Backhaul signal compression through spatial-temporal linear prediction

Abstract

The technology in this application compresses multi-antenna, complex-valued signals by exploiting both a spatial and a temporal correlation of the signals to remove redundancy within the complex-valued signals and substantially reduce the capacity requirement of backhaul links. At a receiver, the compressed signal is received, and a decompressor decompresses the received signal over space and over time to reconstruct the multiple antenna stream.

Claims

1. A decompression method, comprising the steps of: receiving a compressed radio signal that corresponds to a multi-antenna signal, the multi-antenna signal including information associated with a user communication received over multiple radio antennas; decompressing the compressed radio signal based on correlations in both space and in time to reconstruct a representation of the multi-antenna signal that is complex-valued, the correlations in both space and in time operable to remove redundancy within the complex-valued signals, and providing a reconstructed representation of the multi-antenna signal for further processing or output.

2. The method of claim 1, wherein the correlations comprise a correlation in space and an independent correlation in time.

3. The method of claim 1, wherein the correlations comprise a joint correlation in space and time.

4. The decompression method in claim 1, wherein the reconstructed representation of the multi-antenna radio signal is sampled and multi-dimensional.

5. The decompression method in claim 1, wherein: the multi-antenna signal includes a plurality of antenna signals, each comprising information received by a different one of the multiple antennas; the compressed radio signal includes, for each antenna signal, an error signal indicating an error between the antenna signal and a prediction of the antenna signal; and the decompressing includes: converting the error signals from a digital format to an analog format applying an inverse spatial linear transform to the error signals to generate corresponding quantized error signals, and performing infinite impulse response filtering on the quantized error signals to generate reconstructed representations of the multiple antenna signals.

6. The decompression method in claim 5, wherein the inverse spatial linear transform includes fixed, predetermined inverse transform coefficients corresponding to an inverse discrete-cosine transform (DCT), an inverse discrete Fourier transform (DFT), or an inverse discrete wavelet transform (DWT).

7. The decompression method in claim 5, wherein the inverse spatial linear transform includes adaptively computed inverse transform coefficients, and wherein the method further comprises receiving the adaptively computed inverse transform coefficients from a transmitting node transmitting the compressed radio signal.

8. The decompression method in claim 5, wherein the inverse spatial linear transform includes inverse transform coefficients corresponding to an inverse Kahunen-Loeve transform (KLT).

9. The decompression method in claim 5, wherein the infinite impulse response filtering includes: summing the error signals with corresponding predicted antenna signals to generate the reconstructed representations of the multiple antenna signals.

10. The decompression method in claim 9, wherein the infinite impulse response filtering further comprises: filtering the reconstructed representations of the multiple antenna signals using a spatial temporal prediction matrix of predictive coefficients to generate the predicted antenna signals.

11. The decompression method in claim 10, wherein the matrix of predictive coefficients is estimated based on empirical moving averages of (1) a cross-correlation of the multiple antenna signals and the reconstructed representations of the multiple antenna signals and (2) an auto-correlation of the reconstructed representations of the multiple antenna signals.

12. The decompression method in claim 10, wherein the matrix of predictive coefficients is estimated based on recursive empirical averages of (1) a cross-correlation of the multiple antenna signals and the reconstructed representations of the multiple antenna signals and (2) an auto-correlation of the reconstructed representations of the multiple antenna signals.

13. The decompression method in claim 11, further comprising receiving the matrix of predictive coefficients from a transmitting node.

14. Decompression apparatus, comprising: a receiver configured to receive a compressed radio signal that corresponds to a multi-antenna signal, the multi-antenna signal including information associated with a user communication received over multiple radio antennas; one or more processors configured to decompress the compressed signal based on correlations in both space and in time to reconstruct a representation of the multi-antenna signal that is complex-valued, the correlations in both space and in time operable to remove redundancy within the complex-valued signals; and an output terminal configured to provide the reconstructed representation of the multi-antenna signal for further processing or output.

15. The decompression apparatus in claim 14, wherein the correlations comprise a correlation in space and an independent correlation in time.

16. The decompression apparatus in claim 14, wherein the correlations comprise a joint correlation in space and time.

17. The decompression apparatus in claim 14, wherein the reconstructed representation of the multi-antenna radio signal is sampled and multi-dimensional.

18. The decompression apparatus in claim 14, wherein: the multi-antenna signal includes a plurality of antenna signals, each comprising information received by a different one of the multiple antennas; the compressed signal includes, for each antenna signal, an error signal indicating an error between the antenna signal and a prediction of the antenna signal, and wherein the decompression apparatus further includes: an analog-to-digital converter configured to convert the error signals from a digital format to an analog format, transform circuitry configured to apply an inverse spatial linear transform to the error signals to generate corresponding quantized error signals, and a filter configured to perform infinite impulse response filtering on the quantized error signals to generate the reconstructed representations of the multiple antenna signals.

19. The decompression apparatus in claim 18, wherein the inverse spatial linear transform includes fixed, predetermined inverse transform coefficients corresponding to an inverse discrete-cosine transform (DCT), an inverse discrete Fourier transform (DFT), or an inverse discrete wavelet transform (DWT).

20. The decompression apparatus in claim 18, wherein the inverse spatial linear transform includes adaptively computed inverse transform coefficients, and wherein the method further comprises receiving the adaptively computed inverse transform coefficients from a transmitting node transmitting the compressed radio signal.

21. The decompression apparatus in claim 18, wherein the inverse spatial linear transform includes inverse transform coefficients corresponding to an inverse Kahunen-Loeve transform (KLT).

22. The decompression apparatus in claim 18, wherein the filter includes a summer configured to sum the error signals with corresponding predicted antenna signals to generate the reconstructed representations of the multiple antenna signals, and wherein the filter is further configured to filter the reconstructed representations of the multiple antenna signals using a spatial temporal prediction matrix of predictive coefficients to generate the predicted antenna signals.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 illustrates a non-limiting example of a multi-antenna radio node communicating compressed multi-antenna signals with a receiver node;

(2) FIG. 2 illustrates a Coordinated multi-point (CoMP) communication system;

(3) FIG. 3 is a flowchart diagram of non-limiting example compression procedures;

(4) FIGS. 4A and 4B illustrate a non-limiting example diagram of multiple antenna signal compression apparatus;

(5) FIG. 5 is a flowchart diagram of non-limiting example decompression procedures; and

(6) FIG. 6 illustrates a non-limiting example diagram of multiple antenna signal decompression apparatus.

DETAILED DESCRIPTION

(7) In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular nodes, functional entities, techniques, protocols, standards, etc. in order to provide an understanding of the described technology. It will be apparent to one skilled in the art that other embodiments may be practiced apart from the specific details disclosed below. In other instances, detailed descriptions of well-known methods, devices, techniques, etc. are omitted so as not to obscure the description with unnecessary detail. Individual function blocks are shown in the figures. Those skilled in the art will appreciate that the functions of those blocks may be implemented using individual hardware circuits, using software programs and data in conjunction with a suitably programmed microprocessor or general purpose computer, using applications specific integrated circuitry (ASIC), and/or using one or more digital signal processors (DSPs). The software program instructions and data may be stored on computer-readable storage medium, and when the instructions are executed by a computer or other suitable processor control, the computer or processor performs the functions.

(8) Thus, for example, it will be appreciated by those skilled in the art that diagrams herein can represent conceptual views of illustrative circuitry or other functional units. Similarly, it will be appreciated that any flow charts, state transition diagrams, pseudocode, and the like represent various processes which may be substantially represented in computer readable medium and so executed by a computer or processor, whether or not such computer or processor is explicitly shown.

(9) The functions of the various illustrated elements may be provided through the use of hardware such as circuit hardware and/or hardware capable of executing software in the form of coded instructions stored on computer-readable medium. Thus, such functions and illustrated functional blocks are to be understood as being either hardware-implemented and/or computer-implemented, and thus machine-implemented.

(10) In terms of hardware implementation, the functional blocks may include or encompass, without limitation, digital signal processor (DSP) hardware, reduced instruction set processor, hardware (e.g., digital or analog) circuitry including but not limited to application specific integrated circuit(s) (ASIC) and/or field programmable gate array(s) (FPGA(s)), and (where appropriate) state machines capable of performing such functions.

(11) In terms of computer implementation, a computer is generally understood to comprise one or more processors or one or more controllers, and the terms computer, processor, and controller may be employed interchangeably. When provided by a computer, processor, or controller, the functions may be provided by a single dedicated computer or processor or controller, by a single shared computer or processor or controller, or by a plurality of individual computers or processors or controllers, some of which may be shared or distributed. Moreover, the term “processor” or “controller” also refers to other hardware capable of performing such functions and/or executing software, such as the example hardware recited above.

(12) The technology described in this application includes an effective, low-complexity way to represent a complex-valued radio signal either received from or to be transmitted to a multiple antenna radio node, e.g., a base station. A spatial-temporal (ST) predictor compresses the data associated with multiple antenna signals thereby reducing their dynamic range. The spatial-temporal (ST) predictor exploits the fact that radio signals received from multiple antennas are often highly correlated in both space (i.e., across antennas) and time and uses a substantially smaller number of bits to represent (quantize) a vector difference signal between the predicted and the original antenna signals while maintaining the same level of incurred quantization distortion. Upon receipt of the quantized difference signal (i.e., the compressed signal) sent over a backhaul channel by the multiple antenna radio node, a reproduction of the original multiple antenna signals may be constructed (e.g., at a receiver) by filtering the difference signal using a vector infinite impulse response (IIR) spatial-temporal filter. The filtering decompresses the received compressed signal. The coefficients associated with the spatial-temporal predictor can be predetermined or determined in real-time based on the spatial and temporal statistics of the multiple antenna radio signals. For the latter case, the predictive coefficients may be sent (preferably infrequently) over the backhaul channel along with the quantized radio signal in order to allow the multiple antenna radio signals to be reconstructed at the receiver. A low-complexity method of adaptively computing the spatial-temporal (ST) predictor based on certain correlation matrix functions of the multiple antenna radio signals is also described.

(13) FIG. 1 illustrates a non-limiting example of a multi-antenna radio node 10 communicating compressed multi-antenna signals with a receiver node 12. The multi-antenna radio node 10 includes two or more antennas for transmitting and/or receiving antenna signals. In particular embodiments, each antenna signal is or was transmitted with the same information though transmission over the air interface distorts that information in ways that are specific to each antenna's location. Collectively, the various antenna signals form a multi-antenna signal. The multi-antenna radio node 10 includes a compressor 15 for compressing this multi-antenna signal before transmitting the compressed multi-antenna signal by a transmitter 16 over a channel 13 to the receiver node 12. The compressor 15 performs operations described below that employ coefficients that in one example embodiment are sent over the channel 13 (shown as a dotted line) to the receiver node 12. Alternatively, those coefficients may be predetermined (fixed) or are determined in the receiver node 12 so as to avoid having to send them over the channel 13. The receiver node 12 includes a receiver 17 for receiving the compressed multi-antenna signal sent over the channel and in one embodiment any transmitted coefficients. A decompressor 18 decompresses these signals using inverse operations from those used in the compressor 15. The decompressed (expanded) multi-antenna signal is then further processed and/or output as indicated in block 19. Compressing the multi-antenna signal saves considerable bandwidth on the channel 13.

(14) One non-limiting example application of the radio node 10 and receiver node 12 is a coordinated multi-point (CoMP) communication system, an example of which is shown in FIG. 2. Mobile radios 20 communicate over an air interface with one or more of the multiple base stations 22. Each base station 22 includes multiple antennas 24, a compressor 15, and a transmitter 16 (as described above for FIG. 1) and communicates with a central processor 26 over a backhaul link 28. The central processor 26 can be a radio network node like a radio network controller (RNC), a core network node like an MSC or an SGSN, or an independent node. The central processor 26 includes a receiver 17, a decompressor 18, and further processing and/or output 19 as described above for FIG. 1.

(15) FIG. 3 is a flowchart diagram of non-limiting example procedures that may be used by the radio node(s) 10. The compressor 15 receives multiple antenna signals that were transmitted with the same information (step S1) and decorrelates those multiple antenna signals over space and time to remove redundancies and generate a compressed signal (step S2). In certain embodiments of radio node 10, the compressor 15 decorrelates the multiple antenna signals in time and space independently, with either occurring first. In alternative embodiments, the compressor 15 decorrelates the multiple antenna signals jointly in time and space. In general, the compressor 15 may decorrelate the multiple antenna signals with respect to time and space jointly, independently, and/or in any other appropriate manner. The transmitter 16 transmits the compressed signal to the receiving node 12 or 26 over a channel 13 or 28 (step S3).

(16) The operations of the compressor in accordance with one example detailed embodiment are now described. First the multiple antenna signals are models as follows: Let y[n]=[y.sub.1[n], y.sub.2[n], . . . , y.sub.n.sub.a[n]].sup.T denote an n.sub.a-dimensional time-domain complex-valued, sampled, multiple antenna signal vector to be communicated through a backhaul link connecting a central processor from or to a particular base station, where n.sub.a denote the number of antennas at the base station and nε{1, 2, . . . , N} denotes the sample time index. The temporal and spatial correlation (and thus redundancy) in the random process {y[n]} is represented using a vector auto-regressive (VAR) model given by:

(17) $\begin{array}{cc}y\left[n\right]=\underset{m=1}{\overset{M}{.Math.}}A\left[m\right]y\left[n-m\right]+e\left[n\right]& \left(1\right)\end{array}$
where M is the model order, {e[n]} is an innovation process which is modeled as a zero-mean, independent identically distributed (IID), vector Gaussian random process with R.sub.e[m]=Ee[n]e[n−m].sup.H=Λ.sub.eδ[m], δ[m] denotes the Kronecker-delta function, and e[n]≡[e.sub.1[n], e.sub.2[n], . . . , e.sub.n.sub.a[n]].sup.T. The VAR model can match any power spectrum of the multi-dimensional radio signal with sufficiently large order M, and it leads to simple (low-complexity) compression methods that incurs little latency, as described below. Non-limiting example values for M might be 2-8. But any suitable value for M may be used. Moreover, the VAR coefficients, as shown below, can also be computed efficiently based on measurements of the second-order statistics of the signal itself, enabling a low-complexity, adaptive implementation.

(18) Based on the VAR model of the multi-antenna radio signal y[n] in equation (1), one approach might be to simply filter {y[n]} with a vector FIR filter with a z-transform given by:

(19) $H\left(z\right)=\left(I-\underset{m=1}{\overset{M}{.Math.}}{A}_m{z}^{-m}\right)$
in order to obtain the innovation process (approximated by an error or difference) {e[n]}, which can then be quantized and sent over the backhaul link. However, since the receiver node does not have access to the original multiple antenna vector {y[n]}, as does the transmitting radio node, the encoding process is modified so as to integrate the FIR filtering with the quantization of the innovation.

(20) FIGS. 4A and 4B illustrate a non-limiting example diagram of multiple antenna signal compression apparatus that may be used to encodecompress {y[n]}. In general, the apparatus computes a predictive multiple antenna vector signal ŷ[n] based on a quantized version y.sub.q[n] of the original multiple antenna vector signal y[n], which is available at both the transmitting and the receiving ends, as

(21) $\hat{y}\left[n\right]=\underset{m=1}{\overset{M}{.Math.}}{A}_m{y}_q\left[n-m\right].$

(22) Since vector y[n] is often correlated in time, the error vector signal ê[n]≡y[n]−ŷ[n], which serves as an estimate of the true innovation e[n], should have much smaller dynamic range than y[n] and can thus be quantized with fewer number of bits to achieve the same level of quantization distortion. The quantized vector signal y.sub.q[n] is simply given by the sum of the predictive vector signal ŷ[n] and the quantized version ê.sub.q[n] of vector ê[n]. Since
y[n]−y.sub.q[n]=y[n]−ŷ[n]−e.sub.q[n]=e[n]−e.sub.q[n],
the fidelity of vector e.sub.q[n] in representing vector ê[n] translates directly into the fidelity of vector y.sub.q[n] in representing the received, multiple antenna signals vector y[n]. The innovator 30 in FIG. 4A includes n.sub.a combiners 31 for determining a difference ê[n] provided to a linear spatial transform 32 and an error covariance calculator 33.

(23) The predictive vector signal ŷ[n] is provided by block 42 shown in FIG. 4B, which in effect applies a vector infinite impulse response (IIR) filter 42 to the quantized error vector signal ê.sub.q[n]. More specifically, block 42 generates the predictive vector signal ŷ[n] for the next time instance by applying a vector finite-impulse-response (FIR) filter functioning as a spatial-temporal predictor 46 to the sum of the predictive vector signal ŷ[n] and the quantized error vector signal ê.sub.q[n] from the previous time instances generated by the adder 44. The quantized error vector signal ê.sub.q[n] is generated by applying an inverse spatial transform 40 shown in FIG. 4B to the output of the decoders 38. The decoders 38 map the bits produced by the analog-to-digital (A/D) encoders 36, e.g., through table lookups, to a reconstructed or quantized version of the transformed error signal, which is then transformed to the quantized error vector signal ê.sub.q[n] through the inverse spatial transform 40.

(24) To minimize the dynamic range of the error vector signal ê[n], the predictive matrix coefficients A≡[A.sub.1, A.sub.2, . . . , A.sub.M] generated by a predictor coefficient calculator 48 shown in FIG. 4B may be computed by minimizing the variance of ê[n]:

(25) $A=\underset{A=\left[A_1,\mathrm{.Math.},A_M\right]}{\arg \min }E{.Math.\hat{e}\left[n\right].Math.}^2=\underset{A=\left[A_1,\mathrm{.Math.},A_M\right]}{\arg \min }E{.Math.y\left[n\right]-\underset{m=1}{\overset{M}{.Math.}}{A}_m{y}_q\left[n\right].Math.}^2.\hspace{1em}$

(26) The orthogonality principle provides:

(27) $E\left(y\left[n\right]-\underset{m=1}{\overset{M}{.Math.}}{A}_m{y}_q\left[n-m\right]\right){y_q\left[n-k\right]}^H=0$
for all k=1, 2, . . . , M. In matrix form, this becomes:

(28) $\begin{array}{cc}A[\begin{array}{cccc}{R}_{y_q}\left[0\right]& R_{y_q}\left[-1\right]& \mathrm{.Math.}& R_{y_q}\left[M-1\right]\\ R_{y_q}\left[1\right]& R_{y_q}\left[0\right]& \ddots & \mathrm{.Math.}\\ \mathrm{.Math.}& \ddots & R_{y_q}\left[0\right]& R_{y_q}\left[-1\right]\\ R_{y_q}\left[M-1\right]& \mathrm{.Math.}& R_{y_q}\left[1\right]& R_{y_q}\left[0\right]\end{array}]=\left[\begin{array}{cccc}{R}_{{\mathrm{yy}}_q}\left[1\right]& R_{{\mathrm{yy}}_q}\left[2\right]& \mathrm{.Math.}& R_{{\mathrm{yy}}_q}\left[M\right]\end{array}\right]& \left(2\right)\end{array}$
where R.sub.yy.sub.q[m]≡Ey[n]y.sub.q[n−m].sup.H and R.sub.y.sub.q[m]≡Ey.sub.q[n]y.sub.q[n−m].sup.H are the multidimensional cross-correlation function of y[n] and y.sub.q[n] and auto-correlation function of y.sub.q[n], respectively. Equation (2) can be efficiently solved by a modified version of the Whittle-Wiggins-Robinson (WWR) algorithm which computes A in an order-recursive fashion, as summarized below. (The WWR algorithm solves equation (2) when its right-hand side is [R.sub.y.sub.q[1]R.sub.y.sub.q[2] . . . R.sub.y.sub.q[M]] instead.)

(29) Let A.sup.(m)≡[A.sub.1.sup.(m), A.sub.2.sup.(m) . . . , A.sub.m.sup.(m)] denote the solution of equation (2) when M=m. In other words, A=A.sup.(M). The following algorithm solves equation (2) by recursively computing A.sup.(m) until m reaches the desired order M. For notational simplicity, let R.sub.y.sub.q[1:m]≡[R.sub.y.sub.q[1],R.sub.y.sub.q[2], . . . , R.sub.y.sub.q[M]] and R.sub.yy.sub.q[1:m]≡[R.sub.yy.sub.q[1],R.sub.yy.sub.q[2], . . . , R.sub.yy.sub.q[M]]. Step 1: Initialization (set m=1) A.sub.1.sup.(1)=R.sub.yy.sub.a[1]R.sub.y.sub.a[0].sup.−1, Ā.sub.1.sup.(1)=R.sub.y.sub.a[1]R.sub.y.sub.a[0].sup.−1, B.sub.1.sup.(1)=R.sub.y.sub.a[1].sup.HR.sub.y.sub.a[0].sup.−1, Q.sub.1=R.sub.y.sub.q[0]−Ā.sub.1.sup.(1)R.sub.y.sub.q[1].sup.H and S.sub.1=R.sub.y.sub.q[0]−B.sub.1.sup.(1)R.sub.y.sub.q[1]. Step 2: Recursively compute the following quantities (until m reaches M)

(30) ${\stackrel{\_}{P}}_m=R_{y_q}\left[m+1\right]-\underset{i=1}{\overset{m}{.Math.}}{\stackrel{\_}{A}}_i^{\left(m\right)}{R}_{y_q}\left[m+1-i\right]$$\begin{array}{cc}{\stackrel{\_}{A}}_i^{\left(m+1\right)}={\stackrel{\_}{A}}_i^{\left(m\right)}-{\stackrel{\_}{P}}_m{S}_m^{-1}{\stackrel{\_}{B}}_{m-i+1}^{\left(m\right)}& \mathrm{for}i=1,2,\mathrm{.Math.},m\end{array}$${\stackrel{\_}{A}}_{m+1}^{\left(m+1\right)}=P_m{S}_m^{-1}$$\begin{array}{cc}{\stackrel{\_}{B}}_{m-i+1}^{\left(m+1\right)}={\stackrel{\_}{B}}_{m-i+1}^{\left(m\right)}-{\stackrel{\_}{P}}_m^H{Q}_m^{-1}{\stackrel{\_}{A}}_i^{\left(m\right)}& \mathrm{for}i=1,2,\mathrm{.Math.},m\end{array}$${\stackrel{\_}{B}}_{m+1}^{\left(m+1\right)}={\stackrel{\_}{P}}_m^H{Q}_m^{-1}$$Q_{m+1}=Q_m-{\stackrel{\_}{P}}_m{S}_m^{-1}{\stackrel{\_}{P}}_m^H\mathrm{and}{S}_{m+1}=S_m-{\stackrel{\_}{P}}_m{Q}_m^{-1}{\stackrel{\_}{P}}_m^H$$P_m=R_{{\mathrm{yy}}_q}\left[m+1\right]-\underset{i=1}{\overset{m}{.Math.}}{A}_i^{\left(m\right)}{R}_{{\mathrm{yy}}_q}\left[m+1-i\right]$$A_{m+1}^{\left(m+1\right)}={P_m\left(R_{y_q}\left[0\right]-\underset{i=1}{\overset{m}{.Math.}}{\stackrel{\_}{A}}_i^{\left(m\right)}{R_{y_q}\left[i\right]}^H\right)}^{-1}$$\begin{array}{cc}{A}_i^{\left(m+1\right)}=A_i^{\left(m\right)}-{\tilde{A}}_{m+1}^{\left(m+1\right)}{\stackrel{\_}{A}}_{m+1-i}^{\left(m\right)}& \mathrm{for}i=1,2,\mathrm{.Math.},m\end{array}$ Step 3: Finally, set A=A.sup.(M). (Note that {Ā.sub.i.sup.(m)}.sub.i=1.sup.m and {B.sub.i.sup.(m)}.sub.i=1.sup.m are auxiliary variables representing, respectively, the forward and backward matrix prediction coefficients satisfying similar (Yule-Walker) equations as (2) except that its right-hand side becomes [R.sub.y.sub.q[1], . . . , R.sub.y.sub.q[m]] instead of [R.sub.yy.sub.q[1], . . . , R.sub.yy.sub.q[m]].)

(31) R.sub.yy.sub.q[m] and R.sub.y.sub.q[m] may be approximated by empirical moving averages {circumflex over (R)}.sub.yy.sub.q[n, m, N.sub.w] and {circumflex over (R)}.sub.y.sub.q[n, m, N.sub.w], respectively, which are given by:

(32) $\begin{array}{c}{\hat{R}}_{{\mathrm{yy}}_q}\left[n,m,N_w\right]\equiv \frac{1}{N_w}\underset{k=n-N_w+1}{\overset{n}{.Math.}}y\left[k\right]y_q^H\left[k-m\right]\\ =\frac{1}{N_w}\left[\begin{array}{c}{N}_w{\hat{R}}_{{\mathrm{yy}}_q}\left[n-1,m,N_w\right]+y\left[n\right]y_q^H\left[n-m\right]\\ -y\left[n-N_w\right]y_q^H\left[n-N_w-m\right]\end{array}\right]\end{array}$$\begin{array}{c}{\hat{R}}_{y_q}\left[n,m,N_w\right]\equiv \frac{1}{N_w}\underset{k=n-N_w+1}{\overset{n}{.Math.}}{y}_q\left[k\right]y_q^H\left[k-m\right]\\ =\frac{1}{N_w}\left[\begin{array}{c}{N}_w{\hat{R}}_{y_q}\left[n-1,m,N_w\right]+y_q\left[n\right]y_q^H\left[n-m\right]-\\ y_q\left[n-N_w\right]y_q^H\left[n-N_w-m\right]\end{array}\right].\end{array}$
for a correlation lag m=0, 1, . . . , M−1, where n denotes the current time index, and N.sub.w denotes the window size. These moving averages can be updated immediately as the latest sample y[n] and y.sub.q[n] become available at the encoding end (the radio node 10). Alternatively, R.sub.yy.sub.q[m] and R.sub.y.sub.q[m] may be approximated by recursive empirical averages {circumflex over (R)}.sub.yy.sub.a[n, m; α] and {circumflex over (R)}.sub.y.sub.q[n, m; α], which are given by:
R.sub.yy.sub.q[n,m;α]≡(1−α){circumflex over (R)}.sub.yy.sub.q[n−1,m;α]+αy[n]y.sub.q.sup.H[n−m]
and
{circumflex over (R)}.sub.y.sub.q[n,m;α]≡(1−α){circumflex over (R)}.sub.y.sub.q[n−1,m;α]+αy[n]y.sub.q.sup.H[n−m],
where αε(0,1) denotes a certain predefined forgetting factor, and {circumflex over (R)}.sub.yy.sub.q[0, m; α] and {circumflex over (R)}.sub.y.sub.q[0,m; α] are initialized to the all-zero matrix for all m. The vectors R.sub.yy.sub.q[m] and R.sub.y.sub.q[m] are calculated in a correlation computer 50 using y[n] and y.sub.q[n], as shown in FIG. 4B, and are provided to the predictor coefficient calculator 48 which uses them to generate the coefficient matrix A.

(33) To reduce the frequency of sending overhead for the VAR coefficients A, the compressor may use these empirical averages to compute A only after each block of T samples. For example, all signal samples between time [kT,(k+1)T−1] will assume the same set of VAR coefficients A computed at time kT based on {circumflex over (R)}.sub.yy.sub.q[kT, m; α] and {circumflex over (R)}.sub.y.sub.q[kT, m; α], or alternatively {circumflex over (R)}.sub.yy.sub.q[kT, m; α] and {circumflex over (R)}.sub.y.sub.q[kT, m; α], for any period index k.

(34) Similar to the actual innovation e[n] at each time n, its estimate ê[n] is also spatially-correlated (across the multiple antennas), and therefore, direct independent quantization of each component ê.sub.i[n], for i=1, 2, . . . , n.sub.a, of ê[n]≡[ê.sub.1[n], ê.sub.2[n], . . . , ê.sub.n.sub.a[n]].sup.T, although possible, is not an efficient way of quantizing ê[n]. To exploit the spatial correlation, a linear spatial transformation is performed in block 32 on the error signal ê[n] using transform coefficients U from transform calculator 35 so that the transformed error vector signal w[n]≡[w.sub.1[n], w.sub.2[n], . . . , w.sub.n.sub.a[n]].sup.T has most of its energy (at a lower amplitude) concentrated in a smaller number K.sub.a of matrix elements representing the error, where K.sub.a≦n.sub.a, and thus the rest of its element can be discarded without affecting the fidelity of the reproduced signal.

(35) The linear transformation 32 may be fixed and pre-computed as, for example, the discrete-cosine transform (DCT), the Discrete Fourier Transform (DFT) or a discrete wavelet transform (DWT). In this case, there is no need to send the transform coefficients U along with the quantized prediction error e.sub.q[n] to the receiving node 12.

(36) Alternatively, the transformation 32 can be computed using adaptively computed matrix coefficients, e.g., using the Kahunen-Loeve Transform (KLT) for the prediction error process {ê[n]} through eigen-decomposition of its marginal covariance matrix Λ.sub.ê=Eê[n]ê.sup.H[n], which is given by Λ.sub.ê=UDU.sup.H, where U is a unitary matrix with columns being the eigenvectors of Λ.sub.ê, and D is a diagonal matrix with diagonal elements {λ.sub.e,i}.sub.i=1.sup.N.sup.a being the eigenvalues of Λ.sub.ê. The KLT transformation matrix is simply given by U. If adaptively computed matrix coefficients are used, then the transform calculator 35 may also compute the matrix coefficients of the inverse transform and send them to the receiving node. If a KLT transformation matrix U is used, then the inverse transform matrix coefficients are given by the Hermitian, or the conjugate transpose, denoted by U.sup.H, of the matrix U.

(37) The eigenvalues {λ.sub.e,i}.sub.i=1.sup.N.sup.a of Λ.sub.ê represent the corresponding variances of the transformed predicted errors, which can be used by the compressor to decide which errors, if any, should be discarded. Alternatively, it may be preferred to allocate different number of bits b.sub.i to each transformed error output w[n] according to its significance as indicated by its variance. If the variance of a component is too small relative to those of the other components, no bits may be allocated to quantize it, i.e., it is discarded.

(38) FIG. 4A shows quantizers 36 (labeled as A/D encoders) which uses the allocated number of bits b.sub.i to quantize their respective inputs w[n] based on information provided by bit allocator 37. The bit allocator 37 determines those allocations using the eigenvalues {λ.sub.e,i}.sub.i=1.sup.N.sup.a of the error covariance matrix Λ.sub.ê calculated by transform computer 35 from the error covariance matrix Λ.sub.ê. There are different methods to determine {b.sub.i}.sub.i=1.sup.N.sup.a depending on the type of quantizer used. For fixed-rate quantization, one can compute {b.sub.i}.sub.i=1.sup.N.sup.a using a high-resolution approximation as, e.g.,

(39) $b_i\cong \frac{b_{\mathrm{total}}}{n_a}+\frac{1}{2}\log \left({\lambda }_{\hat{e},i}/{\left(\underset{j=1}{\overset{n_a}{.Math.}}{\lambda }_{\hat{e},i}\right)}^{1/n_a}\right)$
where b.sub.total denotes the total number of bits available to quantize each sample of w[n]. Alternatively, one can also allocate equal number of bits to the first k components, where σ.sub.k.sup.2>β and k≦n.sub.a. In this case, β is the minimum energy that determines if an error in the vector w[n] from spatial transform 32 should be neglected. After calculating b[n], the compression apparatus transmits b[n] to receiving node 12 as a compressed multi-antenna signal. Although not shown in FIG. 4A, for the sake of simplicity, the compression apparatus may perform a parallel-to-serial conversion on b[n] or otherwise process b[n] in a suitable manner before transmitting the compressed multi-antenna signal to receiving node 12.

(40) Alternatively, one can apply the Breiman, Friedman, Olshen, and Store (BFOS) algorithm to optimally allocate the bits for a given set of component codebooks {C.sub.i}. See Riskin et al., “Optimal bit allocation via the generalized BFOS algorithm,” IEEE Trans. Info. Thy., vol. 37, pp. 400-402, March 1991, incorporated herein by reference. This quantizer for each coefficient described by Riskin et al. is a fixed-rate quantizer, i.e., it generates a fixed total number of bits b.sub.total at each time instance. But the quantizer for each coefficient can also be a variable-rate quantizer. In this example, it is preferred to use a quantizer with a uniform step or cell size in combination with an entropy encoder, such as a Huffman encoder, a Ziv-Lempel encoder, or an arithmetic encoder, which are well known to those skilled in the art, to generate a variable total number of bits at each time instance. See for example chapter 9 in Gersho and Gray, Vector Quantization and Signal Compression, Kluwer Academic Publishers, 1992. The fidelity of the reproduced signal is controlled by the choice of the step or cell size instead of the choice of the total number of bits b.sub.total.

(41) The eigenvalues {λ.sub.e,i}.sub.i=1.sup.N.sup.a may also be needed to scale each error in vector w[n] and to scale the reconstructed components at the decoding end if standard (Gaussian) quantization codebooks designed for probability distributions with unit variance are used.

(42) The marginal covariance matrix A.sub.ê can be approximated in the error covariance calculator 33 by an empirical moving average computed over a window of time samples as:

(43) 0$\begin{array}{c}{\Lambda }_{\hat{e}}\approx {\hat{\Lambda }}_{\hat{e}}\left[n,N_w\right]\equiv \frac{1}{N_w}\underset{i=n-N_w+1}{\overset{n}{.Math.}}\hat{e}\left[i\right]{\hat{e}}^H\left[i\right]\\ =\frac{1}{N_w}\underset{i=n-N_w+1}{\overset{n}{.Math.}}\left(y\left[i\right]-\hat{y}\left[i\right]\right){\left(y\left[i\right]-\hat{y}\left[i\right]\right)}^H\\ =\frac{1}{N_w}\left[\begin{array}{c}\left(N_w-1\right){\hat{\Lambda }}_{\hat{e}}\left[n-1;N_w-1\right]+\\ \left(y\left[n\right]-\hat{y}\left[n\right]\right){\left(y\left[n\right]-\hat{y}\left[n\right]\right)}^H-\\ \left(y\left[N_w\right]-\hat{y}\left[N_w\right]\right){\left(y\left[N_w\right]-\hat{y}\left[N_w\right]\right)}^H\end{array}\right]\end{array}\hspace{1em}$
where N.sub.w denotes the number of time samples within the window, or alternatively, by a recursive empirical average computed as:
Λ.sub.ê≈{circumflex over (Λ)}.sub.ê[n;α]={circumflex over (Λ)}.sub.ê[n−1;α]+(y[n]−ŷ[n])(y[n]−ŷ[n]).sup.H−(y[n−1]−ŷ[n−1])(y[n−1]−ŷ[n−1]).sup.H
where αε(0,1) denotes a certain predefined forgetting factor, and {circumflex over (Λ)}.sub.ê[0; α] is initialized to the all-zero matrix. To minimize the frequency of sending U, thereby saving bandwidth on the backhaul, {λ.sub.e,i}.sub.i=1.sup.K.sup.a, and {b.sub.i}.sub.i=1.sup.K.sup.a (where K.sub.a is the number of errors in w[n] with non-zero number bits allocated), the transform calculator 35 may use these empirical averages to compute (U,{λ.sub.e,1}.sub.i=1.sup.N.sup.a.,{b.sub.i}.sub.i=1.sup.N.sup.a) only after each block of T samples. For example, all signal samples between time [kT,(k+1)T−1] may assume the same spatial transform U and eigenvalues {λ.sub.ê,i}.sub.i=1.sup.N.sup.a computed at time kT based on {circumflex over (Λ)}.sub.ê[n, N.sub.w] or alternatively {circumflex over (Λ)}.sub.ê[n; α], for any period index k.

(44) The receiving node 12 performs a decompression method to recover representations of the multiple antenna signals. FIG. 5 is a flowchart diagram of non-limiting example decompression procedures. First, a compressed signal that corresponds to a multi-antenna signal is received (step S10). Next, the received signal is decompressed based on one or more correlations in space and in time to reconstruct a representation of the multi-antenna signal (step S11). The correlation in space and time may represent a correlation of multiple antenna signals in space and in time performed independently and in any order, or the correlation may represent a joint correlation of the multiple antenna signals in space and time. In general, receiving node 12 may utilize any appropriate form of correlation with respect to time and space, including any suitable joint or independent correlation of the two values. The reconstructed representation of the multi-antenna signal is then provided for further processing or output (step S12).

(45) FIG. 6 illustrates a non-limiting example diagram of multiple antenna signal decompression apparatus using to generate a reconstruction vector {y.sub.q[n]} of the multiple antenna signal vector y[n]. The receiver 18 receives the compressed multi-antenna signal b[n] at respective decoders 50 which generate corresponding analog signals based on respective bit allocations b.sub.1, b.sub.2, . . . , b.sub.n.sub.a that are either predetermined or adaptively selected. If they are adaptive selected, those bit allocations may be received from radio station over the channel. An inverse spatial transform 52 performs the inverse spatial transform using inverse coefficient matrix U.sup.−1 generated by transform calculator 35 and sent to the receiving node 12 over the channel or generated at the receiving node. The inverse spatial transform generates the quantized version ê.sub.q[n] of the error signal ê[n] which is input to a vector IIR filter 54 which combines it in respective combiners 58 with corresponding predictive vector signals ŷ[n] generated by a spatial-temporal predictor 56 using coefficient matrix A operating on reconstruction vector {y.sub.q[n]} of the multiple antenna signal vector y[n]. This is equivalent to filtering ê.sub.q [n] by vector IIR filter 54 with a matrix z-transform H(z) given by generated by the predictor coefficient calculator 48 in the radio node and sent to the receiving node 12 over the channel or generated at the receiving node 12.

(46) $H\left(z\right)={\left(I-\underset{m=1}{\overset{M}{.Math.}}{A}_m{z}^{-m}\right)}^{-1}.$
The output of the vector IIR filter 54 is the reconstruction vector {y.sub.q[n]} that represents the multi-antenna signal now decompressed.

(47) Since the VAR coefficients A computed by the predictor coefficient calculator 48 are minimum-phase (in the sense that the roots of the determinant of the matrix

(48) $\left(I-\underset{m=1}{\overset{M}{.Math.}}{A}_m{z}^{-m}\right)$
are all inside the unit circle), the IIR filter response is stable.

(49) In an example embodiment, the matrix predictive coefficients A are diagonal matrices, which means that in effect, the spatial temporal predictors 46 and 56 do not exploit the spatial correlation but only the temporal correlation of the received compressed multi-antenna signal. The spatial correlation is exploited only through transform coding on the prediction errors. This embodiment reduces the amount of overhead needed to describe the predictive coefficients A (which are scalars) at the expense of some performance degradation. These scalar predictive coefficients can also be further restricted to be identical across different antennas, in which case, the measurement of second-order statistics may be averaged across antennas as well. The modified WWRA algorithm reduces to the Levinson-Durbin algorithm in this case.

(50) While the model order M of the predictor is assumed to be fixed and predetermined, if desired, the adaptive selection of M may be integrated in the order-recursive computation of the predictive coefficients by incrementing the model order only when the resulting reduction in the prediction error variance is sufficiently substantial. In this case, the adaptively selected model order M may be sent to the receiving node.

(51) If the underlying frame structure and timing of the backhaul signaling is known, performance may be improved by using different (smaller) model orders at the start of each frame to avoid mixing potentially different statistics of adjacent frames.

(52) There are multiple advantages provided by this technology including, for example, providing an effective way to compress complex-valued radio signals either received from or to be transmitted to a remote base station with one or more antennas. Both spatial and temporal correlations in the multi-dimensional radio signals are exploited through joint spatial-temporal linear prediction to significantly reduce the amount of data that must be transmitted over the backhaul to communicate the ultimate information to be delivered. This means the capacity of the backhaul is significantly increased. Moreover, the technology is universal and has relatively low implementation complexity. There is no need to assume any particular time or frequency structure in the radio signal, and hence, is applicable for example to all 2G, 3G, and 4G standardized signals. The technology provides for continuous operation with little additional latency to the radio signal. Moreover, using linear prediction to compress analog signals in multiple dimensions (e.g., compressing a multi-antenna radio signal) provides an excellent tradeoff in performance and complexity. Accordingly, the technology may become important in backhaul-signal codecs in the future.

(53) Although various embodiments have been shown and described in detail, the claims are not limited to any particular embodiment or example. None of the above description should be read as implying that any particular element, step, range, or function is essential such that it must be included in the claims scope. The scope of patented subject matter is defined only by the claims. The extent of legal protection is defined by the words recited in the allowed claims and their equivalents. All structural and functional equivalents to the elements of the above-described preferred embodiment that are known to those of ordinary skill in the art are expressly incorporated herein by reference and are intended to be encompassed by the present claims. Moreover, it is not necessary for a device or method to address each and every problem sought to be solved by the technology described, for it to be encompassed by the present claims. No claim is intended to invoke paragraph 6 of 35 USC §112 unless the words “means for” or “step for” are used. Furthermore, no embodiment, feature, component, or step in this specification is intended to be dedicated to the public regardless of whether the embodiment, feature, component, or step is recited in the claims.