Multiple Antennas and Space-Time Communications 10.1 Narrowband MIMO Model In this section we consider a narrowband MIMO channel. A narrowband point-to-point com- munication system of M, transmit and M, receive antennas is shown in Figure 10.1. This system can be represented by the following discrete-time model: [Il = SIE] or simply as y = Hx-+n. Here x represents the M,-dimensional transmitted symbol, n is the M,-dimensional noise vector, and H is the M, x M, matrix of channel gains h, represent- ing the gain from transmit antenna j to receive antenna i. We assume a channel bandwidth iy Figure 10.1: MIMO systems. of B and complex Gaussian noise with zero mean and covariance matrix oy, where typ- [n2] = No/2, the power spectral density of the channel noise. For simplicity, given a transmit power constraint P we will assume an equivalent model with a noise power 0? of unity and transmit power P/o? = p, where p can be interpreted as the average SNR per receive antenna under unity channel gain. This power constraint implies that the input symbols satisfy M, DEL] =p. (10.1) ‘or (equivalently) that Tr( Rx) = p, where Tr(R,) is the trace of the input covariance matrix R. = Elxx1,10.2 Parallel Decomposition of the MIMO Channel ‘We have seen in Chapter 7 that multiple antennas at the transmitter or receiver can be used for diversity gain. When both the transmitter and receiver have multiple antennas, there is an- other mechanism for performance gain called multiplexing gain. The multiplexing gain of a ‘MIMO system results from the fact that a MIMO channel can be decomposed into a number R of parallel independent channels. By multiplexing independent data onto these indepen- dent channels, we get an R-fold increase in data rate in comparison to a system with just ane antenna at the transmitter and receiver. This increased data rate is called the multiplexing gain. In this section we describe how to obtain independent channels from a MIMO system. ‘Consider a MIMO channel with M, x M, channel gain matrix H that is known to both the transmitter and the receiver. Let Ry denote the rank of H. From Appendix C, for any ‘matrix H we can obtain its singular value decomposition (SVD) as H=UEV", (10.2) where the M, x M, matrix U and the M, x M, matrix V are unitary matrices? and where © isan M, x M, diagonal matrix of singular values (¢;) of H. These singular values have the property that 6; = V2; for A; the ith largest eigenvalue of HH”, and Ry of these singular values are nonzero. Because Ry, the rank of matrix H, cannot exceed the number of col- ‘umns of rows of H, it follows that Ru < min(M,, M,). If H is full rank, which is referred to asa rich scattering environment, then Ry = min(M,,.M,). Other environments may lead to a low-rank H: a channel with high correlation among the gains in H may have rank 1. ‘The parallel decomposition of the channel is obtained by defining a transformation on the channel input and output x and y via transmit precoding and receiver shaping. In transmit precoding the input x to the antennas is generated by a linear transformation on input vector ¥ as x = VX. Receiver shaping performs a similar operation at the receiver by multiplying the channel output y by U“, as shown in Figure 10.2. ‘The transmit precoding and receiver shaping transform the MIMO channel into Ry par- allel single-input single-output (SISO) channels with input X and output J, since from the SVD we have that 2 Vand V unitary imply that U"U = Ty, and V"V = Ty,Figure 10.2: Transmit precoding and receiver shaping. Figure 10.3: Parallel decomposition of the MIMO channel.y=U"(Hx +n) = UMUEV"x +n) = UM ULV! VS +n) = UNUEV" VE 4U"n =Et+A, where fi = Un and where © is the matrix of singular values of H with o; on the ith diago- nal and zeros everywhere else. Note that multiplication by a unitary matrix does not change the distribution of the noise — that is, n and ii are identically distributed. Thus, the transmit precoding and receiver shaping transform the MIMO channel into Ry parallel independent channels, where the ith channel has input %;, output i, noise ij, and channel gain a;. Note that the ¢; are related because they are all functions of H, but since the resulting parallel channels do not interfere with each other we say that the channels with these gains are in- dependent - linked only through the total power constraint. This parallel decomposition is shown in Figure 10.3. Since the parallel channels do not interfere with each other, the op- timal ML demodulation complexity is linear in Ry, the number of independent paths that need to be demodulated. Moreover, by sending independent data across each of the parallel channels, the MIMO channel can support Ry times the data rate of a system with just one transmit and receive antenna, leading to a multiplexing gain of Ry. Note, however, that the performance on each one of the channels will depend on its gain o;. The next section will more precisely characterize the multiplexing gain associated with the Shannon capacity of the MIMO channel.10.3 MIMO Channel Capacity This section focuses on the Shannon capacity of a MIMO channel, which equals the maxi: ‘mum data rate that can be transmitted over the channel with arbitrarily small error probability. Capacity versus outage defines the maximum rate that can be transmitted over the channel with some nonzero outage probability. Channel capacity depends on what is known about the channel gain matrix or its distribution at the transmitter and/or receiver. First the static channel capacity under different assumptions about this channel knowledge will be given, which forms the basis for the subsequent section on capacity of fading channels. 10.3.1 Static Channels The capacity of a MIMO channel is an extension of the mutual information formula for a SISO channel given by equation (4.3) to a matrix channel. For static channels a good esti- mate of H can be obtained fairly easily at the receiver, so we assume CSIR throughout this section. Under this assumption, the capacity is given in terms of the mutual information be- tween the channel input vector x and output vector y as C = max I(X; ¥) = max{H(¥) — H(Y | X)] 10.5) ma ny for H(Y) and H(¥ | X) the entropy in y and y | x, as defined in Section 4.1. The definition of entropy yields that H(Y | X) = H(n), the entropy in the noise. Since this noise n has fixed entropy independent of the channel input, maximizing mutual information is equiva- lent to maximizing the entropy Given covariance matrix Rx on the input vector x, the output covariance matrix Ry as- sociated with MIMO channel output y is given by “) = HRW" 4 1y, (10.6) Ely It turns out that, for all random vectors with a given covariance matrix R,, the entropy of y is maximized when y is a zero-mean, circularly symmetric complex Gaussian (ZMCSCG) random vector [5]. But y is ZMCSCG only if the input x is ZMCSCG, and hence this is the optimal distribution on x in (10.5), subject to the power constraint Tr(R,) = p. Thus. we have H(Y) = B log? det{eR,] and H(n) = B log? det| rely, |, resulting in the mutual information 1(X; Y) = Blog? det{Iy, + HR.H"]. (10.7) ‘This formula was derived in (3; 5] for the mutual information of a multiantenna system, and it also appeared in earlier works on MIMO systems (7; 8] and matrix models for ISI channels 19; 10}.‘The MIMO capacity is achieved by maximizing the mutual information (10.7) over all input covariance matrices R, satisfying the power constraint: a= - a C=, max Blogadet{Iu, +HR.H"), (10.8) where det[A] denotes the determinant of the matrix A. Clearly the optimization relative to R, will depend on whether or not H is known at the transmitter. We now consider this opti- mization under different assumptions about transmitter CSI. ‘CHANNEL KNOWN AT TRANSMITTER: WATER-FILLING ‘The MIMO decomposition described in Section 10.2 allows a simple characterization of the MIMO channel capacity for a fixed channel matrix H known at the transmitter and re- ceiver. Specifically, the capacity equals the sum of capacities on each of the independent parallel channels with the transmit power optimally allocated between these channels. This ‘optimization of transmit power across the independent channels results from optimizing the input covariance matrix to maximize the capacity formula (10.8). Substituting the matrix ‘SVD (10.2) into (10.8) and using properties of unitary matrices, we get the MIMO capacity with CSIT and CSIR as on C= max_ YBlogy(1+o7p,), (10.9) pe Liwise where Ry is the number of nonzero singular values o? of H. Since the MIMO channel de- composes into Ry parallel channels, we say that it has Ry degrees of freedom. Since p = P/o?, the capacity (10.9) can also be expressed in terms of the power allocation P; to the ith parallel channel ascy 2 Ru C= max YBioe.(1+ 2) max Yatoe,(1+ (10.10) ia im merrier & PeLs PsP = 07P/a? is the SNR associated with the ith channel at full power. This expres- sion indicates that, at high SNRs, channel capacity increases linearly with the number of degrees of freedom in the channel. Conversely, at low SNRs, all power will be allocated to the parallel channel with the largest SNR (or, equivalently, the largest ?). The capacity for- ‘mula (10.10) is similar to the case of flat fading (4.9) or frequency-selective fading (4.23). Solving the optimization leads to a water-filling power allocation for the MIMO channel: P,_[Vr-Un ne V6 = 1011) P { 0 1 (1 + plihi|?) where the channel matrix H is reduced to a vector ‘h of channel gains, the optimal weight vector e = h"/|[hi|, and p = P/o?.‘CHANNEL UNKNOWN AT TRANSMITTER: UNIFORM POWER ALLOCATION ‘Suppose now that the receiver knows the channel but the transmitter does not. Without chan- nel information, the transmitter cannot optimize its power allocation or input covariance structure across antennas. If the distribution of H follows the ZMSW channel gain model, then there is no bias in terms of the mean or covariance of H. Thus, it seems intuitive that the best strategy should be to allocate equal power to each transmit antenna, resulting in an input covariance matrix equal to the scaled identity matrix: Rx = (/M,)Iu,. It is shown in [4] that, under these assumptions, this input covariance matrix indeed maximizes the mu- tual information of the channel. For an M,-transmit M,-receive antenna system, this yields ‘mutual information given by 139) = Blogs da hy, + fan" Using the SVD of H, we can express this as eu Yaroe( + = Ux: (10.13) o2P/o?. ‘The mutual information of the MIMO channel (10.13) depends on the specific realization of the matrix H,, in particular its singular values (¢;}. The average mutual information of a random matrix H, averaged over the matrix distribution, depends on the probability distri- bution of the singular values of H [5; 11; 12]. In fading channels the transmitter can transmit at a rate equal to this average mutual information and ensure correct reception of the data, as discussed in the next section. But for a static channel, if the transmitter does not know the channel realization (or, more precisely, the channel’s average mutual information) then it does not know at what rate to transmit such that the data will be received correctly. In this ccase the appropriate capacity definition is capacity with outage. In capacity with outage the transmitter fixes a transmission rate R, and the outage probability associated with R is the probability that the transmitted data will not be received correctly or, equivalently, the prob- ability that the channel H has mutual information less than R. This probability is given by Page = o(u Bog2 det} Ly, + fan" < ’). (10.14) This probability is determined by the distribution of the eigenvalues of HH”: these eigen- values are the squares of the singular values of H. The distribution of singular values for matrices is a well-studied problem and is known for common matrices associated with the MIMO channel [12, Sec. 2.1]10.3.2 Fading Channels ‘Suppose now that the channel gain matrix experiences flat fading, so the gains h,; vary with time. As in the case of the static channel, the capacity depends on what is known about the channel matrix at the transmitter and receiver. With perfect CSIR and CSIT the transmitter can adapt to the channel fading, and its capacity equals the average over all channel matrix realizations with optimal power allocation. With CSIR and no CSIT, ergodic capacity and ‘capacity with outage are used to characterize channel capacity. These different characteriza tions are described in more detail in the following sections. ‘CHANNEL KNOWN AT TRANSMITTER: WATER-FILLING With CSIT and CSIR, the transmitter optimizes its transmission strategy for each fading channel realization, as in the case of a static channel. The capacity is then just the average of capacities associated with each channel realization, given by (10.8), with power optimally allocated. This average capacity is called the ergodic capacity of the channel. There are two possibilities for allocating power under ergodic capacity. A short-term power constraint as- sumes that the power associated with each channel realization must equal the average power constraint P. In this case the ergodic capacity becomes = 1 C= En as, Blogs detly, +HR,H 1] = 2 yy =ta[, pax, DB bon( t+ 3 } (10.16) where y; = ¢?P/o?. A less restrictive constraint is a long-term power constraint, where we can use different powers Py for different channel realizations H subject to the average power constraint over all channel realizations Ey{Py] < P. The ergodic capacity under this as- sumption is given by a . a C= lthace™ ML ARB ng Meese + RAH] = max Fat max Latoes( i+ (10.17) PuckulPale? Led, esha S for y; = ?Pu/a?. The short-term power constraint gives rise to a water- ig in space across the antennas, whereas the long-term power constraint allows for a two-dimensional water-filling across both space and time; this is similar to the frequency-time water-filling associated with the capacity of a time-varying frequency-selective fading channel. The brack- ted terms in the second lines of (10.16) and (10.17) are functions of the singular values (0;} of the matrix H.. Thus the expectation with respect to H in these expressions will be based ‘on the distribution of these singular values. This distribution for H a ZMSW matrix, as well as for other types of matrices, can be found in [12, Sec. 2.1].‘CHANNEL UNKNOWN AT TRANSMITTER: ERGODIC CAPACITY “AND CAPACITY WITH OUTAGE Consider now a time-varying channel with random matrix H that is known at the receiver but not the transmitter. The transmitter assumes a ZMSW distribution for H. The two rele- vant capacity definitions in this case are ergodic capacity and capacity with outage. Ergodic capacity defines the maximum rate, averaged over all channel realizations, that can be trans- mitted over the channel for a transmission strategy based only on the distribution of H. This leads to the transmitter optimization problem — that is, finding the optimum input covariance matrix to maximize ergodic capacity subject to the transmit power constraint. Mathemati- cally, the problem is one of characterizing the optimum R, to maximize C=, max EnlBlog2det( ly, +HR,H")); (10.18) here the expectation is with respect to the distribution on the channel matrix H, which for the ZMSW model i: |. zero-mean circularly symmetric and of unit variance. As in the case of static channels, the optimum input covariance matrix that maximizes ergodic capacity for the ZMSW model is the scaled identity matrix Rx = (p/M,)Iw,; that is, the transmit power is divided equally among all the transmit antennas, and independent ‘symbols are sent over the different antennas. Thus the ergodic capacity is given by C=E [? toad, + mn" 10.19) where, as when CSIT is available, the expectation with respect to His based on the distribu- tion of its singular values. The capacity of the static channel grows as M = min(M,, M,) for ‘M large, so this will also be true of the ergodic capacity because it simply averages the static channel capacity. Expressions for the growth rate constant can be found in [4; 25]. When the channel is not ZMSW, capacity depends on the distribution of the singular values for the random channel matrix; these distributions and the resulting ergodic capacity in this more ‘general setting are studied in (11.no vansmlter CSIT win transmitter CST. 5 10 3s 0 15 20 ‘Average Received SNR Figure 10.4: Ergodic capacity of 4x 4 MIMO channel. The ergodic capacity of a4 x 4 MIMO system with i.i.d. complex Gaussian channel gains is shown in Figure 10.4. This figure shows capacity with both transmitter and receiver CSI and with receiver CSI only. There is litte difference between the two plots and the difference decreases with SNR, which is also the case for an SISO channel. Comparing the capacity of this channel to that of the SISO fading channel shown in Figure 4.7, we see that the MIMO ergodic capacity is four times larger than the SISO ergodic capacity, which is just as expected since min(M,,M,) = 4. hy, Ce) os WW x ” vy, zm hua, Figure 10.7: MIMO channel with beamforming.10.4 MIMO Diversity Gain: Beamforming The multiple antennas at the transmitter and receiver can be used to obtain array and diver sity gain (defined in Section 7.1) instead of capacity gain. In this setting the same symbol — weighted by a complex scale factor — is sent over each transmit antenna, so that the input covariance matrix has unit rank. This scheme is also referred to as MIMO beamforming. A beamforming strategy corresponds to the precoding and shaping matrices (described in Section 10.2) being just column vectors: V = v and U = u, as shown in Figure 10.7. As in- dicated in the figure, the transmit symbol x is sent over the ith antenna with weight vj. On the receive side, the signal received on the ith antenna is weighted by u?. Both transmit and receive weight vectors are normalized so that |lu|] = ||v|]| = 1. The resulting received signal is given by y =u" Hvx +u"n, (10.20) where if m = (1),...,2y,) has i.i.d. elements then the statistics of un are the same as the 's for each of these element: Beamforming provides diversity and array gain via coherent combining of the multiple signal paths. Channel knowledge at the rec is required for coherent combining. The performance gain then depends on whether or not the channel is known at the transmitter. When the channel matrix H is known, the received SNR is optimized by choos- ing u and vas the principal left and right singular vectors of the channel matrix H. That is, for o1 = omax the maximum singular value of H, u and v are (respectively) the first columns of U and V. The corresponding received SNR can be shown to equal y = 02,0, where Omax is the largest singular value of H [6; 33]. The resulting capacity is C = B log2(1 + 02,,). st rer is sumed, since t + Unfortunately, beamforming is also used in the smart antenna context of Section 10.8 to describe adjustment of the transmit or receive antenna directivity in a given direction,corresponding to the capacity of a SISO channel with channel power gain 02,,. For Ha ZMSW matrix, it can be shown [6, Chap. 5.4.4] that the array gain of beamforming diversity is between max(M;, M,) and M,M, and that the diversi When the channel is not known to the transn s described in Section 7.3.2 can be used to extract an array gain of M, and the maximum di- versity gain of 2M, [6, Chap. 5.4.3]. For M, > 2, full diversity gain can also be obtained 4 reduced capacity relative to the capscity-achicving tranami presiding and rebciver shap- ing matrices, the demodulation complexity with beamforming is on the order of |4’| instead of || antenna s An even simpler strategy is to use MRC at either the transmitter or receiver and lection on the other end: this was analyzed in [34].