1880
5, MAY 2013
Computationally Efficient Vector Perturbation
Precoding Using Thresholded Optimization
C. Masouros, Member, IEEE, M. Sellathurai, Senior Member, IEEE, and T. Ratnarajah, Senior Member, IEEE
AbstractWe propose a low-complexity vector perturbation
(VP) precoding scheme for the downlink of multi-user multiple
input multiple output (MU-MIMO) systems. While conventional
VP performs a computationally intensive sphere search through
multiple candidate perturbation vectors to minimize the norm of
the precoded signal, the proposed precoder applies a threshold
to the desired norm to reduce the number of search nodes
visited by the sphere encoder. This threshold is determined by
the performance requirements of the mobile users. Once the
threshold is met, the search for the perturbation vectors finishes
thus saving significant computational burden at the transmitter.
To evaluate the advantages of the proposed technique compared
to VP, we further derive the computational complexity in terms
of the volume of the associated search space and the resulting
numerical operations. In addition, we use a new performance-
complexity metric to study the relevant tradeoff and look at the
power efficiency of the system, both of which metrics can be
used to optimize the user-determined threshold accordingly. The
presented analysis and results show that the proposed thresh-
olded VP (TVP) offers a favorable tradeoff between performance
and complexity where significant complexity reduction is attained
while the user threshold performance is guaranteed.
Index TermsVector perturbation, sphere encoding, complex-
ity reduction, multi-user MIMO, non-linear precoding.
I. I NTRODUCTION
HE need to produce power- and cost-efficient mobile de-
T vises has recently stimulated interest towards precoding
schemes for the multiple input multiple output (MIMO) down-
link transmission. Simple forms of precoding schemes already
appear in communication standards such as the 3GPP Long
Term Evolution (LTE) [1] and are expected to dominate future
implementations of telecommunications networks. Capacity
achieving non-linear dirty paper coding (DPC) techniques
[2], [3] have been proposed for pre-subtracting interference
prior to transmission. The DPC methods developed so far
are in general complex as they require sophisticated sphere-
search algorithms [6] to be employed at the transmitter and
Manuscript received August 25, 2012; revised November 19, 2012. The
associate editor coordinating the review of this paper and approving it for
publication was A. Ghrayeb.
C. Masouros is with the Dept. of Electrical and Electronic Eng., Univer-
sity College London, Torrington Place, London, WC1E 7JE, UK (e-mail:
chris.masouros@ieee.org).
M. Sellathurai is with the Department of Electrical, Electronic
and Computer Eng., Heriot-Watt University, Edinburgh, UK (e-mail:
mathini@ieee.org).
T. Ratnarajah is with the Institute for Digital Communications, The Uni-
versity of Edinburgh, The Kings Buildings, Mayfield Road, Edinburgh EH9
3JL, UK (e-mail: t.ratnarajah@ieee.org).
The work of C. Masouros was supported by the Royal Academy of
Engineering, UK.
The work of T. Ratnarajah was supported by the Seventh Framework
Programme (FP7) for Research of the European Commission under grant
HARP-318489.
Digital Object Identifier 10.1109/TCOMM.2013.022713.120632
0090-6778/13$31.00 
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO.
assume codewords with infinite length for the encoding of the
data. Their suboptimal counterparts [4], [5] offer a complexity
reduction at a comparable performance. Still however, the
associated complexity is prohibitive for their deployment in
current communication standards.
On the other hand, linear precoding schemes based on
channel inversion [7]-[10] offer the least complexity, but their
performance is far from achieving the optimum maximum
likelihood bound. Their non-linear adaptation, namely vector
perturbation (VP) precoding [11] provides a performance
improvement by introducing perturbation vectors at the trans-
mitter to better align the data symbols to the eigenvalues
of the inverse channel matrix on an instantaneous basis.
This results in much improved transmit scaling factors and
enhanced receive signal to noise ratios (SNRs) compared to
linear precoding. There are however, three main challenges
associated with VP precoding [12]:
The improved performance comes at the expense of an
increased complexity since the search for the optimal
perturbation vectors is an NP-hard problem, typically
solved by sphere search algorithms at the transmitter.
The power scaling factors at the transmitter are data de-
pendent and, unless an average scaling factor is agreed in
advance, this results in significant transmission overhead
to feedforward the instantaneous scaling factors, required
at the receiver for decoding.
The large dynamic range of the signal at the receiver can
cause the analog receive components to operate in their
non-linear region, thus leading to additional signal noise
and decoding errors.
The above challenges have been the focus of recent work
on VP precoding. The complexity of various sphere search
techniques has been studied in [13]-[16] (among others) in
terms of search nodes visited and search lattice volumes. A
number of techniques have been proposed towards reducing
the complexity of VP precoding (e.g. [17]-[20]). All the above
designs achieve the complexity reduction at the expense of an
inferior performance compared to conventional VP. In [17] a
search over a reduced lattice is proposed, based on empirical
observations of the relation between the instantaneous symbols
and the optimum perturbation vectors. This results in reduced
complexity while at the same time it is shown to alleviate
the modulo loss typically observed in VP designs. Further
work in [18] has proposed the decoupling of the perturbation
optimization in the real and imaginary domain of the data
symbols thus offering a lower complexity compared to the
joint optimization approach.
The received signal dynamic range relative to power scaling
has been the focus of numerous works (e.g. [12], [21]). In [12]
c 2013 IEEE
MASOUROS et al.: COMPUTATIONALLY EFFICIENT VECTOR PERTURBATION PRECODING USING THRESHOLDED OPTIMIZATION
the problem of the increased dynamic range of the signal at
the receiver due to the perturbation vectors is addressed. A
thresholded approach is proposed where, if the norm of the
transmit signal exceeds a predetermined value, a transmission
outage is imposed. The fact that the receiver has to detect the
non-transmission of data imposes an increase in the observed
error rates.
In this paper we adopt a similar thresholded approach
where, instead of a system-level outage threshold, we apply
a search-termination threshold directly to the sphere encoder.
In contrast to [12], the aim here is to reduce the associated
precoding complexity by reducing the number of nodes that
need to be visited by the sphere encoder, for a given target
performance. The threshold value is directly determined by the
operational SNR requirement of the mobile users. To cater for
the clipping noise at the receiver due to the dynamic range of
the received signal we also limit the search to low perturbation
amplitudes. To do this we adapt the Schnorr-Euchner search
(SE) algorithm [15],[22] by introducing a threshold on the
required weight (precoded signal norm) of the optimal node.
If this threshold weight is met, the search is terminated at a
reduced number of visited nodes. If the threshold can not be
met, then the minimum available weight is chosen, in which
case a standard full sphere search has been carried out.
It should be noted that compared to previous works on
complexity reduction for VP [17]-[20], this work explores
a distinct new approach based on the thresholded search as
opposed to a full search. The proposed can be applied on
top of other complexity reduction techniques to further reduce
complexity compared to the case where the full sphere search
is carried out. Since the proposed involves a thresholded
approach where on average the search terminates early, com-
plexity gains will still be observed. However, to keep the focus
of this work on the central idea we only use conventional VP
with a limited search lattice, proposed in [17] as the reference
technique. For reasons of clarity we list the contributions of
the present work:
1) A thresholded approach to the optimization of VP pre-
coding is discussed, based on a receive-SNR threshold,
to greatly reduce the complexity of VP precoding,
2) A new sphere search algorithm is introduced to support
the thresholded VP optimization,
3) The complexity of conventional and proposed tech-
niques is calculated and compared analytically, and the
complexity benefits of the proposed approach are proven
mathematically, based on the volume of the search space
associated with each of the techniques,
4) We quantify the performance-complexity tradeoff of
conventional and proposed VP, by introducing a joint
sum rate and complexity metric, and further compare
the power efficiency of these techniques. Either of these
metrics can be used by the mobile users to determine
the value of the threshold coefficient introduced in the
proposed VP.
The rest of the paper is organized as follows. Section II
illustrates the MIMO channel model used in this paper and
briefly describes conventional vector perturbation precoding.
In Section III the proposed thresholded vector perturbation
is presented. Section IV discusses the thresholded search
1881
algorithm used for the proposed precoder and presents a
complexity analysis. Analytical sum rate expressions are given
in Section V, where a new metric to evaluate the performance-
complexity tradeoff is introduced. The power efficiency is dis-
cussed in Section VI. Finally numerical results are illustrated
and discussed in Section VII and conclusions are summarized
in Section VIII.
In the following we use bold upper case notation to denote
matrices, bold lower case to denote vectors and regular lower
case for scalars. H = HH (HHH )1 denotes the Moore-
Penrose generalized inverse [23] of matrix H and HH is
the Hermitian transpose of matrix H. mod [x] denotes the
complex modulo operation with base . (x), (x) denote
the real and imaginary parts of the complex number x, x
denotes the rounding of x to the closest integer, x denotes
the maximum integer less or equal to x and j = 1.
QR(.) denotes the QR-decomposition of a matrix and sign(x)
denotes the sign of x. Finally, E{.} denotes the expectation
operator.
II. MU-MIMO C HANNEL M ODEL AND V ECTOR
P ERTURBATION P RECODING
A. MIMO Downlink Channel Model
We consider a downlink system comprised of a single base
station (BS) equipped with N transmit antennas and M N
users with a single receive antenna each. The received signals
of all antennas can be represented by the vector
r=Hx+w (1)
Here r CM1 and H CMN CN (0, IM IN )
contains the frequency flat fading channel coefficients with the
(m, n)-th element hm,n being the complex-Gaussian channel
tap between the n-th transmit and the m-th receive antenna.
Also, x CN 1 is the vector with the symbols transmitted
by the N transmit antennas and w CM1 CN (0, 2 IM )
is the vector of the additive white Gaussian noise (AWGN)
components at the M receive antennas. For the purposes of
the analysis below, full knowledge of channel state information
(CSI) is assumed at the BS transmitter, which is a common
assumption for precoding schemes.
B. Vector Perturbation (VP) Precoding
VP precoding employs a channel inversion precoding matrix
and applies a perturbation on the transmitted symbols such
that the useful signal power at the receiver is maximized. The
system block diagram is given in Fig. 1. The transmitted signal
is given by [11]

P
x= H (u + l ) (2)
where u CM1 is the data symbol vector,
= ||H (u + l )||2 (3)
2
is the transmit power scaling factor so that E{||x|| } = P
and l CM1 is the selected perturbation vector with
integer entries that will be detailed in the following. Also
= 2|c|max + , where |c|max is the absolute value of the
1882 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 5, MAY 2013

w1
to find the absolute minimum of the transmit signal norm in
mod y1
P (7), if perturbations that satisfy the threshold performance are
u H P available. Hence, the optimization algorithm need not perform
l*
H
P
mod y2
a full sphere search and can be terminated once the first vector
w2
Sphere
Search
l that satisfies the predetermined threshold is found.
BS Precoding mod yM
Assume that the received SNR excluding the modulo
P
wM
operation is required to be greater or equal to a certain SNR
threshold t
Eb P
Fig. 1. Block diagram of a VP MU-MIMO downlink. = t (8)
N0
where Eb , N0 are the per bit power and noise power spectral
constellation symbol with the maximum magnitude and de- density respectively. This translates to a weight threshold for
notes the minimum Euclidean distance between constellation the sphere encoder as
symbols. For the special case of channel inversion precoding
Eb P
[7] we have l = 0M , where 0M is the all zero M 1  t (9)
vector. Based on the above expressions the received symbol N 0 t
vector can be calculated as It is therefore intuitive to adapt the sphere encoder such
 that the first perturbation vector that is found to satisfy this
P weight threshold is selected and the search is terminated.
r= (u + l ) + w (4)
In the case when the instantaneous channel is such that the
threshold can not be met, the perturbation corresponding to
At the receiver, as shown in Fig. 1, the signal is first scaled
the minimum weight is selected according to the conventional
back to eliminate the effect of the transmit scaling factor and
sphere encoding process. The selected perturbation vector is
then fed to a modulo operator to remove the perturbation
then given as
quantity l . The output of the modulo stage is given as 
     lt , if lt : ||H (u + lt )||2 t
l =

arg min||H (u + l)||2 , otherwise (10)
y = mod r = mod u + l + 
w = u+n lA
P P
(5) Note that, as opposed to (7) where the search lattice can
Where mod [x] = f ((x)) + j f ((x)) and have infinite dimensions, in (10) we constrain the search
  lattice as A = { r + j i | r , i {1, 2, ... }}, where
x + /2
f (x) = x (6) is the bound of the complex integer lattice. This is done

to alleviate clipping noise errors at the receiver due to high
In the above, vector n in (5) denotes the equivalent noise perturbation values as explained in [12]. Apart from the above
vector at the receiver after the scaling and modulo operation. modification on the optimization criterion, the transmit and
receive processing for TVP is identical to the one of VP given
C. Perturbation Vector Selection in (2)-(6).
It should be noted that, to avoid a large transmission
To maximize the signal component in the received symbols
overhead for the feedforwarding of the scaling factor to the
or equivalently minimize the noise amplification, the pertur-
receiver, average scaling factors E{} based on averaging
bation vectors l should be chosen such that is minimized
in (3) over a block of symbols can be employed for both VP
in (5). Hence we have
and TVP. This is the approach adopted in the simulations that
are shown in the following.
l = arg min ||H (u + l)||2 (7) To get a first impression of how the thresholded opti-
lZM +jZM
mization performs, the probability distribution of the attained
For complex symbol alphabets, the optimization in (7) is
scaling factors from (3) for the proposed TVP optimization
a 2M -dimentional real integer lattice problem, known to be
compared to VP is shown in Fig. 2 for transmit SNR=20dB
NP-hard. Sphere search techniques are typically employed to
and various values of the receive SNR threshold t . The
solve the minimization, the complexity of which is known to
dashed lines in the figure represent the resulting weight thresh-
grow exponentially with M . A computationally efficient and
old t on the scaling factor in (9). A leakage of the distribution
flexible implementation of sphere search is the SE algorithm
over the threshold values can be seen, which occurs in badly
[22], which we adopt in this paper and modify for the purposes
conditioned channels when the threshold cannot be met and
of the proposed VP design.
l is chosen according to the second line in (10). This leakage
increases with increasing values of t as the optimization
III. P ROPOSED T HRESHOLDED V ECTOR P ERTURBATION becomes more tight and thus t in (10) likely to be unfulfilled.
(TVP) P RECODING A characteristic example of this is graphically illustrated
The idea behind the proposed approach is based on the in Fig. 3 where the optimization of as a function of the
fact that typically, the precoder is operating under a threshold real and imaginary parts of the perturbation vector l is shown
performance requirement from the mobile users (most com- for an instantiation of a badly conditioned single input single
monly a receive SNR threshold). In this case there is no need output (SISO) channel. The figure also includes the surfaces
MASOUROS et al.: COMPUTATIONALLY EFFICIENT VECTOR PERTURBATION PRECODING USING THRESHOLDED OPTIMIZATION 1883

TABLE I
T HRESHOLDED S CHNORR -E UCHNER A LGORITHM
VP
M, H u, H , t
1000
Input:
500
Output: l , w
0
0 5 10 15 20 25 30 35 40 1. [Q, R] := QR(H ) /QR decompose input/
1000
TVP, =5dB, =31.6228
2. x := QH u
t t
3. w := /initialize weight record/
500
4. k := M /examined layer/
5. dk := 0 /distance to examined layer/
0
0 5 10 15 20 25 30 35 40 6. ek := Rx
1000
TVP, =10dB, =10
7. lk := ek,k  /examined lattice point/
e lk
8. y := k,k
t t
500 rk,k
9. stepk := sign(y) /offset to next layer/
0
0 5 10 15 20 25 30 35 40 10. while k M and w > t do /thresholded search/
1000
TVP, =12dB, =6.3096
11. w := dk + y 2
500
t t
12. if w w then
13. if k = 1 then /not a leaf node/
0 14. for i = 1, . . . k 1 do
1000
0 5 10 15 20 25 30 35 40 15. ek1,i := ek,i y rk,i
TVP, =15dB, =3.1623 16. endfor
k := k 1
t t
500 17. /move down/
18. dk := w
0 19. lk := ek,k 
0 5 10 15 20 25 30 35 40 e lk
20. y := k,krk,k
Fig. 2. Probability distribution of for increasing t , SNR=20dB, N = 21. stepk := sign(y)
M = 4, 4QAM. 22. else /a leaf node/
23. l := l /best lattice point so far/
24. w := w /update weight record/
25. k := k + 1 /move up/
26. lk := lk + stepk /next layer/
e lk
27. y := k,krk,k

10
2 28. stepk := stepk sign(y)
29. endif
30. else
31. if k == M then /returned to root node/
32. return l
33. else
2
log10||H (u+ l)||

t=5dB 34. k := k + 1 /move up/

1
10 35. lk := lk + stepk /next layer/
+

e lk
36. y := k,krk,k
t=10dB
37. stepk := stepk sign(y)
38. endif
39. endif
t=15dB 40. endwhile
0
10
2 3
0 1 2
1 0
2 2
3
Im(l) Re(l)
IV. T HRESHOLDED S PHERE S EARCH A LGORITHM AND
Fig. 3. Optimization of with (l), (l), SNR=20dB, N = M = 1,
C OMPLEXITY A NALYSIS
4QAM.
A. Thresholded Schnorr-Euchner Algorithm
To obtain the optimization in (10) the associated sphere
search algorithm itself needs to be modified. In the following,
we illustrate the modified SE algorithm used in this work,
corresponding to different receive SNR thresholds t . It can where the weight threshold t has been introduced as a
be seen that for the case of t = 15dB the resulting weight termination condition. To save space we only briefly explain
threshold is below the minimum of the optimization, in which the proposed algorithm, using terminology from the sphere
case (10) will result in a full sphere search. This represents decoding literature. For a more detailed analysis of the SE
the leakage cases in Fig. 2. For the cases of t = 10dB, algorithm and the concept of sphere search, the reader is
t = 5dB the weight threshold is satisfied and the areas above pointed to [15] and references therein.
the thresholds represent weights of the candidate lattice points The proposed thresholded SE algorithm is shown in Table I,
to be visited in the sphere encoder. The areas below each closely following the structure in [15] to ease comparison.
surface correspond to lattice points which need not be visited Without loss of generality and to ease notation, we assume
in the search, which provides an initial measure of the saved real channel elements and symbol vectors, yielding a M -
complexity with different values of t for TVP. The resulting dimensional integer search lattice. For the complex channel
complexity is further analyzed in the following sections. and symbol vector case, the channel model can be decomposed
1884 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 5, MAY 2013

into its real and imaginary parts is shown for example in

[18] and the lattice becomes 2M -dimensional. Here, ek is
the M 1 vector ek = [ek,1 , . . . , ek,k ], k = 1, . . . , M and
R has elements rm,n . The algorithm performs a tree search
from the root level with index k = M to the leaf level with
k = 1. At each stage, the algorithm computes the distance
to the target vector H u. If the distance is smaller than the
recorded distance then the current layer is expanded into its
branches. If the distance is larger that the recorded distance
then this is not a preferred branch and the algorithm moves up
one layer to search the next branch. The algorithm is invoked
once it has moved all the way to the bottom layer (a leaf
node) without exceeding the lowest distance. In this case the
lattice point is stored as the best solution so far, the distance
Fig. 4. Geometrical representation of the 2-dimensional integer-lattice sphere
(weight) is updated and the algorithm moves up a layer. The search problem for conventional sphere search and thresholded sphere search.
central modification to the SE search [15] is at line 10: At the
point where the stored solution becomes less or equal to the
weight threshold t or the search comes to a root node (i.e. a
full search has been completed), the algorithm terminates.
+
VT SE = VM () VM ( t )
M +
B. Complexity Analysis - Volume of Search Space
It has been observed in [15] and references therein that the
complexity of the SE search is proportional to the volume
of the region being searched in the lattice space. At the k-
th search layer this volume is a hypersphere with maximum
radius k . A geometrical illustration of this is shown in Fig. 4
for a 2-dimensional lattice. Here the dots represent the lattice
points and the area inside the outer circle denotes the area of
candidate points searched, based on the search radius.
For the proposed thresholded SE (TSE), due to the applied
threshold, the search is carried out to the point where t is
satisfied and no further lattice points closer to the target point
need to be evaluated. The equivalent search space for the 2-
dimentional lattice is denoted by the shaded area in Fig. 4.
It is therefore clear that, the search volume associated with
TSE is the volume of the hypersphere withradius k less
the volume of the hypersphere with radius t . When the
termination threshold t is too strict and a full SE search
is carried out, this signifies that the equivalent volume of the
threshold hypersphere is zero, as it contains zero lattice points.
Note that, since the TSE terminates once the first lattice point
satisfying the threshold is found, the area shown in Fig. 4
consists of an upper bound of the actual volume searched for

M/2 M M/2
t (12)
(M/2 + 1) 2

where k = r1,1 2 + r2 + + r2 , k = 1, . . . , M , [x]+ =
2,2 k,k
max{0, x} and (.) denotes the gamma function.
Proof: The proof is provided in Appendix A
In the above theorem, the complexity reduction with TVP
has been shown by means of search space volume. A complex-
ity comparison between VP and TVP in terms of elementary
operations is shown analytically in the next section and
numerically in the results section in the following.
C. Complexity Analysis - Numerical Operations
The above analysis can be used to obtain an upper bound
on the expected numerical operations associated with the
proposed precoding. Note that, as the search volume is as-
sociated with the search tree complexity, the analysis below
focuses only on the operations required for the search stage,
ignoring the pre- and post-search operations such as the QR
decomposition etc. This is a common practice in the literature
[13]-[15].
It was shown in [14] that for an infinite lattice1 , the expected
number of lattice points contained inside a k-dimensional
hypersphere of radius is given as


TSE. Typically TSE will only search a part of this volume until  2
  2 k
the first lattice point on the surface of the inner hypersphere Ep k, = , rk (q) (13)
2( 2 + q) 2
is met. q=0

In the following, we use these key observations to attain a where  x

complexity evaluation for TSE, following the methodology of t1 t
(x, ) = e dt (14)
[14], [15]. 0 ()
Theorem 1: Denote the search volume for the k-th search is the normalized incomplete gamma function and r (x)
layer with radius k as Vk (k ). For the complexity CT SE denotes the number of ways a non-negative integer x can be
associated with TSE in a M -dimensional lattice we have
1 Note that in the results section we use a search lattice with finite limits. For

+ the case of a finite lattice, the subset of lattice points inside the hypersphere is
CT SE C VT SE = VM () VM ( t ) (11) dependent on the limits of the lattice and no generic closed form expression
for the number of lattice points exists [14]. We therefore look at the generic
case of an infinite lattice to attain a generic complexity expression at this
for which point.
MASOUROS et al.: COMPUTATIONALLY EFFICIENT VECTOR PERTURBATION PRECODING USING THRESHOLDED OPTIMIZATION
represented as a sum of squares of integers. The closed
form definition of r (x) is a classical number-theory problem,
well studied in the literature. While a generic closed form
expression for all does not exist, specific closed form
expressions have been derived for a number of values of
(see [14] and references therein). Indicatively, for the case of
= 4 explored in the following simulations, the solution is
given by Lagranges Four Squares Theorem [14] as

8 m if x odd

m|x
24 m if x even
r4 (x) = (15)
m|x,
m odd
where m|x denotes that m is a divisor of x.
Clearly, based on the above we can derive upper bounds for
expected numbers of nodes visited by SE and TSE.
Theorem 2: The expected number of lattice points visited
by the TSE search at the k-th layer is upper bounded by
1885
V. S UM R ATE - C OMPLEXITY T RADEOFF
In [24] the sum rate performance of VP was studied
analytically and rate lower bounds were derived. Here we use
these to evaluate the sum rates achieved with TVP and propose
a new metric that characterizes the performance-complexity
tradeoff. The results show that the proposed TVP offers a
better tradeoff to conventional VP.
Before moving to the tradeoff analysis, it should be high-
lighted here that the performance threshold used in TVP is
imposed by the mobile users, and this threshold is met by
the proposed scheme at a much reduced complexity. The
tradeoff here refers to the system configuration stage where
the users select the SNR threshold value t . At this stage
the performance-complexity tradeoff below (or the power
efficiency metric discussed in the next section) may or may
not be considered by the users. In any case, it is worth
clarifying that the proposed technique guarantees the required
performance at a lower complexity compared to conventional
VP.
  2

[k /4 t ]+ k
pT SE (k) = , rk (q) (16) A. Sum Rate Performance
q=0
2( 2 + q) 2
The sum rate of a VP system with uniformly distributed
inputs is given as [24]
Proof: The proof is provided in Appendix B  
The complexity in numbers of numerical operations can then Eb E{}
R = M log + 2M Eb
(21)
be calculated using the formula N0 eE{} 2N 0

M
where
C(M ) = p(k)fp (k) (17) 


1
1
2  1 |s|2  2
|t|
k=1
(x) = + e 2x e 2x d
with p(k) = pSE (k) for SE and p(k) = pT SE (k) for TSE, 2 12 s= 2x t=
yielding (22)

M  
is a function that relates to the effect of the receive modulo
2k /4 k
CSE fp (k) , rk (q)  CSE (18) operation [5]. For the high SNR region it was shown in [24]
q=0
2( 2 + q) 2 that  
k=1
E{}
and lim =0 (23)

M   2
Eb
N
2NEb
[k /4 t ]+ k 0 0
CT SE fp (k) , rk (q)  CT SE
q=0
2( 2 + q) 2 for which case a lower bound can be derived as
k=1
(19) Eb
R Rb = M log (24)
where fp (k) = 2k + 11 denotes the number of numerical N0 eE{}
operations pervisited node [14] in the k-th search layer.
Since VM ( t ) 0 and Ep (k, t ) 0, it can be con- B. Performance-Complexity Tradeoff
cluded that the proposed TVP has a strictly reduced complex- To evaluate the tradeoff between performance and com-
ity compared to VP, with a relative complexity improvement plexity for VP and TVP the above complexity and sum rate
lower bounded as expressions can be used to derive a tradeoff metric as
CSE CT SE Rb
= (25)
CSE
M   C


k=1 fp (k)
t k
q=0 2(2 +q) , 2 rk (q) where C denotes the average complexity in terms of numbers
M 
 2
/4
 , if t k
c= k k
k=1 fp (k) q=0 2(2 +q) , 2 rk (q) of operations and can be substituted by the upper bounds in


100%, otherwise (18),(19) for VP and TVP respectively to yield the expressions
(20) log N Ee
b

V P   
0
 (26)
 2 have infinite terms,
While the inner summations in (18)-(20), M 2k /4
2( +q) 2 rk (q)
k
/4 f p (k) ,
it can be seen that the terms 2(k2 +q) , k2 rk (q) and k=1 q=0 2

 
2(2t+q) , k2 rk (q) tend to zeros for large q. We therefore where = E{} for VP and
approximate these expressions using a sum limit of qmax = log N E b

100 when numerically calculating the complexity in terms of T V P   0 eT

 (27)
M [2k /4t ]+ k
numbers of operations in the results section. k=1 fp (k) q=0 2(2 +q) , 2 rk (q)
1886 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 5, MAY 2013

where T = E{} for TVP. The expression in (27) combined 10

6

with (9) can be used by the mobile subscribers to select t VP

such that the performance-complexity tradeoff is optimized 5
TVP, =15dB
t
10 TVP, =10dB
with TVP. In the results section we quantify this tradeoff by t

Number of floating point oprations

TVP, =5dB
means of simulation, using the above expressions. t

4
10

VI. P OWER E FFICIENCY

Next, we look at the transmit power efficiency of TVP 10
3

compared to VP, and its dependence on the SNR threshold t .

Following the modeling of [25] we define the transmit power 10
2

efficiency of the communication link as the total transmit

energy consumption required per bit i.e. ratio of the consumed
1
power at the transmitter over the sum rate achieved 10
2 3 4 5 6 7 8 9 10
Number of antennas N=M

PP A + N P0 + pc C
E= (28) Fig. 5. NOP vs. N = M for VP and TVP with t =5dB, 10dB, 15dB for
Rb
SNR=20dB, 4QAM.
where we have used the rate bound Rb from (24). In
(28), PP A = P in Watts is the power consumed at the 4x4, 4QAM
1400
power amplifier to produce the total transmit signal power P ,
with being the power amplifier efficiency and being the 1200

modulation-dependent peak to average power ratio (PAPR).

Number of floating point oprations

P0 = Pmix + Pf ilt + PDAC is the power related to the 1000

VP
mixers and transmit filters and the digital-to-analog converter TVP, =10dB
t
800
(DAC), assumed constant for the purposes of this work. We TVP, t=5dB
TVP, = 10dB
use practical values of these from [25] as = 0.35 and 600
t dB
TVP, = 15dB
t dB
Pmix = 30.3mW, Pf ilt = 2.5mW, PDAC = 1.6mW yielding
P0 = 34.4mW. Finally, pc in Watts/KOps is the power 400

per 103 elementary operations (KOps) of the digital signal 200

processor (DSP) and C is as discussed above. This term
is used to introduce complexity as a factor the transmitter 0
0 5 10 15 20 25 30

power consumption in the power efficiency metric. Typical Eb/N0

values of pc include pc = 22.88mW/KOps for the Virtex-4

and pc = 5.76mW/KOps for the Virtex-5 FPGA family from
Xilinx [26].
The expression in (28) provides a metric that combines
transmission rate, complexity and transmit signal power, all
in a unified expression. By varying the required SNR thresh-
old, both the resulting complexity and transmission rates are
influenced, as shown above. Therefore, this expression can be
used for the design of the communication link to maximize the
power efficiency by optimizing t . High values of E indicate
that high transmit power consumption is required to achieve a
target sum rate, and thus denote a low power efficiency. The
following results show that TVP provides an increased power
efficiency which converges to the one for VP for high values
of t .
VII. N UMERICAL R ESULTS
This section presents numerical results based on Monte
Carlo simulations of the proposed TVP and conventional VP
for the frequency flat Rayleigh fading statistically uncorrelated
MIMO channel whose impulse response is assumed perfectly
known at the transmitter. Without loss of generality it is
assumed that P = 1. A system employing 4QAM modulation
is explored for the case with N = M while it is clear that the
benefits of the proposed technique extend to non-symmetric
MIMO channels and higher order modulation. We focus on
the performance and complexity comparisons and tradeoff
Fig. 6. NOP vs. SNR for VP and TVP with t =5dB, 10dB, dB -10dB,
dB -15dB for N = M = 4, 4QAM.
between VP and TVP. As explained above and following the
methodology in the literature, here the complexity evaluation
focuses only on the complexity of the tree-search stage,
ignoring pre- and post-processing. For TVP we explore two
types of search termination thresholds t . One design based
on a fixed receive SNR threshold t and one based on a t
proportional to the transmit SNR (txSNR).
To avoid clipping noise errors, the search space is lim-
ited to low amplitude perturbations in the lattice A =
{ r + j i | r , i {1, 2, 3}} as in [12], [17] for both
conventional VP and the proposed TVP. Furthermore the
scaling factors of (3) are averaged over symbol blocks of
Nb = 100 symbols to reduce the transmission overhead for
the feedforwarding of . In the following we use the notation
Eb
dB = 10 log10 N 0
for convenience.
Figure 5 shows the complexity of VP and TVP in terms
of numbers of floating point operations (NOP) for increasing
numbers of antennas N = M , with dB = 20dB. At this
point exact NOP counts are shown based on simulation, while
the upper bounds derived in Section IV are compared to
simulation in following results. The cases of three different
receive SNR thresholds are observed t = 5dB, t = 10dB,
t = 15dB. It can be seen that for t = 15dB only a minor
MASOUROS et al.: COMPUTATIONALLY EFFICIENT VECTOR PERTURBATION PRECODING USING THRESHOLDED OPTIMIZATION 1887

10x10, 4QAM
6
10 1400

1200
5
10
Number of floating point oprations

VP
1000
TVP, t=10dB

Complexity (NOPs)
TVP, =5dB
t
4 TVP, = 10dB 800
10 t dB
TVP, t=dB15dB

600

3
10
400
VP
TVP
200 VPUB
2
10 TVPUB
0 5 10 15 20 25 30
Eb/N0
0
0 5 10 15 20

t
Fig. 7. NOP vs. SNR for VP and TVP with t =5dB, 10dB, dB -10dB,
dB -15dB for N = M = 10, 4QAM.
Fig. 8. NOP vs. t for VP and TVP with txSNR=20dB, N = M = 4,
4QAM.

complexity reduction is achieved, which accounts for the 4x4 10x10

cases with well conditioned channel matrices. The complexity 7.5 22

benefits however improve as the SNR thresholds diminish, 7 20

allowing for a more relaxed precoding optimization in (10). 6.5 18

In fact, while an exponential growth of complexity with M can 6

16
be seen for VP and TVP with high values of t , for t = 5dB 5.5
Sum Rate
14
an almost linear growth can be observed. 5
12
In Fig. 6 the complexity with increasing txSNR is shown 4.5
10
for VP and TVP with various weight thresholds, for a 4 4 4

3.5 8
MIMO (N = M = 4). The cases of fixed SNR thresholds VP VP
3 6 TVP
t = 5dB, t = 10dB and txSNR dependent thresholds TVP

t = dB 10dB,t = dB 15dB are shown. It can be 2.5

seen that for the transmit-SNR dependent thresholds a fixed
complexity reduction is observed for TVP compared to VP.
For t = dB 10dB,t = dB 15dB TVP entails a 33.67%
and 13.31% of the VP complexity respectively. For the fixed
threshold cases, at low txSNR values the required receive SNR
thresholds are infeasible and therefore the full shpere search
is run and no complexity gains can be gleaned. For higher
txSNR values however the complexity benefits increase to a
90% complexity reduction w.r.t. the VP complexity.
For the case with N = M = 10 the complexity is shown in
Fig. 7. Again the same conclusion can be drawn for the fixed
and txSNR dependent thresholds. It is worth noting however
the vast complexity benefits which reach a 99.7% reduction
for TVP compared to VP.
Next, we look at the effect of t on the complexity,
performance-complexity tradeoff and the power efficiency, in
Figs. 8-11 for the 44 and 1010 MIMO systems with 4QAM
modulation for txSNR=20dB. In Fig. 8 the complexity in
numbers of operations is shown for increasing values of t in
dB for the 44 MIMO. For high performance thresholds, TVP
has the same complexity as VP, since full SE search is carried
out as the weight threshold cannot be satisfied. Complexity
benefits can be observed as t reduces below 20dB. The theo-
retical complexity upper bounds are included for comparison
to the numerical results. The difference between theoretical
and simulated complexity, as mentioned above is due to the
fact that constant radius hyperspheres with the maximum
observed radius are considered in the analysis for both VP
and TVP (as opposed to the adaptive radius used in SE).
0 500 1000 1500
4
10
2
10
4 6
10
Complexity (NOPs)
Fig. 9. Sum Rate (bps/Hz) vs. NOP for VP and TVP with txSNR=20dB,
4QAM.
Furthermore, for TVP the bound becomes loose for decreasing
values of t . As explained earlier, this is because the volume
and number of lattice points calculated is an overestimate of
the ones visited, as the search terminates once the first lattice
point on the surface of the threshold hypersphere is met. A
similar trend can be proven for the 10 10 MIMO, but with a
much higher complexity gain as shown in Fig. 7. For reasons
of space the relevant graph is omitted at this point.
In Fig. 9 the sum rate lower bounds of (24) are shown
against increasing complexity by adjusting the threshold t ,
to illustrate the relevant tradeoff. The graphs represent the
values for TVP while the sum rate and complexity for VP
is denoted by the large circles in the plots. It can be seen
that for both 4 4 and 10 10 systems, the sum rates
improve drastically with increasing complexity at the low
complexity regions, while the benefits of added complexity
saturate at higher complexity values. These graphs can be used
for selecting the threshold values t according to the system
rate and complexity requirements in the design stage of the
algorithm.
Next, Fig. 10 shows the tradeoff between complexity and
performance in units of bps/Hz per 103 operations (KOps),
by means of the metric in (25). A better tradeoff between
performance and complexity can be seen for the proposed TVP
1888 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 5, MAY 2013

25
has been studied in terms of the volume of the search space
VP associated with the perturbation optimization and numbers of
TVP
elementary operations. Overall, the proposed scheme has been
20
proven to offer a greatly reduced complexity and a favorable
performance-complexity tradeoff and power efficiency com-
pared to VP, while securing the required performance for the
(bps/Hz/KOps)

15

mobile users.
4x4
10
A PPENDIX A: P ROOF OF T HEOREM I
In this section we provide the proof of Theorem 1. Firstly,
5
10x10 (11) is a direct expression of the fact that, for TSE, according
to Fig. 4 and the above discussion, the complexity upper
0
bound C is proportional to the volume of the hypersphere
0 5 10 15 20
k less the volume
with radius of the hypersphere with radius

t t . If t > k (i.e. the t -hypersphere contains the
k -hypersphere) then the search volume is zero: all lattice
Fig. 10. vs. t for VP and TVP with txSNR=20dB, 4QAM. points within the hypersphere with radius k satisfy the weight
threshold and a random lattice point can be chosen.
1.2 To obtain (12), we continue the proof with an overview of
the calculation for the complexity associated with conventional
1
SE, which gives us the volume of the k -hypersphere - the first
VP
TVP
term on the right side
of (11). We then adapt this methodology
Energy per bit (Watt/bps/Hz)

0.8
for the threshold t -hypersphere to obtain the second term
4x4 in the same expression.
It can be seen in Table I that the SE search starts with
0.6
an infinite search radius for k = M which is decreased at
each layer according to the minimum observed weight in
0.4
the previous layer. Hence, the aforementioned radius ak is
10x10 adaptive on a layer-by-layer basis. For the complexity CSE of
0.2
the conventional SE search in the M -dimensional lattice we
have [15]
0
0 5 10 15 20
t
CSE VSE = VM () (29)

Fig. 11. Energy per bit vs. t for VP and TVP with txSNR=20dB, 4QAM. For SE the radius of the k-th layer is upper-bounded
according to the algorithm in Table I as

for low values of t which converges to the tradeoff of VP for k1 2k yk2 , k = 2, . . . , M (30)
high t . Using this graph, the values of t can be optimized to Therefore, starting with the initial volume of a line spanning
maximize the performance-complexity tradeoff. It can be seen from 1 to +1 for layer k = 1
that for the 4 4 MIMO the value t = 2.5dB maximizes
the tradeoff, while for the 10 10 case the optimum value V1 (1 ) = 21 (31)
increases to t = 5dB. it is shown in [15] using the recursive volume expression
Finally, Fig. 11 shows the power efficiency of VP and TVP  k 
in terms of energy usage per information bit, based on (28) Vk (k ) Vk1 2 2
k y dy, k = 2, . . . , M (32)
for the same system setup. It can be seen that TVP has an k
improved power efficiency compared to VP as lower power that the volume of the search space for the k-th layer follows
needs to be consumed for the same target sum rate. The
k/2 k
k/2
optimal values of t for the depicted scenario are t = 5dB Vk (k ) n k (33)
for the 4 4 and t = 10dB for the 10 10 MIMO system. (k/2 + 1) n=1 (k/2 + 1) k
The second inequality in (33) stems from the fact that
VIII. C ONCLUSION the volume of a hypersphere with radius n in dimension
This paper has shown that under given user performance n = 1, . . . k is less than the volume of the hypersphere with
requirements, significant computational gains can be achieved constant radius k n , n = 1, . . . k in all dimensions.
for vector perturbation precoding. By introducing a search Using (33) and lattice theory we have the following upper-
termination threshold in the sphere search algorithm, based bound for the first term on the right side of (11)
on an SNR requirement from the users served, the proposed
k/2 k
n k/2
design achieves a notable complexity reduction. The complex- Vk () k k (34)
ity efficiency of both conventional and proposed techniques (k/2 + 1) n=1 2 2 (k/2 + 1) k
MASOUROS et al.: COMPUTATIONALLY EFFICIENT VECTOR PERTURBATION PRECODING USING THRESHOLDED OPTIMIZATION 1889

where k is defined as in the theoremabove. For the inner [6] B. Hassibi and H. Vikalo, On the expected complexity of integer least
hypersphere of Fig. 4 with radius 1 = t at the leaf level, squares problems, in Proc. 2002 IEEE Int. Conf. Acoust., Speech, Signal
Process., pp. 14971500.
similarly to (32), we have the following recursion [7] C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, A
 k 
vector-perturbation technique for near-capacity multiantenna multiuser

communicationpart I: channel inversion and regularization, IEEE
V k ( t ) = Vk1 2k y 2 dy, k = 2, . . . , M Trans. Commun., vol. 53, no. 1, pp. 195202, Jan. 2005.
k
(35) [8] C. Masouros and E. Alsusa, Dynamic linear precoding for the exploita-
tion of known interference in MIMO broadcast systems, IEEE Trans.
where k is the radius for the k-th layer of TSE, with starting Wireless Commun., vol. 8, no. 3, pp. 13961404, Mar. 2009.
volume

[9] C. Masouros, Correlation rotation linear precoding for MIMO broad-
cast communications, IEEE Trans. Signal Process., vol 59, no. 1, pp.
V 1 ( t ) = 2 t (36) 252262, Jan. 2011.
[10] C. Masouros and E. Alsusa, Soft transmitter precoding for the downlink
This recursion yields of DS/CDMA communication systems, IEEE Trans. Veh. Technol., vol
59, no. 1, pp. 203215, Jan. 2010.

k/2
 k
k/2 k/2 [11] C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, A
V k ( t ) = t n t vector-perturbation technique for near-capacity multiantenna multiuser
(k/2 + 1) n=2
(k/2 + 1) communicationpart II: perturbation, IEEE Trans. Commun., vol. 53,
(37) no. 3, pp. 537544, Mar. 2005.
In (37) the term on the right denotes the volume of a [12] J. Maurer, J. Jalden, D. Seethaler, and G. Matz, Vector perturbation
precoding revisited, IEEE Trans. Signal Process., vol. 59, no. 1, pp.
hypersphere with constant radius t which is the minimum 315328, Jan. 2011.
radius in the TSE recursion. Substituting (37) and (34) for [13] J. Jalden and B. Ottersten, On the complexity of sphere decoding in
k = M in (11) yields (12) which concludes the proof. digital communications, IEEE Trans. Signal Process., vol. 53, no. 4,
pp. 14741484, Apr. 2005.
[14] B. Hassibi and H. Vikalo, On the sphere-decoding algorithmI:
A PPENDIX B: P ROOF OF T HEOREM II expected complexity, IEEE Trans. Signal Process., vol. 53, no. 8, pp.
28062818, Aug. 2005.
In this section we provide the proof of Theorem 2. The [15] E. Agrell, T. Eriksson, A. Vardy, and K. Zeger, Closest point search in
lattices, IEEE Trans. Inf. Theroy, vol. 48, no. 8, pp. 22012214, Aug.
proof is a derivative of the discussions in Section IV. Using 2002.
(13) for SE, for the volume in (34) of a hypersphere with [16] L. G. Barbero, T. Ratnarajah, and C. Cowan, A comparison of complex
radius k in all dimensions we have the straightforward upper lattice reduction algorithms for MIMO detection, in Proc. 2008 IEEE
ICASSP.
bound [17] H. S. Han, S. H. Park, S. Lee, and I. Lee, Modulo loss reduction for
  vector perturbation systems, IEEE Trans. Commun., vol. 58, no. 12,
pSE (k) Ep k, 2k pp. 33923396, Dec. 2010.
 

2k /4 k [18] S. H. Park, H. S. Han, S. Lee, and I. Lee, A decoupling approach
= , rk (q)  pSE (k) (38) for low-complexity vector perturbation in multiuser downlink systems,
q=0
2( 2 + q) 2 IEEE Trans. Wireless Commun., vol. 10, no. 6, pp. 16971701, June
2011.
while for the associated volume with TSE based on (11), (34), [19] U. P. Rico, E. Alsusa, and C. Masouros, A fast least-squares solution-
seeker algorithm for vector-perturbation, in Proc. 2008 IEEE Globe-
(37) we get the upper bound Com.
   + [20] H. Song, S. Choi, and D. Hong, Line encoding for multiuser MIMO
pT SE (k) Ep k, 2k Ep (k, t ) downlink, IEEE Trans. Veh. Technol., vol. 60, no. 2, pp. 734-739, Feb.
  

+ 2011.
2k /4 k t k [21] F. Boccardi and G. Caire, The p-sphere encoder: peak-power reduction
= , , rk (q)
q=0
2( 2 + q) 2 2( 2 + q) 2 by lattice precoding for the MIMO Gaussian broadcast channel, IEEE
Trans. Commun., vol. 54, no. 11, pp. 20852091, Nov. 2006.
  2
[22] C. P. Schnorr and M. Euchner, Lattice basis reduction: improved prac-
[k /4 t ]+ k
= , rk (q)  pT SE (k) (39) tical algorithms and solving subset sum problems, Math Programming,
q=0
2( 2 + q) 2 vol. 66, pp. 181191, 1994.
[23] R. Penrose, A generalized inverse for matrices, Proc. Cambridge
Philosophical Society, vol. 51, pp. 406413, 1955.
The right hand side of the first line of (39) represents the lattice [24] D. J. Ryan, I. B. Collings, I. V. L. Clarkson, and R. W. Heath,
points contained
in the volume between the hyperspheres with Performance of vector perturbation multiuser MIMO systems with
radius k and t , when the latter is contained in the former limited feedback, IEEE Trans. Commun., vol. 57, no. 9, pp. 2633
2644, Sep. 2009.
hypersphere. [25] S. Cui, A. J. Goldsmith, and A. Bahai, Energy-constrained modulation
optimization, IEEE Trans. Wireless Commun., vol. 4, no. 5, pp. 2349
2360, Sep. 2005.
R EFERENCES [26] D. Curd, Power consumption in 65 nm FPGAs, Xilinx White Paper,
[1] 3GPP TS 36.211, V8.2.0 (2008-03), Evolved universal terrestrial radio Feb. 2007.
access (E-UTRA); physical channels and modulation, Release 8.
[2] M. Costa, Writing on dirty paper, IEEE Trans. Inf. Theory, vol. 29,
pp. 439441, May 1983.
[3] U. Erez, S. Shamai, and R. Zamir, Capacity and lattice strategies for
cancelling known interference, IEEE Trans. Inf. Theory, vol. 51, no.
11, pp. 38203833, Nov. 2005.
[4] C. Windpassinger, R. Fischer, T. Vencel, and J. Huber, Precoding
in multiantenna and multiuser communications, IEEE Trans. Wireless
Commun., vol. 3, no. 4, pp. 13051316, July 2004.
[5] C. Masouros, M. Sellathurai, and T. Rantarajah, Interference optimiza-
tion for transmit power reduction in Tomlinson-Harashima precoded
MIMO downlinks, IEEE Trans. Signal Process., vol. 60, no. 5, pp.
24702481, May 2012.
1890
Christos Masouros (M06) received his Diploma
in electrical and computer engineering from the
University of Patras, Greece, in 2004, the M.Sc. by
research, and the Ph.D. in electrical and electronic
engineering from the University of Manchester, UK,
in 2006 and 2009, respectively. He is currently a
Lecturer in the Dept. of Electrical and Electronic
Eng., University College London. He has previously
held a Research Associate position at the University
of Manchester, UK, and a Research Fellow position
at Queens University Belfast, UK. He was recently
awarded a Royal Academy of Engineering Research Fellowship for 2011
2016.
His research interests lie in the fields of wireless communications and
signal processing, with particular focus on green communications, cognitive
radio, interference mitigation techniques for MIMO, CDMA, and multicarrier
communications.
Mathini Sellathurai received the Technical Licen-
tiate degree from the Royal Institute of Technology,
Stockholm, Sweden, in 1997 and the Ph.D. degree
from McMaster University, Canada, in 2001. She is
currently a Reader with Heriot-Watt University Ed-
inburgh, U.K. Her current research interests include
adaptive and statistical signal processing, space-time
and MIMO communications theory, network coding,
information theory, and cognitive radio. Dr. Sellathu-
rai was the recipient of the Natural Sciences and
Engineering Research Council of Canadas doctoral
award for her Ph.D. dissertation and a co-recipient of the IEEE Communi-
cation Society 2005 Fred W. Ellersick Best Paper Award. Dr. Sellathurai is
currently serving as an Associate Editor for the IEEE T RANSACTIONS ON
S IGNAL P ROCESSING.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 5, MAY 2013
Tharmalingam Ratnarajah (A96-M05-SM05) is
currently with the Institute for Digital Communica-
tions, the University of Edinburgh, Edinburgh, U.K.,
as a Reader in signal processing and communica-
tions.
Since 1993, he has held various positions with
Queens University Belfast, Belfast, U.K.; the Uni-
versity of Ottawa, Ottawa, ON, Canada; Nortel
Networks, Ottawa, ON, Canada; McMaster Univer-
sity, Hamilton, ON, Canada; and Imperial College,
London, U.K. His research interests include random
matrices theory, information theoretic aspects of MIMO channels and ad hoc
networks, wireless communications, signal processing for communication,
statistical and array signal processing, biomedical signal processing, and
quantum information theory. He has published over 175 publications in these
areas and holds four U.S. patents. He is currently the coordinator of the FP7
Future and Emerging Technologies project HIATUS (2.7M) in the area of
interference alignment and the FP7 project HARP (3.2M) in the area of highly
distributed MIMO. Previously, he was the coordinator of the FP7 Future and
Emerging Technologies project CROWN (2.3M) in the area of cognitive radio
networks. Dr. Ratnarajah is a member of the American Mathematical Society
and Information Theory Society.