Professional Documents
Culture Documents
Abstract—Delay matters in future wireless communication. An failure of the provided service? An answer to this question
appropriate limit for rates achievable under delay constraints provides not only an appropriate performance limit for delay
is the delay limited capacity (DLC). In this work, the DLC of sensitive services such as e.g. streaming services in LTE
OFDM systems is investigated. Despite its complicated correla-
tion structure the OFDM DLC is fully characterized for low systems (Long Term Evolution of 3GPP UMTS system). It
and high SNR. It is shown that (under weak assumptions) the also gives structural insights into the general system behavior
OFDM DLC is almost independent of the fading distribution in yielding guidelines for engineering wireless communication
the low SNR region but strongly depends on the delay spread systems.
thereby achieving a capacity gain over AWGN capacity. In the
high SNR region the roles are exchanged. Here, the impact It is known that in general multiple degrees of freedom in
of delay spread is negligible while the impact of the fading fading channels allow reliable communication in each fading
distribution becomes dominant. The relevant quantities and their state under a long term power constraint. This is due to
asymptotic behaviour are derived without employing simplifying the possibility of recovering the information from several
assumptions on the OFDM correlation structure. Using a general
independently faded copies of the transmitted signal. The rate
convergence framework the analysis further shows that if the
delay spread becomes large even the predominant impact of the achievable in each fading state is called zero outage capacity
fading distribution vanishes and DLC capacity loss compared or alternatively delay limited capacity (DLC) [1]. Not only
to AWGN capacity approaches 0.58[nats/s/Hz]. The convergence multiple input multiple output (MIMO) channels but also
speed, the loss due to non-uniform power delay profile, and the frequency selective multi-path channels offer multiple degrees
relation to ergodic capacity is also analyzed and underlined with
of freedom. This is in contrast to single antenna Rayleigh flat
simulations and application examples. The main conclusion here
is that OFDM fully takes advantage of the degrees of freedom of the fading channels, where a DLC does not exist.
underlying fading channel in terms of delay spread and, regardless This work investigates the DLC of frequency selective
of the fading distribution, delay sensitive capacity measures such as multi-path channels using orthogonal frequency division
the DLC converge to the ergodic capacity. Finally, since universal
bounds are obtained which apply to any fading distribution the multiplexing (OFDM) to mitigate inter-symbol interference.
results can also be used for other classes of parallel channels OFDM can be considered as a special case of parallel fading
extending the range of applicability. channels with correlated fading process. Pioneering work on
Index Terms—Delay limited capacity, orthogonal frequency this topic was carried out in [2][3][4][5][6][7][8]. Unfortu-
division multiplexing (OFDM), power control, rate allocation, nately, these results do not carry over to the OFDM case: since
parallel Gaussian channels. the subcarriers are highly correlated due to oversampling of
the channel in the frequency domain the fading distribution
is commonly degenerated which significantly complicates the
I. I NTRODUCTION analysis. This particularly affects the critical impact of the
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
3748 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009
p1 (h) h̃1 n1
is the optimal resource allocation strategy. Similar analysis
x1 y1
is carried out in the high SNR regime where it is shown
that simple channel inversion achieves close-to-optimal per- p2 (h) h̃2 n2
formance also in the (degenerated) OFDM case where this x2 y2
time the DLC depends on the fading distribution; the analysis
culminates in a general convergence theorem again in terms of
(large) delay spread for OFDM where even the impact of the
fading distribution vanishes showing a universal capacity loss pK (h) h̃K nK
of 0.58[nats/s/Hz] compared to AWGN capacity. This shows xK yK
that OFDM fully takes advantage of the degrees of freedom
of the underlying fading channel in terms of delay spread
and, regardless of the fading distribution, short term capacity Fig. 1. General system model: data of K streams xk is sent over parallel
fading channels with arbitrary fading distribution hk generated by eqn. (1)
measures such as the DLC converge to the ergodic capacity. and received under AWGN with iid nk ∼ CN (0, 1).
The remainder of this paper is organized as follows: Section
II presents the OFDM system model. In Section III the OFDM
DLC is introduced and suboptimal power allocation strategies some vector quantizer is covered in our analysis provided that
are discussed. In Section IV-A the behavior at low SNR is the rates defined below are achievable. The general model is
studied while Section IV-B focuses on the high SNR regime. summarized in Fig.1.
We conclude with some final remarks in Section VII. Given the channel gains h = [h1 , ..., hK ]T the rate achiev-
able over all K parallel Gaussian channels with a certain
A. Notations power allocation p = [p1 , ..., pK ]T reads as
All terms will be arranged in boldface vectors. Common 1
K
1
K
vector norms (such as ·1 for the l1 -norm) will be employed. R(h, p) = rk (hk , pk ) = log (1 + pk hk ) , (2)
K K
The expression z ∼ CN (0, 1) means that the complex-valued k=1 k=1
random variable z = x + jy is circular symmetric Gaussian where rk (hk , pk ) denotes the rate achievable on subcarrier
distributed, i.e. the real and imaginary parts are indepen- k. We further introduced the factor 1/K so that all rates
dently Gaussian distributed with zero mean and variance 1/2: are normalized to spectral efficiency and given in [nats/s/Hz].
x, y ∼ N (0, 1/2). A sequence of random variables is called The small impact of the OFDM guard interval on the spectral
iid if 1.) any subset is an independent set and 2.) all random efficiency shall be neglected here.
variables are circular symmetric. The expectation operator Now, assume that the system is subjected to a long term
(e.g. with respect to the fading process) will be denoted as power constraint, i.e.
E (respectively Eh or Eh̃ ). Pr(A) denotes the probability of K
an event A. All logarithms are to the base e unless explicitly
Eh pk (h) ≤ P ∗ , (3)
defined in a different manner.
k=1
II. OFDM C OMMUNICATION M ODEL where we use pk (h) to denote that the power allocation may
depend on the current fading realization. This means that while
Assuming familiarity with the general model consider a there is no peak power constraint per fading state, in average
standard OFDM communication system where a single user the power constraint P ∗ has to be met. Please note that due to
uses K subcarriers for information transmission. The complex non-linear components in the transmitter path such ideal power
channel gain on subcarrier k is by means of Fast Fourier control scheme is difficult to implement in practice. Therefore,
Transform (FFT) given by the results should be seen as a limit for any practical power
L
2πj(l−1) · (k−1)
control scheme.
h̃k = c̃l e− K , k = 1, ..., K, (1)
l=1 III. A P ERFORMANCE M EASURE FOR D ELAY L IMITED
where L ≤ K is the delay spread, and c̃l are the complex T RANSMISSION
path gains that are modeled as independent, zero mean random A. Optimal rate allocation
variables with variance σl > 0 for all l. The vector of variances
σ = [σ1 , ..., σL ]T is called the power delay profile (PDP) We introduce the DLC Cd (P ∗ ) for an OFDM system,
and the channel energy is normalized, i.e. ||σ||1 = 1. We which is a special case of parallel fading channels.
say that the channel has a uniform PDP if σ1 = . . . = σL Definition 1: The delay limited capacity Cd (P ∗ ) of an
and a non-uniform PDP otherwise. Note that in practice the OFDM system under a long term power constraint P ∗ is given
PDP is typically non-uniform. The channel (path) gains are by
defined as hk := |h̃k |2 (respectively cl := |c̃l |2 ) and the Cd (P ∗ ) = sup inf R (h, pk (h)) (4)
p∈P ∗ h∈H
distribution of the channel gains is called the (joint) fading
distribution. It is worth pointing out that we do not make where H ⊆ RK + is the set of possible channel gains and
any assumptions on the fading distribution. Even the case of P ∗ comprises all power allocation policies advising a power
point masses (i.e. discrete fading distributions) induced e.g. by allocation pk (h) ∀k to every h ∈ H such that (3) holds.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
WUNDER et al.: DELAY-LIMITED TRANSMISSION IN OFDM SYSTEMS: PERFORMANCE BOUNDS AND IMPACT OF SYSTEM PARAMETERS 3749
In other words, Cd (P ∗ ) is the maximum rate which can be gains. In order to avoid this complexity we introduce the
achieved for all possible channel gains without violating the notion of rate waterfilling for expected ordered channel gains,
average power constraint P ∗ . or so-called statistical rate waterfilling (SRW) as follows:
Definition 1 implies that in order to achieve Cd (P ∗ ) we for a given vector h of real elements let us introduce the
need to find the power allocation pk (h) ∀k that supports a total ordering hk[K] ≥ hk[K−1] ≥ . . . ≥ hk[1] , i.e. hk[1] is
given rate Cd with minimum power. For any h ∈ H this the minimum value and hk[K] is the maximum value; the
optimization problem is equivalent to: distribution of hk[p] is known to be the p-th order statistics of
a sample h. Based on the order information we can deduce a
K
fixed rate allocation on the subcarriers, avoiding optimal RW.
min pk
p∈RK The key idea is to allocate a fixed rate budget to the p-th
k=1 (5) ordered subcarrier. Defining the terms
1
K
subj. to rk (hk , pk ) ≥ Cd
+∞
K 1
k=1
ζp := dFhk[p] (h)
Using the relation between power and rate on subcarrier k in h
0
eqn. (2) the problem can be easily solved and the resulting
optimal rate allocation is given by where Fhk[p] is the marginal distribution of the p-th ordered
rk − ⎫ channel gain and using these factors in the optimization
e ⎪ problem (5) the SRW rate allocation is given by:
− λ = 0, k = 1, ..., K ⎪ ⎪
⎪
hk ⎪
⎬ rk[p] − ⎫
K e
− = 0, = 1, ⎪
⎪
1 (Rate Waterfilling) λ k ..., K ⎪
⎪
rk = Cd ⎪
⎪
⎪ ζp−1 ⎪
⎬
K ⎪
⎪ K
k=1 ⎭ 1
λ>0 rk[p] = Cd ⎪ ⎪
K p=1 ⎪
⎪
− ⎪
⎭
where [·] := min {·, 0} and λ ∈ R is a Lagrange multiplier. λ>0
The rate allocation is called rate waterfilling (RW) because (Statistical Rate Waterfilling)
substituting λ = log(λ̃) and hk = log(1/h̃k ) yields the The performance of SRW is illustrated in Fig.2 and it can be
classical waterfilling rule. Solving for λ and after some algebra observed that it does particularly well in the low SNR region.
we obtain the single user OFDM delay limited capacity Cd This will be exploited in the low SNR analysis where it is
with power constraint P ∗ (corresponds to Theorem 3.2 in [7]) shown that it becomes optimal as SNR goes to zero.
⎛
⎞ There is an interesting second rate allocation termed chan-
|D (Cd , h)| exp |D(CCd K
,h)| nel inversion (CI) introduced in [3] where the powers asserted
P ∗ =Eh ⎝ ⎠
d
K k∈D(Cd ,h) hk
1/|D(Cd ,h)| to the subcarriers are all the same. It is easy to see then that
⎛ ⎞ the CI rate allocation according to
1 1
− Eh ⎝ ⎠ (6) e C d hk
K hk rk = log 1 + K 1/K
, k = 1, ..., K,
k∈D(Cd ,h)
k=1 hk
where the random variable D (Cd , h) ⊆ {1, ..., K} denotes (Channel inversion)
the set of active subcarriers and |D (Cd , h)| its cardinality. always leads to a rate higher than the requested rate at
Since the numerator in (6) can be bounded by a constant and the expense of power consumption. Hence, this is also a
by applying arithmetic-geometric mean inequality to the last suboptimal solution. CI is illustrated in Fig.2. In contrast to
term, the delay limited capacity Cd is greater than zero if and SRW it performs well in the high SNR region. This will be
only if exploited in the high SNR analysis where it is shown that it
becomes optimal as SNR goes to infinity.
1 The relevant performance measures for SRW and CI play a
1/|D(Cd ,h)|
dFh (h) < ∞. (7)
k∈D(Cd ,h) hk significant role in the forthcoming analysis. Next, we analyze
K R+
existence of DLC in OFDM systems.
Here, Fh denotes the joint fading distribution function. The
class of fading distributions for which (7) holds is called C. Existence of DLC
regular in [3]. It will become apparent in the following that
the correlation structure of the channel gains in OFDM pro- Denote the guaranteed rates achievable under SRW by
vides the main challenge in proving and analyzing regularity CdSRW (P ∗ ). By the suboptimality of SRW the DLC is clearly
−1
according to (7). non-zero if ζK = Eh (h∞ ) < ∞, i.e.
Let us now introduce two important suboptimal power
1
allocation strategies. dFh (h) < ∞. (8)
h∞
RK
+
B. Suboptimal rate allocation and it is therefore of general interest when eqn. (8) holds. The
It is evident from the expression for the DLC that the following theorem states a strikingly weak sufficient condition
major difficulty is the rate waterfilling operation for all channel on the existence of the DLC.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
3750 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009
15
IV. T HE I MPACT OF S YSTEM PARAMETERS
equal power strategy It is now of great interest to understand the impact of
equal rate strategy
rate water−filling strategy
the ergodic fading process and its parameters. So the delay
OFDM DLC spread L and the power delay profile σ as well as the fading
10 distribution itself obviously affect the OFDM delay limited
capacity. Since the expression in (6) is still very complicated,
Cd [bps/Hz]
we focus on the behavior in the low and the high SNR regime
and carry out a detailed analysis.
5
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
WUNDER et al.: DELAY-LIMITED TRANSMISSION IN OFDM SYSTEMS: PERFORMANCE BOUNDS AND IMPACT OF SYSTEM PARAMETERS 3751
−3 −3
x 10
even for moderate L, K ≥ L when the complex path gains are
x 10
5 5
4.5 4.5
4 4 iid; moreover, the upper bound also holds when the PDP is
non-uniform [10]. We can apply this result to the DLC where
3.5 3.5
C [bps/Hz]
3
C [bps/Hz]
2.5
C
2.5
we have to show that from the convergence in probability
d
d
d
2 2
1.5
1st order
2nd order 1.5
C
d
1st order
given in eqn. (13) it follows convergence of the expected
1
0.5
1
0.5
2nd order
maximum of the channel gains as well. Leaving out technical
0
−3.475 −3.47 −3.465 −3.46 −3.455 −3.45 −3.445 −3.44 −3.435 −3.43
0
−6.5 −6.45 −6.4 −6.35
details we can show the following:
E /N [dB] Eb/N0 [dB]
Theorem 3: Suppose that the complex path gains are in-
b 0
−3
x 10
0.01
4 0.009
C
d
1st order
2nd order
dependent. Moreover, assume that the path gain distribu-
tion allows for some bounded Lipschitz constant in an -
0.008
3.5
C 0.007
3 d
1st order 0.006
neighborhood of zero uniformly. Then the following result
C [bps/Hz]
C d [bps/Hz]
2.5 2nd order
0.005
2
holds:
d
0.004
Cd (P ∗ )
1.5
0.003
1
0.002
lim sup lim ∗
≤1 (14)
L→∞ P →0 log (L) P
0.5 0.001 ∗
0 0
−8.2 −8 −7.8 −7.6 −7.4 −7.2 −10 −9 −8 −7 −6 −5
E /N [dB] Eb/N0 [dB]
b 0
Note that the DLC compares favorably by the factor log (L)
p=1/3 p=1/3 with the capacity of AWGN in the low SNR regime.
1 X X
A B In order to make the theorem useful in practice we can write
Cd (P ∗ ) cLo g (L)
lim ≤1+ =: ψ (L) , any L (large),
P ∗ →0 log (L) P ∗ log L
C p=1/3 (15)
X
1 h1 where g (L) := log [log (L)] and cLo > 0 is a constant
independent of K (≥ L); consequently, the limit in eqn. (14)
Fig. 4. Three fading states with equal probability.
regarding L is independent of how K, L scale. On the other
hand for small but non-zero P ∗ we have by Theorem 2 and
eqn. (12)
tight for practical values of L while the range where the
Cd (P ∗ ) cLo g (L) + Kχ−1 P ∗ log (L)
approximation is useful degrades for very large L. ≥1− (16)
Rather interesting, the multiplicity of the maximum subcar- log (L) P ∗ log (L)
rier gain occur in the expressions; a scenario where this mat- which now indeed depends on K. Hence, for fixed L and
ters is discussed in the example in Fig.IV-A1 for one subcarrier growing K the lower bound (16) becomes arbitrarily small.
with three fading states occurring with equal probability. This This effect can be typically alleviated by the fact that for K
mimicks e.g. a mobile
at the cell border employing handover. L the number of subcarriers having approximately the same
Here, clearly Eh h−1 ∞ = 1 but in fact according to Theorem 2 maximum channel gain h∞ is also increasing. A very good
the DLC growth over energy per bit indicated by the sublinear estimate is hk ≥ hk∗ cos πL ∗
K with hk = h∞ and |k − k | <
∗
P ∗ →0
term will be Δd (P ∗ )/P ∗ → 4/3. K/(2L) [12] so that Kχ −1
= O (L) and the lower bound
2) An explicit formula for OFDM: Appealing to Theorem 2 becomes independent of K. Note that there is also an impact
the forthcoming analysis reduces to the study of the expected of the PDP which is treated in Sec.VI.
maximum of the channel gains. However, the expressions do The unknown small constant cLo > 0 can be found numeri-
not show how the DLC depends on the system parameters cally. The approach is demonstrated in Fig.5 where we depict
which we investigate by means of an asymptotic analysis, i.e. Cd (P ∗ ) over Eb /N0 (in dB scale) and the approximations for
for large L, K. This analysis turns out to be quite accurate different L. It is seen that the minimum energy per bit, at
even for very small L. which reliable transmission is possible, goes to minus infinity
We make use of the following result [10][11, Theorem with order − log[log(L)] as indicated by Theorem 3.
1]: under very mild assumptions on the fading distribution
we have that h∞ equals approximately log (L) with large
probability (recall that we set ||σ||1 = 1), i.e. B. The high SNR regime
Pr (log (L) − 4 log [log (L)] ≤ h∞ After the treatment of the low SNR regime we turn towards
≤ log (L) + 4 log [log (L)]) high SNR. In contrast, here not only the delay spread but also
the fading distribution is important. Furthermore CI instead of
= 1 − O log−4 (L) (13) SRW becomes the asymptotically optimal rate allocation.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
3752 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009
−3
x 10
10 14
9
12
8
7
10 log(P*)−0.58
[nats]
6
Cd [bps/Hz]
C [bps/Hz]
5 8
L=2,...,1024
d
2
−12 −10 −8 −6 −4 −2 0 2 10 15 20 25 30 35 40 45
E /N [dB] SNR [dB]
b 0
Fig. 5. The solid curves depict the DLC over Eb /N0 in an OFDM channel Fig. 6. Scaling in the high SNR region: The black line indicates the scaling
with L = 2...1024 taps for Rayleigh fading and uniform PDP. The upper at high SNR given by log (P ∗ ) − 0.58 (for Rayleigh fading). The dashed
bound on the minimum energy per bit marked by the crosses is given by lines give the OFDM DLC for L = K = {2, 4 (non-uniform), 4 (uniform),
eqn. (15) with cLo ≈ 1.2; the constant is found by simple linefitting for the 32}.
smaller L’s while by virtue of Theorem 3 the formula will hold also for the
larger L’s. The gain over of the AWGN channel (−1.59[dB]) is marked by
the dashed curve.
matters for small L the impact quickly vanishes in the limit
for large L. The more independent channel gains that can
1) High SNR rate control: Defining the quantity h := be obtained, the faster will be the convergence, as stated in
K −1/K
k=1 hk we have by using the suboptimal CI rate
control the following theorem. Here, avoiding technicalities in the
law Cd (P ∗ ) ≥ log P ∗ /Eh h provided that Eh h < ∞, proofs (such as L, K being prime numbers) L, K are generally
i.e. for regular fading distributions [3]. We can extend this assumed to be dyadic numbers (or just divisible); this is not too
result to an upper bound without using any simplifying as- restrictive since K is dyadic in practice. Please note that the
sumptions on the fading distribution; it is also remarkable following convergence holds for any performance measure that
that it is tight for large K and large P ∗ regardless whether fulfills the conditions stated in the proof, i.e. monotonicity and
the distribution
is continuous or not. Further, note that the uniform integrability (such as peak-to-average power ratio).
quantity Eh h is not at all always meaningful; a simple The proof technique improves on the approach taken by [13]
counterexample is given in the already discussed three fading where weak convergence of the joint distribution of any finite
state example
in Sec. IV-A where the DLC is non-zero but subset of subcarriers to a joint Gaussian distribution is
shown,
Eh h = ∞. which excludes performance measures such as Eh h defined
Theorem
4 (High SNR optimality of CI): Suppose that on all subcarriers.
Eh h < ∞. Theorem 5: Suppose that complex path gains c̃1 , ..., c̃L are
i.) The following upper limit holds: iid; further suppose that their marginals are circular symmetric
1 and have uniformly bounded and sufficiently smooth densities
lim sup [Cd (P ∗ ) − log (P ∗ )] ≤ log Eh h + . with exponential tails for all L. Then, the following upper
∗
P →∞ K
(17) limit holds:
ii.) The following lower limit holds: ∞
∗ ∗
lim inf [C d (P ∗
) − log (P ∗
)] ≥ log Eh h . (18) lim sup [Cd (P ) − log (P )] ≤ log (h) exp (h) dh (19)
∗ P →∞ L,P ∗ →∞
0
Both bounds coincide for large K (or the distribution is
:=HF ≈−0.58
continuous [3]) and, hence, CI rate control is optimal in
the high SNR regime then. Equality in (19) holds for uniform PDP.
Proof: The proof is given in Appendix
VIII-B. The upper bound becomes tight in the presence of Ko -th
−1/3
Theorem 4 states that as long as Eh h < ∞ the DLC order diversity with convergence rate Ko (for constants see
lies in some target corridor determined by E h h . Hence, it eqn. (34)) where Ko is the number of independent subcarriers.
suffices to evaluate the term Eh h in (17) in the high SNR Proof: see Appendix VIII-C.
regime which characterizes the fixed capacity gap compared Interestingly, there is always a loss in capacity compared to
to the log(P )-scaling of AWGN depending on the fading AWGN under the assumptions of the theorem (e.g. Rayleigh-
distribution. A simple explicit asympotic expression for this , Nakagami-fading etc.). The capacity loss equals that of
gap is now provided. AWGN capacity to ergodic capacity as we will show in the
2) An explicit formula for OFDM: In order to get some next subsection. An illustration is shown in Fig.6 where also
insight let us carry out again an asymptotic analysis. The the non-uniform case is exemplarily incorporated showing
following theorem shows that while the fading distribution almost no impact of the PDP.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
WUNDER et al.: DELAY-LIMITED TRANSMISSION IN OFDM SYSTEMS: PERFORMANCE BOUNDS AND IMPACT OF SYSTEM PARAMETERS 3753
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
3754 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009
TABLE I
M AXIMUM NUMBER OF SUPPORTABLE USERS AT THE CELL BORDER FOR 3GPP P EDESTRIAN A/B CHANNEL , 3 KM / H . T HE VALUES IN ROW 3,4 IS
OBTAINED BY THE SIMULATED DLC AND IN R OW 5,6 IS THE OBTAINED BY UPPER BOUND IN (22).
Real-time Streaming Services Conversational voice High quality streaming audio Videophone
Rate Requirement 13 kb/s 128 kb/s 384 kb/s
Nr. of users (Ped. A, DLC simulated) ≤ 34 ≤3 ≤1
Nr. of users (Ped. B, DLC simulated) ≤ 67 ≤6 ≤2
Nr. of users (Ped. A, bound) ≤ 34 ≤3 ≤1
Nr. of users (Ped. B, bound) ≤ 77 ≤7 ≤2
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
WUNDER et al.: DELAY-LIMITED TRANSMISSION IN OFDM SYSTEMS: PERFORMANCE BOUNDS AND IMPACT OF SYSTEM PARAMETERS 3755
generating a subfamily of complex-valued random variables and by applying central limit theorem for triangular arrays
with cardinality |Em
n
|. Furthermore, we superimpose a natural to the random variable ξ (αi , βi ) (which is the sum of inde-
order on the elements of Em n
such that k1 < k2 < ... < k2m pendent random variables by the iid assumption) proves (30)
for ki ∈ Em . In order to establish convergence we need to
n for any set ki ∈ Emn
, i = 1, ..., M . Note that by the assumed
prove the following steps. exponential decay of the marginals, Lindeberg’s condition for
i.) (Monotonicity): We can frequently use the following triangular arrays will be clearly satisfied and convergence of
n D
well-known inequality [17] for some a, b > 0 dividing K: h̃Em → h̃m G follows (see also [13] where a somewhat similar
convergence is proved).
& ' a1 iii.) (Uniform integrability): The following condition
%
K
1
−K
%
a %
b
− 1b
Eh hk ≤ hk1 +(k2 −1)a dFh (h) holds: ! "
k=1 k1 =1 RK
+ k2 =1 lim sup Ehn hI h ≥ α = 0 (31)
(29) α→∞ n≥1
By using the lower bound in (18) and the integral inequality The proof is sketched: Using (29) and properties of the FFT
in (29) it is straightforward to see that (by the structure of in (1) the expectation can be upperbounded as:
the DFT) if the fading distribution is generated by a complex ( ( ( (−1
( ( −1 ( (
path gain distribution of which the marginal densities can be Ehn h ≤ Ehn (h̃1 ( (h̃K/2 (
written as fc̃i (c̃i ) , c̃i ∈ C, where fc̃i is circular symmetric,
⎛(⎛ ⎞2 ⎛ ⎞2 ((−1 ⎞
i.e. it is invariant under complex phase factors then the fading ( L/2
(
L/2 (
distribution is invariant regarding k1 in (29); hence we arrive ⎜( ( ⎟
= Ehn ⎝(⎝ c̃2l−1 ⎠ − ⎝ c̃2l ⎠ ( ⎠
at ( (
( l=1 l=1 (
Kn (
% ( 2
( (− K n Both sums are independent and converge in distribution to
(h̃k ( dFh̃n h̃
a circular symmetric Gaussian distribution. By assumption
n k=1
R2K
+ on the smoothness on the marginals this implies pointwise
% (( ((− E n2
convergence of densities and by Scheffé’s theorem [18] the
≤
n
(h̃k ( | m1 | Fh̃n h̃ convergence of the expectations. Since the limit expectation
R2K
is finite for two independent circular symmetric Gaussian
+ n
k∈Em 1
% (( ((− E n2
distributed complex path gains (refer to Theorem 1) this in
≤
n
(h̃k ( | m2 | Fh̃n h̃ turn implies eqn. (31).
R2K
iv.) (Lower bound): Using (17) we set Ehn (h) =
+ n
k∈Em 2
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
3756 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009
RK o K2o
/ 0 12
+
⎛⎡ ⎤ Ko
Ko ≤ Pr h(Ko ) > co
Ko − 1
≤
Ko
× ⎝⎣ f a− − f b− ⎦ 1
Ko − 1 i i
Ko The first inequality follows from Cauchy’s inequality and the
i≥1
⎞ Ko bound for the first integral follows from the independence
Ko −1 + of hk , k ∈ {1, ..., Ko} in combination with Prop. 1 with
− ai Ko f b i − f a+ ⎠
i monotonously decreasing density f . The probability can be
i≥1
again tackled with Tschebyscheff’s inequality. It follows
−/+ −/+
/ 0 σo
Here, 0 ≤ ai < bi ≤ ∞ are interval boundaries
+ +such Pr h(Ko ) > co ≤ 2
that f (h) ≤ 0, h ∈ a− , b −
and f
(h) ≥ 0, h ∈ ai , b i . Ko (co + HF )
i i
Proof: This lemma is proved in [11, Proposition 8]. and putting terms together
Now define the following random variable (i.e. partial
sums): C (P ∗ )
co σo
1
Ko ∗
≥ log (P ) + HF + log 1 + 2Ko + −HF
h(Ko ) := − log (hk ) e Ko 2Ko
Ko
k=1 √
eσo
+ −H √
By independence, we have h(Ko ) → (−HF ) in probability and e F Ko (co + HF )
we now show Eh(Ko ) (exp(h(Ko ) )) → exp(−HF ). Defining Ko large 2 co σo
some real co HF and splitting up the events in sets ≥ log (P ∗ ) + HF + √ 3
+ −H √
Ko e F 3 Ko
{h(Ko ) ≤ co } and {h(Ko ) > co } we obtain in the first case √
eσo
the inequality +√ (34)
Ko (co + HF )
(Ko ) / 0
/ (Ko ) 0
Eh eh I h(Ko ) ≤ co ≤ Eh min eh , eco concludes the proof.
(33)
using
( the set function
( I{·}. Defining further the event set A := R EFERENCES
{(h(Ko ) + HF ( ≤ } for some small > 0 the RHS of eqn. [1] S. Hanly and D. Tse, “Multi-access fading channels—part II: delay-
(33) can be upperbounded as follows: since limited capacities," IEEE Trans. Inform. Theory, vol. 44, no. 7, pp.
/ (Ko ) 0 2816-2831, Nov. 1998.
Eh min eh , eco I {A} ≤ e−HF + [2] G. Caire, G. Taricco, and E. Biglieri, “Optimum power control over
fading channels," IEEE Trans. Inform. Theory, vol. 45, no. 5, pp. 1468-
≤ e−HF + 2e−HF 1489, July 1999.
[3] E. Biglieri, G. Caire, and G. Taricco, “Limiting performance of block-
fading channels with multiple antennas," IEEE Trans. Inform. Theory,
for sufficiently small and vol. 47, no. 4, pp. 1273-1289, May 2001.
/ (Ko ) 0 [4] E. Jorswieck and H. Boche, “Delay-Limited capacity of MIMO fading
Eh min eh , eco I {Ac } ≤ eco Pr (Ac ) channels," in Proc. IEEE ITG Workshop Smart Antennas, Duisburg
(Germany), Feb. 2005.
[5] L. Li and A. Goldsmith, “Capacity and optimal resource allocation
we just need a bound on the probability Pr (Ac ) depen- for fading broadcast channels—part II: outage capacity," IEEE Trans.
dent on . The probability can be easily upperbounded by Inform. Theory, vol. 47, no. 3, pp. 1103-1127, Mar. 2001.
[6] C. Huppert and M. Bossert, “Delay-limited capacity for broadcast
Tschebyscheff’s inequality, i.e. channels," in Proc. 11th European Wireless Conf., Nicosia, Cyprus,
2005.
σo [7] E. Jorswieck and H. Boche, “Delay-limited capacity: multiple anten-
Pr (Ac ) ≤
K o 2 nas, moment constraints, and fading statistics," IEEE Trans. Wireless
Commun., vol. 6, no. 12, pp. 4204-4208, Dec. 2007.
where
[8] J. H. Sung and J. R. Barry, “Approaching the zeo-outage capacity of
2 MIMO-OFM without instantaneous waterfilling," IEEE Trans. Inform.
σo = Eh [log (h1 ) − Eh (log (h1 ))] Theory, vol. 54, no. 4, pp. 1423-1436, Apr. 2008.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.
WUNDER et al.: DELAY-LIMITED TRANSMISSION IN OFDM SYSTEMS: PERFORMANCE BOUNDS AND IMPACT OF SYSTEM PARAMETERS 3757
[9] S. Verdu, “Spectral efficiency in the wideband regime," IEEE Trans. Gerhard Wunder received his graduate degree
Inform. Theory, vol. 48, no. 6, pp. 1319-1343, June 2002. of electrical engineering (Dipl.-Ing.) in 1999 and
[10] G. Wunder and S. Litsyn, “Generalized bounds on the cf distribution of the Ph.d degree (on the peak-to-aver power ratio
OFDM signals with application to code design," IEEE Trans. Inform. problem in OFDM) in electrical engineering in 2003
Theory, revised July 2005. from Technische Universität (TU) Berlin, Germany.
[11] G. Wunder and T. Michel, “The delay-limited capacity region of OFDM He is now with the Fraunhofer German-Sino
broadcast channels," arXiv:cs/0610061, Dec. 2006. Lab for Mobile Communications, Heinrich-Hertz-
[12] G. Wunder and H. Boche, “Peak value estimation of band-limited signals Institut, leading several industry and research
from its samples, noise enhancement and and a local characterisation projects in the field of wireless communication sys-
in the neighborhood of an extremum," IEEE Trans. Signal Processing, tems. He is also a lecturer for detection/ estimation
vol. 51, no. 3, pp. 771-780, Mar. 2003. theory, stochastic processes and information theory
[13] S. Wei, D. L. Goeckel, and P. A. Kelly, “The OFDM signal converge at the TU Berlin, department for mobile communications. Recently, he also
weakly to a Gaussian random process: proof and application," in Proc. received the habilitation and Privatdozent degree from the TU Berlin in
39th Annual Allerton Conf. Commun., Control, Computing, Oct. 2001. communication engineering. His general research interests include estimation
[14] 3rd Generation Partnership Project (3GPP), “Feasiblity study for OFDM and information theory as well as crosslayer design problems.
for UTRAN enhancement," 3GPP, Tech. Rep. TR-25.892, v.6.3.0, rel.
6, May 2004. Thomas Michel received his graduate degree in
[15] 3GPP, “Technical specification group radio access network; user equip- Business Administration and Engineering (Dipl.-
ment, radio transmission and reception (FDD)," 3GPP, Technical Spec- Wirtsch.-Ing.) from TU Dresden, Germany in 2003
ification TS-25.101, Dec. 2004, v. 6.3.0, rel. 6. and his Ph.D. degree in Electrical Engineering from
[16] A. Kolmogorov and S. Fomin, Introductory Real Analysis. New York: TU Berlin, Germany in 2008.
Dover Publications, Inc., 1975.
[17] E. Lieb and M. Loss, Analysis. American Mathematical Society, 1997.
[18] P. Billingsley, Convergence of Probability Measures. New York: Wiley,
1968.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:19 from IEEE Xplore. Restrictions apply.