Correlation shaping matched filter receiver

Abstract

The present invention is directed toward an apparatus and method for receiving signals, demodulating the signals and shaping the correlation of the output of a correlation demodulator. The present invention can be used irrespective of whether the received signals have noise components that are Gaussian or non-Gaussian. Moreover, the present invention can be utilized when a predetermined set of received signals is linearly independent or linearly dependent.

Description

BACKGROUND OF THE INVENTION

[0002] Many of our present technologies use signal processing principles to achieve their communication functions. In a typical signal detection application, the receiver performs the function of receiving and processing signals. A design engineer using a receiver to perform signal detection designs the receiver so that it is capable of receiving and processing a predetermined set of signals. These signals vary depending upon the system in which the design engineer is deploying the receiver. Because signals received after transmission typically contain a noise component, the receiver must process the received signal and extract information regarding the transmitted signal. The receiver may contain a detector or similar device, which could be used to determine which signal within the predetermined set of signals was in fact received by the receiver.

[0003] Many prior art receivers perform signal detection by using a correlation demodulator and a detector. One example of a correlation demodulator is a matched filter demodulator. A single matched filter is basically a linear filter whose transfer function has been matched to a particular electronic signal or environment in order to perform a filtering that is optimum for some particular purpose. See generally U.S. Pat. No. 4,044,241 “Adaptive Matched Filter,” the contents of which are hereby incorporated by reference. The filter is particularly matched to a signal plus noise input from which the signal is desired to be extracted. Id. A matched filter demodulator is comprised of a bank of matched filters the output of which is sampled at one time instant to obtain a vector signal.

[0004] The matched filter demodulator can equivalently be implemented as a correlation demodulator with correlating signals equal to the time-reversed versions of the filters' transfer functions.

[0005] When one of a predetermined set of signals is received in additive white Gaussian noise, the matched filter detector is optimal for maximizing the probability of detection.

[0006] The matched filter detector is comprised of a correlation demodulator, where the correlating signals are equal to the predetermined set of signals that could have been received, followed by a detector. The detector declares as the detected signal the one for which the output of the correlator is a maximum. See generally J. G. Proakis, Digital Communications, McGraw-Hill, Inc. 3rd ed 1995, the contents of which are hereby incorporated by reference.

[0007] Although matched filter receivers are optimal when the added noise is white and Gaussian, they are not when the received signal contains added non-Gaussian noise, or other forms of distortion. Design engineers nonetheless continue to use matched filter receivers in operating environments containing non-Gaussian or non-additive noise, in part because optimal receivers for these noise environments are typically nonlinear and, therefore, difficult to implement. Furthermore, in many applications the noise distribution is unknown to the receiver. See generally T. Kailath and V. Poor, Detection of Stochastic Processes, IEEE Transactions on Information Theory, vol. 44, pp. 2230-59, October 1998, the contents of which are hereby incorporated by reference. In light of the fact that many receivers are required to perform in situations where the added noise is non-Gaussian or a combination of Gaussian and non-Gaussian, there is a need for a receiver which is simple to implement, does not rely on the channel parameters and can achieve an acceptable probability of detecting the correctly received signal irrespective of whether the noise added to the received signal is white, Gaussian, or non-Gaussian.

SUMMARY OF THE INVENTION

[0008] According to one aspect, the present invention provides a receiver comprising a correlation demodulator and a correlation shaper. In one embodiment, the correlation shaper is a decorrelator. In another embodiment, the decorrelator is comprised of a whitening transformation performed on a vector signal output from the correlation demodulator. In yet another embodiment the correlation shaper outputs a vector signal whose covariance matrix has the property that the second row and each subsequent row is a permutation of the first row. In another embodiment, a subspace whitening transformation is performed on a vector signal output from the correlation demodulator. In an additional embodiment, the correlation demodulator is comprised of a bank of correlators that cross-correlate the received signal with a set of orthogonal signals. In yet another embodiment, the correlation demodulator is comprised of a bank of correlators that cross-correlate the received signal with a set of projected orthogonal signals. In an additional embodiment, the correlation demodulator is comprised of a bank of correlators that cross-correlate the received signal with a set of geometrically uniform signals. In yet another embodiment, the correlation demodulator is comprised of a bank of correlators that cross-correlate the received signal with a set of projected geometrically uniform signals. In still another embodiment, the correlation shape of a signal that is an output of the correlation demodulator can be varied.

[0009] According to another aspect, the present invention provides a method for processing signals comprising the steps of receiving a member of a predetermined set of signals; demodulating the member of the predetermined set of signals; and, shaping the correlation of the demodulated signal.

BRIEF DESCRIPTION OF THE DRAWING

[0010] The invention is described with reference to the several figures of the drawing, in which,

[0011] FIG. 1 is a block diagram of an illustrative receiver, which may be used with the present invention.

[0012] FIG. 2 is a block diagram of an illustrative matched filter demodulator, which may be used with the present invention.

[0013] FIG. 3 illustrates various embodiments of the receiver of the present invention.

[0014] FIG. 4 illustrates various embodiments of the receiver of the present invention.

[0015] FIG. 5 is a plot of test results for one embodiment of the present invention.

[0016] FIG. 6 is a plot of test results for one embodiment of the present invention.

[0017] FIG. 7 is a plot of test results for one embodiment of the present invention.

[0018] FIG. 8 is a plot of test results for one embodiment of the present invention.

[0019] FIG. 9 is a plot of test results for one embodiment of the present invention.

[0020] FIG. 10 is a plot of test results for one embodiment of the present invention.

[0021] FIG. 11 is a plot of test results for one embodiment of the present invention.

DETAILED DESCRIPTION

[0022] The present invention is directed toward a novel way of receiving and processing a distorted version of a signal, sk(t), which is one of a predetermined set of signals, {s1(t), s2(t) . . . sm(t)} . The one of the predetermined set of signals, sk(t), lies in a real Hilbert space H with inner product <x(t), y(t))> 1$⟨x\left(t\right),y\left(t\right)⟩={\int }_{t=-\infty }^{\infty }x\left(t\right)y\left(t\right)dt.$

[0023] The present invention is particularly useful in signal detection applications where the noise environment is not Gaussian.

[0024] A receiver is configured to receive a predetermined set of signals. This predetermined set of signals is comprised of a finite number of signals, and is hereinafter represented as {sk(t), 1≦k≦m}. The actual signal received by the receiver, r(t), is comprised of one member of the predetermined set of received signals, sk(t), and a noise component, n(t). In order to perform its signal detection function, a receiver must process the received signal prior to detecting which of the predetermined set of received signals it actually received.

[0025] The receiver and method of signal detection of the present invention allows an engineer to design a specific correlation shape for the output of a correlation demodulator. In this way, the present invention overcomes the problems encountered with prior art matched filter processing methods, for example, by overcoming the fact that the output of a matched filter demodulator contains elements that are correlated among themselves.

[0026] In some of the embodiments provided herein, the design engineer could choose to have the output of the correlation demodulator be completely decorrelated. In yet other embodiments the design engineer could choose to shape the correlation of the output of the correlation demodulator such that the covariance matrix of this output has the property that the second row and each subsequent row is a permutation of the first row. In alternative embodiments, an engineer practicing the present invention can choose any correlation shape for the output of the correlation demodulator. By shaping the correlation of the output of the correlation demodulator, the design engineer can design a simple receiver tailored to receive the predetermined set of signals.

[0027] The receiver and method of signal processing of the present invention are highly versatile and can be utilized irrespective of whether the predetermined set of received signals is linearly dependent or linearly independent. In addition, the received signal can be pre-whitened to facilitate signal processing when the received signal has been corrupted by colored noise. Moreover, the invention disclosed herein can be implemented for both continuous-time signals and discrete signals. Embodiments disclosed herein may also be implemented in hardware, for example in DSP chips, or in software using C++ programming language, for example. In addition to disclosing a receiver for processing signals, one aspect of the present invention also provides a novel method wherein a design engineer can alter the correlation shape of the output of a correlation demodulator.

[0028] A basic overview of the present invention is illustrated in FIG. 1. With reference now to FIG. 1, a signal r(t) 20 is first received by a receiver. The signal is then processed by a correlation demodulator 30 which may provide some correlation shaping of the vector signal output. The vector signal output of the correlation demodulator may then be additionally shaped by a correlation shaper 40. The output of the correlation shaper could be passed to a detector or similar device. The detector or similar device could process the output and determine which member of a predetermined set of signals was received by the receiver. The detector or similar device could output signal, sk(t) 60 in FIG. 1.

[0029] The embodiments are generally directed toward different implementations of the present invention. In some embodiments, a matched filter demodulator may be used. In these embodiments, the correlation of the output of the matched filter demodulator is shaped using a linear transformation. In other embodiments, the shaping of the correlation of the demodulator output is done within the demodulator by choosing a different set of signals than those used to carry out the cross-correlation in the matched filter demodulator. These embodiments of the present invention are physically distinct, but mathematically equivalent. This equivalence is shown in the following section. In the notation that follows, the symbol W is used with reference to a whitening transformation. One skilled in the art will appreciate that these equations are generally applicable to the case where the transformation is not a whitening transformation.

[0030] Mathematical Equivalence of Embodiments

[0031] A. Problem Formulation

[0032] Suppose we have a transmitter that transmits one of M signals {sk(t), 1≦m≦M} with equal probability, where the signals lie in a real Hilbert space H with inner product <x(t),y(t))>=∫x(t)y(t)dt. We assume only for simplicity of exposition here that the signals are linearly independent and normalized so that ∫sm2(t)dt=1 for all m. For a discussion of the more general case of linearly dependent signals see Y. C. Eldar, A. V. Oppenheim, and D. Egnor, “Orthogonal and Projected Orthogonal Matched Filter Detection,” submitted to IEEE Trans. on Signal Proc. January 2001. The received signal r(t) is modeled as r(t)=sk(t)+n(t), where n(t) is a stationary white noise process with zero mean and spectral density2.

[0033] We demodulate the signal r(t) using a correlation demodulator as depicted in FIG. 4. The received signal r(t) is cross-correlated with M normalized signals qm(t)εH so that am=(qm(t), r(t)), where the signals qm(t) are to be determined. When the signals qm(t) are orthogonal the receiver is referred to as an orthogonal matched filter receiver. The detected signal is sk(t) where k=argmax am. The difference between the orthogonal matched filter (“OMF”) receiver and the matched filter (“MF”) receiver lies in the choice of the signals qm(t).

[0034] If the transmitted signal is sk(t), then

a m =<q m (t),r(t))=<qm(t),sk(t))+<qm(t),n(t))  (1)

[0035] The detected signal will be the transmitted signal sk(t) if maxm |qm(t), sk(t)+n(t))=((t), sk(t)+n(t)). Therefore, we would like to choose the signals qm(t) to maximize (qm(t), sm(t)) for 1≦m≦M. It is well known that the signals qm(t)=sm(t) maximize this inner product. The resulting demodulator is then equivalent to the well known MF demodulator. We note that qm(t)=sm(t) also maximizes the sum 2$R_{\mathrm{hs}}=\sum _{m=1}^M⟨q_m\left(t\right),s_m\left(t\right)⟩,$

[0036] since the individual terms are maximized by this choice. We will m=l see shortly that when additional constraints are imposed, it will be useful to consider maximizing the sum rather than the individual terms.

[0037] In general, the outputs am of the demodulator are correlated since they share information regarding the noise process n(t). We would like to choose the signals qm(t) so that the outputs am are uncorrelated. When the noise is non-Gaussian this approach does in fact lead to improved performance over conventional MF detection in many cases.

[0038] Let cov (am, ak) denote the cross-covariance of am and ak. Then,

cov (am, ak)=E(<qm(t),n(t))(n(t),qk(t)>)=σ2<q(t),qk(t)>.  (2)

[0039] From (2) it follows that the outputs of the demodulator are uncorrelated if and only if the signals qm(t) are orthonormal, i.e., if and only if <qm(t), qk(t))=δmk for all m, k. We therefore propose to choose the signals qm(t) to be orthonormal.

[0040] As before, we would also like to choose the signals qm(t) to maximize <qm(t), sm(t)) for 1≦m≦M. However, we now have an additional constraint, namely that the signals qm(t) are orthonormal. If the signals sm(t) are not orthonormal, then we cannot maximize the inner products individually subject to this constraint. Instead, we consider maximizing the sum of the inner products. Thus, we seek a set of signals {qm(t), 1≦m≦M} such that 3$\begin{array}{cc}{R}_{\mathrm{hs}}=\sum _{m=1}^M⟨q_m\left(t\right),s_m\left(t\right)⟩& \left(3\right)\end{array}$

[0041] is maximized, subject to the constraint

<qm(t)qk(t)>=δmk, 1≦m≦M.  (4)

[0042] The design problems of (3) and (4) can be formulated in two equivalent ways to provide further insight into the problem. Specifically, it will be shown that the following problems are the same:

[0043] 1. Find a set of orthonormal signals {qm(t}, 1≦m ≦M} that maximize Rhs=Σm<sm(t),qm(t));

[0044] 2. Find an optimal whitening transformation W that minimizes the total mean squared error (MSE) between the whitened output b=Wã and the input ã, where ã denotes the vector output of the conventional MF demodulator. Then choose the signals {qm(t), 1 m M} to be the orthonormal signals given by qm(t)=ΣkW*mksk(t); and,

[0045] 3. Find a set of orthonormal signals {qm(t), 1 m M} that are closest in a least-squares sense to the signals {sm(t), 1 m M}, namely that minimize ε=Σm<sm(t)−qm(t), sm(t)−qm(t)).

[0046] Turning now to showing the equivalence between the three problems above, let M S:CM→H denote the linear transformation defined by 4$\mathrm{Sx}=\sum _{m=1}^M{x}_m{s}_m\left(t\right),$

[0047] where xεCM is an arbitrary M-dimensional vector and xm denotes the m-th component of x. Let S*:H<CM denote the adjoint transformation so that if x=S*y(t) for arbitrary y(t)εH, then xm=<sm(t)y(t)). Let ã denote the vector with the m-th component ãm, where ãm is the output of the correlation demodulator when qm(t)=sm(t). From (2) it follows that the covariance matrix of ã denoted Ca, is given by

C a =σ 2 S*S,   (5)

[0048] where the mk-th element of S*S is <sm(t), sk(t)>. If the signals sm(t) are not orthonormal, then Ca is not diagonal and the elements of ã are correlated. Suppose we whiten ã using a whitening transformation W to obtain the random vector b=Wã, where the covariance matrix of b is given by Cb=σ2I, and then base our detection on b. Thus, the components bm are the inputs to the detector, and the detected signal is sk(t) if k=argmaxbm. Since the detector bases its decision on the vector b, we choose a whitening transformation W that minimizes the MSE given by 5$\begin{array}{cc}{\epsilon }_{\mathrm{mse}}=\sum _{m=1}^{M}E\left({\left(b_m^{\prime }-{\tilde{a}}_m^{\prime }\right)}^2\right),\mathrm{where}{\tilde{a}}_m^{\prime }={\tilde{a}}_m-E\left(\widetilde{a_m}\right)\mathrm{and}{b}_m^{\prime }=b_m-E\left(b_m\right).& \left(6\right)\end{array}$

[0049] That is, from all possible whitening transformations, we seek the one that results in a white vector b as close as possible to the original vector ã.

[0050] We now show that the demodulator depicted in FIG. 3 is equivalent to the correlation demodulator of FIG. 4 where the signals qm(t) are orthonormal and given by qm(t)Σk=k Wmk*Sk(t) where Wmk denotes the mk th element of W. In other words, the outputs of FIG. 3, 4 are equal, provided that qm(t)=ΣkWmk*sk(t)

[0051] The vector output b of FIG. 3 is given by

b=Wã=WS*r (t)=Q*r(t)  (7)

[0052] where Q: CM→H is given by Q=SW*. Therefore, b can be viewed as the output of a correlation demodulator with signals qm(t)=ΣkWmk*sk(t).

[0053] We now need to show that the signals qm(t) are orthonormal. It is sufficient to show that Q*Q=WS*SW*=I. By definition, Cb=σ2I. In addition, Cb=WCaW* and from (5) Ca=σ2S*S. Therefore, 6$Q^{*}Q={\mathrm{WS}}^{*}{\mathrm{SW}}^{*}=\frac{1}{{\sigma }^2}{C}_b=I.$

[0054] In summary, the output of FIG. 3 may be obtained using the correlation demodulator of FIG. 4, where the signals qm(t) are orthonormal and given by qm(t)=ΣkWmk*sk(t)

[0055] We now show the minimization of εmse given by (6), is equivalent to maximization of Rhs given by (3). Using (7) we have

b−ã =(Q*−S*)r(t)=(Q*−S*)(sk(t)+n(t))  (8) and 7 $$ b m ′ - a ~ m ′ = ⟨ q m ⁡ ( t ) - s m ⁡ ( t ) , n ⁡ ( t ) ⟩ . ( 9 )

[0056] Substituting (9) into (6) we have 8$\begin{array}{cc}{\epsilon }_{\mathrm{mse}}=\sum _{m=1}^{M}E\left({⟨q_m\left(t\right)-s_m\left(t\right),n\left(t\right)⟩}^2\right).\mathrm{Let}{e}_m=E\left({⟨s_m\left(t\right)-q_m\left(t\right),n\left(t\right)⟩}^2\right).\mathrm{Then},e_m={E\left(\underset{t}{\int }\left(s_m\left(t\right)-q_m\left(t\right)\right)n\left(t\right)dt\right)}^2& \left(10\right)\end{array}$9$\begin{array}{cc}={\int }_{t,t^1}\left(s_m\left(t\right)-q_m\left(t\right)\right)\left(s_m\left(t^{\prime }\right)-q_m\left(t^{\prime }\right)\right)E\left(n\left(t\right)n\left(t^{\prime }\right)\right)dtd{t}^{\prime }={\sigma }^2\int {\left(s_m\left(t\right)-q_m\left(t\right)\right)}^{2}dt={\sigma }^2⟨s_m\left(t\right)-q_m\left(t\right),s_m\left(t\right)-q_m\left(t\right)⟩& \left(11\right)\end{array}$

[0057] Combining (10) and (11) we see that minimization of εmse is equivalent to minimization of εis, where 10$\begin{array}{cc}{\epsilon }_{\mathrm{ls}}=\sum _{m=1}^M⟨s_m\left(t\right)-q_m\left(t\right),s_m\left(t\right)-q_m\left(t\right)⟩.& \left(12\right)\end{array}$

[0058] Therefore, the optimal whitening problem is equivalent to the problem of finding a set of orthonormal signals {qm(t), 1≦m≦M} that are closest in the least-squares sense to the signals {sm(t), 1≦m≦M}.

[0059] Finally, we show that this least-squares problem is equivalent to our original design problem of (3) and (4). Expanding εis, we have 11$\begin{array}{cc}\begin{array}{c}{\epsilon }_{\mathrm{ls}}=\\ =\sum _{m=1}^M(⟨s_m\left(t\right),s_m\left(t\right)⟩+⟨q_m\left(t\right),q_m\left(t\right)⟩-2⟨s_m\left(t\right),q_m\left(t\right)⟩)\\ =\sum _{m=1}^M\left(2-2⟨s_m\left(t\right),q_m\left(t\right)⟩\right)\end{array}& \left(13\right)\end{array}$

[0060] From (3) and (13) it follows that minimization of εls is equivalent to maximization of Rhs. Since minimization of εmse is equivalent to minimization of εls, we conclude that these three problems are equivalent.

[0061] Note, that if the transmitted signals sm(t) are orthonormal, then the output of the MF demodulator ã is white. Thus, in this case W=I and the OMF detector is equivalent to the MF detector. Alternatively, if the signals sm(t) are orthonormal, then the residual least-squares error εls is minimized when qm(t)=sm(t), and again the OMF detector reduces to the MF detector.

[0062] B. Optimal Whitening

[0063] Since the optimal whitening problem is equivalent to the problem of (3)-(4), we can choose to determine the signals {qm(t), 1≦m≦M} by solving this problem.

[0064] We first restate the optimal whitening problem in its most general form. Let aεRM be a random vector with m-th component am and positive-definite covariance matrix Ca, and let a′m=am−E(am). We seek a whitening transformation W such that the white vector b =Wa has a covariance matrix Cb=σ2I, and is as close as possible to a in the MSE sense. Thus, we seek the transformation W that minimizes 12$\begin{array}{cc}{\epsilon }_{\mathrm{mse}}=\sum _{m=1}^{M}E\left({\left(a_m^{\prime }-b_m^{\prime }\right)}^2\right)& \left(14\right)\end{array}$

[0065] where bm is the mth component of b, and b′m=bm−E(bm), subject to the constraint

C b =WC a W*=σ 2 I   (15)

[0066] where Cb is the covariance matrix of b. Since W must be invertible (15) reduces to

σ2(W*W)−1=Ca  (16).

[0067] We solve this minimization problem using the eigendecomposition of Ca and the singular value decomposition (SVD) of W.

[0068] Let the vectors νk denote the orthonormal eigenvectors of Ca, so that

C a νk=λkνk,1≦k≦M  (17)

[0069] where λK>0. We can then decompose Ca as Ca=VDV* where V denotes the unitary matrix of columns νk and D denotes the diagonal matrix with diagonal elements λK. Then

W*Wν k =σ2Ca−1νk=σ2λk−1νk  (18).

[0070] From the properties of the SVD, it then follows that

W νk=σkuk,1≦k≦M  (19)

[0071] where 13${\sigma }_k=\sigma /\sqrt{{\lambda }_k}$

[0072] and the vectors uk are orthonormal.

[0073] Since the M vectors νkεCM are orthonormal, they span the space CM and any xεCM may be expressed as x=Σk<νk, x>νk where the inner product on CM is defined as <νk, x>=νk*x. Let a′=a−E(a) and b′=b−E(b). Then a′=Σk<νk,a′>νk and

b′−a′=Wa′−a′= Σk, a′)(σkuk−νk),  (20)

[0074] where we used (19). We can now express εmse of (14) as 14$\begin{array}{cc}{\epsilon }_{\mathrm{mse}}=E\left({⟨b^{\prime }-a^{\prime },b^{\prime }-a^{\prime }⟩}^2\right)=\sum _{k,m}E\left(⟨v_m,a^{\prime }⟩⟨a^{\prime },v_k⟩\right)⟨{\sigma }_k{u}_k-v_k,{\sigma }_m{u}_m-v_m⟩& \left(21\right)\end{array}$

[0075] Now,

E (<νm, a′> <a′,νk)=<νm,Caνk>=λk<νm, νk>=λkδmk  (22)

[0076] Substituting (22) in (21) results in 15$\begin{array}{cc}{\epsilon }_{\mathrm{mse}}=\sum _k{\lambda }_k⟨{\sigma }_k{u}_k-v_k,{\sigma }_k{u}_k-v_k⟩=\sum _k{\lambda }_k\left({\sigma }_k^2+1-2{\sigma }_k\Re \left(⟨u_k,v_k⟩\right)\right),& \left(23\right)\end{array}$

[0077] where (•) denotes the real part. From (23) it follows that minimizing εmse is equivalent to maximizing A=Σkσk(<uk, νk>). Using the Cauchy-Schwartz inequality

[0078] we have,

A =Σkσk(<uk,νk>)  (24) 16 $$ ≤ ∑ k ⁢ σ k ⁢ &LeftBracketingBar; ⟨ u k , v k ⟩ &RightBracketingBar; ( 25 ) ≤ ∑ k ⁢ σ k ⁡ ( ⟨ u k , u k ⟩ ⁢ ⟨ v k , v k ⟩ ) 1 / 2 = ∑ k ⁢ σ k ( 26 )

[0079] with equality in (25) if and only if <uk, νk> is real and nonnegative, and equality in (26) if and only if uk=ckνk for some nonzero constants ck. Since the vectors uk are orthonormal <uk,uk=1. We therefore conclude that A≦Σkσk with equality if and only if uk=νk. Thus, εmse is minimized when W is given by 17$\begin{array}{cc}{\mathrm{Wv}}_k={\sigma }_k{v}_k=\frac{\sigma }{\sqrt{{\lambda }_k}}{v}_k& \left(27\right)\end{array}$

[0080] or

W=σVD −½ =V*=σC a −½   (28)

[0081] In summary, the optimal whitening transformation that minimizes the MSE εmse defined in (14) for an input a with covariance Ca and an output b with covariance Cb=σ2I, is W=σCa{fraction (1/2)}.

[0082] In FIG. 3 the input to the whitening transformation is a=ã with Ca=σ2S*S. Thus, the optimal whitening transformation in this case is W=(S*S)−½, and the optimal orthonormal signals qm(t) that maximize Rhs are given by qm(t)=ΣkW*kmsk(t) or qm(t) =Qim where Q=S(S*S)−½ and im(k)=δmk. If the signals sk(t) are linearly dependent, then W=((S*S)½)†where ( )† denotes the Moore-Penrose pseudo-inverse.

[0083] C. Covariance Matrix of the Output is Arbitrary:

[0084] The correlation shaper in FIG. 3 can be chosen so that the covariance matrix is arbitrary within the mathematical constraints imposed on any covariance matrix. In this case the correlation shaper can be chosen so that WCaW*=Cb.

[0085] D. Covariance Matrix of the Output has the Permutation Property:

[0086] Suppose we choose the correlation shaper in FIG. 3 so that the covariance matrix of the output has the property that the second row and each subsequent row is a permutation of the first row.

[0087] Let dk be the elements of the first row of the specified covariance matrix. The correlation shaper that minimizes the MSE between the input and output is given as follows.

[0088] Let D be a diagonal matrix whose diagonal elements are the square-roots of the generalized Fourier transform of the sequence dk; the generalized Fourier transform is defined on a group formed by the elements of the pre-specified covariance matrix. See generally, Y. C. Eldar, G. D. Forney, Jr., “On quantum detection and the square-root measurement”, IEEE Trans. on Inform. Theory, vol. 47, No. 3, March 2001, the contents of which are hereby incorporated by reference.

[0089] Let F be a Fourier matrix representing the generalized Fourier transform over the group formed by the elements of the covariance matrix.

[0090] For linearly independent signals:

W=SFD (DF*S*SFD)−½DF*

[0091] For linearly dependent signals:

W=SFD ((DF*S*SFD)−½)†DF*

[0092] 1. Orthogonal/Projected Orthogonal & Geometrically Uniform & Projected Geometrically Uniform Correlating Signals

[0093] In all cases the closest signals (in a least-squares sense) to the predetermined set of received signals are given by the equation 18$q_k\left(t\right)=\sum _{l=1}^m{S}_l\left(t\right){\left(W\right)}_{\mathrm{lk}}$

[0094] where (W)lk is the lk-th element of W. As was previously the case, one skilled in the art will recognize that the these equations are equally valid when the transformation is not a whitening transformation.

[0095] As previously mentioned, in some of the embodiments, the correlation demodulator 30 is a matched filter demodulator. Reference will now be made to FIG. 2. The matched filter demodulator 130 receives an incoming signal, r(t) 120, and processes the received signal 120 through a plurality of multipliers 132 and integrators 134. The plurality of multipliers 132 and integrators cross-correlate the received signal 120 with the predetermined set of signals {s1(t),s2 (t) . . . sm(t)}. After this cross-correlation, the matched filter demodulator 130 outputs a correlated signal {ã1, ã2 . . . ãm} 136.

[0096] An illustration of a preferred embodiment of the present invention is provided in FIG. 3. FIG. 3 illustrates a received signal r(t) 120, the matched filter demodulator 130 of FIG. 2, demodulator output signal {ã1, ã2 . . . ãm} 136, a correlation shaper 140, and an output of the correlation shaper {b1, b2 . . . bm} 142. Four embodiments will be discussed with reference to FIG. 3. The embodiments vary depending on the correlation shape chosen by the design engineer and upon whether the predetermined set of signals {s1(t), s2(t) . . . sm(t)} is linearly independent or linearly dependent.

[0097] Embodiment No. 1—Linearly Independent Received Signals & Decorrelated Output

[0098] In the first of these four embodiments, it is assumed that the correlation shape chosen is to have the output {b1, b2 . . . bm} be completely decorrelated. In addition, in this embodiment, the predetermined set of signals {s1(t), s2(t) . . . sm(t)} is linearly independent. In this embodiment, the correlation shaper 140 performs a whitening transformation on the matched filter demodulator output signal 136. After the whitening transformation, the output of the correlation shaper 140, which was correlated when it emerged from the matched filter demodulator 130, becomes uncorrelated. In this embodiment, it may be desirable to minimize the mean squared error between the output of the correlation shaper 142 and the output of the matched filter demodulator 136. This embodiment may perform satisfactorily for a given system even if the covariance shaper does not result in the smallest mean squared error value between the output of the correlation shaper 142 and the matched filter demodulator 136.

[0099] Embodiment No. 2—Linearly Independent Received Signals & Output with Specific Correlation Shape

[0100] An alternative embodiment of the present invention, also shown in FIG. 3, allows a design engineer to specify a correlation shape for the output of the correlation shaper 140, depicted as {b1, b2 . . . bm} in FIG. 3. The correlation shape of {b1, b2 . . . bm} is given by the covariance matrix of b, 142, where b is a vector with components bk. An engineer can alter the correlation shape of b 142 by choosing the covariance matrix to have specific properties. In addition, one skilled in the art may decide in certain circumstances to allow the correlation shape of the output of the correlation shaper to be arbitrary. In this instance, the covariance matrix can be comprised of arbitrary values that satisfy the constraints imposed on any covariance matrix.

[0101] This embodiment differs from the one previously described, in which the correlation shaper decorrelated its output by performing a whitening transformation on the input of the correlation shaper. Rather, in this embodiment, the transformation is chosen so that the output of the correlation shaper has an arbitrary correlation shape.

[0102] In this embodiment, it may additionally be desirable to minimize the mean squared error between the output of the correlation shaper 142 and the output of the matched filter demodulator 136. We assume that the second and each subsequent row is chosen to be a permutation of the first. An engineer may alter the correlation shape of the output by altering the permutations of the rows of the covariance matrix.

[0103] This embodiment may perform satisfactorily for a given system even if the correlation shaper does not result in the smallest mean squared error value between the output of the correlation shaper 142 and the matched filter demodulator 136.

[0104] Embodiment No. 3—Linearly Dependent Received Signals & Decorrelated Output

[0105] In an alternative embodiment of the inventive receiver, the output of the correlation shaper 140 can be decorrelated when the predetermined set of signals {s1(t), s2(t) . . . sm(t)} is linearly dependent. When the predetermined set of signals is linearly dependent, the outputs of the matched filter demodulator are deterministically linearly dependent, meaning that vector ã and the elements of b=Wã are also linearly dependent. The linear dependence of the predetermined set of signals renders conventional whitening techniques impossible. Thus, in this alternative embodiment, the outputs of the matched filter demodulator will be whitened on the subspace upon which they lie.

[0106] Making reference now to FIG. 3, the outputs of the matched filter demodulator 136 lie in a subspace V of Rm. In this embodiment, it is desirable to whiten the outputs of the matched filter demodulator {ã1, ã2, . . . ãm} 136 on the subspace V, hereinafter referred to as subspace whitening. In this embodiment, subspace whitening is defined so that the whitened signal lies in V and its representation in terms of any orthonormal basis for V is white.

[0107] In this embodiment, it may be desirable to minimize the mean squared error between the output of the correlation shaper 142 and the output of the matched filter demodulator 136. This embodiment may perform satisfactorily for a given system even if the correlation shaper does not result in the smallest mean squared error value between the output of the correlation shaper 142 and the matched filter demodulator 136.

[0108] Embodiment No. 4—Linearly Dependent Received Signals & Output Has Specific Correlation Shape An additional embodiment of the present invention can be used when the predetermined set of signals {s1(t), s2(t) . . . sm(t)} is linearly dependent. As opposed to Embodiment No. 3 discussed above, an engineer utilizing this embodiment would choose a correlation shape for the output of the correlation shaper 140. In this embodiment, the correlation shaping performed by the correlation shaper 140 is similar to that described in Embodiment No. 2, except that correlation shaping in this embodiment is performed on a subspace spanned by the predetermined set of signals {s1(t),s2(t) . . . sm(t)} . In this embodiment, it may desirable to minimize the mean squared error between the output of the correlation shaper 142 and the output of the matched filter demodulator 136. This embodiment may perform satisfactorily for a given system even if the correlation shaper does not result in the smallest mean squared error value between the output of the correlation shaper 142 and the matched filter demodulator 136.

[0109] Embodiment No. 5-8—Choosing a Set of Orthogonal Signals

[0110] An alternative embodiment of the present invention is depicted in FIG. 4. The receiver 210 comprises a received signal r(t) 220, a correlation demodulator 230, a set of output signals {a1, a2 . . . am} from the correlation demodulator 240, and one of a predetermined set of signals received by the receiver, sk(t) 260. In this embodiment, the correlation shaping is achieved by selecting the signals {q1(t), q2(t) . . . qm(t)} 232. The selection of this set of signals varies depending upon the desired correlation shape of the output vector {a1, a2 . . . a} 240 and upon whether the predetermined set of signals {s1(t), s2(t) . . . sm(t)} is linearly independent or linearly dependent. Alternative embodiments for each of these variations are discussed below.

[0111] Embodiment No. 5—Linearly Independent Received Signals & Orthogonal Signals Making reference now to FIG. 4, the first of these embodiments will be discussed. The output signal {a1, a2. . . am} 240 is decorrelated and the predetermined set of signals {s1(t), s2(t) . . . sm(t)} is linearly independent. If the desired correlation shape is to have the output of the correlation demodulator 230 be decorrelated, the set of signals {q1(t), q2(t). . . qm(t)} 233 should be orthogonal. In this embodiment, it may be desirable to minimize the least squares error between the orthogonal signals {q1(t), q2(t) . . . qm(t)} 233 and the predetermined set of signals {s1(t), s2(t) . . . sm(t)}. While minimizing this least-squares value is preferred in this embodiment, those skilled in the art will recognize that adequate performance may be achieved although the least-squares error may not be at its minimum.

[0112] Embodiment No. 6—Linearly Independent Received Signals & Geometrically Uniform Signals

[0113] If an engineer desired a particular correlation shape for {a1, a2 . . . am} 240, the engineer could achieve this result by altering the choice of the signal set {q1(t), q2(t) . . . qm(t)} 233. If the covariance of the output of the correlation shaper has the property that the second and each subsequent row is chosen to be a permutation of the first, this can be achieved by choosing {q1(t), q2(t) . . . qm(t)} to be geometrically uniform. See generally G. D. Forney, Jr., “Geometrically Uniform Codes,” IEEE Trans. Inform. Theory, vol. IT-37, No. 5, pp. 1241-60, Sep. 1991, the contents of which are hereby incorporated by reference. It may be desirable in this embodiment to minimize the least- squares error between the geometrically uniform signals and the predetermined set of signals {s1(t), s2(t) . . . sm(t)}. Thus, an alternative embodiment of the present invention comprises minimizing the least-squares error between {q1(t), q2(t) . . . qm(t)} and {s1(t), s2 (t) . . . sm (t)} and requiring {q1(t), q2(t) . . . qm(t)} to be geometrically uniform.

[0114] Embodiment No. 7—Linearly Dependent Signals & Projected Orthogonal Signals

[0115] If the desired correlation shape is to have the output of the output of the correlation demodulator be decorrelated on the space in which it lies, the set of signals {q1(t), q2(t) . . . qm(t)} should be chosen as a projected orthogonal signal set. In yet another embodiment to be used when the predetermined set of signals is linearly dependent, the signals, {q1(t), q2(t) . . . qm(t)} 233 in FIG. 4, are the closest projected orthogonal signals to the predetermined set of signals, {s1(t), s2(t) . . . sm(t)}, in a least-squares sense. The projected orthogonal signals are a projection of a set of orthonormal signals onto the space spanned by the predetermined set of signals, {s1(t), s2(t) . . sm(t)}. Analogously to the previous embodiment, skilled artisans will recognize that, depending upon particular system constraints, an absolute minimum least-squares value need not be attained in order to achieve adequate performance.

[0116] Embodiment No. 8—Linearly Dependent Received Signals & Projected Geometrically Uniform Signals

[0117] The present invention can also be used to achieve a particular correlation shape among the members of the output vector a 240 when the predetermined signals {s1(t), s2(t) . . . sm(t)} are linearly dependent. If the covariance matrix of a 240 when represented in terms of an orthonormal basis for the space in which it lies has the property that subsequent rows are permutations of the first, then {q1(t), q2(t) . . . qm(t)} should be selected to be a set of projected geometrically uniform set of signals, i.e., a set of geometrically uniform signals projected onto the space spanned by the predetermined set of signals. In an alternative embodiment, the signals {q1(t), q2(t) . . . qm(t)} 233 are chosen to be the closest projected geometrically uniform signals in a least-squares sense to the signals {s1(t), s2(t) . . . sm(t)}. In this way, an engineer is able to determine a correlation shape for {a1, a2 . . . am} 240, when the predetermined set of signals {s1(t), s2(t) . . . sm(t)} is linearly dependent.

[0118] Embodiment No. 9—Method of Performing Signal Detection

[0119] An additional embodiment comprises a method for processing a signal from a predetermined set of signals, wherein the method includes the steps of receiving a member of a predetermined set of signals; demodulating the member; outputting a demodulated signal; and shaping the correlation of the demodulated signal.

[0120] Test Results for Embodiment Nos. 5 & 7

[0121] The results of tests conducted using an implementation of Embodiment Nos. 5 & 7 are provided below. In each test, signal detection was executed and the results compared to the signal detection capabilities of a matched filter. One skilled in the art will recognize that the signal detection properties tested herein are indicative of the performance that the present invention would exhibit when implemented as part of a receiver or as part of a method of signal detection. Two types of random signals were used, each 100 samples in length: (1) binary signals where each sample takes on the value 0 or 1 with equal probability; and, (2) computer simulated analog signals where each sample is uniformly distributed between 0 and 1. The non-Gaussian noise distribution was chosen as a Gaussian mixture of two components with equal weights.

[0122] For ease of notation, the embodiments of FIG. 4 will be referred to in short hand. In the case where the predetermined set of received signals is linearly independent, the orthogonal matched filter will be referred to as “OMF.” In the case of where the predetermined set of received signals is linearly dependent, the projected orthogonal matched filter will be referred to as “POMF.” Both the OMF and the POMF were tested in Gaussian and non-Gaussian noise environments.

[0123] A Gaussian mixture of two components each with standard deviation 0.25 centered at ±1, corresponding to a signal-to-noise-ratio (“SNR”) close to 0 dB was tested first. Five hundred realizations of each type of signal were generated. For each realization of the signal, a probability of detection was determined by recording the number of successful detections over 500 noise realizations. The histograms of the probability of detection Pd for the different detectors were plotted. These histograms indicated that Pd has a unimodal distribution with a bell-shaped appearance. The results of these tests are presented in terms of the mean and standard deviation of Pd for the various detectors in FIG. 5 and FIG. 6.

[0124] Reference will now be made to FIG. 5 and FIG. 6. These figures contain the mean of Pd for the OMF detector and the matched filter detector as a function of the number of signals in the transmitted constellation. The vertical lines indicate the standard deviation of Pd. The results in FIG. 5 were obtained for binary signals, and the results in FIG. 6 for computer simulated analog signals. Both figures demonstrate the superior results obtained by the OMF detector when compared with the matched filter detector.

[0125] FIG. 7 is a plot of the mean and standard deviation of Pd for the POMF detector and the matched filter detector as a function of the number of signals in the transmitted constellation. In this simulation, the Gaussian mixture of two components had a standard deviation 0.25 centered at ±1 . For each signal constellation, ⅔ (rounded up) of the signals were linearly independent. The dependent signals were generated by randomly mixing a set of linearly independent signals where the signal and mixing elements were uniformly distributed between 0 and 1.

[0126] The qualitative behavior of the POMF detector is similar to that of the OMF detector, even when compared under various Gaussian mixture parameters and SNRs. FIGS. 5-7 demonstrate the superior results obtained with OMF and POMF detectors when compared to the matched filter detector in a non-Gaussian noise environment.

[0127] These tests were repeated in a Gaussian noise environment, the results of which are provided in FIGS. 8-10. The details of the simulations were identical to those of FIGS. 5, 6, and 7, respectively, with the Gaussian noise replacing the Gaussian mixture noise. FIGS. 11(a) and (b) are plots of the difference in the probability of detection (on a scale of 10−3) using the OMF and matched filter detectors for transmitted constellations of 15 signals in Gaussian noise, for binary and computer simulated analog signals respectively. FIG. 11(c) is a plot of the difference in the probability of detection using the POMF and matched filter detectors in Gaussian noise. The qualitative behavior of the OMF and POMF detectors in comparison to the matched filter detector is similar for varying values of the variance of the Gaussian noise.

[0128] For Gaussian noise the matched filter detector outperformed the OMF and the POMF detectors. This appears consistent with the fact that the matched filter detector maximizes the probability of detection for Gaussian noise. However, it is evident from FIGS. 8-11 that the relative improvement in performance using the matched filter detector over the OMF and POMF detectors is insignificant.

[0129] Other embodiments of the invention will be apparent to those skilled in the art from a consideration of the specification or practice of the invention disclosed herein. It is intended that the specification and examples are illustrative only. Although various embodiments have been described in detail, it should be understood that various changes, substitutions, and alterations can be made herein by one of ordinary skill in the art without departing from the scope of the present invention as hereinbefore described and as hereinafter claimed. What is claimed is:

Claims

1. A receiver comprising: a correlation demodulator for receiving signals; and a correlation shaper operating on an output from the correlation demodulator.

2. The receiver of claim 1 wherein the correlation demodulator is a matched filter demodulator.

3. The receiver of claim 1, wherein the correlation shaper is a whitening transformation performed on the output from the correlation demodulator.

4. The receiver of claim 3, wherein the whitening transformation is determined by minimizing the mean squared error between the output from the correlation demodulator and the output from the correlation shaper.

5. The receiver of claim 1, wherein the correlation shaper is comprised of a transformation, said transformation is determined by minimizing the mean squared error between the output of the correlation demodulator and the output of the correlation shaper.

6. The receiver of claim 1, wherein the correlation shaper is chosen so that the covariance matrix of the output has the property that the second and subsequent rows are permutations of the first row.

7. The receiver of claim 6, wherein the correlation shaper is determined by minimizing the mean squared error between the output from the correlation demodulator and the output from the correlation shaper.

8. The receiver of claim 1, wherein the correlation shaper is a subspace whitening transformation performed on the output signal of the correlation demodulator.

9. The receiver of claim 8, wherein the subspace whitening transformation is determined by minimizing the mean squared error between the output from the correlation demodulator and the output from the correlation shaper.

10. The receiver of claim 5, wherein the transformation is performed on a subspace.

11. The receiver of claim 1, wherein the correlation shaper is chosen so that the covariance matrix of the representation of the output of the correlation shaper in the space in which it lies has the property that the second and subsequent rows are permutations of the first row.

12. The receiver of claim 11, wherein the correlation shaper is determined by minimizing the mean squared error between the output from the correlation demodulator and the output from the correlation shaper.

13. The receiver of claim 1, wherein the correlation demodulator is comprised of a bank of correlators that cross-correlate a received signal with a set of orthogonal signals.

14. The receiver of claim 13, wherein the set of orthogonal signals is determined by minimizing the least-squares error between the set of orthogonal signals and the predetermined set of signals.

15. The receiver of claim 1, wherein the correlation demodulator is comprised of a bank of correlators that cross-correlate a received signal with a set of geometrically uniform signals.

16. The receiver of claim 15, wherein the set of geometrically uniform signals is determined by minimizing the least-squares error between the set of geometrically uniform signals and the predetermined set of signals.

17. The receiver of claim 1, wherein the correlation demodulator is comprised of a bank of correlators that cross-correlate a received signal with a set of projected orthogonal signals.

18. The receiver of claim 17, wherein the set of projected orthogonal signals is determined by minimizing the least-squares error between the set of projected orthogonal signals and the predetermined set of signals.

19. The receiver of claim 1, wherein the correlation demodulator is comprised of a bank of correlators that cross-correlate a received signal with a set of projected geometrically uniform signals.

20. The receiver of claim 19, wherein the set of projected geometrically uniform signals is determined by minimizing the least-squares error between the set of projected geometrically uniform signals and the predetermined set of signals.

21. A method for processing signals comprising the steps of: receiving a member of a predetermined set of signals that has undergone some distortion; demodulating the received signal of the predetermined set of signals; outputting a demodulated signal; and shaping the correlation of the demodulated signal.

22. The method of claim 21, wherein the shaping of the correlation of the demodulated signal is comprised of performing a whitening transformation on the demodulated signal.

23. The method of claim 22, wherein the whitening transformation is comprised of the step of minimizing the mean squared error between the demodulated signal and an output signal from the whitening transformation.

24. The method of claim 21, wherein shaping of the correlation of the demodulated signal is further comprised of performing a transformation on the demodulated signal, said transformation is determined by minimizing the mean squared error between the demodulated signal and the output of the correlation shaper.

25. The method of claim 21, wherein shaping of the correlation of the demodulated signal is comprised of performing a transformation of the demodulated signal such that the covariance matrix of the output of the transformation has the property that the second and each subsequent row is a permutation of the first.

26. The method of claim 25, wherein the transformation is comprised of the step of minimizing the mean squared error between the demodulated signal and an output signal from the transformation.

27. The method of claim 21, wherein shaping of the correlation of the demodulated signal is comprised of performing a subspace whitening transformation on the demodulated signal.

28. The method of claim 27, wherein the whitening transformation is comprised of the step of minimizing a mean squared error between the demodulated signal and an output signal from the whitening transformation.

29. The method of claim 21, wherein shaping of the correlation of the demodulated signal is comprised of performing a transformation of the demodulated signal such that the covariance matrix of the representation of the output of the transformation on the space in which it lies has the property that the second and each subsequent row is a permutation of the first.

30. The method of claim 29, wherein the transformation is comprised of the step of minimizing a mean squared error between the demodulated signal and an output signal from the transformation.

31. The method of claim 21, wherein whitening the demodulated signal is comprised of the step of cross-correlating the member of a predetermined set of signals with a set of orthogonal signals.

32. The method of claim 31, further comprising the step of minimizing the least- squares error between the predetermined set of signals and the set of orthogonal signals.

33. The method of claim 21, wherein shaping the correlation of the demodulated signal is comprised of the step of cross-correlating the member of a predetermined set of signals with a set of geometrically uniform signals.

34. The method of claim 33, further comprising the step of minimizing the least- squares error between the predetermined set of signals and the set of geometrically uniform signals.

35. The method of claim 21, wherein shaping the correlation of the demodulated signal on a subspace is comprised of the step of cross-correlating the member of a predetermined set of signals with a set of projected orthogonal signals.

36. The method of claim 35, comprising the step of minimizing the least-squares error between the projected orthogonal signals and the predetermined set of signals.

37. The method of claim 21, wherein shaping the correlation of the demodulated signal on a subspace is comprised of the step of cross-correlating the member of a predetermined set of signals with a set of projected geometrically uniform signals.

38. The method of claim 37, comprising the step of minimizing the least-squares error between the projected geometrically uniform signals and the predetermined set of signals.