Physical Communication 2 (2009) 296–305

Contents lists available at ScienceDirect

Physical Communication
journal homepage: www.elsevier.com/locate/phycom

Full length article

A novel approach to UWB data detection with symbol-level

synchronizationI
Vincenzo Lottici a,∗,1 , Zhi Tian b,2 , Geert Leus c,3
a
Department of Information Engineering, University of Pisa, Pisa, Italy
b
Department of Electrical & Computer Engineering, Michigan Technological University, Houghton, MI, United States
c
Faculty of Electrical Engineering, Mathematics, and Computer Science, Delft University of Technology, Delft, The Netherlands

article info abstract

Keywords: Differentially modulated ultra-wideband (UWB) systems have recently attracted a lot
Ultra-wideband communications
of attention since they can avoid the costly channel estimation required by coherent
Multi-symbol differential detection
schemes. The conventional differential-detector (DD), however, shows an inevitable 3 dB
Symbol-level synchronization
Sphere decoding performance loss and suffers from multiple access and intersymbol interference. Multiple
symbol differential detection (MSDD) provides an attractive solution that alleviates the
SNR loss, but still calls for accurate timing recovery. In this paper, we show how to
relax the severe timing requirements of the MSDD thereby only relying on symbol-level
synchronization. Further, the detection complexity can be kept at an affordable level by
pursuing a sphere decoding approach. Simulation results corroborate the effectiveness of
the proposed system when operating in typical dense multipath propagation scenarios.
© 2009 Elsevier B.V. All rights reserved.

1. Introduction efficiently meet the strict requirements imposed by sev-

eral applications, such as short-range high-rate indoor con-
Ultra-wideband (UWB) impulse radios have been at- nectivity, location-aware wireless networks and low-rate
tracting a growing interest in the field of next-generation communications with high-resolution ranging [2].
wireless communications [1,2]. Conveying information The harsh multipath propagation conditions typically
over a stream of ultrashort pulses at very low spectral den- occurring in wireless environments, however, hamper the
sity, several attractive features are promised including fine extensive deployment of UWB systems. Each transmitted
timing resolution, robustness against multipath, high user pulse, indeed, arrives at the receiver over tens or even
capacity, coexistence with legacy services via frequency- hundreds of delayed paths [3], with possibly severe per-
overlay, low probability of interception and detection and pulse shape distortion due to diffraction and scattering
precise positioning capability [1]. These appealing features effects [4]. In such operating environments, exploiting the
have designated UWB signaling as a viable candidate to rich diversity of UWB channels is evidently revealed as be-
ing very difficult, especially in view of the limited afford-
able receiver complexity. The well-known Rake receiver
I Parts of this paper were presented at the SCVT’2007 Conf. in Nov. 2007
can collect a significant fraction of the received energy
and the ICUWB’2008 Conf. in Sept. 2008. scattered over a dense multipath [5]. But as a matter of
∗ Corresponding author. Tel.: +39 050 2217 536; fax: +39 050 2217 522.
fact, its choice is practically impeded by the large required
E-mail addresses: [email protected] (V. Lottici),
[email protected] (Z. Tian), [email protected] (G. Leus). number of correlator-based fingers combined with the in-
1 Supported by the FP7 Network of Excellence in Wireless COMmuni- tensive computational load involved in the estimation of
cations NEWCOM++ (contract n. 216715). channel parameters [6]. Viable yet sub-optimal alterna-
2 Supported by US NSF CAREER grant CCF-0238174. tives for efficient energy capture have been recently pro-
3 Supported in part by NWO-STW under the VIDI program (DTC.6577). posed in the form of transmitted reference (TR), differential
1874-4907/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.phycom.2009.08.009
 V. Lottici et al. / Physical Communication 2 (2009) 296–305 297

detector (DD) schemes and their variants [7–10]. In the for- on symbol-level synchronization or SLS-MSDD for short,
mer, the received waveform resulting from ‘‘information- that departs from previous works thanks to the following
free’’ reference pulse(s) is used as noisy template in a distinct features.
simple correlation receiver for data detection, whereas
(1) The SLS-MSDD adopts an optimization criterion based
in the latter differential encoding of information data al-
on the generalized likelihood ratio test (GLRT), in
lows one to detect the current symbol using as noisy tem-
which the likelihood function is maximized not only
plate a replica of the signal waveform received within
with respect to the symbols to be searched for, but
the previous symbol interval. These detectors can gather
energy from all multipath components bypassing costly also over all the finite-energy received template
path-by-path channel estimation, but still suffer from waveforms. As such, both the channel response and the
several considerable drawbacks, among which the fact that timing offset are treated by the receiver as unknown
the template waveform recovered from the received signal quantities and hence do not need to be explicitly
is neither noise-free nor interference-free. As a result, both acquired.
of them experience poor detection performance when op- (2) By virtue of the GLRT-based optimization approach,
erating in the presence of severe multiple access and inter- the expensive channel estimation task is bypassed
symbol interference. without affecting the received energy capture mech-
The need for circumventing the inherent weaknesses anism.
of the TR and DD methods has thus prompted the de- (3) Dealing implicitly with the mistiming effect (within
velopment of improved non-coherent receivers. A recent the detection process itself) allows detection of the
solution is based on the idea of jointly detecting a block of burst data only from the easy-to-get information of
consecutive differentially-encoded symbols experiencing where the symbol boundaries are roughly located,
the same unknown channel, which is termed as multiple rather than requiring costly timing synchronization at
symbol differential detection (MSDD) [11–13]. The efficacy frame or even at pulse level.4
of the MSDD approach is confirmed by the considerable re- (4) The implementation of the SLS-MSDD scheme via ex-
silience to severe multipath fading and multiple access in- haustive search exhibits exponential complexity that
terference (MAI). The detection performance obtainable by quickly becomes impractical as the burst length in-
the MSDD at a reasonable complexity exhibits only a small creases. Consequently, in order to exploit the perfor-
gap from the (impractical) conventional Rake process- mance advantages expected at large data block sizes,5
ing, say around 3–4 dB, and favorably combines with the a proper reformulation of the proposed detector based
capability of working independently of the pulse shape dis- on sphere decoding (SD) [23] suggests a fast iterative
tortion and the knowledge of the channel response. The scheme that attains the desired performance of the
advantages of the MSDD can be attained, however, only original formulation but at an appealing polynomial
provided that timing information is appropriately recov- complexity.
ered from the received signal [13]. But designing synchro-
The fast SLS-MSDD algorithm can conveniently make
nization algorithms for UWB receivers (that in practice
salient tradeoffs in performance versus complexity via the
means to identify at frame level where the first frame in
choice of the data block size and the (floating-point or
each symbol starts, and then, to find at the pulse level
integer-based) arithmetic format adopted in the computa-
where a pulse is located within a frame) is quite a real and
tion of the SD metrics. Extensive simulation results under
demanding task, exacerbated also by the adoption of ex-
tremely narrow and low-amplitude pulses. This is demon- typical multipath indoor propagation environments cor-
strated by the large effort spent so far in the literature roborate the effectiveness of our scheme.
on this topic, for example, [14–19]. These synchroniza- The rest of the paper is organized as follows. Section 2
tion techniques offer adequate estimation accuracy in the presents the UWB system model, Section 3 derives the
presence of both dense multipath and severe MAI level. novel SLS-MSDD receiver, while the fast algorithm based
Nevertheless, the strong limitation they incur is the ne- on SD is developed in Section 4. Section 5 is devoted
cessity of training sequences or accumulation of long seg- to evaluating performance and complexity via computer
ments of the received signal that often combines with a simulations, followed by concluding remarks in Section 6.
heavy computational load. Clearly, such issues disagree
with the UWB philosophy that calls for as simple and fast 2. System model
as possible receiver processing schemes.
Starting from the baseline illustrated above, one is well In UWB impulse radio signaling, each symbol is
motivated to make a further step toward an efficient de- conveyed over a block of Nf frames with one pulse p(t )
tector that ensures competitive performance levels at even per frame. The symbol, frame and pulse durations are
lower complexity while avoiding both channel estima- denoted as Ts , Tf and Tp respectively, satisfying Ts = Nf Tf ,
tion and accurate timing synchronization. To this end, this Tf  Tp , and Tp being on the order of (sub-)nanoseconds.
paper develops a new UWB detector within the MSDD
framework under the relaxed assumption that timing in-
formation is roughly acquired at the symbol level only, 4 Accurate timing recovery is mandatory in conventional UWB
or in other words, the initial timing resolution can be as receivers so as to maintain adequate system throughput and capacity
large as one symbol period [20]. Bypassing accurate tim- performance [21,22].
ing estimation, we come up with a novel multi-symbol de- 5 This is due to improved averaging over the noise and MAI
tection scheme, in the sequel referred to as MSDD based components.
298 V. Lottici et al. / Physical Communication 2 (2009) 296–305

Concurrent channel access is enabled by employing be explicitly estimated during detection in order to reduce
user-specific pseudo-random time hopping (TH) codes the overall receiver complexity; (iv) the composite noise
N −1
{cj }j=f 0∈ [0, Nc − 1], which time-shifts pulse positions w(t ), including both ambient noise and MAI, is modeled as
at multiples of the chip period Tc , with Nc Tc < Tf . Assume a wide sense stationary white Gaussian process with two-
that pulse amplitude modulation (PAM) is adopted and the sided power spectral density N0 /2.
independent information-bearing symbols ai ∈ {±1} are To cope with the lack of accurate pulse-level timing
transformed into the channel symbols bi ∈ {±1} through information, our basic idea is to partition the received
the differential encoding rule bi = ai bi−1 . The transmitted symbol-level waveform us (t ) in (4) into the two segments
signal relevant to a burst of M information symbols can be 0, t ∈ [0, τ )

(0)
written as us (t ) , ,
us (t − τ ), t ∈ [τ , Ts )
M
us (t + Ts − τ ), t ∈ [0, τ )
X 
x(t ) = bi ps (t − iTs ), (1) u(s1) (t ) , , (6)
i =0
0, t ∈ [τ , Ts )

where the symbol-long waveform ps (t ) is both of which rely on the unknown timing offset τ .
Making use of (6) and expressing the differentially-
N f −1
encoded channel symbol as bi = b0 k=1 ak , i > 0, the
Qi
X
ps (t ) = p(t − jTf − cj Tc ). (2) received signal (3) can be put in the following alternative
j =0 form
M M +1
After traveling through a slow-fading multipath channel,
bi u(s0) (t − iTs ) + bi−1 u(s1) (t − iTs ) + w(t )
X X
assumed to be time-invariant within each block and with y(t ) =
L paths each with gain αl and delay τl , the received signal i=0 i=1

in the interval 0 ≤ t ≤ (M + 1)Ts can be written according M Y

X i M +1 i−1
XY
to (1) as = ak q(t − iTs ) + ak g (t − iTs ) + w(t ), (7)
i=0 k=0 i=1 k=0
M
(0) (1)
where both q(t ) , b0 us (t ) and g (t ) , b0 us (t ) have non-
X
y(t ) = bi us (t − iTs − τ ) + w(t ). (3)
i=0 zero support Ts and contain the channel parameters, the
timing offset and the initial channel symbol b0 .
The timing offset τ in (3) is the delay of the first path
Now, the fact that q(t ) and g (t ) are both unknown to
due to the signal propagation from the transmitter to
the receiver suggests detecting the information symbols a
the receiver, the additive noise component w(t ) accounts
following the GLRT rule. This amounts to maximizing the
for the contribution of both the thermal noise and MAI,
log-likelihood metric (LLM)
whereas
Z (M +1)Ts
Λ y(t )|ã, q̃(t ), g̃ (t ) = 2 y(t )s̃(t )dt
N f −1
 
X
us ( t ) = u(t − jTf − cj Tc ) (4) 0
j =0 Z (M +1)Ts
− s̃2 (t )dt (8)
is the received symbol-level waveform with non-zero 0
support less than Ts , i.e., ISI-free condition is satisfied, T
over ã = ã1 , ã2 , . . . , ãM and the finite-energy functions

depending on the channel impulse response through
q̃(t ) and g̃ (t ) with support in [0, Ts ], where
L−1
X
u(t ) = αl p(t − τl,0 ), (5) M Y
X i M +1 i −1
XY
l =0 s̃(t ) = ãk q̃(t − iTs ) + ãk g̃ (t − iTs ) (9)
i=0 k=0 i=1 k=0
where τl,0 , τl − τ .
is the signal corresponding to the trial values of ã, q̃(t ) and
3. Multiple symbol differential detection with symbol- g̃ (t ). Exploiting the finite support of q̃(t ) and g̃ (t ) in [0, Ts ),
level synchronization we can obtain
Z (M +1)Ts Z Ts

In this section, we derive the structure of a novel y(t )q̃(t − iTs )dt = y(t + iTs )q̃(t )dt , (10)
0 0
MSDD scheme that aims at recovering M consecutive
differentially-encoded information symbols a , [a1 , a2 ,
Z (M +1)Ts Z Ts

. . . , aM ]T from the received signal y(t ) in the interval 0 ≤ y(t )g̃ (t − iTs )dt = y(t + iTs )g̃ (t )dt , (11)
0 0
t ≤ (M + 1)Ts . The following main assumptions will be
adopted: (i) the timing offset τ is assumed to be within the and
interval [0, Ts ), thus meaning that the timing information (M +1)Ts
Z
is acquired at symbol-level only through some form of q̃(t − iTs )g̃ (t − jTs )dt
0
(rough) coarse synchronization; (ii) the data block size
(M + 1)Ts is smaller than the channel coherence time so
Z Ts
q̃(t )g̃ (t )dt ,

that hereinto the channel is considered as time-invariant; i=j
= 0
(12)
(iii) the channel impulse response is unknown and will not 0, i 6= j.

 V. Lottici et al. / Physical Communication 2 (2009) 296–305 299

Hence, using (10)–(11) yields them to zero, we get

(M +1)Ts
∂ Λ y(t )|ã, q̃(t ), g̃ (t )
Z   
y(t )s̃(t )dt 

0
∂λ



 λ=0
µ=0
" #
Z Ts M Y
X i M +1 i−1
XY 

q̃(t ) ãk y(t + iTs ) + g̃ (t ) ãk y(t + iTs ) dt , (13) T
Z s 
=


z1 (t ; ã) − q0 (t ) − η(ã)g0 (t ) ε(t )dt
 
0 i=0 k=0 i=1 k=0


 =2
0

= 0, ∀ε(t )


whereas according to (12), we get (20)
∂ Λ y(t )|ã, q̃(t ), g̃ (t )
 


(M +1)Ts Ts

∂µ
Z Z 
s̃2 (t )dt = (M + 1) q̃2 (t ) + g̃ 2 (t ) dt
  
λ=0


0 0

 µ=0

 Z Ts
(M +1)Ts
 = 2 0 z2 (t ; ã) − g0 (t ) − η(ã)q0 (t ) ρ(t )dt
M M +1 Y
i j−1 Z 
  
X X Y
q̃(t − iTs )g̃ (t − jTs )dt

+2 ãk ãl 

i=0 j=1 k=0 l=0 0
= 0, ∀ρ(t ).
Z Ts M Z Ts
This equation set is satisfied whenever
X
= (M + 1) q̃2 (t ) + g̃ 2 (t ) dt + 2 q̃(t )g̃ (t )dt . (14)
 
ãi
0 i=1 0

q0 (t ) + η(ã)g0 (t ) = z1 (t ; ã)

(21)
Therefore, substituting (13) and (14) into (8), the LLM η(ã)q0 (t ) + g0 (t ) = z2 (t ; ã),
takes the form (we drop the influent multiplicative factor
M + 1) that is,

Λ y(t )|ã, q̃(t ), g̃ (t ) z1 (t ; ã) − η(ã)z2 (t ; ã)

 
q0 (t ) = , (22)
1 − η2 (ã)
Z Ts
q̃(t )z1 (t ; ã) + g̃ (t )z2 (t ; ã) dt
 
=2
0
z2 (t ; ã) − η(ã)z1 (t ; ã)
Z Ts Z Ts g0 (t ) = . (23)
q̃ (t ) + g̃ (t ) dt − 2η(ã)
2 2
q̃(t )g̃ (t )dt , 1 − η2 (ã)
 
− (15)
0 0
Then, substituting the solutions (22)–(23) of (21) into (15)
where yields (up to an irrelevant multiplicative factor)
Z Ts
M Y
i
Γ y(t )|ã = z12 (t ; ã) + z22 (t ; ã) − 2η(ã)z1 (t ; ã)z2 (t ; ã) dt .
   
1 X (24)
z1 (t ; ã) , ãk y(t + iTs ), t ∈ [0, Ts ) , (16) 0
M + 1 i=0 k=0
Consequently, in line with (19) and with Γ y(t )|ã given
 
M +1 i−1
1 XY by (24), the proposed SLS-MSDD detection rule becomes
z2 (t ; ã) , ãk y(t + iTs ), t ∈ [0, Ts ) , (17)
M + 1 i=1 k=0
â = arg max Γ y(t )|ã .
  
(25)
ã
and
A few remarks about (24)–(25) are now in order.
M
1 X
η(ã) , ãi . (18) (1) Let us take into account the optimal estimates of the
M + 1 i =1 unknown waveform segments q(t ) and g (t ) given by
(22) and (23), respectively. Making use of
Accordingly, the GLRT-based decision strategy on the i
information symbols a can be formulated as
Y
y(t + iTs ) = [q(t ) + ai g (t )] ak + w(t + iTs ),
  k =0

Λ y(t )|ã, q̃(t ), g̃ (t ) . t ∈ [0, Ts ) ,

  
â = arg max max (19) (26)
ã q̃(t ),g̃ (t )
that can be easily derived from (7) and (16)–(17) turn
out to be
In order to solve (19), we first keep ã fixed and compute
the inner term 1
M Y
X i i
Y
z1 (t ; ã) = al [q(t ) + ai g (t )]
ãk
Γ y(t )|ã , max Λ y(t )|ã, q̃(t ), g̃ (t ) .
     M + 1 i=0 k=0 l=0
q̃(t ),g̃ (t ) M Y i
1 X
+ ãk w(t + iTs )
Toward this end, we can resort to standard variational M + 1 i=0 k=0
techniques by letting q̃(t ) = q0 (t ) + λε(t ) and g̃ (t ) = = ϕ(ã)q(t ) + ξ (ã)g (t )
g0 (t )+µρ(t ), q0 (t ) and g0 (t ) being the optimum solutions M Y i
1
to be found, and ε(t ) and ρ(t ) two generic functions with
X
+ ãk w(t + iTs ),
support in [0, Ts ). After taking the first-order derivatives of
M + 1 i=0 k=0
Λ y(t )|ã, q̃(t ), g̃ (t ) with respect to λ and µ and setting t ∈ [0, Ts ) ,

(27)
300 V. Lottici et al. / Physical Communication 2 (2009) 296–305

M +1 i−1 i−1 is asymptotically large and so the assumption ã = a

1 X Y Y
z2 (t ; ã) = [g (t ) + ai q(t )] ãk al stands, evaluation of (27)–(28) leads to
M + 1 i=1 k=0 l =0 Z Ts
M +1 Y
i−1
1 Γ̄ y(t )|ã ã=a = q2 (t ) + g 2 (t ) dt
X    
+ ãk w(t + iTs ) 0
M + 1 i=1 k=0
Ts
= ϕ(ã)g (t ) + ζ (ã)q(t )
Z
= u2s (t )dt , (32)
M +1 i−1
1 XY 0
+ ãk w(t + iTs ),
M + 1 i=1 k=0 that is, the maximum value of (31) in the asymptotic
case is given by the energy of the received symbol-level
t ∈ [0, Ts ) , (28)
waveform us (t ), independent of both the data symbol
where sequence a and the timing offset τ .
1
M Y
X i i
Y (3) The SLS-MSDD circumvents both the explicit esti-
ϕ(ã) , ãk al , mation of the channel parameters and the need for
M + 1 i=0 k=0 l =0 accurate frame-level (or even pulse-level) timing syn-
M i i chronization. The metric (31) to be maximized, indeed,
1
relies on the energy of z1 (t ; ã) and z2 (t ; ã), which can
X Y Y
ξ (ã) , ai ãk al
M + 1 i =0 k=0 l=0 be constructed entirely from the received signal y(t ),
and as shown in (16)–(17).
(4) Since the information symbols take values in {±1}, the
M i i
1 X Y Y metric (31) can be further rearranged to an equivalent
ζ (ã) , ai+1 ãk al .
M + 1 i=0 form that is suited for practical implementations of
k=0 l=0
the SLS-MSDD. Similarly to the approach in [13], it is
Substituting (27)–(28) into (22)–(23) and assuming a possible to show that
high SNR, it can be found that
i−1 Y
M X i −l
ϕ(ã) − η(ã)ζ (ã) q(t ) + ξ (ã) − η(ã)ϕ(ã) g (t )
     X
Γ̄ y(t )|ã = ãk+l (Yl,i + Yl+1,i+1 ),

q0 (t ) = , (29) (33)
1 − η2 (ã) i=1 l=0 k=1
ϕ(ã) − η(ã)ξ (ã) g (t ) + ζ (ã) − η(ã)ϕ(ã) q(t )
   
g0 (t ) = . (30) where the coefficients
1 − η2 (ã) Z Ts
1
Yi,j , y(t + iTs )y(t + jTs )dt (34)
It can be observed from (29)–(30) that, for ã 6= a, the M +1 0
optimal estimates q0 (t ) and g0 (t ) differ from the true
are obtained by correlating symbol-long segments
q(t ) and g (t ) respectively. On the other side, whenever
of the received signal y(t ) up to M symbols apart.
ã = a we get ϕ(a) = 1 and ξ (a) = ζ (a) = η(a), and
It is worth noting that generation of {Yi,j } requires
hence, the estimates q0 (t ) and g0 (t ) coincide with the
waveform storage with delays on the order of multi-
desired segments q(t ) and g (t ), regardless of the val-
ples of symbol intervals. This can be practically done
ues of η(a).
via either analog averaging or digital sampling, even
(2) The occurrence of η(ã) in the objective function (24)
though both of them have drawbacks in terms of hard-
adds considerably to the computational complexity of
ware implementation. Analog averaging requires long
the detector. To simplify the implementation of the
high-bandwidth delay lines that are hard to realize.
SLS-MSDD, a practical approach is to approximate η(ã)
Digital sampling requires accurate high-rate ADCs at
by zero6 and remove it from (24), resulting in the
Nyquist rates, which can be costly in terms of chip
simplified objective function
area and power consumption. Nevertheless, the latter
Z Ts approach can be affordable using low-resolution sam-
Γ̄ y(t )|ã = z12 (t ; ã) + z22 (t ; ã) dt ,
   
(31) plers as suggested by the study in [24] on UWB receiver
0 implementation. As an alternative, one might also con-
where z1 (t ; ã) and z2 (t ; ã) are given again by (16)–(17). sider subsampling techniques for direct signal conver-
Adopting (31) as the objective function in lieu of (24) sion [25] or use subsampling combined with sparsity
means detecting the transmitted symbols by search- considerations to reconstruct the high bandwidth
ing for the maximum of the sum of two partial energy signal [26,27], thus trading off performance against
metrics: the first is related to z1 (t ; ã) and the other complexity.
to z2 (t ; ã). More specifically, whenever the SNR value (5) The basic idea of the SLS-MSDD is to jointly detect a
block of M data symbols, within which the channel
response can be considered as time-invariant. As a
6 This approximation is accurate when information-bearing symbols result, the detection accuracy is expected to improve as
are independent and identically distributed, in which case their mean M increases. Finding out the maximum of the objective
value η(ã) approaches 0 as the data size M increases. Nevertheless, there function (33) via some exhaustive search method,
exists a (limited) percentage of symbol sequences for which η(ã) 
1 is not fulfilled. To make this condition more rigorous, some form of
however, requires high computational complexity
information symbol precoding could help, although it is out of the scope going up exponentially in the number of symbols M to
of the current paper. be jointly detected. Therefore, whenever performance
 V. Lottici et al. / Physical Communication 2 (2009) 296–305 301

(m)
has to be attained at affordable complexity, efficient estimate â to lie inside the sphere of radius δ (m) > 0 is
implementations of the SLS-MSDD are inevitably given by
called for, which will be the focus of next section. " #
j X
i−1 i −l
(m)
Zl,i ≤ δ (m) ,
X Y
1 − σl,i âk+l
4. Sphere decoding for SLS-MSDD i=1 l=0 k=1

1 ≤ j ≤ M, (39)
An effective way to elude the expensive computational
load required by the exhaustive-search-based SLS-MSDD is or more explicitly,
to take advantage of the SD algorithm. Originally proposed
(m)
h i
to solve the so-called shortest vector problem (SVP) in a
j = 1 : 1 − σ0,1 â1 Z0,1 ≤ δ (m) ; (40a)
lattice [23], the SD basically relies upon examining only
those lattice points (assumed to belong to a given finite- (m) (m) (m)
h i h i
j = 2 : 1 − σ0,1 â1 Z0,1 + 1 − σ0,2 â1 â2 Z0,2
alphabet) inside a sphere of radius δ , which is progressively
made smaller and smaller to reduce the search space,
+ 1 − σ1,2 â(2m)
h i
thereby lessening the overall computational complexity. Z1,2 ≤ δ (m) ; (40b)
The SLS-MSDD detection rule developed so far, however, ..
does not yet fit in the SD framework, and accordingly, a .
proper reformulation of the objective function Γ̄ y(t )|ã
 " #
i−1
M X i −l
(m)
Zl,i ≤ δ (m) .
X Y
in (33) is required. To this end, a key observation we j=M : 1 − σl,i âk+l (40c)
make here is that the maximum possible value of (33) is i =1 l =0 k =1
independent of ã as long as they take values in {±1}, that is,
Interesting to note, the condition (40a) for j = 1 contains
i−1
M X (m) (m) (m)
X â1 only, (40b) concerning j = 2 contain â1 and â2
Γ̄MAX [y(t )] , Yl,i + Yl+1,i+1 . (35) only, and so on. This suggests checking the M conditions
i=1 l=0
in (39) one by one at each iteration of the SD algorithm, as
Hence, subtracting (33) from (35), the new objective summarized in the following steps.
function (this time to be minimized) becomes (m)
S1 : starting with (40a), the candidate set for â1 can be
i−1
M X
X found as
Φ y(t )|ã = ϑl,i Yl,i + Yl+1,i+1 ,
 
(36) (m)
= ã1 ∈ {±1}|(1 − σ0,1 ã1 ) Z0,1 ≤ δ (m) ;

i =1 l =0 I1 (41)

where (m) (m)

S2 : after the tentative â1 has been chosen from I1 , it is
i −l substituted into (40b) to produce the candidate set for
(m)
Y
ϑl,i , 1 − σl,i ãk+l (37) â2 as
k =1
(m)
= ã2 ∈ {±1} 1 − σ0,1 â(1m) Z0,1
n h i
I2
takes values in {0, 2} depending on whether σl,i ,
+ 1 − σ0,2 â(1m) ã2 Z0,2 + 1 − σ1,2 ã2 Z1,2
h i
sign{Yl,i + Yl+1,i+1 } has the same or opposite sign with  
Qi−l
respect to k=1 ãk+l . According to (36)-(37), the SLS-MSDD
can thus be put in the alternative form
o
≤ δ (m) ; (42)
â = arg min Φ y(t )|ã ,
  
(38)
ã ..
.
which is structurally amenable to be put into practice (m) (m)
Sj : the candidate set Ij for âj is based on the tentative
through the SD search algorithm for the following reasons: (m) (m)
(i) the objective function (36) is nothing but the sum of the decisions {â1 , â2 , . . . , â(j−m1) } and equals
non-negative coefficient Zl,i , with Zl,i , Yl,i + Yl+1,i+1 , (m)
= ãj ∈ {±1}| 1 − σ0,1 â(1m) Z0,1
n h i
weighted linearly by the non-negative integer-valued un- Ij
knowns ϑl,i ; (ii) the i-th addend in (36), l=0 ϑl,i Zl,i , de-
Pi−1
+ 1 − σ0,2 â(1m) â(2m) Z0,2
h i
pends on the preceding tentative symbols up to ãi , namely,
ã1 , ã2 , . . . , ãi ; (iii) in view of (i)–(ii), (36) defines a sphere
+ 1 − σ1,2 â(2m) Z1,2 + · · ·
h i
in the M-dimensional lattice of the trial vectors ã , [ã1 , ã2 ,
. . . , ãM ]T [28], which fully complies with the SD frame-
+ 1 − σ0,j â1(m) â(2m) · · · ãj Z0,j
h i
work.
Based on the above observations, we now illustrate
+ 1 − σ1,j â2(m) â(3m) · · · ãj Z1,j
h i
how to realize the SD-based SLS-MSDD. For the sake of
simplicity, assume that the initial radius δ (1) > 0 is chosen
+ 1 − σj−1,j ãj Zj−1,j ≤ δ (m) ;
 
to be large enough so that the sphere defined by (36) (43)
contains the optimal â to be searched for. As the generic ..
m-th SD iteration, a necessary condition for any tentative .
302 V. Lottici et al. / Physical Communication 2 (2009) 296–305

(m) (m)
SM : the last candidate set IM is acquired for âM , which environments through computer simulations. First, the
concludes the m-th iteration. Then, the radius δ (m) and BER performance is quantified as a function of the ra-
the optimal estimate âopt are updated according to the tio Eb /N0 in both single-user and MAI scenarios, Eb being
(m) defined as the mean received energy per transmitted
new tentative estimate â
bit. Next, some implementation issues are investigated
(m)
h i
δ (m+1) ← Φ y(t )|â , (44a) with particular emphasis on the savings in computational
complexity which can be obtained with respect to the
(m)
âopt ← â , (44b) SLS-MSDD based on exhaustive search (ES). The follow-
ing conventional schemes are taken as performance bench-
and further, the next (m + 1)-th iteration follows.
marks: (i) single-user Rake receiver with perfect channel
The iterations go on with a smaller and smaller sphere, state information (CSI) and ideal timing recovery (IRake);
(m)
with the candidate estimate â found in the previous (ii) single-user symbol-by-symbol DD with ideal timing re-
iteration lying on its surface. When at a given iteration all covery (IDD); (iii) single-user symbol-by-symbol DD with
nodes have been visited without any success, the iterative symbol-level synchronization (SLS-DD).
process stops, yielding the optimal solution âopt for
which
 the objective function attains the minimum value 5.1. Simulation setup
Φ y(t )|âopt .

Several remarks are now of interest.
In the system setup for all simulation runs, each ac-
(1) The proposed SD-based SLS-MSDD shares with the SD tive user transmits consecutive bursts of M binary PAM
of the existing literature optimal ML performance at information-bearing symbols. The transmission channel is
polynomial complexity (often cubic or higher) in the assumed to be time-invariant within each burst, but ran-
data block size over a wide range of practical SNRs, as domly varying from burst to burst according to the model
will be verified in Section 5.4. in [29]. To be specific, the multipath components arrive in
(2) At a given iteration index m and for some symbol clusters with amplitudes modeled as independent double-
(m) (m)
index j, the candidate set Ij for âj might be empty. sided Rayleigh distributed random variables having mean
In that case, a new iteration starts immediately by square values exponentially decaying with the cluster de-
decrementing the index j and testing a different value lays, as well as with the ray delays within the cluster,
(m+1)
for âj . with decay factors chosen as 30 ns and 5 ns, respectively.
(3) The choice of the initial radius δ (1) is critical, and can The clusters and the rays within each cluster have Pois-
add to the overall complexity if not properly addressed. son distributed arrival times with arrival rates 0.5 ns−1
Indeed, if it is chosen too small, no point lies inside and 2 ns−1 , respectively. The monocycle p(t ) is selected
the sphere and the algorithm must restart with a larger as the second derivative of a Gaussian shape with normal-
radius, whereas, if chosen too large, many more points ized unit energy and pulse width equal to Tp = 1.0 ns. The
have to be checked. A reasonable choice is to set the frame and chip interval are Tf = 100 ns and Tc = 1.0 ns,
initial radius by evaluating (36) in correspondence respectively, Nf = 10 is the number of frames per infor-
with a given tentative sequence, for instance that with mation symbol, while the TH codes cj are randomly picked
pseudo-random binary distributed entries. in the interval [0, Nc − 1] with Nc = 91, so that cj Tc <
(4) To speed up the search procedure, at each iteration âj
(m) Tf , ∀j ∈ [0, Nf − 1]. In the MAI scenario, the Nu − 1 inter-
(m) fering users have the same power Pi , whereas the desired
is picked up from Ij so that the metric to be evalu- one has power Pu , so that the near/far ratio (NFR) results
ated in (39) is minimized. So doing, the SD search space as NFR , Pi /Pu . Further, the time origins of all the active
is reduced, and accordingly, computational complexity users are set randomly over the symbol interval (0, Ts ) to
gets remarkably lower. reproduce an asynchronous access to the channel and in
(5) Since the unknown ϑl,i takes integer values in {0, 2}, line with the assumption that the timing information of the
checking the M conditions in (39) requires no mul- desired user is acquired at the symbol level only.
tiplications, but simply real additions combined with
logical operations.
(6) The real-valued coefficients Zl,i depend on the re- 5.2. BER in the single-user scenario
ceived signal through (34), and therefore, can be
pre-computed before the SD iterations start. As an Fig. 1 illustrates the BER performance of the SLS-MSDD
alternative, we can employ only the sign of Zl,i by sub- in a single-user scenario (Nu = 1) with the coefficients Zl,i
stituting Zl,i = 1 in (40a)–(40c). This strategy leads to taken as real-valued values (soft SD), and for different block
the one-bit hard-quantized version of the SLS-MSDD, sizes, namely, M = 5, 10, 15, 20, 30, 35. As expected, the
which allows a further reduction in complexity in that performance of the SLS-MSDD improves as the block size
simpler integer-format additions and logical opera- M increases: at BER = 10−2 , the performance gap between
tions are now required. the M = 5 scheme and that with M = 10 is more than 6
dB, while M = 35 offers an additional gain of 5 dB over
5. Performance results M = 10. When M is not large enough, however, the BER
curves exhibit noise floors in the high SNR region, due to
In this section, the effectiveness of the SD-based SLS- insufficient averaging that compromises the assumption
MSDD receiver is verified over dense multipath wireless η(a) ≈ 0 we made to reduce the detector complexity.
 V. Lottici et al. / Physical Communication 2 (2009) 296–305 303

1 1

0.1 0.1

0.01 0.01

0.001 0.001

0.0001 0.0001
6 8 10 12 14 16 18 20 22 24 26 28 30 6 8 10 12 14 16 18 20 22 24 26 28 30

Fig. 1. BER of SLS-MSDD with soft SD metrics for Nu = 1 and various M. Fig. 2. BER of SLS-MSDD with hard SD metrics for Nu = 1 and various M.

The above phenomenon is considerably alleviated as M

1
increases. Still, the SLS-MSDD has a significant edge
compared to the conventional DD scheme not only with
mistiming (SLS-DD) but also with ideal timing recovery
(IDD). In fact, the gain of the SLS-MSDD with M = 35 over
the IDD is more than 5 dB in the BER interval of practical
interest while skipping accurate timing synchronization. 0.1
On the other side, the IRake outperforms the SLS-MSDD by
approximately 10 dB, but at a very expensive price paid for
accurate channel and timing estimation.
Fig. 2 refers to the SLS-MSDD in the single-user case
with one-bit hard-quantized coefficients Zl,i (hard SD). In 0.01
accordance with Fig. 1, the BER metrics improve with the
data block size M, flatten out in the high-SNR region at
moderate M, and are still superior to those of both the SLS-
DD and IDD. Nevertheless, the considerable reduction in
computational complexity due to the adoption of integer-
based arithmetic only has to be inevitably traded with a 0.001
given performance degradation. If compared to the soft
version, this is limited, however, to 1–2 dB only, providing
that the burst length takes adequate values, say M ≥ 15.

5.3. BER in the MAI scenario 0.0001

6 8 10 12 14 16 18 20 22 24 26 28 30
The BER curves of Fig. 3 quantify the MAI effects on the
soft version of the SLS-MSDD when the block size is M =
Fig. 3. BER of SLS-MSDD with soft SD metrics for M = 20, 30, NFR = −6
20 (square marks) or M = 30 (circle marks), the number
dB, and various Nu .
of active users is Nu = 5, 10, 15 and the NFR parameter is
set to −6 dB. The single-user cases for both the SLS-MSDD
and the conventional IDD are also plotted as reference. As whereas 2 dB has to be expended additionally when the
one could expect, the BER performance gets worse as the MAI level raises to Nu = 15. Further, the BER robustness of
MAI level increases. At BER = 10−2 , the SLS-MSDD with the SLS-MSDD against the MAI improves adopting a longer
M = 30 can sustain Nu = 5 users at a small price of less data block. At BER = 10−2 , the SLS-MSDD using M = 20
than 0.5 dB in extra SNR (compared to the single-user case), outperforms the IDD operating in a single-user channel up
304 V. Lottici et al. / Physical Communication 2 (2009) 296–305

1 10

8
0.1
7

0.01 5

3
0.001
2

0.0001 0
6 8 10 12 14 16 18 20 22 24 26 28 30 5 10 15 20 25 30 35

Fig. 5. Complexity exponent of SLS-MSDD with soft and hard SD metrics

Fig. 4. BER of SLS-MSDD with soft SD metrics for M = 20, 30, NFR = −3
for Nu = 1.
dB, and various Nu .

complexity of the SLS-MSDD (i) significantly improves over

to Nu = 10 users, whereas the same can be achieved at a
that of the ES-based scheme, especially for larger values
lower BER of 10−3 as well, as long as M = 30 is employed.
of M; (ii) decreases while increasing Eb /N0 for given M,
When the NFR in multi-user scenario increases to
in that the iterative search on the tree gets faster when
−3 dB, the MAI effects turn out to be much stronger, as
the noise level reduces; (iii) flattens out for M larger
shown in Fig. 4. When Nu = 5, the SLS-MSDD can reach
than around 15, thus meaning that (due to the adopted
the BER level of 10−2 only provided that the block size is
logarithmic definition) it is nearly polynomial in M at a
set to M = 30, while choosing M = 20 the BER increases
constant exponent; (iv) decreases when using soft metrics
to (asymptotically) around 2 × 10−2 . Setting the number
instead of hard ones, for given M and Eb /N0 . This is because
of active users to Nu = 15, instead, significantly degrades
the search process in soft SD converges faster thanks to its
the SLS-MSDD performance in that the BER plots for both
performance advantage, at the price of using floating-point
M = 20 and M = 30 flatten out above the 0.1 level.
rather than integer format computations.

5.4. Computational complexity 6. Concluding remarks

The computational load of the SD-based SLS-MSDD can The aim of the current paper has been to derive
be derived from the total number of additions spent in a novel multi-symbol detector for UWB communica-
evaluating the conditions (39) throughout the iterations tions with remarkable robustness to mistiming. A num-
required by the search procedure. Depending on not only ber of appealing features are offered: (i) joint data
the data block size M (or equivalently, how many nodes detection based on symbol-level synchronization only,
are included in the tree to be visited) but also the received (ii) simple receiver structure by circumventing the costly
signal (which affects the sphere radius chosen at each tap-by-tap channel estimation required by conventional
iteration), the complexity is a random variable that has to coherent schemes, and (iii) efficient implementation by re-
be properly described through its average properties. In the sorting to a sphere decoding approach which enables af-
sequel, adhering to the approach followed in [28], we will fordable computational complexity even for large blocks.
refer to the so-called complexity exponent (CE), defined Simulation results obtained for typical dense multipath
as the logarithm of base M of the average number of flop channels in both single-user and MAI-limited scenarios in-
operations per data block. dicate that the adoption of a proper block size enables
The CE metric is illustrated in Fig. 5 as a function of attractive performance-versus-complexity tradeoffs.
M and for various Eb /N0 ratios, for both the soft (solid References
lines) and hard (dotted lines) versions of the SLS-MSDD
operating in the single-user scenario. We also included the [1] R.A. Scholtz, M.Z. Win, Impulse radio: How it works, IEEE Commun.
Lett. 2 (2) (1998) 36–38.
CE of the SLS-MSDD based on ES (solid line without marks) [2] L. Yang, G.B. Giannakis, Ultra-wideband communications: An idea
as a performance benchmark. The results suggest that the whose time has come, IEEE Signal Process. Mag. 21 (6) (2004) 26–54.
 V. Lottici et al. / Physical Communication 2 (2009) 296–305 305

[3] R.J.-M. Cramer, R.A. Scholtz, M.Z. Win, Evaluation of an ultra-wide- [28] B. Hassibi, H. Vikalo, On sphere decoding algorithm. I. Expected
band propagation channel, IEEE Trans. Antennas Propag. 50 (5) complexity, IEEE Trans. Signal Process. 53 (8) (2005) 2806–2818.
(2002) 561–570. [29] A.A.M. Saleh, R.A. Valenzuela, A statistical model for indoor
[4] R.C. Qiu, A study of the ultra-wideband wireless propagation channel multipath propagation, IEEE J. Sel. Areas Commun. 5 (2) (1987)
and optimum UWB receiver design, IEEE J. Sel. Areas Commun. 20 (9) 128–137.
(2002) 1628–1637.
[5] M.Z. Win, R.A. Scholtz, On the energy capture of ultra-wideband
bandwidth signals in dense multipath environments, IEEE Commun.
Lett. 2 (9) (1998) 245–247. Vincenzo Lottici received the Dr. Ing. degree
[6] V. Lottici, A. D’Andrea, U. Mengali, Channel estimation for ultra- (cum Laude) in electronic engineering and the
wideband communications, IEEE J. Sel. Areas Commun. 20 (12) ‘‘Best Thesis SIP Award’’ from the University
(2002) 1638–1645. of Pisa in 1985 and 1986, respectively. From
[7] R. Hoctor, H. Tomlinson, Delay-hopped transmitted-reference RF 1987 to 1993 he worked in the research and
communications, in: IEEE UWBST, 2002 pp. 265–269. development of digital signal processing for
[8] Y.-L. Chao, R.A. Scholtz, Optimal and suboptimal receivers for UWB sonar systems. Since 1993 he has been with
transmitted reference systems, in: IEEE Globecom 2003, vol. 2, 2003, the Department of Information Engineering of
pp. 759–763. the University of Pisa, where he is currently
[9] L. Yang, G.B. Giannakis, Optimal pilot waveform assisted modulation a Research Fellow and Assistant Professor in
for ultra-wideband communications, IEEE Trans. Wireless Commun. Telecommunications. He participated to several
3 (4) (2004) 1236–1249. international and national research projects, and specifically, from 2003
[10] M. Ho, S. Somayazulu, J. Foerster, S. Roy, A differential detector for he has been taking an active role in the standardization activity promoted
UWB communications system, in: IEEE VTC 2002, vol. 4, May 2002, by ETSI (European Telecommunications Standards Institute) of the
pp. 1896–1900. physical layer of the mobile radio system TETRA release 2. His research
[11] V. Lottici, Z. Tian, A sphere decoding approach to multiple symbols interests include the area of wireless multicarrier and UWB systems
differential detection for UWB systems, in: IEEE Globecom 2006, at large, with particular emphasis on synchronization and channel
Nov. 2006. estimation techniques.
[12] V. Lottici, L. Wu, Z. Tian, Inter-symbol interference mitigation in
high-data-rate UWB systems, in: IEEE ICC 2007, June 2007. Zhi Tian received the B.E. degree in Electrical
[13] V. Lottici, Z. Tian, Multiple symbol differential detection for UWB Engineering from the University of Science and
communications, IEEE Trans. Wireless Commun. 7 (5) (2008) Technology of China, Hefei, China, in 1994, the
1656–1666. M. S. and Ph.D. degrees from George Mason Uni-
[14] Z. Tian, V. Lottici, Low-complexity ML timing acquisition for UWB versity, Fairfax, VA, in 1998 and 2000. From 1995
communications in dense multipath channels, IEEE Trans. Wireless to 2000, she was a graduate research assistant
Commun. 4 (6) (2005) 3031–3038. in the Center of Excellence in Command, Con-
[15] Z. Tian, G.B. Giannakis, A GLRT approach to data-aided timing trol, Communications and Intelligence (C3I) of
acquisition in UWB radios – Part I: Algorithms, IEEE Trans. Wireless George Mason University. Since August 2000,
Commun. 4 (6) (2005) 2956–2967. she has been with the department of Electri-
[16] L. Yang, G.B. Giannakis, Timing ultra-wideband signals with dirty cal and Computer Engineering, Michigan Tech-
templates, IEEE Trans. Commun. 53 (11) (2005) 1952–1963. nological University, where she is an Associate Professor. Her current
[17] L. Wu, X. Wu, Z. Tian, Asymptotically optimal uwb receivers with research focuses on signal processing for wireless communications, par-
noisy templates: Design and comparison with rake, IEEE J. Sel. Areas ticularly on ultra-wideband systems, cognitive radios and distributed
Commun. 24 (4) (2006) 808–814. sensor networking. Dr. Tian serves as an Associate Editor for IEEE Trans-
[18] X. Luo, G.B. Giannakis, Low-complexity blind synchronization and actions on Wireless Communications and IEEE Transactions on Signal
demodulation for (ultra-) wideband multi-user ad hoc access, IEEE Processing. She is the recipient of a 2003 NSF CAREER award.
Trans. Wireless Commun. 5 (7) (2006) 1930–1941.
[19] X. Luo, G.B. Giannakis, Raise your voice at a proper pace to
synchronize in multiple ad hoc piconets, IEEE Trans. Signal Process. Geert Leus was born in Leuven, Belgium, in
55 (1) (2007) 267–278. 1973. He received the electrical engineering de-
[20] V. Lottici, Z. Tian, G. Leus, A synchronization-free approach to data gree and the Ph.D. degree in applied sciences
recovery for multiple access UWB communications, in: IEEE ICUWB from the Katholieke Universiteit Leuven, Bel-
2008, Sept. 2008. gium, in June 1996 and May 2000, respectively.
[21] Z. Tian, G.B. Giannakis, BER sensitivity to mistiming in ultra- He has been a Research Assistant and aPost-
wideband impulse radios, Part I: Nonrandom channels, IEEE Trans. doctoral Fellow of the Fund for Scientific Re-
Signal Process. 53 (4) (2005) 1550–1560. search - Flanders, Belgium, from October 1996
[22] Z. Tian, G.B. Giannakis, BER sensitivity to mistiming in ultra- till September 2003. During that period, Geert
wideband impulse radios, Part II: Fading channels, IEEE Trans. Signal Leus was affiliated with the Electrical Engineer-
Process. 53 (5) (2005) 1897–1907. ing Department of the Katholieke Universiteit
[23] U. Fincke, M. Pohst, Improved methods for calculating vectors of Leuven, Belgium. Currently, Geert Leus is an Associate Professor at the
short length in a lattice, including a complexity analysis, Math. Faculty of Electrical Engineering, Mathematics and Computer Science of
Comput. 44 (1985) 463–471. the Delft University of Technology, The Netherlands. During the summer
[24] S. Hoyos, B.M. Sadler, G.R. Arce, Monobit digital receivers for of 1998, he visited Stanford University, and from March 2001 till May
ultrawideband communications, IEEE Trans. Wireless Commun. 4 2002 he was a Visiting Researcher and Lecturer at the University of Min-
(4) (2005) 1337–1344. nesota. His research interests are in the area of signal processing for com-
[25] M. Chen, R. Brodersen, A subsampling UWB impulse radio archi- munications. Geert Leus received a 2002 IEEE Signal Processing Society
tecture utilizing analytic signaling, IEICE Trans. Electron. E88-C (6) Young Author Best Paper Award and a 2005 IEEE Signal Processing Soci-
(2005) 1114–1121. ety Best Paper Award. He is the Chair of the IEEE Signal Processing for
[26] M. Vetterli, P. Marziliano, T. Blu, Sampling signals with finite rate of Communications Technical Committee, and an Associate Editor for the
innovation, IEEE Trans. Signal Process. 50 (2002) 1417–1428. IEEE Transactions on Signal Processing and the EURASIP Journal on Ap-
[27] Y. Vanderperren, G. Leus, W. Dehaene, A recongurable pulsed uwb plied Signal Processing. In the past, he has served on the Editorial Board
receiver sampling below nyquist rate, in: IEEE ICUWB 2008, Sept. of the IEEE Signal Processing Letters and the IEEE Transactions on Wire-
2008. less Communications.