Signal Processing for Comm.

Systems HO #23
Fall 2000-2001/Yumin Lee Page 1

LEAST-SQUARES CHANNEL ESTIMATION

Problem formulation for least-squares channel estimation

! The correlator-estimator is a heuristic channel estimator, in the sense that it does

not attempt to minimize any criterion and is thus not optimal in any sense. The
least-squares channel estimator, on the other hand, is optimal in the least-squares
sense.
! Consider the discrete-time equivalent model
ν
y k = ∑ p j x k − j + η k , k = 0,1,2, L
j =0

Let

 yν + N −1 
y 
y≡  ν + N −2 

 M 
 
 yν 

Then

 xν + N −1 xν + N − 2 ... x N −1   p0  ην + N −1 
x xν + N −3 ... x N − 2   p1  ην + N −2 
y =  ν + N −2   +   ≡ Xp + η
 ... ... ... ...   M   ... 
    
 xν xν −1 ... x0   pν   ην 

Assuming that X is known. We would like to find a channel estimate p̂ based

2
on the observation y, such that y ≈ X pˆ , i.e., y − X p̂ is minimized.

! Note that in order to completely specify X, we need x0, x1, …, xN+ν−1 (a total of
N+ν training symbols).

Solution for the Least-Squares Channel Estimator

! Since Xpˆ is a linear combination of the columns of X with coefficients taken
from the vector p̂ , therefore we are essentially looking for the linear
combination of the columns of X that is “closest” (in Euclidean distance) to the
observation vector y.
! In other words, we want to find the vector in the column space of X that is
closest in Euclidean distance to y. Therefore, the optimal choice of Xpˆ is the
projection of y on the column space of X.

Please report errors to [email protected].

Signal Processing for Comm. Systems HO #23
Fall 2000-2001/Yumin Lee Page 2

! Therefore, the vector p̂ must be chosen in such a way that ( ( y − Xpˆ ) is

orthogonal to every column of X. (Note: This is the orthogonality principle.)
In other words, we must have

(y − Xpˆ )H X = [0 0 L 0],

or
X H (y − Xpˆ ) = 0 .
Thus, the least-squares channel estimate is given by

pˆ = (X H X ) X H y
−1

! The matrix (X∗X)−1XH is referred to as the pseudo-inverse of X if X is a tall

matrix (i.e., more rows than columns). Note that (XHX)−1XH is a (ν+1)×N
matrix that can be pre-computed. Therefore least-squares channel estimation is
simply a matrix-vector multiplication.
! An Example: ν=1, N=7
Assumet that (x0, x1, …, x6, x7) = (1, −1, −1, 1, −1, 1, 1, 1). Then, we have
1 1
1 1
 
 1 − 1
 
X = − 1 1 
 1 − 1
 
− 1 − 1
− 1 1 
Thus
 7 − 1
XHX =  ,
− 1 7 
and
1 7 1 1 1 1 − 1 1 − 1 − 1
(X H
X) XH =
−1

48 1 7 1 1 − 1 1 − 1 − 1 1 

8 8 6 6 6 8 6
 48 48 48 − 48 48 − − .
= 48 48
8 8 6 6 6 8 6 
 − − − 
 48 48 48 48 48 48 48 
Therefore

1 4 y 7 + 4 y 6 + 3 y5 − 3 y 4 + 3 y 3 − 4 y 2 − 3 y1 
pˆ =
24 4 y 7 + 4 y 6 − 3 y5 + 3 y 4 − 3 y 3 − 4 y 2 + 3 y1 

Note that the least-squares channel estimator is a linear and time-varying system.

Please report errors to [email protected].

Signal Processing for Comm. Systems HO #23
Fall 2000-2001/Yumin Lee Page 3

Requirement on Training Sequence for Least-Squares Channel Estimation

! The least-squares channel estimator is given by

pˆ = (X H X ) X H y .
−1

Since
y = Xp + η
therefore

pˆ = (X H X ) X H (Xp + η ) = p + (X H X ) X Hη
−1 −1

! For channel estimation, the number of observations N must be greater than or

equal to the number channel taps ν+1, i.e., X must be a tall matrix. Usually, we
need N ≥ 2ν to get good performance.
! η=0), then pˆ = p regardless of the choice of X.
If there is no noise (η
Furthermore, if E[η η]=0, then we have E (pˆ ) = p , i.e., the least-squares channel
estimate is unbiased regardless of the choice of X.
! The estimation error vector is given by

ε = pˆ − p = (X ∗ X ) X ∗η .
−1

The error covariance matrix is given by

Λ ≡ E εε[ H
] = E [(X X ) X ηη X (X X )
H −1 ∗ H H −1
]
= (X X ) X E [ηη ]X (X X )
H −1 H H H −1

ηηH] = σ2I, then we have

Assume that E[η

Λ = σ 2 (X H X ) (X H X )(X H X ) = σ 2 (X H X )
−1 −1 −1

! What criterion should the training sequence satisfy in order to get good
ν
Note that tr (Λ ) = ε = ∑ E pˆ j − p j
2 2
performance? is the sum of the
j =0

mean-square estimation error of each channel tap. On the other hand, if we

further assume that x0, x1, …, xN+ν−1 is taken from a periodic sequence with
period N, then it is easily verified that all the diagonal elements of (XHX) are
N −1 2

equal to ∑X
j =0
j . Therefore, the tr(XHX) is proportional to the energy of the

training sequence. Thus, we would like to design the training sequence such
Λ) is minimized, under the constraint that tr(XHX) is a constant.
that tr(Λ

Please report errors to [email protected].

Signal Processing for Comm. Systems HO #23
Fall 2000-2001/Yumin Lee Page 4

! Let A = XHX, and let λi, i=0,1,2,…, ν be the eigenvalues of A. Then we have
ν
tr (A ) = ∑ λi
i =0

and
ν
tr (Λ) = σ 2 tr (A −1 ) = σ 2 ∑
1
i =0 λi
ν
1
Therefore, we would like to design the training sequence such that σ 2 ∑ is
i =0 λi
ν
minimized under the constraint that ∑λ
i =0
i is a constant. This goal can be

achieved by requiring the eigenvalues of XHX to be the same. In this case XHX
is a diagonal matrix, and the optimal training sequence should satisfy
 N −1 2

 ∑ xi 0 ... 0
 i =0 
 N −1 2 
X X= 0
H

∑xi =0
i ... 0 .

 ... ... ... ... 
 N −1 2

 0 ... 0 ∑ xi 
 i =0 
! In other words, in order to obtain good performance, the (deterministic)
autocorrelation of the training sequence should be very small within the window
[0,ν]. This requirement is very similar as that for the correlator-estimator, and
will also be referred to as the whiteness requirement.

Performance of the Least-Squares Channel Estimator

! A measure of the performance of a channel estimator is the mean-square

estimation error per channel tap, defined as
2
Eε
σ ≡ 2
p
(ν + 1)
! Assume that the whiteness requirement is satisfied, and that
N −1 2
1
N
∑x
i =0
i = EX ,

then we have
tr (Λ) σ2
σ p2 ≡ =
(ν + 1) NE X

Please report errors to [email protected].

Signal Processing for Comm. Systems HO #23
Fall 2000-2001/Yumin Lee Page 5

! Note that the mean-square estimation error per channel tap is inversely
proportional to the signal-to-noise ratio EX /σ2 as one would expect.
Furthermore, the mean-square estimation error per channel tap is also inversely
proportional to the length N of the training sequence.

Please report errors to [email protected].