On the MIMO Capacity with Imperfect CSI

Safa I. A. Ahmed and Kai-Kit Wong

University College London

Abstract: Multiple-input multiple-output (MIMO) antenna systems provide high capacity

gain over a single-antenna system. This gain is maximised when channel state information
(CSI) is available at the transmitter. In practice, however, only partial information could be
obtained. In this paper, we shall present a study of the MIMO system capacity in the case of
imperfect CSI. It is found that in the case of partial or imperfect CSI the capacity depends
on the statistical properties of the error in the CSI. Based on the knowledge of statistical
distributions of the deviations in CSI knowledge, a pre-coding strategy which maximises a
capacity upper bound is proposed and evaluated.

1 Introduction
Multiple-input multiple-output (MIMO) antenna systems promise high performance improvement for the
next generation wireless communications. Theoretic capacity limits of a MIMO channel have also been
known and widely acknowledged in the last decade [1-2]. It has been illustrated that the system capacity is
maximised when there is full knowledge of channel state information (CSI) at both the receiver and the
transmitter [2]. In practice, CSI may be learned at the receiver by transmitting (known) training sequences
and this CSI could also be fed back to the transmitter for exploiting the full potential of a MIMO channel.
However, the feedback capacity is limited and this will affect the resolution and quality of the CSI at the
transmitter, resulting in partial knowledge of CSI (or imperfect CSI) at the transmitter [4] although studies
have shown that a substantial capacity improvement is possible even with a small amount of feedback [3].
It is possible that the system performance may degrade severely if the system is optimised based on an
inaccurate CSI [4]. Other factors such as Doppler will definitely cause more imperfection on the CSI at
the transmitter if it is not updated frequently enough.

In this paper, a study of MIMO system with imperfect CSI is presented. In particular, we assume that the
statistical knowledge of the CSI error is known in addition to the partial CSI and our aim is to devise an
efficient MIMO pre-coding scheme robust to the CSI error for capacity maximisation. This will be done
by maximising a newly derived capacity upper bound.

The rest of this correspondence is organized as follows. Section 2 lays out the model of the system under
investigation. Section 3 derives a new capacity upper bound with the imperfect CSI. Section 4 presents a
MIMO pre-coding technique that uses the statistical properties of the CSI error for capacity maximisation.
Numerical results will be given in Section 5 and finally, we conclude the paper in Section 6.

2 System Model
Consider an ergodic flat fading Rayleigh channel with Nt transmit antennas and Nr receive antennas as
shown in Figure 1. This system can be described in a vector form as [5]
(1)
where y, H, x and n are, respectively, the received signal vector, the MIMO channel matrix, the transmit
symbol vector and the noise vector. To maximise the mutual information of the system, it is assumed that
x is circularly symmetric Gaussian [2] while the short-term power is constrained by P so that we have
(2)
On the other hand, the channel matrix H is assumed to have independent and identically distributed (i.i.d.)
zero-mean complex Gaussian entries (so that the channel amplitude is in Rayleigh fading). Likewise, n
has zero-mean complex Gaussian i.i.d. entries, each with variance of σ2.
 Transmitter Nt antennas
Nr antennas Receiver

Channel Noise

Figure 1: General MIMO system.

3 A New Capacity Upper Bound for MIMO with Imperfect CSI

~
In this paper, we assume that the transmitter has some knowledge of the channel Η which is different
from the actual channel matrix H by
(3)
where ΔH denotes the error matrix in CSI, which we assume has i.i.d. complex Gaussian entries.

Lemma 1: When there is partial knowledge of CSI at the transmitter, the capacity C becomes dependent on
the statistical error distribution and is given by
(4)

Proof: Equation (4) can be easily obtained from the original log-det capacity formula in [2] by substituting
Equation (3) into the formula.

Noting that ln(.) is a concave function, a capacity upper bound can be derived using Jensen’s inequality
[6], which gives
(5)

It is clear from (4) that the capacity depends on the statistical properties of the error. It suggests that if the
mean and covariance of the channel error are known at the transmitter, then a better capacity maximisation
be possible and this can be realised by maximising the capacity upper bound (5).

With the model (3), it is customary to consider the case where the channel error matrix ΔH has zero mean
and i.i.d. Gaussian entries, each with variance of σe2, which fits well with conventional channel estimation
scenarios. As a result, (5) yields

(6)

4 Robust MIMO Pre-Coding

Given the capacity upper bound in (6), a proper pre-coding design that can make the best use of the
available CSI with reference to its (statistical) confidence level to maximise the capacity can be devised.
In particular, this is achieved by maximising the capacity upper bound and the solution of Rx follows the
same way as the well known MIMO decomposition and water-filling power allocation in [1] but on the
effective channel matrix

Another possible way is that we can diagonalise the mean channel matrix by singular value decomposition
(SVD) and then allocate the power over the sub-channels based on scaling the estimated channel power to
the CSI error variance [or referred to as the channel-to-error ratio (CER)] in a water-filling manner. Both
schemes will be evaluated in Section 5.

5 Numerical Results
Capacity upper bound provided in Section 3 was simulated. Figure 2 shows the capacity bound and the
actual maximum achievable capacity. Results show that for a (2,2) system, the bound is tight especially at
high signal-to-noise ratios (SNR). The results for the two proposed pre-coding techniques are provided in
Figures 3-5. For modelling purposes, the channel matrix was normalised to provide a specified SNR using
Frobenius matrix norm [7]. Results illustrate that optimised MIMO transmission based on imperfect CSI
and error statistical properties results in remarkable capacity gain. In addition, results in Figure 3 reveal
the performance gain of the first proposed scheme while Figures 4 and 5 demonstrate the performance of
the second method for various CER. As can be seen, the proposed methods offer near-maximal capacity
under high CER. On the other hand, at low CER, the second method gives significant capacity gain over
the conventional MIMO water-filling solution without considering the CSI error statistics.

Capacity of 2 by 2 System Capacity of 3 by 3 System

40 8
Capacity Bound Nt=Nr=8 Partial CSI
Capacity Upper bound and M axim um c apacity bits/s/Hz

Partial CSI and Error

35 Capacity Full CSI Nt=Nr=8
7 Equal Power Transmission
Capacity Bound Nt=Nr=4
Full CSI
Capacity Full CSI Nt=Nr=4
30
Capacity Bound Nt=Nr=2 6
Capacity Full CSI Nt=Nr=2
25
C a p a c ity b its /s /H z

20
4
15

3
10

2
5

0 1
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 0 1.5 3 4.5 6 7.5 9 10.5 12 13.5 15
SNR [dB]
SNR [dB]

Figure 2: Capacity upper bound Figure 3: Capacity first scheme ZMG error
MIMO System Capacity 10 dB MIMO System Capacity CER 0dB
25 9
24 Capacity considering error Capacity considering error
23
22 Capacity partial Waterfilling 8 Capacity partial Waterfilling
21 Equal Power Transmission Equal Power Transmission
20
19 Full CSI Full CSI
7
18
17
16 6
Capacity bits/s/Hz

Capac ity bits /s /Hz

15
14
13
12 5
11
10
9 4
8
7
6 3
5
4
3 2
2
1
0 1
0 3 6 9 12 15 18 21 24 27 30 0 3 6 9 12 15 18 21 24 27 30
SNR [dB] SNR [dB]

Figure 4: Capacity second scheme CER 10dB Figure 5: Capacity second scheme CER 0dB
6 Conclusion
In this paper, a study of MIMO system with partial or imperfect CSI at the transmitter has been presented.
A new capacity upper bound for a MIMO channel with CSI error has been given using Jensen’s inequality
and robust MIMO solutions which maximise the channel capacity based on the bound have been devised
and evaluated. Numerical results have revealed significant capacity gains over conventional non-robust
MIMO systems by the proposed schemes exploiting the CSI error statistical knowledge for a wide range
of CER and SNR.

References

[1] G. G. Raleigh and J. M. Cioffi, “Spatio-temporal coding for wireless communication,” IEEE Trans.
Commun., vol 46, no. 3, pp. 357-366, March 1998.
[2] I. Telatar, “Capacity of multi-antenna Gaussian channels,” Bell Labs Technical Memorandum, June
1995.
[3] D. J. Love, R. W. Heath, W. Santipach, and M. L. Honig, “What is the value of limited feedback for
MIMO channels?,” IEEE Commun. Mag., vol. 42, no. 10, pp. 54-59, Oct. 2004.
[4] D. J. Love, “Duplex distortion models for limited feedback MIMO communication,” IEEE Trans. Sig.
Proc., vol. 54, no. 2, pp. 766-774, February 2006.
[5] E. Biglieri , R. Calderbank , A. Constantinides , A. Goldsmith , A. Paulraj , H. V. Poor, MIMO wireless
communications, Cambridge, January 2007.
[6] Jensen, Johan Ludwig William Valdemar (1906). "Sur les fonctions convexes et les inégalités entre les
valeurs moyennes". Acta Mathematica 30: 175-193.
[7] A. B. Gershman and N. D. Sidiropoulos, Space-time processing for MIMO communications, Wiley,
2005.