The present invention relates to an apparatus, system, and method for wireless communications, and, in particular embodiments, to an apparatus, system, and method for self-interference cancellation in wireless communication systems.
Current half-duplex wireless communication systems employ two orthogonal channels to transmit and receive. Full-duplex (FD) systems allow better exploitation of these resources by transmitting and receiving on the same channel. The main deterrent in employing FD systems is the large self-interference (SI) as compared to the intended signal. It is, therefore, desirable to have apparatuses, systems, and methods to reduce the SI in order to allow the intended signal to be detected.
In accordance with an embodiment, a method for reducing self-interference (SI) in a full-duplex capable transceiver includes obtaining an adjusted signal, wherein the adjusted signal is a difference signal between a received signal in an analog domain and an estimated SI, wherein the estimated SI is estimated according to an SI received at a receiver during a half-duplex operation; and obtaining an intended signal, wherein the intended signal is a difference signal between the adjusted signal in a digital domain and an estimated residual SI, and wherein the estimated residual SI is an amount of SI remaining in the adjusted signal after removal of the estimated SI from the received signal.
In accordance with another embodiment, a method for reducing self-interference (SI) in a full-duplex capable transceiver includes obtaining, by the transceiver, an adjusted signal, wherein the adjusted signal is a difference signal between a received signal in an analog domain and an estimated SI signal, wherein the estimated SI signal is estimated according to an SI signal received at a receiver during a training period during a half-duplex operation; and obtaining, by the transceiver, an intended signal according to an estimated residual SI signal and the adjusted signal.
In accordance with another embodiment, a full-duplex capable wireless network component includes an antenna sub-system configured for full-duplex operation; a self-interference (SI) channel estimation component configured to estimate an SI signal during a training phase mode; an radio-frequency (RF) self-interference cancellation stage component configured to obtain an adjusted RF signal according to a difference signal between a received RF signal and the estimated SI signal in a RF domain during a full-duplex operation mode; an analog-to-digital converter (ADC) configured to convert the adjusted RF signal to a digital adjusted signal; and a baseband SI cancellation stage configured to obtain the digital intended signal in a digital domain according to a difference signal between the digital adjusted signal and a residual SI signal.
For a more complete understanding of the present invention, and the advantages thereof, reference is now made to the following descriptions taken in conjunction with the accompanying drawing, in which:
The making and using of the presently preferred embodiments are discussed in detail below. It should be appreciated, however, that the present invention provides many applicable inventive concepts that can be embodied in a wide variety of specific contexts. The specific embodiments discussed are merely illustrative of specific ways to make and use the invention, and do not limit the scope of the invention.
Full-duplex operation by allowing simultaneous transmission/reception over the same channel has the potential to double the transmission rate of half-duplex if the self-interference signal can be perfectly suppressed (or reasonably suppressed) from the received signal. However, as mentioned above, one of the key deterrents in implementing a full-duplex transceiver is the large SI from the wireless device's own transmission. The SI is usually several orders of magnitude higher than the signal of interest because the later signal crosses longer distance than does that of the SI signal. Recent research results showed that, using different cancellation stages, it is possible to sufficiently attenuate the SI such that the signal of interest is properly detected.
In a practical environment, it is difficult, if not impossible, to completely cancel the self-interference due to imperfect channel estimation. Therefore, channel estimation is a critical issue in full-duplex systems. In one system, the coefficients of the self-interference channel are obtained in the frequency domain by dividing the received signal by the known transmit symbol over each subcarrier. However, this approach ignores the sparsity of the channel. In another system, a two-step Least Square (LS)-based estimator is used where a first estimate of the self-interference channel is obtained by considering the actual signal as additive noise. After that, the interference is suppressed and the resulting signal is used to detect the intended data. A more precise estimate of the channel is then obtained by jointly estimating the self-interference and intended signal channels using the known transmitted data and detected data. However, an initial estimate of the intended signal channel is important in the detection of the intended data.
Disclosed herein are apparatuses, systems, and methods for SI reduction in a FD system. In an embodiment, the SI cancellation or reduction is performed in the radio-frequency (RF) level to avoid saturation/overloading of the low noise amplifier (LNA) and analog-to-digital converter (ADC). The residual SI that remains after the RF SI cancellation is reduced in the baseband. An estimate of the SI signal is determined in order to subtract it from the received signal. To obtain this estimate, the transmit SI data is known, but the SI propagation channel may be unknown. Disclosed herein is a transmission protocol for switching from HD to FD in order to estimate the SI channel. In an embodiment, a half-duplex transmission period is used at the beginning of a transmission to estimate the self-interference channel and then used reduce the self-interference without affecting the intended signal when switching to full-duplex transmission at the completion of the estimation period. The mode is switched from HD to FD once the training period is over. This protocol allows for good channel estimation and SI cancellation or reduction performance.
In an embodiment, during a short HD-initialization phase, the wireless node receives only the self-interference from a transmit data and estimates the SI channel that can be used to reduce the SI during the FD period. This HD-initialization period allows accurate estimates of the SI channel to establish SI-cancellation (or reduction) at the RF. The transmitter (Tx) adjusts its Tx power to allow more accurate SI channel estimation using its existing receiver (Rx) ADC.
In an embodiment, in FD operation, the SI is cancelled before the LNA/ADC to avoid LNA/ADC overloading/saturation and further self-interference suppression can be done after ADC at the baseband. Usually, no additional processing can be done before at least some of the SI is cancelled or reduced. A replica of the self-interference for cancellation can be created from the known transmit signal and the estimate of the self-interference channel. The SI-channel estimate obtained in the initial HD period is fed back to the RF cancellation stage to create a cancellation signal and subtract it from the received signal. In an embodiment, residual SI exists due to estimation error. Additional processing is performed in the digital domain to further reduce the SI.
Embodiments of the disclosure can be combined with existing passive cancellation by using passive circuit and antenna combinations.
In an embodiment, disclosed, is a self-interference channel estimation and cancellation system and method in a full-duplex transceiver in two steps. In an embodiment, the first accurate self-interference channel estimate is obtained in a short initial half-duplex period for the radio-frequency (RF) self-interference-cancellation stage prior to the LNA/ADC. Noting the self-interference channel sparse structure dominated by a relatively small number of clusters of significant paths, in an embodiment, its sensing matrix satisfies the restricted isometry property (RIP). Hence, compressed-sensing (CS) theory can be applied to exploit its sparsity by using a mixed-norm optimization criteria to return the non-zero coefficients and to develop an accurate CS-based self-interference channel estimate with much fewer samples than the linear reconstruction method. In an embodiment, the regularization parameter is derived. The regularization parameter can be selected to keep the residual self-interference not exceeding the intended signal level.
In the second step during the full-duplex operation, a subspace-based process is disclosed to jointly estimate the residual self-interference and intended signal channels for the baseband self-interference cancellation stage. Since the channels are obtained up to a matrix ambiguity, disclosed is a method to find the expression of the self-interference channel ambiguity and a phase ambiguity resolution scheme for the intended signal channel estimation with much smaller number of training samples than traditional data-aided estimator. In an embodiment, a substantially minimal amount of training data is used. The small amount of training data used in the disclosed channel estimator can be explained by the fact that the estimator exploits the information bearing in the unknown data to find the subspace of the transmit signal. The knowledge of the signal subspace reduces the number of the remaining parameters to estimate compared to the LS estimator.
In an embodiment, two channel estimation techniques for the RF and baseband self-interference cancellation stages in full-duplex MIMO transceivers are disclosed. The first process for the RF self-interference cancellation stage is based on the concept of compressed sensing to reduce the self-interference power to at least the same level of the intended signal. Then, in the baseband cancellation stage, a subspace-based channel estimator is applied to find the residual self-interference channel and cancel the residual self-interference. This disclosed process performs a joint estimation of the residual self-interference and intended signal channels by exploiting the available knowledge of the self-signal while the intended signal is unknown. Compared to the standard non-blind LS estimator, the disclosed scheme does not require training blocks to find the residual self-interference channel and needs fewer training data to solve the intended signal channel ambiguity and, therefore, offers better bandwidth efficiency. Simulation results have shown that the disclosed process improves the channel estimation accuracy and the cancellation performance.
In an embodiment, a method for reducing self-interference (SI) in a full-duplex capable transceiver is disclosed. The method includes subtracting, with the transceiver, an estimated SI signal from a received signal in an analog domain to produce an adjusted signal, wherein the estimated SI signal is estimated according to a transmitted signal received at the transceiver during a half-duplex operation. The method also further includes subtracting, with the transceiver, an estimated residual SI signal from the adjusted signal in a digital domain to obtain an intended signal, wherein the residual SI is an amount of SI signal remaining in the adjusted signal after removal of the estimated SI from the received signal. In an embodiment, subtracting the estimated SI signal is performed before the adjusted signal arrives to a low noise amplifier and before the adjusted signal arrives to an analog-to-digital convertor. In an embodiment, the transmit power of the transceiver is adjusted according to the transmitted signal received at its own receiver during the half-duplex operation to improve an accuracy of the SI channel estimation.
In another embodiment, a method for reducing self-interference (SI) in a full-duplex capable transceiver is disclosed. The method includes determining, by the transceiver, an estimated SI signal during a training period; subtracting, by the transceiver, the estimated SI signal from a received signal during full-duplex operation to produce an adjusted signal; estimating, by the transceiver, a residual SI signal according to the estimated SI signal, wherein the residual SI signal comprises an error in the estimated SI signal; and subtracting the residual SI signal from the adjusted signal to produce an intended signal. The estimated SI signal is subtracted from the received signal in a radio-frequency (RF) domain before the received signal is amplified and converted into a digital signal. The residual SI signal is subtracted from the adjusted signal in a baseband. In an embodiment, the power of the SI signal is reduced according to the estimated SI signal obtained in the training period.
In another embodiment, a full-duplex capable wireless network component is disclosed. The wireless network component includes an antenna sub-system configured for full-duplex operation; a self-interference (SI) channel estimation component configured to estimate a SI signal during a training phase mode; an radio-frequency (RF) self-interference cancellation stage component configured to subtract the estimated SI signal from a received RF signal in a RF domain to produce an adjusted RF signal during a full-duplex operation mode; an analog-to-digital convertor (ADC) configured to convert the adjusted RF signal to a digital adjusted signal; and a baseband SI cancellation stage configured to subtract a residual SI from the digital adjusted signal in a digital domain. In an embodiment, the estimated SI signal is subtracted from the received signal in software. In an embodiment, the digital intended signal is obtained by subtracting the residual SI signal from the digital adjusted signal in software. The SI channel estimation component is configured to determine the estimated SI according to a compressed-sensing-based procedure and/or according to a mixed-norm optimization criteria that returns non-zero coefficients for a compressed-sensing based self-interference channel estimate. The baseband SI cancellation stage is configured to determine the residual SI according to a maximum likelihood function. The antenna sub-system comprises a multi-antenna sub-system and the training phase mode is a half-duplex mode.
There are many reasons that render it beneficial to develop another process in the second cancellation stage different from the process in the first stage. First, the residual self-interference channel after the first cancellation stage is completely random without any specific sparse structure. Moreover, in an embodiment, it may be desirable to jointly estimate the residual self-interference and the intended signal channels without knowing the data. In this situation, the compressed sensing estimator cannot recover the channel coefficients without a perfect knowledge of the data.
Simulation results show that the disclosed processes outperform the LS processes with better bandwidth efficiency since they do not require any training data to estimate the self-interference channel. The disclosed processes offer the remarkable signal-to-residual-self-interference-and-noise ratio (SINR) after the RF and baseband self-interference-cancellation stages approaching the signal-to-noise ratio (SNR).
In this disclosure, we adopt the following notations. (.)T, (.)H and (.)# refer to matrix transpose, conjugate transpose, and pseudo-inverse, respectively. For a matrix M, we use det(M) and trace(M) to denote the determinant and the trace, respectively. The operator ⊕ refers to the Kronecker product of two matrices. Ip refers to the p×p identity matrix. └x┘ rounds the real x to the largest integer smaller or equal to x. Finally, let ∥.∥1 and ∥.∥2 denote the l1- and the l2-norms, respectively and ∥.∥0 counts the number of nonzero entries of its argument.
In an embodiment, the AP 110 and UEs 120 are configured to operate in FD mode. In order to provide high isolation of transmitter power from on frequency co-located receivers in the AP 110, the AP 110 includes a self-interference cancellation apparatus and system described in more detail below. In an embodiment, the AP 110 is a cellular AP. In another embodiment, the AP 110 is a WiFi AP.
I. Full-Duplex System Model
Returning to
Considering multipath channels, the received nth complex-baseband equivalent sample of the Rx stream r can be written as:
where xq(n) and sq(n), for n=0; . . . ; N−1 are the transmitted samples from the Tx stream q of the same transceiver and from the other intended transmitter, respectively. hr,q(i)(l); l=0; . . . ; Li is the Li-tap impulse response of the self-interference channel from Tx stream q to Rx stream r of the same transceiver and hr,q(s)(l); l=0; . . . ; Ls is the Ls-tap impulse response of the intended signal channel from Tx stream q of the other intended transmitter to Rx stream r. wr(n) is the additive thermal noise in Rx stream r. The first and second terms in (1) represent the self-interference and intended signal, respectively. For simplicity, we assume Li=Ls=L. From equation (1), it follows that the vector y(n) can be written as:
where
y(n)=[y1(n),y2(n), . . . ,yNr(n)]T,
h
(i)(l)=[h1(i)(l),h2,q(i)(l), . . . ,hN
h
q
(i)(l)=[h1,q(i)(l),h2,q(i)(l), . . . ,hN
h
(s)(l)=[h1(s)T(l), . . . ,hN
h
q
(s)(l)=[h1,q(s)(l),h2,q(s)(l), . . . ,hN
w(n)=[w1(n);w2(n), . . . ,wN
In equation (2), X(n−l) is a Nr×NtNr Toeplitz matrix with the first column given by the Nr×1 vector [x1(n−l), 0, . . . , 0] and the first row given by [x1(n−l), x2(n−l), . . . , xNe1 with e1 being the 1×Nr vector having one in the first element and zeroes elsewhere. The matrix S(n−l) is constructed in the same way as X(n−l) but with transmitted samples sq(n−l) from the other intended transmitter. Now let the two N_t N_r (L+1)×1 vectors h(i) and h(s) gather all the coefficients of the self-interference and intended signal channels, respectively, i.e.,
h
(i)
=[h
(i)T(0),h(i)T(1), . . . ,h(i)T(L)]T,
h
(s)
=[h
(s)T(0),h(s)T(1), . . . ,h(s)T(L)]T. (4)
And define:
The NrN×NtNr(L+1) self-signal matrix X includes samples transmitted from the same transceiver and, the NrN×NtNr(L+1) intended signal matrix S contains samples transmitted from the other intended transmitter. Then, the received NrN×1 vector y=[yT(0), . . . , yT(N−1)]T is given by:
y=Xh
(i)
+Sh
(s)
>+w, (6)
where w is the NrN×1 thermal noise vector.
In full-duplex systems, the self-interference, shown by the 1st term in equation (6), is many order of magnitude higher than the intended signal from the other intended transmitter, shown by the 2nd term in equation (6). This imposes different cancellation stages to reduce the self-interference to a sufficiently low level for proper signal detection. The RF cancellation stage aims to suppress the self-interference prior to the LNA/ADC. Since the self-signal matrix X is known, we only need to estimate the self-interference channel h(i) to generate the self-interference replica at RF for cancelation. Remaining self-interference after ADC will be further suppressed by the baseband cancellation stage by digital signal processing at baseband as shown in
II. Compressed-Sensing-Based RF Cancellation Stage
As previously discussed, one major task in the RF cancellation stage is to estimate the self-interference channel vector h(i). Since the self-signal matrix X is known, the straightforward approach to find h(i) is to employ a linear estimator. In general, a linear estimate of h(i) is given by:
ĥ
(i)
=My, (7)
where the NrNt(L+1)×NrN matrix M (to be derived) determines the estimate of h(i). There are a large number of different estimates of h(i). For example, using the least square (LS) criterion, M will be given by (XHX)−1XH, while using minimum mean squared error (MMSE) estimator, M=E{h(i)h(i)H}XH(XE{h(i)h(i)H}XH)−1, where E{.} denotes statistical expectation. While the later needs to knowledge of the second order statistic of the channel, it enjoys substantially lower channel estimate error as compared to the LS estimator. Once an estimate of the self-interference channel is available, the self-interference replica is generated and subtracted from the received signal in equation (6) to obtain:
where we have substituted the expression of y from equation (6) into ĥ(i) in equation (7). In order to suppress the self-interference, one should design M such that the 1st term in equation (8), i.e., (IN
During the initial half-duplex fashion period, the transceiver receives only its own signal. The signal model in equation (6) reduces to:
y=Xh
(i)
+w. (9)
The estimation of the self-interference channel h(i) is equivalent to the traditional problem of training based channel estimation. Usually, the processes to solve this problem rely on linear LS strategies. However, these methods do not exploit the particular structure of the channel. As confirmed by measurements, the self-interference channel between close-by antennas in the same transceiver, exhibits a very strong path component compared to the reflected paths, and hence the vector h(i) contains a few dominant components. Therefore, the problem turns out to estimating a sparse channel from the observation y. Hence, mathematically, we are looking for arg minh∥h∥0 such that y=Xh. This is, however, a difficult combinatorial optimization problem and may be intractable even for small size problem. Recently, it has been shown that when h is sparse enough compared to X, it is possible to replace ∥h∥0 by ∥h∥1 in the optimization problem and we still obtain the exact same solutions for both problems. The new problem:
is a convex optimization problem and can be solved by linear programming. In practice, only noisy measurements are available. Therefore, the constraint y=Xh is replaced by ∥y=Xh∥22≦λ, for some parameter λ, to introduce the additive noise. This optimization problem is computationally tractable since it can be recast as a second-order cone programming.
The parameter λ specifies how much error we wish to allow. In the following, we propose an approach to select the regularization parameter A that is suitable for the following baseband cancellation stage. First, if we are able to obtain the exact value of h, we will have ∥y=Xh∥22=∥w∥22 which can be approximated to σ2NrN for sufficiently large noise vector w, where σ2 is the noise variance. However, the estimated value ĥ cannot exactly match the real channel h(i). Let h(r) denotes the residual channel (h(r)=h(i)−ĥ(i)). In that case, we have:
y=Xĥ
(i)
=Xh
(r)
+w (11)
where the term Xh(r) represents the residual self-interference after the RF cancellation stage. In order to effectively estimate h(r) in the following baseband cancellation stage, the power of the residual interference should be reduced to, at most, the same power of the intended signal. Therefore, using the estimated vector ĥ(i), we want to obtain:
where PS is the power of the received intended signal. To that end, the regularization parameter λ is chosen high enough so that (PS+σ2)NrN≦λ to guarantee that the residual interference is in the same order of magnitude as the intended signal. The attractive feature in compressed sensing theory is that if h(i) is sparse, then a smaller number of measurements than the length of h(i) is sufficient to recover h(i). This reconstruction ability depends on some properties of the matrix X. In particular, it suffices that the matrix X satisfies the restricted isometry property (RIP) as follows. Let S denotes the number of non-zero elements in the vector h(i). According to the definition RIP, X satisfies the RIP2 (the RIP guaranties the uniqueness of the solution to the problem. In fact, for any two different S sparse vectors θ1 and θ2, the vector θ1−θ2 has at most 2S non zero elements (if the non-zero elements of θ1 and θ2 are not in the same positions). According to the RIP inequality, the two images of θ1 and θ2 are different as long as θ1 is different from θ2.) of order 2S with parameter δSε[0,1], for a given integer S, if for every vector θ such that ∥θ∥0≦2S we have:
(1−δS)∥θ∥22≦∥Xθ∥22≦(1+δS)∥θ∥22. (13)
In other words, X satisfies the RIP if the singular values of all the submatrices XT, formed from X by taking the columns indexed by T from X, are in └√{square root over (1−δS)}, √{square root over (1+δS)}┘, where T⊂{1, . . . , NtNr(L+1)} with cardinality no larger than S. It follows that, to prove the RIP for a given matrix, it suffices to bound the eigenvalues of the S×S Grammian matrix GT=XTHXT in the interval [1−δS, 1+δS], for all subsets of column indices T. According to the Ger{hacek over (s)}gorin's Disc theorem, the eigenvalues of GT lie in the union of the S discs di centered at ci=GT(i, i) and with radius ri=Σj≠i, j=1|GT(i, j)|, for i=1, . . . , S. That is, for two δd and δo real in [0,1] and satisfying δd=δo=δS, if all the diagonal elements of GT verify |GT(i, i)−1|<δd| and all the off-diagonal elements satisfy |GT(i,j)−1|<δo/S, then all the eigenvalues of GT contained in the union of the discs di, i=1, . . . , S, are in the range [1−δS, 1+δS]. As shown in Appendix 1, it follows that the matrix X satisfies the RIP with parameter δS with probability exceeding:
where c2 is a constant depending only on δS and specified in Appendix 1.
III. Subspace-Based Baseband Cancellation Stage
Once the two-way communications start full-duplex operation, the self-interference channel estimate obtained during the training period is used to reduce the power of the self-interference. After the RF cancellation stage, the resulting signal in baseband is given by:
where we use the similar vector structures as above. In the baseband cancellation stage, the task is to reduce the residual self-interference signal represented by the first term in equation (15). To that end, we need to estimate the residual self-interference channel from yc(n). Since the self-signal is known, the simplest way to estimate the corresponding channel is to resort to a linear estimator. But this method will suffer from large estimation error since the intended signal appears as additive noise. Therefore, the intended signal also should be considered in the estimation process to jointly estimate the residual self-interference and the intended signal channels. In this section, we develop a subspace-based method for jointly estimating these two channels. Before presenting the channel estimator, we need to have a more tractable representation of the received signal yc(n) to introduce the disclosed process. By defining:
x(n)=[x1(n),x2(m), . . . ,xN
s(n)=[s1(n),s2(m), . . . ,sN
H
(r)(l)=[h1(r)(l),h2(r)(l), . . . ,hN
H
(s)(l)=[h1(s)(l),h2(s)(l), . . . ,hN
the cancelled input signal yc(n) can be expressed as:
Then, we gather the two channel matrices H(s)(l) and H(r)(l) in one matrix H(l)=[H(r)(l)H(s)(l)] and define the NrM×2NtN lower triangular block Toeplitz matrix:
where M=N+L and the transmitted data in one 2NtN×1 vector:
x=[x
T(0),sT(0), . . . ,xT(N−1),sT(N−1)]T, (19)
Using these notations, the received NrM vector over the Nr antennas is given by:
Note that for multi-block transmission, the vector in equation (20) is indexed according to the block number t, i.e., yc(t). We omit this indexation for simplicity and we consider a given number of block to later estimate the covariance matrix of yc.
We assume that the noise samples are uncorrelated, i.e., E(w(n)w*(m))=σ2 if n=m and 0 if n≠m, and the noise and signal samples are also uncorrelated. It follows that, the covariance matrix Ry
where Rx is the 2NNt×2NNt covariance matrix of x.
In practice, the sample estimate, {circumflex over (R)}y
The signal subspace is the span of the columns of the matrix H and the noise subspace is the orthogonal complement to the signal subspace. By assuming independent channels between different antennas, the dimension of the signal subspace is 2NNt (the rank of HRxHH is 2NNt) and the dimension of the noise subspace is p=NrM−2NNt. To guaranty that the noise subspace is nondegenerate (p>0), the number of transmit antenna in each terminal Nt should be smaller than
Therefore, the matrix Ry
As the signal subspace is spanned by the 2NNt columns of the matrix H and by orthogonally between the signal and noise subspace, the columns of H are orthogonal to any vector in the noise subspace. Then we have:
v
i
H
H=0,i=1,2, . . . ,p. (23)
From equation (23), we conclude that vi spans the left null space of H. Knowing the left null space of H, it is possible to determine the space spanned by the column of H, denoted by span(H), i.e., the space containing all the linear combinations of the columns of H. Therefore, knowing the span(H) does not give the exact matrix H since there are infinitely many matrices satisfying equation (23). However, for the specific block Toeplitz matrix that we have at hand in equation (18), it can be shown that if two matrices H1 and H2 have the same form as in equation (18) and satisfy the conditions in equation (23), then there exists a nonsingular 2Nt×2Nt matrix C satisfying:
The proof of the existence of C is similar to that presented in Moulines, et al. with the additional condition of H(0) being full rank matrix. It has been proven that two Toeplitz matrices spanning the same subspace and having all zero elements above the principal diagonal are proportional with a scalar constant of proportionality. In the disclosed case, it turns out that the two matrices are related by a block diagonal matrix.
Recall that we are looking for a matrix that satisfies the set of equations in (23). Since the matrix H is entirely defined by the matrices H(0), . . . , H(L), instead of looking for the whole NrM×2NtN matrix H, we can restrict the search for the Nr×2Nt matrices H(l), l=0, . . . , L. Now considering again the set of equations in (23), each eigenvector vi can be written as:
v
i
=[v
i
T(M),viT(M−1), . . . ,viT(1)]T, (25)
where vi for m=1, M are Nr×1 vectors. Then, each equation in (23) is rearranged as:
or in the following matrix form:
Collecting all the θi matrices in a Np×Nr(L+1) matrix:
θi=[θ1T,θ2T, . . . ,θpT]T, (30)
we can rewrite equation (27) in a more compact form as:
θ=0. (31)
The problem is equivalent to maximize a MUSIC-type spectrum with the spectrum function being
with the additional condition of ≠0 to avoid the all zeroes solution, where ∥.∥F denotes the Frobenius norm. Therefore, the column of
can be obtained by finding a basis of the null space of θ. In practice, we perform the singular value decomposition (SVD) of θ and choose the 2Nt right singular vectors as the columns of
.
As discussed above, the solution is not unique. For 0 obtained from the SVD of θ, the intended signal channel matrix is proportional to
0:
=
0c, (32)
where C is a 2Nt×2Nt invertible matrix. We will next present a method to find the matrix C.
Let H0 denote the block Toeplitz matrix in the form of equation (18) obtained from the estimated matrix 0. Using equation (24), the received vector in equation (20) is reformulated as:
By multiplying the received signal by the pseudo-inverse of H0, the modified 2NtN×1 received signal is given by:
where
c
=[
c
T(0),
we have:
From its definition, the matrix is composed from the concatenation of two matrices,
(r) and
(s), representing the residual self-interference channel and the intended signal channel, respectively (i.e.,
=[
(r)
(s)]). In the same way, we divide C in two 2Nt−Nt matrices C(r) and C(s) where the first one is associated with the residual self-interference channel and the second one is associated with the intended signal channel. Considering this division, we expand equation (34) as follows:
c(n)=C(r)x(n)+C(s)s(n)+
The vector 0H
0)−1. Adopting the Gaussian hypothesis, the log-likelihood function is given by:
The Maximum-Likelihood (ML) estimates of C(r) and C(s) maximize the function (.,.) given in equation (38). The direct maximization of the cost function L(.,.) requires a 4Nt2-dimensional grid search, which is intractable in practice. To overcome this complexity, we look to a closed-form expression of the solution. Noting that L(.,.) is a separable function of the matrices to estimate, we first minimize the cost function with respect to one matrix. The obtained minimum is a function of the other matrix. Then we introduce this minimum back in the expression of the cost function which becomes a single variable function. Minimizing this new function yields the global maximum of the original log-likelihood function. We first maximize the log-likelihood function in equation (38) with respect to P. The solution of this optimization problem is:
Substituting P by PML into the log-likelihood function in equation (38), we obtain the so-called compressed likelihood function, that depends only on the unknown matrix C(r):
The ML estimate of C(r) is given by:
At this point, we need to introduce some definitions. Let {tilde over (C)}(r) denotes the 2Nt2×1 vector obtained by stacking all the columns of C(r)T on top of each other (i.e., {tilde over (C)}(r)=vec(C(r)T)) and {tilde over (x)}(n) be the 2Nt×2Nt2 matrix given by:
{tilde over (x)}(n)=diag(xT(n), . . . ,xT(n)). (42)
Using these notations, the minimization problem in equation (41) is alternatively expressed as:
This modified problem allows us to obtain the following simple least square (LS) solution:
Since we are interested in the ML estimate, we define ΣML as the difference between the ML and LS estimates:
ξML={tilde over (C)}ML(r)−{tilde over (C)}LS(r), (45)
and let ξ={tilde over (C)}(r)−{tilde over (C)}Ls(r) denote the difference between the ML solution and a given value of {tilde over (C)}(r). We also consider the following two notations:
As shown in Appendix 2, the optimization problem at hand is equivalent to:
Its solution is easily obtained by nulling the derivative with respect to f:
Rearranging the expression in equation (48) using the notations given above, the ML estimate of {tilde over (C)}(r) is given by:
Note that the difference between the ML and LS estimates comes from the term {circumflex over (R)}d−1 in equation (49).
For completeness, we present a method to find the ambiguity matrix of the intended signal channel C(s). It can be obtained from the Eigen-decomposition of the matrix PML obtained in equation (39) as follows:
C
ML
(s)
=U
P
D
PΦ, (50)
where DP is a diagonal matrix containing the Nt most significant eigenvalues of the matrix PML and the columns of UP are the corresponding 2Nt×1 eigenvectors. The matrix Φ is a diagonal phase matrix which can be easily found using a small number of training symbols.
Following the discussion in Section II, it is desirable to establish bounds on |GT(i, i)−1| and Σj=1,j≠iS|GT(i,j)|, for all subsets T. In the following proof, the elements of X are Gaussian random variables with mean 0 and variance 1=N. The matrix X also verifies the RIP when its elements have arbitrary variance σx2 by multiplying each term in the inequality in equation (13) by N/σx2. Moreover, we suppose a real matrix X. Using Lemma 5 in Haupt, et al., each diagonal element of GT(i, j)=Σn=1N|xp
Each column of X contains the N transmitted samples from one of the Nt transmitted streams. Therefore, there are exactly Nt different values for GT(i, i). By the union bound, we have for every subset T and for all i=1, . . . , S:
For a given subset T, any off-diagonal element G (i, j) is the inner product between the mi and mj columns of X. For convenience, we write mi as mi=ni+piNr+diNrNt with niε[1, Nr], piε[0, Nt−1] and diε[0, L]. Depending on mi and mj, we distinguish the following different cases:
G
T(i,j)=xp
We gather the previous results along with the union bound to establish an upper bound on the probability that all the elements GT(i,j), for any subset T and i≠j, satisfy
To obtain the result claimed in Section II, let δd=2δS/3, δ0=δS/3 and use equations (53) and (59) to obtain:
Define c1=2(L+1)Nt2+2Nt and for c2<δS2/54, we obtain:
for any
Using the notations introduced in equations (45) and (46), we can write:
and further develop to obtain:
d(n)dH(n)−d(n)({tilde over (x)}(n)ξ)H−{tilde over (x)}(n)ξdH(n)+{tilde over (x)}(n)ξξH{tilde over (x)}H(n). (63)
Injecting equation (63) into the cost function in equation (43), we obtain the following expression:
or the following equivalent cost function:
Noting that, when N is large, the LS and ML estimates are close to the true value. Therefore, the vector ξ can be assumed to be small. And, using the fact that, for ∥M∥<<1,det(I+M)≈1+trace(M) and the property that the trace is invariant under permutations, the minimization problem can be reduced to the one given in equation (47).
The bus 640 may be one or more of any type of several bus architectures including a memory bus or memory controller, a peripheral bus, video bus, or the like. The CPU 610 may comprise any type of electronic data processor. The memory 620 may comprise any type of system memory such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), read-only memory (ROM), a combination thereof, or the like. In an embodiment, the memory 620 may include ROM for use at boot-up, and DRAM for program and data storage for use while executing programs.
The mass storage device 630 may comprise any type of storage device configured to store data, programs, and other information and to make the data, programs, and other information accessible via the bus 640. The mass storage device 630 may comprise, for example, one or more of a solid state drive, hard disk drive, a magnetic disk drive, an optical disk drive, or the like.
The I/O interface 660 may provide interfaces to couple external input and output devices to the processing unit 601. The I/O interface 660 may include a video adapter. Examples of input and output devices may include a display coupled to the video adapter and a mouse/keyboard/printer coupled to the I/O interface. Other devices may be coupled to the processing unit 601 and additional or fewer interface cards may be utilized. For example, a serial interface such as Universal Serial Bus (USB) (not shown) may be used to provide an interface for a printer.
The antenna circuit 670 and antenna element 675 may allow the processing unit 601 to communicate with remote units via a network. In an embodiment, the antenna circuit 670 and antenna element 675 provide access to a wireless wide area network (WAN) and/or to a cellular network, such as Long Term Evolution (LTE), Code Division Multiple Access (CDMA), Wideband CDMA (WCDMA), and Global System for Mobile Communications (GSM) networks. Additional, in some embodiments, the antenna circuit 670 operates in Full Duplex (FD) mode. In some embodiments, the antenna circuit 670 and antenna element 675 may also provide Bluetooth and/or WiFi connection to other devices. In an embodiment, the antenna circuit 670 includes a transmitted signal cancellation system.
The processing unit 601 may also include one or more network interfaces 650, which may comprise wired links, such as an Ethernet cable or the like, and/or wireless links to access nodes or different networks. The network interface 601 allows the processing unit 601 to communicate with remote units via the networks 680. For example, the network interface 650 may provide wireless communication via one or more transmitters/transmit antennas and one or more receivers/receive antennas. In an embodiment, the processing unit 601 is coupled to a local-area network or a wide-area network for data processing and communications with remote devices, such as other processing units, the Internet, remote storage facilities, or the like.
The following references are incorporated herein by reference:
Although the description has been described in detail, it should be understood that various changes, substitutions and alterations can be made without departing from the spirit and scope of this disclosure as defined by the appended claims. Moreover, the scope of the disclosure is not intended to be limited to the particular embodiments described herein, as one of ordinary skill in the art will readily appreciate from this disclosure that processes, machines, manufacture, compositions of matter, means, methods, or steps, presently existing or later to be developed, may perform substantially the same function or achieve substantially the same result as the corresponding embodiments described herein. Accordingly, the appended claims are intended to include within their scope such processes, machines, manufacture, compositions of matter, means, methods, or steps.