Chapter 4

Symbol Detection in MIMO System

4.1 Introduction
In the previous chapter we discussed that using multiple antennas at the transmitter and receiver
enable significant increase in the spectral efficiency and the reliability of a mobile radio channel
without increasing the system transmission power or bandwidth. However, developing a cost
effective MIMO system needs considerable amount of effort. The bottle neck is designing low
computational complexity and efficient receivers that can fully exploit the benefits of the MIMO
architecture without taking a long time to decode the transmitted symbols. In this chapter the
Optimal Maximum-Likelihood MIMO detection problem is formulated and a survey of existing
detection techniques which reduce the ML complexity for MIMO systems in the un-coded
spatial multiplexing system is presented.

4.2 MIMO System Model

Consider a MIMO system as shown in Fig 3.1 where Nt different signals are transmitted and arrive
at an array of Nr (Nt Nr) receivers via a rich-scattering flat-fading environment. Grouping all the
transmitted and received signals into vectors, the system can be viewed as transmitting an Nt x 1
vector signal x through an Nt x Nr matrix channel H, with Nr x 1 Gaussian noise vector v added at
the input of the receiver as depicted in (3.2)

y = Hx + v

(4.1)

Chapter4 Symbol Detection in MIMO System

where y is the received Nr x 1 vector. The (nr, nt)th element of H, hn

nt

, is the complex channel

response from the ntth transmit antenna to the nrth receive antenna. x is zero mean and has
covariance matrix of Rx = E{xx*} = x2 I. The vector v is also zero-mean and Rv = E{vv*} = v2 I.
The entries of channel matrix H are assumed to known at the receiver but not at the transmitter. If
training or pilot signals are sent to get the channel information this assumption is reasonable.
Channel parameters are constant for some coherent interval.

4.3

MIMO Detection Problem Formulation

The task is that of detecting Nt transmitted symbols from a set of Nr observed symbols that have
passed through a non-ideal communication channel, typically modeled as a linear system followed
by an AWGN as shown in Fig 4.1.

Fig 4.1. A simplified linear MIMO communication system showing the following discrete
signals: transmitted symbol vector x N , channel matrix H Nt xNr , additive noise
t

vector v Nt , receive vector y N , and detected symbol vector x N .

t

Transmitted symbols from a known finite alphabet = {x1,,xM} of size M are passed to the
channel. The detector chooses one of the MNt possible transmitted symbol vectors from the
available data. Assuming that the symbol vectors x N are equiprobable, the Maximum
t

Likelihood (ML) detector always returns an optimal solution according to the following:
x arg max P( y is observed x was sent )
*
x Nt

(4.2)

Assuming the additive noise v to be white and Gaussian, the ML detection problem of Fig 4.1 can
be expressed as the minimization of the squared Euclidean distance to a target vector y over Ntdimensional finite discrete search set:
58

Chapter4 Symbol Detection in MIMO System

x arg min y Hx
* x N

(4.3)

Optimal ML detection scheme needs to examine all MNt or 2bNt symbol combinations (b is the
number of bits per symbol). The problem can be solved by enumerating over all possible x and
finding the one that causes the minimum value as in (4.3).
In ML detection, optimization is performed over the space of all possible vectors x. Since the
search space is discrete with x having integer components, this problem is posed in the literature
as an integer least-squares optimization problem [85], and it belongs to the class of
nondeterministic polynomial-time hard, NP-hard, combinatorial optimization problems [86, 87].
A combinatorial optimization (CO) problem involves searching values for discrete variables in
such a way an optimal solution with respect to a selected objective function is detected. A
straight forward approach to the solution of a CO problem would be exhaustive search, i.e. the
enumeration of all possible solutions and choosing the one that minimizes the objective function
in equation (4.3). A naive implementation of this search strategy results in a prohibitive
complexity, as the number of candidate solutions increases exponentially with the problem size.
Therefore, for a Nt x Nr MIMO system with symbols from M-QAM constellation alphabet the
computational complexity increases exponentially with constellation size M and number of
transmitters Nt as can be observed from (4.3).
This work focuses on designing MIMO detection algorithms capable of finding a near optimal
solution with lesser than ML computational complexity. These will be low complexity near
optimal uncoded MIMO detectors.

4.4

Existing MIMO Detection Algorithms

Two classes of algorithms are available for the solution of combinatorial optimization problems:
exact and approximate algorithms. Exact algorithms find the optimal solution for every finite
size of a combinatorial optimization problem; however, for NP-hard problems, exact algorithms
have an exponential worst-case complexity, and they generally suffer from a strong rise in
computation time when the problem size increases.
Approximate algorithms, on the other hand, trade optimality for efficiency; they exploit some
problem-specific knowledge to produce reasonable solutions at a comparatively low
computational complexity with no surety to produce optimal solutions.
59

Chapter4 Symbol Detection in MIMO System

In the following, we will briefly review some of the exact and approximate algorithms that have
been used to solve the MIMO detection problem.

4.4.1 Sphere Decoder

Fincke and Pohst introduced the Sphere Decoder (SD) algorithm [88] as an exact MIMO detection
algorithm which gives ML BER performance [89]. Transmitted symbols x vector are represented
as points on a rectangular integer grid Nt- dimensional lattice. The MIMO channel H is assumed as
a lattice generating matrix. Hx which is Nr- dimensional spans the transformed skewed lattice.
Hence considering a received vector y at Nr receivers and Hx the transformed lattice, ML detection
problem reduces to determining the nearest lattice point to y in terms of minimum Euclidean
distance. The basic concept of SD is to search for the lattice points within a radius R around the
received vector y, instead of traversing the complete search space of the lattice. The SD searches
the point that gives the minimum Euclidean distance within the hypersphere with radius R, as
shown in Fig 4.3.

y
R

Fig 4.2: Sphere Decoder algorithms illustration.

The resultant point in hypersphere that are closest y is the solution returned by the SD algorithm.
The issue of selecting the radius R and finding the nodes that are with this radius actually
determine the efficiency of SD algorithm.
SD is a type of Branch and Bound tree search algorithms [89]. Applying QR factorization on H
matrix channel enables converting it into tree structure inherent to it. A product of unitary matrix Q
and upper triangular matrix R is substituted in place of H in (4.3). The resultant expression is
multiplied with QT to yield the following equivalent problem representation:

60

Chapter4 Symbol Detection in MIMO System

x arg min y' Rx
* x N

(4.4)

here y' = QT y . The above function can be rewritten owing to the upper triangular nature of R in
terms of SD algorithm.
n

j=1

y - rj,l x l
'
j

C0

(4.5)

l= j

where C0 is the squared radius of n-dimensional sphere with origin at y' . This condition needs to
be justified to ensure that the lattice point x falls within the hypersphere for all the components xj,
where j= 1 to n. This constraint is observed in a depth first method meaning that the root node is
searched first till the algorithm reaches the leaf node to determine a solution that satisfies the above
constraint. Similarly reverse traversing in the hunt to find more leaf nodes which fulfill the criteria
is done. SD algorithm then outputs the lattice points that possess the least Euclidean distance with

y' . A tree constructed by SD algorithm for 2x2 4-QAM MIMO system is illustrated in Fig 4.3.

Fig 4.3: Sphere Decoder algorithms tree structure illustration.

The points not fulfilling the constraints imposed are pruned from the tree, the lattice points that are
within the sphere are x1 =(0,1,1,0) and x2 =(0,0,1,1).
The procedure for finding the radius of search C0 which was not originally specified was worked
out by Vikalo and Hassibi in [90], by utilizing noise variance. In [91] it was proposed that C0 be
initially kept as infinity till searching of first lattice point, and then C0 is set equal to the Euclidean
61

Chapter4 Symbol Detection in MIMO System

distance from received signal. A significant research work has been done in SD algorithm
suggesting numerous methods to efficiently search the lattice tree structure, like lower bound
pruning [91], lattice reduction [89], incrementing search radius [90], etc.
The worst case computational complexity of SD remains exponential; however its expected
complexity is regarded as cubic over certain SNR and problem dimensions [90]. An exponential
lower bound on SD average complexity has been worked out by Jalden in [94], however their will
definitely be a problem dimension in which an approximate algorithm with polynomial time
proves to be efficient than SD especially at lower SNR and larger modulation alphabets sizes.
Remainder of this chapter discusses the approximate MIMO detection algorithms such as linear,
non-linear and Semi definite programming methods.

4.4.2 Linear Detection

Linear detection is used in the class of receivers in which the symbol estimate x is taken by a linear
transformation of the received symbol y [28] as:

x Q z ( Wy )

(4.6)

where matrix W is dependent on the channel H and Qz is a quantizer or slicer that maps its
arguments on to the nearest constellation point.

4.4.2.1

Zero-forcing Detection

These detectors solve the integer least square problem by removing the discreteness constraint on
the components of x. Zero-Forcing detection is low complexity linear detection algorithm that
gives the estimate of x as:

x Q z (x ZF )

(4.7)

x ZF H y

(4.8)

and

62

Chapter4 Symbol Detection in MIMO System

where H+ denotes the pseudo-inverse of H and

x ZF

is mapped to nearest integer in the

constellation alphabet from which x is derived [28]. The receiver tries to force the cross correlation
between the estimation error and the transmitted vector X to zero, therefore it is termed as ZeroForcing detector.
ZF detection algorithm is a linear detection algorithm since it behaves as a linear filter separating
different data streams to perform decoding independently on each stream, therefore eliminating the
multi-stream interference. The drawback of ZF detection id retarded BER performance due to
noise enhancement. The AWGN noise v loses its whiteness property it is enhanced and correlated
across the data streams. In addition, ZF detection gives Nr-Nt+1 diversity order in a NtxNr MIMO
system with Nr possible diversity order.
For n x n MIMO system, ZF detector possess a polynomial complexity of cubic order O(n3) which
constitutes the computational complexity of calculating the pseudo-inverse of the matrix channel
H.

4.4.2.2

Minimum Mean Square Error Detection

Minimum Mean Square Error (MMSE) detector estimates the transmitted vector x by applying the
linear transformation to the received vector y. It finds out the estimate x of the transmitted symbol
vector x as:

x Q z (x MMSE )

(4.9)

x MMSE Wy

(4.10)

and

where
1

W HH H
I Nt H H
SNR

2
where W is selected to minimize the mean square error as E Wy x .

63

(4.11)

Chapter4 Symbol Detection in MIMO System

MMSE detectors balances the noise enhancement and multi-stream interference by minimizing the
total error. Its BER performance is superior to ZF detection due to mitigating the noise
enhancement. Its computational complexity is dominated by the matrix inversion in (4.11), which
is cubic order O(n3).

4.4.3 Non-linear (V-BLAST) Detection

V-BLAST MIMO architecture uses Successive Interference Cancellation (SIC) technique to
decode the transmitted symbols. Instead of detecting the transmitted symbol vector with Nt
symbols jointly it decodes the symbols based on the energy of the symbols. V-BLAST decodes the
the first transmitted symbol by fulfilling the ZF or MMSE performance criterion and assuming the
the remaining symbols contribution as interference, subsequently the contribution of the earlier
detected symbol is cancelled out to get a reduced order integer least square problem having Nt-1
unknowns. The V-BLAST is a non-linear recursive detection procedure that that extracts the
components of the transmitted vector x according to a certain ordering (l1, l2,,lNt) of the indices
elements of x based on channel realization H. The ZF based V-BLAST algorithm is a V-BLAST
derivative based on ZF criterion. Its algorithm is given below [95].
Initialization:

W1 H +
i 1
Repetition:

ki arg min Wi j
yki ( Wi ) ki y i
x ki Q Z ( yki )
y i 1 y i x ki (H)ki
Wi 1 H ki
i i 1

64

when j [k1...ki 1 ]

Chapter4 Symbol Detection in MIMO System

H+ is the pseudoinverse [85] of H, jth of Wi, (H ) ki is the kith column of H. H ki is the resultant

matrix by zero forcing the columns of H. Qz(.) is a slicer.

The above algorithm first of all orders the symbols to be decoded, carries out nulling and gets
decision statistics, decides the first symbol. The coefficients of H corresponding to the detected
symbol having maximum energy are cancelled out, inverse of the new Channel matrix H are
computed. The process is repeated to iteratively decode the transmitted symbols.
This ZF VBLAST detection algorithm uses ZF approach discussed earlier. The ordering of the
+

symbols is based on the minimum noise variance (H ) j

N0 .

The VBLAST computational complexity for n x n MIMO systems is O(n4). It possesses enhanced
BER performance than linear detectors like ZF and MMSE, however suffers from error
propagation. VBLAST performance suffers degraded performance when the first symbol is
decoded incorrectly.
P erformance
Analysis of Linear and Non-linear detectors-4-QAM 4x4 MIMO S ystem
0
10
ML
ZF
VB LAS T
-1
10

-2

BER

10

-3

10

-4

10

-5

10

10

15
E b/No

20

25

30

Fig 4.4: Performance of linear and non-linear MIMO detectors.

65

Chapter4 Symbol Detection in MIMO System

4.4.4 ZF-ML Detection

The quadratic form of (4.2) given as:
f (x) y Hx

(4.12)

The function f(x) in (4.12) is convex. This detection algorithm also termed as multi-step reducedconstellation (MSRC) detection performs local search of the target symbols within certain
constraint specified reduced search space. In fact a ZF initial solution estimate is used to define
the radius of search. Constellation points around the ZF solution are searched in steps using (4.2)
to find out the minimum Euclidian distance. This particular method which starts with the ZF
processing is termed as ZFML detection [90].
First y is computed and then a ML search around the neighborhood of y is performed as shown
in Fig 4.6. Each of the Nt symbol generates a neighbor list, then a joint ML search our reduced
constellations is performed. For a 16-QAM 4x4 MIMO and fixed neighbor size of 4 for each
antenna, there are nine entries in the lookup table, with each entry containing 4 constellation
points. For simplicity pivot points can be defined like in the above case nine pivot points are
defined. We determine one pivot point which lies close to the ZF solution found, now the
neighbors of each pivot point are searched jointly.

Fig 4.5: ZF-ML reduced constellation search

Once the first-step search result is generated, we can form a second neighbor list around it and
perform a second-round search, and so on. The neighbor lists for the second round can be

66

Chapter4 Symbol Detection in MIMO System

generated and searched. BER performance comparison of ZF-ML, ML and linear detection
methods is represented in Fig 4.7.
Compared to ML that performs a coarse search over the complete search space the ZFML used a
Nt

reduced constellation, therefore its computational complexity is M M . Where M is the

constellation size, Mn is the neighbors list and Nt is the number of transmitters.

0

10

P erformance Analysis of ZFML Detector in 4-QAM 4x4 MIMO S ystem

ML
ZF
ZFML

-1

10

-2

BER

10

-3

10

-4

10

-5

10

10

15
E b/No

20

25

30

Fig 4.6: Performance of ZFML MIMO detectors.

4.4.5 SDP Detection

Convex relaxation technique is another approach to solve the integer least-squares problems.
In these techniques the objective function to be minimized is expressed in a relaxed form while
keeping the problem convex and the convexified problem is solved using mathematical
programming techniques.
Semi-Definite Programming (SDP) approach has been successfully used for MIMO Detection
[97],[98]. Semidefinite programming technique deals with optimization problems which can be
expressed as [99].
minimize

Tr (CX)

subject to

Tr (A i X) = bi i 1,..., m

X0

67

Chapter4 Symbol Detection in MIMO System

where the space of real symmetric n x n matrices is X Sn , the matrices A i Sn and C Sn define
the problem parameters and b R m . Trance of CX matrix is defined by Tr (CX) and the inequality

X 0 shows that X is positive semidefinite.

The integer least-squares problem of MIMO detection is represented in higher dimensions and
subsequently SDP is obtained by relaxing nonconvex constraints or SDP can be derived as
Lagrangian bidual [100]. The integer least- squares problem can also be converted into binary
quadratic minimization problem with the help of different relaxation models that can be further
solved by SDP [97]. The SDP based MIMO detectors approach near-optimal BER performance
with the computational complexity for n x n MIMO system of O (n5.5 M 5.5 log(1/ )) with
tolerance .

4.5

Summary

A survey of existing linear, non-linear and exact MIMO detectors was presented in this chapter.
Linear detectors like ZF, MMSE have reduced computational complexity with a lower BER
performance. Non-linear techniques like VBLAST present acceptable BER performance with not
very high complexity. Exact techniques such as Sphere Decoder give optimal BER performance
however the complexity is still on higher side and may vary with noise variance. It is believed that
under certain conditions SD complexity can become exponential.

68