A COMPRESSED SENSING TECHNIQUE FOR OFDM CHANNEL ESTIMATION IN MOBILE

ENVIRONMENTS: EXPLOITING CHANNEL SPARSITY FOR REDUCING PILOTS

Georg Tauböck and Franz Hlawatsch

Institute of Communications and Radio-Frequency Engineering, Vienna University of Technology
Gusshausstrasse 25/389, A-1040 Vienna, Austria; e-mail: [email protected]

ABSTRACT N × |T | submatrix comprising those columns of Φ that are indexed

We consider the estimation of doubly selective wireless channels by the elements of T. Then the S-restricted isometry constant δS of
within pulse-shaping multicarrier systems (which include OFDM Φ is defined as the smallest quantity δS such that
systems as a special case). A new channel estimation technique us-
ing the recent methodology of compressed sensing (CS) is proposed. (1−δS )a22 ≤ ΦT a22 ≤ (1+ δS )a22
CS-based channel estimation exploits a channel’s delay-Doppler
sparsity to reduce the number of pilots and, hence, increase spectral for all subsets T with |T | ≤ S and all vectors a ∈ R|T |.
efficiency. Simulation results demonstrate a significant reduction of We will use the estimator of u defined by the convex program
the number of pilots relative to least-squares channel estimation.
û  arg min u1 , (2)
u∈U
Index Terms— OFDM, multicarrier modulation, channel estima-
tion, compressed sensing, sparse reconstruction, basis pursuit where U is the set of all u ∈ RM satisfying Φu − v2 ≤  for a
given  > 0 [5]. This estimator—an extension of basis pursuit —is
1. INTRODUCTION able to recover S-sparse parameter vectors according to the follow-
ing result [5].
The recently introduced principle and methodology of compressed
sensing (CS) allows the efficient reconstruction of sparse signals For a given S, assume that the 3S- and 4S-restricted isometry con-
from a very limited number of measurements (samples) [1, 2]. CS stants of Φ satisfy
has gained a fast-growing interest in applied mathematics. In this δ3S + 3δ4S < 2 . (3)
paper, we apply CS to pilot-based channel estimation in highly mo- Let v = Φu + w with w2 ≤ , and let uS ∈ RM contain the S
bile environments. We consider pulse-shaping multicarrier (MC) components of u with largest absolute values, the remaining M −S
systems, which include orthogonal frequency-division multiplexing components being zero. Then the estimate û in (2) satisfies
(OFDM) as a special case [3]. Conventional methods for channel
estimation (e.g., [4]) are not able to exploit the inherent sparsity of u − uS 1
û − u2 ≤ C1  + C2 √ , (4)
the transmission channel that is due to the sparse distribution of scat- S
terers in space. As we will demonstrate, CS provides a constructive
where the constants C1 and C2 depend only on δ3S and δ4S .
way for exploiting this sparsity in order to reduce the number of pi-
lots and, hence, increase spectral efficiency. For a zero-mean i.i.d. Gaussian noise vector, the conditionw 2 ≤
This paper is organized as follows. In Section 2, some basic facts  is satisfied with high probability for an appropriate . In the noise-
about CS are reviewed. The MC system model is described in Sec- less case ( = 0), (4) shows that û = u for any S-sparse vector u.
tion 3. Section 4 studies the delay-Doppler sparsity of doubly selec- Conditions under which the matrix Φ satisfies (3) are obviously of
tive channels. In Section 5, we present the CS-based channel esti- interest. It has been shown [6] that if a complex-valued measurement
 
mation method. Finally, simulation results in Section 6 demonstrate matrix Φc ∈ CN ×M is constructed by selecting uniformly at random
performance gains relative to least-squares channel estimation. [1] N  rows from a unitary M  × M  matrix U and normalizing the
columns (so that they have unit Euclidean norms), a sufficient con-
2. REVIEW OF COMPRESSED SENSING dition for (3) to be true with overwhelming probability1 is
In a typical sparse reconstruction scenario, one tries to estimate a N  ≥ C3 (ln M  )4 μ2 S . (5)
parameter vector u ∈ RM based on the linear model √
Here, μ  M  maxi,j |Ui,j | (known as the coherence of U) and
v = Φu + w , (1)
C3 is a constant. Later, we shall use the fact that under condition
where v ∈ RN is an observation vector, Φ ∈ RN×M is a measure- (5), the real-valued measurement matrix Φ ∈ RN×M (with N = 2N ,
ment matrix, and w ∈ RN is a noise vector. The reconstruction is M = 2M  ) that is obtained from Φc according to
subject to the constraint that u is S-sparse, i.e., at most S of its en- » –
tries are nonzero. The positions of the nonzero entries are unknown. {Φc } −
{Φc }
Φ  (6)
Typically, the number of parameters is much larger than the number
{Φc } {Φc }
of observations, i.e., M  N .
A key ingredient of CS is the uniform uncertainty principle [5], also satisfies (3) with overwhelming probability. This follows from
which essentially states that the measurement matrix Φ obeys a “re- the special structure of Φ (see, e.g., [7, Lemma 1]) and the fact that
stricted isometry hypothesis.” Let ΦT , T ⊂ {1, . . . , M } be the if a real-valued vector u = [u1 · · · u2M  ]T is S-sparse, then so is the
complex-valued vector uc  [u1 · · · uM  ]T + j[uM  +1 · · · u2M  ]T.
This work was supported by the WWTF projects MOHAWI (MA 44) and
SPORTS (MA 07-004) and by the STREP project MASCOT (IST-026905) 1 “Overwhelming probability” means that the probability of (3) not being
within the Sixth Framework Programme of the European Commission. true decreases exponentially with an increasing number of selected rows, N .

1-4244-1484-9/08/$25.00 ©2008 IEEE 2885 ICASSP 2008

 3. MULTICARRIER SYSTEM MODEL for l = 0, . . . , L−1 and k = 0, . . . , K −1. Here, zl,k = z, γl,k ,
and the system channel coefficients Hl,k can easily be expressed in
3.1. Modulator, Channel, Demodulator terms of g[n], h[n, m], and γ[n] [3].
We will need a “delay-Doppler-domain expression” of the channel
We consider a pulse-shaping MC system for the sake of generality coefficients Hl,k . Let us assume that the receive pulse γ[n] is zero
and because of its advantages over conventional cyclic-prefix (CP) outside [0, Lγ ]. To compute xl,k in (8) for l = 0, . . . , L − 1, r[n]
OFDM [3, 8]; however, CP-OFDM is included as a special case. The
complex baseband domain is considered throughout. Let K, N ≥ K, must then be known for n = 0, . . . , Nr−1, where Nr  (L−1)N +
and L denote the number of subcarriers, the symbol duration, and Lγ + 1. In this interval, we can express r[n] as
the number of transmitted symbol periods, respectively. The MC ∞ NX
X r −1
in
j2π N
modulator generates the discrete-time transmit signal r[n] = Sh [m, i]s[n−m]e r + z[n] , (12)
m=−∞ i=0
X K−1
L−1 X
s[n] = al,k gl,k [n] , (7) with the discrete-delay-Doppler spreading function [11]
l=0 k=0
Nr −1
where al,k with l = 0, . . . , L − 1 and k = 0, . . . , K − 1 denotes 1 X in
−j2π N
k
Sh [m, i]  h[n, m]e r . (13)
the data symbols and gl,k [n]  g[n − lN ]ej2π K (n−lN) is a time- Nr n=0
frequency shifted version of a transmit pulse g[n]. Subsequently,
Combining (8), (12), and (7), we reobtain the system channel rela-
s[n] is converted into the continuous-time transmit signal s(t) =
P ∞ tion (11) with Hl,k expressed as
n=−∞ s[n]f (t−nTs ), where f (t) is an interpolation filter and Ts
is the sampling period. For simplicity, we assume an ideal filter, i.e., ∞ NX
X r −1
p km N li
f (t) = 1/Ts sinc(πt/Ts ) with sinc(x)  sinx x . Hl,k = F [m, i] e−j2π( K − Nr ) , (14)
The channel is assumed doubly selective/dispersive with time- m=−∞ i=0
where “
varying impulse response h(t, τ ). The channel output is i ”
Z ∞ F [m, i]  Sh [m, i] A∗γ,g m, (15)
Nr
r(t) = h(t, τ )s(t−τ )dτ + z(t) , P
−∞ with the cross-ambiguity function [12] Aγ,g (m, ξ)  ∞
n=−∞ γ[n]
where z(t) is zero-mean, stationary, white, rotationally invariant, g ∗ [n − m]e−j2πξn . Using the approximation Nr ≈ LN (which is
complex Gaussian noise with power spectral density N0 . exact for CP-OFDM), we can write (14) as the two-dimensional dis-
R ∞ output r(t) is converted into the dis-
At the receiver, the channel crete Fourier transform
crete-time signal r[n] = −∞ r(t)f (t − nTs )dt, where f (t) now X L−1
K−1 X km − li )
serves as an anti-aliasing filter. Subsequently, the MC demodulator Hl,k = F̃ [m, i] e−j2π( K L , (16)
computes the inner products of r[n] with time-frequency shifted ver- m=0 i=0
k
sions γl,k [n]  γ[n − lN ]ej2π K (n−lN) of a receive pulse γ[n], i.e., with the “pre-aliased” version of F [m, i]
X∞
xl,k = r, γl,k  = ∗
r[n]γl,k [n] , (8) X
N−1

n=−∞
F̃ [m, i]  F [m, i + qL] , i = 0, . . . , L−1 . (17)
q=0
for l = 0, . . . , L − 1 and k = 0, . . . , K − 1. The xl,k are finally
equalized and quantized according to the data symbol alphabet. 4. DELAY-DOPPLER SPARSITY
In practice, CP-OFDM is typically used [9, 10]. This is a special
case of our pulse-shaping MC setting; it is obtained for a rectangular We assume that the channel comprises P propagation paths corre-
transmit pulse g[n] that is 1 on [0, N − 1] and 0 otherwise, and a sponding to P specular (point) scatterers with fixed delays τp and
rectangular receive pulse γ[n] that is 1 on [N − K, N − 1] and 0 Doppler frequency shifts νp for p = 1, . . . , P . This simple structure
otherwise (N −K ≥ 0 is the CP length). is often a good approximation to real mobile radio channels. The
By combining some of the equations presented earlier, a relation channel impulse response then has the form
between the discrete-time signals s[n] and r[n] is obtained as
X
P
X
∞
h(t, τ ) = ηp δ(τ −τp ) ej2πνp t ,
r[n] = h[n, m]s[n−m] + z[n] , (9) p=1
m=−∞
where ηp characterizes the attenuation and initial phase of path p.
with the discrete-time time-varying impulse response The discrete-time impulse response (10) becomes
Z ∞ “ “ τ ””
h[n, m] = h(nTs , τ ) sinc π m − dτ . (10) X
P “ “ τp ””
−∞ Ts h[n, m] = ηp ej2πνp nTs sinc π m − . (18)
p=1
Ts
The discrete-time noise z[n] is zero-mean, stationary, white, rota-
tionally invariant, complex Gaussian with variance σz2 = N0 . Furthermore, inserting (18) into (13) and applying the geometric
sum formula, the delay-Doppler spreading function is obtained as
3.2. System Channel
X
P
jπ(νp Ts − Ni )(Nr −1)
Next, we consider the system channel subsuming the MC modula- Sh [m, i] = ηp e r Λp [m, i] , (19)
tor, the physical channel, and the MC demodulator. Combining (8), p=1
(9), and (7) and neglecting intersymbol and intercarrier interference with
“ “ τp ”” ` ´
(which is justified even in highly mobile environments if g[n] and Λp [m, i]  sinc π m − dirNr π(i − νp Ts Nr ) , (20)
γ[n] are properly designed), we obtain Ts
x P x n
xl,k = Hl,k al,k + zl,k , (11) where dirN (x)  N1 ej N (N−1) N−1
n=0 e
−j2 N sin(x)
= N sin(x/N) .

2886
 In the following, we investigate the sparsity of Sh [m, i]. In view k = 0, . . . , D−1. Furthermore, (16) entails the relation
of (19), we first consider the sparsity of Λp [m, i]. Using the appro-
priate versions of Parseval’s theorem, it can be readily shown that  XX
D−1 I−1
km − li )
P∞ PNr −1 2 Hl ΔL,k ΔK = (−1)l F  [m, i−I/2] e−j2π( D I ,
m=−∞ i=0 Λp [m, i] = 1, i.e., the total energy of Λp [m, i] is
m=0 i=0
1. Next, we calculate an upper bound on the percentage of the energy (24)
of Λp [m, i] that is located outside a rectangular neighborhood of the 
where F [m, i] is a cyclically extended version of F̃ [m, i] that equals
delay-Doppler point of the pth scatterer, (τp /Ts , νp Ts Nr ). To this F̃ [m, i] for i = 0, . . . , L−1 and F̃ [m, i + L] for i = −L, . . . , −1.
end, we first consider the energy of those samples of sinc(π(m − Suppose that pilot symbols al,k = pl,k are transmitted at time-
τp /Ts )) whose distance from τp /Ts is greater than Δm ∈ {2, 3, . . . }. frequency positions (l, k) ∈ P, where P is a subset of the subsam-
Let M denote the set of all integers m ∈ Z except those with pled grid (l ΔL, k ΔK), l = 0, . . . , I −1, k = 0, . . . , D−1. The
|m − τp /Ts | ≤ Δm. We have the bound |P| pilots pl,k and their time-frequency positions are known to the
X ˛˛ “ “ Z ∞
τp ””˛˛2 X receiver. From (11), xl,k = Hl,k pl,k + zl,k for (l, k) ∈ P. The
∞
1 2 dx
˛ sinc π m − ˛ ≤ 2 2
≤ 2 receiver calculates channel coefficient estimates Ĥl,k at the pilot po-
m∈M
Ts
i=Δm
(πm) π Δm−1 x2 sitions (l, k) ∈ P according to
2
= , (21) xl,k zl,k
π 2 (Δm −1) Ĥl,k  = Hl,k + , (l, k) ∈ P . (25)
pl,k pl,k
2
where sin x ≤ 1 and some monotonicity arguments have been used. Thus, these Hl,k are known up to the additive noise terms zl,k /pl,k .
In a similar
` manner, ´we consider the energy of those samples of Next, we use the fact that all Hl ΔL,k ΔK can be expressed via
dirNr π(i − νp Ts Nr ) whose distance (up to the modulo-Nr oper- the Fourier transform relation (24). In vector notation, (24) reads
ation, see below) from νp Ts Nr is greater than Δi ∈ {2, . . . , Nr /2} √
(Nr is assumed even). Let I denote the set {0, . . . , Nr − 1} with h = ID Uuc , (26)
the exception of all i = iZ mod Nr , where iZ is any integer with
|iZ − νp Ts Nr | ≤ Δi. We have the bound where we defined (i) the ID-dimensional complex “parameter” vec-
ˆ ˜T ˆ
X˛ Nr /2 tor uc  uTc (0) · · · uTc (D − 1) with uc (m)  F  [m, −I/2]
` ´˛
˛ dirN π(i − νp Ts Nr ) ˛2 ≤ 2 X 1 ˜ T
` ´ · · · F  [m, I/2−1] ; (ii) the ID-dimensional complex channel vec-
r
Nr i=Δi sin2 Nπ i
2 ˆ ˜T ˆ
i∈I r tor h  h(0) · · · h(D−1)
˜ with h(k )  H0,k ΔK HΔL,k ΔK
Z Nr /2 “ π ”
2 dx
` ´
2 · · · H(I−1)ΔL,k ΔK ; and (iii) the ID × ID block matrix U with
≤ 2 π
= cot (Δi −1) (k−1)(m−1)
Nr2 Δi−1 sin N x Nr π Nr I × I blocks Uk,m  √1D e−j2π D SF for k, m = 1,
r
1 . . . , D, where F is the I × I IDFT matrix with entries (F)l,i =
≤ , (22) (l−1)(i−1)
π(Δi −1) √1 ej2π I for l, i = 1, . . . , I and S is the diagonal I × I
I
where we have used sin2 x ≤ 1, some monotonicity arguments, and matrix with diagonal entries 1, −1, 1, −1, . . .. We note that U is
π
cot x ≤ 2x within [0, π2 ]. Combining (20)–(22), we see that unitary with coherence μ = 1. Furthermore, according to the previ-
X ous section, uc is modeled as P NΛ -sparse.
2 |P| specific entries of the channel vector h are given by the chan-
Λ2p [m, i] ≤ 3 . (23)
π (Δm −1)(Δi −1) nel coefficients Hl,k at the |P| pilot positions (l, k) ∈ P. Let h(p)
(m,i)∈M×I
2 denote the corresponding |P|-dimensional subvector of h. Further-
Thus, at most a percentage of π3 (Δm−1)(Δi−1) · 100 % of the energy more, let Φc denote the |P|×ID submatrix of U obtained by select-
of Λp [m, i] is located outside a rectangular neighborhood of (τp /Ts , ing the corresponding
p |P| rows of U and multiplying the resulting
νp Ts Nr ) comprising (2Δm + 1)(2Δi + 1) samples. Since the do-
main of Λp [m, i] is Z × {0, . . . , Nr − 1}, with Nr typically large,
matrix by ID/|P| (this implies pthat the columns of Φc have unit
Euclidean norm). We then have 1/|P| h(p) = Φc uc , which, up to
we conclude that Λp [m, i] is an approximately sparse function.
a constant factor, is (26) reduced top the pilot positions.
Hereafter, therefore, we will consider Λp [m, i] as NΛ -sparse, with
an appropriately chosen number NΛ of nonzero samples. The energy Next, we define vc as the vector 1/|P| h(p) but with the entries
(p)
bound (23) allows us to choose NΛ such that a prescribed approx- ph replaced
Hl,k of
(p)
by their estimates Ĥl,k . Then, because of (25),
imation quality can be guaranteed. From (19), it then follows that v
p c = 1/|P| h +w c , where wc comprises the scaled noise terms
p
Sh [m, i] is P NΛ -sparse, and the same is true for F [m, i] in (15) 1/|P| zl,k /pl,k , (l, k) ∈ P. Inserting 1/|P| h(p) = Φc uc , we
and, in turn, for F̃ [m, i] in (17). obtain vc = Φc uc + wc . This is equivalently stated as

5. CS-BASED CHANNEL ESTIMATION v = Φu + w , (27)

ˆ ˜ T ˆ ˜ T
To develop the new channel estimation method, we first show that where u  {uTc }
{uTc } , v  {vcT }
{vcT } , w 
ˆ ˜T
pilot-based channel estimation can be formulated as a sparse recon- {wcT }
{wcT } , and Φ (defined in (6)) are the real versions of
struction problem of the type discussed in Section 2. uc etc. According to our construction, u is 2P NΛ -sparse. Thus,
For practical channels as well as transmit and receive pulses, the (27) is seen to be a sparse reconstruction problem of the form (1),
function F [m, i] in (15) is effectively supported in a subdomain of with dimensions M = 2ID and N = 2|P| and sparsity S = 2P NΛ .
the delay-Doppler plane. This allows us to perform a subsampling We can hence use the extended basis pursuit (2) to obtain an esti-
in the time-frequency domain. Thus,` we assume that the support of ´ mate of u or, equivalently, of F  [m, i]. From this estimate, estimates
F̃ [m, i] is contained in [0, D−1] × [0, I/2−1] ∪ [L−I/2, L−1] , of all channel coefficients Hl,k are finally obtained via (16). The
where D ≤ K, I is even, and D and I are chosen such that ΔK  convex program (2) is a simple instance of a second-order cone pro-
K/D and ΔL  L/I are integers. Because of (16), the channel gram, which can be solved efficiently by interior-point methods [13].
coefficient function Hl,k is then uniquely specified by its values on Thus, CS-based channel estimation is computationally feasible, al-
the subsampled grid (l, k) = (l ΔL, k ΔK) with l = 0, . . . , I −1, beit more complex than classical least-squares channel estimation.

2887
 Regarding the choice of the pilot positions (l, k) ∈ P, we recall 10
2
10
0

that these positions correspond to |P| indices within the index range Least−squares (12.5% pilots)
{1, . . . , ID} of the channel vector h. To be consistent with the CS 10
1 CS−based (12.48% pilots)
Least−squares (25% pilots) 10
−1

framework of Section 2, we select these positions uniformly at ran- 0

dom [1]. For good approximation quality in the sense of (4), the 10

MSE

SER
−2
number of pilots should satisfy condition (5), which becomes 10
−1
10
` ´4
|P| ≥ C3 ln(ID) · 2P NΛ −2 10
−3
Least−squares (12.5% pilots)
10 CS−based (12.48% pilots)
Least−squares (25% pilots)
(recall that μ = 1). This bound is not useful for actually determining −3 −4 Known channel
|P| because of the constant C3 . However, the bound suggests that 10
10 15 20 25 30
10
10 15 20 25 30
SNR [dB] SNR [dB]
the required number of pilots scales only linearly with the number
P of channel paths (scatterers) and the sparsity parameter NΛ , and Fig. 1. Performance of CS-based and conventional least-squares
poly-logarithmically with the system design parameters I and D. In channel estimation. Left: MSE versus SNR, right: SER versus SNR.
practice, the pilot positions will be randomly chosen (and communi-
cated to the receiver) only once before the beginning of data trans-
mission. With high probability, they will lead to good performance 7. CONCLUSION
for arbitrary channels with at most P paths.
We have proposed a channel estimation technique based on the re-
cently introduced principle of compressed sensing (CS). Our results
6. SIMULATION RESULTS demonstrate that CS makes it possible to exploit the “delay-Doppler
We next present numerical results to compare the performance of the sparsity” of wireless channels for a reduction of the number of pi-
proposed CS-based channel estimation method with that of classical lots required for channel estimation within multicarrier systems. We
least-squares channel estimation. In accordance with the DVB-T conjecture that the performance of the proposed technique can be
standard [10], we simulated a CP-OFDM system with K = 2048 further improved by using more sophisticated CS methods.
subcarriers and CP length N −K = 512, whence N = 2560. The
system employed a 4-QAM symbol alphabet. 8. ACKNOWLEDGMENT
During blocks of L = 2 transmitted OFDM symbols, we sim-
ulated a noisy doubly selective/dispersive channel whose discrete- The authors would like to thank G. Matz, H. Rauhut, and P. Fertl for
delay-Doppler spreading function Sh [m, i] was computed from (19) fruitful discussions.
and (20). We assumed P = 20 propagation paths whose (continuous- 9. REFERENCES
valued) delays τp /Ts and Doppler frequencies νp Ts were randomly
chosen within [0, 511] × [−0.02/K, 0.02/K] for each block of 2 [1] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles:
exact signal reconstruction from highly incomplete frequency informa-
OFDM symbols (hence, the maximum allowed normalized Doppler tion,” IEEE Trans. Inf. Theory, vol. 52, pp. 489–509, Feb. 2006.
frequency—maximum Doppler divided by the subcarrier spacing— [2] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52,
is ±2%). The complex scatterer amplitudes ηp were randomly cho- pp. 1289–1306, April 2006.
sen from zero-mean, rotationally invariant, complex Gaussian distri- [3] W. Kozek and A. F. Molisch, “Nonorthogonal pulseshapes for multicar-
butions with three different variances. More precisely, we assumed rier communications in doubly dispersive channels,” IEEE J. Sel. Areas
3 strong scatterers of equal power, 7 medium scatterers with power Comm., vol. 16, pp. 1579–1589, Oct. 1998.
10 dB below that of the strong scatterers, and 10 weak scatterers with [4] E. G. Larsson, G. Liu, J. Li, and G. B. Giannakis, “Joint symbol tim-
power 20 dB below that of the strong scatterers. ing and channel estimation for OFDM based WLANs,” IEEE Comm.
Letters, vol. 5, pp. 325–327, Aug. 2001.
For least-squares channel estimation, we used two different rect-
angular pilot constellations, namely, (i) with spacings ΔL = 1 and [5] E. J. Candès, J. Romberg, and T. Tao, “Stable signal recovery from
incomplete and inaccurate measurements,” Comm. Pure Appl. Math.,
ΔK = 4, corresponding to 1024 pilots or 25% of all available trans- vol. 59, pp. 1207–1223, March 2006.
mit symbols, and (ii) with spacings ΔL = 1 and ΔK = 8, corre- [6] M. Rudelson and R. Vershynin, “Sparse reconstruction by convex re-
sponding to only 512 pilots or 12.5% of all symbols. For CS-based laxation: Fourier and gaussian measurements,” in Proc. 40th Annual
channel estimation, we selected uniformly at random 511 pilots from Conf. Information Sciences and Systems, pp. 207–212, March 2006.
the rectangular pilot constellation (i), corresponding to 12.48% of all [7] E. Telatar, “Capacity of multi-antenna Gaussian channels,” European
symbols—i.e., about the number of pilots used in constellation (ii) Trans. Telecomm., vol. 10, pp. 585–596, Nov. 1999.
and about half that used in constellation (i). These three pilot con- [8] G. Matz, D. Schafhuber, K. Gröchenig, M. Hartmann, and
stellations were held fixed during all simulation runs. The CS-based F. Hlawatsch, “Analysis, optimization, and implementation of low-
method used the classical basis pursuit, i.e., (2) with  = 0; this al- interference wireless multicarrier systems,” IEEE Trans. Wireless
Comm., vol. 6, pp. 1921–1931, May 2007.
lows a faster implementation than extended basis pursuit with  > 0.
[9] IEEE P802 LAN/MAN Committee, “The working
The MATLAB function l1eq pd() from the toolbox 1 -MAGIC group for wireless local area networks (WLANs).”
[14] was employed. http://grouper.ieee.org/groups/802/11/index.html.
Fig. 1 depicts the mean square error (MSE) and the symbol er- [10] ETSI, “Digital video broadcasting (DVB); framing structure, channel
ror rate (SER) versus the signal-to-noise ratio (SNR). It is seen that coding and modulation for digital terrestrial television.” EN 300 744,
the CS-based method (with 511 pilots) significantly outperforms the V1.4.1, 2001 (http://www.etsi.org).
least-squares method with 512 pilots. The extremely poor perfor- [11] P. A. Bello, “Characterization of randomly time-variant linear chan-
mance of the least-squares method is due to the fact that the Shannon nels,” IEEE Trans. Comm. Syst., vol. 11, pp. 360–393, 1963.
sampling theorem is violated by the pilot grid. In contrast, the CS- [12] P. Flandrin, Time-Frequency/Time-Scale Analysis. San Diego (CA):
Academic Press, 1999.
based method is able to produce reliable channel estimates even far
below the Shannon sampling rate. Compared with the least-squares [13] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge (UK):
Cambridge Univ. Press, Dec. 2004.
method with 1024 pilots, we observe only a relatively small perfor-
[14] E. J. Candès and J. Romberg. Toolbox 1 -MAGIC, California Inst. of
mance degradation of the CS-based method with 511 pilots, espe- Technol., Pasadena, CA (http://www.acm.caltech.edu/l1magic/).
cially in the low-to-medium SNR regime (up to 20 dB).

2888