ANALYTIC FRAMEWORK FOR PERFORMANCE EVALUATION

OF MULTI-ANTENNA WIMAX SYSTEMS OVER FADING CHANNEL

D. Molteni, R. Bosisio, M. Nicoli

Dip. Elettronica e Informazione, Politecnico di Milano, Italy; E-mail: {molteni,nicoli}@elet.polimi.it

ABSTRACT at the CC output by adapting the concept of effective SINR to M-

QAM based on a set partitioning of the modulation mapping [6].
This paper focuses on the uplink of multicell IEEE 802.16-d Then, the error rate of the overall concatenated code chain is derived
WiMax systems with OFDM modulation and antenna array at the by evaluating the error event correction ability of the outer block
base stations. We propose an analytical framework to assess the code. An ef¿cient description of the long-term features of the multi-
average error probability of all the different transmission modes path channel allows the analysis to cover heterogeneous propagation
over space-time dispersive Rayleigh fading channels. The proposed environments with varying angle-delay spreads. Numerical results
method takes into account the effects of the multilevel modulation, show that the analytical model provides an accurate ¿tting of the
the error correction capability of the concatenated code, the inter- BER over a large range of space-time multipath environments. The
leaving scheme, the power-angle structure of the inter-cell interfer- proposed model can also be extended for the evaluation of other per-
ence and the array processing at the base station. Simulation results formance metrics such as the frame error rate and the outage error
corroborate the proposed analysis for a IEEE 802.16-d cellular sys- probability, or it can be exploited for the design of channel aware
tem over different propagation scenarios. scheduling algorithms.
Index Terms— WiMax, Antenna arrays, Convolutional codes,
2. MODELING THE WIMAX SYSTEM
Concatenated coding, Fading channel, Bit-interleaving.
We focus on the uplink of a IEEE 802.16-d multicell scenario [1][2]
1. INTRODUCTION with ¿xed subscriber stations (SSs). We consider an hexagonal cell
Worldwide Interoperability for Microwave Access (WiMax) is an layout as in [5] where the transmission from the single-antenna sub-
OFDM-based technology [1][2] that provides last mile broadband scriber station SS0 to the O-antenna base station BS0 is impaired
wireless access. High data-rate, large spectral ef¿ciency and im- by the QL out-of-cell interferers {SSl }Q l=1 . The bit sequence {en },
L

proved coverage make WiMax one of the most promising technolo- to be transmitted by SS0 , is coded, bit-interleaved and mapped onto
gies for the near future. The system is expected to operate in het- the sequence of complex symbols {{n }. The standard prescribes
erogeneous scenarios as possible applications range from wireless M-QAM modulation coupled with a concatenated coding scheme
services in rural areas to intensive applications on notebooks and made of an inner CC and an RS outer code. The modulation set
mobile devices. This makes mandatory to assess the WiMax perfor- of size P = 2p s is here indicated by X = {v1 > = = = > I vP }, with
mance in heterogeneous propagation environments. vn = (vT n + mvL
n ) H j and v T L
n > vn M {±1> ±3> = = = > ± P 3 1}
The evaluation of the performance for bit-interleaved coded (Hj is the transmitted waveform energy). The CC has generators
OFDM systems over frequency selective fading channels have re- (171,133) and code rate Ucc = ncc @qcc (a puncturing scheme is
cently attracted a great interest. In [3], the concept of effective provided for some Ucc values), while the RS code is characterized
signal-to-noise ratio (SNR) has been introduced as the SNR of an by the parameters (qrs > nrs > wrs ).
equivalent additive white Gaussian noise (AWGN) channel which The modulated symbols {{n } are transmitted by SS0 through
would yield the same error probability as that of a given frequency- OFDM signalling over a frequency selective fading channel. The
selective channel. The effective SNR can be used to adapt mod- O × 1 baseband signal received by BS0 on the nth subcarrier is:
ulation and coding modes or to assess the average error rate for a
given transmission mode [4]. Going one step further, in [5] a uni¿ed yn = hn {n + nn > (1)
framework has been proposed to analytically assess the bit error rate
(BER) of a multi-antenna WiMax system over space-time dispersive where hn gathers the complex channel gains for the O receiving an-
Rayleigh fading environments with spatially correlated intercell in- tennas, {n M X is the transmitted symbol with Hv = E[|{n |2 ] =
terference. The error rate is expressed as a function of the effective
2
3
(P 3 1)Hj , while nn models the background noise and the inter-
signal to interference plus noise ratio (SINR), rede¿ned with respect cell interference. The noise vector nn is assumed to be zero-mean
to [4] so as to jointly account for the correlation of the inter-cell in- Gaussian, temporally uncorrelated but spatially correlated with spa-
terference over the antennas and the correlation of the channel gains tial covariance matrix Q = E[nn nH n ] =  q IO +QL , where  q is the
2 2
S L
over both the antennas and the subcarriers. variance of the background noise while QL = Q Q
l=1 L>l denotes
The analysis in [5] accounts for a single transmission mode: the contribution from the QL interferers {SSl }Q l=1 . The subcarrier
L

BPSK modulation and convolutional code (CC). However, the push- index, n M K = {n1 > = = = > nN }, ranges over the N subcarriers used
ing demand for high data rate applications makes mandatory the for data transmission, while the total number of subcarriers (which
employment of more ef¿cient transmission modes. In particular, includes pilots, guard-bands and DC subcarrier) is indicated by Q.
WiMax prescribes M-QAM and concatenated coding schemes made See [5, Table 1] for a list of IEEE 802.16-d system parameters.
by an inner CC and an outer Reed-Solomon (RS) code. In this pa- A minimum variance distortionless (MVDR) combiner [7] is ap-
per, we extend the framework [5] to account for the whole set of plied to the O signals (1) in order to reduce the interference effects.
standard-compliant transmission modes. The BER is ¿rst evaluated The output is { ˆn = (h̃H n h̃n )
31 H
h̃n yn , where h̃n = Q3H@2 hn is

1-4244-1484-9/08/$25.00 ©2008 IEEE 3081 ICASSP 2008

the channel vector weighted by the inverse Hermitian Cholesky fac- PEP can be computed as [6]:
tor of the noise spatial covariance Q. As shown in [7], the resulting [ [
{n 3{n |]2 ,
SINR after spatial ¿ltering, de¿ned as  n = E[|{n |2 ]@E[|ˆ Sg () $ S (i) S (x)Sg (|i> x) (4)
is given by  n = Hv ||h̃n || . The sequence {ˆ
2
{n } obtained at the i x

beamformer output is then used for Max-Log-MAP demodulation, where Sg (|i> x) is the PEP conditioned to the set (i> x), while
bit deinterleaving, Viterbi and RS decoding. S (i) = 1@pg is the probability of each label set. Notice that for
The spatial channel vector is modelled as in [5] as: any coded bit sequence to be modulated and any given label set i,
QU   there are 2(p31)g possible symbol sequences x with equal probabil-
I [ n u ity S (x) = 1@2(p31)g =
hn = S0 u a(u )Jn exp 3m2 = (2)
Q W Similarly to [5], Sg (|i> x) is a function of the SINR at the deci-
u=1
sion variable (the effective SINR) that is a combination of the SINR
This is the sum of QU paths, each characterized by a direction of values  g = [ i · · ·  ig ]T observed over the subcarriers ff . How-
1
arrival (DOA) u , a time of arrival (TOA)  u and a complex fad- ever, here each SINR value needs to be scaled by a factor that takes
ing amplitude u . S0 denotes the average power from all the paths, into account the Euclidean distance between the transmitted QAM
Jn is the frequency response of the cascade connection of the trans- symbol and its nearest concurrent in the QAM constellation. This
mitter and receiver ¿lters on the nth subcarrier, W is the sampling factor varies with {n and ln . More speci¿cally, let X0l (and X1l ) be
interval, while the vector a(u ) = [d1 (u ) · · · dO (u )]T represents the subset of all symbols {n whose QAM label has the value 0 (and
the array response to the DOA u . Amplitudes are assumed to 1) in position ln , ln = 1> = = = > p. These subsets are shown in Fig.
be uncorrelated Rayleigh-faded, u q CN (0> ( u > u )), with 1 for P = 16, l = 1 and Gray’s mapping. If the transmitted bit is
variance ( u > u ) that depends on the power-delay-angle pro¿le equal to 0 and it is mapped onto the lth label position, the transmitted
(PDAP) ( > ). We employ an exponential PDAP, ( > ) x symbol {n belongs to the subset X0l . The Max-Log-MAP demod-
exp (3 @ ) exp (3 | 3 0 | @  ), with delay spread  and an- ulator decision is erroneous when the received symbol lies in X1l .
gle spread  . The angle 0 denotes the main DOA connecting SS0 Thereby, the probability of error can be upper bounded by using the
and BSS 0 , while x stands for proportional. The PDAP is normalized distance between {n and the boundary of the area associated to the
so that Q U
u=1 ( u >  u ) = 1.
nearest neighbor in s X1l . For the 16-QAM example in Fig. 1 (l = 1),
We also adopt two simpli¿ed Rayleigh-faded channels as ex- this distance is {n Hj , with {n = 1 for {n belonging to group
treme cases of frequency selectivity: the frequency-Àat (FF) channel 1 and {n = 3 for group 2. The same holds for the label position
where the channel gains are constant all over the bandwidth (as for l = 3, while for l = 2 and l = 4 only the distance factor {n = 1
 < 0) and the maximum frequency diversity (FD) channel where is experienced. As a result, the SINR at the decision variable is a
the channel gains are i.i.d. over the subcarriers (as for  < "). function of the overall sequence { = {(i> x) = [{i1 · · · {ig ] of
s
A multipath model is employed also for the interference sce- the g Euclidean distances (normalized by Hj ) that are selected by
nario. The lth interferer spatial covariance,
S U for l = 1> = = = > QL , the g error bits in the codeword f:
is indeed modelled as QL>l = Sl Q  ( H
u=1 l l>u )a( l>u )a ( l>u ), [ [
3
QU
depending on the DOAs {l>u }u=1 , the normalized power-angle-  eg (|{) = Hj ||h̃n ||2 {2n =  n {2n = (5)
2(P 3 1)
pro¿le l () x exp (3 | 3 l | @ ) and the mean power Sl . The nMff nMf
f

angle l denotes the main DOA connecting SSl and BS0 .

The corresponding PEP is
3. CONVOLUTIONAL CODE PERFORMANCE ANALYSIS s 
Sg (> |i> x) = Sg (|{) = T 2 eg (|{) = (6)
In this Section we analyze the performance at the output of the CC
decoder. The extension to CC-RS concatenation will be carried in It is worth noticing that each distance {n can assume only few
the following Section. Given a channel characterized by the SINR values. In fact, for the 16-QAM case in Fig. 1 it is {n = 1 for
set  = [ 1 · · ·  N ], the bit error probability Secc () = S (en 6= three quarters of the sets (ln > {n ) and {n = 3 for the remainder, i.e.:
en |) can be approximated using the union bound approach as [8]: S ({n = 1) = 3@4 and S ({n = 3) = 1@4. Similar considerations
holds for 64-QAM. Since only few distance values are observed,
1 [
Secc () $  g Sg () (3) it is convenient to gather in (4) all the con¿gurations of (i> x) that
ncc gDg correspond to the same distance sequence {, yielding:
free

[ s  k s l
where gfree is the CC free Hamming distance, Sg () is the pair- Sg () $ S ({)T 2 eg (|{) = E{ T 2 eg (|{) >
wise error probability (PEP) associated to the codeword f having {
Hamming distance g from the transmitted sequence, while  g is the (7)
corresponding input weight. As in [5], we select as error event f the where S ({) is the probability of the distance set { and E{ [·] in-
codeword that corresponds to the worst con¿guration of g error bits; dicates the average with respect to {. The average of the bit error
such a con¿guration needs to be evaluated for the speci¿c code and probability with respect to fading, Secc = E [Secc ()], is ¿nally ob-
interleaver, as done in [5] for the IEEE 802.16-d standard. tained by plugging each PEP (7) into (3) and by integrating over
For the derivation of Sg () we assume that the g erroneous bits  eg (|{) using the pdf sg ( eg |{) of the effective SINR:
in the error event are mapped by the interleaver onto g different sub- ] s  
1 [
carriers f = [i1 · · · ig ]. Let the nth bit (n = 1> = = = > g) belong Secc =  g E{ T 2 eg (> f|{) sg ( eg |{)g eg |{ =
to the QAM set element {n M X and be placed in the position ncc
gDgfree
ln M {1> = = = > p} of the modulation label set; let x = [{1 · · · {g ] (8)
and i = [l1 · · · lg ] be the sequences of the g QAM elements and the Given a distance set {, the pdf sg ( eg |{) can be evaluated
g bit positions in the QAM labels selected by the error event f. The through the eigenvalue decomposition (EVD) of the space-frequency

3082
 'k
0
3 10

'k 1 -1
10

1101 1001 0001 0101 -2 ıIJ 0s

X11 X01 10
(FF)
-3
10
1100 1000 0000 0100
-4
10 ıIJ 0.1ȝs

BER
1110 1010 0010 0110 -5
10

-6
10
ıIJ o f ıIJ 0.3ȝs
1111 1011 0011 0111 (FD)
-7
10
Simulations
Group 2 Group 1 -8
10 Analytical d=6,7
Analytical d=6,7,8 ıIJ 1ȝs
Fig. 1. The 16-QAM constellation with Gray’s mapping. In the sub- -9
10
5 10 15 20 25 30
set X11 there are two groups of bits with different minimum distances SINR [dB]
from the boundary of the subset X01 =
Fig. 2. Performance of a single-antenna WiMax system with QPSK
channel correlation matrix as described in Sect. 3.1. However, in or- modulation. The BER is evaluated at the output of the CC decoder.
der to avoid the expensive EVD for each value of {, we propose
to approximate the expectation E{ [·] by means of a sample aver- For the QPSK and M-QAM modulations, the standard recommends
age: we simulate some values of { as the outcomes of g i.i.d. ran- an inner CC with rate Ucc and memory T, and a RS (qrs > nrs > wrs )
dom variables, having known pdf (see the pdf example above for as outer code. Conventional performance analysis [8] for concate-
16-QAM); for each value, the effective SINR is obtained from (5) nated codes accounts for the error occurrence in the RS codeword as
and its pdf is calculated according to the approach in Sect. 3.1; the a bit-isolated random process. However, the standard here does not
estimate of the average bit error probability is then obtained by av- prescribe an interleaver between the outer and inner codes thus mak-
eraging over some realizations of {. ing mandatory to model the error at the output of the CC decoder as
Remark. It can be easily observed that for QPSK (P = 4) it bursty (any error event in the Viterbi trellis leads to a burst of errors
is S
{n = 1> n, and the effective SINR simpli¿es
S to  eg (> f) = that corrupts one or more consecutive symbols in the outer block).
1
2 nMff  n . For BPSK, it is  eg (> f) = nMff  n as in [5]. Let Gf denote the maximum length (in coded bits) of the er-
ror event of distance g (the length is here evaluated after the pos-
3.1. Computation of the effective SINR pdf sible puncturing inverse process). At the CC decoder output, the
s error burst has length Oout>g = Gf Ucc 3 T> as the error event al-
Let us de¿ne h̃ = Hj [{1 h̃T i1
> · · · > {g h̃T
ig
]T as the vector that ways terminates with T correct bits. To simplify the performance
gathers the scaled Og channel gains associated to the g subcarriers analysis, we divide the RS codeword (made by qrs symbols of
of the error event f. This is a zero-mean complex Gaussian random E = 8 bits each) into Vg = qrs E@(Oout>g + T) disjoint seg-
vector with covariance matrix R = E[h̃ · h̃H ] given by ments. In each segment, the error burst can occur with probability
5 6 sh>g $ (Oout>g + T) Ss > where Ss denotes the probability for an
{21 R(i1 > i1 ) · · · {1 {g R(i1 > ig ) error path to diverge along the Viterbi trellis. It can be easily shown
9 .. .. .. :
that Ss = Secc }g @ g > where }g refers to the number of paths with
R = Hj 7 . . . 8 > (9)
{g {1 R(ig > i1 ) · · · 2
{g R(ig > ig ) Hamming distance g and Secc = E [Secc ()] from (7). The BER is
eventually given by all the possible error event combinations in Vg
where R(n> k) = E[h̃n · h̃H segments:
k ] is the O × O channel crosscorrela-
tion associated with the nth and kth subcarriers. According to the Vg
# $
multipath model in Sect. 2, it is: [ [ l Vg l
E Secc+rs g
sh>g (1 3 sh>g )Vg 3l = (12)
QU   }g Vg l
[ n n 3 k
g l=Wg +1
 ~} €
R(n> k) = S0 ( u > u )ã(u )ãH (u )Jn Jk exp 3m2 Number of possible error paths, weighted by their probability
u=1
Q W
(10)
Notice that the maximum number of symbols that can be corrupted
with ã(u ) = Q3H@2 a(u ). The effective SINR (5) can be now by an error event of distance g in the RS codeword is Qg =
rewritten as  eg (|{) = ||h̃|| , i.e. as the sum of the square mag-
2
bOout>g @Ec + 1. It follows that the RS decoder can correct up to
nitudes of the Og correlated zero-mean Gaussian entries of h̃. We Wg = bwrs @Qg c error events.
can evaluate the pdf of such a sum following the same procedure as Remark. In presence of Àat or almost Àat fading channels, the
in [5]. This leads to the following effective SINR distribution: length of the error bursts can be longer than the error correction abil-
[ Dl ity of the block code. In this case, the RS code does not provide
sg ( eg |{) = exp (3 eg @l ) > (11) any performance bene¿t and the BER in (12) simpli¿es to the CC
l
performance in Sect. 3.
l being the lth eigenvalue of R and Dl = 1@gn=1>n6=l (1 3 nl ).
5. NUMERICAL RESULTS
4. CONCATENATED CODING EXTENSION In this Section we corroborate the analytic framework derived in
In this Section we extend the analytic framework to the concatenated Sect. 3 and 4 through numerical simulations of a IEEE 802.16-d
coding schemes prescribed in the IEEE 802.16-d standard [1] [2]. multicell scenario. The performance is assessed as a function of the

3083
 0 0
10 10

-1
10 Simulations -1
10 ıș 0 deg
Analytical (Concatenated) (FF)
ıIJ 0s
-2 -2
10 10

-3 -3
10 10 ıș o f
ıIJ o f
-4
10
-4
10 ıș 1.1 deg
BER

(FD)

BER
ıIJ 0.1 ȝs
-5 -5
10 10

-6
10
-6
10 ıș 4 deg
-7
AWGN ıIJ 2.5ȝs
ıIJ 1 ȝs
-7
10 10
ıIJ o f Simulations ıș 1.7 deg
-8
10 (FD) -8
10 Analytical d=10 ıIJ 0.4 ȝs
Analytical d=10,12
-9 -9
10 10
0 5 10 15 20 25 -15 -10 -5 0 5 10 15 20 25
SINR [dB] SINR [dB]

Fig. 3. Performance of a single-antenna WiMax system with QPSK Fig. 4. Performance of a 4-antenna WiMax system with 16-QAM
modulation. The BER is evaluated at the output of the RS decoder. modulation. The BER is evaluated at the output of the CC decoder.
S
single-antenna SINR de¿ned as SINR = S0 @(2q + l Sl ) where derivation permits an accurate ¿tting of the simulation results over
S0 @q = 20dB. The active user SS0 is transmitting from the main
2
the whole range of SINR and delay/angle spread.
DOA 0 = 5 deg. All the QL out-of-cell interferers are sup-
6. CONCLUSIONS
posed to transmit at the same power level and their main DOAs are
1 = 346=1 deg> 2 = 0 deg and 3 = 46=1 deg. For further details In this paper an ef¿cient analytical framework has been proposed
on the simulated radio environment please refer to [5]. to evaluate the performance of multi-antenna OFDM systems. The
We ¿rst consider a single-antenna (O = 1) link with P = 4 method allows to calculate, with moderate computational cost, the
(QPSK), Ucc = 2@3 and RS (32> 24> 4). In this case the background average bit error probability of WiMax systems without the need
noise plus co-cell interference simpli¿es to AWGN. The channel is of simulating the whole transmission/propagation/receiving chain.
modelled as in Sect. 2 with exponential power-delay pro¿le (PDP) Simulation results show that the proposed framework provides an ac-
and delay spread  M {0=1; 0=3; 1; 2=5}s. The single path case curate performance ¿tting for different modulations, coding schemes
(FF channel) and the maximum diversity case (FD channel) are sim- and frequency-selective radio scenarios.
ulated as well to provide upper/lower performance references. The
simulated and the analytical BER at the output of the CC decoder are 7. REFERENCES
compared in Fig. 2 as a function of the SINR. The analytical perfor-
[1] IEEE Std 802.16TM -2004, “802.16TM IEEE standard for local
mances are obtained using the union bound (8) truncated to the ¿rst
and metropolitan area networks Part 16: Air interface for ¿xed
two minimum distances g = {6> 7} (g = 6 is the CC free distance),
broadband wireless access systems,” Oct. 2004.
or to the ¿rst three g = {6> 7> 8} (simulations show that this num-
ber is enough for a reasonable performance bound). As expected, [2] IEEE P802.16-2004/Cor1/D5 (Draft Corrigendum to IEEE Std
the performance improves for increasing delay spread, from   = 0 802.16-2004), “Corrigendum to IEEE standard for local and
(FF) to  < " (FD). The analytical BER is shown to be accurate metropolitan area networks - Part 16: Air interface for ¿xed
all over the SINR domain. The BER at the output of the RS decoder broadband wireless access systems,” Sep. 2005.
is shown in Fig. 3. The analytical RS performance is obtained us- [3] S. Nanda, K.M. Rege, “Frame error rates of convolutions codes
ing (12) jointly with the union bound (8) truncated at the distances on fading channels on the concept of effective Ee /N0 ,” IEEE
g = {6> 7> 8}. Since the RS encoder works on a codeword length Trans. Veh. Technol., Vol. 47, No. 4, pp. 1245-1250, Nov. 1998.
equal to the OFDM symbol duration, the coding gain provided by
[4] K. Witrisal, Yong-Ho Kim, R. Prasad, “A novel approach for
the block encoder (as outer code) can be appreciated only for large
performance evaluation of OFDM with error correction coding
 values. Thereby, we selected  = 2=5s and the FD case for RS
and interleaving,” in Proc. IEEE VTC Fall ’99, Vol. 1, pp. 294-
performance evaluation. The AWGN case is shown as a reference.
299, Sep. 1999.
The multi-antenna case is considered in Fig. 4 with 16-QAM
modulation and Ucc = 1@2. The performance is evaluated at the [5] D. Molteni, M. Nicoli, R. Bosisio, L. Sampietro, “Performance
output of the CC decoder. The ULA at the base station BS0 is com- analysis of multiantenna WiMax systems over frequency selec-
posed of O = 4 antennas with inter-element spacing 1=4, where  is tive fading channels,” in Proc. IEEE PIMRC ’07, Sep. 2007.
the carrier wavelength (see [9] for antenna spacing selection). Multi- [6] G. Caire, G. Taricco, E. Biglieri, “Bit-interleaved coded modula-
path channels are simulated with delay spread  M {0=1; 0=4; 1} s tion,” IEEE Trans. Inform. Theory, Vol. 44, No. 3, pp. 927-945,
and angular spread  M {1=1; 1=7; 4} deg. In addition, the FD and May 1998.
FF cases are simulated. The interference covariance matrix is de-
[7] H. L. Van Trees, Optimum Array Processing, Wiley, 2002.
rived using as angular spread  = 4 deg for all the three inter-
ferers. The union bound (8) is here truncated to the error events at [8] J.G. Proakis, Digital Communications, McGraw Hill 2001.
either the free distance g = 10, or at the two ¿rst minimum distances [9] M. Nicoli et al, “Throughput optimization for non-uniform lin-
g = {10> 12}. As already observed for the single-antenna case, the ear antenna arrays in multicell WiMAX systems,” in Proc. IEEE
best performances are obtained with environments which provide an Workshop on Smart Antennas (WSA’06), March 2006.
high grade of diversity. It can be easily seen that the analytical BER

3084