Introduction to Random Matrix Theory:

Theory and Applications

Part II

Boaz Nadler

Department of Computer Science and Applied Mathematics

The Weizmann Institute of Science

May 2013

. . . . . .

Boaz Nadler RMT

Part II: Modern Random Matrix Theory

. . . . . .

Boaz Nadler RMT

Part II: Modern Random Matrix Theory

Starting point:
[Eugene Wigner, 1955, 1958]
[Mehta and Gaudin, 1960]
random matrix models to study quantum phenomena.
Example: Energy levels of atom = eigenvalues of Hermitian
operator
Hψj = Ej ψj
Low energy states can be tackled analytically. For higher energy
levels, analysis becomes intractable.

. . . . . .

Boaz Nadler RMT

Part II: Modern Random Matrix Theory

Starting point:
[Eugene Wigner, 1955, 1958]
[Mehta and Gaudin, 1960]
random matrix models to study quantum phenomena.
Example: Energy levels of atom = eigenvalues of Hermitian
operator
Hψj = Ej ψj
Low energy states can be tackled analytically. For higher energy
levels, analysis becomes intractable.

Wigner replaced operator H by finite large random m × m

Hermitian matrix Hm .

. . . . . .

Boaz Nadler RMT

Wigner Matrix

Wigner Matrix [Hermite/Gaussian ensemble]:

- m × m Hermitian matrix W ,

. . . . . .

Boaz Nadler RMT

Wigner Matrix

Wigner Matrix [Hermite/Gaussian ensemble]:

- m × m Hermitian matrix W ,
- Wi,j independent (except symmetry constraint)
{
σ 2 i ̸= j
E[Wi,j ] = 0 E[Wi,j
2
] =
2σ 2 i = j (but σ 2 also ok)

Assume E[Wi,j
k ] < ∞ (weaking possible)

. . . . . .

Boaz Nadler RMT

Wigner Matrix

Wigner Matrix [Hermite/Gaussian ensemble]:

- m × m Hermitian matrix W ,
- Wi,j independent (except symmetry constraint)
{
σ 2 i ̸= j
E[Wi,j ] = 0 E[Wi,j
2
] =
2σ 2 i = j (but σ 2 also ok)

Assume E[Wi,j
k ] < ∞ (weaking possible)

If σ 2 = 1/m, then this is a standard Wigner Matrix.

. . . . . .

Boaz Nadler RMT

Empirical Spectral Measure

Empirical Spectral Distribution (ESD):

1
Fm (t) = #{i | ℓi (W ) ≤ t}
m

. . . . . .

Boaz Nadler RMT

Empirical Spectral Measure

Empirical Spectral Distribution (ESD):

1
Fm (t) = #{i | ℓi (W ) ≤ t}
m

Question: Does Fm converge to limiting F as m → ∞ ?

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Theorem: As m → ∞, for standard Wigner matrix,

1 √
dFm (x) → 4 − x 2 dx x ∈ [−2, 2]
4π

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Theorem: As m → ∞, for standard Wigner matrix,

1 √
dFm (x) → 4 − x 2 dx x ∈ [−2, 2]
4π

Numerical Illustration:

X = randn(m,m);

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Theorem: As m → ∞, for standard Wigner matrix,

1 √
dFm (x) → 4 − x 2 dx x ∈ [−2, 2]
4π

Numerical Illustration:

X = randn(m,m);
W = 1/sqrt(2*m) (X + X’) ;

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Theorem: As m → ∞, for standard Wigner matrix,

1 √
dFm (x) → 4 − x 2 dx x ∈ [−2, 2]
4π

Numerical Illustration:

X = randn(m,m);
W = 1/sqrt(2*m) (X + X’) ;
L = eig(W);

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Theorem: As m → ∞, for standard Wigner matrix,

1 √
dFm (x) → 4 − x 2 dx x ∈ [−2, 2]
4π

Numerical Illustration:

X = randn(m,m);
W = 1/sqrt(2*m) (X + X’) ;
L = eig(W);
histL = hist(L,x);

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

simulations: 1000 m = 400 Nbins= 64

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−2 −1 0 1 2
. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Wigner’s original proof = method of moments.

Instead of studying empirical spectral distribution Fm , study its

moments. ∫
x k dFm (x)

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Wigner’s original proof = method of moments.

Instead of studying empirical spectral distribution Fm , study its

moments. ∫
x k dFm (x)

Prove that moments converge to some limits.

. . . . . .

Boaz Nadler RMT

Wigner’s Semi-Circle Law

Wigner’s original proof = method of moments.

Instead of studying empirical spectral distribution Fm , study its

moments. ∫
x k dFm (x)

Prove that moments converge to some limits.

Find distribution corresponding to these limiting moments.

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Simple Observation:
∫ 1 ∑ 1
βm,k = x k dFm (x) = m ℓki = k
m Tr [W ].

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Simple Observation:
∫ 1 ∑ 1
βm,k = x k dFm (x) = m ℓki = k
m Tr [W ].

Since Wij are independent zero mean, many entries in W k have

zero mean as well !

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Example:
1 ∑
m
1
Tr [W ] = Wii
m m
i=1

Hence
ETr [W ] = 0.

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Example:
1 ∑
m
1
Tr [W ] = Wii
m m
i=1

Hence
ETr [W ] = 0.
2 ∑
Next, (Tr [W ]) = i,j Wii Wjj .

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Example:
1 ∑
m
1
Tr [W ] = Wii
m m
i=1

Hence
ETr [W ] = 0.
2 ∑
Next, (Tr [W ]) = i,j Wii Wjj .
Hence, ∑
E[( m1 Tr [W ])2 ] = 1/m2 i Wii2 = 1
m2
·m· 2
m = 1
m2

Conclusion: First moment βm,1 has zero mean, and variance tends
to zero as m → ∞. Thus

βm,1 → 0 as m→∞

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Second Moment: ∑ ∑ ∑
Tr [W 2 ] = i (W 2 )ii = i j Wij2
Hence, [ ]
1 1 1
E 2
Tr [W ] = · m2 = 1.
m m m

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Second Moment: ∑ ∑ ∑
Tr [W 2 ] = i (W 2 )ii = i j Wij2
Hence, [ ]
1 1 1
E 2
Tr [W ] = · m2 = 1.
m m m

More complicated calculation,

1
Var [ Tr [W 2 ]] → 0 as m→∞
m
Conclusion:
βm,2 → 1

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

General Moments:
∑
(W k )11 = W1,i1 Wi1 i2 · · · Wik−1 1
i1 ,...,ik−1

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

General Moments:
∑
(W k )11 = W1,i1 Wi1 i2 · · · Wik−1 1
i1 ,...,ik−1

For odd k = 2r + 1, at least one term appears only once,

E[Tr (W 2r +1 )] = 0.

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

General Moments:
∑
(W k )11 = W1,i1 Wi1 i2 · · · Wik−1 1
i1 ,...,ik−1

For odd k = 2r + 1, at least one term appears only once,

E[Tr (W 2r +1 )] = 0.

For even k = 2r , each term must appear even number of times for
non-zero mean.
If each term appears exactly twice, mean is (1/m)k/2
Exact number of such terms is a combinatorial graph enumeration
problem.

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Summary:

m Etr [W
1 2k+1 ] =0
1 2k
m tr [W ] → Catalan Numbers

. . . . . .

Boaz Nadler RMT

Method of Moments / Wigner Matrix

Summary:

m Etr [W
1 2k+1 ] =0
1 2k
m tr [W ] → Catalan Numbers

Distribution with these moments is semi-circle law.

. . . . . .

Boaz Nadler RMT

Universality

Limiting Distribution depends only on underlying moments

Let W be a standardized Wigner matrix whose entries come from
any distribution with finite moments,
Then as m → ∞ eigenvalues of W converge to semi-circle law.

Example: Matrix with entries ±1

X = randn(m,m);
W = sign(X+X’);
W = W / sqrt(m);
L = eig(W);

. . . . . .

Boaz Nadler RMT

Universality / Semi-Circle Law

Binomial iter: 100 m = 400 Nbins= 64

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−2 −1 0 1 2

. . . . . .

Boaz Nadler RMT

Method of Moments

Applicable to other more complicated random matrix models.

Mathematical Foundations:

. . . . . .

Boaz Nadler RMT

Method of Moments

Applicable to other more complicated random matrix models.

Mathematical Foundations:
Lemma 1: Moments uniquely determine a distribution.

. . . . . .

Boaz Nadler RMT

Method of Moments

Applicable to other more complicated random matrix models.

Mathematical Foundations:
Lemma 1: Moments uniquely determine a distribution.∫
Precisely, F - a distribution, with finite moments βk = x k dF (x)
If Carleman condition holds
∑ −1/2k
β2k =∞
k

then F is uniquely determined by {βk }

. . . . . .

Boaz Nadler RMT

Method of Moments

. . . . . .

Boaz Nadler RMT

Method of Moments

Lemma 2: under some technical conditions,

Convergence of moments → convergence of distributions.

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

Very useful tool in RMT.

Definition: f (x) probability density function.
∫ ∞
1
Sf (z) = f (x)dx z ∈ C+
−∞ x − z

where C+ = {z = x + iy | y > 0}

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

Very useful tool in RMT.

Definition: f (x) probability density function.
∫ ∞
1
Sf (z) = f (x)dx z ∈ C+
−∞ x − z

where C+ = {z = x + iy | y > 0}

Inversion Formula:
∫ b
1
F (a, b) = lim Im(S(u + iv ))du
π v →0+ a

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

Lemma: convergence of Stieltjes transforms → convergence of

distributions

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

Lemma: convergence of Stieltjes transforms → convergence of

distributions
Precisely:
- Fn sequence of probability distributions,
- sn (z) their Stieltjes transforms.
If sn (z) → s(z) for all z ∈ C+ and limv →∞ vs(v ) = −1,
then Fn → F .

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

Matrix Inversion Lemma:

∑
n
1
−1
Tr (A )=
k=1
akk − αkT A−1
k αk

where αk is k-th row of A with k-th element removed

Ak is k-th minor of A (remove k-th row and column).

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

Matrix Inversion Lemma:

∑
n
1
−1
Tr (A )=
k=1
akk − αkT A−1
k αk

where αk is k-th row of A with k-th element removed

Ak is k-th minor of A (remove k-th row and column).
Take A = Wm − zIm . Then

1 ∑ 1
m
1
Tr (A−1 ) = = sm (z).
m m ℓj − z
j=1

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

αk = vector of size (m − 1) with i.i.d. zero mean entries, unit

variance.
αk independent of Ak and hence of A−1 k .

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

αk = vector of size (m − 1) with i.i.d. zero mean entries, unit

variance.
αk independent of Ak and hence of A−1 k .
T −1
Quadratic form αk Ak αk concentrates around mean value, which
(k) (k)
is m1 Tr (Wm − zIm ) ≈ sm (z).
akk ≈ −z.

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

αk = vector of size (m − 1) with i.i.d. zero mean entries, unit

variance.
αk independent of Ak and hence of A−1 k .
T −1
Quadratic form αk Ak αk concentrates around mean value, which
(k) (k)
is m1 Tr (Wm − zIm ) ≈ sm (z).
akk ≈ −z.

Overall:
−1
sm (z) ≈
z + sm (z)

. . . . . .

Boaz Nadler RMT

The Stieltjes Transform

αk = vector of size (m − 1) with i.i.d. zero mean entries, unit

variance.
αk independent of Ak and hence of A−1 k .
T −1
Quadratic form αk Ak αk concentrates around mean value, which
(k) (k)
is m1 Tr (Wm − zIm ) ≈ sm (z).
akk ≈ −z.

Overall:
−1
sm (z) ≈
z + sm (z)
Now invert - precisely the semi-circle law.

. . . . . .

Boaz Nadler RMT

The Quarter Circle Law

[Marchenko and Pastur, 1965]

- Let X be m × n random matrix.
- Xij zero mean, unit variance.
- Look at sample covariance matrix
1
Sn = XX T
n
- its eigenvalues ℓ1 , . . . , ℓm .

. . . . . .

Boaz Nadler RMT

The Quarter Circle Law

[Marchenko and Pastur, 1965]

- Let X be m × n random matrix.
- Xij zero mean, unit variance.
- Look at sample covariance matrix
1
Sn = XX T
n
- its eigenvalues ℓ1 , . . . , ℓm .

For fixed dimension m, as n → ∞,

Sn → Σ = Im → ℓi → 1

. . . . . .

Boaz Nadler RMT

The Quarter Circle Law

[Marchenko and Pastur, 1965]

- Let X be m × n random matrix.
- Xij zero mean, unit variance.
- Look at sample covariance matrix
1
Sn = XX T
n
- its eigenvalues ℓ1 , . . . , ℓm .

For fixed dimension m, as n → ∞,

Sn → Σ = Im → ℓi → 1

Question: Eigenvalue spread as m, n → ∞ ?

. . . . . .

Boaz Nadler RMT

Simulation: Eigenvalue Spread

Example:

X = randn(m,n);
S = 1/n X X’ ;
L = eig(S) ;

. . . . . .

Boaz Nadler RMT

Simulation: Eigenvalue Spread

iter: 5000 m = 25 n = 5000 Nbins= 64

0
0.6 0.8 1 1.2 1.4 1.6

. . . . . .

Boaz Nadler RMT

Simulation: Eigenvalue Spread

iter: 5000 m = 25 n = 1000 Nbins= 64

1.5

0.5

0
0.6 0.8 1 1.2 1.4 1.6

. . . . . .

Boaz Nadler RMT

Simulation: Eigenvalue Spread

iter: 5000 m = 25 n = 500 Nbins= 64

1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0.6 0.8 1 1.2 1.4 1.6

. . . . . .

Boaz Nadler RMT

Spread of Sample Eigenvalues

Let {ℓi }m
i=1 be the eigenvalues of a random symmetric matrix H.

Empirical Spectral Distribution Function:

1
Fm (t) = #{ℓi ≤ t}
m

. . . . . .

Boaz Nadler RMT

The Quarter-Circle Law

[Marchenko & Pastur, 1967]

Let H ∼ Wm (n, Σ).
Theorem: For Σ = I , as m, n → ∞ with m/n → c, (c < 1) let ℓi
be sample eigenvalues of H/m, then

1 √
fMP (t) = (b − t)(t − a) t ∈ [a, b]
2πct
√ √
where a = (1 − c)2 , b = (1 + c)2

. . . . . .

Boaz Nadler RMT

The Quarter-Circle Law

[Marchenko & Pastur, 1967]

Let H ∼ Wm (n, Σ).
Theorem: For Σ = I , as m, n → ∞ with m/n → c, (c < 1) let ℓi
be sample eigenvalues of H/m, then

1 √
fMP (t) = (b − t)(t − a) t ∈ [a, b]
2πct
√ √
where a = (1 − c)2 , b = (1 + c)2

If c > 1, then a = 0, and there are m − n sample eigenvalues

exactly at zero.

. . . . . .

Boaz Nadler RMT

Extensions

Non-identity Covariance Matrices: X ∼ Wm (n, Σm ) where Σm

eigenvalue density converges to limiting distribution.

. . . . . .

Boaz Nadler RMT

Extensions

Non-identity Covariance Matrices: X ∼ Wm (n, Σm ) where Σm

eigenvalue density converges to limiting distribution.
Then, as n, m → ∞, empirical density of 1/nXX ′ converges to
limit. Can describe its Stieltjes transform in terms of that of
limiting Σ.

. . . . . .

Boaz Nadler RMT

Extensions

Non-identity Covariance Matrices: X ∼ Wm (n, Σm ) where Σm

eigenvalue density converges to limiting distribution.
Then, as n, m → ∞, empirical density of 1/nXX ′ converges to
limit. Can describe its Stieltjes transform in terms of that of
limiting Σ.

Double Wishart Matrices:

Let A ∼ Wm (n1 , Σ) and let B ∼ Wm (n2 , Σ).

As m, n1 , n2 → ∞, eigenvalue distribution of A−1 B converges to
limiting Marchenko-Pastur type distribution.

Remark: Eigenvalues of A−1 B same as those of

(Σ−1 A)−1 (Σ−1 B), so result independent of Σ.
Can choose Σ = I .
. . . . . .

Boaz Nadler RMT

Largest Eigenvalue of Random Matrices

Let Xm×n be random Xij ∼ N (0, 1).

Marchenko-Pastur law: limiting spectral density of eigenvalues of

1/nXX ′ .

. . . . . .

Boaz Nadler RMT

Largest Eigenvalue of Random Matrices

Let Xm×n be random Xij ∼ N (0, 1).

Marchenko-Pastur law: limiting spectral density of eigenvalues of

1/nXX ′ .

However, how large can the largest eigenvalue be ?

. . . . . .

Boaz Nadler RMT

Largest Eigenvalue of Random Matrices

Let Xm×n be random Xij ∼ N (0, 1).

Marchenko-Pastur law: limiting spectral density of eigenvalues of

1/nXX ′ .

However, how large can the largest eigenvalue be ?

Question: Bound on ℓ1 = ∥1/mXX T ∥2 ?

. . . . . .

Boaz Nadler RMT

Largest Eigenvalue of Random Matrices

Let Xm×n be random Xij ∼ N (0, 1).

Marchenko-Pastur law: limiting spectral density of eigenvalues of

1/nXX ′ .

However, how large can the largest eigenvalue be ?

Question: Bound on ℓ1 = ∥1/mXX

√
T∥ ?
2
Is it true that ℓ1 → (1 + m/n)2 ?

. . . . . .

Boaz Nadler RMT

Largest Eigenvalue of Random Matrices

√ [Wachter, 1978]
ℓ1 → (1 + m/n)2 with probability 1

. . . . . .

Boaz Nadler RMT

Largest Eigenvalue of Random Matrices

√ [Wachter, 1978]
ℓ1 → (1 + m/n)2 with probability 1

[Gordon 1985]
non-asymptotic concentration of measure bounds for largest
eigenvalue of random matrices.
many follow-up works on this topic ...

. . . . . .

Boaz Nadler RMT

Distribution of Largest Eigenvalue

[complex case, Johansson, 00’]

[real case, Johnstone, 01’]
Single Wishart Case:

Pr [ℓ1 < µnm + σnm s] → Fβ (s)

where Fβ is the Tracy-Widom distribution of order β.

Universality
under mild moment conditions, largest eigenvalue converges to
TW distribution
[Soshnikov, Tao & Vu, etc]

. . . . . .

Boaz Nadler RMT

Tracy-Widom Distribution

[Tracy & Widom, 1994, 1996]

( ∫ ∞ )
F2 (s) = exp − (x − s) q(x)dx
2
s
( ∫ ∞ )
F1 (s) = F2 (s) exp −
2
q(x)dx
s

where q(x) is the solution to Painleve-II non-linear 2nd order

differential equation

q ′′ = sq + 2q 3 , q(s) ∼ Ai(s) as s → ∞

q and Fβ are somewhat tricky to compute. Several packages.

Folkmar Bornemann (by request) - matlab code.

. . . . . .

Boaz Nadler RMT

Tracy-Widom Distributions

Tracy−Widom Densities
0.5
β=1
β=2
0.4

0.3
density

0.2

0.1

0
−5 0 5
s

. . . . . .

Boaz Nadler RMT

Tracy-Widom Distributions

Large x asymptotics:
4 3/2 ( )
e− 3 x −3/2
F2 (x) = 1 − 1 + O(x
16πx 3/2
2 3/2 ( )
e− 3 x
F1 (x) = 1 − √ 3/2 1 + O(x −3/2
4 πx

. . . . . .

Boaz Nadler RMT

Second Order Accuracy
[complex case, El-Karoui, 07]
[real case, Ma, Johnstone and Ma]
For Wishart matrices,
with careful choice of centering and scaling parameters, as
n, m → ∞ with n/m → γ,

|Pr[ℓ1 < µnm + σnm s] − Fβ (s)| ≤ Ce −cs m−2/3

. . . . . .

Boaz Nadler RMT

Second Order Accuracy
[complex case, El-Karoui, 07]
[real case, Ma, Johnstone and Ma]
For Wishart matrices,
with careful choice of centering and scaling parameters, as
n, m → ∞ with n/m → γ,

|Pr[ℓ1 < µnm + σnm s] − Fβ (s)| ≤ Ce −cs m−2/3

For real valued data

(√ √ )2
µnm = n− 1
2 + m− 1
2

 1/3
√ 1 1
σnm = µnm  √ +√ 
n− 1
2 m− 1
2
. . . . . .

Boaz Nadler RMT

Largest Eigenvalue Fluctuations

for ℓ1 /m fluctuations are O(1/m2/3 ), much smaller than standard

√
1/ m fluctuations for averages of many random variables.

TW is a new ”universal” distribution - attractor in terms of largest

eigenvalue due to universality.

. . . . . .

Boaz Nadler RMT

Distribution of the Largest Eigenvalue Divided By Trace

In various settings, interest is in following ratio

ℓ1
U= 1 ∑p > threshold
p j=1 ℓj

. . . . . .

Boaz Nadler RMT

Distribution of the Largest Eigenvalue Divided By Trace

In various settings, interest is in following ratio

ℓ1
U= 1 ∑p > threshold
p j=1 ℓj

This random variable plays a role in several different applications:

I Signal Detection [Besson & Scharf 06’, Kritchman & N. 08,
Bianchi et al. 09’]
I Two-way models of interaction [Johnson & Graybill, 72’]
I Models for Quantum Information Channels.

. . . . . .

Boaz Nadler RMT

Ratio of Largest Eigenvalue to Trace

In principle, as p, n → ∞, [Nechita 08’, Bianchi et al. 09’]

[ ]
U − µn,p
Pr < s → TWβ (s)
σn,p

. . . . . .

Boaz Nadler RMT

Ratio Distribution

β = 2, p = 20, n = 500
TW
0.5 ℓ1

0.4
density

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2
s

. . . . . .

Boaz Nadler RMT

Ratio Distribution

β = 2, p = 20, n = 500
TW
0.5 ℓ1
Ratio

0.4
density

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2
s

. . . . . .

Boaz Nadler RMT

Ratio of Largest Eigenvalue to Trace

Theorem: As, p, n → ∞,
[ ] ( )
U − µn,p 1 µn,p 2 ′′
Pr < s ≈ Fβ (s) − Fβ (s)
σn,p βnp σn,p

[N., J. Mult. Anal., 2011]

. . . . . .

Boaz Nadler RMT

Ratio of Largest Eigenvalue to Trace

Theorem: As, p, n → ∞,
[ ] ( )
U − µn,p 1 µn,p 2 ′′
Pr < s ≈ Fβ (s) − Fβ (s)
σn,p βnp σn,p

[N., J. Mult. Anal., 2011]

This distribution also relevant to study performance of various
detection methods
[N. IEEE Tran. Sig. Proc. 10’]
[N. Penna and Garello Int. Conf. Comm. 11’]

. . . . . .

Boaz Nadler RMT

 Part III:
Signal Bearning Matrices

. . . . . .

Boaz Nadler RMT

Spiked Covariance Models

Consider model whereby

Σ = diag (λ1 , λ2 , . . . , λk , 0, . . . , 0) + σ 2 Im

Spiked covariance with k spikes.

Observe n vectors xi ∈ {R, C}m from this model.

Question: What happens to largest sample eigenvalues and

eigenvectors as n, m → ∞, with k, λj fixed ?

. . . . . .

Boaz Nadler RMT

Phase Transition

[complex case, Ben-Arous, Baik, Peche]

[real case, Baik and Silverstein]
Theorem: For spike model with k spikes, as n, m → ∞ with
m/n → c, for j = 1, . . . , k,
{ ( ) √
σ2
(λj + σ 2 ) 1 + m−k λj > σ 2 m/n
ℓj → √ n λ j √
σ 2 (1 + m/n)2 λj < σ 2 m/n

Phenomena known as retarded learning in statistical physics.

. . . . . .

Boaz Nadler RMT

Phase Transition / Eigenvectors

[D. Paul 07’, Nadler 08’]

Theorem: As m, n → ∞
 √

 0 if λ < σ 2 m/n
nλ2
R (m/n) = |⟨vPCA , v⟩| =
2 2
pσ 4
−1 √

 nλ2 λ
if λ > σ 2 m/n
pσ 4
+ σ2

In statistical physics:
[Hoyle and Rattray, Reimann & al, Biehl, Watson]

. . . . . .

Boaz Nadler RMT

Phase Transition / Eigenvectors

[D. Paul 07’, Nadler 08’]

Theorem: As m, n → ∞
 √

 0 if λ < σ 2 m/n
nλ2
R (m/n) = |⟨vPCA , v⟩| =
2 2
pσ 4
−1 √

 nλ2 λ
if λ > σ 2 m/n
pσ 4
+ σ2

In statistical physics:
[Hoyle and Rattray, Reimann & al, Biehl, Watson]
√
Asymptotic n-Gaussian fluctuations [Paul, 07]
√
n(ℓ1 − E[ℓ1 ]) ∼ N (0, σ 2 (λ1 ))

. . . . . .

Boaz Nadler RMT

Phase Transition for finite p as function of σ

First, a ”thought experiment”: Take training set {xν } with finite

m, n and start increasing σ. What should be the expected behavior
of R = ⟨vPCA , v⟩ and of ℓ1 ?

λ1, λ2 as a function of σ <v ,e >

k 1
80 1

0.8
60

0.6
40
0.4

20
0.2

0 0
0 1 2 3 0 1 2 3
σ σ

λ ∼ κ2 + σ 2 (1 + m/n) R ∼ 1 − σ 2 /κ2 m/n

2
n = 50, m = 200, κ = 7.87
. . . . . .

Boaz Nadler RMT

Phase Transition as function of σ

1
∆ λ / 10
R
0.9 Rth

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3
σ

. . . . . .

Boaz Nadler RMT

Tools in Random Matrix Theory

Moment Methods
Stieltjes Transform
Concentration of Measure
Many others...

. . . . . .

Boaz Nadler RMT

 Part IV:
Signal Detection / Cognitive Radio Application

. . . . . .

Boaz Nadler RMT

Signal Detection Model

{xi }ni=1 are n i.i.d. observations from

√
x = λs(t)h + σξ(t)

. . . . . .

Boaz Nadler RMT

Signal Detection as Hypothesis Testing

Given {xi } determine between

H0 : no signal, λ = 0 vs.H1 : signal present λ > 0

. . . . . .

Boaz Nadler RMT

Signal Detection as Hypothesis Testing

Given {xi } determine between

H0 : no signal, λ = 0 vs.H1 : signal present λ > 0

H0 = null hypothesis.
H1 = alternative

. . . . . .

Boaz Nadler RMT

Test Statistics

Test statistic: A function f (x1 , . . . , xn ) → R such that

f (x1 , . . . , xn ) ≤ t → accept H0

. . . . . .

Boaz Nadler RMT

Test Statistics

Test statistic: A function f (x1 , . . . , xn ) → R such that

f (x1 , . . . , xn ) ≤ t → accept H0

f (x1 , . . . , xn ) > t → reject H0

. . . . . .

Boaz Nadler RMT

False Alarm and Power

Two common quantities of interest:

. . . . . .

Boaz Nadler RMT

False Alarm and Power

Two common quantities of interest:

False Alarm Rate:

PFA = Pr[f (x1 , . . . , xn ) > t | H0 true]

. . . . . .

Boaz Nadler RMT

False Alarm and Power

Two common quantities of interest:

False Alarm Rate:

PFA = Pr[f (x1 , . . . , xn ) > t | H0 true]

Detection Power:

PD = Pr[f (x1 , . . . , xn ) > t | H1 true]

. . . . . .

Boaz Nadler RMT

False Alarm and Power

Two common quantities of interest:

False Alarm Rate:

PFA = Pr[f (x1 , . . . , xn ) > t | H0 true]

Detection Power:

PD = Pr[f (x1 , . . . , xn ) > t | H1 true]

Typically, choose threshold t = t(α) such that PFA ≈ α for some

α ≪ 1.

. . . . . .

Boaz Nadler RMT

Optimal Tests

A test is optimal at level α, if

PD [f ] ≥ PD [any other test]

with same false alarm rate α.

. . . . . .

Boaz Nadler RMT

Optimal Tests

A test is optimal at level α, if

PD [f ] ≥ PD [any other test]

with same false alarm rate α.

Neyman-Pearson Lemma: If both H0 and H1 are simple

hypothesis, then optimal test statistic is likelihood ratio test

p(x1 , . . . , xn |H1 )
p(x1 , . . . , xn |H0 )

. . . . . .

Boaz Nadler RMT

Test Statistics for Signal Detection / Cognitive Radio

If σ is known, then H0 is simple.

However, H1 is not simple.
Neyman Pearson Lemma not directly applicable

. . . . . .

Boaz Nadler RMT

Test Statistics for Signal Detection / Cognitive Radio

If σ is known, then H0 is simple.

However, H1 is not simple.
Neyman Pearson Lemma not directly applicable

In practice: Many test statistics were proposed, all based on the

eigenvalues of the sample covariance matrix Sn .

Energy Detection: Tr (Sn ) known σ

Largest Eigenvalue: ℓ1 known σ
Max/Min Ratio: ℓ1 /ℓp unknown σ
GLRT: ℓ1 /Tr (Sn ) unknown σ
etc. etc.

. . . . . .

Boaz Nadler RMT

Optimal Test for Signal Detection ?

Given so many tests, which one to choose ?

Mathematical Principle: Look at simplest possible setting.

. . . . . .

Boaz Nadler RMT

Optimal Test for Signal Detection ?

Given so many tests, which one to choose ?

Mathematical Principle: Look at simplest possible setting.

σ - noise level known.

. . . . . .

Boaz Nadler RMT

Optimal Test for Signal Detection ?

Given so many tests, which one to choose ?

Mathematical Principle: Look at simplest possible setting.

σ - noise level known.
λ - signal strength, known.

. . . . . .

Boaz Nadler RMT

Optimal Test for Signal Detection ?

Given so many tests, which one to choose ?

Mathematical Principle: Look at simplest possible setting.

σ - noise level known.
λ - signal strength, known.

H0 is simple
H1 is still not simple

. . . . . .

Boaz Nadler RMT

Optimal Test for Signal Detection ?

Given so many tests, which one to choose ?

Mathematical Principle: Look at simplest possible setting.

σ - noise level known.
λ - signal strength, known.

H0 is simple
H1 is still not simple
but
if we consider test statistics based
only
on sample eigenvalues then it is !

. . . . . .

Boaz Nadler RMT

Which Test Statistic to use ?

Consider the case of two (nearly) simple hypothesis

H0 : Σ = I vs. H1 : W′ ΣW = I + diag (λ, 0, . . . , 0)

with λ - known. What is unknown is the basis which makes Σ

diagonal in H1 .
Suppose we use only eigenvalues {ℓj } of H as a test statistic.

. . . . . .

Boaz Nadler RMT

Which Test Statistic to use ?

Consider the case of two (nearly) simple hypothesis

H0 : Σ = I vs. H1 : W′ ΣW = I + diag (λ, 0, . . . , 0)

with λ - known. What is unknown is the basis which makes Σ

diagonal in H1 .
Suppose we use only eigenvalues {ℓj } of H as a test statistic.

Neyman-Pearson: optimal method is likelihood ratio test

p(ℓ1 , . . . , ℓm |H1 )
≷ C (α)
p(ℓ1 , . . . , ℓm |H0 )

. . . . . .

Boaz Nadler RMT

Which Test Statistic ?

From multivariate analysis (Muirhead 78’)

∏ ∏ 1
(ℓi − ℓj ) 0 F0 (− nL, Σ−1 )
(n−m−1)/2
p(ℓ1 , . . . , ℓm |Σ) = Cn,m ℓi
2
i<j

0 F0- hypergeometric function with matrix argument.

Key point: asymptotically in sample size n, for dimension m fixed,
( ) ∑
p(ℓ1 , . . . , ℓm |H1 )
log ≈ n(ℓ1 − h(λ)) + O( c1j /(ℓ1 − ℓj ))
p(ℓ1 , . . . , ℓm |H0 )

Asymptotically, as n → ∞, should only look at largest eigenvalue !

. . . . . .

Boaz Nadler RMT

Which Test Statistic ?

From multivariate analysis (Muirhead 78’)

∏ ∏ 1
(ℓi − ℓj ) 0 F0 (− nL, Σ−1 )
(n−m−1)/2
p(ℓ1 , . . . , ℓm |Σ) = Cn,m ℓi
2
i<j

0 F0- hypergeometric function with matrix argument.

Key point: asymptotically in sample size n, for dimension m fixed,
( ) ∑
p(ℓ1 , . . . , ℓm |H1 )
log ≈ n(ℓ1 − h(λ)) + O( c1j /(ℓ1 − ℓj ))
p(ℓ1 , . . . , ℓm |H0 )

Asymptotically, as n → ∞, should only look at largest eigenvalue !

Roy’s Largest Root Test

[Kritchman & N., IEEE-TSP, 09’]
. . . . . .

Boaz Nadler RMT

Using Roy’s Largest Root Test

p=25, K=2, λ = [1 0.4]

1

MDL
0.9
AIC
0.8

0.7
≠ K)

0.6
est

0.5
Pr(K

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
N

. . . . . .

Boaz Nadler RMT

Using Roy’s Largest Root Test

p=25, K=2, λ = [1 0.4]

1

MDL
0.9
AIC
0.8 SCHOTT
MODIFIED−AIC
0.7
≠ K)

0.6
est

0.5
Pr(K

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
N

. . . . . .

Boaz Nadler RMT

Using Roy’s Largest Root Test

p=25, K=2, λ = [1 0.4]

1

MDL
0.9
AIC
0.8 SCHOTT
MODIFIED−AIC
0.7
RMT
≠ K)

0.6
est

0.5
Pr(K

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
N

. . . . . .

Boaz Nadler RMT

Properties of Roy’s Test

Detection Power of Roy’s test

via Random Matrix Theory

. . . . . .

Boaz Nadler RMT

Properties of Roy’s Test

Detection Power of Roy’s test

via Random Matrix Theory

Via largest
√ root test, asymptotically for m, n large only
λ/σ 2 > m/n can be detected (with probability one) !
Via other tests, can detect (but not with probability one) weaker
signals
[Onatsky et. al.]
Final word of caution: Roy’s largest root test is asymptotically
optimal when n → ∞. Not so for very weak signals.

. . . . . .

Boaz Nadler RMT

What we did not cover

A lot !

. . . . . .

Boaz Nadler RMT

What we did not cover

A lot !

- Free Probability (Voiculescu)

- Determinantal Processes
- Kernel random matrices
- Random matrices with heavy tailed distributions
- Random matrices with dependent (correlated) entries
- Random Graphs
- Other transforms
- Concentration of Measure (non-asymptotic finite sample bounds)
- relation to statistical physics
- Rate of convergence
- Linear spectral statistics, Central Limit Theorems - etc etc.

. . . . . .

Boaz Nadler RMT

Some (Recent) References

much more material can be found in:

Multivariate Statistics:
- A. T. James, Distributions of matrix variates and latent roots derived
from normal samples, Ann. Math. Statist., vol. 35, 475-501, 1964.
- T.W. Anderson, An introduction to multivariate statistical analysis,
Wiley, 2003.
- R.J. Muirhead, Aspects of Multivariate Statistical Theory, 2005.
Random Matrices / Recent Books:
- G. Anderson, A. Guionnet, O. Zeitouni, An Introduction to Random
Matrices, Cambridge, 2009.
- A. Tulino, S. Verdu, Random Matrix Methods and Wireless
Communications, 2011.
- Z. D. Bai, J. W. Silverstein, Spectral Analysis of Large Dimensional
Random Matrices, Springer, 2009.

. . . . . .

Boaz Nadler RMT

Some (Recent) References

Largest Eigenvalue Distribution:

- I.M. Johnstone, On the distribution of the largest eigenvalue in principal
component analysis, Ann. Stat., 29:295–327, 2001.
- I.M. Johnstone, Approximate Null Distribution of the Largest Root in
Multivariate Analysis, Ann. Applied Stat, 2009
- N. El Karoui. A rate of convergence result for the largest eigenvalue of
complex white Wishart matrices. Annals of Probability, 2006.
- B. Nadler, On the distribution of the ratio of the largest eigenvalue to
the trace of a Wishart matrix, J. Mult. Anal., 2010.

. . . . . .

Boaz Nadler RMT

Some (Recent) References

Largest Eigenvalue Signal + Noise:

- J. Baik, G. Ben-Arous, S. Peche, Phase transition of the largest
eigenvalue for non-null complex sample covariance matrices, Ann. Prob.,
2005.
- J. Baik, J. W. Silverstein, Eigenvalues of large sample covariance
matrices of spiked population models, J. Mult. Anal., 2006.
- D. Paul, Asymptotics of sample eigenstructure for a large dimensional
spiked covariance model, Stat. Sinica, 2007.
- B. Nadler, Finite sample approximation results for PCA, Ann. Stat.,
2008.
- R.R. Nadakuditi and J. W. Silverstein. Fundamental Limit of Sample
Generalized Eigenvalue Based Detection of Signals in Noise Using
Relatively Few Signal-Bearing and Noise-Only Samples. IEEE Journal of
Selected Topics in Signal Processing, 2010.
- I. M. Johnstone, B. Nadler, Tech. Report, Stanford University, 2011.

. . . . . .

Boaz Nadler RMT

The End

Obrigado pela sua atencao !

more material at
[Link]/∼nadler

. . . . . .

Boaz Nadler RMT