5238 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO.
10, OCTOBER 2009
The RF-Chain Limited MIMO System–Part I:
Optimum Diversity-Multiplexing Tradeoff
Yi Jiang and Mahesh K. Varanasi
Abstract—The large gain promised by the multi-input multi-
output (MIMO) technology comes with a cost. In particular,
multiple analog radio frequency (RF) chains, which are expensive
and power consuming, are required at both the transmitter and
receiver sides. On the other hand, the antennas connecting to the
RF chains are less expensive. Hence, one engineering compromise
is to implement more antennas than RF chains and to use only a
subset of them based on some antenna selection (AS) algorithm.
An interesting question therefore arises: given a RF chain limited
MIMO system, what is the fundamental performance gain by
adding more antennas? In this two-part paper, we answer this
question by using the diversity-multiplexing (D-M) gain tradeoff
metric. Consider a Rayleigh fading channel with 𝑀𝑡 antennas
and 𝐿𝑡 (𝐿𝑡 ≤ 𝑀𝑡 ) RF chains at the transmitter while 𝑀𝑟
antennas and 𝐿𝑟 (𝐿𝑟 ≤ 𝑀𝑟 ) RF chains at the receiver. We
obtain the fundamental D-M tradeoff as a function of 𝑀𝑡 , 𝑀𝑟 ,
and min(𝐿𝑟 , 𝐿𝑡 ). Referring to the special case where 𝐿𝑡 = 𝑀𝑡
and 𝐿𝑟 = 𝑀𝑟 as the RF unlimited system (or full system) and
RF limited system (or pruned system) otherwise, we prove that
the pruned system with optimal channel-dependent AS has the
same D-M tradeoff as the full system if the multiplexing gain
is less than some integer threshold 𝑃 , while it suffers from
some diversity gain loss for multiplexing gains larger than 𝑃 .
In particular, if min(𝐿𝑟 , 𝐿𝑡 ) = 𝐾 ≜ min(𝑀𝑟 , 𝑀𝑡 ), then 𝑃 = 𝐾,
i.e. the D-M tradeoffs of the pruned system and the full system
are the same. Moreover, this result can be extended to more
general fading channels such as Nakagami channel. A fast and
D-M tradeoff-optimal AS algorithm is proposed as a byproduct
of our analysis.
Index Terms—Antenna selection, diversity gain, fast algorithm,
MIMO, Nakagami fading, outage probability, Rayleigh fading,
spatial multiplexing gain, tradeoff.
I. I NTRODUCTION
M ULTIPLE-INPUT multiple-output (MIMO) wireless
communication systems have well-known advantages
over their single-input single-output (SISO) counterparts,
namely, much higher spectral efficiency and greatly improved
reliability [1][2][3][4]. By quantifying spectral efficiency as
multiplexing gain and reliability as diversity gain, it is shown
in [5] that a MIMO system can achieve both diversity gain
Manuscript received October 17, 2008; revised March 29, 2009; accepted
April 15, 2009. The associate editor coordinating the review of this paper and
approving it for publication was I. Collings.
Y. Jiang was with the Department of Electrical and Computer Engineering,
University of Colorado, Boulder, CO, USA. He is now with Qualcomm Inc.,
San Diego, CA, USA (e-mail: yjiang@dsp.colorado.edu).
M. K. Varanasi is with the Department of Electrical and Com-
puter Engineering, University of Colorado, Boulder, CO, USA (e-mail:
varanasi@colorado.edu.).
This work is supported in part by NSF Grants CCF-0434410 and CCF-
0728955.
The material in this paper was presented in part at the IEEE International
Symposium on Information Theory, Nice, France, in June 2007.
Digital Object Identifier 10.1109/TWC.2009.081385
1536-1276/09$25.00 ⃝
c 2009 IEEE
and spatial multiplexing gain simultaneously but there is a
fundamental tradeoff between them. Such a tradeoff is referred
to as the diversity-multiplexing (D-M) gain tradeoff. The D-
M tradeoff metric has since then been a popular performance
measure for the existing MIMO communication schemes [5]
and has motivated some interesting new designs [6][7][8].
The dramatic performance gain of the MIMO system comes
with increased hardware complexity, i.e. more expensive and
power consuming analog radio frequency (RF) chains at both
sides of the channel [9]. On the other hand, the antennas
connecting to the RF chains are less expensive. Hence one
engineering compromise is to implement more antennas than
the RF chains and to use only a subset of them based on
some antenna selection (AS) algorithm [9] [10]. An inter-
esting question therefore arises: given an RF chain limited
MIMO system, what is the fundamental performance gain
by adding more antennas? In this two-part paper, we answer
this question using the D-M gain tradeoff metric. Assuming
the 𝐿𝑡 transmit antennas and 𝐿𝑟 receive antennas are chosen
according to the maximization of the instantaneous channel
mutual information, we derive the D-M tradeoff corresponding
to the optimal AS approach. To facilitate our discussion, we
refer to a channel with 𝑀𝑡 antennas and 𝐿𝑡 (𝐿𝑡 ≤ 𝑀𝑡 ) RF
chains at the transmitter while 𝑀𝑟 antennas and 𝐿𝑟 (𝐿𝑟 ≤ 𝑀𝑟 )
RF chains at the receiver as an (𝑀𝑡 × 𝑀𝑟 , 𝐿𝑡 × 𝐿𝑟 ) channel.
Note that the conventional channel where the number of RF
chains equal to the number of antennas can be denoted as
(𝑀𝑡 × 𝑀𝑟 , 𝑀𝑡 × 𝑀𝑟 ).
One of the major results of this paper is illustrated in Figure
1. We highlight the major points as follows.
P1 The D-M tradeoff curve of the (𝑀𝑡 × 𝑀𝑟 , 𝐿𝑡 × 𝐿𝑟 )
channel (− ∘ −) is strictly higher than (𝐿𝑡 × 𝐿𝑟 , 𝐿𝑡 × 𝐿𝑟 )
(−△−) except for the end point (𝑁, 0), where 𝑁 ≜
𝐿𝑡 ∧ 𝐿𝑟 . 1
P2 The D-M tradeoff curves of the (𝑀𝑡 × 𝑀𝑟 , 𝐿𝑡 × 𝐿𝑟 )
and (𝑀𝑡 × 𝑀𝑟 , 𝑀𝑡 × 𝑀𝑟 ) channels overlap for the
multiplexing gain 𝑟 ∈ (0, 𝑃 ), where 𝑃 is an integer
between 0 and 𝑁 .
P3 The D-M tradeoff curve of (𝑀𝑡 × 𝑀𝑟 , 𝐿𝑡 × 𝐿𝑟 ) is linear
for 𝑟 ∈ (𝑃, 𝑁 ). Moreover, it is a tangent to the D-M
tradeoff curve of the (𝑀𝑡 × 𝑀𝑟 , 𝑀𝑡 × 𝑀𝑟 ) channel.
P4 The D-M tradeoff curve of (𝑀𝑡 × 𝑀𝑟 , 𝐿𝑡 × 𝐿𝑟 ) channel
is a function of 𝑀𝑡 , 𝑀𝑟 , and depends on 𝐿𝑟 and 𝐿𝑡 only
through the minimum of these numbers, 𝑁 .
P5 If 𝑁 = 𝐾 ≜ 𝑀𝑡 ∧ 𝑀𝑟 , then the D-M tradeoffs of (𝑀𝑡 ×
𝑀𝑟 , 𝐿𝑡 × 𝐿𝑟 ) and (𝑀𝑡 × 𝑀𝑟 , 𝑀𝑡 × 𝑀𝑟 ) fully overlap.
1 Throughout this paper, we denote 𝑎 ∧ 𝑏 as min{𝑎, 𝑏} and 𝑎 ∨ 𝑏 as
max{𝑎, 𝑏}
JIANG and VARANASI: THE RF-CHAIN LIMITED MIMO SYSTEM: PART I OPTIMUM DIVERSITY-MULTIPLEXING TRADEOFF
(0,M M )
t r
Tangent Point
Diversity Gain d(r)
(0,L L )
t r
0
0 P
Fig. 1. The optimal D-M tradeoff the standard (𝑀𝑡 × 𝑀𝑟 , 𝑀𝑡 × 𝑀𝑟 )
channel, (𝐿𝑡 × 𝐿𝑟 , 𝐿𝑡 × 𝐿𝑟 ) channel, and RF chain limited (𝑀𝑡 × 𝑀𝑟 , 𝐿𝑡 ×
𝐿𝑟 ) channel. Here 𝑁 = 𝐿𝑡 ∧ 𝐿𝑟 and 𝐾 = 𝑀𝑡 ∧ 𝑀𝑟 .
Channel with M & M antennas but only 2 RF chains at Tx & Rx (L = L = 2)
t r
18
16
14
12
Diversity Gain d(r)
10
8 complicates the distribution of the pruned channel. This gap
6
4 we introduce the MIMO fading channel and the concept of
2 (1,1)
0
0 0.5
Spatial Multiplexing Gain r
Fig. 2. The improvement of the D-M gain tradeoff by introducing more
transmit and receive antennas.
P1-P3 agree with the intuition that the D-M gain tradeoff of
a (𝑀𝑡 × 𝑀𝑟 , 𝐿𝑡 × 𝐿𝑟 ) channel should lie somewhere between
those of (𝑀𝑡 ×𝑀𝑟 , 𝑀𝑡 ×𝑀𝑟 ) and (𝐿𝑡 ×𝐿𝑟 , 𝐿𝑡 ×𝐿𝑟 ) channels.
However, P1-P3 clearly represent a rather optimistic result. It
can be seen that introducing extra antennas can greatly boost
the D-M tradeoff although the maximal spatial multiplexing
gain is limited by the number of RF chains. Figure 2 shows
the simple case of a system with two RF chains at both
transmitter and receiver side. An immediate corollary of P2 is
the known result that pruning a 𝑀𝑡 × 𝑀𝑟 channel to a smaller
𝐿𝑡 × 𝐿𝑟 one can still maintain the maximal diversity gain
𝑀𝑡 𝑀𝑟 (corresponding to the spatial multiplexing gain 𝑟 = 0)
[11][12]. But the D-M tradeoff analysis is clearly a more
complete characterization. It follows from P4 that one can
always design the system with 𝐿𝑡 = 𝐿𝑟 for reduced hardware
complexity but without incurring D-M tradeoff loss.
Throughout this work, it is assumed that the receiver has
5239
perfect knowledge of the channel gains between all pairs of
(M x M ,L x L )
t r t r
𝑀𝑡 transmit antennas and 𝑀𝑟 receive antennas. This means
(Mt x Mr,Mt x Mr)
(Lt x Lr,Lt x Lr)
that channel estimation must be accomplished via training
over a predetermined sequence of sets of transmit and receive
antennas of size no greater than the number of available RF
chains. For example, if 𝑀𝑡 = 2𝐿𝑡 and 𝑀𝑟 = 𝐿𝑟 , the channel
between the first 𝐿𝑡 transmit and the 𝐿𝑟 receive antennas could
be estimated first followed by the second set of 𝐿𝑡 transmit
(P,(M −P)(M −P))
t r
antennas and the 𝐿𝑟 receive antennas. Moreover, in order to
convey the information
( ) about which antennas to select at the
transmitter, log 𝑀 𝐿𝑡
𝑡
bits of feedback are provided from the
receiver to the transmitter for each coherence time (or channel
realization). Channel estimation and the finite-rate feedback,
(provisions for which are generally made anyway for control
N K signaling, are hence the additions to system complexity rel-
Spatial Multiplexing Gain r
ative to a (𝐿𝑡 , 𝐿𝑟 ) MIMO system. The gain in performance
of the RF chain limited MIMO system over the conventional
(𝐿𝑡 , 𝐿𝑟 ) MIMO system must hence be judged in light of this
modest increase in system complexity.
It is not difficult to see that the major points summarized
t r
above also constitute an answer to the following question:
Mt = Mr = 2
what is the optimal D-M tradeoff of an 𝐿𝑟 ×𝐿𝑡 channel pruned
M =M =3
t r
from an 𝑀𝑟 × 𝑀𝑡 original channel? Although considerable
M =M =4
t r research work has been done on quantifying the influence
of AS upon system performance, including channel capacity,
outage probability, and diversity gain (see [9] [10] [13] [14]
and the references therein), the D-M tradeoff analysis of
the pruned systems has remained elusive, which is mainly
(8,1)
due to the fact that the channel-dependent antenna selection
is now closed.
(4,1)
The rest of the paper is organized as follows. In Section II
D-M gain tradeoff. A fast AS algorithm inspired by Greedy
QR decomposition is introduced in Section III. Interestingly,
this algorithm turns out to be optimal in terms of the D-M
1
1.5 2
gain tradeoff. In Section IV, we obtain the fundamental D-
M gain tradeoff of a MIMO system with optimal AS. The
D-M tradeoff result is further extended to the more general
Nakagami fading channel. The summary and conclusions of
the theoretical results and their applications are relegated to
the second part of this paper [15].
II. MIMO C HANNEL M ODEL AND D-M T RADEOFF
A. Channel Model
Consider a communication system with 𝑀𝑡 transmit and 𝑀𝑟
receive antennas in an independent, identically distributed (iid)
Rayleigh flat fading channel. The sampled baseband signal is
given by
y = Hx + z, (1)
where H ∈ ℂ𝑀𝑟 ×𝑀𝑡 is the fading channel matrix, x ∈ ℂ𝑀𝑡 ×1
is the information symbol vector, and y ∈ ℂ𝑀𝑟 ×1 is the
received signal. Without loss of generality, we assume that
z ∼ 𝑁 (0, I𝑀𝑟 ) is the circularly symmetric complex Gaussian
noise where I𝑀𝑟 denotes the identity matrix with dimension
𝑀𝑟 . Given 𝔼[x∗ x] = 𝜌, where 𝔼[⋅] is the expectation and (⋅)∗
is the conjugate transpose, the input SNR is 𝜌.
5240 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 10, OCTOBER 2009

B. Diversity-Multiplexing Gain Tradeoff and Approximate where 𝐾 ≜ 𝑀𝑟 ∧𝑀𝑡 . More recently, the D-M tradeoff analysis
Universality is extended to more general fading channels such as Rician and
In [5], the authors established the framework of D-M Nakagami fading [17]. It is shown that some channel fadings
tradeoff analysis in the asymptotically high SNR regime. such as Nakagami-𝑚 (𝑚 ≥ 1) has better D-M tradeoff than
Denote 𝑅(𝜌) as the data rate of any communication scheme Rayleigh fading channels in the regime of 𝑟 ∈ [0, 1).
with input SNR 𝜌. The diversity gain and multiplexing gain The relationship (5) is true only for the limiting case that
are defined as follows. the frame length (code length) goes to infinity. But in terms
Definition 2.1: A scheme is said to have multiplexing gain of the coarser scale of D-M tradeoff, the constraint of frame
𝑟 and diversity gain 𝑑 if the data rate 𝑅(𝜌) satisfies length can be relaxed. For instance, it is proven in [5] that there
exist tradeoff-optimal codes with length 𝑇 ≥ 𝑀𝑡 + 𝑀𝑟 − 1 in
𝑅(𝜌)
lim = 𝑟, (2) iid Rayleigh fading channels. Furthermore, [18] shows that
𝜌→∞ log 𝜌 (6) holds independent of frame length and channel fading
and the average error probability 𝑃𝑒 (𝜌) satisfies distribution if the codes are approximately universal, i.e. the
log 𝑃𝑒 (𝜌) pairwise error probability for every pair of codewords decays
lim = −𝑑. (3) exponentially with SNR given that the channel realization is
𝜌→∞ log 𝜌
. not in outage. Indeed, based on [18, Theorem 3.1] it can
Following [5], we adopt the symbol = to denote exponential be shown that the recently proposed cyclic division algebra
. 𝑏
equality, i.e., we write 𝑓 (𝜌) = 𝜌 if (CDA) based 𝑀𝑡 × 𝑀𝑡 space-time block codes (STBC)
log 𝑓 (𝜌) with non-vanishing determinant (NDV) property [19][8][7]
lim = 𝑏. is approximately universal [18]. Therefore, the NDV STBCs
𝜌→∞ log 𝜌
. achieve the optimal D-M gain tradeoff for any fading channels
Thus (3) can be rewritten as 𝑃𝑒 (𝜌) = 𝜌−𝑑 . with delay 𝑇 = 𝑀𝑡 .
2
Define the outage probability as
𝑃outage (𝜌) ≜ inf ℙ (log ∣I + HQH∗ ∣ < 𝑅(𝜌)) , (4) III. FAST A NTENNA S ELECTION A LGORITHM
Qર0,tr(Q)<𝜌
Suppose the MIMO system is RF chain limited, i.e. the
where ℙ(ℰ) stands for the probability of the event ℰ. It is number of RF chain is less than the number of antennas.
proven in [16] that in the limit of increasing frame length We consider using only 𝐿𝑡 ≤ 𝑀𝑡 transmit antennas and
𝑃𝑒 = 𝑃outage . (5) 𝐿𝑟 ≤ 𝑀𝑟 receive antennas for data transmission. Denote
𝒮𝑡 ⊂ {1, 2, ⋅ ⋅ ⋅ , 𝑀𝑡 } and 𝒮𝑟 ⊂ {1, 2, ⋅ ⋅ ⋅ , 𝑀𝑟 } the sets of
That is, not only is 𝑃outage (𝜌) achievable in the sense that the indices of antennas selected at the transmitter and receiver
there exist codes with average frame error probability that is sides, respectively. The cardinality of the sets ∣𝒮𝑡 ∣ = 𝐿𝑡 and
arbitrarily close to the outage probability but that it is also ∣𝒮𝑟 ∣ = 𝐿𝑟 . We focus on isotropic transmission since it does
a fundamental limit in that a lower frame error probability not incur the loss of D-M gain tradeoff. To select the antenna
cannot be achieved for arbitrary 𝜖 > 0, there exists a code of subsets incurring the smallest capacity loss, one needs to solve
sufficient length for which 𝑃𝑒 (𝜌) < 𝑃outage (𝜌)+𝜖. Conversely, the following optimization problem
for sufficiently long codes, 𝑃𝑒 (𝜌) > 𝑃outage (𝜌) − 𝜖. Moreover,  
 
this relationship is independent of the channel fading statistics 𝒮𝑡opt , 𝒮𝑟opt = arg max𝒮𝑡 ,𝒮𝑟 log I + H𝒮𝑟 ,𝒮𝑡 H∗𝒮𝑟 ,𝒮𝑡 𝐿𝜌𝑡 
[16], i.e., it applies to any block fading channel, including subject to 𝒮𝑡 ⊂ {1, 2, ⋅ ⋅ ⋅ , 𝑀𝑡 }, 𝒮𝑟 ⊂ {1, 2, ⋅ ⋅ ⋅ , 𝑀𝑟 }
pruned ones. ∣𝒮𝑡 ∣ = 𝐿𝑡 , ∣𝒮𝑟 ∣ = 𝐿𝑟 .
Hence in the limit of long frame length, we have from (3) (9)
and (5) that the optimal diversity gain as a function of the Here H𝒮𝑟 ,𝒮𝑡 ∈ ℂ𝐿𝑟 ×𝐿𝑡 is the submatrix of H obtained by
multiplexing gain 𝑟 is keeping only the rows and columns whose indices are in
log 𝑃outage (𝜌) 𝒮𝑟 and 𝒮𝑡 , respectively. No solution
( 𝑟 ) to( (9) is
) known other
𝑑opt (𝑟) = − lim . (6) than exhaustive search over the 𝑀 ⋅ 𝑀𝑡
combinations,
𝜌→∞ log 𝜌 𝐿𝑟 𝐿𝑡

which makes the optimal solution computationally difficult,

It is easy be shown that water filling power allocation in the
especially in systems with many antennas [20]. Several fast,
spatial domain yields no D-M tradeoff improvement over the
yet suboptimal, AS algorithms have been proposed in the
isotropic transmission with the input covariance matrix 𝑀𝜌𝑡 I
literature, e.g., [21] [22]. We introduce next another new
[5]. Therefore,
(   ) computationally efficient AS algorithm which is later shown
 
log ℙ log I + HH∗ 𝑀𝜌𝑡  < 𝑟 log 𝜌 to be D-M tradeoff-optimal.
𝑑opt (𝑟) = − lim . (7)
𝜌→∞ log 𝜌
For an iid Rayleigh channel given in (1), the optimal D-M A. Algorithm Description
gain tradeoff is shown to be a piece-wise linear curve obtained The fast AS algorithm applies the same AS routine to
by connecting the following 𝐾 + 1 points [5] the transmit antennas and receive antennas separately. We
𝐾 remark that the AS routine is closely related to the Greedy QR
{(𝑟, (𝑀𝑟 − 𝑟)(𝑀𝑡 − 𝑟))}𝑟=0 , (8)
decomposition which plays an important role in the Greedy
2 We write A ર 0 if A is a positive semi-definite (p.s.d.) matrix, and ordering Rate Tailored V-BLAST (GRT-VB) scheme [23]. We
A ર B or B ⪯ A if A − B ર 0. first show how to select 𝐿𝑡 < 𝑀𝑡 transmit antennas. The
JIANG and VARANASI: THE RF-CHAIN LIMITED MIMO SYSTEM: PART I OPTIMUM DIVERSITY-MULTIPLEXING TRADEOFF 5241

transmit AS routine consists of 𝐿𝑡 steps. We elaborate the be the singular values of the pruned channel matrix H𝒮˜𝑟 ,𝒮˜𝑡
first step. The subsequent steps are easily inferred. obtained using the proposed fast AS algorithm. Then
In the first step, we go through the following procedure. ∏
𝑛
1
𝜆2𝑛 ˘ 2𝑛 ≤ 𝜆2𝑛 , 𝑛 = 1, ⋅ ⋅ ⋅ , 𝑁,
≤𝜆
(i) Calculate the Euclidean norms {∥h𝑖 ∥}𝑀 𝑖=1 where h𝑖 is
𝑡
(𝑀𝑟 − 𝑖 + 1)(𝑀𝑡 − 𝑖 + 1)
𝑖=1
the 𝑖th column of H. (13)
(ii) Permute h1 and h𝑗 where 𝑗 = arg max1≤𝑖≤𝑀𝑡 {∥h𝑖 ∥}. and
This operation can be represented by H1 = HΠ1 with
Π1 being a permutation matrix. (Π1 degrades to be I𝑀𝑡 𝑁 𝑁 𝑁
∏ 𝜆2𝑛 ∏ ∏
if 𝑗 = 1.) ≤ ˘2 ≤
𝜆 𝜆2𝑛 . (14)
𝑛
(iii) Apply a Householder matrix Q1 to transform the first 𝑛=1
(𝑀 𝑟 − 𝑛 + 1)(𝑀 𝑡 − 𝑛 + 1) 𝑛=1 𝑛=1
column of H1 to a scaled e1 , where e1 is the first column
An important corollary of Theorem 3.1 is the following;
of I𝑀𝑟 .
compared to the transmission over the 𝑁 strongest eigen-
The procedure (i–iii) can be illustrated as follows subchannels of the full system, the mutual information loss
⎛ ⎞ ⎛ ⎞ of the pruned channel obtained via the fast AS algorithm is
× × × × 𝑟11 × × ×
⎜ × × × × ⎟ Q∗1 HΠ1 ⎜ 0 × × × ⎟ upper bounded by a finite constant without regard to SNR,
⎜ ⎟ ⎜ ⎟
⎝ × × × × ⎠ −−−−−→ ⎝ 0 × × × ⎠ . which can be seen as follows. After transmit AS, the mutual
× × × × 0 × × × information of input and output of the pruned channel with
(10) input SNR 𝜌 is
Note that 𝑟11 = max{∥h𝑖 ∥, 1 ≤ 𝑖 ≤ 𝑀𝑡 }. In the next step,  
 
the same procedure is applied to the trailing (𝑀𝑟 −1)×(𝑀𝑡 − 𝐼(H𝒮˜𝑟 ,𝒮˜𝑡 , 𝜌) = log I + H𝒮˜𝑟 ,𝒮˜𝑡 H∗𝒮˜ ,𝒮˜ 𝐿𝜌𝑡 
∑𝑁 ( ˘2 𝜌
𝑟)𝑡
(15)
1) submatrix on the right hand side of (10), which yields a 𝜆
= 𝑛=1 log 1 + 𝐿𝑡
𝑛
.
permutation matrix Π2 and a Householder matrix Q2 . After
𝐿𝑡 recursive steps, we obtain In contrast, if the input power is uniformly loaded on the
strongest 𝑁 eigen-subchannels of the full system, the channel
HΠ = QR (11) mutual information is
∑𝑁 ( )
where Π = Π1 Π2 ⋅ ⋅ ⋅ Π𝐿𝑡 is a permutation matrix, Q = 𝜆2𝑛 𝜌
Q1 Q2 ⋅ ⋅ ⋅ Q𝐿𝑡 is a unitary matrix, and R is a matrix whose 𝐼(Λ𝑁 , 𝜌) = log 1 + , (16)
𝑛=1
𝑁
first 𝐿𝑡 columns form an upper triangular matrix with positive
diagonal elements {𝑟𝑖𝑖 }𝐿𝑖=1 . If the procedure is repeated for
𝑡
where Λ𝑁 is a diagonal matrix consisting of the 𝑁 largest
𝑀𝑡 ∧ 𝑀𝑟 steps, we obtain a Greedy QR decomposition which eigenvalues of HH∗ . Therefore,
is used in GRT-VB [23]. Denoting Π̃ and R̃ the submatrices
𝐼(Λ𝑁 , 𝜌) − 𝐼(H𝒮˜𝑟 ,𝒮˜𝑡 , 𝜌)
consisting of the first 𝐿𝑡 columns of Π and R, respectively, ⎛ ⎞
we have ∑𝑁 𝜌𝜆2
1 + 𝑁𝑛
HΠ̃ = QR̃. (12) = log ⎝ ˘2
⎠
𝜌𝜆
𝑛=1 1 + 𝐿𝑡 𝑛

We select the transmit antennas whose indices correspond to 𝑁

( )
∑ 𝐿𝑡 𝜆2𝑛
the nonzero rows of Π̃, and denote the set of their indices ≤ log
as 𝒮˜𝑡 . We denote the channel matrix after transmit AS as 𝑁𝜆 ˘2
𝑛=1 𝑛
H:,𝒮˜𝑡 ≜ HΠ̃ ∈ ℂ𝑀𝑟 ×𝐿𝑡 . This algorithm is computationally ∑𝑁 ( )
𝐿𝑡 (𝑀𝑡 − 𝑛 + 1)(𝑀𝑟 − 𝑛 + 1)
quite efficient as it involves only 𝑂(𝐿𝑡 𝑀𝑟 𝑀𝑡 ) complex mul- (by (14)) ≤ log
tiplications, as detailed in the Appendix. 𝑛=1
𝑁
To select the 𝐿𝑟 < 𝑀𝑟 receive antennas, we apply the same ≜ 𝐶. (17)
procedure to H𝑇:,𝒮˜ ∈ ℂ𝐿𝑡 ×𝑀𝑟 . In this case 𝐿𝑟 recursive steps
𝑡
On the other hand, it follows from the upper bound 𝜆𝑛 ≥ 𝜆 ˘𝑛
are involved. We denote by 𝒮˜𝑟 the set of indexes of selected
that 𝐼(Λ𝑁 , 𝜌) ≥ 𝐼(H𝒮˜𝑟 ,𝒮˜𝑡 , 𝜌). Hence we have obtained the
receive antennas. Hence, the pruned channel matrix can be
following bounds:
denoted by H𝒮˜𝑟 ,𝒮˜𝑡 ∈ ℂ𝐿𝑟 ×𝐿𝑡 . In general, 𝒮˜𝑡 ∕= 𝒮𝑡opt and 
 𝜌 
𝒮˜𝑟 ∕= 𝒮𝑟opt , i.e., the fast AS algorithm is suboptimal. But we 0 ≤ 𝐼(Λ𝑁 , 𝜌) − log I + H𝒮˜𝑟 ,𝒮˜𝑡 H∗𝒮˜𝑟 ,𝒮˜𝑡  ≤ 𝐶, (18)
shall see later that this algorithm is optimal in terms of D-M 𝑁
gain tradeoff. where 𝐶 is a finite constant. We deduce from these bounds
that the D-M tradeoff of the pruned channel is the same
as that of the transmission constrained over the 𝑁 strongest
B. Bounds on Singular Values eigen-subchannels of the full system, since a finite mutual
Besides the close-to-optimal performance, a major signifi- information gap amounts to a finite scaling of input SNR,
cance of this fast AS algorithm is that it enables us to reveal which does not influence the SNR exponent of the outage
the important relationship between the singular values of H probability. This observation is indeed the cornerstone of our
and H𝒮˜𝑟 ,𝒮˜𝑡 . analysis in the next section. It is also worth emphasizing that
Theorem 3.1: Let 𝜆1 ≥ 𝜆2 ≥ ⋅ ⋅ ⋅ ≥ 𝜆𝐾 be the singular Theorem 3.1 is purely a result in linear algebra and is true
values of H ∈ ℂ𝑀𝑟 ×𝑀𝑡 . Let 𝜆 ˘1 ≥ ⋅ ⋅ ⋅ ≥ 𝜆
˘ 𝑁 (𝑁 ≜ 𝐿𝑡 ∧ 𝐿𝑟 ) independently of the distribution of H.
5242 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 10, OCTOBER 2009

To prove Theorem 3.1, we need to establish the following Combining (23) and (25), we have proven the lemma.
two lemmas. Now we are ready to prove Theorem 3.1.
Lemma 3.2: Consider a matrix H ∈ ℂ𝑀𝑟 ×𝑀𝑡 with singular
values 𝜆1 ≥ 𝜆2 ≥ ⋅ ⋅ ⋅ ≥ 𝜆𝐾 ≥ 0 (𝐾 = 𝑀𝑡 ∧ 𝑀𝑟 ). Denote Proof: (of Theorem 3.1) We first prove the upper bound
𝑎𝑖 the Euclidean norm of the 𝑖th column of H, and 𝑏𝑖 the in (13). Consider the pruned matrix H:,𝒮𝑡 which consists of
Euclidean norm of the 𝑖th row of H. We have the columns of H whose indices belong to 𝒮𝑡 . Denote 𝜆 ˜1 ≥
𝑘 𝑘 ⋅⋅⋅ ≥ 𝜆 ˜𝑁 the 𝑁 largest singular values of H:,𝒮 . Because
∑ ∑ 𝑡
˜ 2 ≤ 𝜆2 for 1 ≤ 𝑘 ≤ 𝑁 [26]. Because
𝜆2𝑖 ≥ 𝑎2[𝑖] , 𝑘 = 1, 2, ⋅ ⋅ ⋅ , 𝐾, (19) H:,𝒮𝑡 H∗:,𝒮𝑡 ⪯ HH∗ , 𝜆 𝑘 𝑘
𝑖=1 𝑖=1 ˘2 ≤ 𝜆
H∗𝒮𝑟 ,𝒮𝑡 H𝒮𝑟 ,𝒮𝑡 ⪯ H∗:,𝒮𝑡 H:,𝒮𝑡 , 𝜆 ˜2 for 1 ≤ 𝑘 ≤ 𝑁 . Hence
𝑘 𝑘
and ˘ ≤ 𝜆 and the upper bound in (13) is proven.
𝜆 2 2
𝑘 𝑘
𝑘
∑ 𝑘
∑ Let H:,𝒮𝑡 = QR̃ be the QR decomposition where Q is
𝜆2𝑖 ≥ 𝑏2[𝑖] , 𝑘 = 1, 2, ⋅ ⋅ ⋅ , 𝐾, (20) given in (21), and R̃ is the submatrix consisting of the first
𝑖=1 𝑖=1
𝐿𝑡 columns of R there. Clearly R̃ has the same diagonal as
where 𝑎[𝑖] and 𝑏[𝑖] are the 𝑖th largest elements of the sequences 2 𝜆2𝑖
R. It follows from Lemma 3.3 that 𝑟𝑖𝑖 ≥ 𝑀𝑡 −𝑖+1 . Hence,
{𝑎}𝑀 𝑡 𝑀𝑟
𝑖=1 and {𝑏}𝑖=1 , respectively. 𝑛 𝑛
∏ ∏ 𝜆2𝑖
Proof: Note that {𝜆2𝑖 }𝐾 𝑖=1 are the largest 𝐾 singular 2
𝑟𝑖𝑖 ≥ . (26)
values of the Hermitian matrices HH∗ and H∗ H. Also note 𝑖=1 𝑖=1
𝑀𝑡 − 𝑖 + 1
that H∗ H has diagonal elements {𝑎2𝑖 }𝑀 𝑖=1 and HH has
𝑡 ∗
2 𝑀𝑟
diagonal elements {𝑏𝑖 }𝑖=1 . The lemma follows immediately Note that the singular values of R̃ and H:,𝒮𝑡 are the same.
from Schur-Horn’s Theorem [24, Theorems 4.3.26], which Recall the fact that for an upper triangular matrix, the
says that the diagonal elements of a positive semi-definite squared diagonal elements are multiplicatively majorized by
matrix is additively majorized by the singular (eigen) values its squared singular values [27][25], i.e.,
∏𝑛 ∏𝑛 ∏𝑛
[25]. ˜2 ≥ 2 𝜆2𝑖
Lemma 3.3: Applying the Greedy QR decomposition (i.e. 𝜆𝑖𝑖 𝑟𝑖𝑖 ≥ , for 1 ≤ 𝑖 ≤ 𝑁. (27)
𝑖=1 𝑖=1 𝑖=1
𝑀𝑡 − 𝑖 + 1
the procedure (10) is applied 𝐾 times) to H yields
Combining (27) and the proven upper bound that 𝜆 ˜ 𝑖 ≤ 𝜆𝑖 , we
HΠ = QR. (21) have
2 ∏𝑛
Denote 𝑟𝑘𝑘 the squared 𝑘th diagonal elements of R. Then ˜ 2 ≥ 𝜆2 1
𝜆 𝑛 𝑛 , 𝑛 = 1, 2, ⋅ ⋅ ⋅ , 𝑁. (28)
∑𝐾 2 𝑀 𝑡 − 𝑖+1
𝑖=𝑘 𝜆𝑖
𝑖=1
2
𝑟𝑘𝑘 ≥ , 𝑘 = 1, 2, ⋅ ⋅ ⋅ , 𝐾. (22)
𝑀𝑡 − 𝑘 + 1 Applying the same AS procedure to H𝑇:,𝒮𝑡 , we obtain H𝒮𝑟 ,𝒮𝑡
Proof: Recall that the Greedy QR decomposition is with singular values 𝜆˘1 ≥ . . . ≥ 𝜆
˘ 𝑁 . Following the same
achieved by successively applying the procedure illustrated argument leading to (28), we can prove that
in (10) 𝐾 times. According to the procedure (i)-(iii) given in ∏𝑛
1
Section III-A, ˘2 ≥ 𝜆
𝜆 ˜2 , 𝑛 = 1, 2, ⋅ ⋅ ⋅ , 𝑁. (29)
𝑛 𝑛
𝑀 𝑟 − 𝑖+1
∑𝐾 2 𝑖=1
1 𝜆 Combining (28) and (29), we have proven the lower bound in
2 2
𝑟11 = max {∥h𝑖 ∥ } ≥ ∥H∥𝐹 = 𝑖=1 𝑖
2
(23)
1≤𝑖≤𝑀𝑡 𝑀𝑡 𝑀𝑡 (13).
where ∥ ⋅ ∥𝐹 stands for the Frobenius norm. Hence (22) is The upper bound in (14) is trivial given the proven upper
true for 𝑘 = 1. At the 𝑘th step (2 ≤ 𝑘 ≤ 𝐾), we have ˘ 𝑛 ≤ 𝜆𝑛 for ∀𝑛. We see from (27) that
bound 𝜆
HΠ1 ⋅ ⋅ ⋅ Π𝑘−1 = Q1 ⋅ ⋅ ⋅ Q𝑘−1 R(𝑘) with 𝑁
∏ 𝑁
∏ 𝜆2𝑛
⎛ ⎞ ˜2 ≥
𝜆 . (30)
𝑛
𝑟11 × ⋅ ⋅ ⋅ ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅ × 𝑀𝑡 − 𝑛 + 1
𝑛=1 𝑛=1
⎜ .. .. ⎟
⎜ 0 . . ⋅⋅⋅ ⋅⋅⋅ × ⎟ In a similar vein,
⎜ ⎟
⎜ 𝑟𝑘−1,𝑘−1 ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ × ⎟
R(𝑘) = ⎜ 0 0 ⎟ . (24) 𝑁
∏ 𝑁
∏ ˜2
𝜆
⎜ 0 ⋅⋅⋅ 0 ∗ ⋅⋅⋅ ∗ ⎟ ˘2 ≥
𝜆 𝑛
. (31)
⎜ ⎟ 𝑛
𝑀 − 𝑛+1
⎝ 0 ⋅⋅⋅ 0 ∗ ⋅⋅⋅ ∗ ⎠ 𝑛=1 𝑛=1 𝑟

0 ⋅⋅⋅ 0 ∗ ⋅⋅⋅ ∗ It follows from (30) and (31) that

Because left and right multiplying a matrix by any unitary 𝑁
∏ 𝑁
∏ 𝜆2𝑛
matrix does not change its singular values, R(𝑘) has the ˘2 ≥
𝜆 . (32)
𝑛
(𝑀𝑟 − 𝑛 + 1)(𝑀𝑡 − 𝑛 + 1)
same singular values as H. According to Lemma 3.2, ∑the first
𝑛=1 𝑛=1

𝑘 − 1 rows of R(𝑘) have Frobenius norm less than 𝑘−1 2

𝑖=1 𝜆𝑖 .
The theorem is proven.
Therefore the trailing (𝑀𝑟 − 𝑘 + 1) × (𝑀𝑡 − 𝑘 + 1)
∑submatrix
𝐾
(denoted by ∗’s) has Frobenius norm larger than 𝑖=𝑘 𝜆2𝑖 . It IV. D IVERSITY-M ULTIPLEXING T RADEOFF A NALYSIS
follows that In this section, we derive the fundamental D-M tradeoff of
𝐾 an optimally pruned channel. We first establish the result in
1 ∑
2
𝑟𝑘𝑘 ≥ 𝜆2𝑖 , 2 ≤ 𝑘 ≤ 𝐾. (25) the case of iid Rayleigh fading channel. The extension to more
𝑀𝑡 − 𝑘 + 1 general fading channels is presented in Section IV-B.
𝑖=𝑘
JIANG and VARANASI: THE RF-CHAIN LIMITED MIMO SYSTEM: PART I OPTIMUM DIVERSITY-MULTIPLEXING TRADEOFF 5243

D−M Tradeoff, Mt = 5, Mr = 6 antennas, meanwhile the increase of hardware complexity is

30 minimal by using AS.
N=4
N=3 We now prove Theorem 4.1.
N=2
25
N=1
Full Syst. or N=5 Proof: The proof of the theorem contains two parts. The
iid ch. Mt=Mr=3 first part is the derivation leading to (41) which leans heavily
Diversity Gain d(r)

20
on the techniques used in [5]. Hence, we only give a sketch for
this part. The second part is the solution of the optimization
15 problem (41).
Denote
(   )
10  𝜌 
 ∗
𝑃outage,p (𝑟, 𝜌) = ℙ log I + H𝒮˜𝑟 ,𝒮˜𝑡 H𝒮˜𝑟 ,𝒮˜𝑡  < 𝑟 log 𝜌 ,
𝐿𝑡
5
the outage probability of the pruned channel with input SNR
𝜌 and multiplexing gain 𝑟. Here the subscript “p” stands
0
0 1 2 3 4 5 for “pruned”. The diversity gain of the pruned system with
Spatial Multiplexing Gain r multiplexing gain 𝑟 is
Fig. 3. Optimal D-M tradeoffs of full and pruned MIMO systems log 𝑃outage,p (𝑟, 𝜌)
𝑑p (𝑟) = − lim .
𝜌→∞ log 𝜌

A. Rayleigh Fading Channel Since a finite mutual information gap amounts to a finite
scaling of input SNR, which does not influence the SNR
For the pruned iid Rayleigh fading channel, the optimal D- exponent of the outage probability, it follows from (18) that
M tradeoff is summarized in the following theorem. ( 𝑁 )
Theorem 4.1: Consider pruning the 𝑀𝑟 × 𝑀𝑡 Rayleigh . ∑ ( )
fading MIMO channel given in (1) to a smaller one with 𝑃outage,p (𝑟, 𝜌) = ℙ log 1 + 𝜌𝜆2𝑛 < 𝑟 log 𝜌 . (35)
𝑛=1
𝐿𝑡 transmit and 𝐿𝑟 receive antennas using the proposed fast
AS algorithm. The optimal D-M gain tradeoff of the pruned For an i.i.d Rayleigh fading channel, the joint distribution of
system is a piecewise linear curve obtained by connecting the the ordered squared singular values of H, 𝜆21 ≥ 𝜆22 ≥ ⋅ ⋅ ⋅ ≥
following 𝑃 + 2 points 𝜆2𝐾 > 0, is [28]
∏
𝐾
2(𝑀 −𝐾)
∏ ∑
𝜆2
𝑓 (𝜆21 , ⋅ ⋅ ⋅ , 𝜆2𝐾 ) = 𝐶𝑀𝑡 ,𝑀𝑟 ⋅ 𝜆𝑘 (𝜆2𝑘 − 𝜆2𝑗 )2 𝑒− 𝑖 𝑖 ,
{𝑛, (𝑀𝑟 − 𝑛)(𝑀𝑡 − 𝑛)}𝑃
𝑛=0 , (𝑁, 0), (33) 𝑘=1 𝑘<𝑗

where 𝑁 ≜ 𝐿𝑟 ∧ 𝐿𝑡 , and where 𝐶𝑀𝑡 ,𝑀𝑟 is a normalizing constant and 𝑀 ≜ 𝑀𝑟 ∨ 𝑀𝑡 .

Define 𝜆2𝑘 = 𝜌−𝛼𝑘 , 1 ≤ 𝑘 ≤ 𝐾 where 𝛼1 ≤ 𝛼1 ≤ ⋅ ⋅ ⋅ ≤ 𝛼𝐾
(𝑀𝑟 −𝑝)(𝑀𝑡 −𝑝)
as 0 < 𝜆1 ≤ ⋅ ⋅ ⋅ ≤ 𝜆𝐾 . The distribution of 𝜶 is
𝑃 = arg min𝑝 𝑁 −𝑝 (34) 𝐾
subject to 0 ≤ 𝑝 ≤ 𝑁 − 1, 𝑝 ∈ ℤ. ∏
𝑓 (𝛼1 , . . . , 𝛼𝐾 ) = 𝐶𝑀𝑡 ,𝑀𝑟 (log 𝜌)𝐾 𝜌−[(𝑀−𝐾)+1]𝛼𝑘
Moreover, this D-M gain tradeoff is an upper bound for any ∏
𝑘=1
∑
AS strategy. 𝜌−𝛼𝑖
× (𝜌−𝛼𝑘 − 𝜌−𝛼𝑗 )2 𝑒− 𝑖 . (36)
As illustrated in Figure 3, Theorem 4.1 says that a good 𝑘<𝑗
AS algorithm incurs no diversity gain loss if the multiplexing
. +
gain is less than 𝑃 . In particular, if 𝑁 = 1 (𝑁 = 𝐿𝑡 ∧ 𝐿𝑟 ), the At high SNR, (1 + 𝜌𝜆2𝑛 ) = 𝜌(1−𝛼𝑛 ) with (𝑥)+ = 𝑥 ∨ 0. We
tradeoff is the line connecting (0, 𝑀𝑡 𝑀𝑟 ) and (1, 0). On the see from (35) that
other hand, if the antenna selection is made such that 𝑁 = 𝐾 ( 𝑁 )
(𝐾 = 𝑀𝑡 ∧ 𝑀𝑟 ), then 𝑃 = 𝐾 − 1 since the slope of the . ∑
+
𝑃outage,p (𝑟, 𝜌) = ℙ (1 − 𝛼𝑛 ) < 𝑟
curve corresponding to 𝑑 ∈ (𝐾 − 1, 𝐾) is the smallest among 𝑛=1
all the 𝐾 pieces. Therefore the pruned system has the same ∫ ∏
𝐾
D-M tradeoff as the full system. For instance, as in Figure 3, .
= 𝜌−[(𝑀−𝐾)+1]𝛼𝑘
if 𝐿𝑡 = 𝐿𝑟 = 5, the pruned system still has the D-M tradeoff 𝒜 𝑘=1
as shown in − ∘ −. ∏ ∑
−𝛼𝑘 𝜌−𝛼𝑖
⋅ (𝜌 − 𝜌−𝛼𝑗 )2 𝑒− 𝑖 𝑑𝜶.(37)
On the other hand, Theorem 4.1 also quantifies the improve-
𝑘<𝑗
ment of the D-M tradeoff by introducing additional antennas at
transmitter and receiver without increasing the number of RF where
chains and the size of codes. In Figure 3 the D-M tradeoff of a { 𝑁
}
∑
3×3 iid Rayleigh channel is presented (−♢−). Comparing the 𝒜= 𝜶 : 𝛼1 ≤ 𝛼2 ≤ ⋅ ⋅ ⋅ ≤ 𝛼𝐾 , (1 − 𝛼𝑛 ) < 𝑟 . +
lines −♢− and −□−, we see the dramatic improvement by 𝑛=1
introducing additional three receive antennas and two transmit (38)
5244 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 10, OCTOBER 2009

Also note that exp (−𝜌−𝛼𝑛 ) decreases with 𝜌 exponentially for has nonempty interior [29]. Hence according to the convex
any 𝛼𝑛 < 0 and exp (−𝜌−𝛼𝑛 ) → 1 as 𝜌 → ∞ for 𝛼𝑛 > 0. optimization theory [29]
Therefore,
𝑑p (𝑟) = inf sup ℒ(𝜶, 𝜇, 𝜸). (45)
∫ 𝐾
∏
𝜶 𝜇,𝜸
.
𝑃outage,p (𝑟, 𝜌) = 𝜌−[(𝑀−𝐾)+1]𝛼𝑘 According to the complementary slackness condition,
𝒜+ 𝑘=1
∏ 𝑁
∑
−𝛼𝑘
⋅ (𝜌 − 𝜌−𝛼𝑗 )2 𝑑𝜶, (39) 𝛾1 𝛼1 = 𝛾𝑛 (𝛼𝑛 − 𝛼𝑛−1 ) = 𝛾𝑁 +1 (1 − 𝛼𝑁 ) = 0. (46)
𝑘<𝑗 𝑛=2

where Equating to zero the partial derivative of the Lagrangian with

{ 𝑁
} respect to 𝛼𝑛 ’s, we obtain
∑
𝒜+ = 𝜶 : 0 < 𝛼1 ≤ ⋅ ⋅ ⋅ ≤ 𝛼𝐾 , (1 − 𝛼𝑛 )+ < 𝑟 . 𝑐𝑛 = 𝜇 + 𝛾𝑛 − 𝛾𝑛+1 , 1 ≤ 𝑛 ≤ 𝑁. (47)
𝑛=1
Inserting (46) and (47) into (45) yields
Since
∏ 𝛼𝑖 ≤ 𝛼𝑗 for 𝑗 > 𝑖, we can replace the term
−𝛼𝑘 𝑁
∑
𝑘<𝑗 (𝜌 − 𝜌−𝛼𝑗 )2 by 𝜌−2(𝐾−𝑘)𝛼𝑘 at high SNR. Hence
∫ ∏𝐾 𝑑p (𝑟) = inf sup (𝜇 + 𝛾𝑛 − 𝛾𝑛+1 )𝛼𝑛
log 𝜌−(𝑀 +𝐾−2𝑘+1)𝛼𝑘 𝑑𝛼1 ⋅ ⋅ ⋅ 𝑑𝛼𝐾 𝜶 𝜇,𝜸
𝒜+ 𝑘=1 𝑛=1
𝑑p (𝑟) = − lim . 𝑁 𝑁
𝜌→∞ log 𝜌 ∑ ∑
(40) = inf sup 𝜇 𝛼𝑛 + (𝛾𝑛 − 𝛾𝑛+1 )𝛼𝑛 . (48)
𝜶 𝜇,𝜸
According to Laplace’s principle, the integral is dominated 𝑛=1 𝑛=1
by the term corresponding to the largest SNR exponent as ∑𝑁
𝜌 → ∞. Hence
As we can rewrite
∑𝑁 𝑛=1 (𝛾𝑛 − 𝛾𝑛+1 )𝛼𝑛 = 𝛾1 𝛼1 +
𝑛=2 𝛾 𝑛 (𝛼𝑛 − 𝛼𝑛−1 ) − 𝛾 𝑁 +1 𝛼𝑁 , it follows from (46) that
𝐾
∑ 𝑁
∑
𝑑p (𝑟) = inf (𝑀 + 𝐾 − 2𝑘 + 1)𝛼𝑘
𝛼𝑘 (𝛾𝑛 − 𝛾𝑛+1 )𝛼𝑛 = −𝛾𝑁 +1 . (49)
𝑘=1
𝑛=1
subject to {𝛼𝑘 }𝐾
𝑘=1 ∈ 𝒜+ . (41)
Combining (48) and (49), we obtain
We observe that at an optimal solution to (41) it must be 𝑁
∑
true that (i) 𝛼𝑘 ≤ 1 for ∀𝑘 and (ii) 𝛼𝑘 = 𝛼𝑁 for 𝑁 ≤ 𝑘 ≤ 𝐾. 𝑑p (𝑟) = inf sup 𝜇 𝛼𝑛 − 𝛾𝑁 +1
𝜶 𝜇,𝜸
The argument for (ii) is trivial. As for (i), if there are some 𝑛=1
elements of 𝜶 which are greater than one, without violating the = sup 𝜇(𝑁 − 𝑟) − 𝛾𝑁 +1 . (50)
constraint, we can set them to be one and reduce the objective 𝜇,𝜸

function in (41). Also note that 𝑀 + 𝐾 = 𝑀𝑡 + 𝑀𝑟 since It follows from the 𝑁 equations in (47) that
𝑀 = 𝑀𝑡 ∨ 𝑀𝑟 and 𝐾 = 𝑀𝑡 ∧ 𝑀𝑟 . Using these observations 𝑛
∑
and denoting 𝛾𝑛+1 = 𝛾1 + 𝑛𝜇 − 𝑐𝑖 , 1 ≤ 𝑛 ≤ 𝑁. (51)
{ 𝑖=1
𝑀𝑡 + 𝑀𝑟 − 2𝑛 + 1 for 𝑛 = 1, ⋅ ⋅ ⋅ , 𝑁 − 1 ∑𝑁
𝑐𝑛 = In particular 𝛾𝑁 +1 = 𝛾1 + 𝑁 𝜇 −
(𝑀𝑡 − 𝑁 + 1)(𝑀𝑟 − 𝑁 + 1) for 𝑛 = 𝑁 , 𝑖=1 𝑐𝑖 . Therefore,
(42) 𝑁
∑
we can rewrite (41) as 𝑑p (𝑟) = sup 𝜇(𝑁 − 𝑟) − (𝛾1 + 𝑁 𝜇 − 𝑐𝑖 )
𝜇,𝜸
𝑖=1
𝑁
∑ 𝑁
∑
𝑑p (𝑟) = inf 𝑐𝑛 𝛼𝑛
𝛼𝑛 = 𝑐𝑖 − inf (𝛾1 + 𝜇𝑟)
𝑛=1 𝜇,𝜸
𝑖=1
subject to 0 < 𝛼1 ≤ 𝛼2 ≤ ⋅ ⋅ ⋅ ≤ 𝛼𝑁 < 1
𝑁
= 𝑀𝑡 𝑀𝑟 − inf (𝛾1 + 𝜇𝑟). (52)
∑ 𝜇,𝛾1
𝛼𝑛 > 𝑁 − 𝑟. (43) The constraints of 𝜇, 𝛾1 are implied in (51), i.e.,
𝑛=1
𝑛
∑
To make the problem physically meaningful, we focus on the 𝛾1 + 𝑛𝜇 − 𝑐𝑖 ≥ 0, 1 ≤ 𝑛 ≤ 𝑁. (53)
case 0 ≤ 𝑟 ≤ 𝑁 . It is straightforward to show that (43) is a 𝑖=1
convex problem whose Lagrangian is
[ 𝑁 ] Let us consider the optimization problem
∑𝑁 ∑
ℒ(𝜶, 𝜇, 𝜸) = 𝑐𝑛 𝛼𝑛 − 𝜇 𝛼𝑛 − (𝑁 − 𝑟) − 𝛾1 𝛼1 inf 𝜇,𝛾1 (𝛾1 +
∑𝜇𝑟)
𝑛 (54)
𝑛=1 𝑛=1 subject to 𝛾1 + 𝑛𝜇 − 𝑖=1 𝑐𝑖 ≥ 0, 1 ≤ 𝑛 ≤ 𝑁,
∑𝑁
which is depicted in Figure 4. The cost function 𝛾1 + 𝜇𝑟 can
− 𝛾𝑛 (𝛼𝑛 − 𝛼𝑛−1 ) − 𝛾𝑁 +1 (1 − 𝛼𝑁 ), (44) be represented by the dashed line passing the point (0, 𝛾1 )
𝑛=2
with∑slope 𝜇. The 𝑁 + 1 dots in Figure 4 have coordinates
𝑛
where the multipliers 𝜇, 𝜸 ≥ 0. It can been seen that the (𝑛, 𝑖=1 𝑐𝑛 ), 0 ≤ 𝑛 ≤ 𝑁 , which are indexed from left to right
Slater’s condition is satisfied for 𝑟 > 0, i.e., the constraint set as the zero-th to the 𝑁 th point. The constraints (𝛾1 + 𝑛𝜇 ≥
JIANG and VARANASI: THE RF-CHAIN LIMITED MIMO SYSTEM: PART I OPTIMUM DIVERSITY-MULTIPLEXING TRADEOFF 5245

γ1+μr B. Extension to Non-Rayleigh Fading Channel

M M The D-M tradeoff analysis in [5] is based on the assumption
t r
of iid Rayleigh fading channel. Recently, [30] extends the
result of [5] to the more general cases where the entries of H
have distribution:
𝛽
𝑓∣ℎ𝑖𝑗 ∣ (𝑥) = 𝑎𝑥𝑡 𝑒−𝑏∣𝑥−𝑐∣ , 𝑥 ≥ 0, (58)
which includes the Rician and Nakagami fading channels as
special cases. It is shown that for an 𝑀𝑟 × 𝑀𝑡 MIMO channel
(n,c1+...+cn)
with distribution given in (58), the optimal D-M tradeoff is a
piece-wise linear curve obtained by connecting the 𝐾 + 1
points:
γ1 infγ (γ1+μ r) ( ( ) )
,μ 𝑡 𝐾
0, 1 + 𝑀𝑡 𝑀𝑟 , {(𝑘, (𝑀𝑟 − 𝑘)(𝑀𝑡 − 𝑘))}𝑘=1 .
1

0 P N r
2
(59)
∑𝑛 That is, the channel with distribution (58) may have better D-
Fig. 4. Visualization of 𝛾1 + 𝑛𝜇 and 𝑖=1 𝑐𝑖 . M tradeoff than the Rayleigh channel for 𝑟 ∈ [0, 1), while it
has the same tradeoff as the Rayleigh channel for 𝑟 ∈ [1, 𝐾].
∑𝑛 Note that the bound in (18) is obtained without assuming
𝑖=1 𝑐𝑖 ) mean that the dashed line should always be above the the distribution of H. Therefore according to (35),
𝑁 + 1 points. Hence we see that the function inf 𝛾1 ,𝜇 𝛾1 + 𝜇𝑥 ( 𝑁 )
subject to the constraints given in (54) is the upper edge of . ∑ ( )
2
𝑃outage,p (𝑟, 𝜌) = ℙ log 1 + 𝜌𝜆𝑛 < 𝑟 log 𝜌 . (60)
the convex hull spanned by the 𝑁 + 1 points. The slope of the
𝑛=1
straight line passing the two points with indices 𝑝 (𝑝 < 𝑁 )
and 𝑁 is Combining the results in [30] and the derivations leading to
∑𝑁 (39), we can show that
𝑛=𝑝 𝑐𝑛 (𝑀𝑟 − 𝑝)(𝑀𝑡 − 𝑝) ∫ ∏
𝐾
𝜇= = .
𝑁 −𝑝 𝑁 −𝑝 𝑃outage,p (𝑟, 𝜌) = 𝜌−[(𝑀 −𝐾)+1]𝛼𝑘
𝒜+ 𝑘=1
∏ 𝑡
We refer to the points on the edge of the convex hull as “active (𝜌−𝛼𝑘 − 𝜌−𝛼𝑗 )2 𝜌− 2 𝑀𝑟 𝑀𝑡 𝛼1 𝑑𝜶, (61)
points”. The 𝑁 th point is an “active point”. Then the adjacent 𝑘<𝑗

“active point” to the 𝑁 th point must have the index defined Similar to the derivations from (39) to (43), we can obtain the
in (34). Otherwise a line passing the 𝑁 th and the 𝑝th (𝑝 ∕= optimal D-M tradeoff of the pruned channel as
𝑃 ) points would be below the 𝑃 th point, which violates the
𝑁
∑
constraint of (53). Because 𝑐𝑛 decreases as 𝑛 increases for
𝑑p (𝑟) = inf 𝑐𝑛 𝛼𝑛
𝑛 ≤ 𝑁 − 1, all the points with index less than 𝑃 are “active”. 𝛼𝑛
𝑛=1
According to (42), subject to 0 < 𝛼1 ≤ 𝛼2 ≤ ⋅ ⋅ ⋅ ≤ 𝛼𝑁 < 1
{ 𝑁
∑𝑛 ∑
(𝑀𝑡 + 𝑀𝑟 − 𝑛)𝑛 for 1 ≤ 𝑛 ≤ 𝑁 − 1 𝛼𝑛 > 𝑁 − 𝑟. (62)
𝑐𝑖 = (55)
𝑖=1
𝑀𝑡 𝑀𝑟 for 𝑛 = 𝑁 . 𝑛=1

The only difference between (62) and (43) is that here Follow-
Therefore the edge of the convex hull, i.e., inf 𝛾1 ,𝜇 𝜇 + 𝛾1 𝑟, is ing the arguments similar to those in the proof of Theorem 4.1,
obtained by connecting the 𝑃 + 2 points with coordinates
we can show that for the general fading channel, the optimal
𝑃 D-M tradeoff of the pruned channel is also a piecewise linear
{(𝑛, (𝑀𝑡 + 𝑀𝑟 − 𝑛)𝑛)}𝑛=0 , (𝑁, 𝑀𝑡 𝑀𝑟 ). (56)
curve obtained by connecting the following 𝑃 + 2 points
{ }𝑃
Since 𝑑p (𝑟) = 𝑀𝑡 𝑀𝑟 − (inf 𝛾1 ,𝜇 𝜇 + 𝛾1 𝑟) and 𝑀𝑡 𝑀𝑟 − (𝑀𝑡 + 𝑡𝑀𝑡 𝑀𝑟
𝑀𝑟 − 𝑛)𝑛 = (𝑀𝑟 − 𝑛)(𝑀𝑡 − 𝑛), we conclude that 𝑑p (𝑟) is 𝑛, (𝑀𝑟 − 𝑛)(𝑀𝑡 − 𝑛) + 𝛿(𝑛) , (𝑁, 0), (63)
2 𝑛=0
also a piecewise linear curve obtained by connecting
where {
𝑃 1 for 𝑛 = 0
{(𝑛, (𝑀𝑡 − 𝑛)(𝑀𝑟 − 𝑛))}𝑛=0 , (𝑁, 0). (57) 𝛿(𝑛) =
0 for 𝑛 =
∕ 0,
To see that this D-M tradeoff is an upper bound for any AS and
strategy, it is sufficient to note that (𝑀𝑟 −𝑝)(𝑀𝑡 −𝑝)+
𝑡𝑀𝑡 𝑀𝑟
𝛿(𝑝)
𝑃 = arg min𝑝 2
(64)
 𝜌 
𝑁 −𝑝
 subject to 0 ≤ 𝑝 ≤ 𝑁 − 1, 𝑝 ∈ ℤ.
𝐼(Λ𝑁 , 𝜌) ≥ log I + H𝒮𝑟 ,𝒮𝑡 H∗𝒮𝑟 ,𝒮𝑡 
𝑁
Consider an iid channel with the same dimension as the one
holds for any 𝒮𝑟 , 𝒮𝑡 with ∣𝒮𝑟 ∣ = 𝐿𝑟 and ∣𝒮𝑡 ∣ = 𝐿𝑡 , where in Figure 3 but with distribution given in (58) where 𝑡 = 2,
𝐼(Λ𝑁 , 𝜌) is defined in (16). We have proved the theorem. which corresponds to a Nakagami-m channel with 𝑚 = 2
5246
D−M Tradeoff, Mr = 5, Mr = 6, t = 2
60
50
Diversity Gain d(r)
40
30
20
10
0 IEEE Trans. Inform. Theory, vol. 50, pp. 968-985, June 2004.
0 1 2 3
Spatial Multiplexing Gain r
Fig. 5. Optimal D-M tradeoffs of full and pruned non-Rayleigh MIMO
channels whose entries have distribution (58) with 𝑡 = 2.
[[30], Table I]. Figure 5 shows that the optimal D-M gain
tradeoff of the pruned MIMO channels. Comparing Figures
3 and 5, we see that for 𝑁 = 2, the pruned non-Rayleigh
channel has better D-M tradeoff than Rayleigh channel even
for 𝑟 ∈ (1, 2).
We conclude this section by emphasizing that the CDA
based STBCs [8] with minimum delay 𝑇 = 𝐿𝑡 achieve the
optimal D-M gain tradeoff of the RF chain limited (or pruned)
channel since they are approximately universal and hence D-M
tradeoff-optimal in any fading channel [18] including the one
obtained via the optimal antenna selection or the fast antenna
selection algorithm of this paper.
V. C ONCLUSION
The conclusion is given at the end of [15].
A PPENDIX
We analyze the numbers of complex multiplication involved
in the proposed AS algorithm. At the 𝑖th step, the AS
algorithm compares the column norms of an (𝑀𝑟 − 𝑖 + 1) ×
(𝑀𝑡 − 𝑖 + 1) submatrix, which requires (𝑀𝑟 − 𝑖 + 1)(𝑀𝑡 −
𝑖 + 1) multiplications. Calculating the Householder Q𝑖 and
left multiplying the channel matrix by Q𝑖 requires about
3(𝑀𝑟 − 𝑖 + 1)+ 4(𝑀𝑟 − 𝑖 + 1)(𝑀𝑡 − 𝑖 + 1) multiplications (see,
e.g., [31, Sections 5.1.3 and 5.1.4]). Hence the total number
of complex multiplications in the transmit AS is
𝐿𝑡
∑ Inform. Theory Application Colloquium, Feb. 2007.
[5(𝑀𝑟 − 𝑖 + 1)(𝑀𝑡 − 𝑖 + 1) + 3(𝑀𝑟 − 𝑖 + 1)]
𝑖=1
1 (
= 𝐿𝑡 30𝑀𝑟 𝑀𝑡 + 10𝐿2𝑡 − 15𝐿𝑡 𝑀𝑡 − 15𝐿𝑡 𝑀𝑟 + 33𝑀𝑟
6
+15𝑀𝑡 − 24𝐿𝑟 + 14) . (65)
We see that the computational complexity of our proposed
AS algorithm is of order 𝑂(𝐿𝑡 𝑀𝑟 𝑀𝑡 ), which is similar
to the algorithm proposed in [32] which has complexity
𝑂((𝑀𝑟 + 𝑀𝑡 )𝑀𝑡 𝐿𝑡 ). Clearly, the computational complexity
of receiving antennas selection after transmit antenna selection
is of order 𝑂(𝐿𝑟 𝑀𝑟 𝐿𝑡 ).
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 10, OCTOBER 2009
R EFERENCES
min(Lr, Lt)=4 [1] I. E. Telatar, “Capacity of multiple antenna Gaussian channels," AT&T
min(Lr, Lt)=3 Tech. Memorandum, June 1995.
[2] G. J. Foschini, Jr., “Layered space-time architecture for wireless commu-
min(Lr, Lt)=2 nication in a fading environment when using multi-element antennas,"
min(Lr, Lt)=1 Bell Labs Tech. J., vol. 1, pp. 41-59, Autumn 1996.
Full System [3] G. J. Foschini, G. D. Golden, R. A. Valenzuela, and P. W. Wolniansky,
“Simplified processing for high spectral efficiency wireless communica-
tion employing multiple-element arrays," Wireless Personal Commun.,
vol. 6, pp. 311-335, Mar. 1999.
[4] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for
high data rate wireless communications: performance criterion and code
construction," IEEE Trans. Inform. Theory, vol. 44, pp. 744-765, Mar.
1998.
[5] L. Zheng and D. Tse, “Diversity and multiplexing: a fundamental
tradeoff in multiple-antenna channels," IEEE Trans. Inform. Theory,
vol. 49, pp. 1073-1096, May 2003.
[6] H. E. Gamal, G. Caire, and M. O. Damen, “Lattice coding and decoding
achieve the optimal diversity-multiplexing tradeoff of MIMO channels,"
4 5
[7] F. Oggier, G. Rekaya, J.-C. Belfiore, and E. Viterbo, “Perfect space-time
block codes," IEEE Trans. Inform. Theory, vol. 52, pp. 3903-3912, Sept.
2006.
[8] P. Elia, K. R. Kumar, S. A. Pawar, P. V. Kumar, and H.-F. Lu,
“Explicit construction of space-time block codes achieving the diversity-
multiplexing gain tradeoff," IEEE Trans. Inform. Theory, vol. 52,
pp. 3869-3884, Sept. 2006.
[9] A. Molisch and M. Win, “MIMO systems with antenna selection," IEEE
Microwave Mag., vol. 5, pp. 46-56, Mar. 2004.
[10] S. Sanayei and A. Nosratinia, “Antenna selection in MIMO systems,"
IEEE Commun. Mag., pp. 68-73, Oct. 2004.
[11] I. Bahceci, T. M. Duman, and Y. Altunbasak, “Antenna selection for
multiple-antenna transmission systems: performance analysis and code
construction," IEEE Trans. Inform. Theory, vol. 49, pp. 2669-2681, Oct.
2003.
[12] H. Shen and A. Ghrayeb, “Analysis of the outage probability for MIMO
systems with receive antenna selection," IEEE Trans. Veh. Technol.,
vol. 55, pp. 1435-1441, July 2006.
[13] A. Gorokhov, D. A. Gore, and A. J. Paulraj, “Receive antenna selection
for MIMO flat-fading channels: theory and algorithms," IEEE Trans.
Inform. Theory, vol. 49, pp. 2687-2696, Oct. 2003.
[14] H. Zhang, H. Dai, Q. Zhou, and B. L. Hughes, “On the diversity
order of spatial multiplexing systems with transmit antenna selection: a
geometrical approach," IEEE Trans. Inform. Theory, vol. 52, pp. 5297-
5311, Dec. 2006.
[15] Y. Jiang and M. Varanasi, “The RF-chain limited MIMO system with
antenna selection—part II: diversity-multiplexing tradeoff of complexity
constrained architectures," IEEE Trans. Wireless Commun., submitted,
2008.
[16] N. Prasad and M. K. Varanasi, “Outage theorems for MIMO block
fading channels," IEEE Trans. Inform. Theory, vol. 52, pp. 5284-5296,
Dec. 2006.
[17] L. Zhao, W. Mo, Y. Ma, and Z. Wang, “Diversity and multiplexing
tradeoff in general fading channels," IEEE Trans. Inform. Theory,
vol. 53, pp. 1549-1557, Apr. 2007.
[18] S. Tavildar and P. Viswanath, “Approximately universal codes over slow
fading channels," IEEE Trans. Inform. Theory, vol. 52, pp. 3233-3258,
July 2006.
[19] T. Kiran and B. S. Rajan, “STBC-schemes with nonvanishing deter-
minant for certain number of transmit antennas," IEEE Trans. Inform.
Theory, vol. 51, p. 2984–2992, Aug. 2005.
[20] T. L. Marzetta, “The case for many (>16) antennas at the base station,"
[21] A. Gorokhov, D. A. Gore, and A. J. Paulraj, “Receive antenna selection
for MIMO spatial multiplexing: theory and algorithms," IEEE Trans.
Signal Processing, vol. 51, pp. 2796-2807, Nov. 2003.
[22] A. Molisch, M. Win, J. H. Choi, and Y. S. Winters, “Capacity of MIMO
systems with antenna selection," IEEE Trans. Wireless Commun., vol. 4,
pp. 1759-1772, July 2005.
[23] Y. Jiang and M. Varanasi, “Spatial multiplexing architectures with
jointly designed rate-tailoring and ordered v-blast decoding—part I:
diversity-multiplexing trade-off ananlysis," IEEE Trans. Wireless Com-
mun., vol. 8, pp. 3252-3261, July 2008.
[24] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. Cambridge
University Press, 1991.
[25] A. Marshall and I. Olkin, Inequalities: Theory of Majorization. New
York: Academic, 1979.
JIANG and VARANASI: THE RF-CHAIN LIMITED MIMO SYSTEM: PART I OPTIMUM DIVERSITY-MULTIPLEXING TRADEOFF
[26] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University
Press, 1985.
[27] Y. Jiang, W. Hager, and J. Li, “Tunable channel decomposition for
MIMO communications using channel state information," IEEE Trans.
Signal Processing, vol. 54, pp. 4405-4418, Nov. 2006.
[28] R. J. Muirhead, Aspects of Multivariate Statistical Theory. New York:
Wiley Series in Probability and Statistics, 1982.
[29] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge Uni-
versity Press, 2002.
[30] L. Zhao, W. Mo, Y. Ma, and Z. Wang, “Diversity and multiplexing
tradeoff in general fading channels," Conf. Inform. Sciences Systems,
Mar. 2006.
[31] G. H. Golub and C. F. Van Loan, Matrix Computations. Baltimore, MD:
Johns Hopkins University Press, 1983.
[32] M. Gharavi-Alkhansari and A. B. Gershman, “Fast antenna subset
selection in MIMO systems," IEEE Trans. Signal Processing, vol. 52,
pp. 339-347, Feb. 2005.
5247
Yi Jiang received the B.S. degree in electri-
cal engineering and information science from the
University of Science and Technology of China
(USTC), Hefei, China, in 2001. He received the M.S.
and Ph.D. degrees from the University of Florida,
Gainesville, both in electrical engineering, in 2003
and 2005, respectively. In the Summer of 2005, he
was a Research Consultant with Information Science
Technologies Inc. (ISTI), Fort Collins, Colo. From
September 2005 to May 2007, he worked as a Post-
doc with the University of Colorado, Boulder. He is
currently with Corporate R&D, Qualcomm, Inc., San Diego, CA. His research
interests are in the areas of signal processing, wireless communications, and
information theory.
Mahesh K. Varanasi (S’87–M’89–SM’95) received
the Ph.D. degree in electrical engineering from Rice
University, Houston, TX, in 1989. He joined the
Electrical and Computer Engineering of the Univer-
sity of Colorado at Boulder in 1989 as an Assistant
Professor where he was later an Associate Professor
during 1996-2001 and is now a Professor since 2001.
His research and teaching interests are in the areas
of communication and information theory, wireless
communication and coding, detection and estimation
theory, and signal processing. He has published on
a variety of topics in these fields and is a Highly Cited Researcher according
to the ISI Web of Science. He is currently serving as an Editor for the IEEE
T RANSACTIONS ON W IRELESS C OMMUNICATIONS .