You are on page 1of 6

Beamforming with Imperfect Channel

Knowledge: Performance Degradation


Analysis based on Perturbation
Theory
M. Navarro, A. Grant

Publication:
Vol.:
No.:
pp.:
Date:

3rd IEEE Sensor Array and Multichannel Signal Processing Workshop (SAM 2004)
Sitges (Spain). July 18-21, 2004

This publication has been included here just to facilitate downloads to those people asking for personal use copies.
This material may be published at copyrighted journals or conference proceedings, so personal use of the download
is required. In particular, publications from IEEE have to be downloaded according to the following IEEE note:
c
2007
IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material
for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or
lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

BEAMFORMING WITH IMPERFECT CHANNEL KNOWLEDGE: PERFORMANCE


DEGRADATION ANALYSIS BASED ON PERTURBATION THEORY
Monica Navarro
Centre Tecnologic de Telecomunicacions
de Catalunya
Nexus I building, Gran Capita, 2-4
08034 Barcelona. Spain
monica.navarro@cttc.es

ABSTRACT
We analyse the performance degradation of beamforming due to
imperfect channel knowledge based on eigenvector perturbation
theory. Easily computable approximations are given for the average signal to noise ratio loss when an estimate of the channel is
available. The performance degradation of imperfect beamforming is then later assessed for LMMSE channel estimation taking
into consideration power constraints.

Alex Grant
Institute for Telecommunications Research
University of South Australia
Mawson Lakes Blvd, Mawson Lakes SA 5095
Australia

cally quantify the performance loss due to imperfect channel state


information at the transmitter. Next (Section 3), the performance
degradation of imperfect beamforming is assessed for a practical
example taking into consideration power constraints.
2. PERFORMANCE DEGRADATION ANALYSIS
The system under investigation is described by the following algebraic expression

1. INTRODUCTION

r[l] = H[l]w[l]s[l] + n[l]

Antenna array schemes have been exploited to enhance communications in mobile wireless networks using different strategies.
Schemes range from space-time coding solutions, which provide
spatial and temporal diversity in a fading environment without requiring knowledge of the channel at the transmitter, to SVD (Singular Value Decomposition) transmition schemes, which can enhance performance based on coordinated transmission through the
antennas. An SVD communication system requires the transmitter
to have available accurate knowledge of the channel in the form
of its singular vectors. That involves a feedback channel from the
receiver to the transmitter that might be subject to errors itself. In
the particular case of beamforming, that is when the covariance
matrix of the signal has rank 1, the transmission scheme simplifies
to the principal singular vector.
Recently, researchers have drawn their attention to the problem of imperfect channel knowledge. Results indicates that exploiting even partial channel state information at the transmitter
may result in substantial improvements in capacity [14]. Several
authors [35] develop transmission schemes for optimising capacity when the channel feedback available to the transmitter is imperfect. In particular, how the SVD over the estimated channel affects
communications has been investigated by authors in [1, 38] from
the point of view of capacity.
In this paper the focus is on quantifying the performance degradation due to imperfect channel knowledge [9]. The cost function considered is the signal to noise ratio. Independent work by
[10, 11] have also addressed the problem of performance degradation analysis for an OFDM MIMO system. In [12] degradation is
given in terms of pairwise error probability (PEP).
The rest of the paper is organised as follows. First (Section
2), we make use of perturbation theory of eigenvectors to analyti-

where r[l] CNr is the received signal, H[l] CNr Nt is the


channel matrix, w CNt is the beamformer vector, selected by
the tranmitter, s[l] is the (scalar) transmitted signal and n[l]
CNr is Gaussian noise with zero mean and covariance matrix n2 I,
where Nt and Nr are the number of transmit and receive antennas,
respectively.

By definition E kwk2 = 1 and E |s[l]|2 = P , and hence

the transmit power E |s[l]|2 kwk2 = P , where P = 1 for simplicity.


Under the assumption of perfect channel knowledge, the optimum beamforming vector is the principal eigenvector of the correlation channel matrix HH H, where optimality is defined in terms
of maximization of the signal to noise ratio. The received signal to
noise ratio, denoted by , is
=

w[l]H H[l]H H[l]w[l]


n2

(1)

(2)

The performance degradation due to imperfect channel knowledge is evaluated in terms of relative SNR loss, defined as follows
L=

max
.
max

(3)

We assume the following channel estimate is known


=H+E
H

(4)

where H is the true channel and E is the estimation error matrix or


perturbation matrix, assumed Normal distributed E N (0, K) [13,
Definition 2.2.1], where K denotes the covariance matrix of the error matrix.

HH
and vi ,
1 be the largest eigenvector of H
Theorem 3 Let v
i be the i-th eigenvector and associated eigenvalue of HH H.
Then the beamformer average performance degradation measured
in terms of signal to noise ratio loss can be approximated by

Perturbation theory relates the eigenvalues and eigenvectors of


b We shall apply
a given matrix A to a perturbed version of it A.
the following theorem from perturbation theory [14] [15, pg. 346]
modified for Hermitian matrices.
Theorem 1 Let A Cnn be Hermitian with simple eigenvalues
i 6= j i 6= j, where i, j = 1, . . . , n and F Cnn have
unitary 2-norm kFk2 = 1. For all 0 < < 1, the eigenvector of
b = A + F is approximated by,
the perturbed matrix A
vk () vk +

n
X

E [L]E

2
e
1

PNt

i=2

1 + e2

2
i +1+Nr e
(1 i )

PNt

i=2

(8)

2
i +1+Nr e
(1 i )2

(see Appendix for proof).

i=1,i6=k

vi Fvk
vi + O(2 )
k i

(5)

0.3

where unit-norm vectors vi are the eigenvectors of the unperturbed


matrix A and i are the associated eigenvalues.

approximation
imperfect beamforming

max
max

0.25
2 x 2 antennas
4 x 4 antennas
8 x 8 antennas

0.15

Next, variables are identified with the channel model parameters under examination. The estimated channel matrix (4) has
HH
= HH H + F, where F = HH E +
correlation matrix H
EH H + EH E. By definition, F = G where kGk2 = 1.
The following new theorem approximates the performance degradation in terms of receive SNR loss.

0.2

HH
and vi , i
1 be the largest eigenvector of H
Theorem 2 Let v
be the i-th eigenvector and associated eigenvalue of HH H. Then
1 can be approximated by
v
v1 +
1 r
v
1+

0.1

0.05

0
0

0.1

0.2

0.3

0.4

0.5

PNt

vi H Fv1
i=2 (1 i ) vi

PNt |vi H Fv1 |2

(6)

0.7

0.8

0.9

Fig. 1. Numerical evaluation of approximation (7) for different configurations of number of receive and transmit antennas.

i=2 (1 i )2

The beamformer performance degradation measured in terms of


signal to noise ratio loss ( 3) can be approximated by

PNt vi H Fv1 2
1 1i
i=2 1 i
L
P t vi H Fv1 2
1+ N
i=2 1 i

0.6

(7)

(see Appendix for proof).


For a relatively small perturbation, embedded in the norm of
F, (7) manifests the dependence of the received SNR loss on the
separation between consecutive eigenvalues and on the projection
of the perturbation matrix F into the eigenspace.
The difficulty in simplification of the approximation (7) lies
in the correlation between
the true

channel and the estimation error through the term vi H Fv1 . This correlation is unknown and
difficult to find. Based on the following assumptions

1. Channel matrix H has entries Hij N (0, 1) and E kHk2 =


1 in order to compare energy levels between the channel
and the estimation error.
2. The estimation error is assumed to be Gaussian with zero
mean and known covariance matrix RE = e2 I. Thus, estimation error matrix E has entries [E]ij N (0, e2 ). We
do not normalise E. We control the amount of error through
its variance e2 .
and considering expectations in (7) the problem may be simplified. Focusing then on the average performance degradation the
following new theorem characterises the average SNR loss.

To validate the approximations, numerical results are shown


in Fig. 1 and Fig. 2. Fig. 1 depicts the receive SNR loss (average
over 10000 channel realisations) as a function of in the interval
of interest 0 < < 1 for a Nr Nt multiple-antenna system.
In Fig. 2, the approximated average SNR loss (8) is compared to
the sample average SNR loss due to imperfect beamforming for a
4 4 MIMO system where the variance of the estimation error
2
2
matrix elements falls in 0 < E
< 2. Relating the variance E
to the ratio of matrix norms, we can interpret the results in terms
of energy ratio between the channel matrixhand thei error matrix.
For E N (0, e2 I) and H N (0, I), E
1

kEk2
F
kEk2
F

= e2 where

k.kF denotes Frobenious norm . From Fig. 2 we observe that (8)


approximates the loss within 1% for values of e2 < 0.05, ie., for
energy ratios between the error matrix and the channel matrix less
than 5%; within 10% for energy ratios up to 30% and about 15%
when the ratio increases up to 200%.
3. PERFORMANCE DEGRADATION FOR IMPERFECT
BEAMFORMING WITH LMMSE CHANNEL
ESTIMATION
In practice, estimation methods may be employed to obtain the
CSI. For example, the CSI may be obtained in form of an estimated channel matrix via LMMSE filtering. This estimator minimises the mean square error (MSE), in other words minimises the
H. The mean
error matrix norm , E [kEk] where E = H
square error depends on the quality of the channel corrupted with
1 kAk
F

tr(AH A)

0.7

10

0.6

10

imperfect
beamforming

90%

0.5

perfect
beamforming

10

Bit Error Rate

0.3

E
0.2

10

70%

50%

10

1%
5

10

20%

approximation
imperfect beamforming

0.1

0
0

10

10%

0.5

1.5

10

10

Eb /N0

2
E

12

14

16

18

20

Fig. 2. Numerical evaluation of approximation (8) for Nt =, Nr = 4

Fig. 3. BER performance for imperfect beamforming subject to power

antennas.

constraints.
0

n2

and the
additive white Gaussian noise via the noise variance
quality of the training via the power allocated to pilot symbols.
It is interesting then to evaluate performance degradation due to
imperfect beamforming subject to power constraints.
When only partial CSI is available at the transmitter, beam HH
reforming in the direction of the principal eigenvector of H
sults in an energy loss, since part of the energy is dispersed in other
directions of the space. If we increase training a better channel
estimate will be available at the transmitter allowing for a better
stearing of the beamformer. Then a larger portion of the energy
will be conveyed to the receiver. However, improving the quality of the training involves a consumption of energy itself. Where
should energy efforts be concentrated so the overall performance
is optimised?
Fig. 3 and Fig. 4 illustrate via a numerical example the performance degradation due to imperfect beamforming when the total
available power is fixed and shared between transmission of training and information symbols. A typical LMMSE filter is consid The percentage denotes the
ered to obtain the channel estimate H.
relative power allocated to training. Simulation results show that if
sufficiently accurate channel estimate is available to the transmitter, beamforming under imperfect CSI does not suffer considerable
losses. However, when taking into account energy constraints, results indicate that we should carefully evaluate where the energy
efforts should be concentrated so the overall performance is optmised. Allocating a large amount of power to training will accurately steer the transmission beam in the right direction, however
the amount of power left for data transmission may not be sufficient to achieve reliable communication, suffering from an effective energy per bit loss.
4. CONCLUSION
In this paper beamforming for imperfect channel state information has been investigated. Approximations for the performance
degradation in terms of signal to noise ratio are derived. Based on
perturbation theory for eigenvectors we obtained an approximation to the signal to noise ratio loss. The approximation requires
the knowledge of the projection of the perturbation matrix F over

0.5

E wH HH Hw (dB)

max
max

no beamforming

0.4

1.5

2.5

1%
10%
20%
50%
70%
90%

3.5

Eb /N0

10

12

14

16

18

20

Fig. 4. Receive signal level for imperfect beamforming with LMMSE


channel estimation for different power assigments to data and training.

the eigenspace of the true channel. The perturbation matrix F depends on the true channel and the estimation error. If F is the
identity matrix there will be no loss. In this case the approximation will only involve the inner product between the eigenvectors
of the true channel which is zero since the eigenvectors are orthogonal. The perturbation matrix F destroys the orthogonality.
It applies a linear transformation to the principal eigenvector Fv1
which is no longer guaranteed to be orthogonal to the rest of eigenvectors vi . The main difficulty is to determine the exact projection
of the transformed principal eigenvector onto the rest of channel
eigen-modes. This remains an open problem.
To simplify the problem we focus on the average performance
degradation and approximate the mean receive SNR loss. In this
case the approximation is only dependent on the eigenvalues and
the relative separation between them, the variance of the estimation error and the system dimensions via the number of receive antennas. The approximation (8) indicates that increasing the number of receive antennas augments the SNR loss. This was con-

firmed by the numerical results.


The performance degradation of imperfect beamforming has
also been assessed when a power constraint is taken into consideration. In a real scenario some form of channel estimation method
is utilised to obtain the CSI. We consider a training based channel
estimation solution and studied beamforming under certain power
constraints. When the total available power is fixed and shared
between transmission of pilot symbols and information data symbols, simulations indicates that we should carefully evaluate which
is the optimum power distribution. From BER performance simulated curves we see that the system suffers more from decreasing
the power available to transmit the information than from erroneously stearing the beamformer.
Results indicate that if a sufficiently accurate channel estimate
is available to the transmitter, beamforming under imperfect CSI
does not suffer considerable losses. This results relates to previous work [35] where it has been found that in terms of capacity
beamforming is close to optimal for relatively good feedback CSI
information.

A PPENDIX
Proof of Theorem 2 can be found in [9]. Here we would only note
that to derive Theorem 2, (9) has been normalised to unit norm,
otherwise a power mismatch would occur giving the approximated
eigenvector larger power than the true channel eigenvector. Thus,
P t vi H Fv1
v1 + N
i=2 1 i vi
1 r
(9)
v
PNt viH Fv1 2
1 + i=2 1 i
Proof: [Theorem 3] Define

PNt vi H Fv1 2
1 1i
i=2 1 i
e,
L
P t vi H Fv1 2
1+ N
i=2 1 i

(10)

We consider the expectation over the channel estimation error.


Hence, for a given channel what is the average SNR loss for different estimation error variances. The expectation of the quotient
is approximated as the quotient of expectations

E [A]
A
E

.
(11)
B
E [B]
One may justify this approximation observing that the denominator is only a normalisation factor, hence E [f (x)/k] = E [f (x)] /k 0
where k and k0 are constants. The problem is simplified to

PNt vi H Fv1 2
i
1

i=2

1
1
i

E [L]E E
PNt vi H Fv1 2
1 + i=2 1 i
2 i 1 i
PNt h H
1

i=2 E vi Fv1
|1 i |2
h
i

(12)
2
P t E |vi H Fv1 |
1+ N
2
i=2
| |
1

2
To calculate the expectation of vi H Fv1 , the term is expanded
first,

H
H

=
vi H Fv1 vi H Fv1
vi Fv1
=

vi H FB1 FH vi

where B1 = v1 v1 H . Note that FH = F, hence


=

B1 FH = FB1 F

HH E + EH H + EH E B1 HH E + EH H + EH E

=
+
+

HH EB1 HH E + HH EB1 EH H + HH EB1 EH E


EH HB1 HH E + EH HB1 EH H + EH HB1 EH E
EH EB1 HH E + EH EB1 EH H + EH EB1 EH E

To calculate the expectations we used properties of Normal distributed and Whishart distributed matrices [13, pg.60, pg.98] (the
reader is referred to [13] for details of these properties). The expectation is taken for each of the terms in the sum. The mean and
covariance matrix of the error matrix E are M = 0, K =
where denotes the Kronecker tensor product [16], Kij = ij ;
= INt and = e2 INr .
Applying [13, Property 3 of Theorem 2.3.5], where A = B1 HH ,
h
i
E HH EB1 HH E = e2 HH HB1
Applying [13, Property 2 of Theorem 2.3.5], where A = B1
h
i
E HH EB1 EH H = e2 tr(B1 )HH H
Applying [13, Property 1 of Theorem 2.3.6], where A = B1 , B =
I
h
i
h
i
E HH EB1 EH E = HH E EB1 EH E = 0
Applying [13, Property 1 of Theorem 2.3.5], where A = HB1 HH
h
i
E EH HB1 HH E = e2 tr(HB1 HH )INt
Applying [13, Property 3 of Theorem 2.3.5], where X = EH , A =
HB1
h
i
E EH HB1 EH H = e2 B1 HH H
Applying [13, Property 4 of Theorem 2.3.6], where A = HB1 , B =
I
h
i
E EH HB1 EH E = 0
Applying [13, Property 2 of Theorem 2.3.6], where A = I, B =
B1 H H
h
i
E EH EB1 HH E = 0
Applying [13, Property 3 of Theorem 2.3.6], where A = I, B =
B1
h
i
h
i
E EH EB1 EH H = E EH EB1 EH H = 0
Applying [13, Theorem 3.3.15], where S = EH E WNt (Nr , e2 INt )
according to [13, Definition 3.2.1], A = B1
h
i
E EH EB1 EH E =
= Nr e4 INt + Nr e4 tr(B1 )INt + Nr2 e4 B1
= e4 Nr ((1 + Nr )B1 + tr(B1 )It )
Then,

E vi H Fv1 = vi H E [FB1 F] vi

= vi H e2 HH HB1 vi + vi H e2 tr(B1 )HH H vi +

+vi H e2 tr(HB1 HH )INt vi + vi H e2 B1 HH H vi +

+vi H Nr e4 ((1 + Nr )B1 + tr(B1 )INt ) vi

Since the eigenvectors form an orthonormal basis of the channel subspace terms vi H e2 HH HB1 vi , vi H e2 B1 HH Hvi and
vi H Nr e4 ((1 + Nr )B1 vi are 0 (v1 H vi = 0 for i 6= 1). Using
the relation HH Hvi = i vi the expectation (13) yields,

E vi H Fv1
=
E

= e2 tr(B1 )i + e2 tr(HB1 HH ) + Nr e4 tr(B1 )


From tr(B1 ) = 1 and tr(HB1 H)H = 1 = 1 it follows

E vi H Fv1
= e2 (i + 1 + Nr e2 )

[10] A. Pascual-Iserte, A.I. Prez-Neira, and M. A. Lagunas, Performance degradation of an ofdm-mimo system with imperfect channel state information at the transmitter, in IST
Mobile and Wireless Communications Summit 2003, Aveiro,
Portugal, June 2003.
[11] A. Pascual-Iserte, A. I. Prez-Neira, and M. A. Lagunas, On
power allocation strategies for maximum signal to noise and
interference ratio in an ofdm-mimo system, Accepted for
publication at IEEE Trans. on Wireless Communications,
2003.

(13)

[12] J. Choi, Performance analysis for transmit antenna diversity


with/without channel information, IEEE Trans. Veh. Technol., 2002.

Substituting (13) into (12) the average signal to noise ratio is approximated by

[13] A. K. Gupta and D. K. Nagar, Matrix Variate Diatributions,


Chapman and Hall/CRC, 2000.

E [e
]E

1 + e2
1 + e2

[14] J. H. Wilkinson, The algebraic eigenvalue problem, Oxford


University Press, 1965.

PNt

2
(i +1+Nr e
)i
i=2
(1 i )2

PNt

i=2

2)
(i +1+Nr e
(1 i )2

(14)

Finally replacing (14) into (3) we obtained (8).


5. REFERENCES
[1] S. A. Jafar and A. Goldsmith, Transmitter optimization
and optimality of beamforming for multiple antenna systems
with imperfect beamforming, submitted to IEEE. Trans.
Wireless Commun.
[2] I. E. Telatar, Capacity of multi-antenna gaussian channels,
European Transactions on Telecommunications, vol. 10, no.
6, pp. 585595, Nov.-Dec. 1999.
[3] E. Visotsy and U. Madhow, Space-time transmit precoding
with imperfect feeback, IEEE Trans. Inform. Theory, vol.
47, no. 6, pp. 263239, Sept. 2001.
[4] A. Narula, M. J. Lopez, M. D. Trott, and G.W. Wornell, Efficient use of side information in multiple-antenna data transmission over fading channels, IEEE J. Selected Areas Commun., 1998.
[5] A. Narula, M. D. Trott, and G.W. Wornell, Performance
limits of coded diversity methods for transmitter antenna arrays, IEEE Trans. Inform. Theory, 1999.
[6] S. A. Jafar and A. Goldsmith, Transmitter optimization for
multiple antenna cellular systems, in IEEE International
Symposium on Information Theory, Lausanne, Switzerland,
June 2002, p. 50.
[7] S. Zhou and G. B. Giannakis, Optimal transmitter eigenbeamforming and space-time block coding based on channel
mean, in International Conference on Acoustics Speech and
Signal Processing, Orlando, Florida, USA., May 2002.
[8] E. Jorswiek and H. Boche, On transmit diversity with imperfect channel state information, in International Conference on Acoustics Speech and Signal Processing, Orlando,
Florida, USA., May 2002.
[9] M. Navarro, Receive and Transmit Strategies for Multiple
Antenna Systems, Ph.D. thesis, Institute for Telecommunications Research, University of South Australia, 2002.

[15] G. H. Golub and C. F. Van Loan, Matrix Computations, John


Hopkins, second edition, 1989.
[16] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis,
Cambridge University Press, 1991.

You might also like