The invention is directed, in general, to wireless communications and, more specifically, to a system and method for blind identification of multichannel finite impulse response (FIR) filters using an iterative structured total least-squares technique.
Blind channel identification (BCI) refers to the estimation of channel impulse response from only its output. Over the past few years, this area of research has received immense practical interest (see, e.g., Tong, et al., “Multichannel Blind Identification: From Subspace to Maximum Likelihood Methods,” Proc. IEEE, vol. 86, no. 10, pp. 1951-1968, October 1998, and references therein). The primary reason for this is the fact that BCI does not require a training sequence to equalize the channel, thereby saving the channel capacity, which is severely limited in mobile communications. Some early research in this direction focused on using higher-order statistics (HOS), which require the use of long data records for reliable estimates. This limits their usage in mobile communications where the channel is desired to be estimated within a short period of time.
Blind estimation of FIR channels using only second-order statistics (SOS) is first attributed to Tong, et al., who showed that by sampling the received data at a rate higher than the baud (symbol) rate, SOS can suffice to estimate the channel impulse response up to a constant (Tong, et al., “A New Approach to Blind Identification and Equalization of Multipath Channels,” in Proc. 25th Asilomar Conf. Signals, Syst. Comp., 1991, pp. 856-860). Since then, many different statistical and deterministic approaches for BCI have been studied, each having its own relative merits and demerits (e.g., Tong, et al., “Multichannel . . . ,” supra).
One aspect of the invention provides a system for blind identification of multichannel FIR filters. In one embodiment, the system includes: (1) a decomposition generator configured to construct, from noisy symbols containing data, a block-Hankel structured matrix and a decomposition thereof, (2) an iterative matrix re-estimator coupled to the decomposition generator and configured iteratively to re-estimate singular vectors corresponding to a scalar that satisfies the block-Hankel matrix until a sufficient convergence occurs, (3) a structured low-rank estimate matrix generator coupled to the iterative matrix re-estimator and configured to generate a block-Hankel structured low-rank estimate matrix related to the block-Hankel structured matrix and (4) an STLS channel estimator coupled to the structured low-rank estimate matrix generator and configured to obtain STLS channel estimates from the block-Hankel structured low-rank estimate matrix.
Another aspect of the invention provides a method of blind identification of multichannel finite impulse response filters. In one embodiment, the method includes: (1) constructing a block-Hankel structured matrix and a decomposition thereof from noisy symbols containing data, (2) iteratively re-estimating singular vectors corresponding to a scalar that satisfies the block-Hankel matrix until a sufficient convergence occurs, (3) generating a block-Hankel structured low-rank estimate matrix related to the block-Hankel structured matrix and (4) obtaining STLS channel estimates from the block-Hankel structured low-rank estimate matrix.
In another embodiment, the method includes: (1) computing a decomposition of a block-Hankel structured matrix containing noisy data from multiple channels, (2) computing an eigenvector corresponding to the smallest eigenvalue of the block-Hankel matrix to yield initial estimates for singular vectors corresponding to a scalar that satisfies the block-Hankel matrix, (3) using re-estimated singular vectors to re-estimate symmetric positive definite matrices related to the decomposition until sufficient convergence occurs and (4) generating a block-Hankel structured low-rank estimate matrix related to the block-Hankel structured matrix.
For a more complete understanding of the invention, reference is now made to the following descriptions taken in conjunction with the accompanying drawings, in which:
1. Introduction
Described herein are iterative STLS-based techniques for performing BCI without assuming any knowledge of the input signal statistics. This is particularly realistic in mobile communications, where a short data sequence may not yield reliable statistics (see, e.g., Hua, “Fast Maximum Likelihood for Blind Identification of Multiple FIR Channels,” IEEE Trans. Signal Processing, vol. 44, no. 3, pp. 661-672, March 1996). A deterministic least-squares (LS) technique was proposed in Xu, et al., “A Least-Squares Approach to Blind Channel Identification,” IEEE Trans. Signal Processing, vol. 43, no. 12, pp. 2982-2993, December 1995, which is based on the cross-relation property between the outputs of two channels fed by the same input. A striking feature of the LS technique is that for an arbitrary input and in the absence of noise, it yields exact channel estimates. In the presence of noise, the channel parameter estimates are obtained by solving an over-determined set of linear equations which translates to obtaining a reduced rank solution of a block-Hankel structured matrix. Unfortunately, in the presence of noise, the LS technique fails to associate this specific structure with the reduced rank matrix. In the discussion herein, the LS solution is constrained so that the reduced rank data matrix also assumes a block-Hankel structure.
This constrained optimization problem is referred to as structured total least-squares (STLS). STLS provides an advantageous framework for approximating a full-rank affinely-structured matrix with a rank-deficient matrix having the same affine structure. STLS is not new in itself. However, various prior art attempts to employ STLS have either fallen short by failing to provide structured constraint (see, e.g., Gunther, et al., “A Structured Least-Squares Approach to Blind Channel Identification and Equalization,” in Proc. 37th Asilomar Conf. Signals, Syst. Comp., 2003, pp. 45-49)) or requiring the simultaneous solution of a set of nonlinear equations.
Described herein is an iterative STLS-based BCI technique for estimating channel parameters. The novel iterative technique both provides structured constraint and obviates the need to solve simultaneous nonlinear equations. The block-Hankel affine structure of the data matrix, which empowers STLS-based techniques in general, will first be fully exploited. An iterative nonlinear solution to the unknown channel parameters is then derived by using a generalized form of singular value decomposition (SVD). Extensive numerical simulations are then carried out to compare the performance of the iterative STLS-based BCI technique against the well-known LS technique, where the affine structure of the date matrix is overlooked. These results reveal that the iterative STLS-based BCI technique outperforms the LS technique for both well-conditioned and ill-conditioned channels over a wide range of signal-to-noise ratio (SNR).
2. Problem Formulation
The mobile communication device 100 also contains data processing circuitry 130 coupled to the transmitter 110 and the receiver 120 and configured to queue data for transmission, buffer received data and otherwise manipulate data for the benefit of a user (not shown). The data processing circuitry 130 may be employed to carry out various embodiments of the iterative STLS-based technique claimed herein. Thus, the iterative STLS-based technique may be embodied in a sequence of executable software instructions. A keyboard, keypad, microphone, touch-sensitive screen or other input device (not shown) may be associated with the data processing circuitry 130 and configured to accept input from a user (not shown). A screen, speaker, vibrator or other output device (not shown) may also be associated with the data processing circuitry 130 and configured to provide output to a user (not shown).
where L is the order of the P FIR filters, i.e., hi(z)=Σl=0Lhi(l)z−l, assumed to be known. Of course, the P-channel model can be obtained either by using P physical receivers or by oversampling each channel output at P times the baud rate (e.g., Tong, et al., “Multichannel . . . ,” supra).
The overall objective of the iterative STLS-based technique is to estimate the channels hi(n) from N output samples yi(n) of P channels i=1, . . . ,P. For purposes of describing the embodiments expressly set forth herein, the following three identifiability assumptions are made: A(1) the P FIR filters are coprime; i.e., they do not share any common zeros; A(2) the number of modes in the input sequence is L+1, and A(3) the additive noise ni(n) is stationary and white.
For a pair of noise-free outputs of any two sensors i and j, the following can be written:
x
i(n)=hi(n)*s(n), and
x
j(n)=hj(n)*s(n),
which yields:
h
i(n)* xj(n)=hj(n)*xi(n).
This may be restated in a compact matrix notation:
where hm=[hm(L), . . . ,hm(0)]T,
and (·)T denotes the transpose. The cross-relation in Equation (2) is the main idea behind both the LS approach of Xu, et al., supra, and the STLS approach, as will be shown in the next two sections.
3. Motivation for STLS-Based Approach
The cross-relation of Equation (2) can straightforwardly be expressed for each (i,j) pair of channels, which when combined together gives:
or simply Xh=0, where X is of dimension
and h is a vector of length (L+1)P. Each constituent block of X is of dimension (N−2L+1)(P−i)×(L+1)P and is given by:
With the identifiability assumptions (A1)-(A3) satisfied, the matrix X, also sometimes referred to as the data selection matrix, is rank-deficient by one, and h is in its null space. Given a noisy data matrix Y, an estimate of the channel parameters can then be obtained by finding the low-rank approximation Ŷ of the noisy data selection matrix Y. In other words, a LS solution as follows is sought:
where ∥·∥ is the Frobenius norm. The first constraint in Equation (6) guarantees the rank-deficiency of Y, and the second constraint avoids the trivial solution h=0. Using the Eckart-Young-Mirsky theorem (see, e.g., Golub, et al., Matrix Computations, Johns Hopkins University Press, Baltimore, Md., 1996), the LS solution can be obtained from the dyadic SVD of Y, i.e.:
Ŷ=Y−uσv
T, and h=v,
where u and v are the left and right singular vectors of the matrix Y corresponding to its smallest singular value σ.
The matrix Y has a unique block-Hankel structure associated with it. However, its low-rank approximant Ŷ, obtained using Equation (6), is not guaranteed to have the same structure. In the BCI problem, the matrix Ŷ is desired to have a block-Hankel structure as in Equation (4). Motivated by this fact, what is needed is an estimate Ŷ guaranteed to have the same structure as Y. In this case, the solution belongs to a set of matrices S, which satisfies the following property:
By adding another constraint, the optimization problem can no longer be solved simply using SVD. In fact, as will be shown below, a nonlinear estimation procedure is required to solve this problem.
4. Iterative STLS-Based BCI
For the sake of simplicity, a two-channel case, i.e., P=2, will be considered. The matrix X, therefore, consists of only two row sub-blocks X2 and −X1 and is a function of 2(N−L+1) data values xi(L), . . . ,xi(N) for i=1,2. In terms of a set of fixed matrices Bi, i=0, . . . ,2N−2L+1, this affine matrix can be expressed as:
B(x)=X=x2(L)B0+x2(L+1)B1+ . . . +x2(N)BN−L+x1(L)BN−L+1+ . . . +x1(N)B2N−2L+1 (7)
where Bi (i=0, . . . ,N−L+1) has a one along the (i+1)th anti-diagonal of the left (N−2L+1)×(L+1) block and zeros elsewhere. Similarly, Bi (i=N−L+2, . . . ,2N−2L+1) has a minus one along the (i+1)th anti-diagonal of the right (N−2L+1)×(L+1) block and zeros elsewhere. For example:
Based on the above formulation, and given the noisy data yi(n), an STLS criteria is defined to estimate the block-Hankel structured matrix B(ŷ), which consists of minimizing the following cost function:
Using Lagrange's multipliers, a solution to Equation (8) is obtained by finding vectors u and v corresponding to the smallest scalar σ that satisfies Equation (4):
Yv=Dvuσ, uTDvu=1, and (9)
Ytu=Duvσ, vTDuv=1, (10)
where Du and Dv are symmetric positive definite matrices defined as:
The channel estimates are given by ĥ=v/∥v∥. Note that the nonlinearity of Equations (9) and (10) stems from the dependence of Du and Dv on the square of the singular vectors u and v, respectively. This makes an analytical solution to Equations (9) and (10) rather difficult to compute. Abatzoglou, et al., “The Constrained Total Least Squares and Its Application to Harmonic Superresolution,” IEEE Trans. Signal Processing, vol. 39, no. 5, pp. 1070-1087, May 1991, and De Moor, “Total Least Squares for Affinely Structured Matrices and Noisy Realization Problems,” IEEE Trans. Signal Processing, vol. 42, no. 11, pp. 3104-3113, November 1994, propose iterative techniques to solve this nonlinear optimization problem. De Moor's technique, however, possesses better convergence properties. Du and Dv are treated as independent of u and v in each iteration in which a QR decomposition of Y is computed, i.e.:
where Q1 is of dimension (N−2L+1)×2(L+1), Q2 is of dimension (N−2L+1)×(N−1) and R is of dimension 2(L+1)×2(L+1). Equations (9) and (10) can then be manipulated to obtain the following upper triangular form:
where u=Q1z+Q2w for certain arbitrary vectors z and w of lengths 2(L+1) and (N−1), respectively. This set of equations allows one to efficiently compute u and v, which are then used to update Du and Dv for the next iteration.
The novel iterative STLS-based BCI technique can be summarized as follows:
z
(k)
=R
−T
D
u
(k−1)
v
(k−1)σ(k−1) 5.
w
(k)=−(Q2TDv(k−1)Q1)−1(Q2TDv(k−1)Q1)z(k) 6.
u
(k)
=Q
1
z
(k)
+Q
2
w
(k) 7.
v
(k)
=R
−1
Q
1
T
D
v
(k−1)
u
(k)
, v
(k)
=v
(k)
/∥v
(k)∥ 8.
γ(k)=(u(k)
u
(k)
=u
(k)/γ(k), v(k)=v(k)/γ(k) 10.
D
u
(k)
=D
u
(k)/(γ(k))2, Dv(k)=Dv(k)/(γ(k))2 11.
σ(k)=u(k)
13. Convergence: If ∥v(k)−v(k−1)∥≦ε (a predetermined scalar), stop. Otherwise, repeat steps 5-13.
At convergence (k=K), the structured low-rank matrix Ŷ can be estimated using:
for i=0, . . . ,N−L. Finally, the STLS channel estimates h are obtained from the null space of B(ŷ).
The above-described iterative STLS-based technique may be expressed in a flow diagram.
5. Simulation Results
The performance of the novel iterative STLS-based technique will now be compared against the conventional LS technique. To study the effect of ill-conditioned channels, a two-channel second-order FIR model (L=2, P=2) proposed in Hua, et al., supra, is considered:
h
i=[1−2 cos(θ1)1]T, i=1,2,
where θ1=θ, θ2=θ1+δ, and δ is the parameter that controls the angular separation between the zeros of the two channels. The channels are considered ill-conditioned when δ is small. The input s(n) is considered as an independent and identically distributed (i.i.d.) sequence of (+1,−1) having unit variance σs2=1. The output data consist of N=100 samples. The outputs of the two channels are assumed to be corrupted by additive white Gaussian noise of variance σn2. The SNR in dB is defined using:
The performance of STLS and LS techniques is evaluated using the mean-squared error (MSE) in dB:
where ĥi is the ith estimate of channel parameters, and Nr=100 is the number of independent Monte-Carlo runs. To alleviate the effect of scaling in the estimates, the first element of ĥi is normalized to one.
The MSE for ill-conditioned channels (θ=π/10, δ=π/5) at different SNR levels was computed.
For ill-conditioned channels, the STLS technique also proves to be more data-efficient.
Those skilled in the art to which the invention relates will appreciate that other and further additions, deletions, substitutions and modifications may be made to the described embodiments without departing from the scope of the invention.
This application claims the benefit of U.S. Provisional Application Ser. No. 61/026,040, filed by Ikram on Feb. 4, 2008, entitled “Method for Blind Identification of Multichannel FIR Filters Using Structured Total Least-Squares,” commonly assigned with the invention and incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
61026040 | Feb 2008 | US |