
|0iB + |1iB
+ |10 iCA
+ |11 iCA
.
2
2
Indeed, if we could accomplish that, then A would only have to measure her two
qubits along the Bell basis, forcing a collapse into one of the four Bell states and
by the Born rule collapsing Bs register into the one of his four matching states. A
glance at the above expressions reveals that this gets us is 99.99% of the way toward
placing |i into Bs B-register, i.e., manufacturing |iB , a teleported twin to Alices
original |iA . Well see how B gets the last .01% of the way there, but first, we prove
the validity of the hoped-for expansion.
We begin with the desired expression and reduce it to the expression we know
to be our actual starting point, |i3 . (Warning: After the first expression, Ill be
346
dropping the state-space subscripts A/B/C and letting position do the job.)
|0iB + |1iB
|0iB + |1iB
|00 iCA
+ |01 iCA
2
2
|0iB |1iB
|0iB + |1iB
+ |10 iCA
+ |11 iCA
2
2
|0i + |1i
|01i + |10i |0i + |1i
=
+
2
2
2
|01i |10i |0i + |1i
|00i |11i |0i |1i
+
+
2
2
2
2
|00i + |11i
1
2 2
|000i + |110i + |001i + |111i
+ |010i + |100i + |011i + |101i
+ |000i |110i |001i + |111i
|010i + |100i + |011i |101i
Half the terms cancel and the other half reinforce to give
|0iB + |1iB
|0iB + |1iB
|00 iCA
+ |01 iCA
2
2
|0iB |1iB
|0iB + |1iB
+ |10 iCA
+ |11 iCA
2
2
=
=
=
1
2 |000i + 2 |100i + 2 |011i + 2 |111i
2 2
1
1
|0i + |1i |00i + |0i + |1i |11i
2
2
|00i + |11i
( |0i + |1i)
2
|iC |00 iAB ,
a happy ending. This was A s original formulation of the tripartite state in terms of
the z-basis, so it is indeed the same as the Bell expansion we were hoping for.
Next, we take action to make use of this alternate formulation of our system state.
(Remember, we havent actually done anything yet.)
347
A Measures the Registers CA Along the Bell Basis
The last derivation demonstrated that
|iC |00 iAB
=
|00 iCA
+ |10 iCA
|0iB + |1iB
|0iB + |1iB
+ |01 iCA
2
2
|0iB |1iB
|0iB + |1iB
+ |11 iCA
,
2
2
that is, the to-be-teleported |i in Alicess C-register seems to shows up (in a
modified form) in Bs B-register without anyone having taken any action all we
did was rearrange the terms. However, this rearrangement is only valid if A intends
to measure the AC-register along the BELL basis. Such a measurement, as we have
seen, always has two parts.
1. A applies a BELL gate to her AC-registers (the operator that takes the nonstandard basis to the z-basis).
2. A measures the resultant AC registers in a standard z-basis.
The first of these two parts, which effects the instantaneous transfer of |i from the
A-channel to the B-channel, corresponds to the teleportation step of the plan. The
second part is where A s measurement selects one of the four near-|is for Bs
B-register.
The circuit A needs is
register-C
BELL
register-A
or more explicitly,
register-C
H
.
register-A
348
After applying the gate (but before the measurement), the original tripartite state,
|i3 , will be transformed to
|0iB + |1iB
3
BELL 1 |i
= BELL |00 iCA 1
2
|0iB + |1iB
+ BELL |01 iCA 1
2
|0iB |1iB
+ BELL |10 iCA 1
2
|0iB + |1iB
+ BELL |11 iCA 1
2
|0iB + |1iB
|0iB + |1iB
= |00iCA
+ |01iCA
2
2
|0iB |1iB
|0iB + |1iB
+ |10iCA
+ |11iCA
,
2
2
Now when A measures her two qubits along the z-basis, she will actually be measuring
the pre-gate state along Bell basis. After the measurement only one of the four terms
will remain (by collapse) and B will have a near-|i left in his B-register.
The circuit that describes A s local Bell basis measurement with Bs qubit going
along for the ride is
|i
|00 i
A
|00 i
B
where the final meters are, as always, natural z-basis measurements.
One Qubit of an Entangled Pair
In case you hadnt noticed, twice today weve seen something that might seem to be
at odds with our previous lessons. Im talking about a single, entangled, qubit being
fed into one channel of a binary quantum gate, like
|i
|00 i
349
or
|00 i
|00 i
g
BELL
Earlier, I admonished you to not expect to see or be allowed to write a nonseparable bipartite state broken into two parts, each going into the individual channels
of a binary gate. Rather we need to consider it as a non-separable entity going into
both channels at once, as in:
(
|i2
U |i2
However, the new notation that I have provided today, one half of an entangled qubit,
|i2
for the A register (or A s) qubit, and
A
|i2
for the B register (or Bs) qubit,.
allows us to write these symbols as individual inputs into either input of a binary
quantum gate without violating the cautionary note. Why? Because earlier, the
separate inputs we disallowed were individual components of a separable tensor (when
no such separable tensor existed). We were saying that you cannot mentally place a
tensor symbol, , between the two individual inputs. Here, the individual symbols
are not elements in the two component spaces, and there is no danger of treating
them as separable components of a bipartite state, and no is implied.
A Sends Her Measurement Results to B
A now has a two (classical) bit result of her measurement: xy = 00 , 01 , 10
or ,11. She sends xy to B through a classical channel, which takes time to get
there.
350
B Uses the Received Message to Extract |i from His Qubit
Bs qubit is in one of the four collapsed states
|0iB
|0iB
|0iB
|0iB
+
+
|1iB
|1iB
|1iB
|1iB .
This happened as a result of A s Bell basis measurement (instantaneous faster-thanlight speed transfer, ergo teleportation). Thats the 99.99% I spoke of earlier. To
get the final .01% of the way there, he needs to look at the two classical bits he
received (which took time to reach him). They tell him which of those four states
his qubit landed in. If its anything other than 00 he needs to apply a local unary
operator to his B-register to fix it up, so it will be in the original |i. The rule is
B Receives
00
01
10
11
B Applies
(nothing)
X
Z
iY
B Recovers
|i
|i
|i
|i
Nothing helps make an abstract description concrete like doing a calculation, so I
recommend an . . .
[Exercise.
i) Express the operator BELL 1 as an 8 8 matrix with the help of the section
The Matrix of a Separable Operator, in the lesson on tensor products.
ii) Express the state |i3 = |0i + |1i |00 i as an 8 1 column vector
by multiplying it out (use this initial state description, not the re-arranged
version that used the Bell basis).
iii) Multiply the 8 8 operator matrix by the 8 1 state vector to get A s output
state (prior to measurement) in column vector form.
iv) Expand the last answer along the z-basis.
v) Factor that last result in such a way that it looks the same as the answer we
got when we applied BELL 1 to the re-arranged version of |i3 . ]
vi) In the case where B receives the classical message 10 from A , apply the
corresponding fix-it-up operator shown in the table to his collapsed qubit and
thereby prove that he recovers the exact teleported state |i = |0i + |1i.
351
A Circuit Representation of Quantum Teleportation
Just as we used SD to be one of four possible operators in the superdense coding
algorithm, we will use QT to mean one of four operators (1, X, Z, iY ) that B must
apply to his register based on the message he receives from A . With this shorthand,
the quantum teleportation circuit can be expressed as:
|i
|00 i
|00 i
QT
|i
The circuit says that after taking the measurements (the meter symbols), A radios
the classical data (double lines and wavy lines) to B who uses it to control (filled
circles) which of the four operations he will apply to his qubit.
Once again, we use the identity
ZX
iY
to more precisely define the QT operation. We place a controlled-X and controlled-Z
gates into into Bs pipline to get an improved circuit description.
|i
|00 i
|00 i
|i
(Dont forget that operators are applied from left-to-right in circuits, but right-to-left
in algebra.)
Many authors go a step further and add the initial gate that creates the ABchannel Bell state |00 i from CBS kets:
|0i
)
|00 i ,
|0i
352
which leads to
C: |i
A: |0i
B: |0i
|i
Here is the circuit with access points marked for an exercise:
C: |i
A: |0i
B: |0i
|i
Observe that the tripartite state
|i3
|iC |0iA |0iB
going into the entire circuit is transformed by various gates and measurements along
the way. It continues to exist as a tripartite state to the very end, but you may not
recognize it as such due to the classical wires and transmission of classical information
around access points R and S, seemingly halting the qubit flow to their right. Yet
the full order-3 state lives on. It is simply unnecessary to show the full state beyond
that point, because registers C and A, after collapse, will contain one of the four CBS
kets, |xiC |yiA , for xy = 00, 01, 10 or 11. But those two registers never change after
the measurement, and when Bob applies a measurement to his local register B, say
iY perhaps, he will be implicitly applying the separable operator 1 1 iY to the
full separable tripartite state.
[Exercise. Using natural coordinates for everything, compute the state of the
vector |i3 as it travels through the access points, P-U: |i3P , |i3Q , |i3R , |i3S ,
|i3T and |i3U . For points S, T and U you will have to know what measurement A
reads and sends to B, so do those three points twice, once for a reading of CA =01
and once for a reading of CA =11 . HINT: Starting with the easy point P, apply
transformations carefully to the basis kets using separable notation like (1 BELL)
or (BELL 1). When you get to post-measurement classical pipes, apply the Born
Rule which will select exactly one term in the sum. ]
353
Why Teleportation Works.
Consider the main steps in the teleportation algorithm. We begin with three channels,
the first of which contains all the quantum information we want to teleport, and the
last two none of it,
|i3
|iC |00 iAB .
Once we have the idea to entangle channels A and C by converting to the Bell basis
(perhaps driven by the fact that one of the Bell states is in the AB register pair) we
end up with a state in the general form,
|i3
|00 iCA |00 iB + |01 iCA |01 iB
+ |10 iCA |10 iB + |11 iCA |11 iB .
Without even looking at any of four |jk i kets in the B-channel, we are convinced
that 100% of the |i information is now sitting inside that register, waiting to be
tapped. Why?
The reason is actually quite simple.
Quantum gates including basis transformations are always unitary and thus
reversible. If the Bell-basis operator had failed to transfer all the |i information into
the B-register, then, since none of it is left in the AC-registers (they contain only
Bell states), there would be no hope of getting an inverse gate to recover our starting
state which holds the full |i. Thus, producing an expression that left channels A
and C bereft of any a trace of |i information must necessarily produce a B-channel
contains it all.
12.4
Introduction to Quantum Oracles
Superdense coding and quantum teleportation may seem more like applications of
quantum computation than quantum algorithms. They enable us to transmit data
in a way that is not possible using classical engineering. In contrast, Deutsch s
little problem really feels algorithmic in nature and, indeed, its solution provides the
template for the many quantum algorithms that succeed it.
We take a short side-trip to introduce Boolean functions then construct our first
quantum oracles used in Deutschs and later algorithms.
12.4.1
Boolean Functions and Reversibility
Classical unary gates of which there are a grand total of four contain both reversible operators (1 and ) and irreversible ones (the [0]-op and [1]-op). Quantum
operators require reversibility due to their unitarity, so there are no quantum analogs
354
for the latter two. Binary gates provide even more examples of irreversible classical
operations for which there are no quantum counterparts.
These are specific examples of a general phenomenon that is more easily expressed
in terms of classical Boolean functions.
Boolean Functions. A Boolean function is a function that has one or more
binary digits (0 or 1) as input, and one binary digit as output.
From Unary Gates to Boolean Functions
A classical unary gate takes a single classical bit in and produces a single classical
bit out. In the language of functions, it is nothing other than a Boolean function of
one bit, i.e.,
f : { 0, 1 }
{ 0, 1 } ,
or using the notation B { 0, 1 } introduced in the single qubit lecture,
f : B B .
(As we defined it, B had a richer structure than that of a simple set; it had a mod-2
addition operation that we will find useful in the definitions and computations to
come.)
Reversible Example. To avoid becoming unmoored by abstractions, we revisit
the negation operator in the language of Boolean functions. If we define
f (x) x,
for x B ,
then in terms of a truth table, the Boolean function f is
x
f (x)
This f is reversible, and in fact is its own inverse, since f (f (x)) = x.
Irreversible Example. On the other hand, if we define
g(x) 1,
for x B ,
with a truth table of
x
g(x)
1
355
we have the quintessential example of an irreversible function.
Boolean functions (classical implied) will now become the subject of study. Computer science seeks to answer questions about such functions or create Boolean functions that do useful things. On the other hand, we will be using quantum circuits
composed of quantum operators to answer questions about these classical Boolean
functions. In other words, we have not abandoned the classical functions (Ill drop
the modifier Boolean for now) in the least. On the contrary: they are the principle
players in our narrative.
Binary Gates as Boolean Functions
The language naturally extends to functions of more than one input bit. To keep
things simple, lets talk about two bits.
A two bit function (classical and Boolean implied) takes two bits in and produces
one bit out. In other words,
0
0
1
1
f:
,
,
,
{ 0, 1 } ,
0
1
0
1
or in B notation,
f : B2 B .
(Column vs. row is not important here, so Ill use whichever fits better into the written
page without the ()t baggage.)
Note that we are avoiding the term binary, replacing it instead with two bit
to avoid confusion arising from the fact that we are using binary digits for every input
slot, whether a unary input or a multi-bit input.
Irreversibility in the Two (or Greater) Bit Case. Since two-input Boolean
functions functions, like all Boolean functions, have a single bit out, they are inherently irreversible; we cannot undo the destruction that results from the loss of one or
more bits.
(Necessarily) Irreversible Example. A typical two bit function that, like all
two+ bit functions, is necessarily irreversible is the XOR, i.e.,
x
f (x, y) x y, for
B2 ,
y
with the truth table
(x, y)
f (x, y)
(0, 0)
(0, 1)
(1, 0)
(1, 1)
0
356
12.4.2
The Quantum Oracle of a Boolean Function
Although our quantum algorithms will use quantum gates, they will often have to
incorporate the classical functions that are the center of our investigations. But
how can we do this when all quantum circuits are required to use unitary and
therefore reversible gates? There is a well known classical technique for turning an
otherwise irreversible function into one that is reversible. The technique pre-dates
quantum computing, but well look at it only in the quantum context, and if youre
interested in the classical analog, you can mentally down-convert ours by ignoring
its superposition capability and focus only on the CBS inputs.
Oracles for Unary Functions
Suppose we are given a black box that computes some unary function, f (x), even one
that may be initially unknown to us. The term black box suggests that we dont know
whats on the inside or how it works.
x
f (x)
It can be shown that, using this black box along with certain fundamental quantum
gates one can build a new gate that
takes two bits in,
has two bits out,
is unitary (and therefore reversible),
computes the function f when presented the proper input, and
does so with the same efficiency (technically, the same computational complexity,
a term we will define in a later lesson), as the black box f , whose irreversible
function we want to reproduce.
We wont describe how this works but, instead, take it as a given and call the new,
larger circuit Uf , the quantum oracle for f . Its action on CBS kets and its circuit
diagram are defined by
Data register:
|xi
|xi
,
Uf
Target register: |yi
|y f (x)i
which also gives the name data register to the A-channel and target register to the
B-channel.
First, notice that the output of the target register is a CBS; inside the ket we
are XOR-ing two classical binary values y and f (x), producing another binary value
which, in turn, defines a CBS ket: either |0i or |1i.
357
Example. We compute the matrix for Uf when f (x) = 0, the constant (and
irreversible) [0]-op. Starting with the construction of the matrix of any linear transformation and moving on from there,
Uf |00i , Uf |01i , Uf |10i , Uf |11i
Uf =
|0i | 0 f (0) i , |0i | 1 f (0) i , |1i | 0 f (1) i , |1i | 1 f (1) i
=
=
|0i |f (0)i , |0i | f (0) i , |1i |f (1)i , |1i | f (1) i ,
where we use the alternate notation for negation, a = a. So far, everything we did
applies to the quantum oracle for any function f , so well put a pin in it for future
use. Now, going on to apply it to f = [0]-op,
|0i |0i , |0i |1i , |1i |0i , |1i |1i
U[0]-op =
1 0 0 0
0 1 0 0
=
0 0 1 0 ,
0 0 0 1
an interesting result in its own right, U[0]-op = 1, but nothing to which we should
attribute any deep meaning. Do note, however, that such a nice result makes it selfevident that Uf is not only unitary but its own inverse, as we show next it always
will be.
Uf is Always its Own Inverse. We compute on the tensor CBS, and the result
will be extensible to the entire H H by linearity:
Uf Uf |xyi = Uf Uf |xyi
= Uf |xi | y f (x) i
E
= |xi y f (x) f (x)
E
= |xi |yi = |xyi QED
= |xi y f (x) f (x)
[Exercise. Why is f (x) f (x) = 0?]
Uf Computes f (x). This is a simple consequence of the circuit definition,
because if we plug 0 y, we get
|xi
|xi
.
Uf
|0i
|f (x)i
[Exercise. What do we get if we plug 1 y?]
[Exercise. Compute the quantum oracles for the other three unary functions
and observe that their matrices reveal them each to be unitary, not to mention self
inverses.]
358
Notice that output of Uf for a CBS is always a separable state
|xi | y f (x) i ,
since y f (x) is always either 0 or 1.
[Exercise. CBS input kets producing separable (even CBS) output kets do not a
separable operator make. Prove this by looking at the three (Uf )s you computed in
the last exercise, and finding two which are patently non-separable.]
Oracles for Functions of Two Boolean Inputs
Everything extends smoothly to more than one-bit gates, so we only need outline
the analysis for two-qubits. We are given black box of a two-input Boolean function,
f (x0 , x1 ),
x0
f
f (x0 , x1 )
x1
This time, we assume that circuit theory enables us to build a three-in, three-out
oracle, Uf , defined by
|x0 i
|x0 i
,
|x1 i
|x1 i
Uf
|yi
| y f (x0 , x1 ) i
usually shortened by using the encoded form of the CBS kets, |xi2 , where x
{ 0, 1, 2, 3 },
|xi2
|xi2
.
Uf
|yi
|y f (x)i
The key points are the same:
Uf is its own inverse.
Uf emulates f (x) by setting y = 0,
Uf |xi2 |0i
|xi2 |f (x)i .
[Exercise. Compute the quantum oracle (in matrix form) for the classical AND
gate.]
359
12.5
Deutschs Problem
Our first quantum algorithm answers a question about an unknown unary function
f (x). It does not find the exact form of this function, but seeks only to answer a
general question about its character. Specifically, we ask whether the function is
one-to-one (distinct inputs produce distinct outputs) or constant (both inputs are
mapped to the same output.)
Obviously, we can figure this out by evaluating both f (0) and f (1), after which
we would know the answer, not to mention have a complete description of f . But
the point is to see what we can learn about f without doing both evaluations of the
function; we only want to do one evaluation. In a classical world if we only get to
query f once we have to choose between inputs 0 or 1, and getting the output for our
choice will not tell us whether the function is one-to-one or constant.
All the massive machinery we have accumulated in the past weeks can be brought
to bear on this simple problem very neatly to demonstrate how quantum parallelism
will beat classical computing in certain problems. It will set the stage for all quantum
algorithms.
12.5.1
Definitions and Statement of the Problem
For this and a subsequent algorithm, we will define a property that a Boolean function
might (or might not) have. We continue to assume that function means Boolean
function.
Balanced and Constant Functions
Balanced Function. A balanced function is one that takes on the value 0 for exactly
half of the possible inputs (and therefore 1 on the other half).
Two examples of balanced functions of two inputs are and XOR and 1y : (x, y) 7
y:
(x, y)
XOR(x, y)
(x, y)
1y (x, y)
(0, 0)
(0, 0)
(0, 1)
(0, 1)
(1, 0)
(1, 0)
(1, 1)
(1, 1)
Two unbalanced function of two inputs are AND and the [1]-op:
360
(x, y)
AN D(x, y)
(x, y)
[1](x, y)
(0, 0)
(0, 0)
(0, 1)
(0, 1)
(1, 0)
(1, 0)
(1, 1)
(1, 1)
Constant Functions. Constant functions are functions that always produce the
same output regardless of the input. There are only two constant functions for any
number of inputs: either the [0]-op or the [1]-op. See the truth table for the [1]-op,
above; the truth table for the [0]-op would, of course, have 0s in the right column
instead of 1s.
Balanced and Constant Unary Function
There are only four unary functions. Therefore the terms balanced and constant might
seem heavy handed. The two constant functions are obviously the [0]-op or the [1]-op,
and the other two are balanced. In fact, the balanced unary functions already have
a term that describes them: one-to-one. Theres even a simpler term in balanced
functions in the unary case: not constant. To see this, lets lay all of our cards on
the table, pun intended.
x
[0]
[1]
So exactly two of our unary ops are constant and the other two are balanced = oneto-one = not constant.
The reason we complicate things by adding the vocabulary constant vs. balanced
is that we will eventually move on to functions of more than one input, and in those
cases,
not all functions will be either balanced or one-to-one (e.g., the binary AND
function isnt either), and
balanced functions will not be one-to-one (e.g., binary XOR function is balanced
but not one-to-one)
Deutschs Problem
We are now ready to state Deutschs problem using vocabulary that will help when
we go to higher-input functions.
361
Deutschs Problem. Given an unknown unary function that we are told is either
balanced or constant, determine which it is in one query of the quantum oracle,
Uf .
Notice that we are not asking to determine the exact function, just which category
it belongs to. Even so, we cannot do it classically in a single query.
12.5.2
Deutschs Algorithm
The algorithm consists of building a circuit and measuring the A-register once.
Thats it. Our conclusion about f is determined by the result of the measurement:
if we get a 0 the function is constant, if we get 1 the function is balanced. We
will have gotten an answer about f with only one query of the oracle and thereby
obtained a 2 improvement over a classical algorithm.
The Circuit
We combine the quantum oracle for f with a few Hadamard gates in a very small
circuit:
|0i
H
Uf
|1i
(ignore)
Because there are only four unary functions, the temptation is to simply plug each
one into Uf and confirm our claim. Thats not a bad exercise (which Ill ask you to
do), but lets understand how one arrives at this design so we can use the ideas in
other algorithms.
The Main Ideas Behind Deutschs Solution
Classical computing is embedded in quantum computing when we restrict our attention to the finite number of CBS kets that swim around in the infinite quantum
ocean of the full state space. For the simplest Hilbert space imaginable the firstorder space, H = H(1) those CBS kets consist of the two natural basis vectors
{ |0i , |1i }, corresponding to the classical bits [0] and [1]. We should expect that no
improvements to classical computing can be achieved by using only z-basis states (i.e.,
the CBS) for any algorithm or circuit. Doing so would be using our Hilbert space
as though it were the finite set { |0i , |1i }, a state of affairs that does nothing but
simulate the classical world.
There are two quantum techniques that motivate the algorithm.
#1: Quantum Parallelism. This is the most general of the two ideas and will
be used in all our algorithms. Any non-trivial superposition of the two CBS kets
|0i
|1i ,
both and 6= 0 ,
362
takes us off this classical plane into quantum hyperspace where all the fun happens.
When we send such a non-trivial superposition through the quantum oracle, we are
implicitly processing both z-basis kets and therefore both classical states, [0] and
[1] simultaneously. This is the first big idea that fuels quantum computing and
explains how it achieves its speed improvements. (The second big idea is quantum
entanglement, but well feature that one a little later.)
The practical impact of this technique in Deutschs algorithm is that well be
sending a perfectly balanced (or maximally mixed ) superposition,
1
1
|0ix = |0i + |1i ,
2
2
through the data register (the A-channel) of the oracle, Uf .
#2: The Phase Kick-Back Trick. This isnt quite as generally applicable as
quantum parallelism, but it plays a role in several algorithms including some well
meet later in the course. It goes like this. If we feed the other maximally mixed state,
1
1
|1ix = |0i |1i ,
2
2
into the target register (the B-channel) of Uf , we can transfer or kick-back 100%
of the information about the unknown function f (x) from the B-register output to
the A-register output.
Youve actually experienced this idea earlier today when you studied quantum
teleportation. Recall that by merely rearranging the initial configuration of our input
state we were able to effect a seemingly magical transfer of |i from one channel
to the other. In the current context, presenting the x-basis ket, |1ix , to the target
register will have a similar effect.
Temporary Change in Notation
Because we are going to make heavy use of the x-basis kets here and the variable x is
being used as the Boolean input to the function f (x), I am going to call into action
our alternate x-basis notation,
|+i
|i
|0ix
and
|1ix .
Preparing the Oracles Input: The Two Left Hadamard Gates
Together, the two techniques explain the first part of Deutschs circuit (in the dashedbox),
|0i
Uf
|1i
(ignore)
363
We recognize H as the operator that takes z-basis kets to x-basis kets, thus manufacturing a |+i (i.e., |0ix ) for the data register input and |i (i.e., |1ix ) for the target
register input,
|0i
|+i
.
|1i
|i
In other words, the Hadamard gate converts the two natural basis kets (easy states to
prepare) into superposition inputs for quantum oracle. The top gate sets up quantum
parallelism for the circuit, and the bottom one sets up the phase kick-back. For
reference, algebraically these two gates perform
H |0i
H |1i
|0i + |1i
= |+i and
2
|0i |1i
= |i .
2
Analyzing the Oracle
The real understanding of how the algorithm works comes by analyzing the kernel of
the circuit, the oracle (in the dashed-box),
|0i
H
.
Uf
|1i
(ignore)
Step 1. CBS Into Both Channels. We creep up slowly on our result by first
considering a CBS ket into both registers, a result we know immediately by definition
of Uf ,
Data register:
|xi
|xi
Uf
Target register: |yi
|y f (x)i
or algebraically,
Uf (|xi |yi)
|xi |y f (x)i
Step 2. CBS Into Data and Superposition into Target. We stick with a
CBS |xi going into the data register, but now allow the superposition |i to go into
364
the target register. Extend the above linearly,
Uf (|xi |i)
|xi
|0i |1i
!
=
Uf |xi |0i Uf |xi |1i
Uf
|xi |0 f (x)i |xi |1 f (x)i
|xi |f (x)i |xi | f (x) i
|xi
|f (x)i | f (x) i
!
.
This amounts to
Uf (|xi |i)
|xi
|0i |1i
when f (x) = 0
|1i |0i
, when f (x) = 1
2
|0i |1i
f (x)
|xi (1)
.
2
Since its a scalar, (1)f (x) can be moved to the left and be attached to the A-registers
|xi, a mere rearrangement of the terms,
|0i |1i
f (x)
f (x)
Uf (|xi |i) = (1)
|xi
=
(1)
|xi |i ,
2
and we have successfully (like magic) moved all of the information about f (x) from
the B-register to the A-register, where it appears as an overall phase factor in the
scalars exponent, (1)f (x) .
The Oracles part of the circuit would process this intermediate steps data as
follows.
(1)f (x) |xi
|xi
Uf
|i
|i
Although we have a ways to go, lets pause to summarize what we have accomplished so far.
Quantum Mechanics: This is a non-essential, theoretical observation to test
your memory of our quantum mechanics lesson. We have proven that |xi |i is
an eigenvector of Uf with eigenvalue (1)f (x) for x = 0, 1.
Quantum Computing: The information about f (x) is encoded kickedback in the A (data) registers output. Thats where we plan to look for it
365
in the coming step. Viewed this way, the B-register retains no useful information; just like in teleportation, a rearrangement of the data sometimes creates a
perceptual shift of information from one channel to another that we can exploit
by measuring along a different basis something we will do in a moment.
Step 3. Superpositions into Both Registers. Finally, we want the state |+i
to go into the data register so we can process both f (0) and f (1) in a single pass.
The effect is to present the separable |+i |i to the oracle and see what comes out.
Applying linearity to the last result we get
|0i + |1i
Uf (|+i |i) = Uf
|i
2
Uf |0i |i + Uf |1i |i
=
2
=
(1)f (0) |0i |i + (1)f (1) |1i |i
(1)f (0) |0i + (1)f (1) |1i
|i .
2
By combining the phase kick-back with quantum parallelism, weve managed to get an
expression containing both f (0) and f (1) in the A-register. We now ask the question
that Deutsch posed in the context of this simple expression, What is the difference
between the balanced case (f (0) 6= f (1)) and the constant case (f (0) = f (1))?
Answer: When constant, the two terms in the numerator have the same sign and
when balanced, they have different signs, to wit,
|0i + |1i
(1)
|i , if f (0) = f (1)
2
Uf (|+i |i) =
|0i |1i
(1)
|i , if f (0) 6= f (1)
2
We dont care about a possible overall phase factor or (1) in front of all this since
its a unit scalar in a state space. Dumping it and noticing that the A-register has
x-basis kets in both cases, we get the ultimate simplification,
|+i |i , if f (0) = f (1)
Uf (|+i |i) =
|i |i , if f (0) 6= f (1)
the perfect form for an x-basis measurement. Before we do that, lets have a look at
the oracles input and output states,
|+i
|+i or |i
,
Uf
|i
|i
366
Measurement
We only care about the A-register, since the B-register will always collapse to |i.
The conclusion?
After measuring the A-register along the x-basis, if we collapse to |+i, f
is constant, and if we collapse to |i, f is balanced.
Of course an x-basis measurement is nothing more than a z-basis measurement
after applying the x z basis transforming unitary H. This explains the insertion
of the final Hadamard gate in the upper right (dashed,
|0i
H
.
Uf
|1i
(ignore)
Deutschs Algorithm in Summary
Weve explained the purpose of all the components in the circuit and how each plays
a role in leveraging quantum parallelism and phase kick-back. The result is extremely
easy to state. We run the circuit
|0i
H
Uf
|1i
(ignore)
one time only and measure the data register output in the natural basis.
If we read 0, f is constant.
If we read 1, f is balanced.
This may not seem like game changing result; a quantum speed up of 2 in a problem that is both trivial and without any real world application, but it demonstrates
that there is a difference between quantum computing and classical computing. It
also lays the groundwork for the more advanced algorithms to come.
12.6
n Qubits and More Algorithms
Were just getting started, though. Next time we attack a general n-qubit computer
and two algorithms that work on that kind of system, so get ready for more fun.
367
Chapter 13
Multi-Qubit Systems and
Algorithms
13.1
Moving Up from 2 Qubits to n Qubits
This week we generalize our work with one and two-qubit computation to include
n-qubits for any integer n > 2. Well start by defining nth order tensor products, a
natural extension of what we did for 2nd and 3rd order products and then apply that
to nth order state spaces, H(n) .
It is common for non-math major undergraduates to be confounded by higher
order tensors due to the vast number of coordinates and large dimensions, so I will
give you a running start by doing a short recap of both the second and third order
tensor products first. This will establish a pattern that should make the larger orders
go down more smoothly.
13.2
General Tensor Products
13.2.1
Recap of Order-2 Tensor Products
Objects of the Product Space and Induced Basis
We learned that the tensor product of two vector spaces, A and B, having dimensions
dA and dB , respectively, is the product space,
W
AB,
368
whose vectors (a.k.a. tensors) consist of objects, w, expressible as weighted sums of
the separable basis, i.e.,
dA 1,
dB 1
ckj (ak bj ) ,
k=0,
j=0
where the ckj are the scalar weights and also serve as the coordinates of w along the
states basis.
The separable basis tensors appearing in the above linear combination are the
dA dB vectors
n
o
ak bj k = 0, . . . , (dA 1) and j = 0, . . . , (dB 1) ,
induced by the two component bases,
A
ak
dA 1
bj
dB 1
and
k=0
j=0
The sums, products and equivalence of tensor expressions were defined by the required
distributive and commutative properties, but can often be taken as the natural rules
one would expect.
Separable Operators in the Product Space
A separable operator on the product space is one that arises from two component
operators, TA and TB , each defined on its respective component space, A and B. This
separable tensor operator is defined first by its action on separable order-2 tensors
[TA TB ] (a b)
TA (a) TB (b)
and since the basis tensors are of this form, it establishes the action of TA TB on
the basis which, in turn, extends the action to the whole space.
13.2.2
Recap of Order-3 Tensor Products
We also outlined the same process for a third-order tensor product space
W
ABC
in order to acquire the vocabulary needed to present a few of the early quantum
algorithms involving three channels. Here is a summary of that section.
369
Objects of the Produce Space and Induced Basis
Assuming A, B and C have dimensions dA , dB and dC , respectively, the basis for the
product space is the set of dA dB dC separable tensors
n
o
ak bj cl ,
where
ak is the k th vector in the basis for A ,
bj is the j th vector in the basis for B
cl is the l
th
and
vector in the basis for C .
A general tensor w in the product space is expressible as a weighted sum of these
basis tensors,
dA 1,
dB 1,
dC 1
ckjl (ak bj cl ) ,
k=0,
j=0,
l=0
where the ckjl are the scalar weights (or coordinates) that define w.
Separable Operators in the Product Space
A separable operator on the product space is one that arises from three component
operators, TA , TB and TC , each defined on its respective component space, A, B and
C. This separable tensor operator is defined first by its action on separable order-3
tensors
[TA TB TC ] (a b c)
TA (a) TB (b) TC (c)
and since the basis tensors are of this form, that establishes the action of TA TB TC
on the basis which, in turn, extends the action to the whole space.
13.2.3
Higher Order Tensor Products
We now formally generalize these concepts to any order tensor product space
W
A0 A1 An2 An1 ,
370
for n >= 2. Well label the dimensions of the component spaces by
dim (A0 )
dim (A1 )
=
=
d0 ,
d1 ,
=
=
dn2 and
dn1 .
..
.
dim (An2 )
dim (An1 )
The tensor product space will have dimension
dim (W )
d0 d1 dn2 dn1
n1
Y
dk
k=0
which seems really big (and is big in fields like general relativity), but for us each
component space is H which has dimension two, so dim(W ) will be the still large
but at least palatable number 2n .
Objects of the Product Space and Induced Basis
The vectors a.k.a. tensors of the space consist of those w expressible as weighted
sums of the separable basis
n
od0 1, d1 1, ..., dn1 1
a0k0 a1k1 a2k2 a(n1)kn1
.
k0 , k1 , ..., kn1 = 0, 0, ..., 0
The somewhat daunting subscript notation says that
a0k0 is the (k0 )th vector in the basis for A0 ,
a1k1 is the (k1 )th vector in the basis for A1 ,
a2k2 is the (k2 )th vector in the basis for A2 ,
..
.
a(n1)kn1 is the (kn1 )th vector in the basis for An1 ,
If we write this algebraically, the typical w in W has a unique expansion along the
tensor basis weighted by the scalars ck0 k1 ...kn1 ,
w
ck0 k1 ...kn1 a0k0 a1k1 a2k2 a(n1)kn1 .
This notation is an order of magnitude more general than we need, but it is good
to have down for reference. Well see that the expression takes on a much more
manageable form when we get into the state spaces of quantum computing.
The sums, products and equivalence of tensor expressions have definitions analogous to their lower-order prototypes. Youll see examples as we go.
371
Separable Operators in the Product Space
A separable operator on the product space is one that arises from n component operators, T0 , T1 . . ., Tn1 , each defined on its respective component space, A0 , A1 ,
. . ., An1 . This separable tensor operator is defined first by its action on separable
order-n tensors
[T0 T1 Tn1 ] (v0 v1 vn1 )
T0 (v0 ) T1 (v1 ) Tn1 (vn1 ) ,
and since the basis tensors are always separable,
a0k0 a1k1 a(n1)kn1 ,
this establishes the action of T0 T1 Tn1 on the basis,
[T0 T1 Tn1 ] a0k0 a1k1 a(n1)kn1
T0 (a0k0 ) T1 (a1k1 ) Tn1 (a(n1)kn1 ) ,
which, in turn, extends the action to the whole space.
Notation
Q
N
Sometimes we use the
or
notation to shorten expressions. In these forms, the
product space would be written in one of the two equivalent ways
W
n1
O
Ak
n1
Y
k=0
the induced basis in one of
( n1
)d0 1, d1 1, ..., dn1 1
O
ajkj
j=0
Ak ,
k=0
( n1
Y
)d0 1, d1 1, ..., dn1 1
ajkj
j=0
k0 , k1 , ..., kn1 = 0, 0, ..., 0
,
k0 , k1 , ..., kn1 = 0, 0, ..., 0
a separable operator as one of
n1
O
Tk
k=0
n1
Y
Tk ,
k=0
and a separable operators action on a separable state as either
" n1 # n1 !
n1
O
O
O
Tk
vk
=
Tk (vk )
k=0
k=0
k=0
or
"n1
Y
k=0
#
Tk
n1
Y
!
vk
k=0
n1
Y
k=0
372
Tk (vk ) .
13.3
n-Qubit Systems
The next step in this lecture is to define the precise state space we need for a quantum
computer that supports n qubits. I wont back up all the way to two qubits as I did for
the tensor product, but a short recap of three qubits will be a boon to understanding
n qubits.
13.3.1
Recap of Three Qubits
A three qubit system is modeled by a third order tensor product of three identical
copies of our friendly spin-1/2 Hilbert space, H. We can use either the order -notation
or component space label notation to signify the product space,
H(3)
HA HB HC .
The dimension is 2 2 2 = 8.
Three Qubit CBS and Coordinate Convention
The natural three qubit tensor basis is constructed by forming all possible separable
products from the component space basis vectors, and we continue to use our CBS
ket notation. The CBS for H(3) is therefore
n
|0i |0i |0i , |0i |0i |1i , |0i |1i |0i , |0i |1i |1i ,
o
|1i |0i |0i , |1i |0i |1i , |1i |1i |0i , |1i |1i |1i ,
with the often used shorthand options
|0i |0i |0i |0i |0i |0i |000i
|0i3
|0i |0i |1i |0i |0i |1i |001i
|1i3
|0i |1i |0i |0i |1i |0i |010i
|2i3
|0i |1i |1i |0i |1i |1i |011i
|3i3 .
|1i |0i |0i |1i |0i |0i |100i
|4i3
|1i |0i |1i |1i |0i |1i |101i
|5i3
|1i |1i |0i |1i |1i |0i |110i
|6i3
|1i |1i |1i |1i |1i |1i |111i
|7i3
373
The notation of the first two columns admits the possibility of labeling each of the
component kets with the H from which it came, A, B or C,
|0iA |0iB |0iC |0iA |0iB |0iC ,
|0iA |0iB |1iC |0iA |0iB |1iC ,
etc.
The densest of the notations expresses the CBS ket as an integer from 0 to 7. This
can all be summarized by looking at the third order coordinate representation of these
eight tensors:
|000i
|0i3
1
0
0
0
0
0
0
0
|001i
|1i3
0
1
0
0
0
0
0
0
|010i
|2i3
0
0
1
0
0
0
0
0
|011i
|3i3
0
0
0
1
0
0
0
0
|100i
|4i3
0
0
0
0
1
0
0
0
|101i
|5i3
0
0
0
0
0
1
0
0
|110i
|6i3
0
0
0
0
0
0
1
0
|111i
|7i3
0
0
0
0
0
0
0
1
A typical three qubit value is a normalized superposition of the eight CBS, e.g.,
|2i3 + e99i |3i3 + i |7i3
|3i3 + |5i3
,
,
2
3
.1 |0i3 + .6 |2i3 .05 |4i3 ( .05 + i .2) |6i3 ,
or most generally,
7
X
ck |ki3 ,
k=0
where
7
X
|ck |2
1.
k=0
13.3.2
Three Qubit Logic Gates; the Toffoli Gate
Order three quantum logic gates are unitary operators on H(3) . They can be constructed by taking, for example,
374
the separable product of three first order operators,
the separable product of a second order and a first order operator, or
any operator defined by a possibly non-separable unitary matrix.
In this course, we wont be studying third order gates other than the ones that are
separable products of first order gates (like H 3 discussed in the next section). However, lets meet one the Toffoli gate which plays a role in reversible computation
and some of our work in the later courses CS 83B and CS 83C.
:
The Symbol
The gate is usually drawn without an enclosing box as in
but can be boxed to emphasize its overall effect on a tripartite state,
The A and B registers are the control bits, and the C register the target bit,
control bits
target bit
two terms that will become clear in the next bullet point.
At times Ill use all caps, as in TOFFOLI, to name the gate in order to give it the
status of its simpler cousin, CNOT.
|xi |yi : Action on the CBS
The TOFFOLI gate has the following effect on the computational basis states:
|xi
|xi
|yi
|yi
|zi
| (x y) z i
375
In terms of the eight CBS tensors, it leaves the A and B registers unchanged and
negates the C register qubit or leaves it alone based on whether the AND of the
control bits is 1 or 0:
(
|zi ,
if x y = 0
|zi 7
| zi ,
if x y = 1
It is a controlled-NOT operator, but the control consists of two bits rather than one.
Remember, not every CBS definition we can drum up will result in a unitary
operator, especially when we start defining the output kets in terms of arbitrary
classical operations. In an exercise during your two qubit lesson you met a bipartite
gate which seemed simple enough but turned out not to be unitary. So we must
confirm this property in the next bullet.
: The Matrix
We compute the column vectors of the matrix by applying TOFFOLI to the CBS
tensors to get
!
MTOFFOLI
TOFFOLI |000i , TOFFOLI |001i , , TOFFOLI |111i
!
|000i , |001i , |010i , |011i , |100i , |101i , |111i , |110i
1
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
0
1
0
which is an identity matrix until we reach the last two rows (columns) where it swaps
those rows (columns). It is unitary.
[Exercise. Prove that this is not a separable operator.]
376
|i2 : Behavior on General State
Applying TOFFOLI to the general
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
state,
0
0
0
0
0
1
0
0
0
c0
c1
0
0
ce
0
c3
0
c4
0
c5
1 c6
0
c7
0
0
0
0
0
0
0
1
c0
c1
ce
c3
.
c4
c5
c7
c6
This is as far as we need to go on the Toffoli gate. Our interest here is in higher order
gates that are separable products of unary gates.
13.3.3
n Qubits
Definition and Notation
Definition of n Qubits. n qubits are, collectively, the tensor product
of n identical copies of the 2-D Hilbert space H, and the value of the n
qubits at any given moment is a particular tensor in this space having unit
length.
Note. As usual, we consider two unit-tensors which differ by a phase factor, ei
for real , to be the same n qubit value.
We can designate this state space using the notation
n
H(n)
}|
{
z
H H ... H
n1
O
H.
k=0
n Qubit Computational Basis States
The natural n qubit tensor basis is constructed by forming all possible separable
products from the component space basis vectors. The CBS for H(n) is therefore
n
|0i |0i |0i |0i , |0i |0i |0i |1i , |0i |0i |1i |0i ,
|0i |0i |1i |1i , |0i |1i |0i |0i , . . . ,
...
, |1i |1i |1i |0i , |1i |1i |1i |1i
377
with the shorthand options
|0i |0i |0i
|0i |0i |0i
|0 000i
|0in
|0i |0i |1i
|0i |0i |1i
|0 001i
|1in
|0i |1i |0i
|0i |1i |0i
|0 010i
|2in
|0i |1i |1i
|0i |1i |1i
|0 011i
|3in
|1i |1i |0i
|1i |1i |0i
|1 110i
|2n 2in
|1i |1i |1i
|1i |1i |1i
|1 111i
|2n 1in
..
.
Dimension of H(n)
As you can tell by counting, there are 2n basis tensors in the product space, which
makes sense because the dimension of the product space is the
product of the dimensions of the component spaces; since dim(H) = 2, dim H(n) = 2 2 2 = 2n ,
X.
For nth order CBS kets we usually label each component ket using the letter x
with its corresponding space label,
|xn1 i |xn2 i |xn3 i |x0 i ,
xk {0, 1} ,
with the more common and denser alternatives
|xn1 i |xn2 i |x1 i |x0 i
|xn1 xn2 . . . x3 x2 x1 x0 i ,
and densest of them all,
|xin ,
x {0, 1, 2, 3, . . . , 2n 1} .
For example, for n = 5,
|0i5
|00000i ,
|1i
|00001i ,
|2i5
|00010i ,
|8i5
|01000i ,
|23i
and, in general,
|10111i ,
|xi5
378
|x4 x3 x2 x1 x0 i .
13.3.4
n Qubit Logic Gates
Quantum logic gates of order n > 3 are nothing more than unitary operators of order
n > 3, which we defined above. Theres no need to say anything further about a
general nth order logic gate. Instead, lets get right down to the business of describing
the specific example that will pervade the remainder of the course.
The nth Order Hadamard Gate, H n
We generalize the two-qubit Hadamard gate, H 2 , to n-qubits naturally. It is the local
operator that behaves like n individual unary H gates if presented with a separable
input state,
H
,
n copies
.
.
H
or described in terms of the CBS,
| xn1 xn2 x1 x0 i
H |xn1 i
H |xn2 i
..
.
H n
H |x1 i
H |x0 i
Algebraically, we developed the formula for the second order Hadamard,
H
|xi
3
1 X
(1)x y |yi2 ,
2 y=0
where stands for the mod-2 dot product. It is only a matter of expending more
time and graphite to prove that this turns into the higher order version,
n 1
n 2X
1
n
(1)x y |yin .
H n |xi
=
2
y=0
Putting this result into the circuit diagram gives us
|xi
H n
/
1
2
n
n 1
2X
y=0
379
(1)x y |yin .
Vector Notation
Well sometimes present the formula using vector dot products. If x and y are considered to be vectors of 1s and 0s, we would represent them using boldface x and
y,
xn1
yn1
xn2
yn2
..
..
x x = . ,
y y = . .
x1
y1
x0
y0
When so expressed, the dot product between vector x and vector y is considered the
mod-2 dot product,
xy
xn1 yn1 xn2 yn2 x1 y1 x0 y0 .
This results in an equivalent form of the Hadamard gate using vector notation,
H n |xi
=
n 1
n 2X
(1)x y |yin .
y=0
Ill remind you about this when the time comes.
Higher Order Basis Conversion
Well be doing higher level basis conversions frequently, especially between the z-basis
and the x-basis. Lets review and extend our knowledge.
The induced z-basis for H(n) is
|0i |0i |0i |0i , |0i |0i |0i |1i , |0i |0i |1i |0i ,
. . . , |1i |1i |1i |1i ,
while the induced x-basis for H(n) is
|0ix |0ix |0ix |0ix , |0ix |0ix |0ix |1ix , |0ix |0ix |1ix |0ix ,
. . . , |1ix |1ix |1ix |1ix .
Using our alternate order-one x-basis notation,
|+i
|i
|0ix
and
|1ix ,
the induced x-basis CBS can also be written without using the letter x as a label,
|+i |+i |+i |+i , |+i |+i |+i |i , |+i |+i |i |+i ,
. . . , |i |i |i |i .
380
Since H converts to and from these the x and z bases in H, it easy to confirm that
the separable H n converts to-and-from these two basis in H(n) .
[Exercise. Do it.]
Notation. In order to make the higher order x-CBS kets of H(n) less confusing
(we need x as an encoded integer specifying the CBS state), Im going to call to
duty some non-standard notation that I introduced in our two qubit lecture: Ill use
the subscript to indicate a CBS relative to the x-basis:
|yin
nth -order encoded input, y, relative to the x-basis.
That is, if y is an integer from 0 to 2n 1, when you see the subscript on the CBS
ket you know that its binary representation is telling us which x-basis (not z-basis)
ket it represents. So,
|0in =
|1in =
|2in =
|3in =
etc.
|+i |+i |+i |+i |+i
|+i |+i |+i |+i |i
|+i |+i |+i |i |+i
|+i |+i |+i |i |i
,
,
,
,
(Without the subscript , of course, we mean the usual z-basis CBS.) This frees
up the variable x for use inside the ket,
|xin
nth -order encoded input, x, relative to the x-basis.
With this notation, the change-of-basis is described simply as
H n |xin
|xin
H n |xin
|xin .
and in the other direction as
Natural Coordinates for the x-Basis
In the last section we were looking at the separable form of the CBS for an nth order
Hilbert space. Lets count for a moment. Whether we have an x-basis, z-basis or any
other basis induced from the n component Hs, there are n factors in the separable
factorization, i.e.,
n components
n components
z
}|
{
|0i |1i |1i |0i |1i
z
}|
{
|+i |i |i |+i |i
or
But when expanded along any basis these states have 2n components (because the
product space is 2n dimensional). From our linear algebra and tensor product lessons
381
we recall that a basis vector, bk , expanded along its own basis, B, contains a single 1
and the rest 0s. In coordinate form that looks like
bk
0
..
.
1 kth element .
.
..
0 B
This column vector is very tall in the current context, whether a z-basis ket,
0
..
|xin
=
2n rows
1 xth element ,
..
0
z
an x-basis CBS ket,
|xin
..
2n rows
1
..
xth element ,
or more to the point of this exploration, an x-basis
n
n
=
2 rows
|xi
ket expanded along the z-basis,
?
..
.
? .
.
..
?
z
Actually, we know what those ?s are because it is the H n which turns the z-CBS
into an x-CBS,
|xin
H n
|xin ,
and we have already seen the result of H n applied to any |xin , namely,
|xin
|xi
382
n 1
n 2X
y=0
(1)x y |yin ,
which, when written out looks something like
n
|0in |1in |2in |3in . . . |2n1 1i
n
.
2
In coordinate form thats
|xin
2n rows
1
1
.
1
..
n .
1
2
.
..
1 z
But we can do better. Not all possible sums and differences will appear in the sum,
so not all possible combinations of +1 and -1 will appear in an x-basis kets column
vector (not counting the scalar factor ( 12 )n ). An x-CBS ket, |xi , will have exactly
the same number of +s as s in its expansion (and +1s, 1s in its coordinate vector)
except for |0i , which has all +s (+1s). How do we know this?
We start by looking at the lowest dimension, H = H(1) , where there were two
easy-to-grasp x-kets in z-basis form,
|+i
|i
|0i + |1i
and
2
|0i |1i
.
2
The claim is easily confirmed here with only two kets to check. Stepping up to second
order, the x-kets expanded along the z-basis were found to be
|+i |+i
|0i |0i + |0i |1i + |1i |0i + |1i |1i
2
|+i |i
|0i |0i |0i |1i + |1i |0i |1i |1i
2
|i |+i
|0i |0i + |0i |1i |1i |0i |1i |1i
2
|i |i
|0i |0i |0i |1i |1i |0i + |1i |1i
,
2
also easily seen to satisfy the claim. Lets make a prediction.
Little Lemma. For an order-n H(n) , a typical x CBS ket look like
n
|0in |1in |2in |3in . . . |2n1 1i
n
,
2
383
where except for |0in which has all plus signs the sum will always have an equal
numbers of +s and s.
[Caution. This doesnt mean that every sum with an equal number of positive
and negative coefficients is necessarily an x CBS ket; there are still more ways to
distribute the +s and s equally than there are CBS kets, so the distribution of the
plus and minus signs has to be even further restricted if the superposition above is to
represent an x-basis ket. But just knowing that all x CBS tensors, when expanded
along the z-basis, are balanced in this sense, will help us understand and predict
quantum circuits.]
Proof of Lemma. We already know that the lemma is true for first and second
order state spaces because we are staring directly into the eyes of the two x-bases,
above. But lets see why the Hadamard operators tell the same story. The matrix for
H 2 , which is used to convert the second order z-basis to an x-basis, is
1 1
1
1
1
1 1 1 1 .
H H =
1 1 1
21
1 1 1
1
If we forget about the common factor 12 , it has a first column of +1s, and all its
remaining columns have equal numbers of +1s and 1s. If we apply H 2 to |0i2 =
( 1, 0, 0, 0 )t we get the first column, all +1s. If we apply it to any |xi2 , for x > 0,
say, |2i2 = ( 0, 0, 1, 0 )t , we get one of the other columns, each one of which has an
equal numbers of +1s and 1s.
To reproduce this claim for any higher order Hadamard, we just show that the
matrix for H n (which generates the nth order x-basis) will also have all +1s in the
left column and equal number of +1s and 1s in the other columns. This is done
formally by recursion, but we can get the gist by noting how we extend the claim
from n = 2 to n = 3. By definition,
H 3 = H H H = H H 2 ,
or, in terms of matrices,
H 3
1
1
1
1
1 1 1 1 1
1 1 1
.
1 1 1
21
2 1 1
1 1 1
1
By our technique for calculating tensor product matrices, we know that the matrix
on the right will appear four times in the 8 8 product matrix, with the lower right
copy being negated (due to the -1 in the lower right of the smaller left matrix). To
384
wit,
H 3
1 1
1
1
1 1 1 1
1
1 1 1
1 1 1
1
3
2
1
1
1
1
1 1 1 1
1
1 1 1
1 1 1
1
1
1
1
1
1 1
1 1
1
1 1 1
1 1 1
1
1 1 1 1
1
1 1
1
1 1
1
1
1
1 1 1
Therefore, except for the first column (all +1s), the tensor products columns will all
be a doubling (vertical stacking) of the columns of the balanced 4 4 (or a negated
4 4). Stacking two balanced columns above one another produces a column that is
twice as tall, but still balanced.
QED
[Exercise. Give a rigorous proof by showing how one extends an order (n 1)
Hadamard matrix to an order n Hadamard matrix.]
13.3.5
Oracles for n Qubit Functions
We saw that quantum oracles for functions of a single boolean variable helped produce
some early quantum algorithms. Now that we are graduating to n qubit circuits useful
in studing Boolean functions of n binary inputs, its time to upgrade our definition
of quantum oracles to cover multi-input f s. (We retain the assumption that our
functions produce only a single Boolean output value theyre not vector functions).
We are given a black box for an n-input Boolean function, f ( xn1 , xn2 , . . . , x1 , x0 ),
xn1
xn2
..
.
x1
x0
f ( xn1 , xn2 , . . . , x1 , x0 ) .
We assume that circuit theory enables us to build an (n + 1)-in, (n + 1)-out oracle,
Uf , defined on the 2n+1 CBS kets
n
o
|xin |yi ,
where x { 0, 1, 2, . . . , 2n 1 } is in encoded form, y { 0, 1 }, and the circuits
385
action on these CBS is
|xin
|xin
.
Uf
|yi
|y f (x)i
A consequence of the definition is that Uf is its own inverse (try it).
It is also easy to see that Uf emulates f (x) by setting y = 0,
Uf (|xin |0i)
|xin |f (x)i .
We assume (it does not follow from the definition) that the oracle is of the
same spatial circuit complexity as f (x), i.e., it grows in size at the same rate as
f grows relative to the number of inputs, n. This is usually demonstrated to
be true for common individual functions by manually presenting circuits that
implement oracles for those functions.
13.4
Significant Deterministic Speed-Up: The DeutschJozsa Problem
Deutschs algorithm enabled us to see how quantum computing could solve a problem
faster than classical computing, but the speed up was limited to 2, forget about
the expense required to build the quantum circuit; its not enough to justify the
investment. We now restate Deutschs problem for functions of n Boolean inputs and
in that context call it the Deutsch-Jozsa Problem. We will find that the classical
solution grows (in time) exponentially as n increases, not counting the increasing
oracle size, while the quantum algorithm we will present next, has a constant time
solution. This is a significant speed-up relative to the oracle and does give us reason
to believe that quantum algorithms may be of great value. (The precise meaning of
relative vs. absolute speed-up will be presented in our up-coming lesson devoted to
quantum oracles, but well discuss a couple different ways to measure the speed-up
informally later today.)
We continue to study functions that have Boolean inputs and outputs, specifically
n binary inputs and one binary output,
f : {0 , 1}n {0 , 1} .
The Deutsch-Jozsa Problem. Given an unknown function,
f ( xn1 , xn2 , . . . , x1 , x0 ) ,
of n inputs that we are told is either balanced or constant, determine
which it is in one query of the quantum oracle, Uf .
386
13.4.1
Deutsch-Jozsa Algorithm
The algorithm consists of building a circuit very similar to that in Deutschs circuit
and measuring the data register once. Our conclusion about f is the same as in the
unary case: if we get a 0 the function is constant, if we get 1 the function is
balanced. Well analyze the speed-up after we prove this claim.
The Circuit
We replace the unary Hadamard gates of Deutschs circuit with nth order Hadamard
gates to accommodate the wider data register lines, but otherwise, the circuit layout
is organized the same:
|0in
H n
H n
Uf
|1i
(ignore)
The circuit reveals that we will be
applying quantum parallelism by allowing the upper Hadamard to produce a
perfectly mixed input state, |0in , into Uf s data register, and
using the phase kick-back trick by putting |i into Uf s target register.
Preparing the Oracles Input: The Two Left Hadamard Gates
The first part of the Deutsch-Jozsa circuit (in the dashed box) prepares states that are
needed for quantum parallelism and phase kick-back just as the lower-order Deutsch
circuit did,
|0in
H n
H n
Uf
|1i
(ignore)
The H and H n operators take z-basis kets to x-basis kets in the first order H, and
the nth order H(n) spaces, respectively, thus manufacturing a |0in for the data register
input and |i for the target register input,
|0in
|1i
H n
|0in
|i
387
The top gate sets up quantum parallelism and the bottom sets up the phase kick-back.
For reference, here is the algebra:
H
|0i
H |1i
|0in
=
n 1
n 2X
|yi
2X
1
1
|yin
n
2 y=0
y=0
and
|0i |1i
= |i .
2
Analyzing the Oracle
Next, we consider the effect of the oracle on these two x-basis inputs (dashed box),
|0in
H n
H n
Uf
|1i
(ignore)
Well do it in stages, as before, to avoid confusion and be sure we dont make mistakes.
Step 1. CBS Into Both Channels. When a natural CBS ket goes into both
registers, the definition of Uf tells us what comes out:
Data register:
|xin
|xin
,
Uf
Target register:
|yi
|y f (x)i
algebraically,
Uf (|xin |yi)
|xin |y f (x)i .
Step 2. CBS Into Data and Superposition into Target. Continuing on
with a general CBS |xi into the data register, we next allow the superposition |i
into the target register.
n
n
!
U
|xi
|0i
U
|xi
|1i
f
f
|0i |1i
=
Uf (|xin |i) = Uf |xin
2
2
=
|xin |0 f (x)i |xin |1 f (x)i
|xin |f (x)i |xin | f (x) i
388
|xin
|f (x)i | f (x) i
!
.
This amounts to
Uf (|xin |i)
|0i |1i
, when f (x) = 0
2
|xin
|1i |0i
, when f (x) = 1
2
|0i |1i
n
f (x)
|xi (1)
.
2
The scalar, (1)f (x) can be kicked back,
|0i |1i
n
n
f (x)
Uf (|xi |i) = (1)
|xi
2
(1)f (x) |xin |i ,
and once again the information about f (x) is converted into an overall phase factor
in the data register, (1)f (x) |xin .
From a circuit standpoint, we have accomplished
|xin
Data register:
(1)f (x) |xin
.
Uf
|i
Target register:
|i
Step 3. Superpositions into Both Registers. Finally, we send the full output
of H n |0in ,
|0in
|+ + + + +i ,
into the data register so we can process f (x) for all x in a single pass and thereby
leverage quantum parallelism. The net effect is to present the separable |0in |i to
the oracle. Applying linearity to the last result we find
!
!
n 1
2X
1
n
n
Uf |0i |i
|yi
|i
= Uf
2n y=0
=
n 1
2X
1
Uf |yin |i
2n y=0
n 1
2X
1
(1)f (y) |yin |i
2n y=0
!
n 1
2X
1
n
(1)f (y) |yi |i .
2n y=0
389
The Final Hadamard Gate
[Warning. The version of the argument I give next is easy enough to follow and will
prove the algorithm, but it may leave you with a huh? feeling. Thats because
it does not explain how one arrives at the decision to apply the final Hadamard. Ill
present a more illuminating alternative at the end of this lesson that will be more
satisfying but which requires that you activate a few more little gray cells.]
We are ready to apply the nth order Hadamard gate in the upper right (dashed
box),
|0in
H n
H n
Uf
|1i
(ignore)
To that end, we consider how it changes the state at access point P into a state at
the final access point Q:
|0in
H n
H n
Uf
|1i
(ignore)
H
In this phase, we are subjecting the full (n + 1)st order separable
!
n 1
2X
1
(1)f (y) |yin |i
2n y=0
to the separable transformation H n 1, which allows us to consider each component
of the separable operation individually. We only care about the Hadamard part, since
390
it is the output of the data register we will test. It produces the output
!
n 1
n 1
2X
2X
1
1
n
n
f (y)
H
(1)
|yi
=
(1)f (y) H n |yin
n
n
2 y=0
2 y=0
! #
n 1 "
n 1
2X
2X
1
1
=
(1)f (y)
(1)y z |zin
2n y=0
2n z=0
1
2n
n 1
2X
z=0
n 1
2X
!
|zin
(1)f (y) (1)y z
y=0
{z
G(z)
where we have regrouped the sum and defined a scalar function, G(z), of the summation index z. So, the final output is an expansion along the z-basis,
n
data register at access point Q
2 1
1 X
G(z) |zin .
2n z=0
We now look only at the coefficient, G(0), of the very first CBS ket, |0in . This will
tell us something about the other 2n 1 CBS coefficients, G(z), for z > 0. We break
it into two cases.
f is constant. In this case, f (y) is the same for all y, either 0 or 1; call it c.
.
We evaluate the coefficient of |0in in the expansion, namely G(0)
2n
n
G(0)
2n
2 1
1 X
(1)c (1)y 0
2n y=0
(1)c
2n
2n
1 ,
thereby forcing the coefficients of all other z-basis kets in the expansion to be 0
(why?). So in the constant case we have a CBS ket |0in at access point Q with
certainty and are therefore guaranteed to get a reading of 0 if we measure
the state.
f is balanced. This time the coefficient of |0in in the expansion is
n
G(0)
2n
2 1
1 X
(1)f (y) (1)y 0
2n y=0
2 1
1 X
(1)f (y) ,
2n y=0
but a balanced f promises an equal number f (y) = 0 and f (y) = 1, so the
sum has an equal number of +1s and 1s, forcing it to be 0. Therefore, the
probability of a measurement causing a collapse to the state |0in is 0 (the
amplitude-squared of the CBS state |0in ). We are guaranteed to never get a
reading of 0 when we measure the data register at access point Q.
391
The Deutsch-Jozsa Algorithm in Summary
Weve explained the purpose of all the components in the circuit and how each plays
a role in leveraging quantum parallelism and phase kick-back. The result is extremely
easy to state. We run the circuit
|0in
H n
H n
Uf
|1i
(ignore)
one time only and measure the data register output in the natural basis.
If we read 0 then f is constant.
If we read x for any other x, (i.e., x [1, 2n 1]), then f is balanced.
13.4.2
Quantum vs. Classical Time Complexity
What have we actually accomplished in efficiency?
The Classical Time Complexity
In order to know the answer to the Deutsch-Jozsa problem deterministically, i.e., with
100% certainty, we would have to evaluate the function f for more than half of the
possible inputs, i.e., at least
2n
+1
2
2n1 + 1
times. That is, wed plug just over half of the possible x values into f (say, x =
0, 1, 2, ..., 2n1 + 1), and if they were all the same, wed know the function must
be constant. If any two were distinct, we know it is balanced. Of course, we may get
lucky and find that f (0) 6= f (1), in which case we can declare victory (balanced ) very
quickly, but we cannot count on that. We could be very unlucky and get the same
output for the first 2n /2 computations, only to know the answer with certainty, on
the 2n /2 + 1st. (if its the same as the others: constant, if not: balanced.)
While we have not had our official lecture on time complexity, we can see that
as the number of binary inputs, n, grows, the number of required evaluations of f ,
2n1 + 1, grows exponentially with n. However, when we consider that there are
N = 2n encoded integers that are allowed inputs to f , then as N grows, the number
of evaluations of f , N2 + 1, grows only linearly with N .
The classical problem has a solution which is exponential in n (the number of
binary inputs, or linear in N = 2n (the number of integer inputs).
392
The Quantum Time Complexity
We have solved the problem with one evaluation of Uf which is assumed to have the
same the same circuit complexity as f . Now you might say that this is a constant
time solution, i.e., it does not grow at all with n, because no matter how large n is, we
only need to evaluate Uf once. In that light, the quantum solution is constant-time;
it doesnt grow at all with n. We simply measure the output of the data register, x,
done using
if (x > 0)
and were done.
You might argue that we have overstated the case, because in order to detect
the output of the circuit, we have to query all n bits of the data registers to know
whether we get 000 00 or an integer other than that. No computer can do that
for arbitrarily large n without having an increasingly large circuit or increasingly long
testing algorithm. So in practical terms, this is an evaluation of Uf followed by n
one-bit queries, something that requires n if statements. That grows linearly with
n or, using encoded integer counting (N = 2n ), logarithmically (even better). Either
way, the quantum algorithm has a better time complexity (linear vs. exponential in
n or logarithmic vs. linear in N ) than its classical counterpart. So the speed-up is
real.
But theres another way to view things that puts the quantum algorithm in an even
more favorable light. Whether quantum or classical, the number of binary registers
to test is the same: n. So we can really ignore that hardware growth when we speak
of time complexity relative to the classical case; the quantum algorithm can be said
to solve the problem in constant time relative to the classical algorithm.
Reminder. Ill define terms like logarithmic, linear and exponential time complexity in the next lecture.
The Deterministic Result
Either way you look at it, if you require 100% certainty of the solution, we have found
an algorithm that is faster than the classical solution.
The Non-Deterministic (Probabilistic) Result
If, however, we allow a small error possibility for the classical case, as is only fair
since we might expect our quantum circuit to be prone to error (a topic of the next
course), then the classical algorithm grows neither exponentially with n nor linearly
with N , but in fact is a constant time algorithm, just like the Deutsch-Jozsa. Ill give
you an outline of the reason now, and after our probability lesson, we can make it
rigorous.
Classical Algorithm Admitting a Small Error Probability << 1.
393
Lets consider the following classical algorithm.
The M -and-Guess Algorithm. Let M be some positive integer (think 20).
Given a Boolean function, f (x) of x [ 0, n 1 ] which is either balanced or constant,
i.e., one that satisfies the Deutsch-Jozsa hypothesis, we evaluate f (x) M times, each
time at a random x [ 0, n 1 ]. We call each evaluation a trial. If we get two
different outputs, f (x0 ) 6= f (x00 ), by the time we complete our M trials, we declare
victory: f is balanced without a doubt. On the other hand, if we get the same output
for all M trials, we declare near victory: We report that f is constant, with a pretty
good certainty.
How often will we get the wrong answer using this algorithm?
The only way we can fail is if the function is balanced yet we declare it to be
constant after M trials. That only happens if we are unlucky enough to get M
straight 0s or M straight 1s from a balanced f .
Well call that eventuality, the event
S B,
which is a symbolic way to say, f was Balanced yet (technically AND) all trial
outcomes were the S ame.
Since a balanced f means there is a 50-50 chance of getting a 1 or a 0 on any trial,
this unlucky outcome is akin to flipping a fair coin M times and getting either all
heads or all tails. As you can intuit by imagining 20 heads or 20 tails in a sequence of
20 fair coin tosses, this is quite unlikely. Well explain it rigorously in the upcoming
lesson on probability, but the answer is that the probability of this event occurring,
designated P (S B), is
M
1
1
1
=
.
P (S B) = 2
2
2
2M
The factor of 2 out front is due to the fact that the error on a balanced function
can occur two different ways, all 1s or all 0s. The final factor 1/2 is a result of an
assumption which could be adjusted if not true that we are getting a constant
function or balanced function with equal likelihood.
So we decide beforehand the error probability we are willing to accept, say some
<< 1, and select M so that
1
2M
This will allow our classical algorithm to complete (with the same tiny error probability, , in a fixed number of evaluations, M , of the function f regardless of the
number of inputs, n. To give you an idea,
(
0.000001 ,
for M = 20
P (S B)
.
16
9 10 ,
for M = 50
394
Since the error probability does not increase with increasing n, the classical algorithm has a constant time solution, meaning that we can solve it with the same
time complexity as the quantum Deutsch-Jozso algorithm. (We will define terms
like complexity and constant time precisely very soon, but you get the general idea.)
Therefore, no realistic speed-up is gained using quantum computing if we accept a
vanishingly small error result.
This does not diminish the importance of the deterministic solution which does
show a massive computational speed increase, but we must always temper our enthusiasm with a dose of reality.
13.4.3
Alternate Proof of the Deutsch-Jozsa Algorithm
Id like to offer a slighty more elaborate but also more illustrative argument for the
final Hadamard gate and how we might guess that it is the correct way to complete
the circuit (dashed box),
|0in
H n
H n
Uf
|1i
(ignore)
Recall that we had established that at access point P,
|0in
H n
H n
Uf
|1i
(ignore)
H
P
the data register was in the state
n
2X
1
1
(1)f (y) |yin .
n
2 y=0
The hypothesis of Deutsch-Jozsa tells us that f is either balanced or constant which
has implications about this state.
1. If f is constant, then all the coefficients, (1)f (y) , are the same, and we are
looking at |0in (or possibly -1 times this state, observationally equivalent).
2. If f is balanced, then half the (1)f (y) are +1 and half are 1. Now, our little
lemma reminds us that this condition suggests but does not guarantee that
a balanced state might, at access point P, be an x-CBS state other than |0in .
395
The constant case 1, guarantees that we land in the x-CBS state, |0in . The balanced
case 2, suggests that we might end up in one of the other x-CBS states, |xin , for
x > 1. Lets pretend that in the balanced case we are lucky enough to land exactly
in one of those other CBS states. If so, when we measure at access point P along the
x-basis,
1. a measurement of 0 would imply that f was constant, and
2. a measurement of x, for x > 0, would imply that f was balanced.
This is because measuring any CBS state along its own basis gives, with 100% probability, the value of that state; that states amplitude is 1 and all the rest of the CBS
states amplitudes are 0.
The Bad News. Alas, we are not able to assert that all balanced f s will produce
x-CBS kets since there are more ways to distribute the + and signs equally than
there are x-CBS kets.
The Good News. We do know something that will turn out to be pivotal: a
balanced f will never have the CBS ket, |0in in its expansion. Lets prove it.
If we give the data registers state at access point P the name |in ,
n
|i
2X
1
1
(1)f (y) |yin ,
n
2 y=0
then we know that a balanced f means that
1
|in =
equal numbers of +1 and 1 .
..
2n
.
1
Furthermore, its |0in coefficient is given by the dot-with-the-basis-ket trick (all coefficients are real, so we can use a simple dot-product),
1
1
1
1
1
1
1
n
n
..
h0
|
i
=
. .
.
n ..
2n
2
1
1
1
1
Aside from the scalar factors 1/ 2, the left vector has all 1s, while the right vector
has half +1s and half 1s, i.e., their dot product is 0: we are assured that there is
no presence of the 0th CBS ket |0in in the expansion of a balanced f . X
396
We have shown that the amplitude of the data registers |0in is 0 whenever f is
balanced, and we already knew that its amplitude is 1 whenever f is constant, so
measuring at access point P along the x-basis will
collapse to |0in if f is constant, guaranteed, and
never collapse to |0in if f is balanced, guaranteed.
Conclusion: If we measure along the x-basis at access point P, a reading of 0
means constant and a reading of x, for any x > 0, means balanced.
Measurement
The x-basis measurement we seek is nothing more than a z-basis measurement after
applying the nth order x z basis transforming unitary H n . This explains the
final nth order Hadamard gate in the upper right (dashed box),
|0in
H n
H n
Uf
|1i
(ignore)
After that, when we measure the data register it will either be
|0in , which corresponds to the pre-H n state of |0in , and therefore indicated a
constant f , or
anything else, which corresponds to a pre-H n state that did not contain even
trace amounts of |0in before the final gate and therefore indicates a balanced
f.
The argument led to the same conclusion but forced us to think about the direct
output of the oracle in terms of the x-basis, thereby guiding the decision to apply the
final Hadamard gate. Not only that, we get a free algorithm out of it, and Ill let you
guess what it is.
A New Problem and a New Algorithm: You Discover it.
[Exercise. While not all balanced functions lead to x CBS kets at access point P,
several do. Describe them in words or formulas.]
Lets call the collection of functions in the last exercise Bx .
[Exercise. How many functions are in the set Bx ?]
[Exercise. If you are told that an unknown function is in the set Bx , formulate
an algorithm using the Deutsch-Jozsa circuit that will, in a single evaluation of Uf ,
determine the entire truth table of the unknown function.]
397
[Exercise. How many evaluations of the unknown function f would be needed
to do this deterministically using a classical approach?]
[Exercise. If you were to allow for a non-deterministic outcome classically, would
you be able to get a constant time solution (one whose number of evaluations of f
would be independent of n for a fixed error, )?]
[Exercise. After attempting this problem, read the next section (BernsteinVazirani) and compare your results and algorithm with that seemingly distinct problem. Are the two truly different problems and algorithms or is there a relationship
between them?]
13.5
True Non-Deterministic Speed-Up: The BernsteinVazirani Problem
The quantum Deutsch-Jozsa algorithm offers a deterministic exponential speed-up
over the classical algorithm, but realistically when we accept a small error, both it
and its classical alternative are constant time, i.e, they can be solved in a fixed
time independent of the number of inputs. We saw this when analyzing the classical
algorithm.
The first problem that shows a clear separation in time complexity is the BerstainVazirani problem, in which classically, even when one accepts an error the solution
still grows in time with the number of inputs, n. Meanwhile, the quantum solution
is constant time only one evaluation of the oracle, Uf , is needed regardless of the
number of inputs.
Reminder. Well cover probability and time complexity in formal lessons, but
for our current purpose we can let intuition guide our computations, just as we did
earlier with classical Deutsch-Jozsa analysis.
We continue to study functions that have Boolean inputs and outputs, specifically
n binary inputs and one binary output,
f : {0 , 1}n {0 , 1} .
The Bernstein-Vazirani Problem, Given an unknown function,
f ( xn1 , xn2 , . . . , x1 , x0 ) ,
of n inputs that are known to be defined by a mod-2 dot product with an
n (binary) digit constant, a,
f (x)
a x,
find a in one query of the quantum oracle, Uf .
398
13.5.1
The Bernstein-Vazirani Algorithm
The algorithm uses the same circuit with the same inputs and the same single data
register measurement as Deutsch-Josza. However this time, instead of asking whether
we see a 0 or a non-0 at the output, we look at the full output: its value will be
our desired unknown, a.
The Circuit
For quick reference, here it is again:
|0in
H n
H n
Uf
|1i
(ignore)
The |0i going into the top register provides the quantum parallelism and the |1i
into the bottom offers a phase kick-back that transfers information about f from the
target output to the data output.
Preparing the Oracles Input: The Two Left Hadamard Gates
Same as Deutsch-Jozsa. The first part of the circuit prepares states that are needed
for quantum parallelism and phase kick-back,
|0in
H n
H n
Uf
|1i
(ignore)
The two registers going into the oracle are, again
n
|0i
2X
1
1
|yin
n
2 y=0
H |1i
|0i |1i
= |i .
2
and
Analyzing the Oracle
The oracle (dashed box) has the same output,
|0in
H n
H n
Uf
|1i
(ignore)
H
399
namely,
n
|0in
Uf
|i
2X
1
1
(1)f (y) |yin
n
2 y=0
!
|i ,
but we can plug in the exact formula for f (y),
Uf |0in |i
!
n 1
2X
1
(1)a y |yin |i ,
n
2 y=0
The Final Hadamard Gate
We apply the final nth order Hadamard gate. At access point P,
|0in
H n
H n
Uf
|1i
(ignore)
H
the data register holds the ket
n
2X
1
1
(1)a y |yin
n
2 y=0
while at point Q, it ends up as (refer to the individual steps in the Deutsch-Josza
derivation, with a y in place of f (y)),
!
n 1
2X
1
n
H n
(1)a y |yi
2n y=0
1
2n
n 1
2X
z=0
n 1
2X
!
|zin
(1)a y (1)y z)
y=0
{z
G(z)
which also defines a scalar function G(z), used to simplify the analysis. So, the final
output is an expansion along the z-basis,
n
data register at access point Q
Consider G(z) in the two cases:
400
2 1
1 X
G(z) |zin .
n
2
z=0
z = a.
G(a)
n 1
2X
a y
(1)
(1)
y a
n 1
2X
y=0
2n ,
y=0
so the amplitudes of the CBS ket |ai is
G(a)
2n
2n
2n
1.
z 6= a. We dont even have to sweat the computation for the amplitudes for the
other kets, because once we know that |ai has amplitude 1, the others have to
be 0. (Why?)
We have shown that at access point Q, the CBS state |ai is sitting in the data register.
Since it is a CBS state, it wont collapse to anything other than what it already is,
and we are guaranteed to get a reading of a, our sought-after n-bit binary number.
Time Complexity
Because the quantum circuit evaluates Uf only once, this is a constant time solution.
What about the classical solution?
Deterministic. Classically we would need a full n evaluations of f in order to
get all n coordinates of a. That is, we would use the input value
0
..
.
ek 1 kth element .
.
..
0
in order to compute the kth coordinate of a based on
0
a0
.. ..
. .
f (ek ) = 1 ak =
. .
.. ..
0
an1
ak .
After n passes we would have all n coordinates of a and be done. Thus, the classical
algorithm grows linearly with the number of inputs n. This kind of growth is called
linear growth or linear time complexity as it requires longer to process more inputs,
but if you double the number of inputs, it only requires twice as much time. This
is not as bad as the exponential growth of the classical deterministic Deutsch-Jozsa
algorithm.
401
Alternatively, we can measure the classical deterministic solution to the current
problem in terms of the encoded integer size, N = 2n . In that case the classical
algorithm is logarithmic in N , which doesnt sound as bad as linear, even though this
is just a different accounting system.)
Non-Deterministic. What if, classically, we evaluate f a fixed number of times,
M , and allow for some error, close to 0? Can we succeed if M is independent of
the number of inputs, n? No. In fact, even if allowed M to grow with n by taking
M = n 1, we would still be forced to guess at the last coordinate. This would
produce a 50% error since the last coordinate could be 1 or 0 with equal probability.
We cant even make a good guess (small error close to 0) if we skip a measely one of
the n evaluations, never mind skipping the many evaluations that would be hoisted
on us if we let M be constant and watched n grow far beyond M .
So in practical terms, the classical solution is not constant time, and we have
a clear separation between quantum and classical solutions to the question. This
is a stronger result than quantum computing provided to the Duetsch-Josza problem where, when we allowed a small error, there was no real difference between the
quantum and classical solutions.
13.6
Generalized Born Rule
We cant end this lesson without providing a final generalization of Traits #15 and
#15, the Born rule for bipartite and tripartite systems. Well call it Traits #15,
the generalized Born rule. The sentiment is the same as its smaller order cousins.
In rough language it says that when we have a special kind of sum of separable
states from two high-dimensional spaces, A and B, an A-measurement will cause the
overall state to collapses to one of the separable terms, thereby selecting the B-state
of that term.
13.6.1
Trait #15 (Generalized Born Rule for (n+m)th order
States)
Assume that we have an (n + m)th order state, |in+m , in the product space A B
= H(n) H(m) with the property that |in+m can be written as the following kind
of sum:
|in+m
n
m
n
m
n
|0inA |0 im
B + |1iA |1 iB + + |2 1iA |2n 1 iB .
In this special form, notice that each term is a separable product of a distinct CBS
ket from A and some general state from B, i.e., the kth term is
|kinA |k im
B .
402
We know by QM Trait #7 (post-measurement collapse), that the state of the component space A = H(n) must collapse to one of the CBS states, call it
|k0 inA .
The generalized Born rule assures us that this will force the component space B =
H(m) to collapse to the matching state,
|k0 im
B ,
only it will become normalized after the collapse to the equivalent
|k0 im
p
.
hk0 | k0 i
(Note that in the last expression I suppressed the superscripts m in the denominator
and subscripts, B, everywhere, to avoid clutter.)
Discussion
The assumption of this rule is that the component spaces A and B are in an entangled
state which can be expanded as a sum, all terms of which have A basis factors. Well,
any state in A B can be expressed expressed this way; all we have to do is express
it along the full 2n+m product basis kets, then collect terms having like A-basis kets
and factor out the common ket in each term. So the assumption isnt so much about
the state |in+m as it is about how the state is written.
The next part reminds us that when an observer of the state space A takes a
measurement along the natural basis, her only possible outcomes are one of the 2n 1
basis kets:
{ |0inA , |1inA , |2inA , . . . , |2n 1inA } ,
so only one term in the original sum survives. That term tells us what a B-state
space observer now has before him:
n
A & |0i
A & |1i
A & |2i
&
|0 im
p
,
h0 | 0 i
&
| im
p 1
,
h1 | 1 i
&
| im
p 2
,
h2 | 2 i
and, in general,
A & |kin
&
|k im
p
hk | k i
403
for k = 0, 1, . . . , 2n 1 .
This does not tell us what a B-state observer would measure, however, since the state
he is left with, call it
| im
p k0
,
hk0 | k0 i
is not assumed to be a CBS ket of the B space. It is potentially a superposition
of CBS kets, itself, so has a wide range of possible collapse probabilities. However,
just knowing the amplitudes of the state |k0 im
B (after normalization) narrows down
what B is likely to find. This measurement collapse, forced by an observer of A, but
experienced by an observer of the entangled B, is the crux of the remainder of the
course.
Size of the Superposition. Weve listed the sum as potentially having the
maximum number of 2n terms, based on the underlying assumption that each term
has an A-basis ket. However, it often has fewer terms, in which case the rule still
applies, only then there are fewer collapse possibilities.
Role of A and B. There was nothing special about the component state space
A. We could have expanded the original state,
|in+m
in such a way that each term in the sum had a B-space CBS ket, |kim
B and the A-space
n
partners were general states, |k iA .
13.7
Towards Advanced Quantum Algorithms
Even though Bernstein V azirani provided a separation between quantum and
classical computing, the classical problem was still easy, meaning it did not grow
exponentially with n; the improvement is not yet dramatic enough to declare quantum
computing a game changing technology. For that, we will need to study two landmark
algorithms,
the quantum solution to Simons problem, and
the quantum solution to Shors problem.
For those we need to add a little more math to our diet, so we take a small-butinteresting side trip in next time.
404
Chapter 14
Probability Theory
14.1
Probability in Quantum Computing
14.1.1
Probability for Classical Algorithms
In the few quantum algorithms weve had, the quantum circuit solved the problem in
a single pass of the circuit: one query of the oracle. We didnt need probability for
that. We did use probability, informally, to estimate how long a classical algorithm
would take to complete if we allowed for experimental error. In the Deutsch-Jozsa
problem, when we considered a classical solution using the M -and-guess method, we
expressed the event of failure (all trial outcomes were the S ame yet (read and)
f is Balanced ) by
S B,
and argued intuitively that the probability of this happening was
P (S B)
1
.
2M
That analysis was necessary to evaluate the worthiness of the quantum alternatives
constant-time solution.
14.1.2
Probability for Quantum Algorithms
Soon, we will study quantum algorithms that require multiple queries of a quantum
circuit and result in an overall performance which is non-deterministic, i.e., probabilistic. Heres a preview of a small section of Simons quantum algorithm:
... (previous steps) ...
Repeat the following loop at most n + T times or until we get n 1 linearly
independent vectors, whichever comes first.
405
... (description of loop) ...
If the above loop ended after n + T full passes, we failed.
Otherwise, we succeeded. Add an nth vector, wn1 , which is ...
... (following steps) ...
Estimating the probabilities of this algorithm will require more than intuition; it will
require a few probability laws and formulas.
Today well cover those laws and formulas. Well use them to solidify our earlier
classical estimations, and well have them at-the-ready for the upcoming probabilistic
quantum algorithms.
14.2
The Essential Vocabulary: Events vs. Probabilities
When we build and analyze quantum circuits, well be throwing around the terms
event and probability. Events (word-like things) are more fundamental than probabilities (number-like things). Here is an informal description (rigor to follow).
14.2.1
Events
An event is something that happens, happened, will happen or might happen.
Events are described using English, Mandarin, Russian or some other natural
language. They are not numbers, but descriptions. There is no such thing as the
event 8. There is the event that Salim rolls an 8 at dice, Or Han missed the 8 PM
train.
The Higgs particle might decay into four muons.
An application error occurred at 6:07 PM.
Ms. Park shot a 1-under-par yesterday at Pebble Beach.
406
We will often use a script letter like A , B, C , . . . to designate events, as in
... Let E be the event that two sample measurements of our quantum circuit
are equal ...,
... Let I be the event that all the vectors we select at random are linearly
independent ..., or
... Let C be the event that the number given to us is relatively prime to 100
....
14.2.2
Probabilities
Probabilities are the numeric likelihoods that certain events will occur. They are
always positive numbers between 0 and 1, inclusive. If the probability of an event is
0, it cannot occur, if it is 1, it will occur with 100% certainly, if it is .7319, it will
occur 73.19% of the time, and so on. We express the probabilities of events using P ()
notation, like so
P (Ms. Park will shoot a birdie on the 18th hole)
P (at least one measurement is > -13.6eV)
=
=
.338,
.021 .
or
In the heat of an analysis, well be using the script letters that symbolize the events
under consideration. They will be defined in context, so youll always know what
they stand for. The corresponding probabilities of the events will be expressed, then,
using syntax like
P (E )
P (I )
P (C )
14.3
>
.99996,
.1803, or
.99996 .
The Quantum Coin Flip
Many probability lessons use the familiar coin flip as a source of elementary examples.
But we quantum computer scientists can use a phenomenon which behaves exactly like
a coin flip but is much more interesting: a measurement of the quantum superposition
state
|0i + |1i
.
2
407
If we prepare this state by applying a Hadamard gate, H, to the basis state |0i, our
coin is waiting at the output gate. In other words, the coin is H |0i:
|0i
|0i + |1i
.
2
[Exercise. We recognize this state under an alias; it also goes by the name |+i.
Recall why.]
Measuring the output state H |0i is our actual toss. It causes the state to collapse
to either |0i or |1i, which we would experience by seeing a 0 or 1 on our meter.
|0i + |1i
2
Here, &
& |0i or |1i
means collapses to.
Thats equivalent to getting a heads or tails. Moreover, the probability of getting
either one of these outcomes (the eigenvalues) is determined by the amplitudes
of
their respective CBS kets (the eigenvectors). Since both amplitudes are 1/ 2, the
probabilities are
1 2
1
=
,
and
P (measuring 0) =
2
2
1 2
1
P (measuring 1) =
=
.
2
2
So we have a perfectly good coin. Suddenly learning probability theory seems a lot
more appealing, and as a bonus well be doing a little quantum computing along the
way.
[Exercise. We could have used H |1i as our coin. Explain why.]
[Exercise. The above presupposes that we will measure the output state along
the z-basis, {|0i , |1i}. What happens if, instead, we measure the same state along
the x-basis, {|+i , |i}. Can we use this method as a fair coin flip?]
14.4
Experimental Outcomes and the Sample Space
We now give the definitions of events and probabilities a bit more rigor.
408
14.4.1
Outcomes
Every experiment or set of measurements associated with our quantum algorithms
will consist of a set of all possible outcomes. An outcome is the most basic result we
can imagine.
A Single Qubit Coin Flip
If the experiment is to prepare exactly one quantum coin, i.e., the state H |0i, and
then measure it, there are two possible outcomes:
Meter reads 0 state collapsed to |0i and
Meter reads 1 state collapsed to |1i.
Usually, though, we simply say that the possible outcomes are 0 and 1.
A Ten Qubit Coin Flip
A more involved experiment would be to prepare ten identical-but-distinct quantum
coins, labeled # 0 - # 9 (because we are computer scientists), then measure each one,
resulting in ten measurements.
10
quantum
coins
#0 :
1:
|0i
|0i + |1i
|0i
|0i + |1i
|0i + |1i
..
.
9:
|0i
Now, defining the outcomes isnt so clear.
Maybe we care about the number of 0s and 1s, but not which preparations (say,
1, # 7, # 9) measured 0 and which (say, # 0, # 2, # 3, # 4, # 5, # 6, # 8) measured 1. The
outcomes could be defined by 11 possible head counts, where head means getting
a 0:
#
Total 0s = 0
Total 0s = 1
Total 0s = 2
409
..
.
Total 0s = 10
Again, we could abbreviate this by saying, the possible outcomes are 0-10.
One problem with this definition of outcome is that some are more likely than
others. It is usually beneficial to define outcomes so that they are all equally
or nearly equally likely. So, we change our outcomes to be the many ten-tuples
consisting of
Results were (0,0,0,0,0,0,0,0,0,0)
Results were (0,0,0,0,0,0,0,0,0,1)
Results were (0,0,0,0,0,0,0,0,1,0)
..
.
Results were (1,1,1,1,1,1,1,1,1,1)
There are now a lot more outcomes (210 = 1024), but they each have the same
likelihood of happening. (If you dont believe me, list the eight outcomes for three
coins and start flipping.) A shorter way to describe the second breakdown of outcomes
is,
a possible outcome has the form
( x0 , x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 , x9 ),
where xk is the kth measurement result.
14.4.2
Requirements when Defining Outcomes
While there are many ways to partition all the possible results of your experiment,
there is usually one obvious way that presents itself. In the ten qubit coin toss, we saw
that the second alternative had some advantages. But there are actually requirements
that make certain ways of dividing up the would-be outcomes illegal. To be legal, a
partition of outcomes must satisfy two conditions.
1. Outcomes must be mutually exclusive (a.k.a. disjoint).
2. Outcomes must collectively represent every possible result of the experiment.
[Exercise. Explain why the two ways we defined the outcomes of the ten qubit
coin toss are both legal.]
410
14.4.3
An Incorrect Attempt at Defining Outcomes
Its natural to believe that organizing the ten qubit coin toss by looking at each
individual qubit outcome (such as the 4th qubit measured a 0) as being a reasonable
partition of the experiment. After all, there are ten individual circuits. Here is the
(unsuccessful) attempt at using that as our outcome set.
The outcomes are to be the individual measurement results
Z3
measurement of qubit # 3 is 0 ,
Z8
measurement of qubit # 8 is 0 ,
O5
measurement of qubit # 5 is 1 ,
O0
measurement of qubit # 0 is 1 ,
and so on.
(Z would mean that the event detects a Zero, while O, script-O, means the event
detects a One. Meanwhile, the subscript indicates which of the ten measurements we
are describing.)
There are ten measurements, each one can be either zero or one, and the above
organization produces 20 alleged outcomes. However, this does not satisfy the two
requirements of outcome.
[Exercise. Explain why?]
14.4.4
Events
Definitions
Event. An event is a subset of outcomes.
Simple Event. An event that contains exactly one outcome is called a
simple event (a.k.a. elementary event).
Please recognize that outcomes are not events. A set containing an outcome is an
event (a simple one, to be precise).
Compound Event. An event that contains more than one outcome is
called a compound event.
Describing Events
Events can be described either by the actual sets, using set notation, or an English
(or French or Vietnamese) sentence.
Examples of simple event descriptions for our ten qubit coint toss experiment are
411
{ (0, 0, 1, 1, 1, 0, 0, 0, 1, 1) }
{ (1, 1, 1, 1, 1, 1, 0, 0, 1, 0) }
{ (0, 0, 0, 0, 0, 0, 0, 0, 0, 0) }
The first five qubits measure 0 and the last five measure 1.
All ten qubits measure 1.
Qubit
nate.
0 measures 0, and as the qubit
increases, the measurements alter-
Examples of compound event descriptions for our ten qubit coint toss experiment
are
{ (0, 0, 1, 1, 1, 0, 0, 0, 1, 1), (1, 1, 1, 1, 1, 1, 0, 0, 1, 0) }
The first five qubits measure 0.
The fourth qubit measures 1.
As the qubit
increases, the measurements alternate.
[Exercise. Describe five simple events, and five compound events. Use some
set notation and some natural English descriptions. You can use set notation that
leverages formulas rather than listing all the members, individually.]
14.4.5
The Sample Space
Sample Space. The sample space is the set of all possible outcomes.
It is referred to using the Greek .
In our ten qubit coin toss, is the set of all ordered ten-tuples consisting of 0s and
1s,
n
o
=
(x0 , x1 , x2 , . . . , x9 ) xk {0, 1} .
At the other extreme, there is the null event.
Null Event. The event consisting of no outcomes, a.k.a. the empty set, is the
null event. It is represented by the symbol, .
412
14.4.6
Set Operations
Unions
One way to express is as a compound event consisting of the set union of simple
events. Lets do that for the ten qubit coin flip using big- notation, which is just
like summation notation, , only for unions,
{ (x0 , x1 , x2 , . . . , x9 ) } .
xk {0, 1}
Example. We would like to represent the event, F , that the first four quantum
coin flips in our ten qubit experiment are all the same. One expression would be
F =
{ (w, w, w, w, x0 , x1 , . . . , x5 ) } .
w, xk {0,1}
Example. We would like to represent the event, F 0 , that the first four quantum
coin flips in our ten qubit experiment are all the same, but the first five are not all
the same. One expression would be
F0 =
{ (w, w, w, w, w 1, x0 , x1 , . . . , x4 ) } ,
w, xk {0,1}
where is addition mod-2.
[Exercise. Find a representation of the event, E , that the first and last quantum
coin flip in our ten qubit experiment are different.]
[Exercise. Find a representation of the event, O, that the number of 1s in our
quantum coin flip in our ten qubit experiment is odd.]
[Exercise. Find a representation of the event, H , that the Hamming weight, i.e.,
the sum of the outcomes, of our ten qubit quantum coin flip is 6.]
Intersections
We can use intersections to describe sets, too. For example, if we wanted an event
A in our ten qubit coin flip in which both a) the first four are the same and also b)
the first and last are different, we might express that using
A
= F E ,
where F and E were defined by me (and you) a moment ago.
Of course, we could also try representing that event directly, but once we have a
few compound events defined, we often leverage them to produce new ones.
[Exercise. Use intersections and the events already defined above to describe
the event, R, in our ten qubit coin flip in which both a) the Hamming weight has
absolute value 5 and b) the sum of the 1s is odd.
413
Differences
What if we wanted to discuss the ten qubit coin flip event, D, which had odd sums,
but whose first four flips are not equal ? We could leverage our definitions of O and
F and use difference notation, or \, like so:
D = O F or
D = O \F .
That notation instructs the reader to start with O, then remove all the events that
satisfy (or are ) F .
Complements and the Operator
When we start with the entire sample space , and subtract and event, S ,
S ,
we use a special term and notation: the complement of S , written in several ways,
depending on the author or context,
S0,
S,
Sc,
S .
or
All are usually read not S , and the last reprises the logical negation operator, .
14.5
Alternate Views of the Ten Qubit Coin Flip
This is a good time to introduce other ways to view the outcomes of this ten qubit
experiment that will help us when we get to some quantum algorithms later in the
course.
A 10-Dimensional Space With Coordinates 0 and 1
For each outcome,
(x0 , x1 , x2 , . . . , x9 ),
we look at it as a ten component vector whose coordinates are 0 or 1. The vectors
can be written in either row or column form,
x0
x1
(x0 , x1 , x2 , . . . , x9 ) = x2 .
..
.
x9
414
An outcome is already in a vector-like format so its not hard to see the correspondence. For example, the outcome (0, 1, 0, . . . , 1) corresponds to the vector
0
1
0
.
..
.
1
This seems pretty natural, but as usual, Ill complicate matters by introducing the
addition of two such vectors. While adding vectors is nothing new, the concept
doesnt seem to have much meaning when you think of them as outcomes. (What
does it mean to add two outcomes of an experiment?) Lets not dwell on that for
the moment but proceed with the definition. We add vectors in this space by taking
their component-wise mod-2 sum or, equivalently, their component-wise XOR,
0
1
01
1
1
0
1 0
1
0
1
0 1
+
= 1 .
..
..
..
..
.
.
.
.
1
1
1 1
0
To make this a vector space, Id have to tell you its scalars, (just the two numbers
0 and 1), operations on the scalars (simple multiplication, and mod-2 addition,
), etc. Once the details were filled in we would have the ten dimensional vectors
mod-2, or (Z2 )10 . The fancy name expresses the fact that the vectors have 10
components (the superscript 10 in (Z2 )10 ), in which each component comes from the
set {0, 1} (the subscript 2 in (Z2 )10 ). You might recall from our formal treatment of
classical bits that we had a two dimensional mod-2 vector space, B = B2 , which in
this new notation is (Z2 )2 .
We can create mod-2 vectors of any dimension, of course, like the five-dimensional
(Z2 )5 or more general n-dimensional (Z2 )n . The number of components 10, 5 or n,
tells you the dimension of the vector space.
The Integers from 0 to 210 1
A second view of a ten qubit coin flip is that of a ten bit integer from 0 to 1023,
constructed by concatenating all the results,
x0 x1 x2 x9 .
415
Examples of the correspondence between these last two views are:
mod-2 vector
(0, 0, . . . 0, 0, 0)
(0, 0, . . . 0, 0, 1)
(0, 0, . . . 0, 1, 0)
(0, 0, . . . 0, 1, 1)
(0, 0, . . . 1, 0, 0)
..
.
(x0 , x1 , x2 , . . . x9 )
binary
00 000
00 001
00 010
00 011
00 100
..
.
x0 x1 x2 , x9
integer
0
1
2
3
4
..
.
x
This gives us additional vocabulary to describe events. A few examples of some
interesting event-descriptions are:
All outcomes between 500 and 700. (Uses integer interpretation.)
All outcomes divisible by 7. (Uses integer interpretation.)
All outcomes which are relatively prime (a.k.a. coprime) to a given number.
(Uses integer interpretation.)
All outcomes that are linearly independent of the set (event) containing the
three vectors
(0, . . . 0, 0, 1), (0, . . . 0, 1, 0) , (0, . . . 1, 0, 0)
(Uses vector interpretation.) See below, for a review of linear independence.
The last two are of particular importance when we consider quantum period-finding.
14.6
Mod-2 Vectors in Probability Computations
for Quantum Algorithms
In order to manufacture some particularly important examples for our upcoming
algorithms, we will take a moment to shore-up our concept of linear independence,
especially as it pertains to the mod-2 vector space (Z2 )n .
14.6.1
Ordinary Linear Independence and Spanning
Definition of Linear Independence (Review)
416
Linear Independence. In a vector space, a set of vectors {v0 , . . . , vn1 }
is linearly independent if you cannot form a non-trivial linear combination of them which produces the zero vector, i.e.,
c0 v0 + c1 v1 + . . . + cn1 vn1
=
all ck = 0.
Another way to say this is that no vector in the set can be expressed as a linearcombination of the others.
[Exercise. Prove that the last statement is equivalent to the definition.]
[Exercise. Show that the zero vector can never be a member of a linearly independent set.]
[Exercise. Show that a singleton (a set consisting of any single non-zero vector)
is a linearly independent set.]
Definition of Span of a Set of Vectors (Review)
The span of a set of vectors is the (usually larger) set consisting of all vectors that
can be constructed by taking linear combinations of the original set.
The Span of a Set of Vectors. The span of a set of m vectors S =
{vk }m1
k=0 is
n
o
c0 v0 + c1 v1 + . . . + cm1 vm1 ck are scalars .
The set S does not have to be a linearly independent set. If it is not, then it means
we can omit one or more of its vectors without reducing its span.
[Exercise. Prove it.]
When we say that a vector, w, is in the span of S, we mean that w can be written
as a linear combination of the vk s in S. Again, this does not require that the original
{vk } be linearly independent.
[Exercise. Make this last definition explicit using formulas.]
When a vector, w, is not in the span of S, adding w to S will increase Ss span.
14.6.2
Linear Independence and Spanning in the Mod-2 Sense
Because the only scalars available to weight each vector in a mod-2 linear combination are 0 and 1, the span of any set of mod-2 vectors reduces to simple sums.
The Span of a Set of Mod-2 Vectors. The span of a set of m mod-2
vectors S = {vk }m1
k=0 is
n
o
vl 0 + vl 1 + . . . + vl s vl k S .
417
Abstract Example
Say we have four mod-2 vectors (of any dimension),
S
{v0 , v1 , v2 , v3 } .
Because the scalars are only 0 or 1, vectors in the span are all possible sums of these
vectors. If one of the vectors doesnt appear in a sum, thats just a way to say its
corresponding weighting scalar is 0. If it does appear, then its weighting scalar is 1.
Here are some vectors in the span of S:
0
v2
v0
v0
v1
(the zero vector)
+ v3
+ v1 + v2 + v3
+ v1 ( = 0, also the zero vector)
[Exercise. For a mod-2 vector space, how many vectors are in the span of the
empty set, ? How many are in the span of {0}? How many are in the span of a set
consisting of a single (specific) non-zero vector? How many are in the span of a set
of two (specific) linearly independent vectors? Bonus: How many in the span of
a set of m (specific) linearly independent vectors? Hint: If youre stuck, the next
examples will help sort things out.]
Concrete Example
Consider the two five-dimensional vectors in (Z2 )5
(1, 0, 0, 1, 1), (1, 1, 1, 1, 0) .
Their span is all possible sums (and lets not forget to include the zero vector):
0 =
(1, 0,
(1, 1,
(1, 0,
(0, 0, 0, 0, 0)
0, 1, 1)
1, 1, 0)
0, 1, 1) + (1, 1, 1, 1, 0) = (0, 1, 1, 0, 1)
Two Linearly Independent Vectors
A set of two mod-2 vectors, {v0 , v1 } (of any dimension, say 10), are linearly independent in the mod-2 sense when
neither vector is zero (0), and
the two vectors are distinct, i.e., v0 6= v1 .
[Exercise. Prove that this follows from the definition of linear independence using
the observation that the only scalars available are 0 and 1.]
418
Finding a Third Independent Vector Relative to an Independent Set of
Two
Consider the same two five-dimensional vectors in (Z2 )5
(1, 0, 0, 1, 1), (1, 1, 1, 1, 0) .
This set is linearly independent as we know from the above observations. If we wanted
to add a third vector, w, to this set such that the augmented set of three vectors,
(1, 0, 0, 1, 1), (1, 1, 1, 1, 0), w ,
would also be linearly independent, what would w have to look like? It is easiest
to describe the illegal vectors, i.e., those w which are linearly dependent on the two,
then make sure we avoid those. For w to be linearly dependent on these two, it would
have to be in their span. We already computed the span and found it to be the four
vectors
(0,
(1,
(1,
(0,
0,
0,
1,
1,
0,
0,
1,
1,
0,
1,
1,
0,
0) ,
1) ,
0) and
1) .
Thus, a vector w is linearly independent of the original two exactly when it is not in
that set,
w
/
0, (1, 0, 0, 1, 1), (1, 1, 1, 1, 0), (0, 1, 1, 0, 1) ,
or taking the complement, when
w
0, (1, 0, 0, 1, 1), (1, 1, 1, 1, 0), (0, 1, 1, 0, 1) .
(I used the over-bar, E , rather than the E c notation to denote complement, since the
tends to get lost in this situation.)
How many vectors is this? Count all the vectors in the space (25 = 32) and
subtract the four that we know to be linearly dependent. That makes 32 4 = 28
such w independent of the original two.
Discussion. This analysis didnt depend on which two vectors were in the original
linearly independent set. If we had started with any two distinct non-zero vectors,
they would be independent and there would be 28 ways to extend them to a set of
three independent vectors. Furthermore, the only role played by the dimension 5 was
that we subtracted the 4 from 25 = 32 to get 28. If we had been working in the
7-dimensional (Z2 )7 and asked the same question starting with two specific vectors,
we would have arrived at the conclusion that there were 27 4 = 128 4 = 124
vectors independent of the first two. If we started with two linearly independent 10dimensional vectors, we would have gotten 210 4 = 10244 = 1020 choices. And if we
419
started in the space of 3-dimensional vectors, (Z2 )3 , there would be 23 4 = 8 4 = 4
independent w from which to choose.
[Exercise. If we had two independent vectors in the 2-dimensional (Z2 )2 , how
many ways would there have been to select a third vector linearly dependent from
the first two?]
14.6.3
Probability Warm-Up: Counting
Were trying to learn how to count, a skill that every combinatorial mathematician
needs to master, and one that even we computer scientists would do well to develop.
Finding a Third Vector that is Not in the Span of Two Random Vectors
Lets describe an event that is more general than the one in the last example.
E the event that, after selecting two 5-dimensional mod-2 vectors
x and y at random, a third selection w will not be in the span of the first
two.
Notation. Since we are selecting vectors in a specific sequence, well use the
notation
x, y
to represent the event where x is the first pick and y is the second pick. (To count
accurately, we must consider order, which is why we dont use braces: { or }.)
Similarly, the selection of the third w after the first two could be represented by
x, y, w .
How many outcomes are in event E ?
This time, the first two vectors are not a fixed pair, nor are they required to be
linearly independent. We simply want to know how many ways the third vector (or
qubit coin flip if you are remembering how these vectors arose) can avoid being in
the span of the first two. No other restrictions.
We break E into two major cases:
1. x and y form a linearly independent set (mostly answered), and
2. x and y are not linearly independent. This case contains three sub-cases:
(i) Both x and y are 0,
(ii) exactly one of x and y is 0, and
(iii) x = y, but neither is 0.
420
Lets do the larger case 2 first, then come back and finish up what we started
earlier to handle case 1.
Harder Case 2. For x and y not linearly independent, we count each sub-case
as follows.
(i) There is only one configuration of x and y in this sub-case, namely 0, 0 . In
such a situation, the only thing we require of w is that it not be 0. There
are
32 1 = 31 such w. Therefore, there are 31 simple events, 0, 0, w , in this
case.
(ii) In this sub-case there are 31 configurations of the form 0, y y6=0 and 31 of
the form x, 0 x6=0 . Thats a total of 62 ways that random choices of x and y
can lead us to this sub-case. Meanwhile, for each such configuration there are
32 2 = 30 ways for w to be different from x and y. Putting it together, there
are 62 30 = 1860 simple events in this sub-case.
(iii) There are 31 configurations of x = y 6= 0 in this sub-case, namely x, x x6=0 .
Meanwhile, for each such configuration any w that is neither 0 nor x will work.
There are 32 2 = 30 such w. Putting it together, there are 31 30 = 930
simple events in this sub-case.
Summarizing, the number of events with x and y not linearly independent and w not
in their span is 31 + 1860 + 930 = 2821.
Easier Case 1. We get into this situation with x, y x6=y6=0 . That means the
first choice cant be 0 so there are 31 possibilities for x, and the second one cant be
0 or x, so there are 30 choices left for y. Thats 31 30 ways to get into this major
case. Meanwhile, there are 28 ws not in the span for each of those individual outcomes
(result of last section), providing the linearly-independent case with 313028 = 26040
simple events.
Combining Both Cases. We add the two major cases to get 2821 + 26040 =
28861 outcomes in event E . If you are thinking of this as three five qubit coin flip
outcomes 15 individual flips, total there are 28861 ways in which the third group
of five will be linearly independent of the first two.
Is this likely to happen? How many possible flips are there in the sample space
? The answer is that there are 215 (or, if you prefer, 323 ) = 32768. (The latter
expression comes from 3 five qubit events, each event coming from a set of 25 = 32
outcomes.) That means 28861 of the 32768 possible outcomes will result in the third
outcome not being in the span of the first two. A simple division tells us that this
will happen 88% of the time.
A third vector selected at random from (Z2 )5 is far more likely to be independent of any two previously selected vectors than it is to be in their
span.
This was a seat-of-the-pants calculation. Wed better get some methodology so
we dont have to work so hard every time we need to count.
421
14.7
Fundamental Probability Theory
14.7.1
The Axioms
Consider the set of all possible events of an experiment,
Events
{ E }.
A probability measure of an experiment is a function, P (), with domain = Events and
range = non-negative real numbers,
P : Events R0 ,
which satisfies the following three axioms:
1. All P () values are non-negative (already stated),
P (E )
0.
2. The probabilities of something happening is certain,
P ()
1.
3. The probabilities of mutually exclusive (disjoint) events, E0 , . . . , En1 , can be
added,
!
n1
n1
X
[
P (Ek ) .
Ek
=
P
k=0
k=0
[Exercise. Prove the following consequences from the above axioms:
For any event, P (E ) 1.
P ()
0.
If E F , P (E ) P (F ). ]
14.7.2
Assigning Probabilities in Finite, Equiprobable, Sample Spaces
Definitions
It may not always be obvious how to assign probabilities to events even when they
are simple events. However, when the sample space is finite and all simple events are
equiprobable, we can always do it. We just count and divide.
422
Caution. There is no way to prove that simple events are equiprobable. This
is something we deduce by experiment. For example, the probability of a coin flip
coming up tails (or the z-spin measurement of an electron in state |0ix being 1 ) is
said to be .5, but we dont know that it is for sure. We conclude it to be so by doing
lots of experiments.
Size of an Event. The size of an event, written |E |, is the number of
outcomes that constitute it,
|E |
# outcomes E .
Probability of an Event. The probability of an event, E , when all
simple events are equiprobable, is
P (E )
|E |
.
||
This is not a definition, but a consequence of the axioms plus the assumption that
simple events are finite and equiprobable.
Example 1
In the ten qubit coin flip, consider the event, F , in which the first four results are
all equal. We compute the probability by counting how many simple events (or,
equivalently, how many outcomes) meet that criterion. What we know about these
events is that the first four are the same, so they are either all 0 or all 1.
All 0.
form,
The number of events in this case is the number of ten-tuples of the
(0, 0, 0, 0, x0 , x1 , . . . , x5 ) ,
which we can see is the same as the number of integers of the form x0 x1 x2 x3 x4 x5 .
Thats 000000 through 111111, or 0 63 = 64.
All 1.
form,
The number of events in this case is the number of ten-tuples of the
(1, 1, 1, 1, x0 , x1 , . . . , x5 ) ,
also equal to 64.
So the number of outcomes in this event is 128. Meanwhile the sample space has
1024 outcomes, giving
P (F )
128
1024
=
423
1/8
.125 .
Example 2
c in which the first four individual flips come up the same
Next, consider the event F
(= Example 1), but with the additional constraint that the fifth qubit be different
from those four. Now, the outcomes fall into the two categories
Four 0s followed by a 1.
(0, 0, 0, 0, 1, x0 , x1 , . . . , x4 ) ,
Four 1s followed by a 0.
(1, 1, 1, 1, 0, x0 , x1 , . . . , x4 ) ,
Again, we look at the number of possibilities for the free range bits x0 , . . . , x4 ,
which is 32 for each of the two categories, making the number of outcomes in this
event 32 + 32 = 64, so the probability becomes
c)
P (F
64
1024
1/16
.0625 .
Example 3
We do a five qubit coin flip twice. That is, we measure five quantum states once,
producing a mod-2 vector, x = (x0 , x1 , . . . , x4 ), then repeat, getting a second vector,
y = (y0 , y1 , . . . , y4 ). Its like doing one ten qubit coin flip, but we are organizing
things naturally into two equal parts. Instead of the outcomes being single vectors
with ten mod-2 components, outcomes are pairs of vectors, each member of the pair
having five mod-2 components,
o
n
(x0 , x1 , . . . , x4 ), (y0 , y1 , . . . , y4 ) xk , yk {0, 1}
or, more concisely,
n
o
x, y x, y (Z2 )5 .
Consider the event, I , in which the two vectors form a linearly independent set.
Weve already discussed the exact conditions for a set of two mod-2 vectors to be
linearly independent:
neither vector is zero (0), and
the two vectors are distinct, i.e., x 6= y.
We compute P (I ) by counting events. For an outcome ( x, y ) I , x can be
any non-0 five-tuple, (x0 , x1 , . . . , x4 ), and y must be different from both 0 and x.
424
That makes 31 vectors x 6= 0, each supporting 30 linearly independent ys. Thats
31 30 = 930 outcomes in I . || continues to be 1024, so
P (I )
|I |
||
930
1024
.908 .
As you can see, it is very likely that two five qubit coin flips will produce a linearly
independent set; it happens > 90% of the time.
[Exercise. Revise the above example so that we take three, rather than two, five
qubit coin flips. Now the sample space is all triples of these five-tuples, (x, y, w).
What is the probability of the event, T , that all three flip-tuples are linearly
independent? Hint: We already covered the more lenient case in which x and y were
allowed to be any two vectors and w was not in their span. Repeat that analysis but
exclude the cases where x and y formed a linearly dependent set. ]
[Exercise. Make up three interesting event descriptions in this experiment and
compute the probability of each.]
14.8
Big Theorems and Consequences of the Probability Axioms
14.8.1
Unions
The third axiom tells us about the probability of unions of disjoint events, {Ek },
namely,
!
n1
n1
X
[
=
P (Ek ) .
P
Ek
k=0
k=0
What happens when the events are not mutually exclusive? A simple diagram in the
case of two events tells the story. If we were to add the probabilities of two events
E and F which had non-empty intersection, we would be counting the intersection
twice. To fix the error we just subtract off the probability of the intersection,
P (E F )
P (E ) + P (F ) P (E F ) .
425
When there are more than two sets, the intersections involve more combinations and
get harder to write out. But the concept is the same, and all we need to know is that
!
n1
n1
[
X
P
Ek
=
P (Ek ) P (various intersections) .
k=0
k=0
The reason this is always enough information is that we will be using the formula to
bound the probability from above, so the equation, as vague as it is, clearly implies
!
n1
n1
[
X
P
Ek
P (Ek ) .
k=0
14.8.2
k=0
Conditional Probability and Bayes Law
Conditional Probability
Very often we will want to know the probability of some event, E , under the assumption that another event, F , is true. For example, we might want to know the
probability that three quantum coin flips are linearly independent under the assumption that the first two are (known to be) linearly independent. The notation for the
event E given F is
E F ,
and the notation for the events probability is
P E F .
This is something we can count using common sense. Start with our formula for an
event in a finite sample space of equiprobable simple events (always the setting for
us),
P (E )
|E |
.
||
Next, think about what it means to say under the assumption that another event,
F , is true. It means that our sample space, , suddenly shrinks to F , and in that
smaller sample space, we are interested in the probability of the event E F , so
P E F
|E F |
,
|F |
whenever F 6= .
Of course, if F = everything is zero so there is no formula needed.
426
Bayes Law
Study this last expression until it makes sense. Once you have it, divide the top
and bottom by the size of our original sample space, ||, to produce the equivalent
identity,
P E F
P (E F )
,
P (F )
whenever P (F ) > 0.
This is often taken to be the definition of conditional probability, but we can view
it as a natural consequence of the meaning of the phrase E given F . It is also a
simplified form of Bayes law, and I will often refer to this as Bayes law (or rule or
formula), since this simple version is all we will ever need.
14.8.3
Statistical Independence
Two Independent Events
Intuitively, two events are statistically independent or simply independent if they
dont affect one another. If one is known to be true, the probability of the other is
unaffected. Stated in terms of conditional probabilities, this is nothing more than the
declaration that
P E F
= P (E ) ,
or, looking at the mirror image,
P F E
P (F ) .
If we substitute the first of these into Bayes law we find
P (E )
P (E F )
,
P (F )
whenever P (F ) > 0 ,
and rearranging gets us
P (E F )
P (E )P (F ),
which, by the way, is true even for the degenerate case, F = . This is the official
definition of statistically independent events, and working backwards you would derive
the intuitive meaning that we started with. In words, two events are independent
the probability of the intersection is equal to the product of the probabilities.
Multiple Independent Events
The idea carries over to any number of events, although the notation becomes thorny.
Its easier to first say it in words, then show the formula. In words,
427
n events are independent
the probability of the intersection [of any subset of the n events] is
equal to the product of the probabilities [of events in that subset].
Formulaically, we have to resort to double indexing.
n events, {Ek }n1
k=0 , are independent
14.8.4
Ek i
=
1 < k0 < k1 <
< kl < n
P (Eki ) .
1 < k0 < k1 <
< kl < n
Other Forms and Examples
Ill list a few useful consequences of the definitions and the formulas derived above.
Unless otherwise noted, they are all easy to prove and you can select any of these as
an exercise.
Events
(E F )c
E (F G )
=
=
E c Fc
(E F ) (E G )
Probabilities
P (E )
P (E F )
P (E F ) + P (E F c )
P E F P (F )
Example 1
In our ten qubit coin flip, what is the probability that the 3rd, 6th and 9th flips are
identical?
Well call the event E . It is the union of two disjoint events, the first requiring
that all three flips be 0, Z , and the second requiring that all three be 1, O,
E
P (E )
= Z O,
with Z O =
=
= P (Z ) + P (O) .
428
There is no mathematical difference between Z and O, so we compute P () of either
one, say O. It is the intersection of the three statistically independent events, namely
that the 3rd, 6th and 9th flips, individually, come up 1, which we call O3 , O6 and
O9 , respectively. The probability of each is, of course, .5, so we get,
P (O)
=
=
=
P (O3 O6 O9 )
P (O3 ) P (O6 ) P (O9 ) .
.5 .5 .5 = .125
Plugging into the sum, we get
P (E )
P (Z ) + P (O)
.125 + .125
= .25 .
Example 2
We do the five qubit coin flip five times. We examine the probability of the event I5
defined as the five five-tuples are linearly independent. Our idea is to write P (I5 )
as a product of conditional probabilities. Let
Ij
event that the vector outcomes
#
0, # 1, . . . # (j 1)
are linearly-independent.
Our goal is to compute the probability of I5 .
(It will now be convenient to use the over-bar notation E to denote complement.)
Combining the basic identity,
P I5
= P I5 I4
+
(exercise: why?), with the observation that
P I5 I4
P I5 I4
=
(exercise: why?), we can write
P I5
= P I5 I4
= P I5 I4 P I4
(exercise: why?).
Now apply that same process to the right-most factor repeatedly, and you end up
with
P I5
= P I5 I4 P I4 I3
P I3 I2
P I2 I1 P I1
=
5
Y
P Ij Ij1 ,
j=1
429
where, the j = 1 term contains the curious event, I0 . This corresponds to no coin-flip
= no vector being selected or tested. Its different from the zero vector = (0, 0, 0, 0, 0)t ,
which is an actual flip possibility; I0 is no flip at all, i.e., the empty set, . But we
know that is linearly independent always because it vacuously satisfies the condition
that one cannot produce the zero vector as a linear combination of vectors from the
set since there are no vectors to combine. So P (I0 ) = 1. Thus, the last factor in
the product is just
P I1 I0
= P I1 ,
but its cleaner to include the conditional probability (LHS of above) in all factors
when using product notation.
[Note. Well be computing the individual factors in Simons algorithm. This is
as far as we need to take it today.]
14.9
Wedge and Vee Notation and Lecture Recap
Some texts and papers often use the alternate notation, instead of for intersections, and instead of for unions. When and are used, is typically the
negation (complement) operator of choice, so the three go together. This is seen more
commonly in electrical engineering and computer science than math and physics. I
introduced the concepts in this lesson using more traditional , , and c notation but
well use both in upcoming lectures. Normally, I will reserve , , and c for non-event
sets (like sets of integers) and , , and for events.
Well take this opportunity to repeat some of the more important results needed
in future lectures using the new notation.
14.9.1
Disjoint Events
The probabilities of mutually exclusive (disjoint) events, E0 , . . . , En1 , can be added,
!
n1
n1
_
X
P
Ek
=
P (Ek ) .
k=0
14.9.2
k=0
Partition of the Sample Space by Complements.
Any event, F , and its complement, F , partition the space, leading to the identity
P E
= P E F
+ P E F .
430
14.9.3
Bayes Law.
Bayes law (or rule or formula) in its simple, special case, can be expressed as
P (E F )
P E F
=
,
whenever P (F ) > 0.
P (F )
14.9.4
Statistical Independence
n events, {Ek }n1
k=0 , are statistically independent
^
Y
P
Eki
=
P (Eki )
for any subset of the indexes, { kl } { 0, 1, . . . , n 1 }.
This completes the probability theory results needed for the Foothill College
courses in Quantum computing. You are now fully certified to study the statistical aspects of the algorithms.
14.10
Application to Deutsch-Jozsa
In a recent lecture, I gave you an informal argument that the Deutsch-Jozsa problem
had a constant time solution using a classical algorithm if we accept a small but
constant error. We now have the machinery to give a rigorous derivation and also
show two different sampling strategies. It is a bit heavy handed to apply such rigorous
machinery to what seemed to be a relatively obvious result, but the practice that we
get provides a good warm-up for times when the answers are not so obvious. Such
times await in our upcoming lessons on Simons and Shors algorithms.
First, a summary of the classical Deutsch-Jozsa algorithm:
The M -and-Guess Algorithm. Let M be some positive integer. Given a
Boolean function, f (x) of x [ 0, n 1 ] which is either balanced or constant, we
evaluate f (x) M times, each time at a random x [ 0, n 1 ]. If we get two different
outputs, f (x0 ) 6= f (x00 ), we declare f balanced. If we get the same output for all M
trials, we declare f constant.
The only way we fail is if the function is balanced (event B) yet all M trials
produce the same output (event S ). This is the event
S B.
whose probability we must compute.
Assumption. Since we only care about the length of the algorithm as the number
of possible encoded inputs, 2n , gets very large, there is no loss of generality if we
assume M < 2n . To be sure, we can adjust the algorithm so that in those cases in
which M 2n we sample f non-randomly at all 2n input values. In those cases we
will know f completely after the M trials and will have zero error. It is only when
M < 2n that we have to do the analysis.
431
14.10.1
Sampling with Replacement
The way the algorithm is stated, we are admitting the possibility of randomly selecting
the same x more than once during our M trials. This has two consequences in the
balanced case the only case that could lead to error:
1. If we have a balanced function, it results in a very simple and consistent probability for obtaining a 1 (or 0) in any single trial, namely P = 12 .
2. It is a worst case scenario. It produces a larger error than we would get if we
were to recast the algorithm to be smarter. So, if we can guarantee a small
constant error for all n in this case it will be even better if we improve the
algorithm slightly.
The algorithm as stated uses a sampling with replacement technique and is the
version that I summarized in the original presentation. Well dispatch that rigorously
first and move on to a smarter algorithm in the section that follows.
14.10.2
Analysis Given a Balanced f
The experiment consists of sampling f M times then announcing our conclusion:
balanced or constant. We wish to compute the probability of error, which we take
to mean the unlucky combination of being handed a balanced function yet observing
either all 0s or all 1s in our M samples. Before attacking the desired probability of
error,
P (S B) ,
we consider the case in which we are given a balanced function the only way we
could fail. The probability to be computed in that case is expressed by
P S B .
So we first do our counting under the assumption that we have a balanced function.
In this context, we are not asking about a wrong guess, since under this assumption
there is no guessing going on; we know we have a balanced function. Rather, we are
computing the probability that, given a balanced function, we get M identical results
in our M trials.
Another way to approach this preliminary computation is to declare our sample
space, , to be all experimental outcomes based on a balanced function. Shining the
light on this interpretation, we get to leave out the phrase given that f is balanced
throughout this section. Our choice sample space implies it.
432
Notation for Outcomes Yielding a One
Let Ok be the event that we observe f (x) = 1 on the kth trial . (Script O stands
for One.) Because f is balanced (in this section),
1
.
2
Let O (no index)
be
the
unlucky
event
that
we get f (x) = 1 for all M trials,
k = 0, . . . , M 1 . This corresponds to the algebraic identity
P (Ok )
M
1
\
Ok .
k=0
Notation for Outcomes Yielding a Zero
Likewise, let Zk be the event that we observe f (x) = 0 on the kth trial . (Script
Z for Z ero.) Of course, we also have
1
.
2
Finally, let Z (no index) be the event that we get f (x) = 0 for all M trials .
P (Zk )
The Probability of M Identical Readings
The event S means all outcomes were the S ame, i.e, we got either all 1s or all 0s.
S is the disjoint union,
S
O Z
(disjoint),
so
P (S )
P (O Z )
2 P (O)
P (O) + P (Z )
!
M
1
M
1
\
Y
2P
Ok
= 2
P (Ok ) .
k=0
k=0
The first line uses the mutual exclusivity of O and Z , while the last line relies on
the statistical independence of the {Ok }. Plugging in 21 for the terms in the product,
we find that
M
M
1
Y
1
1
1
P (S ) = 2
= 2
=
.
M 1
2
2
2
k=0
433
14.10.3
Completion of Sampling with Replacement for Unknown f
M 1
We might be tempted to say that 21
is the probability of failure in our M -andguess algorithm, but that would be rash. We computed it under the assumption of
a balanced f . To see the error clearly, imagine that our function provider gives us a
balanced function only 1% of the time. For the other 99% of the functions, when we
guess constant we will be doing so on a constant function and will be correct; our
chances of being wrong in this case are diminished to
.01
1
2M 1
To see how this comes out of our theory, we give the correct expression for a wrong
guess. Let W be the event consisting of a wrong guess. It happens when both f is
balanced (B) AND we get M identical results (S ),
W
S B.
Bayes law gives us the proper path,
P (W ) = P (S B) = P S B P (B) .
It was actually, P S B , not P (W ), that we computed in the last section, so we
plug in that number,
1
P (B) ,
P (W ) =
2M 1
and now understand that the probability of being wrong is attenuated by P (B),
which is usually taken to be 1/2 due to fair sampling. But understand that the world
is not always fair, and we may be given more of one kind of function than the other.
Using the usual value for P (B), this yields the proper probability of guessing wrong,
1
1
1
=
.
P (W ) =
2M 1
2
2M
There are many different versions of how one samples the function in a classical
Deutsche-Jozsa solution, so other mechanisms might yield a different estimate. However, they are all used as upper bounds and all give the same end result: classical
methods can get the right answer with exponentially small error in constant time.
14.10.4
Sampling without Replacement
Now, lets make the obvious improvement of avoiding duplicate evaluations of f for
the randomly selected x. Once we generate a random x, we somehow take it out
of contention for the next random selection. (I wont go into the details on how we
434
guarantee this, but you can surmise that it will only add a constant time penalty
something independent of n to the algorithm.)
[Exercise. Write an algorithm that produces M distinct x values at random.
You can assume a random number generator that returns many more than M distinct
values (with possibly some repeats), since a typical random number generator will
return at least 32,768 different ints, while M is on the order of 20 or 50. Hint: Its
okay if your algorithm depends on M , since M is not a function of n. It could even
be on the order of M 2 or M 3 , so long as it does not rely on n.]
Intuitively, this should reduce the error because every time we remove another x
whose f (x) = 1 from our pot, there are fewer xs capable of leading to that 1, while
n
the full complement of 22 xs that give f (x) = 0 are still present. Thus, chances of
getting f (x) = 1 on future draws should diminish each trial from the orginal value of
1
. We prove this by doing a careful count.
2
Analysis Given a Balanced f
As before, we will do the heavy lifting assuming the sample space = B, and when
were done we can just multiply by 1/2 (which assumes equally likely reception of
balanced vs. constant function.).
M = 2 Samples
The {Ok } are no longer independent. To see how to deal with that, well look at
M = 2 which only contains the two events O0 and O1 .
The first trial is the easiest. Clearly,
P (O0 )
1
.
2
Incorporating second trial requires conditional probability. We compute with care.
Our interest is P (O), and
P (O) = P (O1 O0 ) = P O1 O0 P (O0 ) .
What is P O1 O0 ? We get it by counting. There is one fewer x capable of producn
ing f (x) = 1 (specifically, 22 1), and the total number of xs from which to choose
is also smaller by one, now at 2n 1, so
P O1 O0
2n
2
2n
1
1
2n1 1
,
2n 1
making
P (O)
P O1 O0 P (O0 )
435
=
2n1 1
2n 1
1
.
2
Lets write
1
2
in a more complicated but also more suggestive way, to give
n1
n1
2
2
1
.
P (O) = P (O1 O0 ) =
2n 1
2n
M = 3 Samples
You may be able to guess what will happen next. Here we go.
P (O) = P (O2 O1 O0 ) = P O2 [O1 O0 ] P (O1 O0 ) .
We know the value P (O1 O0 ), having
computed
it in the M = 2 case, so direct
your attention to the factor P O2 [O1 O0 ] . The event O1 O0 means there are
now two fewer xs left that would cause f (x) = 1, and the total number of xs from
which to choose now stands at 2n 2, so
P O2 [O1 O0 ]
2n
2
2n
2
2
2n1 2
.
2n 2
Substituting into our expression for P (O) we get
n1
n1
n1
2
2
2
1
2
P (O) = P (O2 O1 O0 ) =
.
n
n
2 2
2 1
2n
The General Case of M Samples
It should now be clear that after M trials,
P (O)
=
=
P (OM 1 O2 O1 O0 )
n1
n1
n1
n1
2
(M 1)
2
2
2
1
2
,
n
n
n
2 (M 1)
2 2
2 1
2n
or, more compactly,
P (O)
M
1
Y
k=0
2n1 k
.
2n k
This covers the eventually of getting all 1s when f is balanced, and if we include the
alternate way to get unlucky, all 0s, we have
P ( S without replacement )
M
1
Y
k=0
436
2n1 k
2n k
Completion of Analysis for Unknown f
We already know what to do. If we believe well be getting about equal numbers of
balanced and constant functions, we multiply the last result by 1/2,
M
1
Y
1
2n1 k
P ( W without replacement ) =
2
2
2n k
k=0
=
M
1
Y
k=0
2n1 k
.
2n k
Bounding the Wrong-Guess Probability
The probability of guessing wrong in the with replacement algorithm (under equal
likelihood of getting the two types of functions) was
P ( W with replacement )
M
1
Y
k=0
1
,
2
so we want to confirm our suspicion that we have improved our chances of guessing
correctly. We compare the current case with this past case and ask
M
1
1
< MY
n1
Y
2
k
1
=
?
n
2 k
2
k=0
k=0
>
Intuitively we already guessed that it must be less, but we can now confirm this with
hard figures. We simply prove that each term in the left product is less than (or in
one case, equal to) each term in the right product.
We want to show that
1
2n1 k
.
n
2 k
2
Now is the time we realize that the old grade school fraction test they taught us from
our childhood, usually called cross-multiplication,
a
c
ad cb .
b
d
actually has some use. We apply it by asking
2 2n1 k
2n k ?
2n 2k 2n k ?
2k k ?
The answer is yes. In fact, except for k = 0 where both sides are equal, the LHS is
strictly less than the RHS. Thus, the product and therefore the probability of error is
actually smaller in the without replacement algorithm. The constant time result of
the earlier algorithm therefore guaranteed a constant time result here, but we should
have fewer wrong guesses now.
437
14.11
A Condition for Constant Time Complexity
in Non-Deterministic Algorithms
This section will turn out to be critically important to our analysis of some quantum
algorithms, ahead. Im going to define a few terms here before our official coverage,
but theyre easy and well give these new terms full air time in future lectures.
14.11.1
Non-Deterministic Algorithms
At the start of this lesson, I alluded to a type of algorithm, A, that was nondeterministic (or as I sometimes say to avoid the hyphen, probabilistic). This means
that given any acceptably small error tolerance, > 0, the probability that A will
give an inaccurate result is < .
Well say that A is probabilistic with error tolerance .
There is a theorem that we can easily state and prove which will help us in our
future quantum algorithms. It gives a simple condition that a probabilistic algorithm
will succeed in constant time. First, a few more definitions.
14.11.2
Preview of Time Complexity An Algorithms Dependence on Size
Assume an algorithm, A, is potentially dependent on an integer, N , that describes
the size of the problem to be solved.
Example 1
A might be an algorithm to sort the data and N is the number of data records to be
sorted.
Example 2
A might be an algorithm to determine whether a function, f , is balanced and N is
the number of inputs to the function.
Well say that the algorithm A has size N .
This does not mean that it will take N units of time for A to execute, or even
that its execution time is dependent in any way on N ; it only means that A solves
a problem associated with a certain number, N , of data elements or inputs, and N
is allowed to take on an increasingly large value, making the algorithm potentially
more difficult (or not). A might complete very quickly, independent of its size, or it
might take longer and longer to complete, as its size, N , grows.
438
14.11.3
Looping Algorithms
Often in quantum computing, we have an algorithm that repeats an identical measurement (test, experiment) in a loop, and that measurement can be categorized in
one of two ways: success (S ) or failure (F ). Assume that a measurement (test, experiment) only need succeed one time in any of the loop passes to end the algorithm
with a declaration of victory: total success. Only if it fails after all loop passes is the
algorithm considered to have failed. Finally, the events S and F for any one loop
pass are usually statistically independent of the outcomes on previous loop passes, a
condition we will assume is met.
Well say that A is a looping algorithm.
14.11.4
Probabilistic Algorithms with Constant Time Complexity
Say we have a looping algorithm A of size N that is probabilistic and can be shown
to complete with the desired confidence (error tolerance) in a fixed number of loop
passes, T , where T is independent of N .
Well call A a constant time algorithm.
14.11.5
A Constant Time Condition for Looping Algorithms
We can now state the theorem, which Ill call the CTC theorem for looping algorithms
(CTC for constant time complexity). By the way, this is my terminology. Dont
try to use it in mixed company.
The CTC Theorem for Looping Algorithms. Assume that A is a
probabilistic, looping algorithm having size N . If we can show that the
probability of success for a single loop pass is bounded away from zero,
i.e.,
P (S )
>
0,
with p independent of the size N , then A is a constant time algorithm.
Proof. Let Sk be the event of success in the kth loop pass. The hypothesis is
that
P (Sk )
p , for all k 1.
We are allowing the algorithm to have an error with probability . Pick T such that
(1 p)T
439
<
a condition we can guarantee for large enough T since (1 p) < 1. Note that p being
independent of the size N implies that T is also independent of N . After having
established T , we repeat As loop T times. The event of failure of our algorithm at
the completion of all T loop passes, which well call Ftot , can be expressed in terms
of the individual loop pass failures,
Ftot
(S1 ) (S2 ) (ST ) .
Since the events are statistically independent, we can convert this to a probability
using a simple product,
P (Ftot )
P (S1 ) P (S2 ) P (ST )
(1 p)T
<
We have shown that we can get A to succeed with failure probability < if we allow
As loop to proceed a fixed number of times, T , independent of its size, N . This is
the definition of constant time complexity.
QED
Explicit Formula for T
To solve for T explicitly, we turn our condition on the integer T into an equality on
a real number t,
(1 p)t
solve for t by taking the log1p of both sides,
t
log1p () ,
then pick any integer T > t. Of course, taking a log having a non-standard base like
1 p, which is some real number between 0 and 1, is not usually a calculator-friendly
proposition; calculators, not to mention programming language math APIs, tend to
give us the option of only log2 or log10 . No problem, because ...
[Exercise. Show that
logA x
log x
,
log A
where logs on the RHS are both base 2, both base 10, or both any other base for that
matter.]
Using the exercise to make the condition on t a little more palatable,
t
log ()
,
log (1 p)
and combining that with the need for an integer T > t, we offer a single formula for
T,
log ()
T =
+ 1,
log (1 p)
where bxc is notation for the floor of x, or the greatest integer x.
440
Examples with Two Different ps for Success
Its important to realize that we dont care whether the event of success in each loop
pass, S , is highly probable or highly improbable. We only care that it is bounded
away from 0 by a fixed amount, p, independent of the size of the algorithm, N . Two
examples should crystallize this. In both examples, we assume that we would like to
assure an error probability less than = .000001 = 106 thats one in a million.
How many times do we have to loop?
Example 1. P (S ) = .002 (Very Improbable). Well use log10 , since my
calculator likes that.
log (106 )
6
T =
+ 1 =
+ 1
log (.998)
.00086945871262889
=
b6900.8452188c + 1
6900 + 1
6901 ,
or, if we want to pick a nice round number, 7000 loop passes.
Example 2. P (S ) = .266 (Reasonably Probable). If we have a more
reasonable chance of success for each measurement in a single pass a little better
than 25% we would expect to get the same level of confidence, error = 106 , with
far fewer loop passes. Lets see how much faster the algorithm converges.
6
log (106 )
+ 1 =
+ 1
T =
log (.734)
0.134303940083929467
=
b44.67478762c + 1
44 + 1
45 ,
or, rounding up, 50 loop passes.
Although our required number of loop passes is understandably sensitive to the
desired accuracy of our algorithm, it is not dependent on the amount of data or
number of inputs on which our algorithm works. The algorithm terminates in constant
time regardless of the amount of data within each of two cases above.
441
Chapter 15
Computational Complexity
15.1
Computational Complexity in Quantum Computing
Algorithms are often described by their computational complexity. This is a quantitative expression of how the algorithms time and space requirements grow as the data
set or problem size grows. Weve already seen examples where the separation between
the quantum and classical algorithms favors quantum, but the arguments given at
the time were a little hand-wavy; we didnt have formal definitions to lean on. We
filled in our loosely worded probability explanations by supplying a lesson on basic
probability theory, and now its time to do the same with computational complexity.
15.2
An Algorithms Sensitivity to Size
We design an algorithm to be applied to an unknown function, a number or a set of
data. It might determine the period of the function, factor the number or sort the
data. The number of inputs to the function (or size of the number to be factored
or amount of data to process) could be anything from one to ten trillion or beyond.
There are many possible solutions to our problem, but once we settle on one that
seems good, we ask the question, how does our algorithmss running time increase
as N increases, where N represents the number of inputs (or number to be factored
or amount of data) on which the algorithm operates?
15.2.1
Some Examples of Time Complexity
The way an algorithms running time grows as the size of the problem increases
its growth rate is known in scientific disciplines as its time complexity. To get
a feel for different growth rates, lets assume that it acts on data sets (not function
inputs or a number to be factored ). The following do not define the various time
442
complexities we are about to study, but they do give you taste of their consequences.
Constant Time Complexity
The algorithm does not depend on the size of the data set, N . It appears to
terminate in a fixed running time (C seconds) no matter how large N is. Such
an algorithm is said to have constant time complexity (or be a constant-time
algorithm).
Polynomial Time Complexity
The algorithm takes C seconds to process N data items. We double N and the
running time seems to double it takes 2C seconds to process 2N items. If we
apply it to 8N data items, the running time seems to take 8C seconds. Here,
the algorithm exhibits linear time complexity (or be a linear algorithm).
The algorithm takes C seconds to process N data items. We double N and
now the running time seems to quadruple it takes C (22 ) = 4C seconds to
process 2N items. If we apply it to 8N data items the running time seems to
take C (82 ) = 64C seconds. Now, the algorithm will likely have quadratic time
complexity (or be a quadratic algorithm).
The algorithm takes C seconds to process N data items. We double N and
now the running time seems to increase by a factor of 23 it takes C (23 ) = 8C
seconds to process 2N items. If we apply it to 8N data items the running time
seems to take C (83 ) = 512C seconds. Now, the algorithm will likely have cubic
time complexity.
The previous examples constant, linear, quadratic all fall into the general
category of polynomial time complexity which includes growth rates limited by some
fixed power of N (N 2 for quadratic, N 3 for cubic, N 5 for quintic, etc.).
Non-Polynomial Time Complexity
Sometimes we cant find a p such that N p reflects the growth in time as the data grows.
We need a different functional form. Examples include logarithmic and exponential
growth. I wont give an example of the former here well define it rigorously in the
next section. But heres what exponential growth feels like.
The algorithm processes N items in C seconds (assume N is large enough that
C > 1). When we double N the running time seems to takes C 2 seconds.
If we apply it to 3N data items the running time takes C 3 seconds and to
7N data items the running time takes C 7 . This is longer/slower/worse than
polynomial complexity for any polynomial size. Now, the algorithm probably
has exponential growth).
443
(This last example doesnt describe every exponential growth algorithm, by the
way, but an algorithm satisfying this for some C > 1 would likely be exponential.)
15.2.2
Time Complexity vs. Space Complexity
Ive limited our discussion to the realm of time. However, if we were allowed to make
hardware circuity that grows as the data set grows (or utilize a larger number of
processors from a very large pool of existing farm of computers) then we may not
need more time. We would be trading time complexity for space complexity.
The general term that describes the growth rate of the algorithm in both time
and space is computational complexity.
For the purposes of our course, well usually take the hardware out of the picture
when measuring complexity and only consider time. There are two reasons for this:
Our interest will usually be in relative speed-up of quantum over classical methods. For that, we will be using a hardware black box that does the bulk of
the computation. We will be asking the question, how much time the quantum algorithm saves us over the classical algorithm when we use the same or
spatially equivalent black boxes in both regimes?
Even when we take space into consideration, the circuitry in our algorithms for
this course will grow linearly at worst, (often logarithmically) and we have much
bigger fish to fry. Were trying to take a very expensive exponential algorithm
classically and find a polynomial algorithm using quantum computation. Therefore, the linear or logarithmic growth of the hardware will be overpowered by
the time cost in both cases and therefore ignorable.
For example the circuit we used for both Deutsch-Josza and Bernstein-Vazirani,
|0in
H n
H n
Uf
|1i
(ignore)
had n + 1 inputs (and outputs), so it grows linearly with n (and only logarithmically
with the encoded N = 2n ). Furthermore, since this is true for both classical and
quantum algorithms, such growth can be ignored when we compare the two regimes.
15.2.3
Notation
To kick things off, we establish the following symbolism for the time taken by an
algorithm to deal with a problem of size N (where you can continue to think of N as
444
the amount of data, while keeping in mind it might be the number of inputs or the
size of a number to be factored).
TQ (N )
time required by algorithm Q to process N elements.
We now explore ways to quantify TQ (N ).
15.3
Big-O Growth
15.3.1
Conflicting Ways to Measure an Algorithm
When you begin to parse the meaning of an algorithms running time, you quickly
come to a realization. Take a simple linear search algorithm on some random (unsorted) data array, myArray[k]. We will plow through the array from element 0
through element N 1 and stop if and when we find the search key, x:
f o r ( k = 0 , found = f a l s e ;
k < N && ! found ;
k++
)
{
i f ( x == myArray [ k ] )
{
found = t r u e ;
foundPosition = k ;
}
}
i f ( found )
c o ut << x << found a t p o s i t i o n << k << e n d l ;
If x is in location myArray[0], the algorithm terminates instantly independent of the
array size: constant time. If it is in the last location, myArray[N-1] (or not in the
list at all), the algorithm will take N 1 steps to complete, a time that increases
linearly with N .
If we cant even adjudicate the speed of an algorithm for a single data set, how
do we categorize it in terms of all data sets of a fixed size? We do so by asking three
of four types of more nuanced questions. The most important category of question is
what happens in the worst case? This kind of time complexity is called big-O.
To measure big-O time complexity, we stack the cards against ourselves by constructing the worst possible data set that our algorithm could encounter. In the
search example above, that would be the case in which x was in that last position
searched. This clears up the ambiguity about where we might find it and tells us that
the big-O complexity is going to be linear. But wait we havent officially defined
what it means to be linear.
445
15.3.2
Definition of Big-0
Lets say we have an ordinary function of N , call it f (N ). f (N ) could be anything.
One example is
f (N )
N 2 + 3N + 75 .
f (N )
N log N + 2 .
Another might be
We wish to compare the growth rate of TQ (N ) with the function f (N ). We say that
TQ (N ) = O (f (N ))
TQ grows no faster than f (N ) .
We are giving our timing on algorithm Q an upper bound using the function f (N ).
In words, we say that the timing for algorithm Q is big-O of f . But we still have a
loose end; the wording grows no faster than is not very rigorous, so lets clear that
up.
TQ (N ) = O (f (N ))
there exist positive constants n0 and c such that
TQ (N ) c |f (N )| ,
for all N n0 .
This means that while TQ (N ) might start out being much greater than c|f (N )| for
small N , eventually it improves as N increases to the extent that TQ (N ) will become
and stay c |f (N )| for all N once we get past N = n0 .
Note. Since our comparison function, f (x), will always be non-negative, I will
drop the absolute value signs in many of the descriptions going forward.
15.3.3
Common Terminology for Certain Big-O Growth Rates
Now we can officially define the terms like quadratic or exponential growth.
446
Descriptions of Growth Functions
f (N )
Informal term for O(f (N ))
constant
log N
logarithmic
log2 N
log-squared
linear
N log N
(no special term)
N2
quadratic
N3
cubic
N k,
k 0, integer
polynomial
N k log N l ,
k, l 0, integer
15.3.4
(also) polynomial
2N
exponential
N!
factorial
Factors and Terms We Can Ignore
Youll note that the table of common big-O terminology doesnt contain functions
like 10N 2 or N 2 + N . Theres a good reason for this.
Ignore a Constant Factor K
Instead of declaring an algorithm to be O(1000N ), we will say it is O(N ) (i.e., linear ).
Instead of O (1.5N 3 ) we will call it O (N 3 ) (i.e., cubic). Heres why.
Theorem. If TQ (N ) = O Kf (N ) , for some constant K, then TQ (N ) =
O(f (N )) .
The theorem says we can ignore constant factors and use the simplest version of the
function possible, i.e., O(N 2 ) vs. O(3.5N 2 ).
[Exercise. Prove it. Hint. Its easy.]
447
For Polynomial big-O complexity, Ignore all but the Highest Power Term
We only need monomials, never binomials or beyond, when declaring a big-O. For
example if TQ is O(N 4 + N 2 + N + 1) its more concisely O(N 4 ). Heres why.
Lemma. If j and k are are non-negative integers satisfying j < k, then
N j N k , for all N 1.
[Exercise. Prove it. Hint. Its easy.]
Now we prove the main claim of the section: for big O, we can ignore all but
the highest power term in a polynomial.
Theorem. If
TQ
k
X
!
aj N j
j=0
then
TQ
O Nk .
Proof. We have positive constants n0 and c such that
k
X
TQ (N ) c
aj N j , for all N n0 .
j=0
Pick a positive number a greater than all the coefficients aj , which will allow us to
write
k
k
X
X
j
c
aj N ca
Nj .
j=0
j=0
Next, if the n0 happened to be = 0 in the big-O criteria, upgrade it to 1. Now n0
is 1 so we can apply the lemma and replace all the N j with the higher power, N k ,
ca
k
X
j=0
ca
k
X
Nk
ca(k + 1) N k ,
for all N n0 ,
j=0
and we have a new constant, c 0 ca(k + 1), with
TQ (N ) c 0 N k , for all N n0 ,
making TQ big-O of N k .
QED
448
15.4
Growth
15.4.1
Definition of
Sometimes we want the opposite relationship between algorithm Q, and a function f .
We want to demonstrate that Q has worse (or at least no better) performance than
f when the data set on which it operates grows. We say that
TQ (N ) = (f (N ))
TQ grows at least as fast as f (N ) .
We are giving our timing on algorithm Q a lower bound using the function f (N ). In
words, we say that the timing for algorithm Q is omega of f . Quantitatively,
TQ (N ) = (f (N ))
there exist positive constants n0 and c such that
TQ (N ) c |f (N )| ,
for all N n0 .
This means that while TQ (N ) might start out being much smaller than c|f (N )| for
small N , eventually it will degrade as N increases to the extent that TQ (N ) will
become and stay c |f (N )| for all N once we get to N = n0 .
There are similar theorems as those we proved for big-O complexity that would
apply to time complexity, but they are not critical for our needs, so Ill leave it as
an exercise.
[Exercise. State and prove lemmas and theorems analogous to the ones we proved
for big-O, but applicable to growth.]
15.5
Growth
Next, we express the fact that an algorithm Q is said to grow at exactly (a term
which is not universally used because it could be misinterpreted) the same rate as
some mathematical expression f (N ) using the notation
TQ (N ) = (f (N )) .
We mean that TQ grows neither faster nor slower than f (N ). In words, we say the
timing for algorithm Q is theta f (N ). Officially,
TQ (N ) = (f (N ))
both
TQ (N ) = O (f (N )) and TQ (N ) = (f (N ))
449
Ideally, this is what we want to know about an algorithm. Sometimes, when programmers informally say an algorithm is big-O of N or N log N , they really mean
that it is theta of N or log N , because they have actually narrowed down the growth
rate to being precisely linear or logarithmic. Conversely, if a programmer says an algorithm is linear or logarithmic or N log N , we dont know what they mean without
qualification by one of categories, usually big-O or .
15.6
Little-o Growth
Less frequently, you may come across little-Oh notation, that is, TQ = o(f (N )). This
simply means that we not only have an upper bound in f (N ), but this upper bound
is, in some sense, too high. One way to say this is that
TQ (N ) = o (f (N ))
TQ (N ) = O (f (N )) , but TQ (N ) 6= (f (N )) , .
15.7
Easy vs. Hard
Computational theorists use the term easy to refer to a problem whose (known)
algorithm has polynomial time complexity. We dont necessarily bother specifying
the exact power of the bounding monomial; usually the powers are on the order of
five or less. However in this course, well prove exactly what they are when we need
to.
Hard problems are ones whose only known algorithms have exponential time complexity.
A large part of the promise of quantum computing is that it can use quantum
parallelism and entanglement to take problems that are classically hard and find
quantum algorithms that are easy. This is called exponential speed-up (sometimes
qualified with the terms relative or absolute).
For the remainder of the course we will only tackle two remaining algorithms, but
they will be whoppers. They both exhibit exponential speed up.
15.8
Wrap-Up
This section was a necessarily brief and incomplete study of time complexity because
we only needed the most fundamental aspects of big-O and growth for the most
obvious and easy-to-state classes: exponential growth, polynomial growth (and a few
cases of N log N growth). When we need them, the analysis we do should be selfexplanatory, especially with this short section of definitions on which you can fall
back.
450
Chapter 16
Computational Basis States and
Modular Arithmetic
|15 3i4
16.1
|12i4
|1100i
(1, 1, 0, 0)t
Different Notations Used in Quantum Computing
In the quantum computing literature youll encounter alternative ways to express the
states of a qubit or the inner workings of an algorithm. If youre schooled only in
one style, you might be puzzled when you come across unfamiliar verbiage, especially
when the author changes dialects in mid-utterance. Today, I want to consolidate a few
different ways of talking about Hilbert space and computational basis states. This will
prepare you for the algorithms ahead and enable you to read more advanced papers
which assume the reader can make such transitions seamlessly. Its also convenient to
have this single resource to consult if youre reading a derivation and suddenly have
a queasy feeling as the notation starts to get away from you.
16.2
Notation and Equivalence of Three Environments
16.2.1
First Environment n-qubit Hilbert Space, H(n)
Single Qubit Hilbert Spaces
Recall that a one-qubit Hilbert space, H, consists of the 2-D complex vector space with
basis vectors |0i and |1i, making its typical state a superposition (always normalized,
451
of course) of those two vectors, as in
|i = |0i + |1i .
Multi Qubit Hilbert Spaces
Multi-qubit computers operate in a tensor product of such spaces, one for each qubit.
This product is a 2n -dimensional Hilbert space, which I sometimes label with the
subscript (n) for clarity, as in H(n) . We can think of it as
n
H(n)
z
}|
{
H H ... H
n1
O
H.
k=0
The computational basis states, or CBS, of this product space are the 2n vectors of
the form
n
|xi
|xn1 i |xn2 i |xn3 i |x0 i
n1
O
|xk i ,
k=0
where each |xk i is either |0i or |1i, i.e., a CBS of the kth one qubit space. We index in
decreasing order from xn1 to x0 because well want the right-most bit to correspond
to the least significant bit of the binary number xn1 . . . x1 x0 .
Different CBS Notations
One shorthand we have used in the past for this CBS is
|xn1 i |xn2 i |x1 i |x0 i ,
and two other common notations well need are the decimal integer (encoded ) form,
x, and its binary representation,
|xin ,
x {0, 1, 2, 3, . . . , 2n 1} or
|xn1 xn2 . . . x3 x2 x1 x0 i ,
xk {0, 1} .
For example, for n = 3,
|0i3
|000i ,
|1i3
|001i ,
|2i
|010i ,
|3i
|011i ,
|4i
and, in general,
|100i ,
|xi3
452
|x2 x1 x0 i .
The 2-dimensional H = H(1) and its 2n -dimensional products H(n) are models we
use for quantum computing. However, the problems that arise naturally in math and
computer science are based on simpler number systems. Lets have a look at two such
systems and show that they are equivalent to the CBS of our Hilbert space(s), H.
16.2.2
Z2 )n
The Second Environment: The Finite Group (Z
Simple Integers
Were all familiar with the integers, Z,
Z
{ . . . , 3, 2, 1, 0, 1, 2, 3, . . . }, usual +,
where I have explicitly stated the operation of interest, namely ordinary addition.
(Were not particularly interested in multiplication at this time.) This is sometimes
called the group of integers, and as a set we know its infinite, stretching toward
in the two directions.
Stepping Stone: Z N , or mod N Arithmetic
Another group that you may not have encountered in a prior course is the finite group
ZN , consisting of only the N integers from 0 to N 1,
ZN
{ 0, 1 , 2, . . . N 1},
+ is (+
mod N ),
and this time we are using addition modulo N as the principal operation; if x+y N ,
we bring it back into the set by taking its remainder after dividing by N :
x + y
(mod N )
(x + y) % N.
The negative of each x ZN is defined by
x
(mod N )
(N x).
Subtraction is defined using the above two definitions, as you would expect,
x y
(mod N )
453
(x + y) % N.
To make this concrete, we consider the group Z15 .
Example Operations mod-15
7 + 2 = 9
7 + 10 = 2
14 + 14 = 13
1 = 14
7 = 8
10 = 5
14 2 = 12
2 8 = 9
4 4 = 0
Stepping Stone: Z 2 with Arithmetic
While ZN is important in its own right (and well be using it in the future), an
important special case youll want to embrace is ZN for N = 2. The above definition
of ZN works for Z2 , but when N = 2 some special notation kicks in:
Z2
{ 0, 1 },
is (+
mod 2) .
A few consequences of mod-2 addition are
0
1
0
1
1
0
0
1
1
0 1
= 1
= 1
= 0
= 0
= 1
= 0 1
Of course is nothing other than the familiar XOR operation, although in this
context, we get subtraction and negative mod-2 numbers defined, as well. Also, while
we should be consistent and call subtraction , the last example shows that we can,
and usually do, use the ordinary operator, even in mod-2 arithmetic. If there is
the potential for confusion, we would tag on the parenthetical (mod 2).
An Old Friend with a New Name
We studied Z2 under a different name during our introductory lecture on a single
qubit (although you may have skipped that optional section; it was a study of classical
bits and classical gates using the formal language of vector spaces in preparation for
defining the qubit). At the time we didnt use the symbolism Z2 for the two-element
group, but called it B, the two-element field on which we built the vector space B
for the formal definition of classical bits and operators.
454
Connection Between Z 2 and H
While not officially recognized by any government or municipality, it will help to
observe a very simple relationship between the mod-2 group, Z2 , and the single-qubit
space, H = H(1) . Its little more than the fact that the dimension of H(1) , 2, equals
the size of the group Z2 , also 2. Moreover, each CBS of H(1) is labeled using the
digits 0 and 1:
Z2
0
1
H(1)
|0i
|1i
I hasten to add that this connection does not go beyond the 1:1 correspondence listed
above and, in particular, does not extend to the mod-2 addition in Z2 vs. the vector
addition in H; those are totally separate and possess no similarities. Also, only the
basis states in H are part of this correspondence; the general state, |i = |0i + |1i
has no place in the analogy. As tenuous as it may seem, this connection will help us
in the up-coming analysis.
Z2 )n with Arithmetic
Second Environment Completed: (Z
As a set, (Z2 )n is simply the n-tuples that have either 0 or 1 as their coordinates,
that is,
(Z2 )n
(xn1 , xn2 , . . . x2 , x1 , x0 ) ,
each xk = 0 or 1,
or, in column vector notation,
(Z2 )n
0
0
0
0 0 0
.. .. ..
. , . , . ,
0 0 1
0
1
0
xn1
x
n2
..
x2
x1
..
.
Notice that we label the 0th coordinate on the far right, or bottom, and the (n 1)st
coordinate on the far left, or top. This facilitates the association of these vectors with
binary number representations (coming soon) in which the LSB is on the right, and
the MSB is on the left (as in binary 1000 = 8, while 0001 = 1).
455
The additive operation stems from the Z2 in its name: its the component-wise
mod-2 addition, or, equivalently XOR, e.g.,
0
1
1
1
0
1
=
0
1
1
1
1
0
Ill usually write the elements of (Z2 )n in boldface, as in x or y, to emphasize the
vector point-of-view.
Common Notation. This set is often written { 0, 1 }n , especially when we dont
care about the addition operation, only n-tuples of 0s and 1s that are used as inputs
to Boolean functions.
Z2 )n is a Vector Space
(Z
Ive already started calling the objects in (Z2 )n vectors, and this truly is an official
designation. Just as Rn is an n-dimensional vector space over the reals, and H(n) is a
2n -dimensional vector space over the complex numbers, (Z2 )n is a vector space over
Z2 . The natural question arises, what does this even mean?
You know certain things about all vector spaces, a few of which are
There is some kind of scalar multiplication, cv.
There is always some basis and therefore a dimension.
There is often an inner product defining orthogonality.
All this is true of (Z2 )n , although the details will be defined as they crop up. For now,
we only care about the vector notation and addition. That is, unless you want to
do this ...
[Exercise. Describe all of the above and anything else that needs to be confirmed
to authorize us to call (Z2 )n a vector space.]
Caution. For general N , (ZN )n is not a vector space. The enlightened among
you can help the uninitiated understand this in your forums, but it is not something
we will need. What kind of N will lead to a vector space?
Recall. Once again, think back to the vector space that we called B = B2 . It was
the formal structure that we used to define classical bits. Using the more conventional
language of this lesson, B is the four-element, 2-dimensional vector space (Z2 )2 .
Z2 )n and H (n)
Connection Between (Z
Lets punch up our previous analogy, bringing it into higher dimensions. We relate
the n-component vectors, (Z2 )n , and the multi-qubit space, H = H(n) . As before,
456
the dimension of H(n) , 2n , equals the size of the group (Z2 )n , also 2n . The vectors in
(Z2 )n corresponds nicely to the CBS in H(n) :
(Z2 )n
H(n)
(0, 0, . . . , 0, 0, 0)
|00 000i
(0, 0, . . . , 0, 0, 1)
|00 001i
(0, 0, . . . , 0, 1, 0)t
|00 010i
(0, 0, . . . , 0, 1, 1)
|00 011i
(0, 0, . . . , 1, 0, 0)
..
.
|00 100i
..
.
|xn1 x2 x1 x0 i
(xn1 , . . . , x2 , x1 , x0 )t
Again, there is no connection between the respective addition operations, and the
correspondence does not include superposition states of H(n) . Still, the basis states in
H(n) line up with the vectors in (Z2 )n , and thats the important thing to remember.
16.2.3
The Third Environment: The Finite Group Z 2n with
Arithmetic
One of our stepping stones above included the finite group ZN with mod-N + as its
additive operation,
ZN
{ 0, 1 , 2, . . . N 1}.
Now were going to make two modifications to this. First, well restrict the size, N ,
to powers of 2, i.e., N = 2n , for some n,
Z2n
{ 0, 1 , 2, . . . 2n 1},
so each x in ZN has exactly n binary digits.
x
x
xn1 xn2 x2 x1 x0 , notationally, or
n1
X
xk 2k , as a sum of powers-of-2.
k=0
The second change will be to the addition operation. It will be neither normal addition
nor mod-N addition. Instead, we define x + y using the bit-wise operator.
If x
and y
then x y
=
=
xn1 xn2 x2 x1 x0 ,
yn1 yn2 y2 y1 y0 ,
(xn1 yn1 ) (x2 y2 )(x1 y1 )(x0 y0 ) .
457
Note that the RHS of the last line is not a product, but the binary representation
using its base-2 digits (e.g., 11010001101 ). Another way to say it is
xy
n1
X
(xk yk )2k .
k=0
To eliminate any confusion between this group and the same set under ordinary modN (mod-2n ) addition, lets call ours by its full name,
(Z2n , ) .
Examples of addition in (Z2n , ) are
11
12
15
45
5 11
13 13
15 3
=
=
=
=
=
=
=
0,
3,
4,
1,
14,
0, and
12.
=
=
0, so
x,
Note that for any x (Z2n , ),
xx
x
i.e., x is its own additive inverse, under bit-wise .
Z2 )n and (Z
Z2n , )
Connection Between (Z
(Z2n , ) and (Z2 )n are fundamentally two ways to express the same group they are
isomorphic in group terminology. This is symbolized by
n
(Z2n , )
= (Z2 ) ,
under the set and operator association
xn1
xn2
..
x = (xn1 xn2 x1 x0 ) x = . ,
x1
x0
x y
x y.
[Exercise. For those of you fixating on the vector space aspect of (Z2 )n , you may
as well satisfy your curiosity by writing down why this makes Z2n a vector space over
Z2 , one that is isomorphic to (effectively the same as) (Z2 )n .]
458
16.2.4
Z2 )n and (Z
Z2n , )
Interchangeable Notation of H(n) , (Z
In practical terms, the relationship between the above three environments allows us
to use bit-vectors,
1
1
(1, 1, 0, 1, 0)t =
0 ,
1
0
binary number strings,
11010,
or plain old encoded ints
26
interchangeably, at will. One way well take advantage of this is by using plain int
notation in our kets. For n = 5, for example, we might write any of the four equivalent
expressions,
|26i
|11010i
|1i |1i |0i |1i |0i
|1i |1i |0i |1i |0i ,
usually the first or second. Also, we may add notation to designate the number of
qubits under consideration, as in
|26i5
or, to show one possible breakdown, the equivalent
|3i |2i3 .
Hazard
Why are these last two equivalent? We must be careful not to confuse the encoded
CBS notation as if it gave coordinates of CBS kets in the natural basis it does not.
In other words, |3i2 expressed in natural tensor coordinates is not
1
.
1
It cant be for two reasons:
a) |3i2 is a 4-dimensional vector and requires four, not two, coordinates to express
it, and
459
b) |3i2 is a CBS, and any CBS ket expressed in its own (natural) basis can have
only a single 1 coordinate, the balance of the column consisting of 0 coordinates.
The correct expression of |3i2 in tensor coordinates is
0
0
,
0
1
as can be seen if we construct it from its component tensor coordinates using
0
0
0
0
|3i2 = |1i |1i =
=
0 .
1
1
1
Therefore, to answer this last question why are the last two expressions equivalent?
we must first express all vectors in terms of natural coordinates. That would produce
three column vectors (for |3i2 , |2i3 and |26i5 ) in which all had a single 1 in its
respective column, and only then could we compute the product of two of them,
demonstrating that it was equal to the third.
To head off another possible source of confusion, we must understand why the
tensor product dimension of the two component vectors is not 2 3 = 6 contradicting
a possibly (and incorrectly) hoped-for result of 5. Well, the dimensions of these spaces
are not 2, 3, 5 or even 6. Remember that exponent to the upper-right of the ket
designates the order of the Hilbert space. Meanwhile, the dimension of each space
is (2)order , so these dimensions are actually 22 , 23 and 25 . Now we can see that the
product space dimension, 32, equals the product of the two component dimensions,
4 8, as it should.
Sums inside Kets
Most importantly, if x and y are two elements in Z2n , we may take their mod-2 sum
inside a ket,
|x yi ,
which means that we are first forming x y, as defined in (Z2n , ), and then using
that n-bit answer to signify the CBS associated with it, e.g.,
|1 5i
|5 11i
4
|15 3i
|21 21i
=
=
|4i ,
|14i
|12i
or, designating a qubit size,
4
and
|0i .
460
Example of Different Notations Applied to Hadamard
Recall that the nth order Hadamard operators definition, usually given in encoded
binary form, is
H
|xi
=
n 1
n 2X
(1)x y |yin ,
y=0
where is the mod-2 dot product based on the individual binary digits in the base-2
representation of x and y,
x y
xn1 yn1 xn2 yn2 x1 y1 x0 y0 .
If x and y are represented as vectors in (Z2 )n using the boldface x and y,
xn1
yn1
xn2
yn2
..
..
x x = . ,
y y = . .
x1
y1
x0
y0
the dot product between vector x and vector y is also assumed to be the mod-2 dot
product,
xy
xn1 yn1 xn2 yn2 x1 y1 x0 y0 ,
giving the alternate form of the Hadamard gate,
H
|xi
=
n 1
n 2X
(1)x y |yin .
y=0
This demonstrates the carefree change of notation often uncounted in this and other
quantum computing presentations.
We have completed our review and study of the different notation and language
used for CBS. As we move forward, well want to add more vocabulary that straddles
these three mathematical systems, most notably, periodicity, but thats best deferred
until we get to the algorithms which require it.
461
Chapter 17
Quantum Oracles
|0in
H n
H n
Uf
|0in
|
{z
}
Quantum Oracle
17.1
Higher Dimensional Oracles and their Time
Complexity
Weve seen oracles in our circuits for Deutsch, Deutsh-Jozsa and Bernstein-Vazirani,
but today we will focus on the oracle itself, not a specific algorithm. Our goals will
be to
extend our input size to cover any dimension for each of the two oracles channels,
get a visual classification of the matrix for an oracle, and
define relativized and absolute time complexity, two different ways of measuring
a quantum algorithms improvement over a classical algorithm.
The last item relies on an understanding of the oracles time complexity, which is why
it is included in this chapter.
Well continue to use Uf to represent the oracle for a Boolean function, f . Even
as we widen the input channels today, Uf will still have two general inputs, an upper,
A or data register and a lower, B or target register.
462
At the top of the page Ive included a circuit that solves Simons problem (coming
soon). It reminds us how the oracle relates to the surrounding gates and contains a
wider (n qubit) input to the target than weve seen up to now.
17.2
Simplest Oracle: a Boolean Function of One
Bit
At the heart of our previous quantum circuits has lurked the unitary transformation
weve been calling the quantum oracle or just the oracle. Other gates may be wired
into the circuit around the oracle, but they are usually standard parts that we pull
off the shelf like CNOT, Hadamard and other simple gates. The oracle on the other
hand is custom designed for the problem to be solved. It typically involves some function, f , that the circuit and its algorithm are meant to explore/discover/categorize.
17.2.1
Circuit and Initial Remarks
The simplest quantum oracle is one that arises from a unary Boolean f . We defined
such a Uf in the Deutsch circuit. Its action on a general CBS |xi |yi is shown in the
following circuit:
|xi
|xi
Uf
|yi
|y f (x)i
In terms of the effect that the oracle has on the CBS |xi |yi, which we know to be
shorthand for |xi |yi, the oracle can be described as
|xi |yi
Uf
|xi |y f (x)i .
There are a number of things to establish at the start, some review, others new.
Initial Remarks (Possibly Review)
1. is the mod-2 addition:
00
01
10
11
=
=
=
=
0
1
1
0
2. This gate can be viewed as a reversible operator for a typically irreversible
boolean function, f .
463
3. The separable state |xi |yi coming in from the left represents a very special and
restricted input: a CBS. Whenever any gate is defined in terms of CBS, we
must remember to use linearity and extend the definition to the entire Hilbert
space. We do this by expanding a general ket, such as a 4-dimensional |i2 ,
along the CBS,
|i2
c0 |0i |0i + c1 |0i |1i + c2 |1i |0i + c3 |1i |1i
c0 |0i2 + c1 |1i2 + c2 |2i2 + c3 |3i2 ,
reading off the output for each of the CBS kets from our oracle description, and
combining those using the complex amplitudes, ck .
4. When a CBS is presented to the input of an oracle like Uf , the output happens
to be a separable state (something not true for general unitary gates as we saw
with the BELL operator). In this case, the separable output is |xi |y f (x)i.
Considering the last bullet, we cant expect such a nice separable product when
we present the oracle with some non-basis state, |i2 , at its inputs. Take care
not to make the mistake of using the above template directly on non-CBS inputs.
Initial Remarks (Probably New)
Oracles are often called black boxes, because we computer scientists dont care how
the physicists and engineers build them or whats inside. However, when we specify
an oracle using any definition (above being only one such example), we have to check
that certain criteria are met.
1. The definition of Uf s action on the CBS inputs as described above must result
in unitarity. Should you come across a putative oracle with a slightly off-beat
definition, a quick check of unitarity might be in order a so called sanity
check.
2. The above circuit is for two one-qubit inputs (thats a total of 4-dimensions for
our input and output states) based on a function, f , that has one bit in and
one bit out. After studying this easy case, well have to extend the definition
and confirm unitarity for
multi-qubit input registers taking CBS of the form |xin and |yim , and
an f with domain and range larger than the set {0, 1}.
3. The function that we specify needs to be easy to compute in the complexity
sense. A quantum circuit wont likely help us if a computationally hard function
is inside an oracle. While we may be solving hard problems, we need to find
easy functions on which to build our circuits. This means the functions have to
be computable in polynomial time.
4. Even if the function is easy, the quantum oracle still may be impractical to build
in the near future of quantum computing.
464
17.2.2
A Two-Qubit Oracles Action on the CBS
The nice thing about studying single bit functions, f , and their two-qubit oracles, Uf ,
is that we dont have to work in abstracts. There are so few options, we can compute
each one according to the definitions. The results often reveal patterns that will hold
in the more general cases.
Notation Reminder If a is a single binary digit, a = a is its logical negation
(AKA the bit-flip or not operation),
(
0, if a = 1,
a (or a)
1, if a = 0.
A short scribble should convince you that 0 a = a and 1 a = a. Therefore,
Uf
|xi |0i 7 |xi |0 f (x)i =
Uf
|xi |1i 7 |xi |1 f (x)i =
17.2.3
|xi |f (x)i , and
|xi |f (x)i .
Case Study #1: f (x) 1
CBS Table for the Oracle
If f 1, a constant function, we can list all the possible outputs in a four-line table:
|xi |yi
Uf (|xi |yi)
|0i |0i = |0i2
|0i |1i = |1i2
|0i |1i = |1i2
|0i |0i = |0i2
|1i |0i = |2i2
|1i |1i = |3i2
|1i |1i = |3i2
|1i |0i = |2i2
(Remember, | i2 does not mean ket squared, but is an indicator that this is an
n-fold tensor product state, where n = 2.)
The Matrix for Uf
Its always helpful to write down the matrix for any linear operator. At the very least
it will usually reveal whether or not the operator is unitary even though we suspect
without looking at the matrix that Uf is unitary since it is real and is its own inverse.
However, self-invertibility does not always unitarity make, so its safest to confirm
unitarity by looking at the matrix.
This is a transformation from 4-dimensions to 4-dimensions, so we need to express
the 4-dimensional basis kets as coordinates. Lets review the connection between the
four CBS kets of H(2) and their natural basis coordinates:
465
|0i |0i
|0i2
1
0
0
0
component
encoded
coordinate
|0i |1i
|1i2
0
1
0
0
|1i |0i
|2i2
0
0
1
0
|1i |1i
|3i2
0
0
0
1
To obtain the matrix, express each Uf (|xi |yi) as a column vector for each tensor
CBS, |ki2 , k = 0, . . . , 3:
Uf |0i2 Uf |1i2 Uf |2i2 Uf |3i2
|1i2 |0i2 |3i2 |2i2
=
0
1
0
0
1
0
0
0
=
0
0
0
1
0
0
1
0
Aha. These rows (or columns) are clearly orthonormal, so the matrix is unitary
meaning the operator is unitary.
It will be useful to express this matrix in terms of the 2 2 Pauli matrix, x ,
associated with the X gate.
Uf
x
0
=
0
x
x
x
This form reveals a pattern that you may find useful going forward. Whenever a
square matrix M can be broken down into smaller unitary matrices along its diagonal
(0s assumed elsewhere), M will be unitary.
[Exercise. Prove the last statement.]
17.2.4
Case Study #2: f (x) x
CBS Table for the Oracle
Now, let f x, the identity function. We repeat the process of the first case study
and list all the possible outputs. For reference, here are the key mappings again:
Uf
|xi |0i 7 |xi |f (x)i ,
Uf
|xi |1i 7 |xi |f (x)i .
466
Now, the table for f x:
|xi |yi
Uf (|xi |yi)
|0i |0i = |0i2
|0i |0i = |0i2
|0i |1i = |1i2
|0i |1i = |1i2
|1i |0i = |2i2
|1i |1i = |3i2
|1i |1i = |3i2
|1i |0i = |2i2
The Matrix for Uf
We obtain the matrix by expressing each Uf (|xi |yi) as a column vectors:
Uf |0i2 Uf |1i2 Uf |2i2 Uf |3i2
|0i2 |1i2 |3i2 |2i2
=
1
0
0
0
0
1
0
0
=
0
0
0
1
0
0
1
0
=
1
x
This time, weve enlisted the help of the 2 2 identity matrix, 1. Again, we see how
Uf acts on its inputs while confirming that it is unitary.
17.2.5
Remaining Cases #3 and #4
There are only two more 1-bit functions left to analyze. One of them, f (x) 0 was
covered in our first quantum algorithms lesson under the topic of Oracle for the [0]-op.
We found its matrix to be
Uf
1.
The last, f (x) x, has the matrix
Uf
[Exercise. Derive the last matrix.]
467
17.3
Integers Mod-2 Review
17.3.1
The Classical f at the Heart of an Oracle
We will be building oracles based on classical Boolean functions that might have many
inputs and one output,
f : { 0, 1 }n { 0, 1 } .
or many inputs and many outputs,
f : { 0, 1 }n { 0, 1 }m .
To facilitate this, well need more compact language than { 0, 1 }n , especially as we
will be adding these multi-bit items both inside and outside our kets. We have
such vocabulary in our quiver thanks to the previous lesson on CBS and modular
arithmetic.
17.3.2
Mod-2 Notation for { 0, 1 }n
We learned that { 0, 1 }n can be thought of in one of three equivalent ways.
Encoded Integer Notation: (Z2n , )
This is the group containing the first 2n integers
Z2n
{0, 1, 2, 3, . . . 2n 1 }
with mod-2 arithmetic, . There is always the possibility of confusion when we
look at a set like ZN , for some integer N > 0. Usually, it means the numbers
{ 0, 1, 2, . . . , N 1, } with the mod-N addition operation. However, when N = 2n
is a power-of-2, it can often mean and, for us, does that we are using a bitwise
mod-2 addition, . For this lesson, Ill use Z2n rather than the more cumbersome
(Z2n , ) when I want to signify these encoded integers with mod-2 addition.
n
Binary Vector Notation: (Z2 )
Here we mean the group consisting of the 2n binary vectors,
0
0
0
1
.. .. ..
..
(Z2 )n
. , . , . , , . ,
0 0 1
1
0
1
0
1
with being component-wise mod-2 addition.
468
The Natural Basis for H(n)
Finally, we have the 2n basis kets for H(n) ,
CBS for H(n)
{ |xin ,
x = 0, 1, 2, 3, . . . , 2n 1 } .
Seen below, we can use either encoded or binary form to write these basis kets.
|0in
|1in
|2in
|3in
|4in
..
.
| 2n 1 in
17.3.3
|0 000i ,
|0 001i ,
|0 010i ,
|0 011i ,
|0 100i ,
|1 111i .
Inside the Ket
This review might be more than we need, but it will nip a few potentially confusing
situations in the bud, the biggest being the notation
|y f (x)i
when y and f (x) are more than just labels for the CBS of the 2-dimensional H, 0 and
1. Of course, when they are 0 and 1, we know what to do:
|1 0i
|1 1i
=
=
|1i or
|0i .
But when we are in a higher dimensional Hilbert space that comes about by studying
a function sending Z2n into Z2m , well remember the above correspondence. For
example,
|15 3i4
|21 21i
=
=
|12i4
and
|0i .
With that little review, we can analyze our intermediate and advanced multi-qubit
oracles.
469
17.4
Intermediate Oracle: f is a Boolean Function
of a Multiple Input Bits
17.4.1
Circuit
A Wider Data Register
Next, we study the circuit specification
|xin
|xin
Uf
|yi
|y f (x)i
which arises when the function under study is an f which maps { 0, 1 }n { 0, 1 }.
We came across this oracle in Deutsh-Jozsa and Bernstein-Vazirani. Well review and
dig deeper.
Such an f would require an n-qubit |xin input be sent to the A register, even as we
maintained the single-qubit |yi going into the register B. This should be intuitively
clear because the oracles very definition calls for the output of the B register to be
|y f (x)i, something that can only be accomplished if both
i) x has the right number of bits to satiate f (namely, n), and
ii) the y to be -ed with f (x) is commensurate with the output value of f
(namely, a single binary bit).
The analysis of the simplest oracle taught us that Uf was expressible as a unitary
4 4 matrix with cute, little unitary 2 2 matrices (either 1 or x ) along its diagonal.
It will turn out this intermediate-level oracle is nothing more than many copies of
those same 2 2s affixed to a longer diagonal, one traversing a 2n+1 2n+1 matrix.
17.4.2
An (n + 1)-Qubit Oracles Action on the CBS
The general mapping we found so useful in the simple case continues to work for us
here. Even though x is now in the larger domain, Z2n , f (x) is still restricted to 0
or 1, so we can reuse the familiar identities with a minor change (the first separable
component gets an order n exponent),
Uf
|xin |0i 7 |xin |f (x)i ,
Uf
|xin |1i 7 |xin |f (x)i .
470
17.4.3
Analyzing Uf for x = 0
For the moment, we wont commit to a specific f , but we will restrict our attention
to the input x = 0.
Uf
|0in |0i 7 |0in |f (0)i ,
Uf
|0in |1i 7 |0in |f (0)i .
f (0) can be either 0 or 1, and y can be either 0 or 1, giving four possible combinations:
|0in |yi
f (0)
Uf (|0in |yi)
|0in |0i = |0in+1
|0in |f (0)i = |0in |0i = |0in+1
|0in |1i = |1in+1
|0in |f (0)i = |0in |1i = |1in+1
|0in |0i = |0in+1
|0in |f (0)i = |0in |1i = |1in+1
|0in |1i = |1in+1
|0in |f (0)i = |0in |0i = |0in+1
This is a little different from our previous table. Rather than completely determining
a 4 4 matrix for a particular f , it gives us two possible 2 2 sub-matrices depending
on the value of f (0).
1. When f (0) = 0, the first two columns of the matrix for Uf will be (see upper
two rows of table):
Uf |0in+1 Uf |1in+1
|0in+1 |1in+1
=
1 0
0 1
= 0 0
?
.. ..
. .
0 0
?
1
0
?
2. When f (0) = 1, the first two columns of the matrix for Uf will be (see lower
471
two rows of table):
Uf |0in+1 Uf |1in+1
|1in+1 |0in+1
=
0 1
1 0
= 0 0
?
.. ..
. .
0 0
x
?
0
?
The ? represents the yet-to-be-studied portion of Uf .
So we see that regardless of the value f (0), the upper-left 2 2 sub-matrix will be
one of our two familiar unitary matrices, 1 or x . There was nothing special about
x = 0. We might have analyzed, x = 1, 2, 3, or any x up to 2n 1. In fact, well do
that next.
17.4.4
Analyzing Uf for General x
We continue studying a non-specific f and, like the x = 0 case, work with a fixed x.
However, this time x Z2n can be any element in that domain. If it helps, you can
think of a small x, like x = 1, 2 or 3. The mappings that will drive the math become
Uf
|xin |0i 7 |xin |f (x)i ,
Uf
|xin |1i 7 |xin |f (x)i .
As before, there two possible values for f (x) and two possible values for y, giving four
possible combinations for this x:
|xin |yi
f (x)
Uf (|xin |yi)
|xin |0i
|xin |f (x)i = |xin |0i
|xin |1i
|xin |f (x)i = |xin |1i
|xin |0i
|xin |f (x)i = |xin |1i
|xin |1i
|xin |f (x)i = |xin |0i
472
We computed the first two columns of the matrix for Uf before, and now we
compute two columns of Uf further to the right.
[In Case You Were Wondering. Why did the single value x = 0 produce two
columns in the matrix? Because there were two possible values for y, 0 and 1, which
gave rise to two different basis kets |0in |0i and |0in |1i. It was those two kets that
we subjected to Uf to produce the first two columns of the matrix. Same thing here,
except now the two basis kets that correspond to the fixed x under consideration are
|xin |0i and |xin |1i, and they will produce matrix columns 2x and 2x + 1.]
1. When f (x) = 0, columns 2x and 2x + 1 of the matrix for Uf will be (see upper
two rows of table):
Uf (|xin |0i) Uf (|xin |1i)
|xin |0i |xin |1i
2x
2x+1
z }| {
0 0
.. ..
. .
1 0
0 1
.. ..
. .
0 0
2x
2x
2x+1
2x+1
z }| {
0
1
2x
2x+1
2. When f (0) = 1, the first two columns of the matrix for Uf will be (see lower
473
two rows of table):
Uf (|xin |0i) Uf (|xin |1i)
|xin |1i |xin |0i
2x
2x+1
z }| {
0 0
.. ..
. .
0 1
1 0
.. ..
. .
0 0
2x
2x
2x+1
2x+1
z }| {
2x
2x+1
As you can see, for any x the two columns starting with column 2x contain all 0s
away from the diagonal and are either the 2 2 identity 1 or the Pauli matrix x on
the diagonal, giving the matrix the overall form
[1 or x ]
[1 or x ]
...
0
[1 or x ]
[1 or x ]
The oracle is as we strongly suspected it would be unitary, and we now have a
nice visual image of what it looks like.
474
17.5
Advanced Oracle: f is a Multi-Valued function of a Multiple Input Bits
17.5.1
Circuit and Vocabulary
A Wider Target Register
Finally, well need to understand the full circuit specification
|xin
|xin
Uf
|y f (x)im
|yi
which arises when the function under study is an f that maps Z2n Z2m . In this
case it only makes sense to have an m-qubit B register, as pictured above, otherwise
the sum inside the bottom right output, |y f (x)i would be ill-defined.
Sometimes m = n
Often, for multi-valued f , we can arrange things so that m = n and, in that case the
circuit will be
|xin
|xin
Uf
n
|y f (x)in
|yi
17.5.2
Smallest Advanced Oracle: m = 2 (Range of f Z 22 )
In this case, f (x) { 0, 1, 2, 3 } for each x Z2n . There will be (n + 2) qubits going
into Uf , (n qubits into the A register and 2 qubits into the B register).
|xin
|xin
Uf
2
|y f (x)i2
|yi
This gives a total of 2n+2 possible input CBS going into the system, making the
dimension of the overall tensor product space, H(n+2) , = 2n+2 . The matrix for Uf
will, therefore, have size (2n+2 2n+2 ).
17.5.3
The 4 4 Sub-Matrix for a Fixed x
We follow the pattern set in the intermediate case and look at a non-specific f but
work with a fixed x Z2n . Because we are assuming m = 2, this leaves four possible
images for this one x, namely f (x) = 0, 1, 2 or 3. As well see, this leads to a 4 4
475
sub-matrix compared to the 2 2 sub-matrix we got when studying a fixed x in the
intermediate case.
The maneuver that we have to apply here, which was not needed before, is the
application of the operator between y and f (x) on a bit-by-bit basis. For m = 2
(and using the notation f (x)k to mean the kth digit of the number f (x)), we get
|y f (x)i2
| y1 y0 f (x)1 f (x)0 i2
| (y1 f (x)1 ) (y0 f (x)0 ) i2 ,
| y1 f (x)1 i | y0 f (x)0 i2 .
The second line is the definition of in Z22 , and the ( )( ) inside the ket is not
multiplication but the binary expansion of the number y1 y0 f (x)1 f (x)0 . Thus, our
four combinations of the 2-bit number y with the fixed value f (x) become
Uf
|xin |0i |0i 7 |xin |0 f (x)1 i |0 f (x)0 i ,
Uf
|xin |0i |1i 7 |xin |0 f (x)1 i |1 f (x)0 i ,
Uf
|xin |1i |0i 7 |xin |1 f (x)1 i |0 f (x)0 i , and
Uf
|xin |1i |1i 7 |xi |1 f (x)1 i |1 f (x)0 i .
Applying 0a = a and 1a = a to the RHS of these equalities produces the identities
that will drive the math,
Uf
|xin |0i |0i 7 |xin |f (x)1 i |f (x)0 i ,
Uf
|xin |0i |1i 7 |xin |f (x)1 i |f (x)0 i ,
Uf
|xin |1i |0i 7 |xin |f (x)1 i |f (x)0 i , and
Uf
|xin |1i |1i 7 |xin |f (x)1 i |f (x)0 i .
Of course, for a general f , there are now four possible values for f (x), and we
have to combine those with the four possible values for y, yielding a whopping 16
combinations for this fixed x, as the following table reveals.
476
|xin |yi2
f (x)
Uf |xin |yi2
|xin |0i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |0i
|xin |1i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |1i
|xin |2i2
|xin |f (x)1 i |f (x)0 i = |xin |1i |0i
|xin |3i2
|xin |f (x)1 i |f (x)0 i = |xin |1i |1i
|xin |0i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |1i
|xin |1i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |0i
|xin |2i2
|xin |f (x)1 i |f (x)0 i = |xin |1 |1ii
|xin |3i2
|xin |f (x)1 i |f (x)0 i = |xin |1i |0i
|xin |0i2
|xin |f (x)1 i |f (x)0 i = |xin |1i |0i
|xin |1i2
|xin |f (x)1 i |f (x)0 i = |xin |1i |1i
|xin |2i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |0i
|xin |3i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |1i
|xin |0i2
|xin |f (x)1 i |f (x)0 i = |xin |1i |1i
|xin |1i2
|xin |f (x)1 i |f (x)0 i = |xin |1i |0i
|xin |2i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |1i
|xin |3i2
|xin |f (x)1 i |f (x)0 i = |xin |0i |0i
Building the matrix for Uf when m = 2 is accomplished by considering the four
columns associated with the four possible f (x) values. Because of the 16 possible
sub-matrices of this form, we will just do a couple cases and see if a pattern emerges.
1. When f (x) = 0, columns 4x through 4x + 3 of the matrix for Uf will be (see
topmost four rows of table):
Uf |xin |0i2
Uf |xin |1i2
Uf |xin |2i2
Uf |xin |3i2
|xin |0i |0i |xin |0i |1i |xin |1i |0i |xin |1i |1i ,
=
477
which translates to the matrix
4x 4x+3
}|
0 0 0 0
.. .. .. ..
. . . .
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
..
.
0
..
.
0
..
.
0
..
.
0
4x
4x+3
or, more concisely,
4x4x+3
z }| {
1
0
4x
4x+3
As youll see in a moment, it will facilitate the analysis if we break the 4 4
identity matrix into four smaller 2 2 matrices, and write this as
4x 4x+3
}|
0
1
4x
4x+3
This 4 4 matrix that appears in columns 4x 4x + 3 is unitary by inspection.
478
2. Lets skip to the case f (x) = 3, and use the table to calculate columns 4x
through 4x + 3 of the matrix for Uf (see bottommost four rows of table):
Uf |xin |0i2
Uf |xin |1i2
Uf |xin |2i2
Uf |xin |3i2
|xin |1i |1i |xin |1i |0i |xin |0i |1i |xin |0i |0i ,
=
which translates to the matrix
4x 4x+3
}|
0 0 0 0
.. .. .. ..
. . . .
0
0
0
1
0
0
1
0
0
1
0
0
1
0
0
0
..
.
0
..
.
0
..
.
0
..
.
0
4x
4x+3
or, more concisely,
4x 4x+3
}|
0
0
4x
4x+3
This 4 4 matrix that appears in columns 4x 4x + 3 is unitary by inspection.
479
3. Exercise: When f (x) = 1, show that columns 4x 4x + 3 of Uf are
4x 4x+3
}|
4x
4x+3
This 4 4 matrix that appears in columns 4x 4x + 3 is unitary by inspection.
4. Exercise: When f (x) = 2, show that columns 4x 4x + 3 of Uf are
4x 4x+3
}|
4x
4x+3
This 4 4 matrix that appears in columns 4x 4x + 3 is unitary by inspection.
Summary of Case m = 2
Weve covered all four possible values for f (x) and done so for arbitrary x = 0, . . . , 2n
1. What we have discovered is that
each x controls four columns in the final matrix;
one of four possible 4 4 unitary (sub-)matrices is positioned in these four
columns, with 0s above and below that 4 4 matrix;
the four possible sub-matrices for f (x) = 0, 1, 2 and 3, respectivly, are
x
x
1
480
and
x
x
since, for each x, its 4 4 unitary matrix is in rows 4x 4x + 3, with zeros
above and below, it follows (exercise) that there can only be zeros to the left
and right on those rows.
Once again, even though we expected Uf to be unitary, its matrix can be seen
to exhibit this property on its own merit, and the only non-zero elements are on
4 4 sub-matrices that lie along the diagonal. This characterizes all oracles for
f : Z2n Z22 .
17.5.4
Easy Way to See Unitarity
Now that weve analyzed a few oracles, we can see why all Uf will be unitary. In fact,
you might have noticed this even before having studied the examples. Here are the
key points, and Ill let you fill in the details as an [Exercise].
Every column is a CBS ket since it is the separable product of CBS kets.
All CBS kets are normal vectors (all coordinates are 0 except one, which is 1).
If two columns were the identical, Uf would map two different CBS kets to the
same CBS ket.
Any matrix that maps two different CBS kets to the same CBS ket cannot be
invertible.
Since Uf is its own inverse, it is invertible, so by last two bullets, all columns
are distinct unit vector.
We conclude that distinct columns have their solitary 1 in different positions.
The inner product of these columns with themselves is 1 and with other columns
is 0: the matrix has orthonormal columns.
QED
Of course, we learn more by describing the form of the matrices, so the activities
of this chapter have value beyond proving unitarity.
17.5.5
The General Advanced Oracle: Any m (Range of f
Z 2m )
Extending Conclusions to Larger Uf Presents No Difficulties
This is where we should apply our intuition and extrapolate the m = 2 case to all
m. Each time m increases by 1, the number of columns in Uf controlled by a single
input value x doubles. For m = 3, we would have 8 8 unitary matrices along the
diagonal, each built from appropriate combinations of x and 1. For m = 4, we would
have unitary 16 16 matrices along the diagonal. And when m = n, we would have
481
2n 2n unitary sub-matrices along the diagonal of a really large 22 22 matrix
for Uf . Our explicit demonstrations in the m = 1 and 2 cases can be duplicated to
any m with no theoretical difficulty. Wed only be dealing with lengthier tables and
larger sub-matrices. So if the number of output bits of f is m, for m > 2, the results
are the same: Uf is unitary and it consists of all 0s except near the diagonal where
sub-matrices are built from combinations of x and 1.
The Diagonal Sub-Matrices are Separable Products
However, if you are interested in one approach to a more rigorous understanding of
Uf for general m, start here. In the m = 2 case, our four possible matrices comprising
the 44 sub-matrices along the diagonal are actually tensor products of two matrices:
1
x
1 1
1 x
1
= x 1
x
x x
Each component of the tensor product, 1 and x , were the two possible 2 2 matrices
that appeared along the diagonal of Uf for m = 1.
The tensor product of unitary matrices is easily shown to be unitary. And one
could proceed inductively to demonstrate that the Uf for output size m + 1 consists
of sub-matrices along the diagonal, each of which is a tensor products of the potential
sub-matrices available to the Uf for size m. This will give both unitarity as well as
the visual that we have already predicted.
[Caution. If we designate some function with 2n inputs and 2m outputs as fn,m ,
we are not saying that Ufn,(m+1) = Ufn,m Ugn,m for two smaller oracles it isnt. That
statement doesnt even track, logically, since it would somehow imply we could generate an arbitrary Uf , which can be exponentially complex, out of repeated products
of something very simple. Rather, we merely noted the fact that the sub-matrices
along the diagonal are, individually, tensor products of the next-smaller ms possible
diagonal sub-matrices. We still need the full details of f to compute all the smaller
sub-matrices, which will be different (in general) from one another.]
482
17.6
The Complexity of a Quantum Algorithm Relative to the Oracle
In this course, we are keeping things as simple as possible while attempting to provide
the key ideas in quantum computation. To that end, well make only one key classification of oracles used in algorithms. You will undoubtedly explore a more rigorous
and theoretical classification in your advanced studies.
Complexity of the Oracle
The above constructions all demonstrated that we can take an arbitrary function, f ,
and, in theory, represent it as a reversible gate associated with a unitary matrix. If
you look at the construction, though, youll see that any function which requires a full
table of 2n values to represent it (if there is no clear analytical short-cut we can use
to compute it) will likewise end up with a similarly complicated oracle. The oracle
would need an exponentially large number of gates (relative to the number of binary
inputs, n). An example will come to us in the form of Simons algorithm where we
have a Z2n -periodic function (notation to be defined) and seek to learn its period.
However, there are functions which we know have polynomial complexity and can
be realized with a correspondingly simple oracle. An example of this kind of oracle
is that which appears in Shors factoring algorithm (not necessarily Shors periodfinding algorithm). In a factoring algorithm, we know a lot about the function that
we are trying to crack and can analyze its specific form, proving it to be O(n3 ). Its
oracle will also be O(n3 ).
Two Categories of a Quantum Algorithms Complexity
Therefore, whenever presenting a quantum algorithm, one should be clear on which
of the two kinds of oracles are available to the circuit. That distinction will be made
using the following, somewhat informal (and unconventional) language.
Relativized Time Complexity - This is the time complexity of a {circuit +
algorithm} without knowledge of the oracles design. If we were later given the
complexity of the oracle, we would have to modify any prior analysis to account
for it. Until that time we can only speak of the algorithm and circuit around
the oracle. We can say that some algorithm is O(n3 ), e.g., but that doesnt
mean it will be when we wire up the oracle. It is O(n3 ) relative to the oracle.
Absolute Time Complexity - This is the time complexity of a {circuit +
algorithm} with knowledge of the oracles design. We will know and incorporate
the oracles complexity into the final {circuit + algorithm}s complexity.
483
Chapter 18
Simons Algorithm for
Period-Finding
18.1
The Importance of Simons Algorithm
Simons algorithm represents a turning point in both the history of quantum computer
science and in every students study of the field. It is the first algorithm that we study
which represents a substantial advance in relativized time complexity vs. classical
computing. The problem is exponential classically, even if we allow an approximate
rather than a deterministic solution. In contrast, the quantum algorithm is very fast,
O(log5 N ), where N is the size of the domain of f . In fact, you will see a specific
algorithm, more detailed than those usually covered, which has complexity O(log3 N ).
Well prove all this after studying the quantum circuit.
[Before going any further, make sure you didnt overlook the word relativized
in the first paragraph. As you may recall from a past lecture, it means that we
dont have knowledge, in general, about the performance of the circuits oracle Uf .
Simons quantum algorithm gives us a lower bound for the relative complexity, but
that would have to be revised upward if we ended up with an oracle that has a larger
big-oh.. This is not the fault of QC; the kinds of periodic functions covered in
Simons treatment are arbitrarily complex. The function that we test may have a
nice O(n) or O(n2 ) implementation, in which case well be in great shape. If not,
it will become our bottleneck. That said, even with a polynomial-fast oracle, there
is no classical algorithm that can achieve polynomial time solution, so the quantum
solution is still a significant result.]
While, admittedly a toy problem in the sense that its not particularly useful,
it contains the key ingredients of the relevant algorithms that follow, most notably
Shors quantum factoring and encryption-breaking. Even better, Simons algorithm
is free of the substantial mathematics required by an algorithm like Shors and thus
embodies the essence of quantum computing (in a noise-free environment) without
distracting complexities.
484
In the problem at hand, we are given a function and asked to find its period. However, the function is not a typical mapping of real or complex numbers, and the period
is not the thing that you studied in your calculus or trigonometry classes. Therefore,
a short review of periodicity, and its different meanings in distinct environments, is
in order.
18.2
Periodicity
Well first establish the meaning of periodicity in a typical mathematical context
the kind you may have seen in past calculus or trig course. Then well define it in a
n
more exotic mod-2 environment Z2n
= (Z2 ) .
18.2.1
Ordinary Periodicity
A function defined on the usual sets, e.g.,
R
C
S
(S could be R, Z2 , etc.) ,
f:
Z
is called periodic if there is a unique smallest a > 0 (called the period),
with
f (x + a)
f (x) ,
for all x in the domain of f .
(Unique smallest is redundant, but it forces us to remember the second of two
separate requirements. First, there must exist some a > 0 with the property. If there
does, we would call the smallest such a its period.) For example,
sin(x + 2) = sin(x) for all x,
and 2 is the smallest positive number with this property, so a = 2 is its period. 4
and 12 satisfy the equality, but theyre not as small as 2, so theyre not periods.
Z2 )n Periodicity
(Z
18.2.2
Lets change things up a bit. We define a different sort of periodicity which respects
not ordinary addition, but mod-2 addition.
A function defined on (Z2 )n ,
f:
(Z2 )n
(S is typically Z or
485
(Z2 )m , m n 1) ,
Z2 )n periodic if there exists an a (Z2 )n , a 6= 0, such that,
is called (Z
for all x 6= y in (Z2 )n , we have
f (x) = f (y)
y = x a.
a is called the period of f , with (Z2 )n periodicity understood by context.
(Notice that in our definition, we rule out period a = 0. We could have allowed it,
in which case one-to-one functions see definition below are periodic with period
0, but this adds a trivial extra case to our discussion of Simons algorithm which has
no pedagogical or historical value.)
There is a subtle but crucial message hidden in the definition. The if-and-only-if
() aspect implies that if we find even one pair of elements, x0 6= y0 with f (x0 ) =
f (y0 ), then it must be true that y0 = x0 a. Once we know that, we have the
pleasant consequence that this one pair can be combined to recover the period using
a = x0 y0 . (Confirm this for yourself.) This is very different from the more common
periodicity of, say, sin x, wherein sin 0 = sin , but is not its period since for many
other pairs (like .2 and .2 + ) sin x 6= sin (x + ). It turns out that we wont make
use of this property in the quantum algorithm, but it will help when analyzing the
time complexity of the classical algorithm.
Due to the isomorphism between (Z2 )n and (Z2n , ), we can use the simple integer
notation along with the operator to effect the same definition, as we demonstrate
next.
18.2.3
Z 2n Periodicity
f is Z 2n periodic if there exists an a Z2n , a 6= 0, such that,
for all x 6= y in Z2n , we have
f (x) = f (y)
y = x a.
a is called the period of f , with Z2n periodicity understood by context.
As I mentioned, this is the same periodicity defined above and has the same implications. We will use the two periodicity definitions interchangeably going forward.
Z2 )n Periodicity
Examples of (Z
We implied that the range of f could be practically any set, S, but for the next few
examples, lets consider functions that map Z2n into itself,
f : Z2n
486
Z2n
Collapsing One Bit Periodic
A simple (Z2 )n periodic function is one that preserves all the bits of x except for one,
say the kth, and turns that one bit into a constant (either 0 or 1).
Let n = 5, k = 2, and the constant, 1. This is a 2nd bit collapse-to-1, algebraically
x4
x3
f (x)
1 = x4 x3 1 x1 x0
x1
x0
If n = 5, k = 4, and the constant was 0, wed have a 4th bit collapse-to-0,
0
x3
g(x)
x2 = 0 x3 x1 x1 x0 .
x1
x0
Lets show why the first is (Z2 )5 periodic with period 4, and this will tell us why all
the others of its ilk are periodic (with possibly a different period), as well.
Notation - Denote the bit-flip operation on the kth bit, xk , to mean
(
0, if xk = 1,
xk
1, if xk = 0.
[Exercise. Demonstrate that you can effect a bit-flip on the kth bit of x using x2k .]
I claim that the 2nd-bit collapse-to-1 function, f , is (Z2 )n periodic with period
a = 4. We must show
f (x) = f (y)
:
y = x 4.
Assume y = x 4. That means (in vector notation)
x4
y4
x4
x3
y 3
x3
x =
and
y =
x2
y 2 = x2 .
x1
y 1
x1
x0
y0
x0
Then,
x4
x3
f (x) =
1 ,
x1
x0
but, also
proving f (x) = f (y).
487
x4
x3
f (y) =
1 ,
x1
x0
: Assume f (x) = f (y) for x 6= y. Since f does not modify any bits other
than bit 2, we conclude yk = xk , except for (possibly) bit k = 2. But since
x 6= y, some bit must be different between them, so it has to be bit-2. That is,
y4
x4
y 3
x3
= x2 = x 4,
y
y =
2
y 1
x1
x0
y0
showing that y = x 4.
QED
Of course, we could have collapsed the 2nd bit to 0, or used any other bit, and gotten
the same result. A single bit collapse is (Z2 )n periodic with period 2k , where k is the
bit being collapsed.
Note - With single bit collapses, there are always exactly two numbers from the
domain Z2n that map into the same range number in Z2n . With the example, f , just
discussed,
)
00100
7 00100 ,
00000
)
10101
7 10101 ,
10001
)
11111
7 11111 ,
11011
and so on.
Collapsing More than One Bit Not Periodic
A collapse of two or more bits can never be periodic, as we now show. For illustration,
lets stick with n = 5, but consider a simultaneous collapse-to-1 of both the 2nd and
0th bit,
x4
x3
f (x)
1 = x4 x3 1 x1 1
x1
1
In this situation, for any x in the domain of f , you can find three others that map to
the same f (x). For example,
00000 = 0
00001 = 1
7 00101 = 5 .
00100 = 4
00101 = 5
488
If there were some period, a, for this function, it would have to work for the first two
listed above, meaning, f (0) = f (1). For that to be true, wed need 1 = 0 a, which
forces a = 1. But a = 1 wont work when you consider the first and third x, above:
f (0) = f (4), yet 4 6= 0 1.
As you can see, this comes about because there are too many xs that get mapped
to the same f (x).
Lets summarize: bit collapsing gives us a periodic function only if we collapse
exactly one bit (any bit) to a constant (either 0 or 1). However, a function which is
a multi-bit-collapser (preserving the rest) can never be periodic.
So, what are the other periodic functions in the Z2n milieu?
18.2.4
1-to-1, 2-to-1 and n-to-1
The answer requires a simple definition. First some more terminology.
Domain and Range
If f is a function, well use dom(f ) to designate the domain of f , and ran(f ) to
designate the range of f . If w ran(f ), the pre-image of w is the set {x | f (x) = w}.
n-To-One
One-To-One: We say f is 1-to-1 when x 6= y
all x, y in dom(f ).
f (x) 6= f (y), for
If S dom(f ) then, even if f is not 1-to-1, overall, it can still be 1-to-1 on S.
(Student to fill in the adjustments to the definition).
Two-To-One: We say f is 2-to-1 when every w in ran(f ) has exactly
two pre-images in dom(f ).
n-To-One: We say f is n-to-1 when every w in ran(f ) has exactly n
pre-images in dom(f ).
Lots of Periodic Functions
(In this small section we will set up a concept that youll find useful in many future
quantum algorithms: the partitioning of dom(f ) into subsets.)
We now characterize all Z2n periodic functions.
Lets say you have one in your hands, call it f , and you even know its period, a.
Buy three large trash bins at your local hardware store. Label one R, one Q and the
third, S (for S ource pool). Dump all x dom(f ) (which is Z2n ) into the source pool,
489
S. Were going to be moving numbers from the source, S, into R and Q according to
this plan:
1. Pick any x S = Z2n . Call it r0 (0, because its our first pick).
2. Generate r0 s partner q0 = r0 a (partner, because periodicity guarantees that
f maps both r0 and q0 to the same image value and also that there is no other
x Z2n which maps to that value.)
3. Toss r0 into bin R and its partner, q0 , (which you may have to dig around in S
to find) into bin Q.
4. We have reduced the population of the source bin, S, by two: S = S {r0 , q0 }.
Keep going . . . .
5. Pick a new x from whats left of S. Call it r1 (1, because its our second pick).
6. Generate r1 s partner q1 = r1 a. (Again, we know that f maps both r1 and
q1 to the same image value, and there is no other x Z2n which maps to that
value.)
7. Toss r1 into bin R and its partner, q1 , into bin Q. S is further reduced by two
and is now S = S {r0 , q0 , r1 , q1 }.
8. Repeat this activity, each pass moving one value from bin S into bin R and its
partner into bin Q until we have none of the original domain numbers left in S.
9. When were done, half of the xs from dom(f ) will have ended up in R and the
other half in Q. (However, since we chose the first of each pair at random, this
was not a unique division of dom(f ), but thats not important.)
Here is the picture, when were done with the above process:
Z2n =
R
Q
= { . . . , rk , . . .} { . . . , qk , . . . }
where R and Q are of equal size, and
f (rk )
f (qk )
but f is otherwise one-to-one on R and Q, individually.
From this construction, we can see a couple things:
1. Every periodic function in the Z2n sense is 2-to-1. If a function is not 2-to-1, it
doesnt have a prayer of being periodic.
490
2. We have a way to produce arbitrary Z2n periodic functions. Pick a number
you want to be the period and call it a. Start picking rk numbers from Z2n
at random, each pick followed immediately by the computation of its partner,
qk = rk a. Assign any image value you like to these two numbers. (The
assigned image values dont even have to come from Z2n they can be the
chickens in your yard or the contacts in your phone list. All that matters is
that once you use one image value for an rk qk pair, you dont assign it to any
future pair.) This will define a Z2n periodic function with period a.
[Exercise. Generate two different Z2n periodic functions.]
18.3
Simons Problem
Weve developed all the vocabulary and intuition necessary to understand the statement of Simons problem, and the quantum fun can now begin.
Statement of Simons Problem
Let f : (Z2 )n (Z2 )n be (Z2 )n periodic.
Find a.
(I made a bold because I phrased the problem in terms of the vector space (Z2 )n ;
had I used the integer notation of Z2n , I would have said find a, without the bold.)
Its not really necessary that ran(f ) be the same (Z2 )n as that of its domain. In
fact, the range is quite irrelevant. However, the assumption facilitates the learning
process, and once we have our algorithm we can lose it.
Lets summarize some of the consequences of (Z2 )n periodicity.
1. f (y) = f (x)
y = xa
2. f (y) = f (x)
a = xy
3. f is two-to-one on (Z2 )n
We can also state this in the equivalent Z2n language.
1. f (y) = f (x)
y = xa
2. f (y) = f (x)
a = xy
3. f is two-to-one on Z2n
Pay particular attention to bullet 2. If we get a single pair that map to the same
f (x) = f (y), we will have our a. This will be used in the classical (although not
quantum) analysis.
491
18.4
Simons Quantum Circuit Overview and the
Master Plan
18.4.1
The Circuit
A birds eye view of the total circuit will give us an idea of whats ahead.
|0in
H n
H n
Uf
|0in
You see a familiar pattern. There are two multi-dimensional registers, the upper
(which I will call the A register, the data register or even the top line, at my whim),
and the lower (which I will call the B register, target register or bottom line, correspondingly.)
This is almost identical to the circuits of our recent algorithms, with the following
changes:
The target channel is hatched, reflecting that it has n component lines rather
than one.
We are sending a |0in into the bottom instead of a |1i (or even |1in ).
We seem to be doing a measurement of both output registers rather than ignoring the target.
In fact, that third bullet concerning measuring the bottom register will turn out to be
conceptual rather than actual. We could measure it, and it would cause the desired
collapse of the upper register, however our analysis will reveal that we really dont
have to. Nevertheless, we will keep it in the circuit to facilitate our understanding,
label it as conceptual, and then abandon the measurement in the end when we are
certain it has no practical value.
Here is the picture Ill be using for the remainder of the lesson.
|0in
H n
H n
Uf
|0in
| {z }
(actual)
| {z }
(conceptual)
[Note: I am suppressing the hatched quantum wires to produce a cleaner circuit.
Since every channel is has n lines built-into it and we clearly see the kets and operators
labeled with the exponent n, the hatched wires no longer serve a purpose.]
492
18.4.2
The Plan
We will prepare a couple CBS kets for input to our circuit, this time both will be |0in .
The data channel (top) will first encounter a multi-dimensional Hadamard gate to
create a familiar superposition at the top. This sets up quantum parallelism which we
found to be pivotal in past algorithms. The target channels |0in will be sent directly
into the oracle without pre-processing. This is the first time we will have started with
a |0i rather than a |1i in this channel, a hint that were not going to get a phase
kick-back today. Instead, the generalized Born rule, (QM Trait #15) will turn out
to be our best friend.
[Preview: When we expect to achieve our goals by applying the Born rule to a
superposition, the oracles target register should normally be fed a |0in rather than
a |1in .]
After the oracle, both registers will become entangled.
At that point, we conceptually test the B register output. This causes a collapse
of both the top and bottom lines states (from the Born rule), enabling us to know
something about the A register. Well analyze the A registers output which resulted from this conceptual B register measurement and discover that it has very
special properties. Post processing the A register output by a second re-organizing
Hadamard gate will seal the deal.
In the end, we may as well have measured the A register to begin with, since
quantum entanglement authorizes the collapse using either line, and the A register is
what we really care about.
Strategy
Our strategy will be to load the dice by creating a quantum circuit that spits out
measured states which are orthogonal to the period a, i.e., z a = 0 mod-2. (This is
not a true orthogonality, as well see, but everyone uses the term and so shall we. Well
discover that states which are orthogonal to a can often include the vector a, itself,
another reason for the extra care with which we analyze our resulting orthogonal
states.)
That sounds paradoxical; after all, we are looking for a, so why search for states
orthogonal to it? Sometimes in quantum computing, its easier to back-into the
desired solution by sneaking up on it indirectly, and this turns out to be the case in
Simons problem. You can try to think of ways to get a more directly, and if you
find an approach that works with better computational complexity, you may have
discovered a new quantum algorithm. Let us know.
Because the states orthogonal to a are so much more likely than those that are not,
we will quickly get a linearly independent set of n 1 equations with n 1 unknowns,
namely, awk = 0, for k = 0, . . . , n2. We then augment this system instantly (using
a direct classical technique) with an nth linearly independent equation, at which point
we can solve the full, non-degenerate n n system for a using fast and well-known
493
techniques.
18.5
The Circuit Breakdown
We need to break the circuit into sections to analyze it. Here is the segmentation:
|0in
H n
H n
Uf
| {z }
(actual)
|0in
| {z }
(conceptual)
18.6
Circuit Analysis Prior to Conceptual Measurement: Point B
We now travel carefully though the circuit and analyze each component and its consequence.
18.6.1
Hadamard Preparation of the A Register
This stage of the circuit is identical in both logic and intent with the Deutsch-Jozsa
and Bernstein-Vazirani circuits; it sets up quantum parallelism by producing a perfectly mixed entangled state, enabling the oracle to act on f (x) for all possible x,
simultaneously.
|0in
H n
H n
Uf
|0in
Hadamard, H n , in H(n)
494
It never hurts to review the general definition of a gate like H n . For any CBS |xin ,
the 2n -dimensional Hadamard gate is expressed in encoded form using the formula,
n 1
n 2X
1
n
H n |xi
=
(1)x y |yin ,
2
y=0
where is the mod-2 dot product. Today, Ill be using the alternate vector notation,
n 1
n 2X
1
n
n
(1)x y |yin ,
H |xi
=
2
y=0
where the dot product between vector x and vector y is also assumed to be the mod-2
dot product. In the circuit, we have
n 1
n 2X
n
1
(1)x y |yin ,
|xi
H n
2
y=0
which, when applied to |0in , reduces to
|0in
H n
1
2
n 1
n 2X
|yin ,
y=0
or, returning to the usual computational basis notation, |xin , for the summation index
is
n 1
n 2X
n
1
|0i
|xin .
H n
2
x=0
Youll recognize the output state of this Hadamard operator as the nth order x-basis
CBS ket, |0in . It reminds us that not only do Hadamard gates provide quantum
parallelism but double as a z x basis conversion operator.
18.6.2
The Quantum Oracle on CBS Inputs
Next, we look at this part of the circuit:
|0in
H n
H n
Uf
|0i
|
{z
}
Quantum Oracle
495
Due to the increased B channel width, we had better review the precise definition of
the higher dimensional oracle. Its based on CBS kets going in,
|xin
|xin
Uf
|y f (x)in
|yi
Uf
|xin |yin
|xin |y f (x)in ,
and from there we extend to general input states, linearly. We actually constructed
the matrix of this oracle and proved it to be unitary in our lesson on quantum oracles.
Today, we need only consider the case of y = 0:
|xin
|xin
Uf
|f (x)in
|0i
|xin |0in
Uf
|xin |f (x)in
In Words
We are
1. taking the B register CBS input |yin , which is |0in in this case, and extracting
the integer representation of y, namely 0,
2. applying f to the integer x (of the A register CBS |xin ) to form f (x),
3. noting that both 0 and f (x) are Z2n ,
4. forming the mod-2 sum of these two integers, 0 f (x), which, of course is f (x),
and
5. using the result to define the output of the oracles B register, |f (x)in .
6. Finally, we recognize the output to be a separable state of the two output
registers, |xin |f (x)in .
Just to remove any lingering doubts, assume n = 5, x = 18, and f (18) = 7. Then
the above process yields
1. |0i5 0
f (18)=7
2. |18i5 18 7
496
3. 0 7 = 00000 00111 = 00111 = 7,
4. 7 |7i5
Uf
5. |18i5 |0i5 |18i5 |7i5
18.6.3
The Quantum Oracle on Hadamard Superposition Inputs
Next, we invoke linearity to the maximally mixed superposition state |0in going into
the oracles top register.
Reminder. Ive said it before, but its so important in a first course such as
this that Ill repeat myself. The bottom registers output is not f applied to the
superposition state. f only has meaning over its domain Z2n , which corresponds to
the finite set of z-basis CBS kets { |xi}. It has no meaning when applied to sums,
especially weighted sums (by real or complex amplitudes) of these preferred CBSs.
By linearity, Uf distributes over all the terms in the maximally mixed input |0in ,
at which point we apply the result of the last subsection, namely
Uf |xin |0in
= |xin |f (x)in ,
to the individual summands to find that
2n 1
n X
|xin |0in
Uf
x=0
2n 1
n X
|xin |f (x)in .
x=0
This is a weighted sum of separable
(The weights are the same for each
products.
n
separable product in the sum: (1/ 2) .) That sum is not, as a whole, separable,
which makes it impossible to visualize directly on the circuit diagram unless we combine the two outputs into a single, entangled, output register. However we do have
an interpretation that relates to the original circuit.
The output state is a superposition of separable terms |xin
exactly the kind of sum the generalized Born rule needs,
|in+m
|f (x)in
.
( 2)n
But this is
n
m
n
m
n
|0inA |0 im
B + |1iA |1 iB + + |2 1iA |2n 1 iB ,
so an A-measurement of x would imply a B-state collapse to its (normalized)
partner, |f (x)in .
A similar conclusion would be drawn if we chose to measure the B register first,
but well get to that slightly more complicated alternative in a moment. (Sneak
preview: there are two not one pre-images x for every f (x) value.)
497
n
Each of the 2n orthogonal terms in the superposition has amplitude 1/ 2 ,
so the probability that
by A will collapse the superposition to
ameasurement
n 2
any one of them is
1/ 2
= 1/2n .
[Exercise. Why are the terms orthogonal? Hint: inner product of tensors.]
n 1
2P
[Exercise. Look at the sum in our specific situation:
|xin |f (x)in . QM
x=0
Trait #6 (Probability of Outcomes) assumes we start with an expansion along some
computational basis, but this expansion only has 2n terms, so cant be a basis for
A B which has 2n 2n = 22n basis vectors. Why can we claim that the scalars
n
1/ 2 are amplitudes of collapse to one of these states? Hint: While the kets
in the sum do not comprise a full basis, they are all distinct CBS kets. (Are they
distinct? each f (x) appears twice in the sum. But that doesnt destroy the linear
independence of the tensor CBS kets in that sum, because . . . .) Therefore, we can
add the missing CBS kets into the sum as long as we accompany them by 0 scalar
weights. Now we can apply Trait #6.]
n 1
2P
[Exercise. Look at the the more general sum in the Born rule:
|xinA |x im
B .
x=0
QM Trait #6 (Probability of Outcomes) assumes we start with an expansion along
some computational basis, but this expansion only has 2n terms, so cant be a basis
for
A B which has 2n 2m = 2n+m basis vectors. Why can we claim that the scalars
n
1/ 2 are amplitudes of collapse to one of these states? Hint: Force this to be
an expansion along the tensor CBS by expanding each |x im
B along the B-basis then
distributing the |xinA s. Now we meet the criterion of Trait #6.]
18.6.4
Partitioning the Domain into Cosets
Its time to use the fact that f is Z2n periodic with (unknown) period a to help us
rewrite the output of the Oracles B register prior to the conceptual measurement.
Z2n periodicity tells us that the domain can be partitioned (in more than one way)
into two disjoint sets, R and Q,
Z2n =
R
Q
= { , x, } { , x a, } .
Cosets
Q can be written as R a, that is, we translate the entire subset R by adding a
number to every one of its members resulting in a new subset, Q. Mathematicians
say that Q = R a is a coset of the set R. Notice that R is a coset of itself since
R = R 0. In this case, the two cosets, R and Q = R a partition the domain of f
into two equal and distinct sets.
498
18.6.5
Rewriting the Output of the Oracles B Register
Our original expression for the oracles complete entangled output was
2n 1
n X
|xin |f (x)in ,
x=0
but our new partition of the domain will give us a propitious way to rewrite this.
Each element x R has a unique partner in Q satisfying
x
f
7 f (x) .
x a xR
Using this fact, we only need to sum the B register output over R (half as big as Z2n )
and include both x and x a in each term,
2n 1
n X
x=0
n X
1
|xin + |x ain |f (x)in
|xi |f (x)i
=
2
xR
n1 X n
1
|xi + |x ain
=
|f (x)in
2
2
n
xR
I moved one of the factors of 1/ 2 into the sum so we could see
1. how the new sum consists of normalized states (length 1), and
2. the common amplitude remaining on the outside, nicely produces a normalized
state overall, since now there are half as many states in the sum, but each state
has twice the probability as before.
Swapping the roles of channels A and B for the Born Rule
The last rearrangement of the sum had a fascinating consequence. While the terms
still consist of separable products from the A and B channels, now it is the B channel
that has basis CBS kets, and the A channel that does not.
How can we see this? Each |f (x)in was always a CBS state f (x) is an integer
from 0 to 2n 1 and so corresponds to a CBS ket but the original sum was plagued
by its appearing twice (destroying potential linear independence of the terms), so we
couldnt view the sum as an expansion along the B-basis. After consolidating all
the pre-image pairs, we now have only one |f (x)in term for each x in the sum: we
have factored the expansion along the B-basis. In the process, the A-factors became
superpositions of A CBS kets, now mere |x inA s in the general population. This
reverses the roles of A and B and allows us to talk about measuring B along the
CBS, with that measurement selecting one of the non-CBS A factors.
499
The upshot is that we can apply the Born rule in reverse; well be measuring the B
register and forcing the A register to collapse into one of its binomial superpositions.
Lets do it. But first, we should give recognition to a reusable design policy.
The Lesson: |0in into the Oracles B Register
This all worked because we chose to send the CBS |0in into the oracles B register.
Any other CBS into that channel would not have created the nice terms |xin |f (x)in
of the oracles entangled output. After factoring out terms that had common |f (x)in
components in the B register, we were in a position to collapse along the B-basis and
pick out the attached sum in the A register.
Remember this. Its a classic trick that can be tried when we want to select a
small subset of A register terms from the large, perfectly mixed superposition in that
register. It will typically lead to a probabilistic outcome that wont necessarily settle
the algorithm in a single evaluation of the oracle, but we expect it to give a valuable
result that can be combined with a few more evaluations (loop passes) of the oracle.
This is the lesson we learn today that will apply next time when we study Shors
period-finding algorithm.
In contrast, when we were looking for a deterministic solution in algorithms like
Deutsch-Jozsa and Bernstein-Vazirani, we fed a |1i into the B register and used the
phase kick-back to give us an answer in a single evaluation.
|1i into oracles B register
|0in into oracles B register
phase kick-back ,
Born rule .
18.7
Analysis of the Remainder of the Circuit: Measurements
18.7.1
Hypothetical Measurement of the B Register
Although we wont really need to do so, lets imagine what happens if we were to
apply the generalized Born rule now using the rearranged sum (that turned the B
channel into the CBS channel).
|0in
H n
H n
Uf
|0in
|
{z
Conceptual
Each B register measurement of f (x) will be attached to not one, but two, input A
register states. Thus, measuring B first, while collapsing A, actually produces merely
500
a superposition in that register, not a single, unique x from the domain. It narrows
things down considerably, but not completely,
n1 X |xin + |x ain
1
|f (x)in
2
2
xR
|x0 in + |x0 ain
|f (x0 )in
&
2
Here, & means collapses to.
Well thats good, great and wonderful, but if after measuring the post-oracle B register, we were to measure line A, it would collapse to one of two states, |x0 i or |x0 + ai,
but we wouldnt know which nor would we know its unsuccessful companion (the one
to which the state didnt collapse). There seems to be no usable information here.
As a result we dont measure A ... yet.
Lets name the collapsed but unmeasured superposition state in the A register
|x0 in , since it is determined by the measurement f (x0 ) of the collapsed B register,
|x0 in + |x0 ain
n
.
|x0 i
2
Guiding Principle: Narrow the Field. We stand back and remember this
stage of the analysis for future use. Although a conceptual measurement of B does not
produce an individual CBS ket in register A, it does result in a significant narrowing
of the field. This is how the big remaining quantum algorithms in this course will
work.
18.7.2
Effect of a Final Hadamard on A Register
In an attempt to coax the superposition ket |x0 in H(n) to cough up useful information, we take H n |x0 in . This requires that we place an H n gate at the output
of the oracles A register:
|0in
H n
H n
Uf
|0in
[Apology. I cant offer a simple reason why anyone should be able to intuit a
Hadamard as the post-oracle operator we need. Unlike Deutsch-Jozsa, today we
are not measuring along the x-basis, our motivation for the final H n back then.
However, there is a small technical theorem abouta Hadamard applied to a binomial
superposition of CBSs of the form (|xin + |yin ) / 2 which is relevant, and perhaps
this inspired Simon and his compadres.]
501
Continuing under the assumption that we measure an f (x0 ) at the B register output, thus collapsing both registers, we go on to work with the resulting superposition
|x0 i in the A register. Lets track its progress as we apply the Hadamard gate to it.
As with all quantum gates, H n is linear so moves past the sum, and we get
|x0 in + |x0 ain
H n |x0 in + H n |x0 ain
H n
The Hadamard of the individual terms is
H n |x0 in =
H n |x0 ain
2n 1
n X
(1)y x0 |yin
y=0
and
2n 1
n X
1
=
(1)y (x0 a) |yin
2
y=0
Focus on the integer coefficient
(1)y (x0 a) .
This can be simplified by observing two facts
1. The mod-2 dot product distributes over the mod-2 sum , and
2. (1)p q = (1)p (1)q , for p and q in Z2n . (Danger: while its true for a
base of -1, it is not for a general complex base c).
[Exercise. Prove the second identity. One idea: Break it into two cases, p q
even and p q odd, then consider each one separately.]
[Exercise. Find a complex c, for which (c)p q
simplest case, n = 1.]
6=
(c)p (c)q . Hint: Use the
Combining both facts, we get
(1)y (x0 a) = (1)y x0 (1)y a ,
502
so
H n |x0 in + H n |x0 ain
1
2
n 1
n 2X
y=0
=
18.7.3
(1)y x0
2n 1
n+1 X
1 + (1)y a
|yin
(1)y x0
1 + (1)y a
|yin .
y=0
The Orthogonality of A Register Output Relative to
the Unknown Period a
Heres where we are:
n
n
n |x0 i + |x0 ai
H
2
2n 1
n+1 X
1
y x0
ya
=
(1)
1 + (1)
|yin .
2
y=0
The integer expression in the parentheses is seen to be
0, if y a = 1 (mod 2)
ya
1 + (1)
=
,
2, if y a = 0 (mod 2)
so we can omit all those 0 terms which correspond to y a = 1 (mod 2), leaving
n1 X
n
n
1
n |x0 i + |x0 ai
H
=
(1)y x0 |yin .
2
2
ya = 0
(mod 2)
Note that the sum is now over only 2
n1
, or exactly half, of the original 2n CBSs.
[Exercise. Show that, for any fixed number a Z2n (or its equivalent vector
a (Z2 )n ) the set of all x with x a = 0 (or x with x a = 0) is exactly half the
numbers (vectors) in the set.]
Mod-2 Orthogonality vs. Hilbert Space Orthogonality
Avoid confusion. Dont forget that these dot products (like y a) are mod-2 dot
products of vectors in (Z2 )n . This has nothing to do with the Hilbert space inner
product nhy | ain , an operation on quantum states.
503
When we talk about orthogonality of vectors or numbers at this stage of the
analysis, we mean the mod-2 dot product, not the state space inner product. Indeed,
there is nothing new or interesting about Hilbert space orthogonality at this juncture:
CBS kets always form an orthonormal set, so each one has inner product = 1 with
itself and inner product = 0 with all the rest. However, that fundamental fact doesnt
help us here. The mod-2 dot product does.
Pause for an Example
If n = 4, and
0
a3
1
a2
a = 5 =
0 = a1
1
a0
then
y | y a = 0 (mod 2)
a3
0
=
a1
a1 ,a3 {0,1}
a3
1
a1
,
a1 ,a3 {0,1}
which consists of eight of the original 24 = sixteen (Z2 )4 vectors associated with the
CBSs.
Therefore, there
are 2n1= 23 = 8 terms in the sum, exactly normalized by
the (1/ 2)n1 = (1/ 2)3 = (1/ 8) out front.
Note: In this case, a is in the set of mod-2 vectors which are orthogonal to it.
[Exercise. Characterize all a which are orthogonal to themselves, and explain
why this includes half of all states in our total set.]
Returning to our derivation, we know that once we measure the B register and
collapse the states into those associated with a specific |f (x0 )i, the A register can be
post-processed with a Hadamard get to give
|2n1 |
|x0 in + |x0 ain
H n
n1
(1)y x0 |yin .
ya = 0
(mod 2)
All of the vectors in the final A register superposition are orthogonal to a, so we can
now safely measure that mixed state and get some great information:
|0in
H n
H n
Uf
|0in
504
&
|y0 i, y0 a
We dont know which y0 among the 2n1 ys we will measure that depends on the
whimsy of the collapse, and theyre all equally likely. However, we just showed that
theyre all orthogonal to a, including y0 .
Warning(s): y = 0 Possible
This is one possible snag. We might get a 0 when we test the A register. It is a possible
outcome, since 0 = 0 is mod-2 orthogonal to everything. The probabilities are low
1/2n and you can test for it and throw it back if that happens. Well account for
this in our probabilistic analysis, further down. While were at it, remember that a,
itself, might get measured, but thats okay. We wont know its a, and the fact
that it might be wont change a thing that follows.
18.7.4
Foregoing the Conceptual Measurement
We havent yet spoken of the actual A register measurement without first measuring the B register. Now that youve seen the case with the simplifying B register
measurement, you should be able to follow this full development that omits that
step.
If we dont measure B first, then we cant say we have collapsed into any particular f (x0 ) state. So the oracles output must continue to carry the full entangled
summation
n1 X n
|xi + |x ain
1
|f (x)in
2
2
xR
through the final [H n 1n ]. This would add an extra sum
to all of our
xR
expressions.
n
H 1n
=
=
n X
|xin + |x ain
|f (x)in
xR
n1 X
n
n
|xi + |x ain
n
n
H 1
|f (x)i
2
n1 X n
H |xin + H n |x ain
|f (x)in .
2
xR
xR
We can now cite the still valid result that led to expressing the Hadamard fraction
505
as a sum of CBS kets satisfying y a = 0 (mod 2), so the last expression is
n1 X
n1 X
1
1
n
yx
=
(1)
|yi
|f (x)in
2
2
ya = 0
xR
(mod 2)
=
2n2
|yin
ya = 0
(mod 2)
(1)y x |f (x)in .
xR
While our double sum has more overall terms than before, they are all confined to
those y which are (mod-2) orthogonal to a. In fact, we dont have to apply the Born
rule this time, because all that were claiming is an A register collapse to one of the
2n1 CBS kets |yin which we get compliments of quantum mechanics: third postulate
+ post measurement collapse.
Therefore, when we measure only the A register of this larger superposition, the
collapsed state is still some y orthogonal to a.
|0in
H n
H n
&
|yi, y a
Uf
|0in
A Small Change in Notation
Because
I often use the variable y for general CBS states, |yi, or summation variables,
P
, Im going to switch to the variable z for the measured orthogonal output state,
y
as in |zi , z a. Well then have a mental cue for the rest of the lecture, where z
will always be a mod-2 vector orthogonal to a. With this last tweak, our final circuit
result is
|0in
H n
H n
&
|zi, z a
Uf
|0i
18.8
Circuit Analysis Conclusion
In a single application of the circuit, we have found our first z, with z a. We would
like to find n 1 such vectors, all linearly independent as a set, so we anticipate
506
sampling the circuit several more times. How long will it take us to be relatively
certain we have n 1 independent vectors? We explore this question next.
I will call the set of all vectors orthogonal to a either a (if using vector notation)
or a (if using Z2n notation). It is pronounced a-perp.
[Exercise. Working with the vector space (Z2 )n , show that a is a vector subspace. (For our purposes, its enough to show that it is closed under the operation).]
[Exercise. Show that those vectors which are not orthogonal to a do not form a
subspace.]
18.9
Simons Algorithm
18.9.1
Producing n 1 Linearly Independent Vectors
[Notational Alert. By now, were all fluent in translating from the vectors (a, wk )
of Z2n to the encoded decimals (a, wk ) of (Z2 )n notation, so be ready to see me switch
between the two depending on the one I think will facilitate your comprehension. Ill
usually use encoded decimals and give you notational alerts when using vectors.]
What we showed is that we can find a vector orthogonal to a in a single application
of our circuit. We need, however, to find not one, but n 1, linearly-independent
vectors, z, that are orthogonal to a. Does repeating the process n 1 times do the
trick? Doing so would certainly manufacture
{z0 , z1 , z2 . . . , zn2 }
with
zk a
for k = 0, . . . , n 2,
however, thats not quite good enough. Some of the zk might be a linear combination
of the others or even be repeats. For this to work, wed need each one to not only be
orthogonal to a, but linearly independent of all the others as well.
Pause for an Example
If n = 4, and
a3
0
a2
1
a = 5 =
0 = a1
a0
1
suppose that the circuit produced the
0
1
,
0
1
three vectors
0
0
1
0
and
1
1
0
1
507
after three circuit invocations. While all three are orthogonal to a, they do not form
a linearly-independent set. In this case, 3 = n 1 was not adequate. Furthermore, a
fourth or fifth might not even be enough (if, say, we got some repeats of these three).
Therefore, we must perform this process m times, m n 1, until we have n 1
linearly-independent vectors. (We dont have to say that they must be orthogonal to
a, since the circuit construction already guarantees this.) How large must m be, and
can we ever be sure we will succeed?
18.9.2
The Algorithm
We provide a general algorithm in this section. In the sections that follow, well
see that we are guaranteed to succeed in polynomial time O(n4 ) with probability
arbitrarily close to 1. Finally, Ill tweak the algorithm just a bit and arrive at an
implementation that achieves O(n3 ) performance.
Whether its O(n4 ), O(n3 ) or any similar big-O complexity, it will be a significant
relative speed-up over classical computing which has exponential growth, even if we
accept a non-deterministic classical solution. (Well prove that in the final section of
this lesson.) The problem is hard, classically, but easy quantum mechanically.
Here is Simons algorithm for Z2n -period finding. To be sure that it gives a
polynomial time solution, we must eventually verify that
1. we only have to run the circuit a polynomial number of times (in n) to get
n 1 linearly independent zs which are orthogonal to a with arbitrarily good
confidence, and
2. the various classical tasks like checking that a set of vectors is linearly independent and solving a series of n equations are all of polynomial complexity
in n, individually.
Well do all that in the following sections, but right now lets see the algorithm:
Select an integer, T , which reflects an acceptable failure probability of 1/2T .
Initialize a set W to the empty set. W will eventually contain a growing number
of (Z2 )n vectors,
Repeat the following loop at most n + T times.
1. Apply Simons circuit.
2. Measure the output of the final H n to get z.
3. Use a classical algorithm to determine whether or not z is linearly dependent on the vectors in W.
If it is independent, name it wj , where j is the number of elements
already stored in W, and add wj to W.
508
if j = n 2, we have n 1 linearly-independent vectors in W
and are done; break the loop.
If it is not independent (which includes special the case z = 0, even
when W is still empty), then continue to the next pass of the loop.
If the above loop ended naturally (i.e., not from the break ) after n + T full
passes, we failed.
Otherwise, we succeeded. Add an nth vector, wn1 , which is linearly independent to this set (and therefore not orthogonal to a, by a previous exercise), done
easily using a simple classical observation, demonstrated below. This produces
a system of n independent equations satisfying
(
0, k = 0, . . . , n 2
wk a =
1, k = n 1
which has a unique non-zero solution.
Use a classical algorithm to solve the systems of n equations for a.
Note: By supplying the nth vector (a fast, easy addition of cost O(n)), we get a full
system requiring no extra quantum samples and guaranteeing that our system yields
a, unequivocally.
Time and Space Complexity
We will run our circuit n + T = O(n) times. There will be some classical algorithms
that need to be tacked on, producing an overall growth rate of about O(n3 ) or O(n4 ),
depending on the cleverness of the overall design. But lets back-up a moment.
Ive only mentioned time complexity. What about circuit, a.k.a. spatial, complexity?
There is a circuit that must be built, and we can tell by looking at the number
of inputs and outputs to the circuit (2n each) that the spatial complexity is O(n).
I mentioned during our oracle lesson that we can ignore the internal design of the
oracle because it, like its associated f , may be arbitrarily complex; all we are doing in
these algorithms is getting a relativized speed-up. But even setting aside the oracles
internals, we cant ignore the O(n) spatial input/output size leading to/from the
oracle as well as the circuit as a whole.
Therefore, if I operate the circuit in an algorithmic while loop of n + T = O(n)
passes, the time complexity of the quantum circuit (not counting the classical tools
to be added) is O(n). Meanwhile, each pass of the circuit uses O(n) wires and gates,
giving a more honest O(n2 ) growth to the { algorithm + circuit } representing (only)
the quantum portion of the algorithm.
So, why do I (and others) describe the complexity of the quantum portion of the
algorithm to be O(n) and not O(n2 )?
509
We actually touched on these reasons when describing why we tend to ignore
hardware growth for these relatively simple circuits. Ill reprise the two reasons given
at that time.
Our interest is in relative speed-up of quantum over classical methods. Since
we need O(n) inputs/outputs in both the classical and quantum circuit, it is
unnecessary to carry the same cost of O(n) on the books it cancels out when
we compare the two regimes.
The quantum circuits grow slowly (linearly or logarithmically) compared to
more expensive classical post processing whose time complexity is often at least
quadratic and (this is important) in series with the quantum circuit. These
attending sub-algorithms therefore render an O(n) or O(n2 ) quantum growth
invisible.
So, be aware of the true overall complexity, but understand that its usually unnecessary to multiply by a linear circuit complexity when doing computational accounting.
18.9.3
Strange Behavior
This section is not required but may add insight and concreteness to your understanding of the concepts.
[Notational Alert. Ill use vector notation (a, wk ) rather than encoded decimal
(a, wk ) to underscore the otherwise subtle concepts.]
We know that (Z2 )n is not your grandparents vector space; it has a dot-product
that is not positive definite (e.g., a a might = 0), which leads to screwy things.
Nothing that we cant handle, but we have to be careful.
Well be constructing a basis for (Z2 )n in which the first n1 vectors, w0 , . . . , wn2
are orthogonal to a, while the nth vector, wn1 , is not. In general, this basis will not
be and cannot be an orthonormal basis. It is sometimes, but not usually. To see
cases where it cannot be, consider a period a that is orthogonal to itself. In such a
case, since a a we might have a = wk for some k < n 1, say a = w0 . Therefore,
we end up with
w0 w0
w0 wn1
=
=
aa = 0
a wn1 =
and
1,
either condition contradicting orthonormality.
In particular, we cannot use the dot-with-basis trick to compute expansion coefficients, a trick that we saw required an orthonormal basis to work. So, in this regime,
even when expanding a vector v along the algorithms resulting w-basis,
v
n1
X
k=0
510
ck w k ,
we would (sadly) discover that, for many of the coefficients, ck ,
ck
6=
v wk .
This is not a big deal, since we will not need the trick, but its a good mental exercise
to acknowledge naturally occurring vector spaces that give rise to non-positive definite pairings and be careful not to use those pairing as if they possessed our usual
properties.
Example #1
Take n = 4 and
0
1
=
0 .
1
We may end up with a basis that uses the 3-dimensional subspace, a , generated by
the three vectors
1
0
0
0
0
1
, , = a ,
0 1 0
0
0
1
augmented by the nth vector not orthogonal to a,
0
1
wn1 = w3 =
0 .
0
These four vectors form a possible outcome of our algorithm when applied to (Z2 )4
with a period of a = ( 0, 1, 0, 1 )t , and you can confirm the odd claims I made above.
Example #2
If youd like more practice with this, try the self-orthogonal
1
1
a =
1
1
511
and the basis consisting of the three
1
1 ,
0
vectors orthogonal to a,
1
1
0 1
, = a
0 1
1
1
plus a fourth basis vector not orthogonal (to a)
wn1
w3
1
0
.
0
0
This isnt an othonormal basis, nor will the dot-with-basis trick work.
[Exercise. Create some of your own examples in which you do, and do not, get
an orthonormal basis that results from the algorithms desired outcome of n 1 basis
vectors being orthogonal to a and the nth, not.]
18.10
Time Complexity of the Quantum Algorithm
18.10.1
Producing n 1 Linearly-Independent wk in Polynomial Time Argument 1
This section contains our officially sanctioned proof that one will get the desired n 1
vectors, orthogonal to a, with polynomial complexity in n. It is a very straightforward
argument that considers sampling the circuit n + T times. While it has several
steps, each one is comprehensible and contains the kind of arithmetic every quantum
computer scientist should be able to reproduce.
Theorem. If we randomly select m + T samples from Z2m , the probability
that these samples will contain a linearly-independent subset of m vectors
is > 1 (1/2)T .
Notice that the constant T is independent of m, so the process of selecting the
m independent vectors is O(m + T ) = O(m), not counting any sub-algorithms or
arithmetic we have to apply in the process (which well get to). Well be using m =
n 1, the dimension of a (
= Z2n1 ).
18.10.2
Proof of Theorem Used by Argument 1
[Notational Alert. In the proof, Ill use vector notation, as in z (Z2 )m or c
(Z2 )m+T , which presents vectors in boldface.]
Pick m + T vectors, { z0 , z1 , z2 , . . . zm+T 1 }, at random from (Z2 )m .
512
Step 1: Form a matrix and look at the column vectors. We begin by
stacking the m + T vectors atop one another, forming the matrix,
z0
z00
z01
z0(m1)
z1
z10
z11
z1(m1)
z2
z20
z
21
2(m1)
=
.
..
..
..
..
.
.
.
.
.
.
.
zm+T 1
z(m+T 1)0 z(m+T 1)1 z(m+T 1)(m1)
The number of independent rows is the row rank of this matrix, and by elementary linear algebra, the row rank = column rank. So, lets change our perspective
and think of this matrix as set of m column vectors, each of dimension m + T .
We would be done if we could show that all m of column vectors
z00
z01
z0(m1)
z10 z11
z1(m1)
z20 z21
z2(m1)
,
,...,
..
..
..
.
.
.
z(m+T 1)0
z(m+T 1)1
z(m+T 1)(m1)
c0 ,
c1 ,
cm1
were linearly independent with probability > 1 (1/2)T +1 . (That would mean
the column rank was m.)
[This row rank = column rank trick has other applications in quantum computing, and, in particular, will be used when we study Neumarks construction for
orthogonalizing a set of general measurements in the next course. Neumarks
construction is a conceptual first step towards noisy-system analysis.]
Step 2: Express the probability that all m column vectors, ck , are
independent as a product of m conditional probabilities.
[Note: This is why we switched to column rank. The row vectors, we know, are
not linearly independent after taking T + m > m samples, but we can sample
many more than m samples and continue to look at the increasingly longer
column vectors, eventually making those linearly independent.]
Let
I (j) event that c0 , c1 , . . . cj1 are linearly-independent.
Our goal is to compute the probability of I (m). Combining the basic identity,
P I (m)
= P I (m) I (m 1)
+ P I (m) I (m 1) ,
513
with the observation that
P I (m) I (m 1)
we can write
P I (m)
= P I (m) I (m 1)
= P I (m) I (m 1) P I (m 1) ,
the last equality from Bayes rule.
Now, apply that same process to the right-most factor, successively, and you
end up with
P I (m)
= P I (m) I (m 1) P I (m 1) I (m 2)
P I (2) I (1) P I (1)
m
Y
=
P I (j) I (j 1) .
j=1
(The j = 1 term might look a little strange, because it refers to the undefined
I (0), but it makes sense if we view the event I (0), i.e., that in a set of no
vectors theyre all linearly-independent, as vacuously true. Therefore, we see
that I (0) can be said to have probability 1, so the j = 1 term reduces to
P ( I (1) ), without a conditional, in agreement with the line above.)
Step 3: Compute the j th factor, P I (j) I (j 1) , in this product.
P I (j) I (j 1)
= 1 P I (j) I (j 1)
But what does I (j) I (j 1) mean? It means two things:
1. It assumes the first j 1 vectors, {c0 , . . . , cj2 }, are independent and
therefore span the largest space possible for j 1 vectors, namely, one of
size 2j1 , and
2. The jth vector, cj1 , is the span of the first j 1 vectors {c0 , . . . , cj2 }.
This probability is computed by counting the number of ways we can select a
vector from the entire space of 2m+T vectors (remember, our column vectors
have m + T coordinates) that also happen to be in the span of the first j 1
vectors (a subspace of size 2j1 ).
[Why size 2j1 ? Express the first j 1 column vectors in their own coordinates, i.e., in a basis that starts with them, then adds more basis vectors to get
the full basis for (Z2 )m+T . Recall that basis vectors expanded along themselves
514
always have a single 1 coordinate sitting in a column of 0s. Looked at this way,
how many distinct vectors can be formed out of various sums of the original
j 1?]
The probability we seek is, by definition of probability, just the ratio
2j1
2m+T
m+T j+1
1
,
2
so
P I (j) I (j 1)
m+T j+1
1
,
2
for j = 1, 2, . . . , m.
Sanity Check. Now is a good time for the computer scientists first line of
defense when coming across a messy formula: the sanity check.
Does this make sense for j = 1, the case of a single vector c0 ? The formula tells
us that the chances of getting a single, linearly-independent vector is
P I (1)
m+T 1+1
1
2
m+T
1
.
2
Wait, shouldnt the first vector be 100% certain? No, we might get unlucky and
pick the 0-vector with probability 1/2m+T , which is exactly what the formula
predicts.
That was too easy, so lets do one more. What about j = 2? The first vector,
c0 , spans a space of two vectors (as do all non-zero single vectors in (Z2 )m ).
The chances of picking a second vector from this set would be 2/(size of the
space), which is 2/2m+T = 1/(2m+T 1 ). The formula predicts that we will get
a second independent vector, not in the span of c0 with probability
P I (2) I (1)
m+T 2+1
1
2
m+T 1
1
,
2
exactly the complement of the probability that c1 just happening to get pulled
from the set of two vectors spanned by c0 , as computed.
So, spot testing supports the correctness of the derived formula.
Step 4: Plug the expression for the jth factor (step 3) back into the
full probability formula for all m vectors (step 2).
In step 2, we decomposed the probability of success, as a product,
P I (m)
m
Y
j=1
515
P I (j) I (j 1) .
In step 3, we computed the value for each term,
P I (j) I (j 1)
m+T j+1
1
.
2
Combining the two, we get
P I (m)
m
Y
m+T j+1 !
1
,
2
j=1
which can easily be re-indexed into the form
P I (m)
m
Y
T +i !
1
.
2
i=1
[Exercise. Prove the last assertion about re-indexing.]
Step 5: We prove a cute mathematical lemma. Lets take a pleasant
off-trail walk to prove a needed fact.
Lemma.
0 i 1
p
Y
then
(1 i )
If
i=1
p
X
i .
i=1
Proof by Induction.
Case p = 1:
1 1
1 1 .
Consider any p > 1 and assume the claim is true for p 1. Then
p
Y
(1 i )
(1 p )
i=1
p1
Y
(1 i )
i=1
(1 p ) 1
1
p1
X
!
i
i=1
p1
i p +
i=1
p
>
p1
X
p i
i=1
i .
X QED
i=1
[Exercise. Where did we use the hypothesis
i 0? Same question, i 1.]
P
[Exercise. Is the stronger hypothesis
i > 0 needed for this lemma? Why
or why not?
516
Step 6: Apply the lemma to the conclusion of step 4 to finish off the
proof. Using m for the p of the lemma, we obtain
T +i !
m
Y
1
P I (m)
=
1
2
i=1
m T +i
X
1
1
2
i=1
#
T "X
m i
1
1
= 1
.
2
2
i=1
But that big bracketed sum on the RHS is a bunch of distinct and positive
powers of 1/2, which can never add up to more than 1 (think binary floating
point numbers like .101011 or .00111 or .1111111), so that sum is < 1, i.e.,
" m #
X 1 i
< 1 , so
2
i=1
#
T
T "X
m i
1
1
1
, and we get
<
2
2
2
i=1
#
T "X
T
m i
1
1
1
1
> 1
.
2
2
2
i=1
Combining the results of the last two equation blocks we conclude
P I (m)
>
T
1
1
.
2
This proves that the column vectors, cj , are linearly-independent with probability greater than 1 1/2T , and therefore the row vectors, our zk also have
at least m linearly independent vectors among them (row rank = column rank,
remember?), with that same lower-bound probability.
QED
18.10.3
Summary of Argument 1
We have demonstrated that by sampling m + T vectors in a Z2 space of dimension
m, we can be confident that the sample set contains m linearly independent vectors
with probability 1 (1/2)T , T independent of m.
In Simons problem, we were looking for n 1 vectors that spanned the n 1dimensional a . By applying the above result to m = n 1, we see that we can find
those n 1 vectors with arbitrarily high probability by running our circuit O(T + n
1) = O(n) times.
517
18.10.4
Producing n 1 Linearly-Independent wk in Polynomial Time Argument 2
This argument is seen frequently and is more straightforward than our preferred one,
but it gives a weaker result in absolute terms. That is, it gives an unjustifiably
conservative projection for the number of samples required to achieve n 1 linearly
independent vectors. Of course, this would not affect the performance of an actual
algorithm, since all we are doing in these proofs is showing that well get linear
independence fast. An actual quantum circuit would be indifferent to how quickly
we think it should give us n 1 independent vectors; it would reveal them in a time
frame set by the laws of nature, not what we proved or didnt prove. Still, its nice to
predict the convergence to linear independence accurately, which this version doesnt
do quite as well as the first. Due to its simplicity and prevalence in the literature, I
include it.
Here are the steps. Well refer back to the first proof when we need a result that
was already proven there.
Theorem. If we randomly select m samples from Z2m , the probability that
we have selected a complete, linearly-independent (and therefore basis) set
is > 1/4.
This result (once proved) estimates a probability of at least 1 1/4T of getting
a linearly independent set of vectors after sampling Z2m mT times. The reason we
have to take take the product, mT , is that the theorem only computes the probability
that results when we take exactly m samples; it does not address the trickier math
for overlapping sample sets or a slowly changing sample set that would come from
adding one new sample and throwing away an old one. Nevertheless, it proves O(m)
complexity.
Keep in mind that well be applying this theorem to m = n 1, the dimension of
a (
= Z2n1 ).
18.10.5
Proof of Theorem Used by Argument 2
[Notation. As with the first proof, well use boldface vector notation, z (Z2 )m .]
Pick m vectors, { z0 , z1 , z2 , . . . zm1 }, at random from (Z2 )m .
Step 1: Express the probability that the m vectors, zk , are independent as a product of m conditional probabilities.
Let
I (j) event that z0 , z1 , . . . zj1 are linearly-independent.
Our goal is to compute the probability of I (m).
518
Using the exact argument from argument 1, step 2 , we conclude,
P I (m)
m
Y
P I (j) I (j 1) .
j=1
(If interested, see argument 1, step 2 to account for the fact that I (0)
can be said to have probability 1, implying that the j = 1 term reduces to
P ( I (1) ), without a conditional.)
Step 2: Compute the j th factor, P I (j) I (j 1) , in this product.
P I (j) I (j 1)
= 1 P I (j) I (j 1)
But what does I (j) I (j 1) mean? It means two things:
1. It assumes the first j 1 vectors, {z0 , . . . , zj2 }, are independent and
therefore span the largest space possible for j 1 vectors, namely, one of
size 2j1 , and
2. The jth vector, zj1 , is the span of the first j 1 vectors {z0 , . . . , zj2 }.
This probability is computed by counting the number of ways we can select a
vector from the entire space of 2m vectors which happens to also be in the span
of the first j 1 vectors, a subspace of size 2j1 . But thats just the ratio
mj+1
1
2j1
=
,
m
2
2
so
P I (j) I (j 1)
mj+1
1
,
2
for j = 1, 2, . . . , m.
Sanity Check. Does this make sense for j = 1, the case of a single vector z0 ?
The formula tells us that the chances of getting a single, linearly-independent
vector is
m1+1
m
1
1
= 1
.
P I (1)
= 1
2
2
Wait, shouldnt the first vector be 100% certain? No, we might get unlucky
and pick the 0-vector with probability 1/2m , which is exactly what the formula
predicts.
That was too easy, so lets do one more. What about j = 2? The first vector,
z0 , spans a space of two vectors (as do all non-zero single vectors in (Z2 )m ).
The chances of picking a second vector from this set would be 2/(size of the
519
space), which is 2/2m = 1/(2m1 ). The formula predicts that we will get a
second independent vector, not in the span of z0 with probability
m2+1
m1
1
1
= 1
,
P I (2) I (1)
= 1
2
2
exactly the complement of the probability that z1 just happening to get pulled
from the set of two vectors spanned by z0 , as computed.
So, spot testing supports the formulas claim.
Step 3: Plug the expression for the jth factor (step 2) back into the
full probability formula for all m vectors (step 1).
In step 1, we decomposed the probability of success, as a product,
P I (m)
m
Y
P I (j) I (j 1) .
j=1
In step 2, we computed the value for each term,
P I (j) I (j 1)
mj+1
1
.
2
Combining the two, we get
P I (m)
m
Y
j=1
mj+1 !
1
,
2
which can easily be re-indexed into the form
P I (m)
m
Y
i=1
i !
1
.
2
[Exercise. Prove the last assertion about re-indexing.]
Step 4: Apply a result from q-Series to finish off the proof.
The expression we have can be multiplied by values < 1 if we are interested in a
lower bound (which we are). Therefore, we can include all the factors for i > m
without harm:
i !
m
Y
1
P I (m)
=
1
2
i=1
i !
Y
1
1
2
i=1
From here one could just quote a result from the theory of mathematical q-series,
namely that this infinite product is about .28879. As an alternative, there are
520
some elementary proofs that involve taking the natural log of the product,
splitting it into a finite sum plus and infinite error sum, then estimating the
error. Well accept the result without further ado, which implies
i !
m
Y
1
P I (m)
=
1
2
i=1
>
.25
This proves that the m vectors, zj , are linearly-independent with probability
> 1/4.
18.10.6
Summary of Argument 2
We have demonstrated that a random sample of exactly m vectors in a Z2 space of
dimension m will be linearly independent with probability at least 1/4.
QED
Corollary. After T full cycles, each cycle pulling m random samples from Z2m
(not reusing vectors from previous cycles), the probability that none of the T sets of
m-vectors is linearly independent is < (3/4)T .
Proof. This follows immediately from the theorem since the chances of failure
for selecting one set of m independent vectors = 1 - chance of success < 3/4, as the
theorem showed. Each cycle is independent so the probability that all of them fail is
the product of the probabilities that each one fails, and therefore < (3/4)T .
QED
In Simons problem, we were looking for n 1 vectors that spanned the n 1dimensional a . By applying the above result to m = n 1, we see that we can find
those n 1 vectors with arbitrarily high probability by running our circuit O(T (n
1)) = O(n) times.
18.10.7
Discussion of the Two Proofs Complexity Estimates
Both proofs give the correct O(n) complexity (of the algorithms quantum processing)
for finding n 1 independent vectors spanning a in the context of Z2n .
The second proof is appealing in its simplicity, but part of that simplicity is due
to its handing off a key result to the number theorists. (I have found no sources that
give all the details of the > 1/4 step). Also, the number of circuit samples argument
2 requires for a given level of confidence is, while still O(n), many times larger than
one really needs, and, in particular, many more than the first proof. This is because
it does not provide probabilities for overlapping sample sets, but rather tosses out
all n 1 samples of any set that fails. One can adjust the argument to account for
this conditional dependence, but thats a different argument; if we know that one
set is not linearly independent, then reusing its vectors requires trickier math than
this proof covers. Of course, this is mainly because it is only meant to be a proof of
polynomial time complexity, and not a blueprint for implementation.
521
For example, for n = 10, T = 10, the second proof predicts that 10 9 = 90
samples would produce at least one of the 10 sets to be linearly independent with
probability greater than 1 (3/4)10 .943686. In contrast, the first proof would only
ask for 9 + 10 = 19 samples to get confidence > .999023. Thats greater confidence
with fewer samples.
18.11
The Hidden Classical Algorithms and Their
Cost
18.11.1
Unaccounted for Steps
As far as we have come, we cant claim victory yet, especially after having set the
rather high bar of proving, not just uttering, the key facts that lead to our end result.
We have a quantum circuit with O(n) gates which we activate O(n) times to find
the unknown period, a, with arbitrarily high probability. Weve agreed to consider
the time needed to operate that portion of the quantum algorithm and ignore the
circuits linear growth. Therefore, we have thus far accounted for a while loop which
requires O(n) passes to get a.
However, there are steps in the sampling process where we have implicitly used
some non-trivial classical algorithms that our classical computers must execute. We
need to see where they fit into the big picture and incorporate their costs. There are
two general areas:
1. The test for mod-2 linear independence in (Z2 )n that our iterative process has
used throughout.
2. The cost of solving the system of n mod-2 equations,
(
0, k = 0, . . . , n 2
wk a =
1, k = n 1
For those of you who will be skipping the details in the next sections, Ill reveal
the results now:
1. The test for mod-2 linear independence is handled using mod-2 Gaussian elimination which we will show to be O(n3 ).
2. Solving the system of n mod-2 equations is handled using back substitution
which we will show to be O(n2 ).
3. The two classical algorithms will be applied in series, so we only need to count
the larger of the two, O(n3 ). Together, they are applied once for each quantum
sample, already computed to be O(n), resulting in a nested count of O(n4 ).
522
4. Well tweak the classical tools by integrating them into Simons algorithm so
that their combined cost is only O(n2 ), resulting in a final count of O(n3 ).
Conclusion. Our implementation of Simons algorithm has a growth rate of
O(n3 ). It is polynomial fast.
18.12
Solving Systems of Mod-2 Equations
18.12.1
Gaussian Elimination and Back Substitution
Conveniently, both remaining classical tasks are addressed by an age old and well
documented technique in linear algebra for solving systems of linear equations called
Gaussian elimination with back substitution. Its a mouthful but easy to learn and,
as with all the ancillary math we have been forced to cover, applicable throughout
engineering. In short, a system of linear equations
5x + y + 2z + w
x y + z + 2w
x + 2y 3z + 7w
=
=
=
7
10
3
is symbolized by the matrix of its constant coefficients, as in
x
5 1
2 1
7
1 1 1 2 y = 10
z
1 2 3 7
3
w
As the example shows, theres no requirement that the system have the same number
of equations as unknowns; the fewer equations, the less you will know about the solutions. (Instead of the solution being a unique vector like (x, y, z, w)t = (3, 0, 7, 2)t ,
it might be a relation between the components, like ( , 4, .5, 3 )t , with free
to roam over R). Nevertheless, we can apply our techniques to any sized system.
We break it into the two parts,
Gaussian elimination, which produces a matrix with 0s in the lower left triangle,
and
back substitution, which uses that matrix to solve the system of equations as
best we can, meaning that if there are not enough equations, we might only get
relations between unknowns, rather than unique numbers.
523
18.12.2
Gaussian Elimination
Gaussian Elimination for Decimal Matrices
Gaussian Elimination seeks to change the matrix for the equation into (depending
on who you read), either
row echelon form, in which everything below the diagonal is 0, as in
3 3 0 5
0 6 2 4 , or
0 0 5 7
reduced row echelon form, which is echelon form with the additional requirement that the first non-zero element in each row be 1, e.g.,
1 1 0
5/3
0 1 1/3 2/3 .
0 0 1 7/5
In our case, where all the values are integers mod-2 (just 0 and 1), the two are actually
equivalent: all non-zero values are 1, automatically.
Properties of Echelon Forms
Reduced or not, row echelon forms have some important properties that we will need.
Lets first list them, then have a peek at how one uses Gaussian elimination (GE), to
convert any matrix to an echelon form.
Geometrically, the diagonal, under which
a square matrix:
0 0
.. ..
. .
0 0
all the elements must be 0, is clear in
..
.
0
..
.
When the the matrix is not square, the diagonal is geometrically visualized
524
relative to the upper left (position (0, 0)):
0
.
.
.
0
.
..
0
0
..
.
..
.
0
0
..
.
0
0
..
.
..
0
..
.
0
..
.
0
0
..
.
..
.
..
.
0
..
.
In any case, a diagonal element is one that sits on position (k, k), for some k.
The first non-zero element on row k is to the right of the first non-zero element
of row k1, but it might be two or more positions to the right. Ill use a reduced
form which has the special value, 1, occupying the first non-zero element in a
row to demonstrate this. Note the extra 0s, underlined, that come about as a
result of some row being pushed to the right in this way.
0 1
0 0 0 0 1
0 0 0 0 0 1
.. .. .. .. .. .. . .
..
. . . . . .
.
.
0 0 0 0 0 0 [0/1/]
Any all-zero rows necessarily appear at the bottom of the matrix.
[Exercise. Show this follows from the definition.]
All non-zero row vectors in the matrix are, collectively, a linear independent
set.
[Exercise. Prove it.]
If we know that there are no all-zero row vectors in the echelon form, then the
number of rows number of columns.
[Exercise. Prove it.]
Including the RHS Constant Vector in the Gaussian Elimination Process
When using GE to solve systems of equations, we have to be careful that the equations
that the reduced echelon form represents are equivalent to the original equations, and
to that end we have to modify the RHS column vector, e.g., (7, 10, 3)t of our
example, as we act on the matrix on the LHS. We thus start the festivities by placing
525
the RHS constant vector in the same house as the LHS matrix, but in a room of
its own,
5 1
2 1 7
1 1 1 2 10 .
1 2 3 7 3
We will modify both the matrix and the vector at the same time, eventually resulting
in the row-echelon form,
3 3 0 5 13
0 6 2 4 832 ,
15
0 0 5 7 17
5
or, if we went further, in the reduced row echelon form
1 1 0 35
13
3
0 1 1 2 832 .
3
3
45
0 0 1 75 17
25
The Three Operations that Produce Echelon Forms
There are only three legal operations that we need to consider when performing GE,
1. swapping two rows,
2. multiplying a row by a nonzero value, and
3. adding a multiple of one one row to another.
[Exercise. Prove that these operations produce equations that have the identical
solution(s) as the original equations.]
526
For example,
5 1
2 1 7
7
5 1 2
1
52nd row
1 1 1 2 10 5 5 5 10 50
3
1 2 3 7 3
1 2 3 7
5 1 2
1
7
0 6 3 9 43
3
1 2 3 7
add 1st to 2nd
3
1
2
3
7
swap 1st and 3rd
0 6 3 9 43
7
5 1 2
1
3
1
2
3
7
add 51st to 3rd
0 6 3 9 43
0 9 17 34 22
1 2 3 7
3
add 32 2nd to 3rd
0 6 3 9 43 ,
0 0 25
95
85
2
2
2
etc. (These particular operations may not lead to the echelon forms, above; theyre
just illustrations of the three rules.)
The Cost of Decimal-Based Gaussian Elimination
GE is firmly established in the literature, so for those among you who are interested,
Ill prescribe web search to dig up the exact sequence of operations needed to produce
a row reduced echelon form. The simplest algorithms with no short-cuts use O(n3 )
operations, where an operation is either addition or multiplication, and n is the
larger of the matrixs two dimensions. Some special techniques can improve that, but
it is always worse than O(n2 ), so well be satisfied with the simpler O(n3 ).
To that result we must incorporate the cost of each multiplication and addition
operation. For GE, multiplication could involve increasingly large numbers and, if
incorporated into the
full accounting, would change the complexity to slightly better
2
3
than O n (log m) , where m is the absolute value of the largest integer involved.
Addition is less costly and done in series with the multiplications so does not erode
performance further.
The Cost of Mod-2 Gaussian Elimination
For mod-2 arithmetic, however, we can express the complexity without the extra variable, m. Our matrices consist of only 0s and 1s, so the multiplication in operations
2 and 3 reduce to either the identity (1 a row) or producing a row of 0s (0 a
row). Therefore, we ignore the multiplicative cost completely. Likewise, the addition
527
operation that is counted in the general GE algorithm is between matrix elements,
each of which might be arbitrarily large. But for us, all matrix elements are either 0
or 1, so our addition is a constant time XOR between bits.
All this means that in the mod-2 milieu, each of our n3 GE operations requires
some constant-time XORs (for the additions) and constant time if -statements (to
account for the multiplication by 0 or 1). Evidently, the a total mod-2 GE cost is
untarnished at O(n3 ).
With some fancy footwork in a special section, below, well improve it to O(n2 ),
but we dont really care about the exact figure once we have it down this far. All we
require is polynomial time to make Simon a success story. However, we are exercising
our ability to evaluate algorithms, historical or future, so this continues to be an
exercise worthy of our efforts.
18.12.3
Back Substitution
Back Substitution for Decimal Matrices
It is easiest and for us, enough to explain back substitution in the case when the
number of linearly independent equations equals the number of unknowns, n. Lets
say we begin with the system of equations in matrix form,
c00
c01
...
c0(n1)
b0
x0
c10
b1
c11
...
c1(n1)
x1
=
..
.. ,
..
.
.
.
.
.
.
.
.
.
.
.
.
c(n1)0 c(n1)1 . . . c(n1)(n1)
assumed to be of maximal rank, n
This would result is a reduced row
1 c001 c002 c003
0 1 c012 c013
0 0
1 c023
0 0
1
. .
.
..
..
.. ..
.
0 0
0
0
0
bn1
xn1
all rows (or columns) are linearly independent.
echelon form
c00(n1)
. . . c00(n2)
b00
. . . c01(n2)
c01(n1)
b01
. . . c02(n2)
c02(n1)
b02
. . . c03(n2)
c03(n1)
b03
,
..
..
..
..
.
.
.
.
...
1
c0(n2)(n1) b0n2
...
0
1
b0n1
where, the c0kj and b0k are not the original constants in the equation, but the ones
obtained after applying GE. In reduced echelon form, we see that the (n 1)st
unknown, xn1 , can be read off immediately, as
xn1
b0n1 .
From here, we substitute back into the (n 2)nd equation to get,
xn2 + b0n1 c0(n2)(n1) =
528
b0n2 ,
which can be solved for xn2 (one equation, one unknown). Once solved, we substitute
these numbers into the equation above, getting another equation with one unknown.
This continues until all rows display the answer to its corresponding xk , and the
system is solved.
The Cost of Decimal-Based Back Substitution
The bottom row has no operations: it is already solved. The second-from-bottom has
one multiplication and one addition. The third-from-bottom has two multiplications
and two additions. Continuing in this manner and adding things up, we get
1 + 2 + 3 + ... + (n 1) =
(n 1) n
2
additions and the same number of multiplications, producing an overall complexity of
O(n2 ) operations. As noted, the time complexity of the addition and multiplication
algorithms would degrade by a factor of (log m)2 , m being the largest number to be
multiplied, making the overall bit complexity O n2 (log m)2 .
The Cost of Mod-2 Back Substitution
For mod-2 systems, we have no actual multiplication (the b0k are either 1 and 0) and
the additions are single-bit mod-2 additions and therefore, constant time. Thus, each
of the approximately n(n + 1)/2 (= O(n2 ) ) addition operations in back substitution
uses a constant time addition, leaving the total cost of back substitution un-degraded
at O(n2 ).
18.12.4
The Total Cost of the Classical Techniques for Solving Mod-2 Systems
We have shown that Gaussian elimination and back substitution in the mod-2 environment have time complexities, O(n3 ) and O(n2 ), respectively. To solve a system of
n mod-2 equation the two methods can be executed in series, so the dominant O(n3 )
will cover the entire expense.
However, this isnt exactly how Simons algorithm uses these classical tools, so we
need to count in a way that precisely fits our needs.
18.13
Applying GE and Back Substitution to Simons Problem
We now show how these time tested techniques can be used to evaluate the classical
post processing costs in Simons algorithm. The eventual answer we will get is this:
529
They will be used in-series, so we only need to count the larger of the two, O(n3 ),
and these algorithms are applied once for each quantum sample, already computed
to be O(n), resulting in a nested count of O(n4 ).
18.13.1
Linear Independence
Gaussian elimination without back substitution can be used to determine whether
a vector z is linearly independent of a set W = {wk } of m vectors which is, itself,
known to be linearly independent. Therefore, we will apply GE after taking each new
sample z, and if we find z to be independent of the existing W, well add it to W
(increasing its size by one) and go on to get our next z. If z turns out to not be
independent of W, well throw it away and get another z without changing W. This
process continues until either W contains the maximum n 1 vectors (the most that
can be orthogonal to the unknown period a) or we have exceeded the n + T sampling
limit.
The following process demonstrates one way to apply GE to determine linearly
independence and build an increasingly larger set of independent vectors in W.
1. Stack the wk on top of one another to form an m n matrix, which we assume
to already be in reduced echelon form (our construction will guarantee it):
1 w01 w02 w03 . . . w0(n3)
w0(n2)
w0(n1)
0 0
1 w13 . . . w1(n3)
w1(n2)
w1(n1)
0 0
0
1
.
.
.
w
w
w
2(n3)
2(n2)
2(n1)
.
..
..
..
.. . .
..
..
..
.
.
.
.
.
.
.
.
0 0
0
0 ...
1
w(m1)(n2) w(m1)(n1)
2. Observations. Notice that m < (n 1), since the vectors in W are independent, by assumption, and if m were equal to n 1, we would already have a
maximally independent set of vectors known to be orthogonal to a and would
have stopped sampling the circuit. That means that there are at least two more
columns than there are rows: the full space is n-dimensional, and we have n 2
or fewer linearly independent vectors so far.
As a consequence, one or more rows (the second, in the above example) skips
to the right more than one position relative to the row above it and/or the final
row in the matrix has its leading 1 in column n 3 or greater.
3. Put z at the bottom of the stack and re-apply GE.
If z is linearly independent of W this will result in a new non-zero
vector row. Replace the set W with the new set whose coordinates are the
rows of the new reduced-echelon matrix. These new row vectors are in the
span of the old W plus z added. [Caution. It is possible that none of
the original wk or z explicitly appear in the new rows, which come from
530
GE applied to those vectors. All that matters is that the span of the new
rows is the same as the span of W plus z, which GE ensures.] We have
increased our set of linearly independent vectors by one.
If z is not linearly independent of W the last row will contain all 0s.
Recover the original W (or, if you like, replace it with the new reduced
matrix row vectors, leaving off the final 0 row the two sets will span the
same space and be linearly independent). You are ready to grab another z
based on the outer-loop inside which this linear-independence test resides.
[Exercise. Explain why all the claims in this step are true.]
4. Once n 1 vectors populate the set W =
complete. We call the associated row-reduced
w0
w1
W =
..
{w0 , . . . , wn2 }, the process is
matrix W ,
wn2
and W satisfies the matrix-vector product equation
W a
0,
where a is our unknown period and 0 is the (n 1)-dimensional 0-vector in
(Z2 )n1 . This condenses our system of n 1 equations, a wk = 0, into a single
matrix relation.
Cost of Using GE to Determine Linear Independence
In summary, we have employed the classical algorithm GE to handle our linearindependence test. GE is O(n3 ) and is applied once for each pass of our quantum
circuit. Since that circuit was O(n), the new total cost (so far) for Simons algorithm
is O(n4 ). (Well do better.)
18.13.2
Completing the Basis with an nth Vector Not Orthogonal to a
Once we are finished assembling the set of n 1 vectors, W = {w0 , . . . wn2 }, we
will add a final nth vector, wn1 , to the set a vector which is not orthogonal to a
(remember, a might not be such a vector it will be orthogonal to itself exactly half
the time in our strange mod-2 dot product). wn1 can be quickly determined and
inserted into the correct position of the reduced W ( ' W) matrix as follows.
1. Starting from the top row, w0 , look for the last (lowest) wk which has its leading
1 in column k (i.e., it has its leading 1 on W s diagonal, but wk+1 , directly below
it, has a 0 in its (k + 1)st position).
531
If such a wk can be found, define
2n2k
wn1
(0, 0, . . . , 0, 1, 0, . . . , 0)
k + 1st position from left
and place this new wn1 directly below wk , pushing all the vectors in the
old rows k + 1 and greater down to accommodate the insertion. Call the
augmented matrix, W 0 .
Before:
1 w01 w02 w03 w04 w05
0 1 w12 w13 w14 w15
wk
0
0
1
w
w
w
W =
23
24
25
0 0
0
0
1 w35
0 0
0
0
0
1
After:
W0
1 w01 w02 w03 w04
0 1 w12 w13 w14
0 0
1 w23 w24
0 0
0
1
0
0 0
0
0
1
0 0
0
0
0
w05
w15
w25
0
wn1
w35
1
If there is no such wk , then either
W has all 1s on its diagonal, or
W has all 0s on its diagonal.
If W has all 1s on its diagonal, the final row of W is of the form
wn2
(0, 0, . . . , 0, 1, ) .
Define
wn1
20
(0, 0, . . . , 0, 0, 1) ,
and place this new wn1 after wn2 last old row, making wn1 the new
bottom row of W . Call the augmented matrix, W 0 .
Before:
1 w01 w02 w03 w04 w05
0 1 w12 w13 w14 w15
1 w23 w24 w25
W =
0 0
0 0
0
1 w34 w35
0 0
0
0
1 w45 wn2
532
After:
1 w01 w02 w03
0 1 w12 w13
0 0
1 w23
0 0
0
1
0 0
0
0
0 0
0
0
w04
w14
w24
w34
1
0
w05
w15
w25
w35
w45
wn1
1
Exercise. If W has all 0s on its diagonal, explain what to do, showing
example matrices.
[Exercise. Prove all outstanding claims in step 1.]
2. Into the (n 1)-dimensional 0 vector comprising the RHS of the equation
W a
0,
insert a 1 into the position corresponding to the new row in W . Push any 0s
down, as needed, to accommodate this 1. It will now be an n-dimensional vector
corresponding to 2k for some k.
0
0
0
0
0 1
0
0
0
0
0
That will produce a full set of n linearly independent vectors for Z2n in reduced
echelon form, which we call W 0 .
Cost of Completing the Basis
The above process consists of a small number of O(n) operations, each occurring in
series with each other and with the loops that come before. Therefore, its O(n) adds
nothing to the previous complexity, which now stands at O(n4 ).
533
18.13.3
Using Back-Substitution to Close the Deal
We are, metaphorically, 99% of the way to finding the period of f . We want to solve
0
0
.
.
.
W a = 1 ,
.
..
0
0
but W is already in reduced-echelon form. We need only apply mod-2 back-substitution
to extract the solution vector, a.
Cost of Back Substitution to the Algorithm
This is an O(n2 ) activity done in series with the loops that come before. Therefore,
its O(n2 ) adds nothing to the previous complexity, which still stands at O(n4 ).
18.13.4
The Full Cost of the Hidden Classical Algorithms
We have accounted for the classical cost of testing the linear independence and solving
the system of equations. In the process, we have demonstrated that it increases the
complexity by a factor n3 , making the full algorithm O(n4 ), not counting the oracle,
whose complexity is unknown to us.
We will do better, though, by leveraging mod-2 shortcuts that are integrated into
Simons algorithm. Youll see.
The melody that keeps repeating in our head, though, is the footnote that this
entire analysis is relative to the quantum oracle Uf , the reversible operator associated
with the Z2n periodic function, f . Its complexity is that of the black box for f , itself.
We do not generally know that complexity, and it may well be very bad. Fortunately,
in some special cases of great interest, we know enough about the function to be able
to state that it has polynomial time complexity, often of a low polynomial order.
18.14
Adjusted Algorithm
18.14.1
New Linear Independence Step
We now integrate Gaussian elimination into Simons algorithm during the test for
linear independence. Here is the step, as originally presented:
534
3. Use a classical algorithm to determine whether or not z is linearly dependent
on the vectors in W.
If it is independent, name it wj , where j is the number of elements already
stored in W, and add wj to W.
if j = n 2, we have n 1 linearly-independent vectors in W and
are done; break the loop.
If it is not independent (which includes the special case z = 0, even when
W is still empty), then continue to the next pass of the loop.
The entire block can be replaced by the following, which not only tests for linear
independence, but keeps Ws associated matrix, W , in reduced echelon form and does
so leveraging the special simplicity afforded by mod-2 arithmetic.
The algorithm is more readable if
date, W is represented by
1 0 1
0 0 0
0 0 0
0 0 0
we carry an example along. Assume that, to
1
1
0
0
0
1
0
0
1
1
1
0
0
1
0
1
0
1
0
0
1
0
0
1
and we just pulled
0 0 0 1 1 0 0 1 1
from our circuit.
New Step 3
3. Loop (new inner loop) until either z has been added to W ( ' W)
or z = 0 produced.
(a) m most significant bit (MSB) position of z that contains a 1.
(O(n))
[In our example, m = bit #5 (dont forget that we count from the right,
not the left)]
(b) Search W for a wk row with the same non-0 MSB position, m.
(O(n))
[In our example, w1 is the (only) row with an MSB of 1 6= 0 in bit #5]
If a row vector wk with same non-0 MSB is not found, insert
z between the two rows of W that will guarantee the preservation of its reduced echelon form, effectively adding it to W.
End this loop.
[In our example, this case does not occur.]
535
If a row vector wk with same non-0 MSB is found replace
z z wk ,
effectively replacing its old MSB with a 0. Continue to next
pass of this loop. (O(n) for )
[In our example,
z z w1 = (000001101)]
[Exercise. Finish this example, repeating the loop as many times as needed, and
decide whether the original z produces an augmented W or turns z into 0.]
[Exercise. Try the loop with z = 000001100.]
[Exercise. Try the loop with z = 010001100.]
Why this Works
In the proposed loop, we are considering whether or not to add some z to our set, W.
If we do not find a wk whose MSB matches the MSB of z, then adding z to
W in the prescribed fashion produces an additional row for our matrix, while
preserving its echelon form. Therefore, the rows are still linearly independent.
(This was an exercise from the Gaussian elimination section.) We have increased W by one vector, and thereby broadened its span to a subspace with
twice as many vectors, all are orthogonal to a.
[Exercise. Fill in the details.]
Say we do find a wk whose MSB matches the MSB of z. z wk is, by definition,
in the span of W {z}. Since z is orthogonal to a, so, too is the new z = z wk ,
but the new z has more 0s, bringing us one loop pass closer to either adding it W
(the above case) or arriving at a termination condition in which we ultimately
produce z = 0, which is never independent of any set, so neither was the previous
zs that got us to this point (including the original z produced by the circuit.)
[Exercise. Fill in the details, making sure to explain why z and z wk are
either a) both linearly independent of W or b) both not.]
Cost of New Step 3
This step has a new inner-loop, relative to the outer quantum sampling loop, which
has, at most, n 1 passes (we move the MSB to the right each time). Inside the loop
we apply some O(n) operations in series with each other. Therefore, the total cost of
the Step 3 loop is O(n2 ), which includes all arithmetic.
536
Summary
We are not applying GE all at once to a single matrix. Rather, we are doing an O(n2 )
operation after each quantum sample that keeps the accumulated W set in eternal
echelon-reduced form. So, its a custom O(n2 ) algorithm nested within the quantum
O(n) loop, giving an outer complexity of O(n3 ).
18.14.2
New Solution of System Step
Finally, we integrate back-substitution into the final step of the original algorithm.
The original algorithms final step was
Otherwise, we succeeded. Add an nth vector, wn1 , which is linearly independent to this set (and therefore not orthogonal to a, by a previous exercise), done
easily using a simple classical observation, demonstrated below. This produces
a system of n independent equations satisfying
(
0, k = 0, . . . , n 2
wk a =
1, k = n 1
which has a unique non-zero solution.
Replace this with the new final step,
Otherwise, we succeeded and W is an (n1)n matrix in reduced echelon form.
Add an nth row vector, wn1 , which is linearly independent to W s rows (and
therefore not orthogonal to a), using the process described in Solving the Final
Set of Linear Equations, above. That was an O(n) process that produced an
n n W , also in reduced echelon form. We now have a system of n independent
equations satisfying
(
0, k = 0, . . . , n 2
wk a =
1, k = n 1
which is already in reduced echelon form. Solve it using only back-substitution,
which is O(n2 ).
The take-away is that we have already produced the echelon form as part of the
linear-independence tests, so we are positioned to solve the system using only backsubstitution, O(n2 ).
18.14.3
Cost of Adjusted Implementation of Simons Algorithm
We detailed two adjustments to the original algorithm. The first was the test for
linear independence using a mod-2 step that simulateously resulted in GE at the end
all the quantum sampling. The second was the solution of the system of equations.
537
Cost of the mod-2 Test for Independence
As already noted, the step-3 loop is repeated at most n 1 times, since we push
the MSB of z to the right at least one position each pass. Inside this loop, we have
O(n) operations, all applied in series. The nesting in step 3, therefore, produces an
O(n2 ) complexity. Step 3 is within the outer quantum loop, O(n), which brings the
outer-most complexity to O(n3 ), so far . . . .
Cost of Solving the System
We saw that we could solve the system using only back-substitution, O(n2 ). This is
done outside the entire quantum loop.
The total cost, quantum + classical, using our fancy footwork is, therefore, O(n3 ),
relativized to the oracle.
18.15
Classical Complexity of Simons Problem
.
Classically, this problem is hard, that is, deterministically, we certainly need a
number of trials that increases exponentially in n to get the period, and even if we
are satisfied with a small error, we would still need to take an exponential number
of samples to achieve that (and not just any exponential number of samples, but a
really big one). Lets demonstrate all this.
18.15.1
Classical Deterministic Cost
Recall that the domain can be partitioned (in more than one way) into two disjoint
sets, R and Q,
Z2n =
R
Q
= { , x, } { , x a, } ,
with f one-to-one on R and Q, individually. We pick xs at random (avoiding duplicates) and plug each one into f or a classical oracle of f if you like,
x
Classical f
f (x) .
If we sample f any fewer than (half the domain size) + 1, that is, (2n /2) + 1 =
2n1 + 1, times, we may be unlucky enough that all of our outputs, f (x), are images
of x R (or all Q), which would produce all distinct functional values. There is
no way to determine what a is if we dont get a duplicate output, f (x0 ) = f (x00 ) for
some distinct x0 , x00 dom(f ). (Once we do, of course, a = x0 x00 , but until that
time, no dice.)
Therefore, we have to sample 2n1 + 1 times, exponential in n, to be sure we get
at least one x from R and one from Q, thus producing a duplicate.
538
18.15.2
Classical Probabilistic Cost
However, what if we were satisfied to find a with a pre-determined probability of
error, ? This means that we could choose an as small as we like and know that if
we took m = m(, n) samples, we would get the right answer, a, with probability
P ( figure out a, correctly)
1 .
The functional dependence m = m(, n) is just a way to express that we are allowed
to let the number of samples, m, depend on both how small an error we want and
also how big the domain of f is. For example, if we could show that m = 21/ worked,
then since that is not dependent on n the complexity would be constant time. On the
other hand, if we could only prove that an m = n4 21/ worked, then the algorithm
would be O(n4 ).
The function dependence we care about does not involve , only n, so really, we
are interested in m = m(n).
What we will show is that even if we let m be a particular function that grows
exponentially in n, we wont succeed. Thats not to say every exponentially increasing
sample size which is a function of n would fail we already know that if we chose
2n1 + 1 we will succeed with certainty. But well see that some smaller exponential
function of n, specifically m = 2n/4 , will not work, and if that wont work then no
polynomial dependence on n, which necessarily grows more slowly than m = 2n/4 ,
has a chance, either.
An Upper Bound for Getting Repeats in m Samples
For the moment, we wont concern ourselves with whether or not m is some function
of or n. Instead, lets compute an upper bound on the probability of getting a
repeat when sampling a classical oracle m times. That is, well get the probability
as a function of m, alone. Afterwards, we can stand back and see what kind of
dependence m would require on n in order that the deck be stacked in our favor.
We are looking for the probability that at least two samples f (xi ), f (xj ) are equal
when choosing m distinct inputs, {x0 , x1 , . . . , xm1 }. Well call that event E . The
more specific event that some pair of inputs, xi and xj , yield equal f (x)s, will be
referred to as Eij . Since Eij and Eji are the same event, we only have to list it once,
so we only consider the cases where i < j. Clearly,
m1
Eij
i, j=0
i<j
The probability of a union is the sum of the probabilities of the individual events
539
minus the probabilities of the various intersections of the individual events,
P (E )
m1
X
P (Eij )
i, j=0
i<j
m1
X
P (Eij Ekl . . .)
various
combinations
P (Eij ) .
i, j=0
i<j
The number of unordered pairs, {i, j}, i 6= j, when taken from m things is (look up
n choose k if you have never seen this)
m
m!
m (m 1)
=
=
.
2
2!(m 2)!
2
This is exactly the number of events Eij that we are counting since our condition,
0 i < j m 1, is in 1-to-1 correspondence with the set of unordered pairs {i, j},
i and j between 0 and m 1, inclusive and i 6= j.
Meanwhile, the probability that an individual pair produces the same f value is
just the probability that we choose the second one, xj , in such a way that it happens
to be exactly xi a. Since were intentionally not going to pick xi a second time this
leaves 2n 1 choices, of which only one is xi a, so that gives
1
.
P (Eij ) =
n
2 1
Therefore, weve computed the number of elements in the sum, m(m 1)/2, as well
as the probability of each element in the sum, 1/(2n 1), so we plug back into our
inequality to get
1
m (m 1)
n
.
P (E )
2
2 1
[Exercise. We know that when we sample m = 2n1 + 1 times, we are certain to
get a duplicate. As a sanity check, make sure that plugging this value of m into the
derived inequality gives an upper bound that is no less than one. Any value 1 will
be consistent with our result. ]
This is the first formula we sought. We now go on to see what this implies about
how m would need to depend on n to give a decent chance of obtaining a.
What the Estimate Tells Us about m = m(n)
To get our feet wet, lets imagine the pipe dream that we can use an m that is
independent of n. The bound we proved,
m (m 1)
1
P (E )
n
,
2
2 1
tells us that any such m (say an integer m > 1/1000000 ) is going to have an exponentially small probability as n . So that settles that, at least.
540
An Upper Bound for Getting Repeats in m = 2n/4 Samples
But, what about allowing m to be a really large proportion of our domain, like
m = 2n/4 ?
Thats an exponential function of n, so it really grows fast. Now were at least being
realistic in our hopes. Still, they are dashed:
P (E )
=
=
=
<
m (m 1)
1
2n/4 (2n/4 1)
1
n
=
n
2
2 1
2
2 1
1
2n/2
1
2n/2 2n/4
n
<
n
2
2 1
2
2 1
n/2
n/2
1
2
1
2
n
<
n
2 2 1
2 2 2n/2
1
2n/2
1
1
n/2 n/2
=
n/2
2 2 (2 1)
2 2 1
1
,
n/2
2 1
a probability that still approaches 0 (exponentially fast unnecessary salt in the
wound) as n . Although we are allowing ourselves to sample the function, f , at
m = m(n) distinct inputs, where m(n) grows exponentially in n, we cannot guarantee
a reasonable probability of getting a repeat f (x) for all n. The probability always
shrinks to 0 when n gets large. This is exponential time complexity.
This problem is hard, even probabilistically, in classical terms. Simons algorithm
gives a (relativized) exponential speed-up over classical methods.
541
Chapter 19
Real and Complex Fourier Series
19.1
The Classical Path to Quantum Fourier Transforms
Our previous quantum algorithms made propitious use of the nth order Hadamard
transform, H n , but our next algorithm will require something a little higher octane.
The fundamental rules apply: a gate is still a gate and must be unitary. As such, it
can be viewed as a basis change at one moment and a tool to turn a separable input
state like |0in into a superposition, the next. Well have occasion to look at it in both
lights.
Our objective in the next four lessons is to study the quantum Fourier transform
a.k.a. the QFT . This is done in three classical chapters and one quantum chapter.
Reading and studying all three classical chapters will best prepare you for the fourth
QFT chapter, but you can cherry pick, skim, or even skip one or more of the three
depending on your interest and prior background.
An aggressive short cut might be to try starting with the third, the Discrete and
Fast Fourier Transforms, read that for general comprehension, then see if it gives you
enough of a foundation to get through the fourth QFT chapter.
If you have time and want to learn (or review) the classic math that leads to
QFT , the full path will take you through some beautiful topics:
This Chapter [Real Fourier Series Complex Fourier Series]
Chapter 20 [Continuous Fourier Transform]
Chapter 21 [Discrete Fourier Transform Fast Fourier Transform]
Chapter 22 [Quantum Fourier Transform]
If you do decide to invest time in the early material, you will be incidentally adding
knowledge applicable to many fields besides quantum computing. You will also have
542
some additional insight into the QFT .
In this first lesson we introduce the concepts frequency and period and show how
a periodic function can be built using only sines and cosines (real Fourier series) or
exponentials (complex Fourier series). The basic concepts here make learning the
Fourier transform (next chapter) very intuitive.
19.2
Periodic Functions and Their Friends
Fourier Series apply to functions that are either periodic or are only defined on
a bounded interval. Well first define the terms, periodicity, bounded domain and
compact support, then we can get on with Fourier Series.
19.2.1
Periodic Functions over R
A function f whose domain is (nearly) all the real numbers, R, is said to be periodic
if there is a unique smallest positive real number T , such that
f (x + T )
f (x),
for all x.
The number, T , is called the period of f .
(Unique smallest is a little redundant, but it condenses two ideas. Officially,
wed say that there must exist some T > 0 with the stated property and, if there did,
we would call the smallest such T (technically the GLB or greatest lower bound ) its
period.)
The simplest examples we can conjure up instantly are the functions sine and
cosine. Take sine, for example:
sin (x + 2)
sin (x),
for all x,
and 2 is the smallest positive number with this property, so a = 2 is its period. 4
and 12 satisfy the equality, but theyre not as small as 2, so theyre not periods.
Its graph manifests this periodicity in the form of repetition (Figure 19.1).
I said that the domain could be nearly all real numbers, because its fine if the
function blows-up or is undefined on some isolated points. A good example is the
tangent function, y = tan x, whose period is half that of sin x but is undefined for
/2, 3/2, etc. (Figure 19.2).
543
Figure 19.1: Graph of the quintessential periodic function y = sin x
(period = 2)
Figure 19.2: The function y = tan x blows-up at isolated points but is still periodic
(with period )
19.2.2
Functions with Bounded Domain or Compact Support
A function which is defined only over a bounded interval of the real numbers (like
[0, 100] or [, ]) is said to have bounded domain. An example is:
x2 ,
if x [1, 3)
f (x) =
undefined,
otherwise
Figure 19.3: Graph of a function defined only for x [1, 3)
[Note the notation [1, 3) for the half-open interval that contains the left endpoint, 1, but not the right endpoint, 3. We could have included 3 and not included
544
1, included both or neither. It all depends on our particular goal and function.
Half-open intervals, closed on the left and open on the right, are the most useful to
us.]
A subtly different concept is that of compact support. A function might be defined
on a relatively large set, say all (or most) real numbers, but happen to be zero outside
a bounded interval. In this case we prefer to say that it has compact support.
The previous example, extended to all R, but set to zero outside [1, 3) is
2
x , if x [1, 3)
f (x) =
0,
otherwise
(See Figure 19.4)
Terminology
The support of the function is the closure of the domain where f 6= 0. In the last
function, although f is non-zero only for [1, 0) (0, 3), we include the two points 0
and 3 in its support since they are part of the closure of the set where f is non-zero.
(I realize that I have not defined closure, and I wont do so rigorously. For us, closure
means adding back any points which are right next to places where f is non-zero,
like 0 and 3 in the last example.)
Figure 19.4: Graph of a function defined everywhere, but whose support is [1, 3],
the closure of [1, 0) (0, 3)
19.2.3
The Connection Between Periodicity and Bounded
Domain
If a function is periodic, with period T , once you know it on any half-open interval of
length T , say [0, T ) or [T /2, T /2), you automatically know it for all x, complements
of f (x) = f (x + T ). So we could restrict our attention to the interval. [T /2, T /2)
imagining, if it suited us, that the function was undefined off that interval. Our
understanding of the function on this interval would tell us everything there is to
545
know about the function elsewhere, since the rest of the graph of the function is just
a repeated clone of what we see on this small part.
Likewise, if we had a (non-periodic) function with bounded domain, say [a, b], we
could throw away b to make it a half-open interval [a, b) (we dont care about f at one
point, anyway). We then convert that to an induced periodic function by insisting
that f (x) = f (x + T ), for T b a. This defines f everywhere off that interval, and
the expanded function agrees with f on its original domain, but is now periodic with
period T = b a.
As a result of this duality between periodic functions and functions with bounded
domain, I will be interchanging the terms periodic and bounded domain at will over
the next few sections, choosing whichever one best fits the context at hand.
19.3
The Real Fourier Series of a Real Period Function
19.3.1
Definitions
Figure 19.5: A periodic function that can be expressed as a Fourier series
(period = 2)
Figure 19.6: A function with bounded domain that can be expressed as a Fourier
series (support width = 2)
546
Until further notice we confine ourselves to functions with domain R and range
R.
Any well-behaved function of the real numbers that is either periodic (See Figure 19.5), or has bounded domain (See Figure 19.6), can be expressed as a sum of
sines and cosines. This is true for any period or support width, T , but we normally
simplify things by taking T = 2.
The Real Fourier Series. The real Fourier series of a well-behaved
periodic function with period 2 is the sum
f (x) =
X
X
1
+
an cos nx +
bn sin nx .
a0
2
n=1
n=1
The sum on the RHS of this equation is called the Fourier Series of the function f .
The functions of x (that is, {sin nx}n , {cos nx}n and the constant function 1/2) that
appear in the sum are sometimes called the Fourier basis functions.
Study this carefully for a moment. There is a constant term out front (a0 /2),
which simply shifts the functions graph up or down, vertically. Then, there are two
infinite sums involving cos nx and sin nx. Each term in those sums has a coefficient
some real number an or bn in front of it. In thirty seconds well see what all that
means.
[The term well-behaved could take us all week to explore, but every function
that you are likely to think of, that we will need, or that comes up in physics and
engineering, is almost certainly going to be well-behaved.]
19.3.2
Interpretation of the Fourier Series
Each sinusoid in the Fourier series has a certain frequency associated with it: the n
in sin nx or cos nx. The larger the n, the higher the frequency of that sine or cosine.
(See Figure 19.7)
Of course, not every term will have the same amount of that particular frequency. Thats where the coefficients in front of the sines and cosines come into
play. The way we think about the collection of coefficients, {an , bn }, in the Fourier
expansion is summarized in the bulleted list, below.
When small -n coefficients like a0 , a1 , b1 or b2 are large in magnitude, the function
possesses significant low frequency characteristics (visible by slowly changing,
large curvature in the graph of f ).
When the higher -n coefficients like a50 , b90 or b1000 are large in magnitude, the
function has lots of high frequency characteristics (busy squiggling) going on.
The coefficients, {an } and {bn } are often called the weights or amplitudes of
their respective basis functions (in front of which they stand). If |an | is large,
547
Figure 19.7: A low frequency (n = 1 : sin x) and high frequency (n = 20 : sin 20x)
basis function in the Fourier series
theres a lot of cos nx needed in the recipe to prepare a meal of f (x) (same
goes for |bn | and sin nx). Each coefficient adds just the right about of weight of
its corresponding sinusoid to build f .
As mentioned, the functions {sin nx} and {cos nx} are sometimes called the
Fourier basis functions, at other times the normal modes, and in some contexts
the Fourier eigenfunctions. Whatever we call them, they represent the individual ingredients used to build the original f out of trigonometric objects, and
the weights instruct the chef how much of each function to add to the recipe:
a pinch of cos 3x, a quart of sin 5x, three tablespoons of sin 17x, etc.
Caution: A sharp turn (f 0 blows up or there is a jump discontinuity) at even a
single domain point is a kind of squiggliness, so the function may appear smooth
except for one or two angled or cornered points, but those points require lots of
high frequencies in order to be modeled by the Fourier series.
548
19.3.3
Example of a Fourier Series
To bring all this into focus, we look at the Fourier series of a function that is about
a simple as you can imagine,
f (x) = x,
x [, ).
f s graph on the fundamental interval [, ) is shown in Figure 19.8. When viewed
as a periodic function, f will appear as is in Figure 19.9. Either way, we represent it
as the following Fourier series
X
2
n+1
x =
(1)
sin nx .
n
n=1
Figure 19.8: f (x) = x, defined only on bounded domain [, )
Figure 19.9: f (x) = x as a periodic function with fundamental interval [, ):
The expression of such a simple function, x, as the infinite sum of sines should
strike you as somewhat odd. Why do we even bother? There are times when we
need the Fourier expansion of even a simple function like f (x) = x in order to fully
analyze it. Or perhaps we wish to build a circuit to generate a signal like f (x) = x
electronically in a signal generator. Circuits naturally have sinusoids at their disposal,
constructed by squeezing, stretching and amplifying the 60 Hz signal coming from the
AC outlet in the wall. However, they dont have an f (x) = x signal, so that one must
built it from sinusoids, and the Fourier series provides the blueprint for the circuit.
Why is the Fourier series of f (x) = x so complicated? First, f , has a jump
discontinuity a sharp edge so it takes a lot of high frequency components to shape
it. If you look at the periodic sawtooth shape of f (x) = x, youll see these sharp
points. Second, f s graph is linear (but not constant). Crafting a non-horizontal
straight line from curvy sines and cosines requires considerable tinkering.
549
Finite Approximations
While the Fourier sum is exact (for well-behaved f ) the vagaries of hardware require
that we merely approximate it by taking only a partial sum that ends at some finite
n = N < . For our f under consideration, the first 25 coefficients of the sines are
shown in Figure 19.10 and graphed in Figure 19.11.
Figure 19.10: First 25 Fourier coefficients of f (x) = x
Figure 19.11: Graph of the Fourier coefficients of f (x) = x
The Spectrum
Collectively, the Fourier coefficients (or their graph) is called the spectrum of f . It is
a possibly infinite list (or graph) of the weights of the various frequencies contained
in f .
Viewed in this way, the coefficients, themselves, represent a new function, F (n).
The Fourier Series as an Operator Mapping Functions to Functions
The Fourier mechanism is a kind of operator, FS, applied to f (x) to get a new
function, F (n), which is also called the spectrum.
550
F (n) = FS [f (x)]
l
{an , bn }
The catch is, this new function, F , is only defined on the non-negative integers. In
fact, if you look closely, its really two separate functions of integers, a(n) = an and
b(n) = bn . But thats okay we only want to get comfortable with the idea that
the Fourier operator takes one function, f (x), domain R, and produces another
function, its spectrum F (n), domain Z0 .
f : R R
F : Z0 R
FS
f 7 F
F contains every ounce of information of the original f , only expressed in a different
form.
Computing Fourier Coefficients
The way to produce the Fourier coefficients, {an , bn } of a function, f , is through these
easy formulas (that I wont derive),
Z
1
f (x) dx , n = 0 ,
a0 =
Z
1
an =
f (x) cos nx dx , n > 0 , and
Z
1
bn =
f (x) sin nx dx , n > 0 .
They work for functions which have period 2 or bounded domain [, ). For some
other period T , we would need to multiply or divide T /(2) in the right places (check
on-line or see if you can derive the general formula).
Using these formulas, you can do lots of exercises, computing the Fourier series of
various functions restricted to the interval [, ).
In practice, we cant build circuits or algorithms that will generate an infinite
sum of frequencies, but its easy enough to stop after any finite number of terms.
Figure 19.12 shows what we get if we stop the sum after three terms.
Its not very impressive, but remember, we are using only three sines/cosines to
approximate a diagonal line. Not bad, when you think of it that way. Lets take the
first 50 terms and see what we get (Figure 19.13).
Now we understand how Fourier series work. We can see the close approximation to
the straight line near the middle of the domain and also recognize the high frequency
551
Figure 19.12: Fourier partial sum of f (x) = x to n = 3
Figure 19.13: Fourier partial sum of f (x) = x to n = 50
effort to get at the sharp teeth at each end. Taking it out to 1000 terms, Figure 19.14
shows a stage which we might find acceptable in some applications.
552
Figure 19.14: Fourier partial sum of f (x) = x to n = 1000
Note: The Fourier series
f (x) =
X
X
1
a0
+
an cos nx +
bn sin nx
2
n=1
n=1
always produces a function of x which is periodic on the entire real line, even if we
started with (and only care about) a function, f (x) with bounded domain. The RHS
of this equation matches the original f over its original domain, but the domain
of the RHS may be larger. To illustrate this, if we were modeling the function
f (x) = x2 , restricted to [, ), the Fourier series would converge on the entire real
line, R, beyond the original domain (See Figure 19.15).
Figure 19.15: f (x) has bounded domain, but its Fourier expansion is periodic.
The way to think about and deal with this is to simply ignore the infinite number of
periods magnanimously afforded by the Fourier series expression (as a function of x)
and only take the one period that lies above f s original, bounded domain.
Compact Support, Alone, is Not Enough
In contrast, if we defined a function which is defined over all R but had compact
support, [, ], it would not have a Fourier series; no single weighted sum of sinusoids
could build this function, because we cant reconstruct the flat f (x) = 0 regions on
the left and right with a single (even infinite) sum. We can break it up into three
regions, and deal with each separately, but thats a different story.
553
19.4
The Complex Fourier Series of a Periodic Function
19.4.1
Definitions
We continue to study real functions of a real variable, f (x), which are either periodic
or have a bounded domain. We still want to express them as a weighted sum of special
pure frequency functions. I remind you, also, that we are restricting our attention
to functions with period (or domain length) 2, but our results will apply to functions
having any period T if we tweak them using factors of T or 1/T in the right places.
To convert from sines and cosines to complex numbers, one formula should come
to mind: Eulers formula,
ei = cos + i sin .
Solving this for cosine and sine, we find:
ei + ei
2
ei ei
sin =
2i
cos =
While I wont show the four or five steps, explicitly, we can apply these equivalences
to the real Fourier expression for f ,
f (x) =
X
X
1
bn sin nx,
an cos nx +
+
a0
2
n=1
n=1
to get an equivalent expression,
f (x) =
a0
X
X
1
1
1
+
(an ibn )einx +
(an + ibn )einx .
2
2
2
n=1
n=1
Now, let n runneth negative to form our Complex Fourier Series of the (same) function f .
The Complex Fourier Series. The Complex Fourier series of a
well-behaved periodic function with period 2 is the sum
f (x) =
cn einx ,
n =
where
2 (an ibn ) ,
cn 21 (an + ibn ) ,
1
a ,
2 0
554
n>0
n<0 .
n=0
The complex Fourier series is a cleaner sum than the real Fourier expansion which
uses sinusoids, and it allows us to deal with all the coefficients at once when doing
computations. The price we pay is that the coefficients, cn , are generally complex
(not to mention the exponentials, themselves). However, even when they are all
complex, the sum is still real. We have been and continue to be interested in
real-valued functions of a real variable. The fact that we are using complex functions
and coefficients to construct a real-valued function does not change our focus.
19.4.2
Computing The Complex Fourier Coefficients
I expressed the cn of the complex Fourier series in terms of the an and bn of the real
Fourier series. That was to demonstrate that this new form existed, not to encourage
you to first compute the real Fourier series and, from that, compute the cn of the
complex form. The formula for computing the complex spectrum, {cn } is
Z
1
einx f (x) dx .
cn =
2
We can learn a lot by placing the complex Fourier expansion of our periodic f on the
same line as the (new, explicit) expression for the cn .
Z
X
1
inx
einx f (x) dx
f (x) =
cn e ,
where cn =
2
n =
Make a mental note of the following observations by confirming, visually, that theyre
true.
We are expressing f as a weighted-sum (weights cn ) of complex basis functions
einx . But the integral is also a kind of sum, so we are simultaneously expressing
the cn as a weighted-sum (weights f (x)) of complex basis functions einx
Under the first sum, the x in the nth basis function, einx , is fixed (thats the
number at which we are evaluating f ), but n is a summation variable; under
the second integration, it is the n of einx which is fixed (thats the index at
which we are evaluting the coefficient, cn ), and x is the itegration variable. So
the roles of x and n are swapped.
The sequence of complex weights, {cn }, is nothing more than a function c(n)
on the set of all integers, Z. This emphasizes that not only is f () a function of
x, but c() is a function of n. This way of thinking makes the above expression
look even more symmetric
f (x) =
c(n) einx ,
n =
while,
1
c(n) =
2
555
f (x) einx dx .
If we know a function, f (x), we compute a specific spectrum value, cn , by
multiplying each f (x) by the complex basis function einx and integrating x
from to . If we know a spectrum, {cn }, we compute a specific functional
value, f (x), by multiplying each cn by the complex basis functions einx and
summing n from to .
The correspondence between functions over R and Z,
f (x)
c(n)
is, conceptually, its own inverse. You do (very roughly) the same thing to get the
spectrum from the function as you do to build the function from its spectrum.
Example
Lets expandf (x) = x along the complex Fourier basis.
Our goal is to find the complex coefficients, cn , that make the following true:
x =
cn einx .
n=
Always begin with the easy one:
Z
Z
1
1
0
c0 =
x e dx =
x dx = 0 .
2
2
The others, both positive and negative, are done in a single breath,
Z
1
cn =
xeinx dx
2
inx
1 e
=
(inx + 1)
2
2 n
=
=
=
=
=
1 1
in
in
in
in
in
e
+
e
+
e
e
2 n2
1 1
[in (2 cos n) (2i sin n)]
2 n2
1 1
2i [n (cos n) 0]
2 n2
i
cos n
n
i
(1)n .
n
Putting it all together, we get
x =
i
(1)n einx .
n
n=
n6=0
556
19.5
Periods and Frequencies
19.5.1
The Frequency of any Periodic Function
This short section will be very useful in motivating the approach to Shors periodfinding algorithm. In order to use it, Ill temporarily need the letter f to mean
frequency (in keeping with the classical scientific literature), so were going to call our
periodic function under study g(x).
Weve been studying periodic functions, g(x) of real x which have periods T = 2,
and we have shown how to express them as a sum of either real sines and cosines,
g(x) =
X
X
1
a0
+
an cos nx +
bn sin nx ,
2
n=1
n=1
or complex exponentials,
g(x) =
cn einx .
n =
Each term in these sums has a certain frequency: the n. You may have gotten the
impression that the term frequency only applies to functions of the form sin (nx),
cos (nx) or e(nx) . If so, Id like to disabuse you of that notion (for which my presentation was partly responsible). In fact any periodic function has a frequency, even
those which are somewhat arbitrary looking.
Well relax the requirement that our periodic functions have period T = 2. That
was merely a convenience to make its Fourier sum take on a standard form. For the
moment, we dont care about Fourier series.
We will define frequency twice, first in the usual way and then using a common
alternative.
19.5.2
Ordinary Frequency
We know what the period, T , of a periodic g(x) means. The frequency, f , of a periodic
g(x) is just the reciprocal of the period,
f
1
.
T
The interpretation is whats important.
The frequency f tells you how many periods of g(x) fit into any unit interval
(like [0, 1) or [1/2, 1/2)).
When the period is on the small side, like 1/10, the frequency will be large (10).
In this case wed see 10 repeated patterns if we graphed g(x) between 1/2 and
1/2. (See Figure 19.16)
557
Figure 19.16: f = 10 produces ten copies of the period in [.5, .5)
When the period is larger, like 10, the frequency will be small (1/10). In this
case wed see only a small portion of the graph of g(x) between 1/2 and 1/2,
missing the full view of its curves and gyrations, only apparent if we were to
graph it over much larger interval containing at least one full period, say -5 to
+5 or 0 to 10.(See Figure 19.17)
S
Figure 19.17: f = .1 only reveals one tenth of period in [.5, .5)
The take-away here is that
f T
and if you know gs period, you know its frequency (and vice versa).
19.5.3
Angular Frequency
When we ask the question How many periods fit an interval of length l? theres
nothing forcing us to choose l = 1. Thats a common choice in physics only because
it produces answers per second, per unit or per radian of some cyclic phenomenon.
If, instead, we wanted to express things per cycle or per revolution we would choose
l = 2. Its a slightly larger interval, so the same function would squeeze 6+ times as
many periods into it; if you were to repeat something 10 times in the space (or time)
of one unit or radian, you would get 62.8 repetitions in the span of a full revolution
of 2 units or radians.
558
Angular frequency, usually denoted by the letter , is therefore defined to be
2
.
T
The relationship between angular frequency and ordinary frequency is
2 f .
The relationship between period and angular frequency has the same interpretation
and form as that for ordinary frequency with the qualitatively unimportant change
that the number 1 now becomes 2. In particular, if you know the functions angular
frequency, you know its period, courtesy of
T
559
2 .
Chapter 20
The Continuous Fourier Transform
20.1
From Series Transform
This is the second of three classical chapters in Fourier theory meant to prepare you
for the quantum Fourier transform or QFT . It fits into the full path according to:
Chapter 19 [Real Fourier Series Complex Fourier Series]
This Chapter [Continuous Fourier Transform]
Chapter 21 [Discrete Fourier Transform Fast Fourier Transform]
Chapter 22 [Quantum Fourier Transform]
In this lesson we learn how any reasonable function (not just periodic ones covered
in the previous chapter) can be built using exponential functions, each having a
specific frequency. The sums used in Fourier series turn into integrals this time
around.
Ironically, we abandon the integrals after today and go back to sums since the
QFT is a finite entity. However, the continuous Fourier transform of this lesson
is correct foundation for the QFT , so if you can devote the time, it is good reading.
20.2
Motivation and Definitions
20.2.1
Non-Periodic Functions
Fourier series assumed (and required) that a function, f , was either periodic or
restricted to a bounded domain before we could claim it was expressible as a weightedsum of frequencies. A natural question to ask is whether we can find such weightedsum expansions for non-periodic functions defined over all R, with or without compact
560
support. Can any (well-enough-behaved-but-non-periodic) function be expressed as
a sum of basis functions like einx ? There are two reasons to be hopeful.
1. Even a non-periodic function over R can be considered to have bounded domain
(and therefore be periodic) if we throw away the very distant wings (say
x < 1010000 and x > 1010000 . Therefore, restricted to this extremely large
interval, at least, it is made up of a weighted-sum of basis functions, either
sinusoids or einx . All we need to do is fix up those pesky two wings that are so
far away.
2. The wild, fast squiggliness or the slow, smooth curvature of a graph are not
exclusive to periodic functions; they are part of any function. We should be
able to build them out of sines and cosines (or exponentials) whether they
appear in functions of compact support (even if unbounded domain) or general
functions over the real line.
The answer to this question is the Fourier Transform (FT ).
20.2.2
Main Result of Complex Fourier Series
We begin with a periodic f and restate the duality between it and its (complex)
Fourier series,
f (x) =
c(n) einx ,
n =
l
1
c(n) =
2
20.2.3
f (x) einx dx .
Tweaking the Complex Fourier Series
The price well have to pay in order to express non-periodic functions as a weighted
sum of frequencies is that the Fourier basis will no longer be a discrete set of functions, {einx }nZ , indexed by an integer, n Z, and neither will their corresponding
weights, {cn }nZ be discretely indexable. Instead, n will have to be replaced by a real
number, s. This means c(n) is going to be a full-fledged function of the real numbers,
c(s), < s < .
The above formulas have to be modified in the following ways:
The integer n must become a real number s.
The sum will have to turn into an integral.
The limits of integration have to be changed from to .
561
While not required, we makep
the formulas symmetric by replacing the normalization constant, 1/(2) by 1/(2), and thus spreading that constant over
both expressions.
These considerations suggest that if we want to express a non-periodic but wellbehaved function f as a sum of frequencies, we would do so using
Z
1
c(s) eisx ds .
f (x) =
2
Since the sum is really an integral, the weights become a function over R, c(s), where
each real s represents a frequency, and c(s) is how much of that frequency the
function f requires.
20.2.4
Definition of the Fourier Transform
This weighting function, c(s), is computable from f (x) using the companion formula,
Z
1
c(s) =
f (x) eisx dx .
2
It is this last function of s that we call the Fourier Transform, or FT , of the function
f (x), and it is usually denoted by the capital letter of the function we are transforming,
in this case F (s),
Z
1
f (x) eisx dx .
F (s) =
2
Notation
We denote the Fourier transform operator using FT , as in
F = FT (f ),
FT
f F
or, using the script notation F ,
F = F (f ),
F
f F .
20.2.5
The Inverse Fourier Transform
We also consider the inverse of the Fourier transform, which allows us to recover f
from F
f = FT 1 (F ),
f = F 1 (F ).
562
or
If we look at our original motivation for the FT , we see that FT 1 (F ) is actually
built-into that formula. Substitute F (s) in for what we originally called it, c(s), and
we have the explicit form of the inverse Fourier transform,
Z
1
f (x) =
F (s) eisx ds .
2
The noteworthy items are:
The FT 1 (F ) and FT differ only by a sign in the exponent.
Alternate Definitions
We have defined the
p FT to provide maximum symmetry. Many authors
omit the factor of 1/(2) in the FT at the cost of having the full factor
1/(2) in the inverse.
There is a second common variant that puts a 2 into the exponent to
make the result of certain computations look cleaner.
In a third variation, the exponent in the forward FT is positive and the
exponent in the inverse is negative.
Be prepared to see the alternate definitions of FT and the correspondingly
modified results they will yield, usually in the form of a constant factor
change in our results.
Sometimes, instead of using a capital letter F to represent F (f ), authors will
use the caret or tilde for that purpose,
fb = F (f ),
fe = F (f ).
20.2.6
or
Real vs. Complex, Even vs. Odd
Although we start with a real-valued function, f , there is no guarantee that F will
be real-valued. After all, we are multiplying f by a complex valued function eisx
prior to integration. However, there are situations in which F will be real, i.e., all the
imaginary parts will cancel out during the integration. Well need a couple definitions.
Even Function. A function, f , is said to be even if it has the property
f (x)
f (x).
A famously even function is cos x.
Odd Function. A function, f , is said to be odd if it has the property
f (x)
=
563
f (x).
Figure 20.1: cos x is even
Figure 20.2: sin x is odd
A famously odd function is sin x.
Theorem. If f is even, F will be real-valued. If f is odd, F will be
purely imaginary-valued. In all other cases, F will take on values that
have both real and imaginary parts.
The domains of both f and F are, in our context, always assumed to be real (x
and s are both real). It is only the values of the functions f and F that we were
concerned about.
Finally, you might notice that there was nothing in any of our definitions that
required the original f be real-valued. In fact, as long as the domain is R, the range
can be C. Well be using this fact implicitly as we consider the symmetric aspects
of the spatial (or time) domain (i.e., where f lives) and the frequency domain
(where F lives). In other words, we can start out with a complex-valued f and end up
with another complex-valued F , or perhaps even a real-valued F . All combinations
are welcome.
20.2.7
Conditions for a function to Possess a Fourier Transform
Unlike Fourier series, where practically any reasonable periodic function possessed a
Fourier series, the same cannot be said of Fourier transforms. Since our function is
564
now free-range over all of R, like a hyper-chicken in a billion acre ranch, it might
go anywhere. We have to be careful that the Fourier integral converges. One oft cited
sufficient condition is that f (x) be absolutely integrable, i.e.,
Z
|f (x)| dx < .
As you can see, simple functions. like f (x) = x or f (x) = x2 x3 + 1 dont pass
the absolute-integrability test. We need functions that tend to zero strongly at both
, like a pulse or wavelet that peters-out at both sides. Some pictures to help you
visualize the graphs of square integrable functions are seen in figure 20.3. The main
characteristic is that they peter-out towards .
(The functions that I am claiming possess a Fourier transform are the fk (x) =
k2 (x), since these have already been squared, and are ready to be declared absolutelyintegrable, a stronger requirement than square-integrable.)
Figure 20.3: Square-integrable wavefunctions from Wikipedia StationaryStatesAnimation.gif, leading to the absolutely integrable k2 (x)
20.3
Learning to Compute
Lets do a couple even functions, since they give real-valued Fourier transforms which
are easy to graph.
20.3.1
Example 1: Rectangular Pulse
Compute the Fourier transform of
(
f (x)
1, |x| .5
0, everywhere else
565
We plug it directly into the definition:
Z
1
F (s) =
f (x) eisx dx
2
Z +.5
1
=
eisx dx
2 .5
+.5
1
1
i(.5s)
i(.5s)
isx
e
=
e
e
=
is 2
is 2
.5
r
r
i(.5s)
2 1 e
2 sin .5s
ei(.5s)
=
=
.
s
2i
s
That wasnt so bad. You can see both functions graphs in Figure 20.4. They demonstrate that while f is restricted to a compact support, F requires the entire real line for
its full definition. Well see that this is no accident, and has profound consequences.
Figure 20.4: A simple function and its Fourier transform
20.3.2
Example 2: Gaussian
A Gaussian function, centered at the origin has the form
f (x) = N ex
2 /(2 2 )
N is the height of its peak at the origin, and is called the standard deviation which
conveys what percentage of the total area under f falls between k, for k = 1, 2, 3,
or any multiple we like. When k = 3, 99.7% of the area is covered. (Figure 20.5
demonstrates this.)
Figure 20.5: Interpretation of from Wikipedia Standard deviation diagram)
566
Its Fourier transform is
N
F (s) =
2
ex
2 /(2 2 )
eisx dx.
This integrand has the form
ex
2 +x
which can be computed by completing the square and using a polar coordinate trick
(look it up its fun). The result is
F (s) = N e
s2
2
2
which, if you look carefully, is another Gaussian, but now with a different height
and standard deviation. The standard deviation is now 1/, rather than . Loosely
speaking, if the spread of f is wide, the spread of F is narrow, and vice versa.
20.4
Interlude: The Delta Function
20.4.1
Characterization of the Delta Function
An invaluable computational tool in the theory of Fourier transforms is the impulse,
(x), a construct that (check your scrutiny at the door, please) behaves like a function
of the real numbers that is zero everywhere except at the origin, where its value is
. It also has the property that its integral over to + (for any positive ) is 1.
In symbols,
if x = 0
,
(x) =
0, otherwise
and
Z
(x) dx
1.
(Since is 0 away from the origin, the limits of integration can be chosen to be any
interval that contains it.)
The delta function is also known as the Dirac delta function, after the physicist,
Paul Dirac, who introduced it into the literature. It cant be graphed, exactly, since
it requires information not visible on a page, but it has a graphic representation as
shown in figure 20.6.
567
Figure 20.6: Graph of the Dirac delta function
20.4.2
The Delta Function as a Limit of Rectangles
There are many ways to make this notation rigorous, the simplest of which is to
visualize (x) as the limit of a sequence of functions, {n (x)}, the nth box function
defined by
1 1
, 2n
n, if x 2n
n (x) =
.
0,
otherwise
Each n (x) satisfies the integration requirement, and as n , n (x) becomes
arbitrarily narrow and tall, maintaining its unit area all the while. (See Figure 20.7)
Figure 20.7: Sequence of box functions that approximate the delta function with
increasing accuracy
And if were not too bothered by the imprecision of informality, we accept the definition
(x)
lim n (x) .
In fact, the converging family of functions {n (x)} serves a dual purpose. In a computer we would select an N large enough to provide some desired level of accuracy and
use N (x) as an approximation for (x), thus creating a true function (no infinities
involved) which can be used for computations, yet still has the properties we require.
568
20.4.3
The Delta Function as a Limit of Exponentials
While simple, the previous definition lacks smoothness. Any function that has unit
area and is essentially zero away from the origin will work. A smoother family of
Gaussian functions, indexed by a real parameter ,
r
x2
(x) =
e
,
does the job nicely:
(x)
lim (x) .
(See Figure 20.8)
Figure 20.8: Sequence of smooth functions that approximate the delta function with
increasing accuracy
Using the integration tricks mentioned earlier, you can confirm that this has the
integration properties needed.
20.4.4
Sifting Property of the Delta Function
Among the many properties of the delta function is its ability to pick out (sift) an
individual f (x0 ) of any function f at the domain point x0 (no matter how mis-behaved
f is),
Z
f (x0 ) =
f (x) (x x0 ) dx ,
or (equivalently),
Z
x0 +
f (x) (x x0 ) dx .
=
x0
569
You can prove this by doing the integration on the approximating sequence {n (x)}
and take the limit.
This sifting property is useful in its own right, but it also gives another way to
express the delta function,
Z
1
(x) =
eisx ds .
2
It looks weird, I know, but you have all the tools to prove it. Here are the steps:
1. Compute F ((x x0 )) with the help of the sifting property.
2. Take F 1 of both sides.
3. Set x0 = 0.
[Exercise. Show these steps explicitly.]
20.5
Fourier Transforms Involving (x)
It turns out to be convenient to take Fourier transforms of functions that are not
absolutely integrable. However, not only are such functions not guaranteed to have
converging FT integrals, the ones we want to transform in fact do not have converging
FT integrals. Thats not going to stop us, though, because we just introduced a nonfunction function, (x), which will be at the receiving end when we start with a
not-so-well-behaved f .
20.5.1
Example 3: A Constant
Consider the function f (x) = 1.
h
F [f (x) = 1] (s)
1 eisx dx
The integral looks very much like an expression of the delta function. (Compare it
to the last of our many definitions of (x), about half-a-page up). If the exponent
did
? =
not have that minus sign, the integral would be exactly 2 (s), making
2 (s). That suggests that we use integration by substitution, setting x0 = x and
wind up with an integral that does match this last expression of the delta function.
That would work:
[Exercise. Try it.]
For an interesting alternative, lets simply guess that the
minus sign in the exponent doesnt matter, implying the answer would still be 2 (s). We then test
570
our hypothesis by taking F 1 2 (s) and confirm that it gives us back f (x) = 1.
Watch.
Z
h
i
1
1
F
2 (s) (s) =
2 (s) eisx ds
2
Z
=
(s) eisx ds
Z
(s 0) eisx ds = ei0x = 1. X
=
The last line comes about by applying the sifting property of (s) to the function
f (s) = eisx at the point x0 = 0.
We have a Fourier transform pair
F
2 (s) ,
and we could have started with a (x) in the spatial domain which would have given a
constant in the frequency domain. (The delta function apparently has equal amounts
of all frequencies.)
Figure 20.9: F [1] is a 0-centered delta function
Also, our intuition turned out to be correct: ignoring the minus sign in the exponent of e did not change the resulting integral, even when we left the integration
boundaries alone; the delta function can be expressed with a negative sign in the
exponent,
Z
1
(x) =
eisx ds .
2
571
20.5.2
Example 4: A Cosine
Next we try f (x) = cos x. This is done by first solving Eulers formula for cosine,
then using the definition of (x),
Z
h
i
1
F [cos x] (s) =
cos(x) eisx dx
2
Z ix
e + eix
1
=
eisx dx
2
2
Z
Z
1
ix(1s)
ix(1+s)
e
dx +
e
dx
=
2 2
1
2 (1 s) + 2 (1 + s)
=
2 2
r
=
(1 s) + (1 + s) .
2
Figure 20.10: F [cos x] is a pair of real-valued delta functions
20.5.3
Example 5: A Sine
f (x) = sin x can be derived exactly the same way that we did the cosine, only here
we solve Eulers formula for sine, giving
r
h
i
F [sin x] (s) = i
(1 + s) (1 s) .
2
Figure 20.11: F [sin x] is a pair of imaginary delta functions
Notice something interesting: this is the first time we see a Fourier transform that
is not real (never mind that (s) isnt even a function ... were inured to that by
572
now). What do you remember from the last few sections that would have predicted
this result?
Does the FT of sin x (or cos x) make sense? It should. The spectrum of sin x
needs only one frequency to represent it: |s| = 1. (We get two, of course, because
theres an impulse at s = 1, but theres only one magnitude.) If we had done
sin ax instead, we would have seen the impulses appear at s = a, instead of 1,
which agrees with our intuition that sin ax requires only one frequency to represent
it. The Fourier coefficients (weights) are zero everywhere except for s = a.
(Use this reasoning to explain why a constant function has an FT = one impulse
at 0.)
As an exercise, you can throw constants into any of the functions whose FT s we
computed, above. For example, try doing A sin (2nx).
20.6
Properties of the Fourier Transform
There are some oft cited facts about the Fourier Transform that we present in this
section. The first is one well need in this course, and the others are results that youll
probably use if you take courses in physics or engineering.
20.6.1
Translation Invariance
One aspect of the FT we will find useful is its shift-property. For real (the only
kind of number that makes sense when f happens to be a function of R),
F
f (x ) eis F (s).
This can be stated in various equivalent forms, two of which are
F
f (x + ) eis F (s) , and
F 1
F (s ) eix f (x) ,
where in the last version we do need to state that is real, since F generally is a
function of C.
If you translate (move five feet or delay by two seconds) the function in the
spatial or time domain, it causes a benign phase-shift (by ) in the frequency domain.
Seen in reverse, if you translate all the frequencies by a constant, this only multiplies
the spatial or time signal by a unit vector in C, again, something that can usually
be ignored (although it may not make sense if you are only considering real f ). This
means that the translation in one domain has no measurable effect on the magnitude
of the signal in the other.
This is a kind of invariance, because we dont care as much about eias F (s) as we
do its absolute-value-squared, and in that case
ias
e F (s) 2 = eias 2 |F (s)|2 = |F (s)|2 .
573
Since well be calculating probabilities of quantum states and probabilities are the
amplitudes absolute-values-squared, this says that both f (x) and f (x + ) have
Fourier transforms which possess the same absolute-values and therefore the same
probabilities. Well use this in Shors quantum period-finding algorithm.
20.6.2
Plancherels Theorem
Theres an interesting and crucial fact about f and F : the area between their graphs
and their domain axes are equal. This turns out to be true of all functions and their
Fourier transforms and the result is called Plancherels Theorem.
Plancherels Theorem. For any Fourier transform pair, F and f , we
have
Z
Z
2
|f (x)| dx =
|F (s)|2 ds .
A picture demonstrating Plancherels Theorem appears in Figure 20.12
Figure 20.12: Area under |f |2 and |F |2 are equal (Plancherels Theorem)
The reason this theorem is so important in quantum mechanics is that our functions are amplitudes of quantum states, and their squared absolute values are the
probability densities of those states. Since we want any state to sum (integrate) to 1
over all space (the particle must be somewhere), we need to know that the Fourier
transform does not disturb that property. Plancherels theorem assures us of this
fact.
20.6.3
Convolution
One last property that is heavily used in engineering, signal processing, math and
physics is the convolution theorem.
A convolution is a binary operator on two functions that produces a third function.
Say we have an input signal (maybe an image), f , and a filter function g that we
want to apply to f . Think of g as anything you want to do to the signal. Do you
want to reduce the salt-and-pepper noise in the image? Theres a g for that. Do you
want to make the image high contrast? Theres a g for that. How about looking at
574
only vertical edges (which a robot would care about when slewing its arms). Theres
another g for that. The filter g is applied to the signal f to get the output signal
which we denote f g and call the convolution of f and g: The convolution of f
and g, written f g, is the function defined by
Z
f () g(x ) d.
[f g] (x)
The simplest filter to imagine is one that smooths out the rough edges (noise). This
is done by replacing f (x) with a function h(x) which, at each x, is the average over
some interval containing x, say 2 from x. With this idea, h(10) would be
Z 12
f () d ,
h(10) = K
8
while h(10.1) would be
Z
h(10.1)
12.1
f () d .
8.1
(Here K is some normalizing constant like 1/(sample interval).) If f had lots of change
from any x to its close neighbors (noise), |f (10) f (10.1)| could be quite large. But
|h(10) h(10.1)|) will be small since the two numbers are integrals over almost the
same interval around x = 10. This is sometimes called a running average and is
used to track financial markets by filtering out the moment-to-moment or day-to-day
noise. Well this is nothing more than a convolution of f and g, where g(x) = K for
|x| #daystoavg., and 0, everywhere else (K often chosen to be 1/(# days to avg.).
20.6.4
The Convolution Theorem
The convolution theorem tells us that rather than apply a convolution to two functions
directly, we can get it by taking the ordinary point-by-point multiplication of their
Fourier transforms, something that is actually easier and faster due to fast algorithms
to compute transforms.
The Convolution Theorem.
f g
2 F 1 [ F (f ) F (g)]
The constant 2 would be replaced by a different constant if we were using one of
the alternate definitions of the Fourier transform.
20.7
Period and Frequency in Fourier Transforms
Lets take a moment to review the relationship between the period and frequency of
the particularly pure periodic function sin nx for some integer n. For
sin nx
575
the period is T = 2/n making the angular frequency
2
= n.
T
Well this is how we first introduced the idea of frequency informally at the start of
this lecture: I just declared the n in sin nx to be a frequency and showed pictures
suggesting that sines with large n squiggled more than those with small n.
The period-frequency relationship is demonstrated when we look at the Fourier
transform of two sinusoids, one with period 2 and another with period 2/3. First,
sin x and its spectrum are shown in figure 20.13 where it is clear that the frequency
domain has only two non-zero frequencies at 1.
Figure 20.13: sin(x) and its spectrum
Next, sin 3x and its spectrum appear in figure 20.14 where the spectrum still has only
two non-zero frequencies, but this time they appear at 3.
Figure 20.14: sin(3x) and its spectrum
Notice that both cases, when we consider to be the absolute value of the delta
spikes position, we confirm the formula,
T
2 ,
reminding us that if we know the period we know the (angular) frequency, and vice
versa.
20.8
Applications
20.8.1
Whats the FT Used For?
The list of applications of the FT is impressive. It ranges from cleaning up noisy audio
signals to applying special filters in digital photography. Its used in communications,
576
circuit-building and numerical approximation. In wave and quantum physics, the
Fourier transform is an alternate way to view a quantum state of a system.
In one example scenario, we seek to design an algorithm that will target certain
frequencies of a picture, audio or signal, f . Rather than work directly on f , where it
is unclear where the frequencies actually are, we take F = F (f ). Now, we can isolate,
remove, enhance the exact frequencies of F (s) that interest us. After modifying F to
our satisfaction, we recover f = F 1 (f ).
A second application, dearer to quantum physicists, is the representation of states
of a quantum system. We might model a system such as an electron zipping through a
known magnetic field (but it could be a laser scattering off a crystal, the spin states of
two entangled particles, or some other physical or theoretical apparatus). We prepare
the system in a specific state the electron is given a particular energy as it enters
the field at time t = 0. This state is expressed by a wavefunction . If we did a
good job modeling the system, tells us everything we could possibly know about
the system at a specific time.
Now is essentially a vector in a Hilbert space, and like all vectors, it can be
expressed in different bases depending on our interest.
Position Basis - (x)
When expressed in the position basis, it has one form, (x). This form reveals
the amplitudes for the position of the electron at some moment in time; we take
its magnitude-squared, |(x)|2 , and integrate that over a region of x-space to
learn the likelihood that the electrons position, if measured, would be in that
region.
Momentum Basis - (p)
When expressed in the momentum basis, it has a different form, (p). This
form reveals the amplitudes for the momentum of the electron: we take its
magnitude-squared, |(p)|2 , and integrate it over a region of p-space to learn
the likelihood that the electrons momentum, if measured, would fall in that
region.
Both (x) and (p) tell the exact same story, but from different points of view.
Again, like any vectors coordinates, we can transform the coordinates to a different
basis using a simple linear transformation. It so happens that the way one transforms
the position representation to the momentum representation in quantum mechanics
is by using the Fourier transform
(p) = F ((x)) .
In other words, moving between position space and momentum space is accomplished
by applying the Fourier transform or its inverse.
577
20.8.2
The Uncertainty Principle
In quantum mechanics, every student studies a Gaussian wave-packet, which has the
same form as that of our example, but with an interpretation: f = is a particles position-state, with (x) being an amplitude. We have seen that the magnitude
squared of the amplitudes tell us the probability that the particle has a certain position. So, if represents a wave-packet in position space, then |(x)|2 reveals relative
likelihoods that the particle is at position x. Meanwhile = F (), as we learned a
moment ago, is the same state in terms of momentum. If we want the probabilities
for momentum, we would graph (s).
(Figures 20.15 and 20.16 show two different wave-packet Gaussians after taking
their absolute value-squared and likewise for their Fourier transforms.
Figure 20.15: A normalized Gaussian with 2 = 3 and its Fourier transform
Figure 20.16: A more localized Gaussian with 2 = 1/7 and its Fourier transform
The second pair represents an initial narrower spread of its position state compared
to the first. The uncertainty of its position has become smaller. But notice what
happened to its Fourier transform, which is the probability density for momentum.
Its uncertainty has become larger (a wider spread). As we set up the experiment to
pin down its position, any measurement of its momentum will be less certain. You are
looking at the actual mathematical expression of the Heisenberg uncertainty principle
in the specific case where the two observables are position and momentum.
20.9
Summary
This is a mere fraction of the important techniques and properties of the Fourier
transform, and Id like to dig deeper, but were on a mission. If youre interested, I
recommend researching the sampling theorem and Nyquist frequency.
578
And with that, we wrap up our overview of the classical Fourier series and Fourier
transform. Our next step in the ladder to QFT is the discrete Fourier transform.
579
Chapter 21
The Discrete and Fast Fourier
Transforms
21.1
From Continuous to Discrete
This is the last of three classical chapters in Fourier theory meant to prepare you for
the quantum Fourier transform. It fits into the full path according to:
Chapter 19 [Real Fourier Series Complex Fourier Series]
Chapter 20 [Continuous Fourier Transform]
This Chapter [Discrete Fourier Transform Fast Fourier Transform]
Chapter 22 [Quantum Fourier Transform]
In this this lesson we turn the integrals of the last chapter back into nice finite
sums. This is allowed because the functions we really care about will be defined on a
finite set of integers, a property shared by the important functions that are the target
of Shors quantum period-finding and factoring algorithms.
After defining the discrete Fourier transform, or DFT , well study a computational improvement called the fast Fourier transform, or FFT . Even though the
DFT is what we really need in order to understand the use of the QFT in our
algorithms, the development of the recursion relations in the FFT will be needed in
the next chapter when we design the quantum circuitry that implements the QFT .
580
21.2
Motivation and Definitions
21.2.1
Functions Mapping ZN C
Sometimes Continuous Domains Dont Work
Instead of functions, f : R C, we turn to functions which are only defined on the
finite set of integers ZN = {0, 1, 2, . . . , N 1}. This is motivated by two independent
needs:
1. Computers, having a finite capacity, cant store data for continuous functions
like 0.5x3 1, cos N x or Ae2inx . Instead, such functions must be sampled at
discrete points, producing arrays of numbers that only approximate the original
function. We are obliged to process these arrays, not the original functions.
Figure 21.1: Continuous functions on a continuous domain
Figure 21.2: Sampling a continuous function at finitely many points
2. Most of the data we analyze dont have a function or formula behind them.
They originate as discrete arrays of numbers, making a function of a continuous
variable inapplicable.
Functions as Vectors
Although defined only on a finite set of integers, such function may still take real or
complex values,
(
R or
f : ZN
,
C
581
and we can convey all the information about a particular function using an array or
vector,
f (0)
c0
f (1)
c1
f
.. .
..
.
.
f (N 1)
cN 1
In other words, the function, f can be viewed as a vector in CN (or possibly RN ).
You will see me switch freely between functional notation, f (k), and vector notation,
fk or ck . The vectors can be 2, 3, or N -dimensional. They may model a 2-D security
cam image, a 3-D printer job or an N -dimensional quantum system.
Applicability of Complex Vectors
Since N -dimensional quantum systems are the stuff of this course, well need the
vectors to be complex: Quantum mechanics requires complex scalars in order to accurately model physical systems and create relative phase differences of superposition
states. Although our initial vector coordinates might be real (in fact, they may come
from the tiny set {0, 1}), well still want to think of them as living in C. Indeed, our
operators and gates will convert such coordinates into complex numbers.
21.2.2
Defining the DFT
The definitions and results of the classical FT carry over nicely to the DFT .
Primitive N th Roots are Central
We start by reviving our notation for the complex primitive N th root of unity,
N
e2i/N .
(I say the, primitive N th root because in this course I only consider this one number
to hold that title. It removes ambiguity and simplifies the discussion.) When clear
from the context, well suppress the subscript N and simply use ,
N .
From the primitive N th root, we generate all N of the roots (including 1, itself):
1, , 2 , 3 , , N 1
or
2i/N
4i/N
1, e
,e
, e6i/N , , e(N 1)2i/N
These roots will be central to our definition of DFT .
582
Figure 21.3: Primitive 5th root of 1 (the thick radius) and the four other 5th roots it
generates
Recap of the Continuous Fourier Transform
Lets look again at the way the FT maps mostly continuous functions to other mostly
continuous functions. The FT , also written F , was defined as a map between functions,
F = F (f )
F
f F,
which produced F from f using the formula
Z
1
f (x) eisx dx.
F (s) =
2
Its easy to forget what the FT is and think of it merely as the above formula, so Ill
pester you by emphasizing the reason for this definition: we wanted to express f (x)
as a weighted-sum of frequencies, s, the weights being F (s),
Z
1
f (x) =
F (s) eisx ds .
2
Adapting the Definition to the DFT
To define the DFT , we start with the FT and make the following adjustments.
583
The integrals will become sums from 0 to N 1. Rather than evaluating f at
real continuous x, we will be taking the N discrete values of the vector (fk ) and
producing a complex spectrum vector (Fj ) that also has N components.
The factor of 1/ 2 will be replaced by a normalizing factor 1/ N . This
is
a choice, just like the specific flavor of FT we used (e.g., the factor of 1/ 2
which some authors omit). In quantum computing the choice is driven by our
need for all vectors to live on the projective sphere in Hilbert space and therefore
be normalized.
We will use N th roots of unity, jk , in place of the general exponential, eisx .
[You can skip the remainder of this bullet, without loss of crucial information,
but for those wishing a more complete explanation, read on.]
Consider eisx to be a function of x which treats s as a constant frequency a
continuous parameter that nails down the function,
s (x)
eisx .
Lets rewrite the spectrum, F , using the symbolism of this s-parameter family,
s (x), in place of eisx ,
Z
1
f (x) s (x) dx .
F (s) =
2
In the discrete case, we want functions of the index k, each of whose constant
frequency is parametrized by the discrete parameter j. The N th roots of unity
provide the ideal surrogate for the continuously parametrized s . To make the
analogy to the FT as true as possible, I will take the negative of all the roots
exponents (which produces the same N roots of unity, but in a different order),
and define
j (k)
jk
N
jk .
N 1
The N functions {j }j=0
replace the infinite {s }sR . In other words, the
continuous basis functions of x, eisx parametrized by s, become N vectors,
j0
vj0
vj1
j1
.
..
.
.
.
vj =
j = 0, 1, . . . , N 1
= jk ,
vjk
..
..
.
vj(N 1)
vj(N 1)
where k is the coordinate index and j is the parameter that labels each vector.
584
Official Definition of DFT
And so we find ourselves compelled to define the discrete Fourier transform, or DFT ,
as follows.
Definition of DFT . An N th order DFT is a function of an N -component
vector f = (fk ) that produces a new N -component vector F = (Fj ) described by the formula giving the output vectors N coordinates,
Fj
N 1
1 X
fk jk ,
N k=0
for j = 0, 1, . . . N 1 ,
where = N is the primitive N th root of 1.
If we need to emphasize the order of the DFT , we could use a parenthesized exponent,
DFT (N ) . However, when the common dimension of the input and output vectors is
understood from the context, we usually omit exponent N as well as the phrase N th
order and just refer to it as the DFT .
Notation
Common notational options that well use for the DFT include
F
DFT (f )
DFT [f ] .
The jth coordinate of the output can be also expressed in various equivalent ways,
Fj
[DFT (f )]j
DFT (f )j
DFT [f ]j .
The last two lack surrounding parentheses or brackets, but the subscript j still applies
to the entire output vector, F .
Note on Alternate Definitions
Just as with the FT , the DFT has many variants all essentially equivalent but
producing slightly different constants
or minus signs in the results. In one case the
forward DFT has no factor of 1/ 2, but the reverse DFT contains a full 1/(2). In
another, the exponential is positive. To make things more confusing, a third version
has a positive exponent of , but is defined to have the minus sign built-into it, so
the overall definition is actually the same as ours. Be ready to see deviations as you
perambulate the literature.
585
Inverse DFT
Our expectation is that {Fj } so defined will provide the weighting factors needed to
make an expansion of f as a weighted sum of the frequencies j ,
fk
N 1
1 X
Fj kj ,
N j=0
valid. If the DFT is to truly represent a discrete rendition of the continuous FT ,
not only should this equality hold, but it will provide the definition of the inverse
transform, DFT 1 . Therefore, we must verify the formula, or stated differently,
consider it to be the definition of DFT 1 and prove that it is consistent with the
DFT in the sense that DFT 1 DFT = 1.
(The 1 in DFT 1 means inverse not order, the latter always appearing in
parentheses. Of course, order (1) doesnt make sense, anyway.)
Proof of Consistency of the Definitions. We substitute the definition of DFT
of a vector (fk ) into the expression of f as a weighted sum of Fj in the definition of
DFT 1 to see if we recover the original vector (fk ). Lets have some fun.
N 1
1 X
Fj kj
N j=0
(replace Fj by its definition, being careful not to re-use spent index, k)
!
N 1
N 1
1 X
1 X
jm
=
fm
kj
N j=0
N m=0
!
N
1
N 1
X
1 X
=
fm
(km)j
N m=0
j=0
From exercise (d ) of the section Roots of Unity (at the end of our complex arithmetic
lecture) the sum in parentheses collapses to N km , so the double sum becomes
N 1
1 X
fm (N km )
N m=0
N
fk
N
586
fk .
QED
21.3
Matrix Representation of the DFT
If we let 1 (and remember, we are using the shorthand = N = e2i/N ), the
following matrix, W , encodes the DFT in a convenient way. Define W as
1
1
1
1
1
1
2
3
N 1
2
4
6
2(N 1)
1
..
..
..
..
1
.
.
.
.
.
W = ..
.
N 1 j
2j
3k
(N 1)j
.
.
.
.
.
..
..
..
..
..
N 1
2(N 1)
3(N 1)
(N 1)(N 1)
1
Now, we can express DFT of the vector (fk ) as
DFT [f ]
W (fk ) .
To see this, just write it out long hand.
1
1
1
1
2
1
1 2
4
6
..
..
..
1 .
..
.
.
.
N
2j
3j
1 j
.
..
..
..
..
.
.
.
N 1
2(N 1)
3(N 1)
1
N 1
f0
f1
f2
..
.
..
(N 1)j fk
.
..
..
.
fN 1
(N 1)(N 1)
2(N 1)
The jth component of this product is the sum
N 1
1 X
fk jk
N k=0
N 1
1 X
fk jk ,
N k=0
which is just DFT [f ]j .
21.4
Properties of DFT
All the results for the continuous FT carry over to the DFT . (If you skipped the
continuous FT chapter, take these as definitions without worry.)
21.4.1
Convolution of Two Vectors
In vector form, convolution of two discrete functions, f and g, becomes
[f g]k
N
1
X
l=0
587
fl gkl .
The convolution theorem for continuous FT holds in the discrete case, where we still
have a way to compute the (this time discrete) convolution by using DFT :
f g =
2 DFT 1 [ DFT (f ) DFT (g)]
21.4.2
Translation Invariance (Shift Property) for Vectors
Translation invariance a.k.a the shift property holds in the discrete case in the form
of
DF T
fkl 7 lk Fk
As you can see, its the same idea: a spatial or time translation in one domain (shifting
the index by l) corresponds to a phase shift (multiplication by a root of unity) in
the other domain.
21.4.3
Computational Complexity of DFT
The formula
[DFT (f )]j
N 1
1 X
fk jk
N k=0
tells us that we have N complex terms to add and multiply for each resulting
component Fj in the spectrum. Since the spectrum consists of N coordinates, thats
N 2 complex operations, each of which consists of two or three real operations, so its
still O (N 2 ) in terms of real operations. Furthermore, we often use fixed point (fixed
precision) arithmetic in computers, making the real sums and products independent
of the size, N , of the vector. Thus, the DFT is O (N 2 ), period.
You might argue that as N increases, so will the precision needed by each floating
point multiplication or addition, which would require that we incorporate the number
of digits of precision, m, into the growth estimate, causing it to be approximately
O (N 2 m2 ). However, well stick with O (N 2 ) and call this the time complexity relative
to the arithmetic operations, i.e., above them. This is simpler, often correct and
will give us a fair basis for comparison with the up-coming fast Fourier transform
and quantum Fourier transform.
21.5
Period and Frequency in Discrete Fourier Transforms
A DFT applies only to functions defined on the bounded integral domain of integers
ZN = {0, 1, 2, . . . , N 1}. Within that interval of N domain numbers, if the function
repeats itself many times that is, if it has a period T , where T << N (T is much
588
less than N ) then the function would exhibit periodicity relative to the domain.
That, in turn, implies it has an associated frequency, f . In the continuous cases, T
and f are related by one of two common relationships, either
Tf
or
Tf
2 .
One could actually make the constant on the RHS different from 1 or 2, although
its rarely done. But in the discrete case we do use a different constant, namely, N .
We still have the periodicity condition that
f (k + T )
f (k) ,
for all k (when both k and k + T are in the domain), but now the frequency is defined
by
Tf
N.
If you were to compute the DFT of a somewhat random vector like
( .1 ,
.1 ,
.1 ,
.3 ,
.1 ,
0,
0,
.1 ,
.4 ,
.3 ,
.1 ,
.1 ,
.2 ,
.1 ,
.1 ,
.1 ,
.5 , 0 , .25 , .15 , .3 , 0 , .1 ,
0 , 0 , .1 , 0 , .25 , .3 , 0 , .6 ,
0 , 0 , .25 , .15 , .3 , 0 , .1 ,
0 , 0 , .1 , .35 , .3 , 0 , .1 ,
.3 , 0 , .5 , .45 , .3 , 0 , .1 ,
0 , 0 , .25 , .15 , .4 , .6 , .6 ,
0 , 0 , .2 , .15 , .3 , .6 , .1 ,
0 , 0 , .25 , .15 , .3 , 0 , .1 ,
0 , 1 , .6 , .15 , .1 , 0 , .1 ,
0 , 0 , .25 , .15 , .3 , 0 , .1 ,
0 , 2 , .25 , .75 , .2 , 0 , .1 ,
0 , 0 , .25 , .05 , .1 , 0 , .6 ,
0 , 0 , .25 , .0 , .3 , 0 , .1 ,
.25 , 0 , .25 , .15 , .0 , .6 , .1 ,
0 , .5 , .25 , .0 , .3 , 0 , .1 ,
0 , 0 , .25 , .15 , .3 , 0 , .1 ) ,
you would see no dominant frequency, which agrees with the fact that the function is
not periodic. (See figure 21.4)
Figure 21.4: The DFT of a non-periodic vector
On the other hand, a periodic vector, like
589
( .1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.15 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
.3 ,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 ,
.1 )
exhibits a very strongly selective DFT , as figure 21.5 shows.
Figure 21.5: The spectrum of a periodic vector
A very pure periodic vector akin to the pure basis functions of the continuous cases,
like sin nx, cos nx and eins would be one like
( 0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0 ) ,
and this one has a DFT in which all the non-zero frequencies in the spectrum have
the same amplitudes, as seen in figure 21.6.
For all these examples I used N = 128, and for the two periodic cases, the period
was T = 8 (look at the vectors), which would make the frequency f = N/8 = 128/8 =
16. You can see that all of the non-zero amplitudes in the spectrum (i.e., the DFT )
are multiples of 16: 16, 32, 48, etc. (There is a phantom spike at 128, but thats
590
Figure 21.6: The spectrum of a very pure periodic vector
beyond the end of the vector and merely an artifact of the graphing software, not
really part of the spectrum.)
The thing to note here is that by looking at the spectrum we can see multiples of
the frequency in the form cf , for c = 1, 2, 3, . . . , 7 and from that, deduce f and from
f , get T .
21.6
A Cost-Benefit Preview of the FF T
21.6.1
Benefit
Although merely a computational technique to speed up the DFT , the fast Fourier
transform, or FFT , is well worthy of our attention. Among its accolades,
1. it is short and easy to derive,
2. it has wide application to circuit and algorithm design,
3. it improves the DFT s O(N 2 ) to O(N log N ), a speed up that changes the
computational time of some large arrays from over an hour to a fraction of a
second, and
4. the recursive nature of the solution sets the stage for studying the quantum
Fourier transform (QFT ) circuit.
21.6.2
Cost
A slightly sticky requirement is that it only operates on vectors which have exactly
N = 2n components, but there are easy work-arounds. One is to simply pad a
deficient f with enough 0s to bring it up to the next power-of-two. For example, we
might upgrade a 5-vector to an 8-vector like so:
f0
f1
f0
f2
f1
f4
f2
0
f4
0
0
591
Therefore, until further notice, we assume that N = 2n , for some n 0.
21.7
Recursive Equation for FF T
The development of a truly fast FFT proceeds in two stages. In this section, we
develop a recursive relation for the algorithm. In the next section, we show how
to turn that into a non-recursive, i.e., iterative, solution which is where the actual
speed-up happens.
21.7.1
Splitting f into f even and f odd
Given a vector f = (fk ) of dimension N = 2n , divide it notationally into two vectors
of size N/2 called f even and f odd , defined by
f0
f1
f2
f3
f4
f5
even
odd
f
and
f
.
..
..
.
.
fN/2
fN/2+1
In terms of coordinates,
fkeven
f2k
and
fkodd
f2k+1 ,
for k = 0, 1, . . . , N/2 1. This leads to a string of equalities.
DFT [f ]j
N 1
1 X
fk jk
N k=0
N
N
1
1
2X
2X
1
fkeven j(2k) +
fkodd j(2k+1)
N
k=0
k=0
N
N
1
1
2X
2X
1
fkeven j(2k) + j
fkodd j(2k)
N
k=0
k=0
N
N
1
1
2X
2X
jk
jk
1
.
fkeven 2
+ j
fkodd 2
N
k=0
k=0
Well rewrite this to our advantage, next.
592
21.7.2
N th Order DFT in Terms of Two (N/2)th Order DFT s
Now, = N was our primitive N th root of unity, so 2 is an (N/2)th root of unity
2
(e.g., ( 8 a) = 4 a ). If we rewrite the final sum using 0 for 2 , N 0 for N/2 and
labeling the outside the odd sum as an N th root, things look very interesting:
!
0 1
N0 1
NX
1 X even 0 jk
1
1
jk
j
DFT [f ]j =
+ N
fk ( )
fkodd ( 0 )
0
0
2
N k=0
N k=0
We recognize each sum on the RHS as an order N/2 DFT , so lets go ahead and
label them as such, using an exponent label to help identify the orders. We get
1
j
DFT (N/2) f odd j
DFT (N ) [f ]j = DFT (N/2) [f even ]j + N
2
Since the j on the LHS can go from 0 (N 1), while both DFT s on the RHS are
only size N/2, we have to remind ourselves that we consider all these functions to be
periodic when convenient.
Start of Side-Trip
Lets clarify that last statement by following a short sequence of maneuvers. First turn an N/2 dimensional vector, call it g, into a periodic
vector that is N or even infinite dimensional by assigning values to
the excess coordinates based on the original N/2 with the help of an
old trick,
g( p + N/2 )
g(p) .
This effectively turns a function g defined only on the bounded domain
ZN into a periodic function of all integers Z.
Next, we prove that the definition of DFT implies that the order-N/2
spectrum of a N/2-periodic g, DF T N/2 (g), is also periodic with period
N/2.
[Exercise. Prove the last claim by considering a j with N > j N/2
and evaluating DFT (N/2) [g]j from the definition of an order N/2 DFT .
Hint: Let j = N/2 + p, where p < N/2. Then use the fact that for an
N/2-order DFT , we use N/2 as our root-of-unity. Reduce the resulting
N/2
expression from the observation that N/2
= 1. That provides a kind
of N/2-periodicity of the N/2-order DFT over the larger interval [0, N ].
Close the deal by extending the logic to larger and larger intervals.]
593
End of Side-Trip
The upshot is that the j on the RHS can always be taken modulo the size of the
vectors on the RHS, N/2. Well add that detail for utter clarity:
1
(N )
DFT [f ]j = DFT (N/2) [f even ](j mod N/2)
2
j
+ N
DFT (N/2) f odd (j mod N/2)
Finally, lets clear the smoke using shorthand like F (N ) = DFT (N ) (f ), FE = F even
and FO = F odd . The final form is due to Danielson and Lanczos, and dates back to
1942.
21.7.3
Danielson-Lanczos Recursion Relation
(N )
j
h
(N/2)
FE
+
(j
mod N/2)
j
N
(N/2)
FO
i
(j
mod N/2)
We have reduced the computation of a size N DFT to that of two size (N/2) DFT s
(and a constant time multiplication and addition). Because we can do this all the
way down to a size 1 DFT (which is just the identity operation check it out), we
are able to compute Fj in log N iterations, each one a small, fixed number of complex
additions and multiplications.
Were Not Quite There Yet
This is promising: We have to compute N output coordinates, and Danielson-Lanczos
tells us that we can get each one using, what appears to be log N operations, so it
seems like we have an O(N log N ) algorithm.
Not so fast (literally and figuratively).
1. The cost of partitioning f into f even and f odd , unfortunately, does require us
running through the full array at each recursion level, so thats a deal breaker.
2. We can fix the above by passing the array in-place and just adding a couple
parameters, start and gap, down each recursion level. This obviates the need
to partition the arrays, but, each time we compute a single output value, we still
end up accessing and adding all N of the original elements (do a little example
to compute F3 for an 8-element f ).
3. Recursion has its costs, as any computer science student well knows, and there
are many internal expenses that can ruin our efficiency even if we manage to
fix the above two items, yet refuse to abandon recursion.
In fact, it took some 20 years before someone (Tukey and Cooley get the credit)
figured out how to leverage this recursion relation to break theO(N 2 ) barrier.
594
21.7.4
Code Samples: Recursive Algorithm
Before we present a non-recursive algorithm based on Tukey and Cooleys work, we
should confirm that the recursive algorithm actually works and has expected time
complexity of N 2 .
Well use an object-oriented (OOP) solution written in C++ whose main utility
class will be FftUtil. A single FftUtil object will have its own FFT size, provided
either at instantiation (by a constructor ) or during mid-life (by an instance method
setSize()). Once a size is established all the roots of unity are stored inside the
object so that subsequent computations using that same object do not require root
calculations. Well have the obvious accessors, mutators and computational methods
in the class, as demonstrated by a simple client, shown next. Because many of the
aspects of the class are independent of the details of the computational algorithm, we
can use this class for the up-coming non-recursive, N logN , solution.
Not all the capabilities of the class will be demonstrated, but youll get the idea.
First, the clients view.
A Clients View of Class FftUtil which Utilizes the Recursive FFT
// client s use of the classes FftUtil and Complex
// to compute a recursive - style FFT
// client - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - int main ()
{
const int FFT_SIZE = 8;
int k ;
bool const TIMING = false ;
Complex * a = new Complex [ FFT_SIZE ];
// load input array a
for ( k = 0; k < FFT_SIZE ; k ++ )
a [ k ] = k * .1;
// instantiate an fft object
FftUtil fft ( FFT_SIZE ) ;
// load the object s input signal with a []
fft . setInSig (a , FFT_SIZE ) ;
// confirm that the object has the right input
cout << " IN SIGNAL ( recursive ) - - - - - - - - - - - - - - - - " << endl ;
cout << fft . toStringIn () << endl ;
// instruct object to compute its FFT using recursion
fft . c a l c F f t R e c u r s i v e () ;
// display the object s output
cout << " FFT OUT ( recursive ) - - - - - - - - - - - - - - - - " << endl ;
cout << fft . toStringOut () << endl ;
// release dynamic memory
delete [] a ;
}
The client is invoking setInSig() and toStringOut() to transfer the signals be595
tween it and object and also calling calcFftRecursive() to do the actual DFT
computation. The client also uses a simple eight-element input signal for testing,
7
fk k=0 { 0, .1, .2, .3, .4, .5, .6, .7 } .
Although the class proves to be only a slow O(N 2 ) solution, we should not shrug-off
its details; as computer scientists, we need to have reliable benchmarks for future
comparison and proof-of-correctness coding runs. Thus, we want to look inside class
FftUtil and then test it.
The publicly exposed calcFftRecursive() called by the client leans on a private
helper not seen by the client: calcFftRecursive(). First the definition of the public
member method:
// public method that client uses to request FFT computation
// assumes signal is loaded into private array inSig []
bool FftUtil :: c a l c F f t R e c u r s i v e ()
{
int k ;
// check for fatal allocation errors
if ( inSig == NULL || outSig == NULL )
return false ;
// calculate FFT ( k ) for each k using recursive helper method
for ( k = 0; k < fftSize ; k ++ )
outSig [ k ] = (1/ sqrt ( fftSize ) )
* c a lc F f t R e c W o r k e r ( inSig , k , fftSize , 1 , 1 ) ;
return true ;
}
Heres the private recursive helper which you should compare with the DanielsonLanczos relation. It is a direct implementation which emerges naturally from that
formula.
//
//
//
//
//
//
//
private recursive method that does the work
a:
array start location
gap :
interval between consec a [] elements ( so we can pass array in - place )
size :
size of a [] in current iteration
rootPos : index in the ( member ) roots [] array where current omega stored
j:
output index FFT we are computing
reverse : true if doing an inverse FFT
Complex fftFHC ( int n , Complex a [] , Complex roots [] ,
int j , int size , int rootPos , int gap , bool reverse )
{
Complex retVal , even , odd ;
int arrayPos , sizeNxtLevel , rootPosNxtLevel , gapNxtLev ;
// DFT of single element is itself
if ( size == 1 )
return a [ j * gap ]; // since using orig arraym need gap
// locals used for clarity
sizeNxtLevel = size >> 1;
r oo tP os N xt Le ve l = rootPos << 1;
gapNxtLev = gap << 1;
// two recursive calls
even = fftFHC ( n , a , roots , j % sizeNxtLevel , sizeNxtLevel ,
596
rootPosNxtLevel , gapNxtLev , reverse ) ;
odd = fftFHC ( n , a + gap , roots , j % sizeNxtLevel , sizeNxtLevel ,
rootPosNxtLevel , gapNxtLev , reverse ) ;
// put the even and odd results together to produce return for current call
// ( inverse FFT wants positive exponent , forward wants negative )
if ( reverse )
arrayPos = ( j * rootPos ) % n ; // j * omega
else
arrayPos = ( j * ( n - rootPos ) ) % n ; // -j * omega
retVal = even + roots [ arrayPos ] * odd ;
return retVal ;
}
Ill let you analyze the code to show that is does predict O(N 2 ) big-O timing, but we
will verify it using benchmarks in the next few lines.
The output confirms that we are getting the correct values.
/* = = = = = = = = = = = = = = = = = = = = = sample run = = = = = = = = = = = = = = = = = = = = =
IN SIGNAL ( recursive )
0 0.1 0.2 0.3
0.4 0.5 0.6 0.7
----------------
FFT OUT ( recursive ) - - - - - - - - - - - - - - - 0.989949 -0.141421+0.341421 i -0.141421+0.141421 i -0.141421+0.0585786 i
-0.141421 -0.141421 -0.0585786 i -0.141421 -0.141421 i -0.141421 -0.341421 i
Press any key to continue . . .
= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = */
And we loop through different input-array sizes to see how the time complexity shakes
out:
FFT
FFT
FFT
FFT
FFT
size
size
size
size
size
1024
2048
4096
8192
16384
Recursive FFT : 0.015 seconds .
Recursive FFT : 0.066 seconds .
Recursive FFT : 0.259 seconds .
Recursive FFT : 1.014 seconds .
Recursive FFT : 4.076 seconds .
The pattern is unmistakable: doubling the array size causes the time to grow fourfold. This is classic N 2 time complexity. Besides the growth rate, the absolute times
(four seconds for a modest sized signal) are unacceptable.
21.8
A Non-Recursive, N log N Solution that Defines the FF T
Strictly speaking the above code is not an FFT since it is, frankly, just not that
fast. A true FFT is an algorithm that produces N log N performance based on
non-recursive techniques.
We now study the improved FFT algorithm which consists of two main phases,
bit-reversal and iterative array-building. Well gain insight into these two sub-algorithms
597
by previewing the very short high-level FFT code that invokes them.
21.8.1
The High-Level FFT Method
To class FftUtil, we add public and private instance methods that will compute the
DFT using a non-recursive algorithm. The highest level public method will be the
(forward) FFT , called calcFft(). It consists of three method calls:
1. Bit Reversal. The first implements bit-reversal by creating a new indexing
window into our data. That is, it generates a utility array of indexes that will
help us reorder the input signal. This utility array is independent of the input
signal, but it does depend on the FFT size, N . This phase has the overall
effect of reordering the input signal.
2. Iterative Array Building. The second applies an iterative algorithm that
builds up from the newly ordered input array to ultimately replace it by the
output DFT .
3. Normalization. Finally, we multiply the result by a normalizing factor.
Heres a birds eye view of the method.
// public method that client uses to request FFT computation
// assumes signal is loaded into private array inSig []
bool FftUtil :: calcFft ()
{
// throwing exception or single up - front test has slight appeal , but
// following is safer for future changes to constituent methods
if ( ! c o p y I n S i g T o A u x S i g B i t R e v () )
return false ;
if ( ! com bineEve nOdd ( false ) )
return false ;
// false = forward FFT
if ( ! normalize () )
return false ;
return true ;
}
21.8.2
Bit-Reversal
The big break in our quest for speed comes when we recognize that the recursive
algorithm leads at its deepest nested level to many tiny order-one arrays, and
that happens after log N method calls. This is the end of the recursion at which point
we compute each of these order-one DFT s, manually. Its the infamous escape valve
of recursion. But computing the DFT of those size-one arrays is trivial:
DFT (1) ( {c} ) = {c} ,
598
that is, the DFT of any single element array is itself (apply the definition). So we
dont really have to go all the way down to that level - theres nothing to do there.
(Those, the size one DFT s are already done, even if they are in a mixed up order
in our input signal.) Instead we can halt recursion when we have size two arrays, at
which point we compute the order two DFT s explicitly. Take a look:
h
i
1
(2)
FEEOE...OE
= fp + (1)j fq ,
j
2
(2)
for some p and q, gives us the jth component of the size two DFT s. FEEOE...OE
represents one of the many order two DFT s that result from the recursion relation
by taking increasingly smaller even and odd sub-arrays in our recursive descent from
size N down to size two.
The Plan: Knowing that the first order DFT s are just the original input signals
array elements (whose exact positions are unclear at the moment), our plan is work
not from the top down, recursively, but from the bottom with the original array
values, and build-up from there. In other words, instead of recursing down from size
N to size one, we iterate up from size one to size N . To do that we need to get the
original signal, {fk }, in the right order in preparation for this rebuilding process.
So our first task is to re-order the input array so that every pair fp and fq that
we want to combine to get a size two DFT end up next to one another. While at
it, well make sure that once they are computed, all size two pairs which need to
be combined to get the fourth order DFT s will also be adjacent, and so on. This
reordering is called bit-reversal. The reason for that name will be apparent shortly.
Lets start with an input signal of size 8 = 23 that we wish to transform and define
it such that itll be easy to track:
7
fk k=0 { 0, .1, .2, .3, .4, .5, .6, .7 } .
We start at the top and, using the Danielson-Lanczos recursion relation, see how the
original f decomposes into two four-element even-odd sets, then four two-element
even-odd sets, and finally eight singleton sets. Figure 21.7 shows how we separate the
(4)
(4)
original eight-element f (8) into fE and fO , each of length 4.
599
Figure 21.7: Going from an 8-element array to two 4-element arrays (one even and
one odd)
(4)
Ill select one of these, fE , for further processing (figure 21.8).
600
Figure 21.8: Decomposing the even 4-element array to two 2-element arrays (one even
and one odd)
(2)
This time, for variety, well recurse on the odd sub-array, fEO (figure 21.9).
Figure 21.9: Breaking a two-element array into two singletons
Remember, our goal is to re-order the input array to reposition pairs such as these
next to one-another. The last picture told us that we want to see f2 repositioned so
it is next to f6 (figure 21.10).
601
Figure 21.10: Final positions of f2 should be next to f4 .
Once this happens, we are at the bottom of recursion. These one-element arrays
are their own DFT s, and as such, are neither even nor odd. Each singleton is some
F0 (more accurately, [FEOE ]0 and[FEOO ]0 .
Aside: Recall that we needed to take j (mod N/2) at each level, but when N/2
is 1, anything mod 1 is 0, which is why we are correct in calling these F0 for different
one-element F arrays (figure 21.11).
Figure 21.11: Singletons are their own DFT s
If we had taken a different path, wed have gotten a different pair. Doing it for
all possibilities would reveal that we would want the original eight values paired as
follows:
f2
f3
f0
f1
f6
f7
f4
f5
Now, we want more than just adjacent pairs; wed like the two-element DFT s
that they generate to also be next to one another. Each of these pairs has to be
positioned properly with respect to the rest. Now is the time for us to stand on the
shoulders of the giants who came before and write down the full ordering we seek.
This is shown in figure 21.12. (Confirm that the above pairs are adjacent).
What you are looking at in figure 21.12 is the bit-reversal arrangement. Its so
named because in order to get f6 , say, into its correct position for transform building,
we reverse the bits of the integer index 6 (relative to the size of the overall transform,
8, which is three bits). Its easier to see than say: 6 = 110 reversed is 011 = 3, and
indeed you will find the original f6 = .6 ends up in position 3 of the bit-reversed,
602
Figure 21.12: Bit-reversal reorders an eight-element array
array. On the other hand, 5 = 101 is its own bit reversed index, so it should and
does not change array positions.
Bit Reversal Mapping for 3-bit Integers
0=
1=
2=
3=
4=
5=
6=
7=
000
001
010
011
100
101
110
111
7
7
000 =
100 =
010 =
110 =
001 =
101 =
011 =
111 =
0
4
2
6
1
5
3
7
Bit Reversal Code
The Code. The bit reversal procedure uses a managing method, allocateBitRev()
which calls a helper reverseOneInt(). First, a look at the methods:
// called once for a given FFT size and not repeated for that size
// produces array bitRev [] used to load input signal
void FftUtil :: allocate BitRev ()
{
int k ;
if ( bitRev != NULL )
delete bitRev ;
bitRev = new int [ fftSize ];
for ( k = 0; k < fftSize ; k ++ )
bitRev [ k ] = reverseOneInt ( k ) ;
603
return ;
}
int FftUtil :: reverseOneInt ( int inVal )
{
int retVal , logSize , inValSave ;
inValSave = inVal ;
// inVal and retVal are array locations , and size of the array is fftSize
retVal = 0;
for ( logSize = fftSize >> 2; logSize > 0; logSize > >= 1)
{
retVal |= ( inVal & 1) ;
retVal < <= 1;
inVal > >= 1;
}
// adjusts for off - by - one last half of array
if ( inValSave >= ( fftSize > >1) )
retVal ++;
return retVal ;
}
Time Complexity
The driver method has a simple loop of size N making it O(N ). In that loop, it calls
the helper, which careful inspection reveals to be O(log N ): the loop in that helper is
managed by the statement logSize >>= 1, which halves the size of array each pass,
an action that always means log growth. Since this is a nesting of two loop the full
complexity is O(N log N ).
Maybe Constant Time? This is done in series with the second phase of FFT
rebuilding so this complexity and that of the next phase do not multiply; we will take
the slower of the two. But the story gets better. Bit reversal need only be done once
for any size N and can be skipped when new FFT s of the same size are computed.
It prepares a static array that is independent of the input signal. In that sense, it is
really a constant time operation for a given FFT or order N .
Either way you look at it, this preparation code wont affect the final complexity
since were about to see that the next phase is also O(N log N ) and the normalization
phase is only O(N ) making the full algorithm O(N log N ) whether or not we count
bit reversal.
21.8.3
Rebuilding from the Bit-Reversed Array
We continue to use the Danielson-Lanczos recursion relation to help guide our next
steps. Once the input array is bit-reversed, we need to build-up from there. Figure 21.13 shows how we do this at the first level, using the singleton values to build
the order-two DFT s.
Note that since we are building a two-element DFT , we use the 2nd root of unity,
a.k.a. -1. After we do this for all the pairs, thus replacing all the singletons with
604
h
i
h
i
(2)
(1)
Figure 21.13: FEO = FEOE
j
j (mod 1)
h
i
(1)
+ (1)j FEOO
j (mod 1)
two-element DFT s, we repeat the process at the next level: we build the four-element
arrays from these two-element arrays. Figure 21.14 shows the process on one of the
(4)
two four-element arrays, FE . This time, we are using the 4th root of unity, i, as the
multiplier of the odd term.
605
i
i
h
h
(2)
(4)
= FEE
Figure 21.14: FE
j (mod 2)
i
h
(2)
+ (i)j FEO
j (mod 2)
(4)
(4)
After the four-element FE we compute its companion four-element DFT , FO ,
then go to the next and final level, the computation of the full output array F (8) . If
the original signal were length 16, 32 or greater, wed need to keep iterating this outer
loop, building higher levels until we reached the full size of the input (and output)
array.
Bottom-Up Array-Building Code
The implementation will contain loops that make it easier to count the big-O of
this part, so first have a casual look at the class methods.
The Code. The private method, combineEvenOdd(), which does all the hard
work, uses an iterative approach that mirrors the diagrams. Before looking at the
full definition, it helps to first imagine processing the pairs, only, which turns those
singleton values into order-two DFT s. Recall, that for that first effort, we use 1 as
the root of unity and either add or subtract,
h
i
h
i
h
i
(2)
(1)
(1)
FEO
=
FEOE
+ (1)j FEOO
j
(mod 1)
(mod 1)
for j = 0, 1,
which replaces those eight one-element DFT s with four two-element DFT s. The
code to do that isnt too bad.
Two-Element DFT s from Singletons:
// this computes the DFT of length 2 from the DFTs of length 1 using
// F0 = f0 + f1
606
//
//
//
//
//
//
F1 = f0 - f1
and does so , pairwise ( after bit - reversal re - ordering ) .
It represents the first iteration of the loop , but has the concepts
of the recursive FFT formula .
the roots [0] , ... , roots [ fftSize -1] are the nth roots in normal order
which implies that -1 would be found at position fftSize /2
rootPos = fftSize /2; // identifies location of the -1 = omega in first pass
for ( base = 0; base < fftSize ; base +=2)
{
for ( j = 0; j < 2; j ++)
{
arrayPos = ( j * ( fftSize - rootPos ) ) % fftSize ; // -j * omega
outSig [ base + j ] = inSig [ base ] + roots [ arrayPos ] * inSig [ base + 1];
}
}
If we were to apply only this code, we would be replacing the adjacent pairs (after
bit-reversal) by their order-two DFT s. For an input signal
fk
7
k=0
{ 0, .1, .2, .3, .4, .5, .6, .7 } ,
the bit reversal would produce
{ 0, .4, .2, .6, .1, .5, .3, .7 } ,
and the the above loop would replace these by the order-two DFT s to produce
{ .4, .4, .8, .4, .6, .4, 1, .4 } .
(We are not normalizing by 1/ N until later.)
In the second generation of the loop we would combine the order two DFT s to
produce size four DFT s. This involves the following adjustments to the code:
base +=2
base +=4
rootPos = fftSize/2
1)
rootPos = fftSize/4
(locates i rather than
inSig[base]
inSig[base + (j % 2)] (original inSig[base] was
equivalent to inSig[base + (j % 1)]), since j % 1 = 0, always).
inSig[base + 1]
j < 2
inSig[base + 2 + (j % 2)]
j < 4
Ill let you write out the second iteration of the code that will combine DFT s of
length two and produce DFT s of length four. After doing that exercise, one can see
that the literals 1, 2, 4, etc. should be turned into a variable, groupsize, over which
we loop (by surrounding the above code in an outer groupsize-loop). The result
would be the final method.
Private Workhorse method combineEvenOdd():
607
// private non - recursive method builds FFT from 2 - elements up
// reverse : true if doing an inverse FFT
bool FftUtil :: combineE venOdd ( bool reverse )
{
// instance members of class - - - - - - - - - - - - - - - - - - // outSig [] , uxSig []:
working member arrays
// roots []:
precomputed roots of unity
// fftSize :
size of current FFT
//
//
//
//
//
key local variables
rootPos :
j:
groupSize :
index in the roots [] array where current omega stored
output index FFT being computed
power - of -2 size whose sub - FFT we are computing .
increases from 1 to fftSize , doubling each time
int base , j , rootPos , arrayPos , groupSize , nextGrou pStart ;
Complex omegaToJ ;
// check for fatal allocation problems
if ( outSig == NULL || auxSig == NULL || roots == NULL )
return false ;
// process entire array in groups of 1 , 2 , 4 , ... up to fftSize
for ( groupSize = 1; groupSize < fftSize ; groupSize < <= 1)
{
nextG roupSta rt = groupSize << 1;
rootPos = fftSize / nextGr oupStart ;
for ( base = 0; base < fftSize ; base += n extGrou pStart )
{
for ( j = 0; j < nex tGroupSt art ; j ++)
{
// allow forward and inverse fft
if ( reverse )
arrayPos = ( j * rootPos ) % fftSize ; // j * omega
else
arrayPos = ( j * ( fftSize - rootPos ) ) % fftSize ; // -j * omega
omegaToJ = roots [ arrayPos ];
// Danielson and Lanczos formula
outSig [ base + j ] = auxSig [ base + ( j % groupSize ) ]
+ omegaToJ * auxSig [ base + groupSize + ( j % groupSize ) ];
}
}
// rather than be clever , copy array back to aux - input for next pass
if ( ( groupSize << 1) < fftSize )
c op y O u t S i g T o A u x S i g () ;
}
return true ;
}
Time Complexity
At first glance it might look like a triple nested loop, leading to some horrific cubic
performance. Upon closer examination we are relieved to find that
the outer loop is a doubling of groupsize until it reaches N , essentially a log N
proposition, and
608
the two inner loops together constitute only N loop passes.
Thus, array building is O(N log N ), as we had hoped.
21.8.4
Normalization
To complete the trio, we have to write the normalize() method. Its a simple linear
complexity loop not adding to the growth of the algorithm.
The Normalize Method
bool FftUtil :: normalize ()
{
double factor ;
int k ;
if ( outSig == NULL )
return false ;
factor = 1. / sqrt ( mSize ) ;
for ( k = 0; k < mSize ; k ++)
outSig [ k ] = outSig [ k ] * factor ;
return true ;
}
21.8.5
Overall Complexity
We have three methods in series, and the most costly of them is O(N log N ), making
the entire FFT O(N log N ).
21.8.6
Software Testing
The only thing left to do is time this and compare with the recursive approach. Heres
the output:
FFT
FFT
FFT
FFT
FFT
FFT
FFT
size
size
size
size
size
size
size
1024
2048
4096
8192
16384
32768
65536
Non - recursive FFT : 0 seconds .
Non - recursive FFT : 0 seconds .
Non - recursive FFT : 0.001 seconds .
Non - recursive FFT : 0.001 seconds .
Non - recursive FFT : 0.003 seconds .
Non - recursive FFT : 0.007 seconds .
Non - recursive FFT : 0.016 seconds .
It is slightly slower than linear, the difference being the expected factor of logN
(although we couldnt tell that detail from the above times). Not only is this evidence
of the N logN time complexity, but it is orders of magnitude faster than the recursive
algorithm (4.076 seconds vs. .003 seconds for a 16k array). We finally have our true
FF T .
This gives us more (far more) than enough classical background in order to tackle
the quantum Fourier transform.
609
Chapter 22
The Quantum Fourier Transform
22.1
From Classical Fourier Theory to the QFT
Today youll meet the capstone of our four chapter sequence, the quantum Fourier
transform, or QFT .
Chapter 19 [Real Fourier Series Complex Fourier Series]
Chapter 20 [Continuous Fourier Transform]
Chapter 21 [Discrete Fourier Transform Fast Fourier Transform]
This Chapter [Quantum Fourier Transform]
If you skipped the three earlier lessons, you can always refer to them on an as-needed
basis. They contain every detail and definition about classical Fourier theory that
will be cited today, all developed from scratch. Search the table of contents for any
topic you want to review if you hit a snag.
22.2
Definitions
22.2.1
From C2 to H(n)
We know that the discrete Fourier transform of order N, or DFT (its order usually
implied by context), is a special operator that takes an N -dimensional complex vector
f = (fk ) to another complex vector fe = (fej ). In symbols,
DFT : CN CN ,
DFT (f ) 7 fe,
610
or, if we want to emphasize the coordinates,
DFT [(fk )] 7 (fej ) .
When we moved to the fast Fourier transform the only change (other than the implementation details) was that we required N = 2n be a power-of-2. Symbolically,
n
FFT : C2 C2 ,
FFT [(fk )] 7 (fej ).
We maintain this restriction on N and continue to take n to be log N (base 2 always
implied).
Well build the definition of the QFT atop of our firm foundation of the DFT ,
so we start with
n
DFT : C2 C2 ,
n
a 2n th order mapping of C2 to itself and use that to define the 2n th order QFT
acting on an nth order Hilbert space,
QFT : H(n) H(n) .
[Order. The word order, when describing the DFT = DFT (N ) means N , the
dimension of the underlying space, while the same word,order, when applied to a
tensor product space H(n) is n, the number of component, single qubit spaces in the
product. The two orders are not the same: N = 2n or, equivalently, n = log N .]
22.2.2
Approaches to Operator Definition
We know that there are various ways to define a unitary operator, U ,
U : H(n) H(n) .
Among them, three are prominent.
1. Describe the action of U on the N = 2n CBS kets, { |xin } = { |0in , . . . ,
|N 1in }, and extend linearly.
2. Describe the action of U on an arbitrary |in H(n) , by expressing U |in in
terms of its N = 2n complex coefficients c0 , . . . , cN 1 .
3. Give the matrix for U in the standard basis, used to convert |in s coefficients
to U |in s coefficients through matrix multiplication.
In each case, we have to confirm that U is both linear and unitary (although in
methods 1 and 3, we get linearity for free since it is built-into those definitions).
Also, we only have to use one technique, since the other two can be surmised using
linear algebra. However, if we use a different method than someone else to define the
same operator, we had better check that our definition and theirs are equivalent.
611
22.2.3
Review of Hadamard Operator
Lets make sure we understand these concepts before defining the QFT by reprising
an example from our past, the Hadamard operator.
First Order Hadamard
We used method 1 to define H = H 1 , by
H |0i
H |1i
|0i + |1i
,
2
|0i |1i
,
2
and extended, linearly. We might have used method 2, this way:
|i
If
then
|0i + |1i ,
+
|0i +
2
1
+
.
2
H |i
|1i
Finally, some authors use method 3 to define the matrix for H,
1 1 1
,
MH
2 1 1
which can then be used to produce the formulas for either method 1 or method 2.
N th Order Hadamard
Meanwhile, for the n-fold Hadamard, we didnt need to define it: we just derived it
using the laws of tensor products. We found (using method 1), that
H
|xi
=
n 1
n 2X
(1)xy |yin .
y=0
(Remember that x y = x y is the mod-2 pairing (a pseudo inner product) of the
two bit strings.)
Because it is so much easier to express operators on the CBS, { |xin }, than to say
what it does on the complex amplitudes, we normally do it that way, just as we did
with H n . Indeed, this is about how half of the authors define the QFT .
Despite the sales pitch for method 1, I think it is more instructive to use method 2
for the QFT , then afterward show how it can be done using method 1, being careful
to show the equivalence of the two approaches.
612
22.2.4
Defining the QFT
As long as were careful to check that our definition provides a linear and unitary
transformation, we are free to use a states coordinates to define it. Consider a general
state |in and its preferred basis amplitudes (a.k.a., coordinates or coefficients),
c0
c1
1
|in (cx )N
=
N = 2n .
.. ,
x=0
.
cN 1
We describe how the order-N QFT = QFT (N ) acts on the 2n coefficients, and this
will define the QFT for any qubit in H(n) .
Concept Definition of the Order-N QFT
|in
1
(cx )N
x=0 ,
QFT (N ) |in
1
(e
cy )N
y=0 .
If
then
In words, starting from |in , we form the vector of its amplitudes, (cx ); we treat (cx )
like an ordinary complex vector of size 2n ; we take its DFT (N ) to get another vector
(e
cy ); we declare the coefficients { e
cy } to be the amplitudes of our desired output state,
n
(N )
QFT
|i . The end.
Expressing the Order of the QFT
If we need it, well display the QFT s order in the superscript with the notation
QFT (N ) . Quantum computer scientists dont usually specify the order in diagrams,
so well often go with a plain QFT and remember that it operates on an order-n
Hilbert space having dimension N = 2n .
Explicit Definition of the Order-N QFT
The concept definition expressed formulaically, says that
if
|i
N
1
X
cx |xin ,
x=0
then
QFT |i
N
1
X
y=0
613
e
cy |yin .
We can really feel the Fourier transform concept when we view the states as complex
vectors |in = c = (cx ) of size 2n to which we subject the standard DFT ,
QFT |in
N
1
X
[DFT (c)]y |yin .
y=0
The definition assumes the reader can compute a DFT , so wed better unwind our
definition by expressing the QFT explicitly. The yth coordinate of the output QFT
is produced using
[QFT |in ]y
e
cy
N 1
1 X
cx yx ,
N x=0
where = N is the primitive N th root of unity. The yth coordinate can also be
obtained using the the dot-with-the-basis-vector trick,
[QFT |in ]y
hy| QFT |in ,
so you might see this notation, rather than the subscript, by physics-oriented authors.
For example, it could appear in the definition or even computation of the yth
coordinate of the QFT ,
n
hy| QFT |in
N 1
1 X
cx yx .
N x=0
Well stick with the subscript notation, [QFT |in ]y , for now.
Notation and Convention
Hold the phone. Doesnt the DFT have a negative root exponent, jk ? The way I
defined it, yes. As I already said, there are two schools of thought regarding forward
vs. reverse Fourier transforms and DFT s. I really prefer the negative exponent for the
forward transform because it arises naturally when decomposing a function into its
frequencies. But there is only one school when defining the QFT , and it has a positive
root exponent in the forward direction (and negative in the reverse). Therefore, I have
to switch conventions.
[If you need more specifics, here are three options. (i ) You can go back and define
the forward DFT using positive exponent from the start ... OR ... (ii ) You can
consider the appearance of DFT in the above expressions as motivational but rely
only on the explicit formulas for the formal definition of the QFT without anxiety
about the exponents sign difference ... OR ... (iii ) You can preserve our original DFT
but modify the definition of QFT by replacing DFT with DFT 1 everywhere
in this section.]
Anyway, we wont be referring to the DFT , only the QFT effective immediately
so the discrepancy starts and ends here.
614
Full Definition of QFT using Method 2
You should also be ready to see the full output vector, QFT |in , expanded along
the CBS,
if
|i
N
1
X
cx |xin
x=0
then
QFT |i
N 1 N 1
1 XX
cx yx |yin .
N y=0 x=0
[Exercise. Show how this is a consequence of the earlier definition.]
Typical Definition QFT Using Definition Method 1
Its more common in quantum computing to take the CBS route, and in that case
the definition of the QFT would be
QFT |xi
N 1
1 X yx n
|yi ,
N y=0
from which we would get the definition for arbitrary states by applying linearity to
their CBS expansion. Our task at hand is to make sure
1. the definitions are equivalent, and
2. they produce a linear, unitary operator.
Vetting a putative operator like this represents a necessary step in demonstrating
the viability of our ideas, so its more than just an academic exercise. You may be
doing this on your own operators some day. Imagine future students studying the
[your name here] transform. It can happen.
Equivalence of the Definitions
Step 1) Agreement on CBS. Well show that the coefficient definition, QFT ours ,
agrees with the typical CBS definition, QFT cbs on the CBS. Consider the CBS |xin .
615
It is a tall 2n component vector with a 1 in position x:
0
0
.
.
.
|xin = 1 xth coefficient
.
..
0
0
=
N
1
X
ck |ki
k=0
N
1
X
kx |kin .
k=0
Now apply our definition to this ket and see where it leads:
n
QFT ours |xi
N 1 N 1
1 XX
ck yk |yin
N y=0 k=0
N 1 N 1
1 XX
kx yk |yin
N y=0 k=0
N 1
1 X yx n
|yi
N y=0
QFT cbs |xin .
QED
Step 2) Linearity. We must also show that QFT ours is linear. Once we do that
well know that both it and QFT cbs (linear by construction) not only agree on the
CBS but are both linear, forcing the two to be equivalent over the entire H(n) .
The definition of QFT ours in its expanded form is
QFT |in
N 1 N 1
1 XX
cx yx |yi ,
N y=0 x=0
which, on close inspection, we recognize
as matrix multiplication of the vector (cx )
xy
by the matrix ( ) (the factor of 1/ N , notwithstanding). If this is not obvious,
heres a slo-mo version of that statement.
c0
c1
|in = c ,
2
..
.
616
so
QFT ours |in
e
c0
e
c1
e
c2
..
.
X
cx 0 x
X
1
x
cx
X
x
2
x
cx
..
.
1 1 1
c0
1 2 c1
1 2 4 c2
.
1 3 6
..
.. ..
.. . .
.
.
. .
.
Whenever an operators coordinates are transformed by matrix multiplication, the
operator is linear.
QED
Step 3) Unitarity. Now that we have the matrix
1 1 1
1 2
1
2
4
MQF T 1
3
6
N
1
.. ..
..
. .
.
for the QFT ,
..
.
we can use MQF T to confirm that QFT is unitary: if the matrix is unitary, so is
the operator. (Caution. This is only true when the basis in which the matrix is
expressed is orthonormal, which { |xin } is.) We need to show that
(MQF T ) MQF T
1,
so lets take the dot product of row x of (MQF T ) with column y of MQF T :
1
1
y
1
y
1
x
2x
x
2x
1
2y
1,
(
)
,
(
)
,
.
.
.
1,
,
.
.
.
=
2y
N
N
N
..
..
.
.
=
N 1
1 X k(yx)
N k=0
1
(xy N )
N
617
xy .
QED
(The second-from-last identity is from Exercise D (roots-of-unity section) in the
early lesson complex arithmetic.)
22.3
Features of the QFT
22.3.1
Shift Property
The shift property of the DFT ,
DF T
fkl 7 lk Fk ,
when plugged into the definition of QFT results in a quantum translation invariance
n
QFT |x zi
zx
N 1
1 X yx n
|yi
N y=0
zx QFT |xin .
We lose the minus exponent of because the QFT uses a positive exponent in its
forward direction.
22.3.2
A Comparison between QFT and H
You may find it enlightening to see some similarities and differences between the
Hadamard operator and the QFT .
We are working in an N = 2n -dimensional Hilbert space, H(n) . We always signify
this on the Hadamard operator using a superscript, as in H n . On the other hand,
when we specify the order of the QFT , we do so using the superscript (N ), as in
QFT (N ) . In what follows, Ill continue to omit the superscript for the QFT initially
and only bring it in when needed.
How do these two operators compare on the CBS?
nth Order Hadamard
|xi
n 1
n 2X
(1)xy |yi ,
y=0
where x y = x y is the mod-2 dot product based on the individual binary digits in
the base-2 representation of x and y. Now 1 = 2 (e2i/2 = ei = 1, X), so lets
replace the 1, above, with its symbol as the square root of unity,
H
|xi
=
n 1
n 2X
618
y=0
2xy |yin .
N -Dimensional QFT
The definition of the N th order QFT for N = 2n is
QFT
(N )
|xi
n 1
n 2X
yx |yin .
y=0
The differences between the two operators are now apparent.
The QFT uses a primitive 2n th root of unity, while H n uses a square root of
unity.
The exponent of the root-of-unity is an ordinary integer product for QFT , but
is a mod-2 dot product for H n .
QFT Equals H for N = 2
A useful factoid is that when N = 2 (n = 1) both operators are the same. Look:
QFT
(2)
|xi
=
=
=
22.3.3
1 X
1
1
(1)yx |yi
2
y=0
1
(1)0x |0i + (1)1x |1i
2
( |0i + |1i
, x=0
2
|0i |1i
x=1
The Quantum Fourier Basis
Any quantum gate is unitary by necessity (see the lecture on single qubits), and a
unitary operator acting on an orthonormal basis produces another orthonormal basis.
Ill restate the statute (from that lecture that provides the rigor for all this).
Theorem (Basis Conversion Property). If U is a unitary operator
and A is an orthonormal basis, then U (A), i.e., the image of vectors A
under U , is another orthonormal basis, B.
n
Applying QF T (2 ) to the preferred z-basis in H(n) will give us another basis for H(n) .
Well write it as { |e
x in }, where
|e
xin
QF T (2 ) |xin .
Well refer to this as the quantum Fourier basis or, once we are firmly back in purely
quantum territory, simply as the Fourier basis or frequency basis. It will be needed
in Shors period-finding algorithm.
619
The Take-Away
Where we used H n to our advantage for Simons (Z2 )n - periodicity, we anticipate
using the QFT to achieve a similar effect when we work on Shors ordinary integer
periodicity.
22.4
The QFT Circuit
Going from an operator definition to a quantum circuit that has an efficient (polynomial growth complexity) circuit is always a challenge. We will approach the problem
in bite-sized pieces.
22.4.1
Notation
Well review and establish some notation applicable to an N = 2n -dimensional H(n) .
The CBS kets in an N -dimensional H(n) are officially tensor products of the individual CBSs of the 2-dimensional qubit spaces,
|xin
|xn1 i |xn2 i |xn3 i . . . |x0 i
n1
O
|xk i ,
k=0
where |xk i is either |0i or |1i of the kth qubit space.
We index in decreasing order from xn1 to x0 because well want the right-most
bit to correspond to the least significant bit of the binary number xn1 . . . x1 x0 .
One shorthand we have used in the past for this CBS is
|xn1 i |xn2 i . . . |x1 i |x0 i ,
but we also know two other common variations, the encoded decimal integer x,
|xin ,
x 0, 1, 2, 3, . . . , 2n1 ,
and its binary representation,
|xn1 xn2 . . . x3 x2 x1 x0 i ,
x {0, 1} .
For example, for n = 3,
|0i3
|000i ,
|1i3
|001i ,
|2i
|010i ,
|3i
|011i ,
|4i
and, in general,
|100i ,
|xi3
620
|x2 x1 x0 i .
Well also use the natural consequence of this notation,
x
n1
X
xk 2k .
k=0
22.4.2
The Math that Leads to the Circuit: Factoring the
QFT
We begin with the definition of the QFT s action on a CBS and gradually work
towards the expression that will reveal the
circuit. It will take a few screens, so here
we go. (Im going to move the constant 1/ N to LHS to reduce syntax.)
N QFT (N ) |xin
N
1
X
xy |yin
y=0
N
1
X
n1
P
yk 2k
k=0
|yn1 . . . y1 y0 i
y=0
N
1
X
n1
Y
y=0
k=0
xyk
2k
|yn1 . . . y1 y0 i
(I displayed the order, N , explicitly in the LHSs QFT (N ) , something that will
come in handy in about two screens.) To keep the equations from overwhelming us,
lets symbolize the inside product by
xy
n1
Y
xyk 2
k=0
providing a less intimidating
N QFT |xin
N
1
X
xy
|yn1 . . . y1 y0 i .
y=0
To make this concrete, we pause to see what it looks like for n = 3.
8 QFT |xi3
=
+
x0 |000i + x1 |001i + x2 |010i + x3 |011i
x4 |100i + x5 |101i + x6 |110i + x7 |111i
621
Separating a DFT into even and odd sub-arrays led to the FFT algorithm, and we
try that here in the hope of a similar profit.
8 QFT |xi3
=
+
x0 |000i + x2 |010i + x4 |100i + x6 |110i
x1 |001i + x3 |011i + x5 |101i + x7 |111i
P
y-even group
P
y-odd group
.
Analysis of y -even Group
The least significant bit, y0 , of all terms in the y-even group is always 0, so for terms
in this group,
y0
0
xy0 2
xy
x01
2
Y
1,
xyk 2k
so for even y we get
2
Y
k=0
xyk 2 .
k=1
Evidently, the xy in the y-even group can start the product at k = 1 rather than
k = 0 since the k = 0 factor is 1. We rewrite the even sum with this new knowledge:
y-even group
=
=
x0 |000i
x0 |00i
+
+
7
X
2
Y
y=0
y even
k=1
x2 |010i + x4 |100i + x6 |110i
x2 |01i + x4 |10i + x6 |11i |0i
!
k
xyk 2
|y2 y1 i |0i
Now that |y0 i = |0i has been factored from the sum we can run through the even y
more efficiently by
halving the y-sum from
P7
even
P3
all
replacing |y2 y1 i |y1 y0 i ,
shifting the k-product down-by-1 so
Q2
replacing 2k 2k+1 .
622
Q1
0
, and
(Take a little time to see why these adjustments make sense.) Applying the bullets
gives
P
y-even group
!
!
3
1
X
Y
k+1
=
|y1 y0 i |0i
xyk 2
y=0
k=0
3
X
1
Y
y=0
k=0
k
2 xyk 2
!
|y1 y0 i
|0i .
Theres one final reduction to be made. While we successfully halved the size of y
inside the kets from its original 0 7 range to the smaller interval 0 3, the x in
the exponent still roams free in the original set. How do we get it to live in the same
smaller world as y? The key lurks here:
2
xyk 2k
The even sub-array rearrangement precipitated a 4th root of unity 2 rather than
the original 8th root . This enables us to replace any x > 3 with x 4, bringing it
back into the 0 3 range without affecting the computed values. To see why, do the
following short exercise.
[Exercise. For 4 x < 7 write x = 4 + p, where 0 p < 3.
Plug 4 + p in for x
in the above exponent and simplify, leveraging the fact that = 8 1.]
The bottom line is that we can replace x with (x mod 4) and the equality is still
holds true,
!
!
3
1
X
Y
k
P
(x
mod
4)
y
2
k
y-even group
=
2
|y1 y0 i |0i .
y=0
k=0
Using the same logic on general n, we get
P
y-even group
!
N/21
n2
X
Y
(x mod N/2) yk 2k
=
2
|yn2 yn3 . . . y0 i |0i .
y=0
k=0
Compare with our latest retooling of the QFT ,
N QFT (N ) |xin
N
1
X
n1
Y
y=0
k=0
xyk
2k
|yn1 . . . y1 y0 i ,
and we are encouraged to see the order-N/2 QFT staring us in the face,
q
P
(n1)
(N/2)
N
y-even group
=
QFT
| x mod (N/2) i
|0i .
2
623
Lets make this as concise as possible by using x to mean x mod (N/2).
q
P
(n1)
(N/2)
N
y-even group
=
QFT
|
x
i
|0i .
2
Weve expressed the even group as a QFT whose order N/2, half the original N .
The scent of recursion (and success) is in the air. Now, lets take a stab at the odd
group.
Analysis of y-odd Group
The least significant bit, y0 , of the all terms in the y-odd group is always 1, so for
terms in this group,
y0
0
xy0 2
xy
x11
2
Y
x,
=
xyk 2k
so for odd y we get
k=0
2
Y
xyk 2 .
k=1
We separated the factor from the rest so that we could start the product at k = 1
to align our analysis with the y-even group, above. We rewrite the odd sum using
this adjustment:
P
y-odd group
=
=
x1 |001i
x1 |00i
x3 |011i + x5 |101i + x7 |111i
x3 |01i + x5 |10i + x7 |11i |1i
+
+
7
X
2
Y
y=0
y odd
!
k
xyk 2
k=1
|y2 y1 i
|1i
Now that |y0 i = |1i and x have both been factored from the sum, we run through
the odd y by
P
P
halving the y-sum from 7odd 3all ,
replacing |y2 y1 i |y1 y0 i ,
shifting the k-product down-by-1 so
Q2
1
Q1
0
, and
replacing 2k 2k+1 .
These bullets give us
y-odd group
=
3
X
1
Y
y=0
k=0
k
2 xyk 2
624
!
|y1 y0 i
|1i ,
and we follow it by the same replacement, (x mod 4) x that worked for the y-even
group (and works here, too):
P
y-odd group
!
!
3
1
X
Y
k
(x mod 4) yk 2
= x
2
|y1 y0 i |1i .
y=0
k=0
Using the same logic on general n, we get
P
y-odd group
!
N/21
n2
X
Y
(x mod N/2) yk 2k
|yn2 yn3 . . . y0 i |1i .
2
= x
y=0
k=0
Once again, we are thrilled to see an (N/2)-Order QFT emerge from the fray,
q
P
(n1)
(N/2)
N
y-odd group
= x
|1i .
QFT
|
x
i
2
The Recursion Relation
P
P
Combine the y-even group
and y-odd group
to get
P
P
N QFT |xin = y-even group
+ y-odd group
q
(n1)
(N/2)
N
=
QFT
|
x
i
|0i
2
q
(n1)
(N/2)
N
QFT
|
x
i
|1i
+ x
2
q
(n1)
(N/2)
N
QFT
|
x
i
(|0i + x |1i)
=
2
The binomial |0i + x |1i had to end up on the right of the sum because we were
peeling off the least-significant |0i and |1i in the even-odd analysis; tensor products
are not commutative. This detail leads to a slightly annoying but easily handled
wrinkle in the end. Youll see.
Dividing out the N , using 2n for N and rearranging, we get an even clearer
picture.
|0i + 2xn |1i
n
n1
(2n )
(2n1 )
QFT
|xi
= QFT
|
xi
2
Compare this with the Danielson-Lanczos Recursion Relation for the DFT which we
turned into an FFT by unwinding recursion. In our current context its even easier,
because we have only one, not two, recursive calls to unwind.
625
If we apply the same math to the lower-order QFT on the RHS and plug the
result into last equation, using x for the x mod (N/2), we find
|0i + 2xn1 |1i
|0i + 2xn |1i
n
(2n )
(2n2 ) n2
QFT
|xi
= QFT
| x i
.
2
2
Now let recursion off its leash. Each iteration pulls a factor of |0i + 2xk |1i out
k
(and to the right) of the lower-dimensional QFT (2 ) until we get to QFT (2) , which
would be the final factor on the left,
n
Y
|0i + 2xk |1i
n
(2n )
.
QFT
|xi
=
2
k=1
First, admire the disappearance of those pesky x factors, so any anxiety about x
mod N/k is now lifted. Next, note that the RHS is written in terms of different
roots-of-unity, 2 , 4 , . . . , 2n = . However, they can all be written as powers of
= N ,
2k
nk
For example, when k = 1 we have 2 = (1), which is 2
the N th order QFT in terms of the N th root-of-unity ,
QFT
(2n )
|xin
n
Y
k=1
(n1)
nk x
|0i + 2
|1i
. Using this, we write
!
.
Interesting Observation. When N = 2 (n = 1), we know QFT = H, so the
first factor the last one that got peeled-off on the far left is literally H |x0 i (by
that time we would be taking x mod 2 which is the least significant bit of x, x0 ). This
final, far left factor is
1
H |x0 i = (1)0x0 |0i + (1)1x0 |1i .
2
and explains why the |
xi, |xi, etc. terms disappeared from the RHS; each factor
gobbled up another qubit from the tensor product, and the final, smallest, QFT (2)
turns the remaining single qubit, |x0 i, into the Hadamard superposition.
Notation and Discussion
Products between kets are tensor products. Each factor in the final product expressing
n
QFT 2 is a single superposition qubit consisting of a mixture of |0i and |1i in its
own component space. Meanwhile, there are n of those small superpositions creating
the final tensor product. It is sometimes written
QFT
(2n )
|xi
n1
O
|0i + 2xnk |1i
k=0
626
I wont use this notation, since it scares people, and when youQmultiply kets by
kets everyone knows that tensors are implied. So, Ill use the
notation for all
products, and you can infuse the tensor interpretation mentally when you see that
the components are all qubits.
However, its still worth remarking that this is a tensor product of kets from the
individual 2-dimensional H spaces (of which there are n) and as such results in a
separable state in the N -dimensional H(n) . This is a special way different from the
expansion along the CBS to express a state in this high dimensional Hilbert space.
But you should not be left with the impression that we were entitled to find a factored
representation. Most states in H(n) cannot be factored theyre not separable. The
result we derived is that when taking the QFT of a CBS we happily end up with a
separable state.
The factored representation and the CBS expansion each give different information
about the output state, and it may not always be obvious how the coefficients or
factors of the two relate (without doing the math).
A simple example is the equivalence of a factored representation and the CBS
expansion of the following |i2 in a two qubit system (n = 2, N = 4):
|0i |1i
|0i + |1i
|0i |0i
2
+
= |0i
|i
=
2
2
2
Here we have both the CBS definition of QFT |xi2 and the separable view.
In the N -dimensional case, the two different forms can be shown side-by-side,
!
n 1
2X
n
2nk x
Y
|0i
+
|1i
n
n
(2 )
xy |yi .
=
QFT
|xi
=
2
y=0
k=1
Of course, there can only be (at most) n factors in the separable factorization, while
there will be up to 2n terms in the CBS expansion.
The reason I bring this up is that the (separable) factorization is more relevant
to the QFT than it was to the FFT because we are basing our quantum work on
the supposition that there will be quantum gates in the near future. These gates
are unitary operators applied to the input CBS qubit-by-qubit, which is essentially a
tensor product construction.
Lets see how we can construct an actual QFT circuit from such unitary operators.
22.4.3
The QFT Circuit from the Math: n = 3 Case Study
In the case of n = 3, the factored QFT is
QFT (8) |xi3
|0i + 4x |1i
|0i + 2x |1i
|0i + x |1i
=
.
2
2
2
627
Good things come to those who calmly examine each factor, separately. Work from
left (most-significant output qubit) to right (least-significant output qubit).
The First (Most-Significant) Output Factor
We already know that this is H |x0 i but well want to re-derive that fact in a way
that can be used as a template for the other two factors. is an 8th root of unity,
so the coefficient of |1i in the numerator can be derived from
4x
4 (4x2 + 2x1 + x0 ) = 16x2 8x1 4x0
x
1 1 4 0 = (1)x0 ,
which means
|0i + 4x |1i
|0i + (1)x0 |1i
|0i + |1i , x0 = 0
2
|0i
|1i
x0 = 1
H |x0 i .
This was the most-significant qubit factor of output ket (the one on the far left of
the product). Lets refer to the output ket as |e
xi and its most significant separable
factor (the one at the far left of our product) as |e
x2 i. We can then rewrite the last
equation as
|e
x2 i
H |x0 i .
[Dont be lulled into thinking this is a computational basis element, though. Unlike
the input state, |xi = |x2 i |x1 i |x0 i, which is a product of CBS, and therefore, itself a
tensor CBS, the output, |e
xi while a product of states, to be sure, is not comprised of
factors which are CBS in their 2-D homes. Therefore, the product, |e
xi, is not a CBS
in the 2n -dimensional product space.]
Summary: By expressing x as powers-of-2 in the most-significant output factor,
|e
x2 i, we were able to watch the higher-powers dissolve because they turned into 1.
That left only the lowest power of , namely 4 = (1), which, in turn, produced a
Hadamard effect on the least significant bit of the input ket, |x0 i. Well do this for
the other two factors with the sober acceptance that each time, fewer high powers
will disappear.
But first, lets stand back and admire our handiwork. We have an actual circuit
element that generates the most significant separable factor for QFT (8) |xi:
|x0 i
|e
x2 i X
First, wow. Second, do you remember that I said pulling the least-significant kets
toward the right during factorization would introduce a small wrinkle? Youre looking
628
at it. The output states most-significant factor, |e
x2 i, is derived from the input states
least significant ket, |x0 i. Make a mental note that this will necessitate a reversal of
the kets once we have the output of the circuit; following the input line for qubit |x0 i
leads not to the output kets 0th factor, as we might have hoped, but rather to the
output kets (n 1)st factor.
The Middle Factor
Again, we remain aware that is an 8th root of unity, making the coefficient of |1i
in the numerator of the middle factor
2x
which means
2 (4x2 + 2x1 + x0 ) = 8x2 4x1 2x0
x
x
1 4 1 2 0 = (1)x1 (i)x0
|0i + 2x |1i
|0i + (1)x1 (i)x0 |1i
H |x1 i ,
x0 = 0
x1 (i)x0 |1i
|0i + (1)
, x0 = 1
2
The good news is that if x0 = 0, then the factor ix0 becomes 1 and we are left
with a Hadamard operator applied to the middle input ket, |x1 i.
The bad news is that if x0 = 1, we see no obvious improvement in the formula.
Fixing the bad news of item 1 is where I need you to focus all your attention and
patience, as it is the key to everything and takes only a few more neurons. Lets go
ahead and take H |x1 i, regardless of whether x0 is 0 or 1. If x0 was 0, we guessed
right, but if it was 1, what do we have to do to patch things up? Not much, it turns
out.
Lets compare the actual state we computed (wrong, if x0 was 1) with the one we
wanted (right, no matter what) and see how they differ. Writing them in coordinate
form will do us a world of good.
What we got when applying H to |x1 i ... :
1
1
x
2 1 1
... but if x0 = 1, we really wanted:
1
1
x
2 1 1 i
629
How do we transform
1
1x1
7
1
1x1 i
Answer: multiply by
1 0
R1
:
0 i
1 0
1
=
0 i
1x1
1
1x1 i
.
Now we have the more pleasant formula for the second factor,
(
H |x1 i ,
x0 = 0
|0i + 2x |1i
=
2
R1 H |x1 i , x0 = 1
A Piece of the Circuit
We found that the two most-significant factors of the (happily separable) QFT |xi
could be computed using the formulas
|e
x2 i
|e
x1 i
H |x0 i , and
(
H |x1 i ,
x0 = 0
R1 H |x1 i ,
x0 = 1.
In words:
1. We apply H to the two least significant kets, |x0 i and |x1 i, unconditionally, since
they will always be used in the computation of the final two most-significant
factors of QFT |xi.
2. We conditionally apply another operator, R1 , to the result of H |x1 i in the
eventuality that x0 = 1.
3. Although we apply all this to the two least significant input kets, |x1 i |x0 i,
what we get is the most-significant portion of the output states factorization,
|e
x2 i |e
x1 i (not the least-significant, so we must be prepared to do some swapping
before the day is done).
Item 2 suggests using a controlled-R1 gate, where bit x0 is the control. If x0 = 0,
the operator being controlled is not applied, but if x0 = 1, it is. Heres the schematic
for that piece:
This leads to the following circuit element:
...
R1
...
|x0 i
...
630
Lets add the remaining components one-at-a-time. First, we want to apply the
unconditional Hadamard gate to |x1 i. As our formulas indicate, this done before R1 ,
(R1 H |x1 i is applied right-to-left). Adding this element, we get:
|x1 i
|e
x1 i X
R1
...
|x0 i
This computes our final value for the |e
x1 i factor.
We have yet to add back in the Hadamard applied to |x0 i. Here the order is
important. We have to make sure we do this after x0 is used to control x1 s R1 gate.
Were we to apply H to x0 before using it to control R1 , x0 would no longer be there
it would have been replaced by the Hadamard superposition. So we place the H-gate
to the right of the control vertex:
|x1 i
|e
x1 i X
R1
|x0 i
|e
x2 i X
That completes the circuit element for the two most-significant separable output
factors. We can now get back to analyzing the logic of our instructional n = 3 case
and see how we can incorporate the last of our three factors.
The Last (Least-Significant) Factor
The rightmost output factor, |e
x0 i, has an with no exponent in the numerator,
x
which means
(4x2 + 2x1 + x0 ) = 4x2 2x1 x0
x
x
4 2 2 1 ()x0 = (1)x2 (i)x1 ()x0
|0i + x |1i
|0i + (1)x2 (i)x1 ()x0 |1i
2
x2 (i)x1 |1i
|0i + (1)
,
x0 = 0
2
|0i + (1)x2 (i)x1 ()x0 |1i , x = 1
0
2
This time, while the output factor does not reduce to something as simple as a H |x2 i
in any cases, when x0 = 0, it does look like the expression we had for the middle
factor, except applied here to |x2 i |x1 i rather than the |x1 i |x0 i of the middle factor.
In other words, when x0 = 0 this least significant factor reduces to
|0i + (1)x2 (i)x1 |1i
,
2
631
while the middle factor was in its entirety
|0i + (1)x1 (i)x0 |1i
.
2
This suggests that, if x0 = 0, we apply the same exact logic to |x2 i |x1 i that we used
for |x1 i |x0 i in the middle case. That logic would be (if x0 = 0)
(
H |x2 i ,
x1 = 0
|0i + x |1i
=
2
R1 H |x2 i , x1 = 1
Therefore, in the special case where x0 = 0, the circuit that works for |e
x0 i looks like
the one that worked for |e
x1 i, applied, this time, to qubits 1 and 2:
|x2 i
...
R1
...
|x1 i
To patch this up, we have to adjust for the case in which x0 = 1. The state we just
generated with this circuit was
1
1
x
x
2 1 2 (i) 1
... but if x0 = 1, we really wanted:
1
1
x
x
x
2 1 2 (i) 1 () 0
How do we transform
1x2
1
(i)x1
7
1x2
1
(i)x1 ()x0
?
Answer: multiply by
1 0
0
R2
:
1 0
1
0
1x2 (i)x1
=
1x2
1
(i)x1 ()x0
.
(Remember, this is in the case when x0 = 1, so = x0 .) This gives us the complete
formula for the rightmost output factor, |e
x0 i,
H |x2 i ,
x0 = 0, x1 = 0
R1 H |x2 i , x0 = 0, x1 = 1
x
|0i + |1i
R2 H |x2 i ,
x0 = 1, x1 = 0
R2 R1 H |x2 i , x0 = 1, x1 = 1
632
In words, we took the tentative circuit that we designed for |e
x0 i under the assumption
that x0 = 0 but tagged on a correction if x0 = 1. That amounts to our newly
introduced operator R2 controlled by x0 , so that the result would be further multiplied
by R2 .
The Circuit Element for |e
x0 i, in Isolation
As long as we only consider |e
x0 i, we can easily use this new information to patch
up the most recent x0 = 0 case. We simply add an |x0 i at the bottom of the picture,
and use it to control an R2 - gate, applied to the end of the |x2 i |e
x0 i assembly line.
|x2 i
R1
|e
x0 i X
R2
...
|x1 i
...
|x0 i
The Full Circuit for N = 8 (n = 3)
In the previous section, we obtained the exact result for the least-significant output
factor, |e
x0 i.
|x2 i
R1
|e
x0 i X
R2
...
|x1 i
...
|x0 i
In the section prior, we derived the circuit for output two most significant factors,
|e
x1 i and |e
x2 i
|x1 i
|e
x1 i X
R1
|x0 i
|e
x2 i X
All thats left to do is combine them. The precaution we take is to defer applying any
operator to an input ket until after that ket has been used to control any R-gates
needed by its siblings. That suggests that we place the |e
x1 i |e
x2 i circuit elements to
the right of the |e
x0 i circuit element, and so we do.
|x2 i
|x1 i
|x0 i
R1
|e
x0 i X
R2
633
|e
x1 i X
R1
H
|e
x2 i X
Prior to celebration, we have to symbolize the somewhat trivial circuitry for reordering the output. While trivial, it has a linear (in n = log N ) cost, but it adds
nothing to the time complexity, as well see.
|x2 i
R1
|e
x2 i X
R2
|x1 i
|x0 i
R1
|e
x1 i X
|e
x0 i X
You are looking at the complete QFT circuit for a 3-qubit system.
Before we leave this case, lets make one notational observation. We defined R1
to be the matrix that patched-up the |e
x1 i factor, and R2 to be the matrix that
patched-up the |e
x0 i factor. Lets look at those gates along with the only other gate
we needed, H:
1 0
1 0
R2
R1
0
0 i
1 1 1
H
2 1 1
The lower right-hand element of each matrix is a root-of-unity, so lets see all three
matrices again, this time with that lower right element expressed as a power of :
1 0
1 0
R2
R1
0
0 2
1 1 1
H
4
2 1
This paves the way to generalizing to a QFT of any size.
22.4.4
The QFT Circuit from the Math: General Case
Minus the re-ordering component at the far right, heres the QFT circuit we designed
for n = 3:
|x2 i
|x1 i
|x0 i
R1
R2
R1
634
[Exercise. Go through the steps that got us this circuit, but add a fourth qubit to
get the n = 4 (QFT (16) ) circuit.]
It doesnt take too much imagination to guess what the circuit would be for any
n:
|xn1 i
|xn2 i
R1
..
.
..
Rn1
..
|xn3 i
|x0 i
R2
..
..
.
...
..
R1
..
..
Rn2
...
..
..
Thats good and wonderful, but we have not defined Rk for k > 2 yet. However, the
final observation of the n = 3 case study suggested that it should be
1
0
Rk
.
nk1
0 2
You can verify this by analyzing it formally, but its easiest to just look at the extreme
cases. No matter what n is, we want R1 s lower-right element = i, and Rn1 s to be
(compare our n = 3 case study directly above), and you can verify that for k = 1
and k = n 1, thats indeed what we get.
Well, we have defined and designed the QFT circuit out of small, unitary gates.
Since youll be using QFT in a number of circuit designs, you need to be able to cite
its computational complexity, which we do now.
22.5
Computational Complexity of QFT
This is the easy part because we have the circuit diagrams to lean on.
Each gate is a single unitary operator. Some are 2-qubit gates (the controlled-Rk
gates) and some are single qubit gates (the H gates). But they are all constant time
and constant size, so we just add them up.
The topmost input line starting at |xn1 i has n gates (count the two-qubit controlledRk s as single gates in this top line, but you dont have to count their control nodes
when you get to them in the lines, below). As we move down to the lower lines
observe that each line has one-fewer gates than the one above until we get to the final
line, starting at |x0 i, which has only one gate. Thats
n
X
k=1
n(n + 1)
gates.
2
635
The circuit complexity for this is O(n2 ). Adding on a circuit that reverses the order
can only add an additional O(n) gates, but in series, not nested, so that does not
affect the circuit complexity. Therefore, we are left with a computational factor of
O(n2 )
O(log2 N ).
You might be tempted to compare this with the O(N log N ) performance of the FFT ,
but thats not really an apples-to-apples comparison, for several reasons:
1. Our circuit computes QFT |xin for only one of the N = 2n basis states. Wed
have to account for the algorithm time required to repeat N passes through
the circuit which (simplistically) brings it to O(N log2 N ). While this can be
improved, the point remains: our result above would have to by multiplied by
something to account for all N output basis states.
2. If we were thinking of using the QFT to compute the DFT , wed need to
calculate the N complex amplitudes, {e
ck } from the inputs {ck }. They dont
appear in our analysis because we implicitly considered the special N amplitudes
1
{xy }N
y=0 that define the CBS. Fixing this is feels like O(N ) proposition.
3. Even if we could repair the above with clever redesign, we have the biggest
obstacle: the output coefficients which hold the DFT information are amplitudes. Measuring the output state collapses them destroying their quantum
superposition.
Although we cannot (yet) use the QFT to directly compute a DFT with growth
smaller than the FFT , we can still use it to our advantage in quantum circuits, as
we will soon discover.
[Accounting for Precision. You may worry that increasingly precise matrix
multiplies will be needed in the 2 2 unitary matrices as n increases. This is a valid
concern. Fixed precision will only get us so far until our ability to generate and compute with increasingly higher roots-of-unity will be tapped out. So the constant time
unitary gates are relative to, or above, the primitive complex (or real) multiplications and additions. We would have to make some design choices to either limit n
to a maximum useable size or else account for these primitive arithmetic operations
in the circuit complexity. Well take the first option: our n will remain below some
maximum n0 , build our circuitry and algorithm to be able to handle adequate precision for that n0 and all n n0 . This isnt so hard to do in our current problem
since n never gets too big: n = log2 N , where N is the true size of our problem, so we
wont be needing arbitrarily large ns in practice. If that doesnt work for us in some
future problem, we can toss in the extra complexity factors and they will usually still
produce satisfactory polynomial big-Os for problems that are classically exponential.]
Further Improvements
Ill finish by mentioning, without protracted analysis, a couple ways the above
circuits can be simplified and/or accelerated:
636
If we are willing to destroy the quantum states and measure the output qubits
immediately after performing the QFT (something we are willing to do in most
of our algorithms), then the two-qubit (controlled-Rk ) gates can be replaced
with 1-qubit gates. This is based on the idea that, rather than construct a
controlled-Rk gate, we instead measure the controlling qubit first and then apply
Rk based on the outcome of that measurement. This sounds suspicious, I know:
were still doing a conditional application of a gate. However, a controlled-Rk
does not destroy the controlling qubit and contains all the conditional logic
inside the quantum gate, whereas measuring a qubit and then applying a 1qubit gate based on its outcome, moves the controlling aspect from inside the
quantum gate to the outer classical logic. It is much easier to build stable
conditioned one-qubit gates than two-qubit controlled gates. Do note, however,
that this does not improve the computational complexity.
If m-bit accuracy is enough, where m < n, then we get improved complexity. In this case, we just ignore the least-significant (n m) output qubits.
That amounts to tossing out the top (n m) lines, leaving only m channels to
compute. The new complexity is now O(m2 ) rather than O(n2 ).
637
Chapter 23
Shors Algorithm
23.1
The Role of Shors Algorithms in Computing
23.1.1
Context for The Algorithms
Shors algorithms are the crown jewels of elementary quantum information theory.
They demonstrate that a quantum computer, once realized, will be able handle some
practical applications that are beyond the reach of the fastest existing super computers, the most dramatic being the factoring of large numbers.
As you may know from news sources or academic reports, the inability of computers to factor astronomically large integers on a human timescale is the key to
RSA encryption, and RSA encryption secures the Internet. Shors quantum factoring
algorithm should be able to solve the problem in minutes or seconds and would be a
disruptive technology should a quantum computer be designed and programmed to
implement it. Meanwhile, quantum encryption, an advanced topic that we study in
the next course, offers a possible alternative to Internet security that could replace
RSA when the time comes.
23.1.2
Period Finding and Factoring
There are many ways to present Shors results, and it is easy to become confused
about what they all say. Well try to make things understandable by dividing our
study into two parts adumbrated by two observations.
1. Shors algorithm for period-finding is a relativized (read not absolute) exponential speed-up over a classical counterpart. Like Simons algorithm, there are
periodic functions that do not have polynomial-fast oracles. In those cases the
polynomial complexity of the {circuit + algorithm} around the oracle will not
help.
2. Shors algorithm for factoring not only makes use of the period-finding algo638
rithm, but also provides an oracle for a specific function that is polynomial-time,
making the entire {circuit + algorithm} an absolute exponential speed-up over
the classical version.
23.1.3
The Period Finding Problem and its Key Idea
Informal Description of Problem
Shors period-finding algorithm seeks to find the period, a, of a periodic function f (k)
whose domain is a finite set of integers, i.e., k ZM = {0, 1, 2, . . . , M 1}.
Typically M can be very large; we can even consider the domain to be all of Z,
although the ZM case will be shown to be an equivalent finitization that makes the
problem tractable. Ill state this explicitly in a couple screens. Here, we want to get
a general picture of the problem and the plan.
First, for this to even make sense a has to be less than M so that such periodicity
is visible by looking at the M domain points available to us.
Second, this kind of periodicity is more traditional than Simons, since here we
use ordinary addition to describe the period a via f (x) = f (x + a) rather than the
exotic mod-2 periodicity of Simon, in which f (x) = f (x a).
Shors Circuit Compared to Simons Circuit
Well see that the circuit and analysis have the same general framework as that of
Simons algorithm with the main difference being a post-oracle QFT gate rather
than a post-oracle Hadamard gate. Whats the big idea behind this change? At the
highest level of analysis it is quite simple even if the details are thorny.
If we put a maximally mixed superposition into the oracle as we have done for all
previous algorithms (and will do here) then a post-oracle measurement in the standard
basis should give all possible values of f with equal probabilities. That wont do. We
have to measure along a basis that will be biased toward giving information. For
Simon, that happened to be the x-basis, thus the use of a final Hadamard operator.
What is the right basis in which to measure when we are hoping to discover an integer
functions period ?
Why the QF T Will Help
Our lectures on classical Fourier series and transforms had many ideas and consequences of which Id like to revive two.
1. Period, T , and frequency, f (not the function, f (x), but its frequency) are
related by
T f
=
639
constant,
where the constant is usually 1 or 2 for continuous functions and the vector
size, M , for discrete functions.
2. If a discrete function is periodic, its spectrum DFT (f ) will have values which
are mostly small or zero except at domain points that are multiples of the
frequency (See Figure 23.1).
Figure 23.1: The spectrum of a vector with period 8 and frequency 16 = 128/8
In very broad and slightly inaccurate terms, this suggests we query the spectrum of
our function, f (x), ascertain its fundamental frequency, m, (the first non-zero spike)
and from it get the period, a = M/m.
But what does it mean to query the frequency? Thats code for take a postoracle measurement in the Fourier basis. We learned that measuring along a nonpreferred basis is actually applying the operator that converts the preferred basis to
the alternate basis, and for frequencies of periodic functions, that gate is none other
than the QFT .
Well this sounds easy, and while it may motivate the use of the QFT , figuring
out how to use it and what to test that will consume the next two weeks.
How Well Set-Up the State Prior to Applying the QF T
Another thing we saw in our Fourier lectures was that we get the cleanest, easiest-toanalyze spectrum of a periodic function when we start with a pure periodic function
in the spatial (or time) domain, and then apply the transform. In the continuous case
that was a sinusoid or exponential, e.g., sin 3x (Figure 23.2).
Figure 23.2: sin(3x) and its spectrum
In the discrete case it was a function that was zero everywhere except for a single
k < a and its multiples, like
640
( 0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0 ) .
Such overtly periodic vectors have DFT s in which all the non-zero frequencies in the
spectrum have the same amplitudes as shown in Figure 23.3.
Figure 23.3: The spectrum of a purely periodic vector with period 8 and frequency
16 = 128/8
We will process our original function so that it produces a purely periodic cousin with
the same period by
1. putting a maximally mixed state into the oracles A register to enable quantum
parallelism, and
2. conceptually collapsing the superposition at the output of the oracle by taking
a B register measurement and applying the generalized Born rule.
This will leave a pure periodic vector in the A register which we can send through
a post processing QFT gate and, finally, measure. We may have to do this more
than once, thereby producing several measurements. To extract m, and thus a, from
the measurements we will apply some beautiful mathematics.
The Final Approach
After first defining the kind of periodicity that Shors work addresses, we begin the
final leg of our journey that will take us through some quantum and classical terrain.
When were done, you will have completed the first phase in your study of quantum
computation and will be ready to move on to more advanced topics.
The math that accompanies Shors algorithms is significant; it spans areas as
diverse as Fourier analysis, number theory, complex arithmetic and trigonometry.
641
We have covered each of these subjects completely. If you find yourself stuck on some
detail, please search the table of contents in this volume for a pointer to the relevant
section.
23.2
Injective Periodicity
23.2.1
Functions of the Integers
The kind of periodicity Shors algorithm addresses can be expressed in terms of functions defined over all the integers, Z, or over just a finite group of integers like ZM .
The two definitions are equivalent, but it helps to define periodicity both ways so we
can speak freely in either dialect. Here well consider all integers, and in the next
subsection well deal with ZM .
Weve discussed ordinary periodicity, like that of a function sin x, as well as (Z2 )n
periodicity, studied in Simons algorithm. Shors periodicity is much closer to ordinary
periodicity, but has one twist that gives it a pinch of Simons more exotic variety.
A function defined on Z,
f:
S,
S Z,
is called periodic injective if there exists an integer a > 0 (called the
period), such that
for all x 6= y in Z, we have
f (x) = f (y)
y = x + ka, some integer k.
The term injective will be discussed shortly. Because of the if and only if () in
the definition, we dont need to say smallest or unique for a. Those conditions
follow naturally.
Figure 23.4: Graph of two periods of a periodic injective function
Caution: Some authors might call f a-periodic to make the period visible,
omitting the reference to injectivity. Others, might just call f periodic and let you
fend for yourselves.
In theory, S, the range of f , can be any set: integers, sheep or neutrinos. Its the
periodicity that matters, not what the functional values are. Still, we will typically
consider the range to be a subset of Z, most notably, Z2r (for some positive integer
r). This will make our circuit analysis clearer.
642
23.2.2
Functions of the Group Z M
A little consideration of the previous definition should convince you that any periodic
injective function with period a can be confined to a finite subset of Z which contains
the interval [0, a). To feel f s periodicity, though, wed want M to contain at least
a few copies of a inside it, i.e., we would like M > 3a or M > 1000a. It helps if
we assume that we do know such an M , even if we dont know the period, a, and let
f be defined on ZM , rather than the larger Z. The definition of periodic injective in
this setting would be as follows.
A function defined on ZM ,
f:
ZM
S,
S ZM ,
is called periodic injective if there exists an integer a ZM (called the
period), such that
for all x 6= y in ZM , we have
f (x) = f (y)
y = x + ka, some integer k.
As before, S can be any set, but well be using Z2r .
The phrase
y = x + ka
uses ordinary addition, not mod-M addition. We are not saying that we can let
x + ka wrap around M back to 0, 1, 2, . . . and find more numbers that are part of
the f (x) = f (y) club. For example, the function
f (x) = x % 8
is periodic injective with period 8 according to either definition. However, when we
look at the second definition and take, say, M = 20 as the integer that we proclaim
to be larger than a defining the domain Z20 , then only the numbers { 3, 11, 19 }
which are of the form 3 + k8 should be considered pre-images of 3. Were we were
to allow 3 + k8 (mod 20), then the members of this club would grow (illogically) to
{3, 11, 19, 7, 15 }. This does not track; while f (3) = f (11) = f (19) = 3, it is
not true of f (7), or f (15), both of which are 7.
23.2.3
Discussion of Injective Periodicity
The term injective is how mathematicians say 1-to-1. Also, periodic injective is
seen in the quantum computing literature, so I think its worth using (rather than the
rag-tag 1-to-1-periodicity, which is a hyphen extravaganza). But what, exactly, are
we claiming to be injective? A periodic function is patently non-injective, mapping
643
multiple domain points to the same image value. Where is there 1-to-1-ness? It
derives from the following fact which is a direct consequence of the definition:
The if-and-only-if () condition in the definition of periodic injective implies
that, when restricted to the set [0, a) = { 0, 1, 2, . . . , a 1 }, f is 1-to-1. The same
is true of any set of <= a consecutive integers in the domain.
[Exercise. Prove it.]
Figure 23.5: Example of a periodic function that is not periodic injective
This is the twist I promised had some kinship to Simons periodic functions. Recall
that they were also 1-to-1 on each of their two disjoint cosets R and Q in that conversation. The property is required in our treatment of Shors period-finding as well; we
must be able to partition f s domain into disjoint sets on which f is 1-to-1. This is
not to say that the property is necessary in order for Shors algorithm to work (there
may be more general results that work for vanilla flavored periodicity). However, the
majority of historical quantum period-finding proofs make use of injective periodicity, whether they call it that, or not. For factoring and encryption-breaking, that
condition is always met.
While were comparing Simon to Shor, lets discuss a difference. In Simons case,
if we found even one pair of elements, x0 6= y 0 with f (x0 ) = f (y 0 ), then we knew
a. However, the same cannot be said of Shors problem. All we would know in the
current case is that x0 and y 0 differ by a multiple of a, but we would know neither a
nor the multiple.
23.3
Shors Periodicity Problem
Statement of Shors Periodicity Problem
Let f : ZM Z be injective periodic.
Find a.
The M in the problem statement is not about the periodicity (a describes that) as
much as it is about how big the problem is; it gives us a bound on a. It is M which
is used to measure the computational complexity of the algorithm; how does the
{algorithm + circuit} grow as M gets larger?
644
Relativized vs. Absolute Speed-Up
I feel compelled to say this again before we hit-the-road. Our goal is to produce a
quantum algorithm that completes in polynomial time when its classical counterpart
has exponential complexity. The time complexity of Shors period-finding algorithm
will end up being limited by both the circuits/algorithms we build around our oracle
Uf , as well as f , itself. So if f cannot be computed in polynomial time, at least
by some quantum circuit, we wont end up with an easy algorithm. We will prove
all the circuits and algorithms around Uf to be of polynomial complexity, specifically O(log3 M ). We will also show (in a future lesson) that the f needed for RSA
encryption-breaking is O(log4 M ), so the factoring problem is quantum-easy compared
with a hard problem using classical circuits/algorithms.
The point is that there are two problems: period-finding, which has quantum
relativized speed-up and factoring, which has quantum absolute speed-up. Our job is
to make sure that the quantum machinery around Uf is polynomial, so that when
f , itself, is polynomial (as is the case in factoring) we end up with a problem that is
absolutely easy in the quantum realm.
One Convenient Assumptions
In the development of the algorithm it will also help to add the very weak assumption
a < M/2,
i.e., the periodicity cycles at least twice in the interval [0, M 1].
Figure 23.6: We add the weak assumption that 2(+) a-intervals fit into [0, M )
In fact, when we apply Shors periodicity algorithm to RSA encryption-breaking or
the factoring problem, this added assumption will automatically be satisfied. Usually,
we have an even stronger assumption, namely that a M .
In the general case, well see that even if we are only guaranteed that a .9999M ,
we still end up proving that Shors problem is solvable in polynomial time by a
quantum circuit. But theres no reason to be so conservative. Even a < M/2 is
overdoing it for practical purposes, yet it makes all of our estimates precise and easy
to prove without any hand-waving. Also, it costs us nothing algorithmically, as we
will also see.
645
Figure 23.7: Typical application provides many a-intervals in [0, M )
Figure 23.8: Our proof will also work for only one a interval in [0, M )
23.3.1
Definitions and Recasting the Problem
As stated, the problem has relatively few moving parts: an unknown period, a, a
known upper bound for a, M , and the injective periodicity property. To facilitate
the circuit and algorithm, well have to add a few more letters: n, N and r. Here are
their definitions.
A Power of 2, 2n , for the Domain Size
Let n be the exponent that establishes closest power-of-2 above M 2 ,
2n1 < M 2 2n .
Well use the integer interval [0, 1, . . . , 2n 1] as our official domain for f , and well
let N be the actual power-of-2,
N 2n .
Since M > 2a we are guaranteed that [0, N 1] will contain at least as many intervals
of size a within it.
[Youre worried about how to define f beyond the original domain limit M ? Stop
worrying. Its not our job to define f , just discover its period. We know that f is
periodic with period a, even though we dont know a yet. That means its definition
can be extended to all of Z. So we can take any size domain we want. Stated another
way, we assume our oracle can compute f (x) for any x.]
The reason for bracketing M 2 like this only becomes apparent as the plot unfolds.
Dont be intimidated into believing anyone could predict we would need these exact
646
Figure 23.9: N = 2n chosen so (N/2, N ] bracket M 2
bounds so early in the game. Even the pioneers certainly got to the end of the
derivation and noticed that these limits would be needed, then came back and added
them up front. Thats exactly how you will do it when you write up the quantum
algorithms that you discover.
The bracketing of M 2 , when written in terms of N , looks like
N
2
< M 2 N.
A Power of 2, 2r , for the Range Size
Also, without loss of generality, we assume that the range of f is a subset of Z2r for
some sufficiently large r > 0, i.e.,
ra(f ) [ 0, 2r 1 ], also written
0 f (x) < 2r .
Summarizing all this,
f : [ 0, 2n 1 ]
f : [ 0, N 1 ]
[ 0, 2r 1 ], also written
[ 0, 2r 1 ], N 2n .
The Equivalence of N and M in Time Complexity
Lets be certain that using N for the time complexity is the same as using M . Taking
the log (base 2) of the inequality that brackets M 2 , we get
N
2
log N log 2
log
< log M 2
log N ,
< 2 log M
log N .
or
The big-O of every expression in this equation will kill the constants and weaken <
to , producing
O(log N )
O(log M )
647
O(log N ) ,
and since the far left and far right are equal, they both equal middle, i.e.
O (log N )
O (log M ) .
Therefore, a growth rate of any polynomial function of these two will also be equal,
O (logp N )
O (logp M ) .
Thus, bracketing M 2 between N/2 and N allows us to use N to compute the complexity and later replace it with M . Specifically, well eventually compute a big-O of
log3 N for Shors algorithm, implying a complexity of log3 M .
23.3.2
Z2 )n CBS Connection
The Z N (Z
Were all full of energy and eager to start wiring this baby up, but there is one final
precaution we should take lest we find ourselves adrift in a sea of math. On the one
hand, the problem lives in the world of ordinary integer arithmetic. We are dealing
with simple functions and sums like,
f (18)
6 + 7
=
=
6, and
13 .
On the other hand, we will be working with a quantum circuit which relies on mod-2
arithmetic, most notably the oracles B register output,
|y f (x)ir ,
or, potentially more confusing, ordinary arithmetic inside a ket, as in the expression
|x + jain .
Theres no need to panic. The simple rule is that when you see +, use ordinary
addition and when you see use mod-2.
|y f (x)ir . Were familiar with the mod-2 sum and its use inside a ket,
especially when we are expressing Uf s target register. The only very minor
adjustment well need to make arises from the oracles B channel being r qubits
wide (where 2r is f s range size) instead of the same n qubits of the oracles A
channel (where 2n is f s domain size). Well be careful when we come to that.
|x + jain . As for ordinary addition inside the kets, this will come about when
we partition the domain into mutually exclusive cosets, a process that Ill
describe shortly. The main thing to be aware of is that the sum must not extend
beyond the dimension of the Hilbert space in which the ket lives, namely 2n .
Thats necessary since an integer x inside a ket |xin represents a CBS state,
and there are only 2n of those, |0in , . . . |2n 1in . Well be sure to obey that
rule, too.
Okay, Ive burdened you with eye protection, seat belts and other safety equipment, and I know youre bursting to start building something. Lets begin.
648
23.4
Shors Quantum Circuit Overview and the
Master Plan
23.4.1
The Circuit
The total circuit looks very much like Simons.
|0in
H n
QFT (N )
| {z }
(actual)
Uf
|0ir
| {z }
(conceptual)
[Note: As with Simon, I suppressed the hatching of the quantum wires so as to
produce a cleaner looking circuit. The A channel has n lines, and the B channel has
r lines, as evinced by the kets and operators which are labeled with the exponents
n, N and r.]
There are two multi-dimensional registers, the upper A register, and the lower
B register. A side-by-side comparison of Shors and Simons circuits reveals two
differences:
1. The post-oracles A register is processed by a quantum Fourier transform instead of a multi-order Hadamard gate.
2. Less significant is the size of the B register. Rather than it being an n-fold tensor
space of the same dimension, 2n , as the A register, it is an r-fold space, with
a smaller dimension, 2r . This reflects the fact that we know f to be periodic,
with period a, forcing the number of distinct image values of f to be exactly a
because of injective periodicity. Well, a < M < M 2 2n , so there you have it.
These images can be reassigned to fit into a vector space of dimension, smaller,
usually much smaller, than As 2n . (Remember that we dont care what the
actual images are sheep, neutrinos so they may as well be 0, 1, ..., 2r1 .) An
exact value for r may be somewhat unclear at this point all we know is that
it need never be more than n, and wed like to reserve the right to give it a
different value by using a distinct variable name.
23.4.2
The Plan
Initial State Preparation
We prepare the separable state |0in |0ir as input.
Data Channel. The top line uses a multi-dimensional Hadamard to turn its |0in
into a maximally mixed (and perfectly balanced) superposition that it passes on to
the oracles top input, thus setting up quantum parallelism.
649
Target Channel. The bottom line forwards its |0ir directly on to the quantum
oracles B register, a move that (we saw with Simon) anticipates an application of
the generalized Born rule.
At that point, we conceptually test the B register output, causing a collapse of
both registers (Born rule). Well analyze whats left in the collapsed A registers
output, (with the help of a re-organizing, QFT gate). Well find that only a very
small and special set of measurement results are likely. And like Simons algorithm,
we may need more than one sampling of the circuit to get an adequate collection of
useful outputs on the A-line, but itll come very quickly due to the probabilities.
Strategy
Up to now, Ive been comparing Shor to Simon. Theres an irony, though, when we
come to trying to understand the application of the final post-oracle, pre-measurement
gate. It was quite difficult to give a simple reason why a final Hadamard gate did
the trick for Simons algorithm (I only alluded to a technical lemma back then.) But
the need for a final QFT , as we have already seen, is quite easy to understand:
we want the period, so we measure in the Fourier basis to get the fundamental frequency, m. Measuring in the Fourier basis means applying a z basis-to-Fourier basis
transformation, i.e., a QFT . m gets us a and we go home early.
One wrinkle rears its head when we look at the spectrum of a periodic function,
even one that is pure in the sense described above. While the likely (or in some cases
only) measurement possibilities may be limited to a small subset {cm}a1
c=0 where
m = N/a is the frequency associated with the period a, we dont know which cm we
will measure; there are a of them and they are all about equally likely. Youll see why
we should expect to get lucky.
Figure 23.10: Eight highly probable measurement results, cm, for N = 128 and a = 8
So, while well know we have measured a multiple cm of the frequency, we wont know
which multiple. As it happens, if we are lucky enough to get a multiple c that has the
bonus feature of being relatively prime (coprime) to a, we be able to use it to find a.
A second wrinkle is that despite what Ive led you to believe through my pictures,
the likely measurements arent exact multiples cm of the frequency, m. Instead they
will be a values yc , for c = 0, 1, , (a 1) which are very close to cm. Well have
to find out how to lock-on to the nearby cm associated with our measured, yc . Still,
when we do, a c coprime to a will be the the most desirable multiple that will lead
to a.
650
Two Forks
As we proceed, well get to a fork in the road. If we take the right fork, well find an
easy option. However the left fork will require much more detailed math. Thats the
hard option. In the easy case the spectrum measurement will yield an exact multiple,
cm, of m that I spoke of above. The harder, general case, will give us only a yc close
to cm. Then well have to earn our money and use some math to hop from the yc on
which we landed to the nearby cm that we really want.
Thats the plan. It may not be a perfect plan, but I think thats what I like about
it.
23.5
The Circuit Breakdown
Well segment the circuit into the familiar sub-sections.
|0in
H n
QFT (N )
| {z }
(actual)
Uf
|0ir
| {z }
(conceptual)
Since many of the sections are identical to what weve done earlier, the analysis is
also the same. However, Ill repeat the discussion of those common parts to keep this
lecture somewhat self-contained.
23.6
Circuit Analysis Prior to Conceptual Measurement: Point B
23.6.1
The Hadamard Preparation of the A register
I suppose we would be well advised to make certain we know what the state looks like
at access point A before we tackle point B, and that stage of the circuit is identical
to Simons; it sets up quantum parallelism by producing a perfectly mixed entangled
651
state, enabling the oracle to act on f (x) for all possible x, simultaneously.
|0in
H n
QFT (N )
Uf
|0ir
Hadamard, H n , in H(n)
Even though we are only going to apply the 2n -dimensional Hadamard gate to the
simple input |0in , lets review the effect it has on any CBS |xin .
|xin
H n
1
2
n 1
n 2X
(1)x y |yin ,
y=0
where the dot product between vector x and vector y is the mod-2 dot product. When
applied to |0in , reduces to
n 1
n 2X
n
1
|yin ,
|0i
H n
2
y=0
or, returning to the usual computational basis notation |xin for the summation,
n 1
n 2X
n
1
n
|0i
|xin .
H
2
x=0
The output state of this Hadamard operator is the nth order x-basis CBS ket, |+in =
|0in , reminding us that Hadamard gates provide both quantum parallelism as well as
a z x basis conversion operator.
23.6.2
The Quantum Oracle
Next, we consider the oracle:
|0in
H n
QFT (N )
Uf
|0ir
|
{z
}
Quantum Oracle
652
The only difference between this and Simons oracle is the width of the oracles B
register. Today, it is r qubits wide, where r will be (typically) smaller than the n of
the A register.
|xin
|xin
Uf
|0i
|xin |0ir
23.6.3
|0 f (x)ir
Uf
|f (x)ir
|xin |f (x)ir
The Quantum Oracle on Hadamard Superposition Inputs
Next, we invoke linearity.
We remarked in Simons circuit that the oracles B register output is not simply
f applied to the superposition state, a nonsensical interpretation since f only has a
meaning over its domain, ZN . Instead, we apply linearity to the maximally mixed
superposition |0in going into the oracles top register and find that
2n 1
n X
|xi |0i
Uf
x=0
2n 1
n X
|xin |f (x)ir ,
x=0
which
see is a weighted sum of separable products, all weights being equal to
we
n
(1/ 2) . The headlines are these:
The output is a superposition of separable terms |xin
the generalized Born rule needs,
|in+r
|f (x)in
,
( 2)n
the kind of sum
|0inA |0 irB + |1inA |1 irB + + |2n 1inA |2n 1 irB .
An A-measurement of x would imply a B-state collapse to its (normalized)
partner, |f (x)in .
If we chose to measure the B register first, a similar collapse to its entangled
A partner would result, but in that scenario there would be (approximately)
N/a = m, not one, pre-images x for every f (x) value. We will study those
details shortly.
23.7
Fork-in-the Road: An Instructional Case Followed by the General Case
At this point, we will split the discussion into two parallel analyses:
653
1. an instructional/easy case that has some unrealistic constraints on the period,
a, and
2. the general/difficult case.
The easy case will give us the general framework that we will re-use in the general
analysis. Dont skip it. The general case requires that you have your mind already
primed with the key steps from the easy case.
23.8
Intermezzo Notation for GCD and Coprime
Well be using two classical concepts heavily when we justify Shors quantum algorithm, and well also need these result for time complexity estimation.
Basic Notation
We express that fact that one integer, c, divides another integer, a, evenly (i.e. with
remainder 0) using the notation
c a .
Also, we will symbolize the non-negative integers using the notation
Z0 .
Now, assume we have two distinct non-negative integers, a and b, i.e.,
a, b Z0 ,
a > b,
The following definitions are fundamental in number theory.
23.8.1
Greatest Common Divisor
gcd(a, b)
23.8.2
largest integer, c, with c a and c b.
Coprime (Relatively Prime)
If gcd(a, b) = 1 we say that a is coprime to b (or a and b are coprime).
b
a
is my shorthand for a is coprime to b.
b is my shorthand for a is not coprime to b.
a
654
23.9
First Fork: Easy Case (a N )
We
case in which a is a divisor of N = 2n (in symbols,
now consider the special
a N ). This implies a = 2l . Immediately, we recognize that theres really no need for
Figure 23.11: Easy case covers a N , exactly
a quantum algorithm in this situation because we can test for periodicity using classical means by simply trying 2l for l = 1, 2, 3, . . . , (n 1), which constitutes O(log N )
trials. In the case of factoring, to which well apply period-finding in a later lecture,
well see that each trial requires a computation of f (x) = y x (mod N ) O(log4 N ),
y some constant (to be revealed later). So the classical approach is O(log N ) relative
to the oracle, and O(log5 N ) absolute including the oracle, all without the help of a
quantum circuit. However, the quantum algorithm in this easy case lays the foundation for the difficult case that follows, so we will develop it now and confirm that QC
leads to at least O(log5 N ) classical complexity (well do a little better).
23.9.1
Partitioning the Domain into Cosets
Its time to use the fact that f is periodic injective with (unknown) period a to help
rewrite the output of the Oracles B register prior to the conceptual measurement.
Injective-periodicity tells us that the domain can be partitioned (in more than one
way) into many disjoint cosets of size a, each of which provides a 1-to-1 sub-domain
for f . Furthermore, because we are in the easy case, these cosets fit exactly into the
big interval [0, N ). Heres how.
Figure 23.12: [0, N ) is the union of distinct cosets of size a
655
[ 0, N 1 ]
[ 0, ma 1 ]
= [0, a 1] [a, 2a 1] [2a, 3a 1] [(m 1)a, ma 1]
=
R + a
R + 2a
R + (m 1)a,
where
R [0, a 1] = {0, 1, 2, ... , a 1},
a = period of f, and
m = N/a is the number of times a divides N = 2n .
Definition of Coset. R + ja is called the jth coset of R.
We rewrite this decomposition relative to a typical element, x, in the base coset R.
[ 0, N 1 ] =
0, 1, . . . x, . . . a 1
a, 1 + a, . . . x + a, . . . 2a 1
2a, 1 + 2a, . . . x + 2a, . . . 3a 1
(m 1)a, 1 + (m 1)a, . . . x + (m 1)a, . . . ma 1
m1
[n
o
x + ja x [0, a)
j=0
m1
[n
x + ja
oa1
x=0
j=0
23.9.2
Rewriting the Output of the Oracle
It seems like a long time since we saw the original expression for oracles output, so
lets write it down again. It was
2n 1
n X
1
|xin |f (x)ir .
2
x=0
Our new partition of the domain gives us a nice way to express this. Each element
x R has a unique partner in each of the cosets, R + ja, satisfying
x+a
x + 2a
..
f
.
7 f (x) .
x + ja
..
x + (m 1)a xR
656
Using this fact (and keeping in mind that N = 2n ) we only need to sum over the a
elements in R and include all the x + ja siblings in each terms A register factor,
N 1
1 X
|xin |f (x)ir
N
x=0
a1
1 X n
|xi + |x + ain + |x + 2ain + + |x + (m 1)ain |f (x)ir
N
x=0
m1
a1
X
X
1
|x + jain |f (x)ir
=
N
j=0
x=0
a1
m1
r X
X
m
1
=
|x + jain |f (x)ir .
N
m
x=0
j=0
I moved a factor of 1/ m to the right of the outer sum so we could see that
1. each term in that outer sum is a normalized state (there are
m CBS terms in
the inner sum, and each inner term has an amplitude of 1/ m) and
2. the common amplitude remaining on the outside produces a normalized state
overall (therep
are a normalized
terms in the outer sum, and each term has an
amplitude of m/N = 1/ a).
23.9.3
Implication of a Hypothetical Measurement of the B
register Output
Although we wont really need to do so, lets imagine what happens if we were to
apply the generalized Born rule now using the rearranged sum in which the B channel
now plays the role of Borns CBS factors and A channel holds the general factors.
|0in
H n
QFT (N )
Uf
|0ir
|
{z
Conceptual
As the last sum demonstrated, each B register measurement of f (x) will be attached
to not one, but m, input A register states. Thus, measuring B first, while collapsing
A, merely produces a superposition of m states in that register, not a single, unique
x from the domain. It narrows things down, but not enough to measure,
657
m1
a1
X
X
m
1
|x + jain |f (x)ir
N x=0
m
j=0
&
Here, &
m1
1 X
|x0 + jain
m j=0
means collapses to.
!
|f (x0 )ir
If after measuring the post-oracle B register we were to go on to measure the A
register, its resulting collapse would give us one of the m values, x0 + ja, but we
would have no way to extract a from that measurement; there is no real information
for us here, so we dont measure A yet.
Lets name the collapsed but unmeasured superposition state in the A register
|x0 in , since it is determined by the measurement f (x0 ) of the collapsed B register,
m1
X
1
|x0 in
|x0 + jain .
m
j=0
Motivation for Next Step
Stand back and youll see that weve accomplished one of the goals of our introductory
key idea section. The conceptual measurement of the B register leaves an overall
state in the A register in which all of the amplitudes are zero except for m that have
equal amplitude 1m . Furthermore, those non-zero terms are spaced at intervals of
a in the N -dimensional vector: this is a pure periodic vector with the same period
a as our function f . We have produced a vector whose DFT is akin to that shown
in Figure 23.13. All non-zero amplitudes of the DFT are multiples of the frequency
m, i.e., of the form cm, c = 0, 1, . . . , (a 1). (Due to an artifact of the graphing
software the 0 frequency appears after the array at phantom position N = 128.)
Figure 23.13: The spectrum of a purely periodic vector with period 8 and frequency
16 = 128/8
The Big Picture
This strongly suggests that we apply the QFT to the A register in order to produce
a state that looks like Figure 23.13.
658
The Details
Well get our ideal result if we can produce an A register measurement, cm, where c
is coprime to a. The following two thoughts will guide us.
The shift property of the QFT will turn the sum x+ja into a product involving
a root-of-unity,
QFT
(N )
|x x0 i
x0 x
N 1
1 X yx n
|yi .
N y=0
Sums of roots-of-unity are famous for canceling themselves out to produce a
whole lotta 0.
23.9.4
Effect of a Final QFT on the A Register
The much advertised QFT is applied to the conceptually semi-collapsed |x0 in at
the output of the oracles A register:
|0in
H n
QFT (N )
Uf
|0ir
The QFT , being a linear operator, distributes over sums, so it passes right through
the ,
m1
1 X
|x0 + jain
m
m1
1 X
QFT (N ) |x0 + jain
m
QFT (N )
j=0
j=0
The QFT of each individual term is
QFT
(N )
|x0 + jai
N 1
1 X (x0 + ja) y
=
|yin
N
y=0
N 1
1 X x0 y jay
=
|yin ,
N
y=0
659
so the QFT of the entire collapsed superposition is
QFT
(N )
m1
1 X
|x0 + jain
m
j=0
N 1
m1
1 X 1 X x0 y jay
|yin
m
N
j=0
y=0
N 1 m1
1 X X x0 y jay
|yin
mN
y=0
j=0
N 1
m1
1 X x0 y X jay
=
|yin
mN
j=0
y=0
Rather than invoke QFT s shift property, we actually re-derived it in-place.
Summary. This, then, is the superposition state and our preferred organization of that expression just prior to sampling the final A register output:
N 1
m1
X
X
1
x0 y
jay |yin
mN
y=0
j=0
We can measure it at any time, and we next look at what the probabilities say we
will see when we do.
Foregoing the B Register Measurement. Although we analyzed this under
the assumption of a B measurement, an A channel measurement really doesnt care
about a conceptual B channel measurement. The reasoning is the same as in
Simons algorithm. If we dont measure B first, the oracles output must continue to
carry the full entangled summation
m1
a1
r X
X
m
1
|x + jain |f (x)ir .
N
m
j=0
x=0
through the final QFT (N )
i
r
1 . This would add an extra outer-nested sum
1 X
|f (x)ir
x[0.a)
to our Summary expression, above, making it the full oracle output, not just that
of the A register. Even leaving B is unmeasured, the algebraic simplification we
get below will still take place inside the big parentheses above for each x, and the
probabilities wont be affected. (Also, note that an A register collapse to one specific
|x0 + jain will implicitly select a unique |f (x0 )i in the B register.) With this overview,
try carrying this complete sum through the next section if youd like to see its (non)
effect on the outcome.
660
23.9.5
Computation of Final Measurement Probabilities (Easy
Case)
We are now in an excellent position to analyze this final A register superposition and
see much of it disappear as a result of some of the properties of roots-of-unities that
we covered in a past lecture. After that, we can analyze the probabilities which will
lead to the algorithm. We proceed in five steps that will
1. identify a special set of a elements, C = {yc = cm}a1
c=0 of certain measurement
likelihood,
2. observe that each of yc = cm will be measured with equal likelihood,
3. prove that a random selection from [0, a 1] will be coprime-to-a 50% of the
time,
4. observe that a y = cm associated with c coprime-to-a will be measured with
probability 1/2, and
5. measure y = cm associated with c coprime-to-a with arbitrarily high confidence
in constant time complexity.
23.9.6
STEP I: Identify a Special Set of a Elements, C =
{yc }a1
c=0 of Certain Measurement Likelihood
After (conceptual) measurement/collapse of the B register to state |f (x0 )ir , the postQFT A register was left in the state:
m1
1 X
|x0 + jain
QFT (N )
j=0
1
mN
N 1
X
y=0
x0 y
m1
X
jay |yin
j=0
We look at the inner sum in parentheses in a moment. First, lets recap some facts
about .
N
was our primitive N th root of unity, so
N = 1.
Because m is the number of times a divides (evenly) into N ,
m = N/a,
661
we conclude
1 = N = am = ( a )m .
In other words, we have shown that
a = m
is the primitive mth root of unity. Using m in place of a in the above sum produces a
form that we have seen before (lecture Complex Arithmetic for Quantum Computing,
section Roots of Unity, exercise (d )),
m1
m1
m, if y 0 (mod m)
X
X
jy
jay
m
=
.
0, if y 6 0 (mod m)
j=0
j=0
This causes a vast quantity of the terms in the QFT output (the double sum) to
disappear: only 1-in-m survives,
m1
1 X
r
|x0 + jain
QFT (N )
m X
x0 y |yin .
N
y0
(mod m)
j=0
Lets think about how to quantify the property that y is 0 mod-m:
y = cm,
y 0 (mod m)
for some c = 0, 1, 2, ... , a 1 .
This defines the special set of size a which will be certain to contain our measured y:
C
23.9.7
{cm}a1
c=0 .
Step II: Observe that Each cm Will be Measured with
Equal Likelihood
Since the easy case assumes N = am, we know that
equation becomes
m1
1 X
|x0 + jain
m
QFT (N )
j=0
p
p
m/N = 1/a, so the last
a1
1 X x0 cm
|cmin ,
a
c=0
and we see that each cm is measured with probability 1/a.
662
Example
Consider a function that has period 8 = 23 defined on a domain of size 128 = 27 . Our
problem variables for this function become
n = 7,
N = 2n = 128 ,
a = 23 = 8
and
N
27
m =
= 3 = 16 .
a
2
Lets say that we measured the B register and got the value f (x0 ) corresponding to
x0 = 3. According to the above analysis the full pretested superposition,
127
1 X
|xi7 |f (x)ir
128
x=0
7
1 X 7
|xi + |x + 8i7 + |x + 16i7 + + |x + 120i7 |f (x)ir
128
x=0
when subjected to a B register measurement will collapse to
1 7
|3 i7 |f (3)ir =
|3i + |11i7 + |19i7 + + |123i7 |f (3)ir ,
16
leaving the following vector in the A register (which I have stacked so as to align
the 16 non-zero coordinates):
( 0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
.25 ,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0 ) .
We are interested in learning the period but cant really look at the individual coordinates of this vector, it being a superposition which will collapse unpredictably.
But the analysis tells us that its QFT will produce a vector with only eight non-zero
amplitudes,
r
r
m X
1 X
7
n
(128)
x0 y
QFT
|3 i
=
|yi
=
x0 y |yi7 ,
N
8
y0
y0
(mod m)
(mod 16)
663
each corresponding to a y which is a multiple of m = 16 having an amplitude-squared
(i.e., probability of collapse) = 1/8 = .125, all other probabilities being zero. Graphing
the square amplitudes of the QFT confirms this nicely. (See Figure 23.14.) So we do
Figure 23.14: Eight probabilities, .125, of measuring a multiple of m = 16
find that the only possible measurements in the frequency domain lie in the special
set
C
{ 0, 16, 2(16), 3(16), . . . , 7(16) } .
Match this with the graph to verify the spikes are at positions 16, 32, 48, etc. (Due
to an artifact of the graphing software the 0 frequency appears after the array at
phantom position 128.) If we can use this set to glean the frequency m = 16, we will
be able to determine the period a = N/m = 128/16 = 8.
Measuring along the frequency basis means applying the basis-transforming QFT
after the oracle and explains its presence in the circuit.
23.9.8
Step III: Prove that a Random Selection from [0, a1]
will be Coprime-to-a 50% of the Time
Specifically, we prove that the probability of a randomly selected c [0, a 1] being
coprime to a is 1/2 (in the easy case).
Recall that, in this easy case, N = 2n , and a|N , so, a = 2r must also be a
power-of-two.
a) = # coprimes to a [0, a 1] ,
P (c
a
# odds < 2r
,
2r
1
= . QED
2
=
23.9.9
since a = 2r
Step IV: Observe that a y = cm Associated with c
Coprime-to-a Will be Measured with Probability 1/2
Our next goal will be to demonstrate that we not only measure some value in the
set C = {cm} in constant time, but that we can expect to get a special subset B C
664
in constant time, namely those cm corresponding to c a (i.e., c coprime to a). This
will enable us to find m (details shortly), and once we have m we get the period a
instantly from a = N/m.
In Step I, we proved that the likelihood of measuring one of the special C =
{yc } = {cm} was 100%. Then, in Steps II and III, we demonstrated that
each of the cm will be measured with equal likelihood, and
the probability of selecting a number that is coprime to a at random from the
numbers between 0 and a 1 is (in this special easy case) 1/2, i.e., a constant,
independent of a.
We combine all this with the help of a little probability theory. The derivation may
seem overly formal given the simplicity of the probabilities in this easy case, but it
sets the stage for the difficult case where we will certainly need the formalism.
First we reprise some notation and introduce some new.
C
{cm}a1
c=0
a}
{ yb yb = bm C and b
(Note: B C.)
P (C)
P (B)
P (we measure some y C)
P (we measure some y B)
P (B|C)
P we measure some y B
given that
the measured y is known to be C
We would like a lower bound (specifically, 1/2) on the probability of measuring a
yc = cm which also has the property that its associated c is coprime to a. In symbols,
we would like to show:
Claim. P (B)
1/2, independent of a, M, N, etc.
Proof. We know that we will measure a y C with 100% certainty, so P (C) = 1.
We also showed that each cm was equally likely to be measured, so this is equivalent
to randomly selecting a c < a. Finally, we saw that randomly selecting a c < a would
also produce one that was coprime to a 50% of the time. So,
1
1
1 =
. QED
P (B) = P (B|C) P (C) =
2
2
We proved, in fact, P (B)
= 1/2, exactly. Of course this means that
a)
P ( c
a)
1 P (c
665
1
.
2
23.9.10
Step V: Observe that yc Associated with c Coprime
to a Will be Measured in Constant Time
Result from the CTC Theorem for Looping Algorithms
Step V follows immediately from a past result.
We have a circuit that we plan on using in an algorithm. Furthermore, our algorithm is going to loop, querying the circuit each pass of the loop, obtaining a
statistically independent result each time. It produces a number c, each pass, and
for the sake of terminology well consider c a to be a success (and a c which is not
coprime to a to be a failure). In the last subsection we proved that P (success)= 1/2,
independent of the size, N .
These are the exact conditions of the CTC theorem for looping algorithms that we
covered at the end of our probability
lesson. That theorem tells us that the algorithm
will succeed, i.e., yield a c a, in constant time, T , and the discussion that followed
told us what T would be,
log ()
T =
+ 1,
log (1 12 )
where is our desired (small) error tolerance and bxc is notation for the floor of x,
or the greatest integer x.
Result from First Principles
We can reproduce the effect of the CTC theorem without invoking it, and for practice
with probability theory lets go ahead and do it that way, too.
If we run the circuit T times, obtaining T samples,
mc1 , mc2 , . . . , mcT ,
the probability that none of the associated c were coprime to a is
a
P
c1 a c2 a . . . cT
T
1
, constant time, T , independent of N .
=
2
This means that the probability of at least one mc having a c a is:
T
1
P (at least one c a) = 1
.
2
Conclusion
After an adequate number of measurements (independent of a, M ), which produce
yc1 , yc2 , . . . , ycT , we can expect at least one of the yck = cm to correspond to c,
666
with c a that is, yck B with high probability. In a moment, well see why its
so important to achieve this coprime condition.
Example:
If we measure the output of our quantum circuit repeatedly, insisting
on yc c a with P > .999999, or error tolerance = 106 , we would need
log (106 )
6
T =
+ 1 =
+ 1 = 19 + 1 = 20
log (.5)
.30103
measurements.
We can instruct our algorithm to cycle 100 times, to get an even better confidence,
but if the data determined that only eight measurements were needed to find the
desired cm, then it would return with a successful cm after eight passes. Our hyperconservative estimate costs us nothing.
23.9.11
Algorithm and Complexity Analysis (Easy Case)
A Classical Tool: The Euclidean Algorithm
The Euclidean Algorithm (EA) is a famous technique from number theory takes two
input integers P, Q with P > Q in Z0 and produces the greatest common divisor of
P and Q, also represented by gcd(P, Q). This is the largest integer that divides both
P and Q, evenly (without remainder). EA(P, Q) will produce gcd(P, Q) in O(log3 P )
time. We will be applying EA to N and cm, to get m0 = gcd(N, cm) so the algorithm
will produce m0 with time complexity O(log3 N ) = O(log3 M ).
Well assume that EA(N, cm) delivers as promised and go into its inner workings
in a later chapter. EA is a crucial classical step in Shors algorithm as we see next.
Why Finding yc B Solves Shors Problem
We proved that we will measure yc B with near certainty after some predetermined
number, T of measurements. Why does such a yc solve Shors period-finding problem
(in the easy case)?
Here is our circuit, after measuring cm:
|0in
H n
QFT (N )
Uf
|0i
After the measurement, this is what we know and dont know:
Known Unknown
N
c
cm
m
a
667
&
|cmi
So the first thing we do is use EA to produce m0 = gcd(N, cm). Now, if cm B then
c a, and
a
c
m0 gcd(N, cm) =
gcd(am, cm) = m
That gives us the unknown m and from that we can get a:
a
N
.
m
But this only happens if we measure y B, which is why finding such a y will give
us our period and therefore solve Shors problem.
How do we know whether we measured an cm B? We dont, but we try m0 ,
anyway. We manufacture an a0 using m0 , by
a0
N
,
m0
hoping a0 = a, which only happens when m0 = m (and, after all, finding a, not m, is
our goal). We test this by asking whether f (x + a) = f (x) for any x and x = 1 will
do. a is the only number that will produce an equality here by our requirement of
injective periodicity. If it does, were done. If not, we try T 1 more times because,
with .999999 probability, well succeed after T = 20 times no matter how big M
(or N ) is.
As for the cost of the test f (x + a) = f (x), that depends on f . We are only
asserting relativized polynomial complexity for period-finding but know (by a future
chapter) that for factoring, f will be polynomial fast (O(log4 (M )), and maybe better).
We can now present the algorithm.
Shor-like Algorithm (Easy Case)
Select an integer, T , that reflects an acceptable failure rate based on any known
aspects of the period. E.g., for a failure tolerance of .000001, We might choose
T = 20
Repeat the following loop at most T times.
1. Apply Shors circuit.
2. Measure output of QFT and get cm.
3. Compute m0 = EA(N, cm), and set a0 N/m0 .
4. Test a0 : If f (1 + a0 ) = f (1) then a0 = a, (success) break from loop.
5. Otherwise continue to the next pass of the loop.
If the above loop ended naturally (i.e., not from the break ) after T full passes,
we failed. Otherwise, we have found a.
668
Computational Complexity (Easy Case)
We have already seen that O(log(N )) = O(log(M )) as well as any powers of the logs,
so I will use M in the following.
The Hadamard gates are O(log(M )).
The QFT is O(log2 (M )).
Complexity of Uf is same as that of f , which could be anything. We dont count
that here, but we will show in a separate lecture that for integer factoring, it is
certainly at least O(log4 M ).
The outer loop is O(T ) = O(1), since T is constant.
The classical EA sub-algorithm is O(log3 M ).
The four non-oracle components, above, are done in series, not in nested loops,
so the overall relativized complexity will be the worst case among them, O(log3 (M )).
In the case of factoring needed for RSA encryption breaking (order-finding) the
actual oracle is O(log4 (M )) or better, so the absolute complexity in that case
will be O(log4 (M ))
So the entire Shor circuit for an f O(log4 (M )) (true for RSA/order-finding) would
have an absolute complexity of O(log4 (M )). Notice that, while not an exponential
speed-up over the O(log5 (M )) easy case classical algorithm presented, it is faster by
a factor of log N . Thats due to the fact that the quantum circuit requires a constant
time sampling, while the classical function must be sampled O(log N ) times. To be
fair, though, the classical algorithm was deterministic, and the quantum algorithm,
probabilistic. You can debate whether or not this counts in your forums.
This completes the easy case fork-in-the-road. It contains all the way points that
we will need for the general case without the subtle and tricky math. You are now
well positioned to understand those additional details, so ... onward.
23.10
Second Fork: General Case (We do not Assume a N )
Now, a need not be a power-of-2, so the classical approach no longer admits an
O(log 5 N ) algorithm. Also, if we take the integer quotient, m = bN/ac, there will
usually be a remainder representing those excess integers that dont fit into a final,
full period interval.
669
Figure 23.15: There is (possibly) a remainder for N/a, called the excess
23.10.1
Partitioning the Domain into Cosets
Like the easy case, f s injective periodicity helps us rewrite the output of the oracles
B register prior to the conceptual measurement. The domain can still be partitioned
into many disjoint cosets of size a, each of which provides a 1-to-1 sub-domain for f ,
but now there is a final coset which may not be complete:
Figure 23.16: [0, N ) is the union of distinct cosets of size a, except for last
Express [ 0, N 1 ] as the union of a union,
[ 0, N 1 ]
excess
}|
{
z
= [0, a 1] [a, 2a 1] [(m 1)a, ma 1] [ ma, N 1 ]
R + a
670
R + (m 1)a,
partial
}| {
z
R + ma
where
R [0, a 1] = {0, 1, 2, ... , a 1},
a = period of f, and
m = bN/ac, the number of times a divides (unevenly) into N.
Cosets. As before, R + ja is called the jth coset of R, but now we have
R + ma [R + ma] ,
the partial mth coset of R from ma to N 1.
Figure 23.17: The final coset may have size < a
We express the decomposition of our entire domain relative to a typical element x in
the base coset R,
m1
[ 0, N 1 ]
[n
x + ja
oa1
x + ma
oN ma1
x=0
x=0
j=0
but now we had to slap on the partial coset, { x + ma }, to account for the possible
overflow.
Notation to Deal with the Partial Coset
We have to be careful about counting the family members of each element x R, i.e.,
those x + ja who map to the same f (x) by periodicity. We sometimes have a member
in the last, partial, coset, and sometimes not. If x is among the first few integers of
R, i.e., [0, N ma), then there will be m + 1 partners (including x) among its kin.
However, if x is among the latter integers of R, i.e., [N ma, a), then there will be
only m partners (including x) among its kin.
Well use m
e to be either m or m + 1 depending on x,
(
m + 1, for the first few x in [0, a 1]
m
e =
m, for the remaining x in [0, a 1]
671
Figure 23.18: If 0 x < N ma, a full m + 1 numbers in ZN map to f (x)
Figure 23.19: If N ma x < a, only m numbers in ZN map to f (x)
23.10.2
Rewriting the Output of the Oracles output
The original expression for oracles B register output was
2n 1
n X
|xin |f (x)ir ,
x=0
and our new partition of the domain gives us a nice way to rewrite this. First, note
that
x+a
x + 2a
..
.
f
7 f (x) .
x + ja
..
x + (m 1)a
and sometimes . . .
x + ma
xR
Now we make use of our flexible notation, m,
e to keep the expressions neat without
sacrificing precision of logic:
r ! X
N 1
a1
m1
e
X
X
m
e
1
1
|xin |f (x)ir =
|x + jain |f (x)ir ,
N
N
m
e
x=0
x=0
j=0
where (to repeat for emphasis),
(
m + 1, x [ 0, N ma )
m
e = m(x)
e
=
m,
x [ N ma, a )
672
The factor of 1/ m
e inside the sum normalizes each term in the outer sum. However,
the common amplitude remaining on the outside is harder to symbolize in a formula,
which is why I used to describe it. (m
e doesnt even make good sense outside the
sum, but it gives us an idea of what the normalization factor is.) It turns out that
q we
pm
dont care about its exact value. It will be some number between N and m+1
,
N
the precise value being whatever is needed to normalize the overall state.
23.10.3
Implication of a Hypothetical Measurement of the B
register Output
Weve seen that it helps to imagine a measurement of the oracles B registers output.
|0in
H n
QFT (N )
Uf
|0ir
|
{z
Conceptual
Each B register measurement of f (x) will be attached to not one, but m,
e input A
register states. The generalized Born rule tells us that measuring B will cause the collapse of A into a superposition of m
e CBS states, narrowing things down considerably.
r
m1
e
a1
X
X
m
e
1
|x + jain |f (x)ir
N x=0
m
e
j=0
&
Here, &
m1
e
1 X
|x0 + jain
m
e j=0
means collapses to.
!
|f (x0 )ir
If after measuring the post-oracle B register we were to go on to measure the A
register, it would collapse, giving us a reading of one of the m
e values, x0 + ja, but
that value would not get us any closer to knowing a, so as with the easy case, we dont
measure A yet. Instead, we name the collapsed but unmeasured superposition
state in the A register |x0 in , since it is determined by the measurement f (x0 ) of
the collapsed B register,
m1
e
X
1
|x0 in
|x0 + jain
m
e
j=0
Foregoing the B Register Measurement. By now we have twice seen why
we can avoid this conceptual measurement yet still analyze the A register as if it had
673
occurred. There is no harm in pretending we measured and collapsed to a |f (x0 )ir
first. (See the Easy Case or our lesson Simons Algorithm for longer explanations of
the same argument.)
Motivation for Next Step
The conceptual measurement of the B register leaves an overall state in the A register
in which all the amplitudes are zero except for m
e of them which have amplitude 1m
.
e
Furthermore, those non-zero terms are spaced at intervals of a in the N -dimensional
vector: this is a pure periodic vector with the same period a as our function f .
In contrast to the easy case, however, m
e is sometimes m and sometimes m + 1,
never mind that a is not a perfect power-of-2 or that it doesnt divide into N evenly.
All this imperfection destroys the arguments made in the easy case.
A look at the DFT of the vector whose components match the amplitudes of
a typical collapsed state left in the A register (see Figure 23.20) confirms that a
frequency domain measurement of QFT (N ) |x0 in no longer assures us of seeing one
of the a numbers cm0 , c = 0, . . . a1, with m0 the true and typically non-integer
frequency N/a. As for our integer quotient m bN/ac close to the actual frequency
N/a, its possible that none of the frequency domain points cm have high amplitudes.
Figure 23.20: The spectrum of a purely periodic vector with period 10 and frequency
12.8 = 128/10
The situation appears grim.
Lets look at the bright side, though. This picture of a typical DFT applied to an
N dimensional vector, 0 except for amplitudes 1m
at m
e time domain points, suggests
e
that there are still only a frequencies, y0 , y1 , . . . , ya1 which have large magnitudes.
And theres even more reason for optimism. Those likely yk appear to be at least
close to multiples of the integer-ized frequency m, i.e., they are near frequency
domain points of the form cm. (Due to an artifact of the graphing software the 0
frequency appears after the array at phantom position N = 128.)
674
The Big Picture
It still seems to be a good idea to apply a QFT to the A register in order to produce
a state that looks like Figure 23.20.
The Key Steps
Well show that there are only a likely A register measurements, yc , c = 0, 1, . . . , (a
1). And even if a given yc 6= cm exactly, it will at least lead us to a cm. In fact, we
can expect the resulting c to be relatively prime to a with good probability, a very
desirable outcome. Here are the general steps.
We will identify a special set of a elements, C = {yc }a1
c=0 of high measurement
likelihood.
We will show that each yc is very close to in fact, uniquely selects a point of
the form cm. Well describe a fast (polynomial time) algorithm that takes any
of the likely measured yc to its unique partner cm.
Well prove that we can expect to get a yc (and thus cm) with c coprime to
the period a in constant time. Such a c will unlock the near-frequency, m, and
therefore the period, a.
23.10.4
Effect of a Final QFT on the A Register
The QFT is applied to the conceptually semi-collapsed |x0 in at the output of the
oracles A register:
|0in
H n
QFT (N )
Uf
|0ir
The linear QFT passes through the ,
m1
e
1 X
|x0 + jain
m
e
QFT (N )
j=0
m1
e
1 X
QFT (N ) |x0 + jain
m
e
j=0
Note that it doesnt matter that m
e is sometimes m and sometimes m + 1. Once
it collapses, it becomes one of the two depending on which x0 is selected by the
measurement, and the formulas all still hold true. The QFT of each individual term
675
is
QFT
(N )
N 1
1 X (x0 + ja) y
=
|yin
N
|x0 + jai
y=0
N 1
1 X x0 y jay
|yin ,
=
N
y=0
so the QFT of the entire collapsed superposition is
QFT
(N )
m1
e
1 X
|x0 + jain
m
e
j=0
m1
e
N 1
1 X 1 X x0 y jay
|yin
N
m
e
j=0
1
mN
e
y=0
N 1 m1
e
XX
y=0
j=0
N 1
1
mN
e
x0 y jay |yin
x0 y
y=0
m1
e
X
jay |yin
j=0
1
In this expression, the normalizing factor mN
is precise. Thats in contrast to the
e
pre-collapsed state in which we had an approximate factor outside the full sum. The B
register measurement picked out one specific x0 , which had a definite m
e associated
with it. Whether it was m or m + 1 doesnt matter. It is one of the two, and that
value is used throughout this expression.
Summary. This is the preferred organization of our superposition state prior
to sampling the final A register output:
N 1
m1
e
X
X
1
x0 y
jay |yin .
mN
e
y=0
j=0
The next several sections explore what the probabilities say we will see when we
measure this state. And while we analyzed it under the assumption of a prior B
measurement, the upper (A) channel measurement wont care about that conceptual
measurement, as well see. We continue the analysis as if we had measured the B
channel.
676
23.10.5
Computation of Final Measurement Probabilities (General Case)
This general case, which I scared you into thinking would be a mathematical horror
story, has been a relative cakewalk so far. About all we had to do was replace the
firm m with the slippery m,
e and everything went through without incident. Thats
about to change.
In the easy case, we were able to make the majority terms in our sum vanish (all
but 1-in-m). Lets review how we did that. We noted that N , the primitive
N th root, so
N = 1 .
Then we replaced N with ma, to get
ma = 1
and realized that this implied that a was a primitive mth root of unity. From there
we were able get massive cancellation due to the facts we developed about sums of
roots-of-unity.
The problem, now, is that we cannot replace N with ma. We have an m,
e but
even resolving that to m or m + 1 wont work, because neither one divides N evenly
(by the general case hypothesis). So well never be able to manufacture an mth root
of unity at this point and cannot watch those big sums dissolve before our eyes. So
sad.
We can still get what we need, though, and have fun with math, so lets rise to
the challenge.
As with the easy case our job is to analyze the final (post QFT ) A register
superposition. While none of the terms will politely disappear the way they did in
the easy case, we will find that certain y states will be much more likely than others,
and this will be our savior.
Computing the final measurement probabilities will require the following five steps.
1. identify (without proof ) a special set of a elements, C = {yc }a1
c=0 of high measurement likelihood,
2. prove that the values in, C = {yc }a1
c=0 have high measurement likelihood,
a1
3. associate {yc }a1
c=0 with {c/a}c=0 ,
4. describe an O(log3 N ) algorithm that will produce c/a from yc , and
5. observe that yc associated with c coprime to a will be measured in constant
time.
677
23.10.6
STEP I: Identify (Without Proof ) a Special Set of
a Elements, C = {yc }a1
c=0 of High Measurement Likelihood
In this step, we will merely describe the subset of y that we want to measure. In the
next step, well provide the proof.
In the easy case we measured y, which had the special form
y = cm,
c = 0, 1, 2, ... , a 1,
with 100% certainly in single measurement. From there we tested whether the c was
coprime to a, (which it was with high probability), and so on. This time we cant be
100% sure of anything even after post-processing with the QFT , but thats normal
for quantum algorithms we often have to work the numbers and be satisfied to
get what we want in constant or polynmoial time. I claim that in the general case we
will measure an equally small subset of y, again a in all, that we label
y = yc ,
c = 0, 1, 2, ... , a 1
with probability P = 1 in O(1) measurements. (Whenever you see me use a ,
it represents a positive number which is as small as we want, say .000000001, for
example.) The following construction will lead us to these special a values.
Consider the (very) long line [ 0, aN 1 ].
Figure 23.21: A very long line consisting of a copies of N = 2n
Around the a integral points on this line,
0, N, 2N, , cN, , (a 1)N,
place (relatively small) half-open intervals of width a.
h a
h
a
a
a
, +
,
N , N+
,
2
2 h
2
2
a
a
cN , cN +
2h
2
a
a
(a 1)N , (a 1)N +
2
2
Each interval contains exactly one integral multiple, ya, of a in it. Well label the
multiplier that gets us into the cth interval yc . (y0 is easily seen to be 0.)
h a
h
a
a
a
0a , +
, , yc a cN , cN +
,
2
2
2
2
h
a
a
,
ya1 a (a 1)N , (a 1)N +
2
2
678
Figure 23.22: Half-open intervals of width a around each point cN
Figure 23.23: Exactly one integral multiple of a falls in each interval
Since yc a < aN , for all c = 0, 1, , (a 1), we are assured yc < N , so it is a possible
result of our measurement after the QFT .
Claim. The {yc } just described are the a values most likely to be measured.
Well see that with probability P = 1 , arbitrarily close to 1, we will measure one
of these yc in constant time, T , independent of N . (The proof appears in the next
step. In this step we only want to establish our vocabulary and goals.)
Example 1
In this first of two examples we illustrate the selection of yc for an actual periodic
function defined on integers [0, N 1], where N = 25 = 32, and the period a = 3.
(Remember, it doesnt matter what the values f (k) are, so long as were told it is
periodic.)
c = 0: The center of the interval is 0N = 0 32 = 0. We seek y0 such that
3y0 [ 1.5, 1.5 )
y0 = 0,
since
3 0 = 0 [ 1.5, 1.5 ) .
c = 1: The center of the interval is 1N = 1 32 = 32. We seek y1 such that
3y1 [ 30.5, 33.5 )
y1 = 11,
since
3 1 = 33 [ 30.5, 33.5 ) .
679
c = 2: The center of the interval is 2N = 2 32 = 64. We seek y2 such that
3y2 [ 62.5, 65.5 )
y2 = 21,
since
3 21 = 63 [ 62.5, 65.5 ) .
Etc.
Example 2
In this second example, we test the claim that the yc so described really do have high
relative measurement likelihood for an actual periodic function.
Consider a function that has period 10 defined on a domain of size 128 = 27 . Our
problem variables for this function become
n = 7,
N = 2n = 128 ,
a = 10
and
N
128
m =
=
= 12 .
a
10
Lets say that we measured the B register and got the value f (x0 ), corresponding to
x0 = 3. For this x,
m
e = m + 1 = 13 ,
since
x0 + ma =
3 + 12 10
123
< 128 .
According to the above analysis, the full pretested superposition,
127
1 X
|xi7 |f (x)ir
128
x=0
sometimes
7
z }| {
1 X 7
|xi + |x + 10i7 + |x + 20i7 + + |x + 120i7 |f (x)ir
128
x=0
when subjected to a B register measurement will collapse to
1 7
|3 i7 |f (3)ir =
|3i + |13i7 + |23i7 + + |123i7 |f (3)ir ,
13
leaving the following vector in the A register (which I have stacked so as to align
the 13 non-zero coordinates):
680
( 0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
.27735 ,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0, 0,
0 ) .
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
We are interested in learning f s period and, like the easy case, the way to get at it is
by looking at this state vectors spectrum, so we take the QFT . Now, unlike the easy
case, this vectors QFT will not create lots of 0 amplitudes; generally all N = 128 of
them will be non-zero. Thats because the resulting sum
N 1
m1
e
X
X
1
x0 y
jay |yin .
QFT (128) |3 i7 =
mN
e
y=0
j=0
did not admit any cancellations or simplification. Instead, the above claim which
we will prove in the next section is that for x0 = 3 only m
e = 13 of them will be
likely, the special {yc }, for c = 0, 1, 2, . . . , 12 we described in our last analysis. Lets
put our money where our mouth is and at least show this to be true for the one
function under consideration.
We take three yc as examples: y4 , y5 and y6 . Well do this in two stages. First
we identify the three yc values. Next, we graph the probabilities of QFT (128) |3 i7
around those three values to see how they compare with nearby y values.
Stage 1. Compute y4 , y5 and y6
c = 4: The center of the interval is 4N = 4 128 = 512. We seek y4 such that
10y4 [ 507, 517 )
y4 = 51,
since
10 51 = 510 [ 507, 517 ) .
c = 5: The center of the interval is 5N = 5 128 = 640. We seek y5 such that
10y5 [ 635, 645 )
y5 = 64,
since
10 64 = 640 [ 635, 645 ) .
c = 6: The center of the interval is 6N = 6 128 = 768. We seek y6 such that
10y6 [ 763, 773 )
y6 = 77,
since
10 77 = 770 [ 763, 773 ) .
681
2
7
(128)
Stage 2. Look at the Graph of
QFT
|3 i
Around These Three y
Here is a portion of the graph of the QFT s absolute value squared showing
the probabilities of measuring over 35 different y value in the frequency domain. It
exhibits in a dramatic way how much more likely the yc are to be detected than
are the non-yc values. (See Figure 23.24.) Even though the non-yc have non-zero
Figure 23.24: Probabilities of measuring y4 = 51, y5 = 64 or y6 = 77 are dominant.
probabilities of measurement, they barely register except for a few.
If we apply the math to all 10 yc for this x0 (which, remember is 3), we find that
the measurements in the frequency domain will usually yield values that lie in the
special set
C
{ 0, 13, 26, 38, 51, 64, 77, 90, 102, 115 }
[Exercise. Show this for the yc not computed in our example.]
Measuring along the frequency basis means applying the basis-transforming QFT
after the oracle and explains its presence in the circuit.
Introducing the ybc
To assist our analysis well define
a setof ybc associated with the desired yc such that
the ybc all live in the interval a2 , + a2 (Figure 23.25).
We
ado this
by subtracting cN from each yc a to bring it into the base interval
a
2 , + 2 . More precisely,
h a
a
ybc yc a cN , +
, c = 0, 1, 2, ..., a 1 ,
2
2
which can be stated using modular arithmetic by
ybc = yc a
(mod N ) .
While we will be looking to measure the yc , it
be important to remember that
will
a
they correspond ( mod N ) to these ybc 2 , + a2 .
682
Figure 23.25: ybc all fall in the interval [a/2, a/2)
This is the vocabulary we will need. Next, well take a lengthy but leisurely
mathematical cruise to prove our claim of high probability measurement for the set
C.
23.10.7
STEP II: Prove that the Values in, C = {yc }a1
c=0 Have
High Measurement Likelihood
We gave a convincing argument that the a distinct frequency domain values {yc }
constructed in the last section do produce highly likely measurement results, but we
didnt even try to prove it. Its time to do that now. This is the messiest math of the
lecture. Ill try to make it clear by offering pictures and gradual steps.
When we last checked, the (conceptual) measurement/collapse of B register to
state |f (x0 )i left the post-QFT A register in the state
N 1
m1
e
X
X
1
x0 y
jay |yin .
mN
e
y=0
j=0
The probability of measuring the state |yi is the amplitudes magnitude squared:
P (measurement yields y)
Letting
ay ,
683
2
m1
e
X
1
x0 y 2
jay
| |
mN
e
j=0
2
m1
e
X
1
jay
mN
e
j=0
the summation factor on the right (prior to magnitude-squaring) is
m1
e
X
e
1 + + 2 + + m1
e
m
1
.
1
j=0
Having served its purpose, can now be jettisoned, and
e
m
1
1
2i
e
aym
1
=
ay
1
iy m
e
e
1
,
eiy 1
eN
e
ay m
e
2i
ay
N
where we are defining the angle
y
2ay
.
N
In other words, the probability of measuring any y is
P (measurement yields y)
2
e
1 eiy m
1
eiy 1
mN
e
The next several screens are filled with the math to estimate the magnitude of the
fraction,
iy m
e e 1
eiy 1 .
We will do it by bounding numerator and denominator, separately. Its a protracted
side-trip because I go through each step slowly, so dont be intimidated by the length
of the derivation.
A General Bound for ei 1
It will help to bracket the expression:
?
i
e 1
An Upper Bound for ei 1
In the complex plane, ei 1 is the length of the chord from 1 to ei , and this is
always the arc length from 1, counterclockwise, to ei , all on the unit circle. But
the arc length is, by definition, the (absolute value of the) angle, , itself, so:
i
e 1 || .
684
Figure 23.26: The chord is shorter than the arc length
A Lower Bound for ei 1
By the Euler identity
1 ei
1 (cos + i sin ) .
By the addition laws for sine and cosine applied to = 2 + 2 ,
2
2
sin
cos = cos
2
2
sin = 2 cos
sin
2
2
So
i
1 e
=
=
2
1
cos
sin
2
2
+
i 2 cos
sin
.
2
2
2
2
1 cos
+ sin
2
2
i 2 cos
sin
2
2
2
i 2 cos
sin
2 sin
2
2
2
2 sin
sin
i cos
2
2
2
2
Noting that
2
sin
i
cos
= 1,
2
2
we have shown that
i
e 1
2 sin
.
2
685
Figure 23.27: |sin(x/2)| lies above |x/| in the interval ( , )
By simple calculus and solving / = sin (/2) for , or noting where the graph of the
sine curve and the line intersect, we conclude that when is in the interval [, ],
we can bound 2 |sin (/2)| from below,
2 ||
,
for [, ] .
2 sin
2
This gives us a lower bound for ei 1 :
2 ||
i
e 1 ,
for [, ] .
Combining both bounds when is in the interval [, ], we have bracketed the
expression under study,
2 ||
i
e 1
|| ,
for [, ] .
Applying Both Bounds to the Fraction
Our goal was to estimate
iy m
e e 1
eiy 1 ,
especially when y = yc .We prefer to work with the ybc which are in an a-sized interval
around 0:
h a
a
ybc = yc a cN , +
, c = 0, 1, 2, ..., a 1
2
2
686
So, lets first convert the exponentials in the numerator and denominator using the
above relation between those special {yc } (which we hope to measure with high probability) and their corresponding mod N equivalents, the {b
yc }, close to the origin.
ei yc
= ei (2ayc /N )
= ei (2byc /N )
= ei (2 [ ybc +cN ]/ N )
= ei (2byc /N ) ei 2 c
= ei ybc /a
This allows us to rewrite the absolute value of the amplitudes we wish to estimate as
iy m
i m/a
e e 1
ybc e 1
= e
eiy 1
eiybc /a 1 .
We want a lower bound for this magnitude. This way we can see that the likelihood
of measuring these relatively few ys is high. To that end, we get
a lower bound for the numerator, and
an upper bound for the denominator.
Upper Bound for Denominator
The denominator is easy. We derived an upper bound for all angles, , so:
i /a
e ybc 1 ybc .
a
Lower Bound for Numerator
The numerator is
i m/a
e ybc e 1
i (2by / N ) m
e
c
e
1 .
Remember that m
e is sometimes m and sometimes m + 1.
f is m
Sub-Case 1: m
When m
e is m, we have things easy: 2mb
yc / N (, ) because
a
a
ybc <
,
2
2
or
am
2m
am
ybc . <
.
N
N
N
And, since
am
< 1,
N
687
we get
<
2m
ybc . < .
N
This allows us to invoke the lower bound we derived when [, ], namely
i 2mby / N
c
e
1
2 |2mb
yc / N |
.
Re-applying the notation
y
2ay
,
N
we write it as
i 2my / N
c
e
1
2 m
y
b
c
a
.
Combining the bounds for the numerator and denominator (in the case where m
e = m),
we end up with
m ,
i m/a
e ybc e 1
ybc
2 a ybc
= 2m
eiybc /a 1
a
f is m + 1 (Deferred)
Sub-Case 2: m
That only worked for m
e = m; the bounds argument we presented wont hold up when
m
e = m+1, but well deal with that kink in a moment. Lets pretend the above bound
works for all m,
e both m and m + 1, and finish computing the probabilities under this
white lie. Then well come back and repair the argument to also include m
e = m + 1.
f = m Bounds
Simulating the Probabilities Using the m
If the above bounds worked for all m,
e we could show that this leads to a desired
lower bound for our probabilities of getting a desired y C in constant time using
the following argument.
For any y, we said
P (measurement yields y)
2
e
1
1 eiy m
eiy 1 ,
mN
e
but for y = yc , one of the (a) special ys that lie in the neighborhoods of the cN s,
we can substitute our new-found lower bound for the magnitude of the fraction.
(Remember, we are allowing that the bound holds for all m,
e even though we only
proved it for m
e = m), so
P (measurement yields yc )
688
1 4m
e2
mN
e
2
m 4
N 2
The last inequality holds because some of the m
e are m + 1 and some are m. Now,
there are a such yc , namely, y0 , y1 , , ya1 , each describing a distinct, independent,
measurement (we cant get both y32 and y81 ). Thus, the probability of getting any one
of them is the sum of the individual probabilities. Since those individual probabilities
are all bounded below by the same constant, we can multiply it by a to get the
collective lower bound,
am 4
N 2
P (measurement yields one of the yc )
[Exercise. Make this last statement precise.]
In this hard case, allowing a|N , we defined m to be the unique integer satisfying
ma N < (m + 1)a, or quoting only the latter inequality,
(m + 1)a
>
N.
Its now time to harvest that weak additional assumption requiring at least two
periods of a to fit into M ,
a
<
M
,
2
<
N
.
2
which also implies
We combine those to conclude
am
N
>
1
.
2
[Exercise. Do it.]
We can plug that result into our probability estimate to get
P (measurement yields one of the yc )
>
2
.
2
Note-to-file. If a << M < N , as is often the case, then
am
N
1,
and the above lower bound improves to
P (measurement yields one of the yc )
where
0 as m
[Exercise. Compute the lower bound when m = 100.]
689
4
,
2
Our last lower bound for the p of success, 2/ 2 > 0, was independent of M (or
N or a), so by repeating the random measurement a fixed number, T , times, we can
assure we measure one of those yc with any desired level of confidence. This follows
from the CTC theorem for looping algorithms (end of probability lesson), but for a
derivation that does not rely on any theorems, compute directly,
P (we do not get a yc after T measurements)
=
T
^
!
measurement Tk yields a yc
k=1
T
Y
P measurement Tk yields a yc
k=1
T
Y
1 P measurement Tk yields a yc
k=1
T
Y
k=1
2
1 2
The last product can be made arbitrarily small by making T large enough, independent of N , M , a, etc. This would prove the claim of Step 2 if we could only use the
bound we got for m
e = m. But we cant. So we must soldier on ...
f = m + 1 Bounds
Repeating the Estimates when m
We now repair the convenient-but-incorrect assumption that m
e is always m. To do
so, lets repeat the estimates, but do so for m
e = m + 1. This is a little harder
sub-case. When were done, well combine both cases.
Remember where m
e came from, and why it could be either m or m + 1. In our
general (hard) case, a does not divide N evenly; the first few x R = [0, a 1] will
generate m + 1 mod-a relatives within [0, N 1] that map to the same f (x), while
the last several x R = [0, a 1] will only produce m such mod-a relatives within
[0, N 1]. m
e represented however many mod-a relatives of x fit into the [0, N 1]
interval: m for some x, and m + 1 for others.
We retrace our steps.
The probability of measuring the state |yi is the amplitudes magnitude squared,
P (measurement yields y)
690
2
m1
e
X
1
jay
mN
e
j=0
2
e
1
1 eiy m
eiy 1 .
mN
e
So far, were okay; we had not yet made any assumption about the particular choice
of m.
e To bound this probability, we went on to get an estimate for the fraction
i m/a
e ybc e 1
eiybc /a 1 .
The denominators upper bound worked for any , so no change needed there. But
the numerators lower bound has to be recomputed, this time, under the harsher
assumption that m
e = m + 1.
Earlier, we showed that 2mb
yc / N was confined to the interval (, ), which
gave us our desired result. Now, however, we replace m with m + 1, and well see
that 2(m + 1)b
yc / N wont be restricted to (, ). What, exactly, are its limits?
Start, as before,
a
a
ybc <
,
2
2
but we need to multiply by a different constant this time,
a(m + 1)
2(m + 1)
a(m + 1)
ybc <
.
N
N
N
By our working assumption, a/N < 2 (which continues to reap benefits) we can assert
that
a(m + 1)
N
am
a
+
N
N
<
1 +
1
,
2
giving us the new absolute bounds
3
2(m + 1)
3
<
ybc <
.
2
N
2
The next step would have been to apply the general result we found for any
[, ], namely that
2 ||
,
for [, ] .
2 sin
2
But now our = 2b
yc (m + 1)/N lives in the enlarged interval [(3/2), (3/2)],
so that general result no longer applies. Sigh. We have to go back and get new
general result for this larger interval. It turns out that old bound merely needs to be
multiplied by a constant:
2 ||
K
2 sin
,
2
3
3
for , ,
2
2
where K is the constant
K
2 sin 3
4
3
691
.4714 .
Figure 23.28: |sin(x/2)| lies above |Kx/| in the interval ( 1.5, 1.5 )
Again, this can be done using calculus and solving K/ = sin (/2) for K. Visually,
you we see where the graphs of the sine and the line intersect, which confirms that
assertion. Summarizing the general result in the expanded interval,
i
3
3
e 1 = 2 sin 2K ||
for , .
2
2
2
This gives us the actual bound we seek in the case m
e = m + 1, namely
i 2(m+1)by / N
2K |2(m + 1)b
yc / N |
c
e
1
.
Re-applying the notation
2ay
y
,
N
we get
m+1
i 2(m+1)y / N
2K
ybc
a
c
e
1
.
Combining the bounds for the numerator and denominator (in the case where m
e =
m + 1),
m+1 ,
i m/a
e ybc e 1
ybc
2K
ybc
a
eiybc /a 1
a
2K(m + 1)
2K m
e
.
f = m and
Finishing the Probability Estimates by Combining Both Cases m
f =m+1
m
Remember that when m
e = m, we had the stronger bound:
i m/a
e ybc e 1
2m
2m
e
=
,
eiybc /a 1
692
so we can use the new, weaker, bound to cover both cases. For all m,
e both m and
m + 1, we have
i m/a
e ybc e 1
2K m
e
,
eiybc /a 1
for
K
2 sin 3
4
3
.4714 .
We can reproduce the final part of the m
e = m argument, incorporating K into the
inequalities
P (measurement yields yc )
e2
1 4K 2 m
mN
e
2
m 4K 2
.
N 2
The last inequality still acknowledges that some of the m
e are m + 1 and some are m.
Again, there are a such yc , namely, y0 , y1 , , ya1 , so the probability of getting any
one of them is () a times this number,
P (measurement yields one of the yc )
am 4K 2
N 2
1 4K 2
2 2
.04503 .
>
2K 2
2
[Exercise.] Compute K when m = 100.
Note-to-file. If a << M < N (i.e., m gets very large), as is often the case, two
things happen. First,
(m + 1)
N
m
,
N
so we can use our first estimate, m
e = m, for measuring a yc , even in the general case.
That estimate gave us
P (measurement yields one of the yc )
2
.
2
>
Second, our earlier Note-to-file suggested that, for m
e = m,
P (measurement yields one of the yc )
where
>
4
,
2
0 as m .
Putting these two observations together, we conclude that when a << N , the lower
bound for any measurement is 4/ 2 . (any means it doesnt matter whether
the state to which the second register collapsed, |f (x0 )i, is associated with an x0 for
which there were m or m + 1 mod-a equivalents in [0, N 1]).
693
So we have both a hard bound assuming worst case scenarios (the period, a, cycles
no more than twice in M ) and the more likely scenario, a << M . Symbolically,
P (measurement yields one of the yc )
.04503, worst case: 3a > M
.
>
.40528,
typically
In the worst case, the smaller lower bound doesnt change a thing; its still a constant
probability boundeed away from zero, independent of N, M, a, etc., and it still gives us
a constant time, T , for detecting one of our special yc with arbitrarily high confidence.
This, again, follows form the CTC Theorem for looping algorithms, or you can simply
apply probability theory directly.
In practice, a << M , so we can use .40528 as our constant p of success bounded
away from 0. If we are satisfied with an = 106 of failure, the CTC theorem tells
us that the number of passes of our circuit would be
log (106 )
6
T =
+ 1 =
+ 1 = 26 + 1 = 27 .
log (.59472)
.2256875
If we sampled y 27 times our chances of not measuring at least one yc is less than of
one in a million.
This completes the proof of the fact that we will measure one of the a yc in constant
time. Our next task is to demonstrate that by measuring relatively few of these yc we
will be able to determine the period. This will be broken into small, bite-sized steps.
23.10.8
a1
STEP III: Associate {yc }a1
c=0 with {c/a}c=0
We now know that well measure one of those yc with very good probability if we
sample enough times (O(1) complexity). But, whats our real goal here? Wed like to
get back to the results of the easy case where we found a number c that was coprime
to a in constant time and used that to compute a. In the general case, however, c
is merely an index of the yc , not a multiple cm, of m. What can we hope to know
about the c which just indexes the set {yc }? Youd be surprised.
In this step, we demonstrate that each of these special, likely-measured yc values is
bound tightly to the fraction c/a in a special way: yc /N will turn out to be extremely
(and uniquely) close to c/a. This, in itself, should feel a little like magic: somehow
the index of the likely-measured set of ys shows up in the numerator of a fraction
that is close to yc /N . Lets pull back the curtain.
Do you remember those (relatively small) half-open intervals of width a around
694
the points cN ,
h
h
a
a
a
a
, +
,
N , N+
,
2
2 h
2
2
a
a
cN , cN +
2h
2
a
a
(a 1)N , (a 1)N +
,
2
2
each of which contained exactly one integral multiple, yc a, of a in it?
h a
h
a
a
a
0a , +
, , yc a cN , cN +
,
2
2
2
2
h
a
a
,
ya1 a (a 1)N , (a 1)N +
2
2
Lets use the relationship of those yc with their host intervals. Start with the obvious,
Figure 23.29: Exactly one integral multiple of a falls in each interval
a
2
yc a
<
cN
c
1
a
2N
yc
N
<
c
1
+
.
a
2N
cN
a
,
2
then divide by N ,
In other words, yc /N is within 1/(2N ) of c/a, or symbolically
c
yc
a
N
1
.
2N
Remember, the relationship between M and N was defined by
N
2
< M 2 N,
so N M 2 . Applying that to the above inequality involving yc /N , we get,
c
yc
a
N
695
1
.
2M 2
Figure 23.30: N = 2n chosen so (N/2, N ] bracket M 2
Compare that with the distances between consecutive c/a values (which I will prove
in a moment),
c
1
c + 1
,
a
a
M2
and we can conclude that each of our special yc , when divided by N , is closer to c/a
than it is to any other fraction with denominator a. Thus, when we measure one of
these yc (which we already showed we can do in constant time with arbitrarily good
probability) we will be picking out the rational number c/a. In fact, we will show
something stronger, but first, we fill in the gap and demonstrate why consecutive
fractions c/a and (c + 1)/a differ by, at least, 1/M 2 .
Lemma. If two distinct integers satisfy 0 < p, q M , then for any
distinct rationals pl and kq , we are guaranteed that
k
l
1
.
p
q
M2
Use this with p
lower bound for
=
q a, k c and l c + 1, and you get the aforementioned
c/a (c + 1)/a .
Proof. Assume k/p > l/q. (Reverse the following if not).
k
l
p
q
kq lp
pq
kq lp
M2
1
.
M2
QED
This lemma tells us that c/a is not only the best fractional approximation to yc with
denominator a, but its best among all fractions having denominators M . For
letting n/d be any fraction with denominator d M . The lemma says
c
n
1
,
a
d
M2
which places n/d squarely outside the 1/ (2M 2 ) neighborhood that contains both c/a
and yc /N .
696
Conclusion: Of all fractions n/d with denominator d M , c/a is the only one
that lies in the neighborhood of radius 1/ (2M 2 ) around yc /N . Thus yc /N strongly
selects c/a and vice versa.
[Interesting Observation. We showed that
yc
N
is uniquely close to
c
.
a
But multiply both sides by N to get to the equivalent conclusion,
N
yc is uniquely close to c
.
a
What is N/a? It is f req, the exact frequency of our function f relative to the interval
[0, N 1]. (In general, its a real number between m and m + 1.) In other words,
each likely-measured yc is uniquely close to an integer multiple of the functions exact
frequency,
yc
is uniquely close to
c f req .
Our math doesnt require that we recognize this fact, but it does provide a nice
parallel with the easy case, in which our measured {yc } were exact multiples, {cm},
of the true integer frequency, f req = m.]
23.10.9
STEP IV: Describe an O(log3 N ) Algorithm that Will
Produce c/a from yc
Highlights of Continued Fractions
The continued fractions algorithm (CFA) will be developed in an optional lesson next
week. In short, it takes a real number x as input and produces a sequence of (reduced)
fractions, {nk /dk } that approach (get closer and closer to) x. We will be applying
CFA to x = yc /N , a rational number itself, but we still want these other fractional
approximations because among them well find one, n/d, which is identical to our
sought-after c/a.
Caution: There is no guarantee that the c yc satisfies c a. So when CFA tells
us that it has found the reduced fraction n/d = c/a, we will not be able to conclude
that n = c and d = a. We will deal with that wrinkle in Step V.
The plan is to list a few results about CFA, then use those facts to show that
it will produce (return) the unique a-denominator fraction, c/a, closest to yc /N , in
O(log3 M ) operations. (Reminder: M is the known upper-bound we were given for
a).
Here are the properties of CFA that well need. Some of them are only true when
x is rational, but for us, thats the case.
697
1. During execution, CFA consists of a loop that produces a fraction in reduced
form, nk /dk , at the end of the kth iteration. nk /dk is called the kth convergent
for x.
2. For any real number, x, the convergents approach x (thus justifying their name),
that is
lim
nk
dk
= x.
(n stands for numerator here, as in numerator/d enomintor, not the n in the
exponent of Shors N = 2n .)
3. For rational x, the above limit is finite, i.e., there will be a K < , with
nK /dK = x, exactly, and no more fractions are produced for k > K.
4. In our version of CFA, we will supply a requested degree of accuracy , and ask
CFA to return n/d, the first fraction it generates which is within of x.
Depending on and x, CFA either returns n/d = nK /dK = x, exactly, as its
final convergent, or returns an -approximation n/d 6= x, but within of it.
5. The denominators {dk } are strictly increasing and, for rational x, all the
denominator of x (whether or not x was given to us in reduced form).
6. If a fraction n/d differs from x by less than 1/(2d2 ) then n/d will appear in the
list of convergents for x. Symbolically, if
n
1
x <
,
d
2d2
then
n
d
nk0
dk0
nk
dk
K
(the convergents for x) .
k=0
7. When x = p/q is a rational number, CFA will complete in O(log3 q). (Sharper
bounds exist, but this is enough for our purposes, and is easy to explain.)
Procedure for Using CFA to Produce c/a
We apply CFA to x = yc /N and = 1/(2M 2 ).
Claim. CFA will produce and return the unique c/a within 1/(2M 2 ) of yc /N .
Proof. a < M , so
1
2M 2
<
698
1
,
2a2
but in step III we showed
c
yc
a
N
<
1
,
2M 2
c
yc
a
N
<
1
.
2a2
so
As this is the hypothesis of bullet 6, c/a must appear among the convergents of
yc /N . Since it is within 1/(2M 2 ) of yc /N , we know that CFA will terminate when it
reaches c/a, if not before.
We now show that CFA cannot terminate before its loop produces c/a. If CFA
returned a convergent nk /dk that preceded c/a, we would have
1
n
k
x
dk
2M 2
by bullet 4. But since the dk are strictly increasing (bullet 5), and we are saying
that the algorithm terminated before getting to c/a, then
dk < a .
That would give us a second fraction, nk /dk with denominator dk < M within
1/(2M 2 ) of yc /N , a title uniquely held by c/a (from Step III). Therefore, when we
give CFA the inputs x = yc /N and = 1/(2M 2 ), it must produce and return c/a.
QED
CFA is O(log3 M )
By Bullet 7 CFA has time complexity O(log3 N ), which we have already established
to be equivalent to O(log3 M ).
23.10.10
Step V: Measure yc Associated with c Coprime to
a in Constant Time
In Steps I and II we proved that the likelihood of measuring one of the special
{yc }a1
c=0 was always bounded below by a constant, independent of N . (The constant
may depend on how many periods, a, fit into M , but for RSA encryption-breaking,
we will see that the special case we require will assure us at least two periods, and
normally, many more).
In the easy case, the {yc } were all equally likely, and we also knew that there were
exactly a/2 coprimes < a. We combined those two facts to put the proof to bed.
Here, neither condition holds, so we have to work harder (which is why this isnt
called the easy case).
699
What we can say is that, from Steps III and IV, each of the measured {yc }s
leads in constant time, with any predetermined confidence, to a partner fraction
in {c/a} with the help of some O(log3 M ) logic provided by CFA.
In this step, we demonstrate that not only do we measure some yc (constant time)
and get a partner, {c/a} (O(log3 M )), but that we can even expect to get a special
subset B {yc } {c/a} in constant time, namely those yc corresponding to c a
(i.e., c coprime to a). This will enable us to extract the period a from c/a.
We do it all in three steps, the first two of which correspond to the missing
conditions we enjoyed in the easy case:
Stated loosely, in our quantum circuit, the difference between measuring the
least likely yc and the most likely yc is a fixed ratio independent of a. (This
corresponds to the equi-probabilities of the easy case.)
The probability of selecting a number, c, which is coprime to a, at random,
from the numbers between 0 and a 1 is constant, independent of a. (This
corresponds to the 50% likelihood of the easy case.)
in
We combine the first two bullets to show that the probability of getting a c a
any single pass of the circuit is bounded away from 0 by a constant independent
of the size of the algorithm. Thats the requirement of the CTC theorem for
looping algorithms and therefore guarantees we obtain such a c with small error
tolerance in constant time, i.e., after a fixed number of loop passes independent
of N .
Proof of First Bullet
This can be demonstrated by retooling an analysis analysis we already did. We
established that the amplitude squared of a general |yi in our post-QFT A register
was
2
e
1
1 eiy m
P (measurement yields y) =
eiy 1 .
mN
e
There are clearly two measurement-dependent parameters that will affect this probability: m,
e which is either m or m + 1, depending on the collapsed state of the second
register, and y , which depends on the measured y. When a << M the probabilities
are very close to being uniform, but to avoid hand-waving, lets go with our worst-case
scenario the assumption that we added to Shors hypothesis in order to get hard
estimates for all our bounds: M > 2a, but not necessarily any larger.
When we computed our lower bound on the probability of getting a yc , we used an
inequality that contained an angle under the assumption that was restricted to
an interval wider than [, ], spefically [(3/2), (3/2)]. The inequality we found
applicable was
700
2 ||
K
2 sin
,
2
3
3
for , ,
2
2
where K is the constant
K
2 sin 3
4
3
.4714 .
The key to our current predicament is to get an upper bound on measuring any (even
the most likely) yc . This will lead us to get an inequality for general restricted to a
narrower range, namely, [/2, /2]. This can be easily determined using the same
graphing or calculus techniques, and is
Figure 23.31: |sin(x/2)| lies above |Lx/| in the interval ( .4714, .4714 )
2 ||
L
2 sin
,
2
h i
for ,
,
2 2
where L is the constant
L
2 sin
1.4142 .
Lets see how this gives us the upper bound we seek.
The probability we we wish to bound, this time from above, comes from the
amplitude whose absolute value is
e
e
1 eiy m
1
1 eiybc m/a
1
=
.
mN
e eiy 1
mN
e eiybc /a 1
We want an upper bound for the magnitude of the fractional term. To that end, we
get an upper bound for the numerator and a lower bound for the denominator.
701
Upper Bound for Numerator
The numerator is easy. We derived an upper bound for all angles, , so:
i m/a
ybc m
e
e
e ybc
.
1
a
Lower Bound for Denominator
The denominator is
i /a
e ybc 1
i (2by / N )
c
e
1 ,
but 2b
yc / N (/2, /2) because
a
a
ybc <
2
2
or
a
N
2
a
ybc . < .
N
N
And since
a
< 1/2 ,
N
we get
<
ybc . <
.
2
N
2
This allows us to invoke the latest lower bound just mentioned for [/2, /2]:
i 2by / N
e c
1
2 |2b
yc / N |
, L = 2 sin 1.4142 .
Re-applying the notation
y
2ay
N
we can write
i 2y / N
e c
1
702
2 |ybc /a|
.
Applying Both Bounds
Combining the bounds for the numerator and denominator,
i m/a
,
e ybc e 1
ybc m
e
2 |ybc /a|
L
=
eiybc /a 1 a
2m
e
.
L
Finally,
P (measurement yields yc )
1 4L2 m
e2
mN
e
2
m + 1 4L2
.
N
2
The last inequality still acknowledges that some of the m
e are m + 1 and some are
m. This time, we are only interested in the probabilities for each individual yc , not
getting one of the set {yc }, so instead of summing all a of the yc s, we combine this,
as is, with the earlier one, which I repeat here,
P (measurement yields yc )
e2
1 4K 2 m
mN
e
2
m 4K 2
,
N 2
to get
P (least-likely yc )
P (most-likely yc )
mK 2
.
(m + 1)L2
Our assumption has been that a M/2 so m = (integer quotient) N/a is > 2 (usually
much greater). This, and the estimates for L 1.4142 and K .4714, result in a
ratio which is independent of a, M or N , between the probability of measuring the
least likely yc to that of measuring the most likely yc .
P (least-likely yc )
P (most-likely yc )
mK 2
(m + 1)L2
2K 2
.072 .
3L2
QED
This covers the bullet about the ratio of the least likely and the most likely yc .
Note-to-file. if a << M < N (i.e., m gets very large), as is often the case, Both
K and L will be close to 1 (review the derivations and previous notes-to-file). This
means all yc are roughly equi-probable. Also m/(m + 1) 1. Taken together, the
ratio of the least likely to most likely is approximately 1.
Summarizing, we have a hard minimum for the worst case scenario (only two
intervals of size a fit into [0, M )) as well as an expected minimum for the more
realistic one (a << M ). That is,
.072, worst case: 3a > M
P (least-likely yc )
P (most-likely yc )
1 ,
typically
703
Proof of Second Bullet
We will show that the probability of randomly selecting a number c a is > 60%.
First, we express
this as a product of terms, each of which is the probability that
a specific prime, p c (p divides evenly into c).
Note that, a is an unknown: no one a is favored over any other a, a priori. Also,
c is selected at random, by hypothesis of this bullet. So both a and c are considered
selected at random with respect to any particular prime, p. We use this fact in the
derivation.
a)
P (c
P ( There is no prime that divides both c and a )
=
P
2 c 2 a 3 c 3 a 5 c 5 a . . .
...
a)
P (c
pk c pk a
P
finite
^
...
!
p c p a
where pk = kth prime .
finite
Y
P p c p a
p prime
p prime
Since (p c p a) is true for p > a or p > c, the probabilities for those higher primes,
p, are all 1, which is why the product is finite.
Next, we compute these individual probabilities. For a fixed prime, p, the probability that it divides an arbitrary non-negative c chosen randomly from all non-negative
integers is actually independent of c,
1
P p c = .
p
[Exercise. Justify this.] This is also true for p a, so,
P
p c p a
1
p2
p c p a
and
1
1
.
p2
Remember, both a and c can be considered randomly selected relative to a fixed
prime, p. If, instead, we restrict ourselves to non-negative values chosen randomly
from a finite set, such as c < a (or a < M ), then the probability that p|c is actually
< 1/p. To see this, consider the worst cases, say a < p (P = 0) or p a < 2p
(P = 1/a < 1/p). So the above equalities become bounds which work even better for
704
us,
1
P p c < ,
p
1
p c p a
<
p2
p c p a
P
P
and
1
1
.
p2
Finally, we plug this result back into the full product, to get
a) =
P (c
finite
Y
P (p|c p|a)
p prime
p prime
1
p2
= (2),
where (s) is the most famous function you never heard of, the Riemann zeta function,
defined by
Y
(s) =
p prime
1
ps
in Euler product form. The value of (2) has to be handed-off to the mathematical
annals, and well simply quote the result,
(2)
.607 .
so we have shown that
a)
P (c
(2)
.607. QED
That proves our second bullet, which is all we will need, but notice what it implies.
Since
a)
P ( c
<
.393 ,
after T random integer selections c1 , c2 , . . . , cT in the interval [0, a 1],
a c2 a . . . cT a
P
c1
< .393T , constant time, T , independent of N .
This can be
made arbitrarily small by choosing a large enough T , so we can expect
to get a c a with arbitrarily high probability after T random samples from the set
[0, a 1].
705
Combine Both Bullets to Arrive at Step V
To reduce the syntax, it will help to introduce some notation.
C
{yc }a1
c=0
a}
{ yb yb C and b
(Note: B C.)
|B|
|C|
P (yc )
the size (number of elements) in set B
(same) = a
|B|
the ratio
|C|
P (we measure a specific yc C)
(Note: In the last definition, yc may or may not also be B.)
P (C)
P (B)
P (we measure some y C)
P (we measure some y B)
P (B|C)
P we measure some y B
given that
the measured y is known to be C
We would like a lower bound on the probability of measuring a yc which also has the
property that its associated c is coprime to a. In symbols, we would like to show:
Claim: P (B)
some constant independent of a, M, N, etc.
Proof:
P (B)
=
=
P (B|C) P (C)
,
!
X
X
P (yb )
P (yc ) P (C)
cC
bB
Let
yBmin y B with P (yBmin ) minimum over B
yBmax y B with P (yBmax ) maximum over B
yCmin , yCmax same, except over all of C
If there is more than one y that produce minimum or maximum probabilities, choose
706
any one. Then,
,
X
!
X
P (yb )
P (yc )
cC
bB
|B| P (yBmin )
|C| P (yCmax )
|B| P (yCmin )
|C| P (yCmax )
q .072 ,
so
P (B)
q .072 P (C) .
From the proof of the second bullet of this step, q > .607, and from Step II P (C) >
.04503, so
P (B)
.607 .072 .04503 .002.
This is independent of a, M, N, etc. and allows us to apply the CTC theorem for
looping algorithms to aver that after
a fixed number, T , of applications of the quantum
circuit we will produce a yc with c a with any desired error tolerance. Well compute
T in a moment.
Remember that we used a worst case bounds above. As we demonstrated, normally the ratio .072 is very close to 1, and P (C) > .40528, so we can expect a better
constant lower bound:
P (B)
.607 1 .40528 .266.
Conclusion of Step V
After an adequate number of measurements (independent of a, M ), which produce
yc1 , yc2 , . . . , ycT , we can expect at least one of the yck = yc to correspond to c/a,
with c a with high probability.
Examples that Use Different Assumptions about a
How many passes, T , do we need to get an error tolerance of, say = 106 (one in
a million)? It depends on the number of times our period, a, fits into the interval
[0, N ). Under the worst case assumption that we formally required only two we
would need a much larger number of whacks as the circuit than in a typical problem
that fits hundreds of periods in the interval. Lets see the difference.
The p of success bounded away from 0 for a single pass of our algorithms loop
in the CTC theorem, along with the error tolerance , gives us the number of required
passes, the formula provided by the theorem,
log ()
T =
+ 1.
log (1 p)
707
Worst Case (P (B) .002): We solved this near the end of the probability
lesson and we found
log (106 )
T =
+ 1 = 6901 ,
log (.998)
or, more briefly and conservatively, 7000 loop passes.
Typical Case (P (B) .266):
This was also an example in the earlier
chapter,
log (106 )
+ 1 = 45 ,
T =
log (.734)
or, rounding up, 50 loop passes.
Its important to remember that the datas actual probability doesnt care
about us. We can instruct our algorithm to cycle 7000 times, but if the data
determined that only 15 loop passes were needed to find the desired c/a, then it
would return with a successful c/a after 15 passes. Our hyper-conservative estimate
costs us nothing.
On the other hand, if we are worried that a < N/2 is too risky, and want to allow
for only one period of a fitting into M or N , the math works. For example, we could
require merely a < .999N and still get constant time bounds for all probabilities. Just
repeat the analysis replacing 1/2 with .999 to find the more conservative bounds. You
would still find that proofs all worked, albeit with P (B) > a constant much smaller
than .002.
23.10.11
Algorithm and Complexity Analysis (General Case)
Why Finding yc B Solves Shors Problem
In the final step of the previous section we proved that we will measure yc B with
near certainty after some predetermined number, T of measurements. Why does such
a yc solve Shors period-finding problem? CFA returns n/d = c/a close to yc /N , but
the actual numerators and denominators do not have to match: n/d is reduced,
but
c/a may not be for general c. However, we are safe when yc B, for then, c a, and
we are assured that c/a is a reduced fraction. In that case, we can assert that the
numerators and denominators match; for the denominators, that means d = a and
we are done.
This happens if we measure y B, and this is why finding such a y will give us
our period and therefore solve Shors problem. (Wait. Might we end up getting d = a
even if we happened to measure some y
/ B? In other words, is the if in the last
sentence if and only if or just if ? Well, there are two ways to deal with this question.
You could chase down the logic to see whether a y
/ B used to send x = y/N to the
CFA could possibly give us d = a, but why bother? I prefer the second method to
dispatch the question: Our algorthm will test d = a after every measurement, so if
(possible or not) we found a for the wrong reason, we still found a.)
708
Shors Period-Finding Algorithm, General Case
After some hard work we have assembled all the results. Two things need to go right
to be certain we measured a y B:
1. We measure y C = {yc }.
2. the associated c satisfies c a.
We could fail in either or both, but the test of overall success is that the fraction,
n/d, returned, by CFA( yc /N, 1/(2M 2 ) ) has the property that d = a. And we can
test that in a single evaluation of our given function f . (Remember f ? It was given
to us in Shors hypothesis, and we used it to design/control Uf ).
Testing whether, say, f (1 + d) = f (1) is enough to know whether d is the period,
a.
The complexity of this test may or may not be polynomial time, depending on f .
For RSA encryption-breaking the f in question is O(log4 (M )), as we will prove, so for
that application we will have absolute speed-up over classical solutions. In general,
this step confines our quantum period-finding to only relativized speed-up.
So, the short description of the algorithm is this:
Run the circuit producing yc s, use those to get (n/d)s, stop when we
confirm f (1 + d) = f (1), report that d = a (our period) and declare
victory.
In the unlikely event that we run out of time (exceed our established
T ), we admit defeat.
The full version follows.
Select an integer T that reflects an acceptable failure rate based on any known
aspects of the period. (E.g., for a failure tolerance of .000001, we might choose
T = 7000 if we expect only 2 periods, a, to fit into [0, M 1) or T = 45 if we
know a << M . If we are happy with failure at .001, then these values would
be adjusted downward.)
Repeat the following loop at most T times.
1. Apply Shors circuit.
2. Measure output of QFT to get y.
3. Apply CFA to y/N and 1/(2M 2 ) to produce n/d.
4. Test d: If f (1 + d) = f (1) then d = a, (success) break from loop.
5. Otherwise continue to the next pass of the loop.
If the above loop ended naturally (i.e., not from the break ) after T full passes,
we failed. Otherwise, we have found a.
709
Computational Complexity (General Case)
(This is nearly word-for-word identical to the easy case.)
We have already seen that O(log(N )) = O(log(M )) as well as any powers of the
logs, so I will use M in the following.
The Hadamard gates are O(log(M )).
The QFT is O(log2 (M )).
Complexity of Uf is same as that of f , which could be anything. We dont count
that here, but we will show in a separate lecture that for integer factoring, it is
certainly at least O(log4 M ).
The outer loop is O(T ) = O(1), since T is constant.
The classical CFA sub-algorithm is O(log3 (M )).
The four non-oracle components, above, are done in series, not in nested loops,
so the overall relativized complexity will be the worst among them, O(log3 (M )).
In the case of factoring needed for RSA encryption breaking (order-finding)
the actual oracle is O(log4 (M )) or better, yielding an absolute complexity of
O(log4 (M ))
So the entire Shor circuit for an f O(log4 (M )) (true for RSA/order-finding) would
have an absolute complexity of O(log4 (M )).
23.10.12
Epilogue on Shors Period-Finding Algorithm
Other than a discussion of Euclids algorithm for computing the greatest common
divisor and some facts about continued fractions (both covered in this volume) you
have completed a rather in-depth development of Shors periodicity algorithm for
quantum computers.
You studied the circuit, the algorithm and computational complexity for quantum
period-finding.
There remains the question of how quantum period-finding can be applied to RSA
encryption-breaking, which is a form of order-finding. Before we close out this first
course, Ill cover that as well.
710
Chapter 24
Euclidean Algorithm and
Continued Fractions
24.1
Ancient Algorithms for Quantum Computing
The Euclidean algorithm is a method for computing the greatest common divisor
of two integers. The technique was described (although probably not invented) by
the Greek mathematician Euclid in about 300 B.C. Two thousand years later it
continues to find application in many computational tasks, one of which is Shors
quantum period finding algorithm. In fact, we apply it twice in this context, once
directly, then again indirectly when we apply a second, centuries old technique called
continued fractions.
In this lesson, we will study both of these tools and compute their time complexities.
24.2
The Euclidean Algorithm
24.2.1
Greatest Common Divisor
The Euclidean Algorithm (EA) takes two positive integers, P and Q, and returns the
largest integer that divides both P and Q evenly (i.e., without remainder). Its output
is called the greatest common divisor of P and Q and is written as a function of its
two input integers, gcd(P, Q). Of special note, we will learn that
EA(P, Q) O(log3 X), where X is the larger of the two integers passed to it
(although sharper/subtler bounds exist), and
EA will be used as a basis for our Continued Fractions Algorithm, CFA.
711
24.2.2
Long Division
A long division algorithm (LDA) for integers A and B, both > 0, produces A B in
the form of quotient, q and remainder, r satisfying
A
qB + r.
The big-O time complexity of LDA(A, B) in its simplest form is O(log2 X), where X
is the larger of {A, B}. If you research this, youll find it given as O(N 2 ), where N is
the number of digits in the larger of {A, B}, but that makes N = log10 X, and log10
has the same complexity as log2 .
24.2.3
The Euclidean Algorithm, EA(P, Q)
Although the final result is independent of which number occupies the first parameter
position, we will assume P is first, as this will affect the intermediate calculations
which are used by CFA, coming up next.
I will describe the algorithm without proof it is well documented and very easyto-follow in all the short descriptions you will find online.
General Idea
To produce EA(P, Q), for P, Q > 0, we start by applying long division to P, Q.
LDA(P, Q) returns a quotient, q, and remainder, r, satisfying
P
qQ + r.
Notice that either r = 0, in which case Q|P and we are are done (gcd(P, Q) = Q),
or else we have two new integers to work with: Q and r, where, now, Q > r. We
re-apply long division, this time to the the inputs Q and r to get Q r, and again,
examine the new remainder (call it r0 ),
Q
q 0 r + r0 .
Like before, if r0 = 0 we are done (gcd(P, Q) = r), and if not, we keep going. This
continues until we get a remainder, re = 0, at which point the gcd is the integer
standing next to (being multiplied by) its corresponding quotient, qe, in the most
recent long division (just as Q was standing next to q in the initial division, or r
was standing next to q 0 in the second division).
Lets add some indexing to our general idea before we give the complete algorithm. Define
r0 P,
r1 Q,
712
and
q0
r2
the first quotient of P Q,
the first remainder of P Q.
Using this notation, the initial division becomes
r0
q0 r1 + r2 .
If r2 = 0, return gcd = q0 , otherwise, r1 > r2 > 0, and we repeat, to compute r1 r2 ,
r1
q1 r2 + r3 .
Continue until rk = 0, at which point return gcd = rk1 .
24.2.4
The Algorithm
EA(P, Q):
Initialize
r0 = P
r1 = Q
Loop over k, starting from k = 0
Compute rk rk+1 using LDA(rk , rk+1 ) to compute qk and rk+2
rk
qk rk+1 + rk+2 .
until rk+2 = 0
Return gcd = rk+1
713
Example
W3 use EA to compute gcd(285, 126)
r0 = 285
r1 = 126
r0 = q0 r1 + r2
285 = 2 126 + 33
r2 6= 0, so compute r1 r2
r1 = q1 r2 + r3
126 = 3 33 + 27
r2 6= 0, so compute r2 r3
r2 = q2 r3 + r4
33 = 1 27 + 6
r3 6= 0, so compute r3 r4
r3 = q3 r4 + r5
27 = 4 6 + 3
r4 6= 0, so compute r4 r5
r4 = q4 r5 + r6
6 = 23 + 0
r5 = 0, so return gcd = r4
gcd = 3
24.2.5
Time Complexity of EA
This is easiest seen by looking at the very first division,
P
qQ + r.
Notice that
r <
P
.
2
To see this, break it into two cases.
case i) Q P/2. In this case, since r < Q, we have r < P/2.
case ii) Q > P/2. In this case q = 0 and r = P Q < P P/2 = P/2.
This same logic, applied to the iterative division
rk
qk rk+1 + rk+2 ,
714
gives
rk
.
2
rk+2 <
In other words, every two divisions, we have reduced rk by half, forcing the evenindexed rs to become 0 in at most 2 log P iterations. Therefore, the number of steps
is O(2 log P ) = O(log P ). (Incidentally, the same argument also works for Q, since
Q = r1 , spawning the odd-indexed rs. So, whichever is smaller, P or Q, gives a tighter
bound on the number of steps. However, we dont need that level of subtlety.) We
have shown that the EA main loops complexity is O(log X), where X is the larger
(or either) of P , Q.
Next, we note that each loop iteration, calls the LDA(rk , rk+1 ), which is O(log2 rk ),
since rk > rk+1 . For all k, rk < X (larger of P and Q), so LDA(rk , rk+1 ) is also in
the more conservative class, O(log X).
Combining the last two observations, we conclude that the overall complexity of
EA(P, Q) is the product of complexities of loop, O(log X), and the LDA within the
loop, O(log2 X), X being the larger of P and Q. Symbolically,
EA(P, Q) O(log3 X),
X = larger of {P, Q}.
24.3
Convergents and the Continued Fraction Alorithm
24.3.1
Continued Fractions
Although we wont deal with such dizzying entities directly, you need to see an example of at least one continued fraction so we can define the more useful derivative
(convergents) and present the CFA algorithm. A continued fraction is a nested construct (possibly infinite) of sums and quotients of integers in the form
1
a0 +
a1 +
a2 +
a3 +
.
1
a4 +
..
We will only consider a0 0, and ak strictly positive for all k > 0.
Notice that the sequence of integers { a0 , a1 , a2 , . . . } completely defines this object
because of its special form: all numerators = 1 and all (except for possibly the last)
denominators the sum of exactly two terms, an integer part, ak , and its fractional
partner, 1 over something.
I said that some continued fractions are finite. When that happens there is a final
aK alone in its denominator, terminating the continued fraction.
715
The headlines are:
Any real number,
x, including both rationals (29/48, 7, 8765/1234, etc.) and
irrationals (, 2, e2 , etc.), can be expressed as a continued fraction. (We only
care about positive x.)
When x is rational the continued fraction is finite. When x is irrational the
continued fraction is infinite (it pencils out to the irrational x but takes an
infinite amount of graphite to write).
A Rational Example
An example of a rational number expressed as a continued fraction is (check it yourself):
x =
285
= 2 +
126
1
1
3 +
1 +
1
2
You might ask why we bother with a continued fraction of an (already) rational number, x. The special form of a continued fraction leads to important approximations
to x that are crucial in our proof of the quantum period-finding algorithm.
4 +
An Irrational Example
A famously simple irrational continued fraction is
x =
2 = 1 +
1
1
2 +
2 +
2 +
.
1
.
2 + ..
I will not prove any the above claims or most of what comes below. The derivations
are widely and clearly published in elementary number theory texts and short webpages on continued fractions. But I will organize the key facts that are important to
Shors Algorithm.
24.3.2
Computing the CFA ak s Using the EA if x is Rational
Before we abandon the ugly nested construct above, lets look at one easy way to get
the ak if x is rational. This will tell us something about the computational complexity
of our upcoming continued fractions algorithm (CFA).
716
CFA Using the EA. If
x =
P
,
Q
the {ak } in its continued fraction expansion are exactly the unique {qk }
of the Euclidean Algorithm, EA(P, Q), for finding the gcd(P, Q).
In other words, we already have an algorithm for expanding a rational number as a
continued fraction. Its called the Euclidean algorithm and we already presented it.
We just grab qk from EA and were done.
(Thats why I made a distinction between the first parameter, P and the second,
Q, and also why I labeled the individual qk in the EA, which we didnt seem to need
at the time.)
Time Complexity for Computing a Rational (Finite) Continued Fraction
Theorem. The time complexity for computing all the ak for a rational x
is O(log3 X), where X is the larger of {P, Q}.
Proof : Since the {ak } of continued fractions are just the {qk } of the EA, and
we have proved that the EA O(log3 X), where X is the larger of P and Q, the
conclusion follows.
QED
24.3.3
A Computer Science Method for Computing the ak s
of Continued Fractions
An iterative algorithm is typically used when programming CFA. We use the notation
bxc
greatest integer x.
The CF Method
Specify a termination condition (e.g., a maximum number of loop passes or a
continued fraction within a specified of the target x, etc.).
Loop over k starting at k = 0 and incrementing until we reach the maximum
number of loop passes or we break sooner due to a hit detected in the loop
body.
1. ak bxc
2. f rac x bxc
3. If f rac = 0, break from loop; weve found x.
4. x 1/f rac
Return the sequence {ak }.
717
24.3.4
Convergents of a Continued Fraction
Algorithmic Definition of Convergents
Rather than using the unwieldy continued fractions directly, well work with the much
more tractable simple fractions which derive from them.
The kth Convergent of x. If we forcibly stop the loop in the CF
method after k iterations, we produce (and return) a number which is only
an approximation of x. This approximation is called the kth convergent
of x.
Of course if x is irrational, we can only ever get an approximation, so this is not
big news. But if x is rational, and we halt the process prematurely, before ak =
aK the final denominator, then we only have an approximation of x even though
we might have been more patient and waited to generate the full expression of x,
itself. Surprisingly, these approximations are sometimes more useful than the whole
enchilada. Certainly, for our purposes, they will turn out to be.
When we fully evaluate the kth convergent it will, of course, resolve to a fraction
in its own right, and we always express that fraction in reduced form.
Algebraic Definition of Convergents
We could have defined the kth convergent of x when we first introduced continued
fractions. They are just the partially completed constructs that we see when writing
out (or attempting to write out) the full continued fraction. Explicitly, the first few
convergents are
n0
= a0 ,
d0
1
n1
= a0 +
,
d1
a1
1
n2
= a0 +
,
d2
1
a1 +
a2
and generally,
nk
dk
a0 +
a1 +
...
ak1 +
.
718
.
1
ak
Example
For the rational x = 285/126, whose continued fraction and {ak } we computed earlier,
you can verify that the convergents are
n0
d0
2
,
1
n1
d1
7
,
3
n2
d2
9
,
4
n3
d3
43
,
19
n4
d4
95
.
42
and
Notice that the final convergent is our original x in reduced form. This is very important for our purposes. Figures 24.1 and 24.2 show two graphs of these convergents at
different zoom levels.
Figure 24.1: A wide shot of the of the convergents starting at 2/1
Figure 24.2: A close-up showing some of the latter convergents.
A couple features might pop:
719
1. The convergents converge very fast every two convergents are much closer
together than the previous two, and
2. The convergents bounce back-and-forth around the target, x, alternating lessthan-x values and greater-than-x values.
Before making this precise, lets see if it holds for a second example.
Example
For the rational x = 11490/16384 the convergents are:
n0
d0
0
,
1
n1
d1
1
,
1
n2
d2
2
,
3
n3
d3
5
,
7
n4
d4
7
,
10
n5
d5
54
,
77
n6
d6
1897
,
2705
n7
d7
5745
=
8192
and
11490
.
16384
This time, well need more zoom levels because the later convergents are so close to
x they are impossible to see. Figures 24.3 through 24.7 show various stages of the
convergents.
Figure 24.3: First view of convergents
720
Figure 24.4: Second view of convergents
Figure 24.5: Third view of convergents
Figure 24.6: Fourth view of convergents
Figure 24.7: Fifth view of convergents
721
24.3.5
An Algorithm for Computing the Convergents {nk /dk }
Specify a termination condition (e.g., a maximum number of loop passes or a
convergent within a specified of the target x, etc.).
Invoke the CF method, given earlier, that produces the entire sequence {ak } to
the desired accuracy. (Note: The CF method can be merged into this algorithm
by combining loops carefully.)
Set the first two convergents, manually:
n0 a0 ,
n1 a1 a0 + 1,
d0 1
d 1 a1
Loop over k starting at k = 2 and iterating until k = K, the final index of the
sequence {ak } returned by the CF method.
1. nk ak nk1 + nk2
2. dk ak dk1 + dk2
Return the sequence {nk /dk }.
24.3.6
Easy Properties of the Convergents
The convergents {nk /dk } have the following properties which are derived in most
beginning tutorials in very few steps.
1. For any real number, x, the convergents approach x,
nk
lim
= x.
x
dk
2. For rational x, the above limit is finite, i.e., there will be a K < , with
nK /dK = x exactly, and no more fractions are produced for k > K.
3. They alternate above and below x,
> x, if k is odd
nk
=
(assuming {nk /dk } =
6 x) .
dk
< x, if k is even
4. Each nk /dk (if not the final convergent which is exactly x) differs from x by no
more than 1/(dk dk+1 ),
n
1
k
x
.
dk
dk dk+1
722
5. For k > 0, nk /dk is the best approximation to x of all fractions with denominator
dk ,
n
k
x
x n , for all d dk .
dk
d
6. Consecutive convergents differ from each other by exactly 1/(dk dk+1 ),
nk1
nk
1
dk1 dk = dk dk+1 .
7. The denominators {dk } are strictly increasing and, if x is rational, are all the
denominator of x (whether or not x was given to us in reduced form).
8. When x = P/Q is a rational number, computation of the all convergents is
O(log3 X), where X is the larger of {P, Q}. (This follows from the fact that
the convergent algorithms above are based on EA as well as the details of the
loops used.)
There is one not-so-easy-to-prove fact that we will need. It can be found in An
Introduction to the Theory of Numbers by Hardy and Wright (Oxford U. Press) as
Theorem 184. The proof is rather involved.
If a fraction n/d differs from x by less than 1/(2d2 ) then n/d will appear in the
list of convergents for x. Symbolically:
If
n
x
d
<
1
2d2
then
n
d
24.3.7
nk0
dk 0
nk
dk
K
(the convergents for x) .
k=0
CFA: Our Special Brand of Continued Fraction Algorithm
Our version of a convergent-generating algorithm, which I call CFA, will take as input
parameters a rational target x and requested degree of accuracy . CFA(x, ) will
return n/d, the first convergent (i.e., that with the smallest index k) to x within
of x. It simply wraps the previous algorithm into an envelope that returns a single
fraction (rather than all the convergents).
1. To our previous algorithm for generating the convergents, pass x along with the
terminating condition that it stop looping when it detects that |x nk /dk | .
723
2. Return nk /dk .
Thats all there is to it.
Depending on and x, CFA(x, ) either returns n/d = nK /dK = x, exactly, as its
final convergent or an -approximation n/d 6= x, but within of it.
724
Chapter 25
From Period-Finding to Factoring
25.1
Period Finding to Factoring to RSA Encryption
RSA encryption relies on an interesting computational and mathematical fact. Given
large enough positive integer, M = pq, which is the product of two primes, it would
take the worlds fastest super computer billions of years to compute the two unknown
factors p and q. How large is large enough? An M with around 28 decimal digits
would work.
Youd have to take a course in classical cryptography or find a tutorial on the
Web to learn why the inability to factor large numbers leads to encryption and how
the RSA works. Our task today is to learn how a quantum computer can solve the
same factoring problem for such a large number efficiently (potentially in minutes or
seconds). We have, in our quill, a very powerful arrow: Shors quantum algorithm
for period-finding. Lets see how a hypothetical quantum computer could leverage
it to overcome the computational limitation to which a classical super computer is
subject.
25.2
The Problem and Two Classically Easy Cases
The Factoring Problem. Given some M Z>0 that we know is not
prime, find a q Z>0 that satisfies q|M .
[Variable Name Most number theory publications use the letter N for the
large number to be factored. I have been using N to be the power-of-2 larger than
M 2 , where M is our bound on the period a and the number we want to factor, so I
will use M for the arbitrary integer in this section.]
We will want to narrow down the kind of M that were willing to look at, and there
are two cases which can be tested and factored by ordinary computers efficiently,
725
1. M even, and
2. M = pk , k > 1, is a power of some prime.
Why the Two Cases Are Easy
The test M even? entails a simple examination of the least significant bit (0 or 1),
trivially fast. Meanwhile, there are easy classical methods that determine whether
M = pk for some prime p and produce such a p in the process (thus providing a
divisor of M ).
In fact, we can dispose of a larger class of M : those M for which M = q k , k > 1
for any integer q < M , prime or not, and produce q in the process all using classical
machinery. If we detected that case, it would provide a factor q, do so without
requiring Shors quantum circuit, and cover the more restrictive condition 2, in which
q is a prime.
So why does the second condition only seek to eliminate the case in which M is
a power of some prime p before embarking on our quantum algorithm rather than
using classical methods to test and bypass the larger class of M that are powers of
any integer, q? First, eliminating only those M that are powers of a single prime
is all that the quantum algorithm actually requires. So once we have disposed of
that possibility, we are authorized to move on to Shors quantum algorithm. Second,
knowing we can move on after confirming M 6= pk , for a p prime, gives us options.
We can ask the number theorists to provide a very fast answer to the question
is M = pk , p prime, and let the quantum algorithm scoop up the remaining
cases (which include M = q k , q not prime).
Alternatively, we can apply a fast classical method to search for a q (prime or
not) that satisfies M = q k , thus avoiding the quantum algorithm in a larger
class of M .
One of the above two paths may be faster than the other in any particular {hardware
+ software} implementation, so knowing that we can go either way gives us choices.
Now lets outline why either of the two tests
M = qk , k > 1
or
M = pk , k > 1, p prime
can be dispatched classically.
Classical Algorithm for Larger Class: M = q k ,
k>1
Any such power k would have to satisfy k < log2 M since p > 2 (weve eliminated M
even). Therefore, for every k < log2 M we compute the integral part
j k
k
M
q=
726
(something for which fast algorithms exist) and test whether q k = M . If it does, q is
our divisor, and we have covered the case M = q k , k > 1, for any integer q without
resorting to quantum computation. The time complexity is polynomial fast because
the outer loop has only log2 M passes (one for each k),
j k
k
even a slow, brute force method to compute q =
M has a polynomial
big-Oh, and
taking the power q k is also polynomial fast. (Moreover, the implementation of
previous bullet can be designed to absorb this computation, obviating it.)
[Exercise. Design an algorithm that implements these bullets and derive its
big-Oh.]
Classical Algorithm for Smaller Class: M = pk ,
k > 1, p Prime
But if one wanted to also know whether the produced q in the above process was
prime, one could use an algorithm like AKS (do a search) which has been shown to
be log polynomial, better than O(log 8 (#digits in M )) = O(log 8 (logM )), in M .
This approach was based on first finding a q with M = q k , then going on to
determine whether q was prime. Thats not efficient, and I presented it only because
the components are easy, off-the-shelf results that can be combined to prove the
classical solution is polynomial fast. In practice we would seek a solution that tests
whether M is a power of a prime directly, using some approach that was faster than
testing whether it is a power of a general integer, q.
Why We Eliminate the Two Cases
The reason we quickly dispose of these two cases is that the reduction of factoring
to period-finding, described next, will not work for either one. However, we now
understand why we can be comfortable assuming M is neither even nor a power of a
single prime and can proceed based on that supposition.
25.3
A Sufficient Condition
The next step is to change the the factoring problem into a proposition with which
we can work more easily.
Claim. We will obtain a q M if we can find an x with the property that
x2 = 1 (mod M ), for
x 6= 1 (mod M ) .
727
Proof
x2
= 1 (mod M )
x2 1 = 0 (mod M )
2
M (x 1)
M (x 1) (x + 1) .
That can only happen if M has a factor, p > 1, in common with one or both of (x1)
and (x + 1), i.e.,
p M and p (x 1) or
p M and p (x + 1) .
That factor, p, cannot be M , itself, for if it did, either
M (x 1), contrary to x 6= +1 (mod M ) or
M (x + 1), contrary to x 6= 1 (mod M ) ,
both, outlawed by the hypothesis of the claim. Define
(
gcd(M, x 1), if common factor p (x 1) or
q
gcd(M, x + 1), if common factor p (x + 1) .
Whichever of the above two cases is true (and we just proved at least one must be),
we have produced a q with q M .
QED
The time complexity of gcd(M, k), M k is shown in another lecture to be
O(log3 M ), so once we have x, getting q is fast.
Before we find x, we take a short diversion to describe something called order
f inding.
25.4
A Third Easy Case and Order-Finding in ZM
Pick a y at random
from ZM {0, 1} = {2, 3, 4, . . . , M 1}. Either y will be
coprime to M (y M ) or it wont.
If y M were done, because q gcd(M, y) will be our desired factor of
M (and we dont even need to look for an x). An O(M 3 ) application of GCD
will reveal this fact by either returning a q > 1 (were done) or result in 1 (we
continue).
728
If y M we go on to find the order of y, defined next and which leads to the
rest of the algorithm.
We therefore assume that y M , since if it were not we would have lucked upon the
first bullet and factored M that would be a third easy case we might encounter.
While it may not be obvious, finding the order of y in ZM will be the key. In
this section we define order and learn how we compute it with the help of Shors
quantum period-finding; the final section will explain how doing so factors M .
Order. The order of y ZM , also called the order of y (mod M ), is
the smallest positive integer, b > 1 such that
yb
(mod M ) .
Why are we so sure such an b > 1 exists? Consider all powers, y k , k = 1, 2, 3, . . ..
Since ZM is finite, so is
k
y (mod M ) k=1 ( being a ZM ) ,
and there must be many (infinitely many) distinct pairs k, k 0 with
yk = yk
(mod M ) .
For each pair, take k 0 to the the larger of the two, and write this last equality as
y k = y k+b
(mod M ),
b k 0 k > 0.
We just argued that there are infinitely many k > 1 (with potentially a different b > 0
for each k) for which the above holds. There must be a smallest b that satisfies this
among all the pairs. (Once you find one pair, take the k and b for that pair. Keep
looking for other pairs with different ks and smaller bs. You cant do this indefinitely,
since eventually youd reach b = 0. It doesnt matter how long this takes we only
need the existence of such a b, not to produce it, physically.) Assume this last equality
represents that smallest b > 0 for any k which makes it true. That means, there exists
a pair with
y k y k+b = 0
(mod M ),
b minimal over all k .
Factoring,
yk 1 yb = 0
M y k (1 y b )
M (1 y b ) .
729
(mod M )
The last step holds because y k M , since we are working inside the major case in
which we were unlucky enough to pick a y M . Were done (proving that an order b
of y exists) because the final equality means
yb
(mod M ) .
Define
a
b + 1,
which implies (plug into above to see why)
ya
(mod M ),
a minimal .
That means the function
f (x) = y x
(mod M ) ,
built upon our randomly selected y, is periodic with period a. Furthermore, it is
ZM periodic, which implies the extra condition
f (x0 ) 6= f (x)
whenever
x x0 < a .
(Review the definition of b and its minimality to verify this extra condition.)
Well also be using the fact that the period, a, is less than M . Heres why we can
assume so. The order of y, b = a 1, has to divide M , because the order of every
element in a finite group divides evenly into the size of the group, M in this case (see
elementary group theory, if youd like to research it). So, either b = M or b M/2.
In the former case, a = M 1, and we dispose of that possibility instantly by testing
whether M 1 is the period of f (x) = y x (mod M ) by evaluating it for any x and
x + (M 1). That only leaves the case b M/2 a < M .
Enter Quantum Computing We have a ZM periodic function, f (x) = y x
(mod M ), with unknown period, a < M . This is the hypothesis of Shors quantum
algorithm which we have already proved can be applied in log3 M time. This is exactly
where we would use our quantum computer in the course of factoring M .
We have picked a y ZM {0, 1} at random, defined a function based on that
y, and found its period, a. That gave us the order of y in ZM . The next, and final,
step is to demonstrate how we use the order, a, to factor M .
25.5
Using the Order of y to Find the x of our
Sufficient Condition
Weve already established that if we can find an x that satisfies
x2 = 1 (mod M ),
x 6= 1 (mod M ),
730
and do so efficiently (in polynomial time), we will get our factor q of M . We did
manage to leverage quantum period-finding to efficiently get the order of a randomly
selected y ZM , so our job is to use that order, a, without using excessive further
computation, to factor M . There are the three cases to consider:
1. a is even, and y a/2 6= 1 (mod M ).
2. a is even, and y a/2 = 1 (mod M ).
3. a is odd.
Case 1: Even a, with y a/2 6= 1 (mod M )
Claim.
y a/2
x
satisfies our sufficient condition.
Proof
x2 = y a , where a = order of y
x2 = 1 (mod M ).
(mod M ), so
Thats half the sufficient condition. The other half is x
assumption in this case is that
x
6=
(mod M ),
6=
+1
(mod M ),
6=
1 (mod M ). Our
so we only need to show that
and well have shown that this x satisfies our sufficient condition. Proceed by contradiction. What would happen if
x
+1
(mod M ) ?
y a/2
+1
(mod M ), i.e.,
y (a/2)+1
Then
(mod M ) .
If a = 2, then
y2
y
=
=
y
1
(mod M ), i.e.,
(mod M ),
731
which contradicts that y ZM {0, 1} (1 is not in that set). So we are forced to
conclude that a > 2, which gives
a
> 1, so
2
a
a a
+1 <
+
= a.
2
2 2
But now we have an a0 = (a/2) + 1 < a that satisfies
0
ya
y (a/2)+1
(mod M ) .
That contradicts that a is the order of y (mod )M , because the order is the smallest
integer with that property, by construction.
QED
That dispatches the first case; we have found an x which satisfies the sufficient
condition needed to find a factor, q, of M .
Cases 2 and 3: Even a, with y a/2 = 1 (mod M ) or Odd a
I combine these two cases because, while they are both possible, we rely on results
from number theory (one being the Chinese Remainder Theorem) which tell us that
the probability of both cases, taken together is never more than .5, i.e.,
1
.
2
This result is independent of M . That means that if we repeatedly pick y at random,
T times, the chances that we are unlucky enough to get case 2 or case 3 in all T trials
is
!
T
T
Y
^
1
1
P
case 2 case 3
=
.
2
2T
k=1
k=1
P ( case 2 case 3 )
Therefore, we end up in case 1 at least once in T trials with probability 1 1/(2T ),
independent of M , i.e., in constant time. This means we efficiently (constant time) get
case 1, and once we have that, we use Shors period-finding algorithm on a quantum
computer, which is O(log 3 (M )), to compute the period and produce x. After we have
x, another O(log 3 (M )) algorithm, GCD (on an ordinary computer), gets us M .
25.6
The Time Complexity of f (x) = y x (mod M )
There is one final detail we have not addressed. Shors quantum period-finding was
only as fast as its weakest link. Since the algorithm is log3 M only counting the logic
exterior to the quantum oracle, Uf , then any f which has a larger growth rate would
erode Shors performance accordingly. In other words, it has relativized exponential
speed up. If it is to have absolute speed-up over the classical case we must show that
the oracle, itself, is polynomial time in M . Im happy to inform you that in the case
of the factoring problem, the function f (x) = y x (mod M ) is, indeed, log3 M .
732
25.7
The Complexity Analysis
y ZM and we also saw that a < M , so we only need to consider computing y x for
both y, x < M . Since M < M 2 <= N = 2n , we can express both x and y as a sum
of powers-of-2 with, at most, n = log N terms. Lets do that for x:
x
n1
X
xk 2k ,
k=0
where the xk are xs base-2 digits. So (all products taken mod-M )
P
yk
n1
Y
xk 2k
y xk 2
k=0
k
However long it takes us to compute the general factor, y xk 2 , we need to repeat it
n times, and when those n factors are computed, we multiply them together using
(n 1) multiplications. There are two parts, taken in series, to this computation:
h
i
k
1. n complexity of y xk 2 (mod M )
2. (n 1) [complexity of a multiplication (mod M )]
The slower (not the product) of the two will determine the overall complexity of y x
(mod M ).
Preliminary Note. The computational complexity of integer multiplication is
O(log2 X), where X is the larger of the two numbers. For us, each product is bounded
by N > M , so integer multiplication costs us, at most, O(log2 N ).
25.7.1
Complexity of Step 1
k
y xk 2
y2
xk
,
k1
and in the process of computing y 2 we would end up computing y 2 , so, in order
to avoid repeated calculations from factor-to-factor, we first compute an array of the
n factors
n k on1
y2
=
1, y, y 2 , y 4 , y 8 , ... y n1 , all mod-M.
k=0
Starting with the second element, y, each element in this array is the square of the
one before. Thats a total of n 2 multiplications (we get 1 and y for free). Thus, to
produce the entire array it costs O(log N ) multiplications, with each multiplication
733
(by above note) O(log2 N ). Thats a total complexity of O(log3 N ). This is done
once for each factor in the product
n1
Y
y2
xk
k=0
To complete the computation of each of the k factors, we raise one of our pre-computed
k
y 2 to the xk power. Thats is xk multiplications for each factor. Wait a minute xk
is a binary digit, either 0 or 1, so this is nothing other than a choice; for each n we
tag on an if-statement to finish off the computation for that factor. Therefore, the
computation of each factor remains O(log3 N ).
There are n factors to compute, so this tags on another log N magnitude to the
bunch, producing a final cost for step 1 of O(log4 N ). However, have not done the
big product yet . . . .
25.7.2
Complexity of Step 2
Each binary product in the big is O(log2 N ). The big has n 1 such products,
so the final product (after computing the factors) is O((n 1) log2 N ) = O(log3 N ).
Combining Both Results
The evaluation of all n factors, O(log4 N ), is computed in series with the final product, O(log3 N ), not nested, so the slower of the two, O(log4 N ), determines the full
complexity of the oracle. Note that this was a lazy and coarse computation, utilizing
simple multiplication algorithms and a straightforward build of the f (x) = y x (mod
M ), function, and we can certainly do a little better.
As we demonstrated when covering Shors algorithm, the relationship between M
and N (N/2 < M 2 N ) implies that this is the equal to O(log4 M )
25.7.3
Absolute Complexity of Shors Factoring
(This is also stated in the lecture on Shors quantum period-finding.)
We have now shown the the oracle can be built with an efficiency that is at least
O(log 4 M ). Shors period finding had a relativized complexity of O(log 3 M ), due to
a constant time circuit sampling that contained four computational components in
series,
H n : O(log M ),
the oracle : O(log4 M ),
QF T : O(log2 M ), and
734
classical post-measurement processing (EA/CF) : O(log 3 M ).
The bottleneck is the oracle, at O(log4 M ), which we will use as our polynomial time
complexity, proving the absolute speed-up of quantum factoring. We acknowledge the
existence of faster oracles than the construction provided above, improving overall
algorithm, accordingly.
735
List of Figures
1
After measuring one location, we know them all . . . . . . . . . . . .
22
After measuring one location, we dont know much . . . . . . . . . .
23
1.1
a few numbers plotted in the complex plane . . . . . . . . . . . . . .
32
1.2
visualization of complex addition . . . . . . . . . . . . . . . . . . . .
34
1.3
the connection between cartesian and polar coordinates of a complex
number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
1.4
conjugation as reflection . . . . . . . . . . . . . . . . . . . . . . . . .
37
1.5
modulus of a complex number . . . . . . . . . . . . . . . . . . . . . .
38
1.6
ei after seconds, @ 1 rad/sec . . . . . . . . . . . . . . . . . . . . .
40
1.7
multiplication moduli multiply and args add . . . . . . . . . . . . .
43
1.8
the fifth roots-of-unity . . . . . . . . . . . . . . . . . . . . . . . . . .
46
1.9
Three of the fourth roots-of-unity (find the fourth) . . . . . . . . . .
47
2.1
A vector in R2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
2.2
Vector addition in R . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
54
2.3
Scalar multiplication in R . . . . . . . . . . . . . . . . . . . . . . . .
55
2.4
Orthogonal vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
2.5
A vector in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
2.6
A vector expressed as linear combination of x
and y
. . . . . . . . . .
61
2.7
A vector expressed as linear combination of b0 and b1 . . . . . . . . .
64
2.8
A vector expressed as linear combination of c0 and c1 . . . . . . . . .
65
3.1
Dot-product of the first row and first column yields element 1-1 . . .
75
3.2
Dot-product of the second row and second column yields element 2-2
75
3.3
Minor of a matrix element . . . . . . . . . . . . . . . . . . . . . . . .
80
3.4
The numerator of Cramers fraction . . . . . . . . . . . . . . . . . . .
85
4.1
The Cauchy sequence 1 k1 k=2 has its limit in [0, 1] . . . . . . . .
98
736
4.2
The Cauchy sequence 1 k1 k=2 does not have its limit in (0, 1) . .
4.3
Triangle inequality in a metric space.svg from Wikipedia . . . . . . . 100
4.4
A 3-D quantum state is a ray in its underlying H = C3
4.5
Dividing a vector by its norm yields a unit vector on the same ray
5.1
T mapping a vector v in R3 to a w in R3 . . . . . . . . . . . . . . . . 108
5.2
A scaling transformation . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3
3 . . . . . . . . . . . . . . . . 110
Projection onto the direction z, a.k.a. x
5.4
. . . . . . . . . . . . . . . . 110
Projection onto an arbitrary direction n
5.5
Rotation of x
counter-clockwise by /2 . . . . . . . . . . . . . . . . . 112
5.6
Rotation of y
counter-clockwise by /2 . . . . . . . . . . . . . . . . . 113
6.1
Classical angular momentum
6.2
A classical idea for spin: A 3-D direction and a scalar magnitude
6.3
Polar and azimuthal angles for the (unit) spin direction . . . . . . . . 136
6.4
A soup of electrons with randomly oriented spins . . . . . . . . . . . 137
6.5
The z-projection of one electrons spin . . . . . . . . . . . . . . . . . 138
6.6
The Classical range of z-projection of spin . . . . . . . . . . . . . . . 138
6.7
The measurements force Sz to snap into one of two values. . . . . . 139
6.8
Near vertical spin measurements give illegally accurate knowledge of
Sx , Sy and Sz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.9
Plans for a follow-up to experiment # 1 . . . . . . . . . . . . . . . . 141
98
. . . . . . 101
. 102
. . . . . . . . . . . . . . . . . . . . . . 135
. . 136
6.10 Results of follow-up measurements of Sz on the +z (|+i) and z (|i)
groups. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.11 Experiment #2: |+iz electrons enter an Sx measurement apparatus. . 142
6.12 The input state for experiment # 2 . . . . . . . . . . . . . . . . . . . 143
6.13 The x-projection of spin . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.14 Viewed from top left, the classical range of x-projection of spin. . . . 143
6.15 A guess about the states of two groups after experiment #2 . . . . . 144
6.16 |i electrons emerging from a group of |+i electrons . . . . . . . . . 145
6.17 The destruction of |+i Sz information after measuring Sx . . . . . . . 146
6.18 A spin direction with polar angle from +z, represented by |i . . . 148
6.19 The prepared state for experiment #3, prior to measurement . . . . . 149
6.20 Classical, or semi-classical expectation of experiment #3s measurement149
6.21 Probabilities of measuring |i from starting state |i, from +z
. . 150
19.1 Graph of the quintessential periodic function y = sin x . . . . . . . . 544
737
19.2 The function y = tan x blows-up at isolated points but is still periodic
(with period ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
19.3 Graph of a function defined only for x [1, 3) . . . . . . . . . . . . 544
19.4 Graph of a function defined everywhere, but whose support is [1, 3],
the closure of [1, 0) (0, 3) . . . . . . . . . . . . . . . . . . . . . . 545
19.5 A periodic function that can be expressed as a Fourier series . . . . . 546
19.6 A function with bounded domain that can be expressed as a Fourier
series (support width = 2) . . . . . . . . . . . . . . . . . . . . . . . 546
19.7 A low frequency (n = 1 : sin x) and high frequency (n = 20 : sin 20x)
basis function in the Fourier series . . . . . . . . . . . . . . . . . . . . 548
19.8 f (x) = x, defined only on bounded domain [, ) . . . . . . . . . . 549
19.9 f (x) = x as a periodic function with fundamental interval [, ): . 549
19.10First 25 Fourier coefficients of f (x) = x . . . . . . . . . . . . . . . . . 550
19.11Graph of the Fourier coefficients of f (x) = x . . . . . . . . . . . . . . 550
19.12Fourier partial sum of f (x) = x to n = 3 . . . . . . . . . . . . . . . . 552
19.13Fourier partial sum of f (x) = x to n = 50
. . . . . . . . . . . . . . . 552
19.14Fourier partial sum of f (x) = x to n = 1000 . . . . . . . . . . . . . . 553
19.15f (x) has bounded domain, but its Fourier expansion is periodic. . . . 553
19.16f = 10 produces ten copies of the period in [.5, .5) . . . . . . . . . . 558
19.17f = .1 only reveals one tenth of period in [.5, .5) . . . . . . . . . . . 558
20.1 cos x is even . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
20.2 sin x is odd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
20.3 Square-integrable wavefunctions from Wikipedia StationaryStatesAnimation.gif, leading to the absolutely integrable k2 (x) . . . . . . . . . 565
20.4 A simple function and its Fourier transform . . . . . . . . . . . . . . 566
20.5 Interpretation of from Wikipedia Standard deviation diagram)
. . 566
20.6 Graph of the Dirac delta function . . . . . . . . . . . . . . . . . . . . 568
20.7 Sequence of box functions that approximate the delta function with
increasing accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568
20.8 Sequence of smooth functions that approximate the delta function with
increasing accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
20.9 F [1] is a 0-centered delta function . . . . . . . . . . . . . . . . . . . 571
20.10F [cos x] is a pair of real-valued delta functions . . . . . . . . . . . . . 572
20.11F [sin x] is a pair of imaginary delta functions . . . . . . . . . . . . . 572
20.12Area under |f |2 and |F |2 are equal (Plancherels Theorem) . . . . . . 574
20.13sin(x) and its spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 576
738
20.14sin(3x) and its spectrum . . . . . . . . . . . . . . . . . . . . . . . . . 576
20.15A normalized Gaussian with 2 = 3 and its Fourier transform . . . . 578
20.16A more localized Gaussian with 2 = 1/7 and its Fourier transform . 578
21.1 Continuous functions on a continuous domain . . . . . . . . . . . . . 581
21.2 Sampling a continuous function at finitely many points . . . . . . . . 581
21.3 Primitive 5th root of 1 (the thick radius) and the four other 5th roots
it generates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
21.4 The DFT of a non-periodic vector . . . . . . . . . . . . . . . . . . . 589
21.5 The spectrum of a periodic vector . . . . . . . . . . . . . . . . . . . . 590
21.6 The spectrum of a very pure periodic vector . . . . . . . . . . . . . . 591
21.7 Going from an 8-element array to two 4-element arrays (one even and
one odd) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
21.8 Decomposing the even 4-element array to two 2-element arrays (one
even and one odd) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
21.9 Breaking a two-element array into two singletons . . . . . . . . . . . 601
21.10Final positions of f2 should be next to f4 . . . . . . . . . . . . . . . . 602
21.11Singletons are their own DFT s . . . . . . . . . . . . . . . . . . . . . 602
21.12Bit-reversal reorders an eight-element array . . . . . . . . . . . . . . . 603
h
i
h
i
h
i
(2)
(1)
(1)
21.13 FEO = FEOE
+ (1)j FEOO
. . . . . . . . 605
j
j (mod 1)
j (mod 1)
i
h
i
i
h
h
(2)
(2)
(4)
j
. . . . . . . . . . . 606
FEO
+ (i)
= FEE
21.14 FE
j
j (mod 2)
j (mod 2)
23.1 The spectrum of a vector with period 8 and frequency 16 = 128/8 . . 640
23.2 sin(3x) and its spectrum . . . . . . . . . . . . . . . . . . . . . . . . . 640
23.3 The spectrum of a purely periodic vector with period 8 and frequency
16 = 128/8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
23.4 Graph of two periods of a periodic injective function . . . . . . . . . . 642
23.5 Example of a periodic function that is not periodic injective . . . . . 644
23.6 We add the weak assumption that 2(+) a-intervals fit into [0, M ) . . 645
23.7 Typical application provides many a-intervals in [0, M ) . . . . . . . . 646
23.8 Our proof will also work for only one a interval in [0, M ) . . . . . . . 646
23.9 N = 2n chosen so (N/2, N ] bracket M 2 . . . . . . . . . . . . . . . . . 647
23.10Eight highly probable measurement results, cm, for N = 128 and a = 8 650
23.11Easy case covers a N , exactly . . . . . . . . . . . . . . . . . . . . . . 655
23.12[0, N ) is the union of distinct cosets of size a . . . . . . . . . . . . . . 655
739
23.13The spectrum of a purely periodic vector with period 8 and frequency
16 = 128/8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
23.14Eight probabilities, .125, of measuring a multiple of m = 16 . . . . . . 664
23.15There is (possibly) a remainder for N/a, called the excess . . . . . 670
23.16[0, N ) is the union of distinct cosets of size a, except for last . . . . . 670
23.17The final coset may have size < a . . . . . . . . . . . . . . . . . . . . 671
23.18If 0 x < N ma, a full m + 1 numbers in ZN map to f (x) . . . . . 672
23.19If N ma x < a, only m numbers in ZN map to f (x) . . . . . . . 672
23.20The spectrum of a purely periodic vector with period 10 and frequency
12.8 = 128/10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674
23.21A very long line consisting of a copies of N = 2n . . . . . . . . . . . . 678
23.22Half-open intervals of width a around each point cN
. . . . . . . . . 679
23.23Exactly one integral multiple of a falls in each interval . . . . . . . . 679
23.24Probabilities of measuring y4 = 51, y5 = 64 or y6 = 77 are dominant. . 682
23.25ybc all fall in the interval [a/2, a/2) . . . . . . . . . . . . . . . . . . . 683
23.26The chord is shorter than the arc length . . . . . . . . . . . . . . . . 685
23.27|sin(x/2)| lies above |x/| in the interval ( , ) . . . . . . . . . . . 686
23.28|sin(x/2)| lies above |Kx/| in the interval ( 1.5, 1.5 ) . . . . . . 692
23.29Exactly one integral multiple of a falls in each interval . . . . . . . . 695
23.30N = 2n chosen so (N/2, N ] bracket M 2 . . . . . . . . . . . . . . . . . 696
23.31|sin(x/2)| lies above |Lx/| in the interval ( .4714, .4714 ) . . . . 701
24.1 A wide shot of the of the convergents starting at 2/1 . . . . . . . . . 719
24.2 A close-up showing some of the latter convergents. . . . . . . . . . . . 719
24.3 First view of convergents . . . . . . . . . . . . . . . . . . . . . . . . . 720
24.4 Second view of convergents . . . . . . . . . . . . . . . . . . . . . . . . 721
24.5 Third view of convergents . . . . . . . . . . . . . . . . . . . . . . . . 721
24.6 Fourth view of convergents . . . . . . . . . . . . . . . . . . . . . . . . 721
24.7 Fifth view of convergents . . . . . . . . . . . . . . . . . . . . . . . . . 721
740
List of Tables
741
You might also like
- OCP IP OpenCoreProtocolSpecification 1.0No ratings yetOCP IP OpenCoreProtocolSpecification 1.0202 pages
- A Course in Robust Control Theory - A Convex Approach100% (1)A Course in Robust Control Theory - A Convex Approach396 pages
- Final Exam in System Identification For F and STS Answers and Brief SolutionsNo ratings yetFinal Exam in System Identification For F and STS Answers and Brief Solutions4 pages
- Linear Matrix Inequalities in System and Control Theory - Stephen Boyd100% (1)Linear Matrix Inequalities in System and Control Theory - Stephen Boyd205 pages
- Fundamentals of Dynamics and Control of Space Systems100% (1)Fundamentals of Dynamics and Control of Space Systems361 pages
- The Fuzzy Logic Control of Aircraft (Farah)No ratings yetThe Fuzzy Logic Control of Aircraft (Farah)120 pages
- An Introduction To Digital Signal Processing: Edmund M-K. LaiNo ratings yetAn Introduction To Digital Signal Processing: Edmund M-K. Lai224 pages
- Fronczak, Reniers - Model-Based Systems Engineering - 4TC00 Dictaat 2014-2015 PDFNo ratings yetFronczak, Reniers - Model-Based Systems Engineering - 4TC00 Dictaat 2014-2015 PDF105 pages
- Download ebooks file (Ebook) Non-Perturbative Field Theory: From Two Dimensional Conformal Field Theory to QCD in Four Dimensions by Yitzhak Frishman, Jacob Sonnenschein ISBN 9780521662659, 0521662656 all chapters100% (4)Download ebooks file (Ebook) Non-Perturbative Field Theory: From Two Dimensional Conformal Field Theory to QCD in Four Dimensions by Yitzhak Frishman, Jacob Sonnenschein ISBN 9780521662659, 0521662656 all chapters76 pages
- DSP Principles Algorithms and Applications Third Edition100% (1)DSP Principles Algorithms and Applications Third Edition1,033 pages
- 4 Introduction To The Divergence Theorem and Stoke's Theorem (Short)No ratings yet4 Introduction To The Divergence Theorem and Stoke's Theorem (Short)116 pages
- Frontiers in Relativistic Celestial MechNo ratings yetFrontiers in Relativistic Celestial Mech372 pages
- Report Multiprocessor Scheduling Algorithm Implementation Using Genetic AlgorithmsNo ratings yetReport Multiprocessor Scheduling Algorithm Implementation Using Genetic Algorithms98 pages
- Regression Model For Tuning PID ControllerNo ratings yetRegression Model For Tuning PID Controller11 pages
- Sham Lu J. Quantum Mechanics (Lecture Notes, 2002) (324s) - PQM0% (1)Sham Lu J. Quantum Mechanics (Lecture Notes, 2002) (324s) - PQM324 pages
- Altair Ebook SignalProcessing 1201 FinalNo ratings yetAltair Ebook SignalProcessing 1201 Final163 pages
- C++ Lecture Notes Department of Computer Science University of Cape TownNo ratings yetC++ Lecture Notes Department of Computer Science University of Cape Town127 pages
- Complete Download Image Processing The Fundamentals Second Edition Maria Petrou PDF All Chapters100% (1)Complete Download Image Processing The Fundamentals Second Edition Maria Petrou PDF All Chapters67 pages
- Lecture Notes On Nonlinear Dynamics PDFNo ratings yetLecture Notes On Nonlinear Dynamics PDF345 pages
- Graphs and Tables of the Mathieu Functions and Their First DerivativesFrom EverandGraphs and Tables of the Mathieu Functions and Their First DerivativesNo ratings yet
- Draft: Lecture Notes On Discrete MathematicsNo ratings yetDraft: Lecture Notes On Discrete Mathematics195 pages
- ECON2125/8013 Maths Notes: John Stachurski March 4, 2015No ratings yetECON2125/8013 Maths Notes: John Stachurski March 4, 2015162 pages
- Linear Algebra and Optimization For Machine LearningNo ratings yetLinear Algebra and Optimization For Machine Learning18 pages
- A Course in Robust Control Theory - A Convex ApproachA Course in Robust Control Theory - A Convex Approach
- Final Exam in System Identification For F and STS Answers and Brief SolutionsFinal Exam in System Identification For F and STS Answers and Brief Solutions
- Linear Matrix Inequalities in System and Control Theory - Stephen BoydLinear Matrix Inequalities in System and Control Theory - Stephen Boyd
- Fundamentals of Dynamics and Control of Space SystemsFundamentals of Dynamics and Control of Space Systems
- An Introduction To Digital Signal Processing: Edmund M-K. LaiAn Introduction To Digital Signal Processing: Edmund M-K. Lai
- Fronczak, Reniers - Model-Based Systems Engineering - 4TC00 Dictaat 2014-2015 PDFFronczak, Reniers - Model-Based Systems Engineering - 4TC00 Dictaat 2014-2015 PDF
- Download ebooks file (Ebook) Non-Perturbative Field Theory: From Two Dimensional Conformal Field Theory to QCD in Four Dimensions by Yitzhak Frishman, Jacob Sonnenschein ISBN 9780521662659, 0521662656 all chaptersDownload ebooks file (Ebook) Non-Perturbative Field Theory: From Two Dimensional Conformal Field Theory to QCD in Four Dimensions by Yitzhak Frishman, Jacob Sonnenschein ISBN 9780521662659, 0521662656 all chapters
- DSP Principles Algorithms and Applications Third EditionDSP Principles Algorithms and Applications Third Edition
- 4 Introduction To The Divergence Theorem and Stoke's Theorem (Short)4 Introduction To The Divergence Theorem and Stoke's Theorem (Short)
- Report Multiprocessor Scheduling Algorithm Implementation Using Genetic AlgorithmsReport Multiprocessor Scheduling Algorithm Implementation Using Genetic Algorithms
- Sham Lu J. Quantum Mechanics (Lecture Notes, 2002) (324s) - PQMSham Lu J. Quantum Mechanics (Lecture Notes, 2002) (324s) - PQM
- C++ Lecture Notes Department of Computer Science University of Cape TownC++ Lecture Notes Department of Computer Science University of Cape Town
- Complete Download Image Processing The Fundamentals Second Edition Maria Petrou PDF All ChaptersComplete Download Image Processing The Fundamentals Second Edition Maria Petrou PDF All Chapters
- Graphs and Tables of the Mathieu Functions and Their First DerivativesFrom EverandGraphs and Tables of the Mathieu Functions and Their First Derivatives
- ECON2125/8013 Maths Notes: John Stachurski March 4, 2015ECON2125/8013 Maths Notes: John Stachurski March 4, 2015
- Linear Algebra and Optimization For Machine LearningLinear Algebra and Optimization For Machine Learning