EURASIP Journal on Applied Signal Processing 2002:3, 330–339

c 2002 Hindawi Publishing Corporation

Channel Estimation and Optimal Power Allocation

for a Multiple-Antenna OFDM System

Tai-Lai Tung
GlobespanVirata, Inc., 5171 California Avenue, Suite 200, Irvine, CA 92612, USA
Email: [email protected]

Kung Yao
Electrical Engineering Department, 68-113 Engineering IV Building, University of California,
Los Angeles, CA 90095-1594, USA
Email: [email protected]

Received 6 June 2001 and in revised form 24 January 2002

We propose combining channel estimation and optimal power allocation approaches for a multiple-antenna orthogonal frequency
division multiplexing (OFDM) system in high-speed transmission applications. We develop a least-square channel estimation
approach, derive the performance bound of the estimator, and investigate the optimal training sequences for initial channel ac-
quisition. Based on the channel estimates, the optimal power allocation solution which maximizes the bandwidth efficiency is
derived under power and quality of service (QoS) (symbol error rate) constraints. It is shown that combining channel tracking
and adaptive power allocation can dramatically enhance the outage capacity of an OFDM multiple-antenna system when severe
fading occurs.

Keywords and phrases: OFDM, multiple-antenna, channel estimation, optimal power allocation, capacity, optimal training se-
quences, MSE bounds.

1. INTRODUCTION that combining the optimal power allocation method with a

closed-loop channel state information updating mechanism
Orthogonal frequency division multiplexing (OFDM) tech- can dramatically increase the system outage capacity in time-
nique has recently gained increasing interest as an effective varying environment.
alternative to conventional single tone modulation meth- The organization of the paper is as follows. In Section 2,
ods for high-rate communications systems [1, 2, 3]. The we first propose least-squares channel estimator. To ana-
main advantage of OFDM is allowing transmission over a lyze the performance of the channel estimator, the mean
frequency-selective fading channel using a low cost receiver. square error (MSE) bound is derived in Section 3. The op-
On the other hand, current research has shown that intel- timal training sequences which can achieve minimum MSE
ligent use of antenna arrays at both receiver and transmit- channel acquisition are designed in Section 4. The tracking
ter side can significantly increase the bandwidth efficiency as and BER performance in time-varying channels are shown
well as improving the transmission reliability [4]. More re- in Section 5. In Sections 6 and 7, we introduce the capacity
cent work then show that combining OFDM technique with of the OFDM multiple-antenna system, and show that the
multiple-antenna architecture can provide high-rate high capacity of the system can be achieved by performing joint
performance transmission [5, 6, 7, 8]. optimal power allocation in spatial and frequency domains.
In this paper, we first propose least-square (LS) channel In Section 8, we show the bandwidth efficiency of the sys-
estimation approach for OFDM multiple-antennas systems. tem using adaptive QAM modulation where the power and
Based on the channel estimates, the bandwidth efficiency symbol error rate are constrained. The numerical results of
is then maximized by performing optimal power allocation the optimal power and bit allocation are shown in Section 9.
over multiple antennas and subbands where some practi- The improvement of system outage capacity in time-varying
cal constraints such as specified quality of service (QoS) re- environment through optimal power allocation with channel
quirements in terms of symbol error probability and lim- tracking are shown in Section 10. Finally, Section 11 contains
ited transmit power constraints are considered. It is shown the conclusions.
Channel Estimation and Optimal Power Allocation for a Multiple-Antenna OFDM System 331

2. CHANNEL ESTIMATION then

 (p)   (p) 
In an OFDM system, the frequency response at the kth tone H h
 0.   0. 
corresponding to the channel from the pth transmit antenna ≡  .  = F  .. 
  
(p) (p)
to the qth receive antenna is given by ([5, 7])
gK ×1 .  ≡ FhL×1 . (6)
(p) (p)
HK −1 hL−1
(qp)

L−1
(qp)
Hk = h
WKk
, (1) The channel can be modeled as an FIR filter with L taps and

=0 sampling interval Ts . Combining (4) and (6), we obtain

where h

(qp)
is the
th channel impulse of

Nt
y= X(p) Fh(p) + ζ = Ah + ζ, (7)
p=1
 (qp)

(qp) H
h(qp) = h0 , . . . , hL−1 , (2)
where
which is the channel impulse response from the pth trans- AK ×LNt = X(1) F, . . . , X(Nt ) F ,
mit antenna to the qth receive antenna, WK = e− j(2π/K) ,   (8)
and K is the total number of spectral channels (tones) in H H H
hLNt ×1 = h(1) , . . . , h(Nt ) .
the OFDM system, Ts is the sampling interval of the sys-
tem, and 1/Ts is the entire channel bandwidth of the OFDM By the LS method, the multiple channels can be estimated
system. T f = (K + v)Ts is the OFDM frame length where by
v is the number of cyclic prefix symbols. We usually have
K
v for bandwidth efficiency and let v ≥ L − 1 to avoid ĥ = A+ y, (9)
interblock interference; therefore, T f
LTs , which also in-
duces flat frequency response in each subchannel. Given Nt where the pseudo-inverse of matrix A+ = (AH A)−1 AH if K ≥
transmit antenna and Nr receive antenna, the received sig- LNt . Because rank(A) = min(K, KNt , LNt ) = min(K, LNt ),
nal in the kth subband at the qth receive antenna can be the necessary and sufficient condition to have unique LS
expressed as channel-estimate is K ≥ LNt . The advantage of this LS ap-
proach is its low-complexity and it requires no a priori sta-
(q)

Nt
(qp) (p) (q) tistical knowledge about the channel and the noise. In the
yk = Hk xk + ζk , (3) training mode, since the training sequences have been spec-
p=1
ified, (AH A)−1 AH can be precomputed and stored such that
(p) the operation of the initial channel acquisition can be further
for q = 1, . . . , Nr , k = 0, . . . , K − 1, where xk is the trans- speeded up.
(q)
mit signal at the pth transmit antenna, and ζk is the noise
at the qth receive antenna. For the sake of conciseness, the
index q for different antennas is omitted from Hk , h
,
(qp) (qp) 3. MSE BOUND OF LS ESTIMATOR
(q) (q)
yk , ζk . Based on (3), we propose a channel estimation ap- Combining (7) and (9), we obtain
proach by forming a matrix equation with all the unknown
channel parameters in one vector, and estimate that vector ĥ = h + A+ ζ, (10)
by the least-square (LS) method. Given K subchannels, the
received vector can be expressed as which indicates that the estimate ĥ is the combination of the
true h plus a term affected by the estimation error. Since


Nt E{ĥ} = h + A+ E{ı} = h, ĥ is an unbiased estimator. The
y= X(p) g(p) + ζ, (4) MSE of estimation is given by
p=1
1 
2
where y = [y0 , . . . , yK −1 ]H , ζ = [ζ0 , . . . , ζK −1 ]H , and X(p) is a MSE = E ĥ − h
(p)
LNt
diagonal matrix with the kth diagonal element as xk , and 1 
2
g(p) is a K × 1 vector in which its kth element is Hk . Define
(p) = E A+ ı
LNt
(11)
1   H 
  = Tr E A+ ζζ H A+
1 1 ··· 1 LNt
 
1 WK1 ··· WK(L−1)  σn2  −1 
  = Tr AH A
F = . .. .. ..  , (5) ,
 .. .  LNt
 . . 
1 WKK −1 · · · WK(K −1)(L−1) K ×L
where the superscript H denotes the Hermitian operation,
Tr{·} denotes the trace of a square matrix and E{ζζ H } =
332 EURASIP Journal on Applied Signal Processing

σn2 IK . Define B = (AH A)−1 , then BLNt ×LNt becomes Nt = 3, Nr = 3, L = 8

100
 H H
−1
FH X(1) X(1) F · · · FH X(1) X(Nt ) F
 H (2) H (1) 
F X X F · · · FH X(2) X(Nt ) F 
H
  10−1

MSE of channel estimator

 .. .. ..  . (12)
 . 
 . . 
H H
FH X(Nt ) X(1) F · · · FH X(Nt ) X(Nt ) F
10−2
(p) 
Suppose xk s, for p = 1, . . . , Nt , all n and k, are constant
(p) H
modulus, and |xk |2 = 1, then FH X(p) X(p) F = KI. Denote
the (i, i)th component of B as bi , then 10−3


σ2

LNt
σ2 LNt
MSE = n bi ≥ n LNt LN
t
bi , (13) 10−4
LNt i=1 LNt i=1 0 1 2 3 4 5 6 7 8 9 10
SNR (dB)
with equality when b1 = b2 = · · · = bLNt . By Hardamard’s K = 32, Gaussian seq. K = 128, Gaussian seq.
inequality, K = 32, NRZ seq. K = 128, NRZ seq.
K = 32, Opt. training seq. K = 128, Opt. training seq.

LNt
 
bi ≥ det(B); K LNt ≥ det B−1 = [det(B)]−1 , (14) Figure 1: Channel acquisition error induced by different training
i=1 sequences.
with equality if and only if B is diagonal. Finally, combining
(13) and (14), we obtain sequences provide the minimum mean square channel ac-
quisition error as compared to randomly generated se-
σn2
MSE ≥ , (15) quences.
K
with equality if and only if B is a diagonal matrix and all the 5. TRACKING PERFORMANCE IN TIME-VARYING
diagonal components bi , i = 1, . . . , LNt are equal. Note that CHANNELS
(15) corroborates the results in [5] where the case Nt = 2 was
considered. The merits of our derivation is that we can ob- In the simulations, channels with three-ray and a Doppler
tain the MSE bound of the channel estimator without com- frequency of 10 Hz and 100 Hz are used to characterize the
puting the inverse of the matrix analytically in (11), which is time-varying multipath propagation environments. Each re-
a computationally-intractable work when Nt > 2. solvable multipath is independent of each other and char-
acterized by Ricean fading where the Rayleigh fading part is
generated using Jake’s model [8]. Two transmitter antennas
4. OPTIMUM TRAINING SIGNAL (Nt = 2) with multiple receiver antennas (Nr = 2, 3, 4) are
used for diversity. The parameters for the simulated OFDM
According to (15), in order to make the channel estimator system are as follows. The entire channel bandwidth fs =
achieve its best performance at the channel acquisition mode, 240 kHz is divided into K = 128 subchannels. QAM signal-
the training signals from multiple antennas should satisfy ing is applied. To make the tones orthogonal to each other,
 the OFDM frame duration is K/ fs = 533 µs. An additional
H (r) H 0L×L , r = s, 83 µs guard interval is used to provide protection from in-
F X X F=
(s)
(16)
cIL×L , r = s, terblock interference. The channel multipath delay spread is
assumed (L − 1)/ fs = 83 µs and L = 3. This results in a to-
where c is a constant, 1 ≤ r, s ≤ Nt . Therefore, B is guaranteed tal block length T f = 616 µs and a subcarrier symbol rate
to be a diagonal matrix with the same diagonal components. 1.623 kBaud. In the data transmission mode, the channel es-
To satisfy the constraints, the optimal training sequence can timates of the previous OFDM block are used to decode the
be designed as current OFDM symbols. The performance of the system is
(p) (p−2) 
measured by bit error rate (BER) and the channel estimation
xk = (−1)k/2 xk(1) , p = 1, . . . , Nt , (17) mean square error (MSE), each averaged over 2500 OFDM
blocks. Figure 2 shows BER performance versus SNR using
where |xk(1) | = c for all k. The verification that the de- beamforming plus successive interference cancellation (de-
signed training sequences can satisfy the constraints is given noted as QR beamforming [7]) with path gain ordering for
in [8]. It is concluded that given Nt transmit antennas, Nt = 2, Nr = 2, 3, 4, and fd = 10 Hz and 100 Hz. Suppose
the channel estimator can achieve MSE lower bound, as that the average received SNR = 8 dB, then for fd = 100 Hz,
long as K ≥ 2Nt −1 L. Figure 1 shows that optimum training the resulting BER is around 10−2 for Nr = 3 and approaches
Channel Estimation and Optimal Power Allocation for a Multiple-Antenna OFDM System 333

f d = 10 Hz, QR + Ordering f d = 100 Hz, QR + Ordering

100 100

10−1 10−1

10−2 10−2
BER

BER
10−3 10−3

10−4 10−4

10−5 10−5

10−6 10−6
0 2 4 6 8 10 0 2 4 6 8 10
SNR (dB) SNR (dB)
(Nt , Nr ) = (2, 2) (Nr , Nr ) = (2, 2)
(Nt , Nr ) = (2, 3) (Nt , Nr ) = (2, 3)
(Nt , Nr ) = (2, 4) (Nt , Nr ) = (2, 4)

Figure 2: BER versus SNR using beamforming with successive interference cancellation (QR beamforming) and path gain ordering.

f d = 10 Hz, QR + Ordering f d = 100 Hz, QR + Ordering

10−1 10−1
MSE of channel estimation

MSE of channel estimation

10−2 10−2

10−3 10−3

10−4 10−4
0 2 4 6 8 10 0 2 4 6 8 10
SNR (dB) SNR (dB)
(Nt , Nr ) = (2, 2) (Nr , Nr ) = (2, 2)
(Nt , Nr ) = (2, 3) (Nt , Nr ) = (2, 3)
(Nt , Nr ) = (2, 4) (Nt , Nr ) = (2, 4)

Figure 3: Channel estimation error versus SNR using QR beamforming and path gain ordering.

10−3 for Nr = 4. For fd = 10 Hz, the BER reduces to 0.5 × 10−3 antennas are used, more spatial diversity can result in better
and 10−6 , respectively. This shows that the BER reduces when chance to successfully detect the data.
more receive antennas are used, especially under low Doppler
frequency conditions. Figure 3 shows the channel estimation 6. CAPACITY OF OFDM MULTIPLE-ANTENNA
error for Nt = 2, Nr = 2, 3, 4, and fd = 10 Hz and 100 Hz SYSTEM
when QR beamforming with path gain ordering is applied.
We observe that the channel estimation performance im- In multiple-antenna OFDM system, the demodulated sig-
proves when Nr increases, SNR increases and fd decreases. nal in the kth subband at the qth receive antenna is de-
(q)  (qp) (p) (q)
It is observed that when the Doppler frequency is small, the noted by yk = Np=t 1 Hk xk + ζk , given Nt transmit an-
rate that channel fades is slow; thus, better channel tracking tennas and Nr receive antennas, the received vector in the
performance can be attained. In addition, when more receive kth subband is given as yk = Hk xk + ık , k = 0, . . . , K − 1,
334 EURASIP Journal on Applied Signal Processing

where yk = [yk(1) , . . . , yk(Nr ) ]H , xk = [xk(1) , . . . , xk(Nt ) ]H , ık = where P is total power allocated to {xk }k=0,K −1 . Since CP is
[ζk(1) , . . . , ζk(Nr ) ]H , and the (q, p)th entry of Hk is Hk . The
(qp) applied, the total transmission power is indeed ((K + v/K)P)
information capacity of the kth subchannel is given as where v is the length of the CP. To solve (24), we first de-
compose Hk by singular value decomposition (SVD), Hk =
 
Ck = max I xk ; yk Uk Sk VH
k
, where Uk UH k
= INr ×Nr , Vk VH k
= INt ×Nt , the (p, p)th
 (p)
f (xk ): Exk ≤P (p)
entry of Sk is sk , k = 0, . . . , K − 1, p = 1, . . . , Nt . Thus, Ck in
    (18)
= max h yk − h ζk . (21) can be expressed as
 (p)
f (xk ): Exk ≤P   
Sk VH D V SH
k k k k
When the noise ζk is Gaussianly distributed, its entropy is Ck = log2 det I+ . (25)
nk
given by
  
Define Wk = VH D V , Ck in (25) is maximized when Wk
k k k
1
h ζk = log2 (2πe)Nr det(Nk , (19) is a diagonal matrix (see the appendix). Since tr(Wk ) =
2 tr(VH D V ) = tr(Vk VH D ) = tr(Dk ), the optimization
k k k k k
where Nk = E{ζk ζkH }. If the noise is independent of the input, problem in (24) is reduced to
the capacity is obtained by maximizing the entropy of the re-  (p) 2 (p) 
ceived signals which occurs when received signals and input

K −1

Nt sk wk
signals have Gaussian distributions. Thus, C= max log2 1+ ,
W0 ,...,WK −1
k=0 p=1
nk
(26)
  1  

K −1

Nt
(p)
max

h yk = log2 (2πe)Nr det Yk , (20) s.t. wk = P,
f (xk ):
(p)
Exk ≤P 2 k=0 p=1

where Yk = E{yk ykH } = Hk Dk HH

k
+ Nk , Dk = E{xk xkH }, Nk = where the (p, p)th entry of Wk is denoted by wk . The ex-
(p)

E{ζk ζk }. Combining (19) and (20), yields

H
tended 2D spatial-spectral water filling solution is then ob-
   tained as
det Hk Dk HH
k
+ Nk  +
Ck = log2   (bit/s/Hz). (21)
det Nk nk 
= v −
(p)
wk , (27)
(p)2
sk
If uniform power allocation is used (i.e., Dk = P/(KNt )) and
the Gaussian noise is spatially white with equal power at dif-
where (•)+ is equal to the argument • if • is positive and set
ferent receive antennas (i.e., Nk = nk I), (21) reduces to
to zero if • is negative. v is chosen so that
  
P  +
Ck = log2 det I + H HH , (22)
−1

KNt nk k k K Nt
v − nk  = P. (28)
(p) 2
k=0 p=1 sk
which is consistent to equation (4) in [4]. Assuming that
the OFDM frames and carriers are synchronized, and cyclic
prefix (CP) is applied, the K subbands in each transmit- Note that making Wk diagonal implies a precoloring process-
ter can be considered as K parallel channels and the overall ing at the transmitter side and a beamforming processing at
channel capacity of the OFDM multiple-antenna system is the receiver side. That is, the transmit signal vector is precol-
given by ored by setting xk = Vk wk . Since the input data stream are
spatially and temporally white, Wk = E{wk wkH } = I attains a


K −1 diagonal form. On the other hand, the beamforming weight
C= Ck vector which steers the beam to the pth transmit antenna and
k=0 places nulls to the others is given by the pth column of Uk .
   (23)


K −1
1 Referring now to [6], a spatio-temporal vector-coding com-
= log2 det I + Hk Dk HH
k .
k=0
nk munication structure is suggested as a means for achieving
capacity. Here, we tackle the capacity-maximization prob-
lem in opposite direction. Upon maximizing the channel ca-
7. OPTIMAL POWER ALLOCATION pacity using Hadamard’s inequality, we show that a spatio-
It is observed that, given the channel transfer functions Hk , temporal vector coding processing (a precoloring processing)
the channel capacity is a function of Dk , whose diagonal el- is required to achieve the capacity.
ements correspond to the allocated power. In order to max-
imize the channel capacity, an optimization problem is for- 8. OPTIMAL BIT ALLOCATION SUBJECT TO POWER
mulated as AND LOS (SER) CONSTRAINTS


K −1

K −1
  In this section, we apply the above optimal power allocation
C = max Ck , s.t. tr Dk = P, (24) method to a practical adaptive QAM modulation scheme for
D0 ,...,DK −1
k0 k=0
Channel Estimation and Optimal Power Allocation for a Multiple-Antenna OFDM System 335

each subband. The symbol error rate (SER) of a 2b − QAM 14

signaling scheme is given in [2] by
12
! 
3Po |H |2
Pe = MQ   , (29) 10

Capacity (Bits/sec/Hz)
N 2b − 1
8
where M is average number of neighbors for symbols in the
QAM constellation. Po , H, N, and 2b are the transmit power, 6
path gain, noise power, and QAM size, respectively. For M =
4, denote γ = Q−1 (Pe /4), then the resultant bits per symbol 4
can be expressed as
2
 
3 Po|H |2
b = log2 1+ . (30) 0
γ2 N 0 2 4 6 8 10 12 14
SNR (dB)
Accordingly, for the multiple antennas OFDM system, the Opti. (4, 4) Opti. (2, 2)
optimal allocated bit in the kth subcarrier for the pth spa- Unif. (4, 4) Unif. (2, 2)
tial channel is given as Opti. (3, 3) Opti. (1, 1)
Unif. (3, 3) Unif. (1, 1)
 
(p) (p) 2
3 w s
= log2 1 + 2 k k . Figure 4: Capacities with optimal power allocation (opti.) and uni-
b pk (31)
γ nk form power allocation (unif.) versus SNR for an OFDM (Nt , Nr )
multiple-antenna system.
Thus, given a total allocated power P, subband bandwidth
W, and the required SER, the total transmission rate R is
maximized by 9. NUMERICAL RESULTS ON OPTIMAL POWER
AND BIT ALLOCATION


K 
−1

Nt

K −1

Nt In the simulations, sampling frequency f0 is normalized to
K (p)
R = max W b pk , s.t. wk = P. 1, the number of subchannels K = 128, cyclic prefix length
(p)
wk K +v k=0 p=1 k=0 p=1
v = 4, OFDM frame duration is (K + v)/ f0 , the maximum
(32) multipath delay spread is (v − 1)/ f0 . The channel from the
pth transmit antenna to the qth receive antenna √
is modeled
The optimal allocated power is then solved as −1θ (pq)
by two rays as h (t) = α (δ(t)+A e
(pq) (pq) (pq) δ(t − τ (pq) )),
 + where τ (pq) ∈ {1/ f0 , 2/ f0 , 3/ f0 }, α(pq) is drawn from a uni-
γ 2 nk  form distributions [0, 1], and under severe frequency se-
= v −
(p)
wk , (33)
(p) 2 lective fading conditions of A(pq) → 1, and θ (pq) → π
3sk
[11]. Note that the total allocated power is defined by P =
K −1 Nt (p) 2
and the corresponding optimal allocated bit is k=0 p=1 E{|xk | }, and the average SNR ratio at the re-
ceiver is defined by
   +  
  (p) 2 (qp) (p) 2 
s 2

K −1

Nt E |Hk x |
b pk = log2 1 +
3  k  v − γ nk   , (34) SNR = "  k #. (36)
γ2 nk (p) 2 K −1
3sk k=0 p=1 Nr k =0 nk

where v is chosen so that Figure 4 shows the capacities of an OFDM (Nt , Nr ) multiple-
 + antenna system using optimal (solid curves using (27) in
(p)


K −1

Nt 2 (26)) and uniform (dash curves using wk = P/(KNt ) in
v − γ nk  = P. (35)
(p) 2 (26)) power allocations. It is observed that a significant in-
k=0 p=1 3sk crease in capacity can be obtained when the optimal power
allocation is applied. It also shows that the channel capacity
Note that the bandwidth efficiency defined by R/KW =
K −1 Nt increases with the total allocated power and the number of
k=0 p=1 b pk /(K + v), where KW is the total bandwidth. antennas (Nt , Nr ). To illustrate the mechanism of optimal
Also note that the optimal bit provided by (31) need not be bit allocation using the adaptive QAM modulation, Figure 5
an integer. To satisfy the integer constraints, an iterative al- shows both the channel gain and the allocated bits in each
gorithm proposed in [3] can be used. To reduce the gap r, subchannel for a specific Nt = Nr = 2 and K = 128 example.
many other practical coding techniques such as trellis coding It is of interest to observe that given fixed transmit power,
[9] and coset coding [10] can also be applied. more power is allocated to higher gain subchannels such that
336 EURASIP Journal on Applied Signal Processing

−15 −15

−20 −20

−25 −25

|sk |2 (dB)
|sk |2 (dB)

−30 −30
(1)

(2)
−35 −35

−40 −40

−45 −45
0 50 100 150 0 50 100 150
kth subband of first spatial channel kth subband of second spatial channel

10 10

8 8

6 6
b1k b2k
4 4

2 2

0 0
0 50 100 150 0 50 100 150
kth subband of first spatial channel kth subband of second spatial channel
SNR = 30 dB SNR = 20 dB

Figure 5: Channel state information and optimal bit allocation for OFDM system with 2 transmit and 2 receive antennas with K = 128 and
SER = 10−5 .

more bits can be efficiently transmitted. Figure 6 shows the To evaluate the effect of CSI updating on outage capacity,
bandwidth efficiency R/KW obtained by different power al- we assume the estimated CSI at the receivers are sent back
location approaches under different QoS (SER) and received to the transmitters every Td seconds, where the loop delay is
power (SNR) conditions. These plots show that under low included which accounts for the two-way signal propagation
SNR and high QoS conditions, optimal power allocation has delay and the time delay involved in generating and execut-
significant advantage over that of uniform power allocation. ing the updating command. Based on this mechanism, the
mean square error of the channel tracking error is given by
E{|h(t) − h(t − Td )|2 }, where h(t) is the path gain at time t.
10. CAPACITY IMPROVEMENT
Since each of the resolvable multipath components is actu-
When the system capacity is observed as a random variable ally a combination of many unresolvable paths with vary-
in a time-varying environment, its complementary cumula- ing amplitudes, phases, and direction of arrivals. The path
tive distribution functions (ccdfs) can be used to show what gain can be modeled as a zero mean complex-valued colored
fraction of time or probability, a specified capacity can be Gaussian random process with an auto correlation function
attained [4]. For example, the capacity that can be attained J0 (2π fd τ), where fd is the maximum Doppler shift, and τ
with a 99% probability results in a 1% probability of outage. is the time difference. Thus, E{h(t)h∗ (t − τ)} = J0 (2π fd τ).
In order to reduce the outage probability and increase the Accordingly, E{|h(t) − h(t − Td )|2 } = 2(1 − J0 (2π fd Td )).
channel capacity in a time-varying environment, we adap- Since h(t) is zero mean complex Gaussian, h(t) − h(t − Td )
tively performing optimal power allocation among multiple- is also zero mean complex Gaussian. Consequently, the vari-
antenna and multiple-subband based on updated channel ance of h(t) − h(t − Td ), denoted as σh2 , is computed as
state information (CSI). In this application, it is important to 2(1 − J0 (2π fd Td )). Therefore, we model the tracked chan-
keep the updating rate high enough so that the time-varying nel as ĥ(t) = h(t) + σh (Td )n(t), where h(t) is the true chan-
channel can be tracked. nel and σh (Td )n(t) is the channel tracking error. Both h(t)
Channel Estimation and Optimal Power Allocation for a Multiple-Antenna OFDM System 337

(Nt , Nr ) = (4, 4) Optimum power distribution (Nt , Nr ) = (4, 4) Uniform power distribution
14 14

12 12

10 10
R/KW (bit/symbol/Hz)

R/KW (bit/symbol/Hz)
8 8

6 6

4 4

2 2

0 0
0 5 10 15 0 5 10 15
SNR SNR
Capacity SER = 1e − 5 Capacity SER = 1e − 5
SER = 1e − 3 SER = 1e − 7 SER = 1e − 3 SER = 1e − 7

Figure 6: Bandwidth efficiency by optimal and uniform power allocations s.t. symbol error rate and power constraints.

101 (Nt , Nr ), SNR = 5 dB, K = 64, fd = 10 Hz

0.99

100
Probability (avg. capacity ¿ outage capacity)

0.98
MSE of channel tracking

0.97
10−1
0.96

10−2 0.95

0.94
10−3
0.93

10−4 0.92

0.91
10−5
0 5 10 15 20 25 30 35 40 45 0.9
OFDM fram (T f ) 4.5 5 5.5 6 6.5 7 7.5
fd = 100 Hz, simulation fd = 10 Hz, simulation Outage capacity (bit/s/Hz)
fd = 100 Hz, analysis fd = 10 Hz, analysis (3, 3), opt. power alloc.
(3, 3), opt. power alloc., Td = 20T f
Figure 7: Analytical and simulated MSE of channel tracking versus (3, 3), opt. power alloc., Td = 60T f
(3, 3), opt. power alloc., Td = 100T f
Td .
(3, 3), uni. power alloc.
(4, 4), opt. power alloc.
(4, 4), opt. power alloc., Td = 20T f
and n(t) are zero mean complex-valued Gaussian random (4, 4), opt. power alloc., Td = 60T f
variable’s with unit variance. Figure 7 shows both analytical (4, 4), opt. power alloc., Td = 100T f
and simulated MSE of channel tracking versus Td (in units (4, 4), uni. power alloc.
of T f ) for fd = 10 Hz and fd = 100 Hz, respectively. Fig-
ures 8 and 9 evaluate the effect of CSI updating on outage Figure 8: Capacity improvement using optimal power allocation
capacity. Figure 8 shows the effect of Td , number of anten- with various Td under fd = 10 Hz, SNR = 5 dB, K = 64, (Nr , Nt ) =
nas, and power allocation schemes on the outage capacity (3, 3) or (4.4).
for Td = 20T f , 60T f , 100T f under fd = 10 Hz, SNR = 5 dB,
K = 64, Nr = 3, and Nt = 3 condition. We observe that
optimal power allocation results in higher channel capacity more antennas are exploited and shorter updating period Td
(p)
compared to uniform power allocation (wk = P/(KNt )). are used. Similarly, Figure 9 shows that the outage capacity
We can also observe that the outage capacity increases when uses Td = 3T f , 6T f , and 9T f under fd = 100 Hz.
338 EURASIP Journal on Applied Signal Processing

$
(Nt , Nr ), SNR = 5 dB, K = 64, fd = 10 Hz diagonal elements, that is, det(K) ≤ i Kii with equality
0.99 if and only if the matrix is diagonal [12, page 233]. Since
Dk = E{xk xkH }, Wk = VH D V aH (I + Sk Wk SH )a = |a|2 +
Probability (avg. capacity > outage capacity)

0.98 k k k k
E{|aSVH x |2 } > 0, for all a = 0, (I + Sk Wk SH
k k k
) is positive
0.97
definite. By Hadamard’s inequality,
0.96   " #
det I + Sk Wk SH
k ≤ 1 + Sk Wk SH
k ii (A.1)
0.95 i

0.94
with equality if and only if (I + Sk Wk SH
k
) is diagonal, which
0.93 happens only when Wk is also diagonal since I and Sk are
0.92
diagonal.

0.91
ACKNOWLEDGMENTS
0.9
4.5 5 5.5 6 6.5 7 7.5 This work was partially supported by UC CoRE grant spon-
Outage capacity (bit/s/Hz) sored by ST Microelectronics, Inc. and NASA-Dryden grant
(3, 3), opt. power alloc. NCC2-374.
(3, 3), opt. power alloc., Td = 3T f
(3, 3), opt. power alloc., Td = 6T f
(3, 3), opt. power alloc., Td = 9T f REFERENCES
(3, 3), uni. power alloc. [1] J.-J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O.
(4, 4), opt. power alloc.
Börjesson, “On channel estimation in OFDM systems,” in
(4, 4), opt. power alloc., Td = 3T f Proc. 45th IEEE Vehicular Technology Conference, vol. 2, pp.
(4, 4), opt. power alloc., Td = 6T f 815–819, Chicago, Ill, USA, July 1995.
(4, 4), opt. power alloc., Td = 9T f
[2] A. R. S. Bahai and B. R. Saltzberg, Multi-carrier Digital
(4, 4), uni. power alloc.
Communications—Theory and Application of OFDM, Kluwer,
1999.
Figure 9: Capacity improvement using optimal power allocation [3] P. S. Chow, J. M. Cioffi, and J. A. C. Bingham, “A practi-
with various Td under fd = 100 Hz, SNR = 5 dB, K = 64, (Nr , Nt ) = cal discrete multitone transceiver loading algorithm for data
(3, 3) or (4, 4). transmission over spectrally shaped channels,” IEEE Trans. on
Communications, vol. 43, no. 2, pp. 773–775, 1995.
[4] G. J. Foschini and M. J. Gans, “On limits of wireless commu-
11. CONCLUSIONS nications in a fading environment when using multiple an-
tennas,” Wireless Personal Communications, vol. 6, no. 3, pp.
In the first part of the paper, we propose channel estima-
311–335, 1998.
tion, acquisition and tracking methods for OFDM multiple- [5] Y. Li, N. Seshadri, and S. Ariyavisitakul, “Channel estima-
antenna system. Then, we derive the lower bound of chan- tion for OFDM systems with transmitter diversity in mobile
nel estimator and design the optimal training sequences for wireless channels,” IEEE Journal on Selected Areas in Commu-
achieving minimum acquisition error. The tracking and BER nications, vol. 17, no. 3, pp. 461–471, 1999.
performance in time-varying channels is then evaluated. [6] G. G. Raleigh and J. M. Cioffi, “Spatio-temporal coding for
In the second part of the paper, we introduce the chan- wireless communication,” IEEE Trans. Communications, vol.
46, no. 3, pp. 357–366, 1998.
nel capacity of an OFDM multiple-antenna system, and the [7] T.-L. Tung and K. Yao, “Spatio-temporal parallel processing
optimal power allocation approach. Then, we apply opti- and complexity reduction for OFDM multiple antenna sys-
mal bit allocation method with adaptive QAM modulation tem,” in Proc. IEEE Sips2000, vol. 2, October 2000.
to maximize the bandwidth efficiency subject to fixed trans- [8] T.-L. Tung, Adaptive space-time signal processing for wireless
mit power and QoS constraints on the system. Simulation re- communication and sensor systems, Ph.D. thesis, University of
sults show that with a multiple-antenna architecture, OFDM California, Los Angeles, Calif, USA, 2001.
modulation, optimal power and bit allocation can dramati- [9] G. Ungerboeck, “Trellis-coded modulation with redundant
signal sets: parts I and II,” IEEE Commun. Mag., vol. 25, no. 2,
cally increase the transmission rate and bandwidth efficiency pp. 5–21, 1987.
when severe spatial and frequency selective fading occur. [10] G. D. Forney Jr., “Coset codes—part I: introduction and ge-
Finally, by combining optimal power allocation with real- ometrical classification,” IEEE Transactions on Information
time CSI updating, the channel outage capacity are shown to Theory, vol. 34, no. 5, pp. 1123–1151, 1988.
be dramatically improved by adaptive power allocation and [11] M. Rice, D. de Gaston, A. Davis, G. German, and C. Bet-
multiple-antenna architecture in the time-varying environ- twieser, “ARTM channel sounding results—an investigation
of frequency-selective fading on aeronautical telemetry chan-
ment.
nels,” in Proc. International Telemetering Conference, Las Ve-
gas, Nev, USA, October 1999.
APPENDIX [12] T. M. Cover and J. A. Thomas, Elements of Information Theory,
John Wiley and Sons, New York, USA, 1991.
Hadamard’s inequality states that the determinant of any
positive definite matrix K is less than the product of its
Channel Estimation and Optimal Power Allocation for a Multiple-Antenna OFDM System 339

Tai-Lai Tung was born in Taidung, Taiwan, Dr. Yao has extensive practical system experiences in digital and
in 1972. He received his B.S. degree in con- satellite communication engineering, DSP, radar system design,
trol engineering from National Chiao Tung link budget analysis, sensor system, and systolic and microphone
University, Hsinchu, Taiwan, in 1994 and array designs. He has worked or consulted for AT&T Bell Labo-
the M.S. and the Ph.D. degrees in electri- ratory, NCR, SRI, TRW, Linknet, Hughes Aircraft Company, and
cal engineering from the University of Cal- Raytheon Systems, and so on.
ifornia, Los Angeles (UCLA), in December
1997 and March 2001, respectively. From
1994 to 1996, he was an Ensign Instructor
at the Naval Communication and Electron-
ics School, Taiwan. Since 1998, he had been a Research Assistant at
UCLA. He is currently a Senior DSP Engineer in GlobespanVirata,
Irvine Office. His research interests include array signal processing,
adaptive signal processing, digital communications, and system op-
timization.

Kung Yao received his B.S.E. (Highest Hon-

ors), M.A., and Ph.D. degrees in electri-
cal engineering all from Princeton Univer-
sity, Princeton, NJ. During the summers, he
has worked at the Princeton-Penn Acceler-
ator, the Brookhaven Nat. Lab. and the Bell
Telephone Lab. in Murray Hill, NJ. He was
a NAS-NRC Post-Doctoral Research Fel-
low at the University of California, Berkeley.
Presently, he is a Professor in the Electrical
Engineering Department at UCLA. In 1969, he was a Visiting Assis-
tant Prof. at the Massachusetts Institute of Technology. In 1973–74,
he was a Visiting Associate Prof. at the Eindhoven Technical Uni-
versity. In 1985–1988, he served as an Assistant Dean of the School
of Engineering and Applied Science at UCLA. His research and pro-
fessional interests include sensor array system, digital communica-
tion theory and system, smart antenna and wireless radio system,
chaos communications and system theory, digital and array signal
and array processing, systolic and VLSI algorithms, architectures
and systems, radar system, and simulation. He has published over
250 journal and conference papers. Dr. Yao received the IEEE Sig-
nal Processing Society’s 1993 Senior Award in VLSI Signal Process-
ing. He is the co-editor of a two volume series of an IEEE Reprint
Book on “High Performance VLSI Signal Processing,” IEEE Press,
1997. He is also the co-author of the book, “Processing and Algo-
rithms in Communication and Radar Systems,” under preparation.
He has served as Program Chair, Secretary, and Chair of the IEEE
Inform. Theory Group in L.A. and served two terms as a member
of the Board of Governors of the IEEE Inform. Theory Group. He
was the Co-Chair of the 1981 Inter. Symposium on Inform. The-
ory and the Representative of the IT-BOG of the 1987 IEEE In-
form. Theory Workshop. Dr. Yao is a member of the VLSI Techni-
cal Committee of the IEEE Signal Processing Society. He was also
the Chair of the Technical Program and the General Chair of the
1990 and 1992 IEEE Workshop on VLSI Signal Processing. He has
served as an Associate Editor for Book Reviews of the IEEE Trans.
on Information Theory and was an Associate Editor of the jour-
nal, “Probability in the Engineering and Informational Sciences.”
In 1991–1993, he was the Associate Editor of VLSI Signal Process-
ing of the IEEE Trans. on Circuits and Systems. Since 1999, he is
an Associate Editor of the IEEE Communications Letters. He is a
member of the Editorial Board of Journal of VLSI Signal Process-
ing and Integration: the VLSI Journal. He is also a Guest Editor of
a Special Issue on “Applications of Chaos in Modern Communi-
cation Systems” of the IEEE Trans. on Circuits and Systems-Part
I to appear in December 2001. In recent years, he has organized
various sessions in array processing, wireless radio, and chaotic
communication for various conferences. He is a Fellow of IEEE.