IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 54, NO.

1, JANUARY 2006 23

Constant-Power Waterfilling: Performance Bound

and Low-Complexity Implementation
Wei Yu, Member, IEEE, and John M. Cioffi, Fellow, IEEE

Abstract—In this letter, we investigate the performance of by Shannon’s Gaussian capacity formula1 SNR .
constant-power waterfilling algorithms for the intersymbol inter- However, because capacity is a logarithmic function of power,
ference channel and for the independent identically distributed the data rate is usually insensitive to the exact power allocation,
fading channel where a constant power level is used across a
properly chosen subset of subchannels. A rigorous performance except when the SNR is low. This motivates the search for
analysis that upper bounds the maximum difference between simpler power-allocation schemes that can perform close to
the achievable rate under constant-power waterfilling and that the optimum.
under true waterfilling is given. In particular, it is shown that Approximate waterfilling schemes often greatly simplify
for the Rayleigh fading channel, the spectral efficiency loss due transmitter and receiver design, and they have been the subject
to constant-power waterfilling is at most 0.266 b/s/Hz. Further-
more, the performance bound allows a very-low-complexity, of considerable study. In the multicarrier context, Chow [1]
logarithm-free, power-adaptation algorithm to be developed. empirically discovered that as long as a correct frequency
Theoretical worst-case analysis and simulation show that the band is used, a constant power allocation has a negligible
approximate waterfilling scheme is very close to the optimum. performance loss compared with true waterfilling. The same
Index Terms—Bit loading, duality gap, waterfilling. phenomenon is observed in the adaptive modulation setting [2].
There have been several performance bounds on constant-power
waterfilling reported in the literature. Aslanis [3] compared
the worst-case difference between a true waterfilling and a
I. INTRODUCTION constant-power waterfilling, and derived a bound based on
the SNR cutoff value. Schein and Trott [4] derived a different
W HEN a communication channel is corrupted by severe
fading or by strong intersymbol interference (ISI), the
adaptation of the transmit signal to the channel condition can
bound, also based on SNR. This letter extends the existing
results in several directions. First, a worst-case performance
typically bring a large improvement to the transmission rate. bound is derived using an approach based on convex analysis.
Adaptation is possible when the channel state is available to The upper bound derived is valid for any arbitrary SNR.
the transmitter, usually by a channel-estimation scheme and a Second, it is shown that the new performance bound can be
reliable feedback mechanism. With perfect channel information, used to design a low-complexity power-allocation algorithm
the problem of finding the optimal adaptation strategy has which is free of logarithm operations, and which has a bounded
been much studied in the past. If the channel can be partitioned worst-case performance. In particular, the algorithm is shown
into parallel independent subchannels, for example, when the to be at most 0.266 bs/s/Hz away from capacity on a Rayleigh
fading statistics for the fading channel is independent and fading channel and often performs much closer to capacity
identically distributed (i.i.d.) or by a discrete Fourier transform in practice.
(DFT) in the case of an ISI channel, the optimal transmit power In this letter, the primary focus is on power adaptation.
adaptation scheme is the well-known waterfilling procedure. In The bit allocation is allowed to vary, and is not restricted to
a waterfilling power spectrum, more power is allocated to better integer values. This approach is justifiable with the use of
subchannels with higher signal-to-noise ratios (SNRs), so that channel coding. In this case, the Shannon capacity for channels
the sum of data rates in all subchannels is maximized, where the with perfect transmitter and receiver side information can be
data rate in each subchannel is related to the power allocation achieved with a concatenation of a standard random Gaussian
codebook and a power-adaptation device [5]. In a related
work [6], schemes with both constant power and constant bit
allocation are investigated.
Paper approved by G. M. Vitetta, the Editor for Equalization and Fading The rest of the letter is organized as follows. In Section II,
Channels of the IEEE Communications Society. Manuscript received July 9, the waterfilling problem is formulated and the new upper bound
2004; revised June 16, 2005. This work was supported in part under a Stanford is derived. In Section III, a new low-complexity power-adap-
Graduate Fellowship and in part by France Telecom. This paper was presented
at the IEEE International Conference on Communications, Helsinki, Finland, tation algorithm is proposed, and its performance analyzed. In
June 2001. Section IV, the performance bound is applied explicitly to the
W. Yu was with the Electrical Engineering Department, Stanford University, Rayleigh fading channel. Simulation results for both wireless
Stanford, CA 94305 USA. He is now with the Electrical and Computer En-
gineering Department, University of Toronto, Toronto, ON M5S 3G4, Canada and wireline applications are presented in Section V. Conclu-
(e-mail: [email protected]). sions are drawn in Section VI.
J. M. Cioffi is with the Electrical Engineering Department, Stanford Univer-
sity, Stanford, CA 94305 USA (e-mail: [email protected]). 1In this letter, “log ” is used to denote logarithm of base 2; “ln ” is
Digital Object Identifier 10.1109/TCOMM.2005.861678 used to denote logarithm of base e.

0090-6778/$20.00 © 2006 IEEE

24 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 54, NO. 1, JANUARY 2006

II. CONSTANT-POWER WATERFILLING where , are convex functions. is

called the primal objective. The Lagrangian of the optimization
A. Problem Formulation problem is defined as
We choose to formulate the problem in the adaptive modu- (7)
lation framework because it is slightly more general than the
multicarrier setting. The communication channel is modeled as where are nonnegative. The dual objective is defined to be
. It is easy to see that is a lower bound
(1) on the optimal

(8)
where is the discrete time index, and are scalar
input and output signals, respectively, is the additive white
Gaussian noise (AWGN), which is i.i.d. with a constant variance (9)
, and is the multiplicative channel fading coefficient.
For simplicity, , the squared magnitude of the fading coeffi- (10)
cient, is assumed to be i.i.d. with a probability distribution .
The capacity of this fading channel under an average transmit So
power constraint when both the transmitter and the receiver have
(11)
perfect and instantaneous channel side information was char-
acterized by Goldsmith and Variaya [7]. They proposed a wa-
terfilling-in-time solution and proved a coding theorem based This is the lower bound that we will use to investigate the op-
on a finite partition of channel fading statistics, i.e., is re- timality of approximate waterfilling algorithms. The difference
stricted to take finite values , with probabilities between the primal objective and the dual objective
. In this case, the maximization problem becomes is called the duality gap. A central result in convex analysis
[10] is that when the primal problem is convex, the duality gap
reduces to zero at the optimum under some general conditions
maximize (2) known as constraint qualifications (which are satisfied for the
problem considered in this letter). In other words, the optimal
value of the primal objective may be obtained by maximizing
subject to (3) the dual objective over nonnegative dual variables .
Thus, for convex problems, the lower bound is tight.
(4)
C. Lower Bound
where is the average transmit power constraint, and the max-
imization is over all power-allocation policies based on the The above general result is now applied to the waterfilling
instant channel fading state . Putting reduces the problem. First, maximizing the data rate is equivalent to mini-
problem to the multicarrier setting. This optimization problem mizing its negative. The capacity is a concave function of power,
has a well-known waterfilling solution. Our interest is in finding so its negative is convex. The constraints are linear, so they are
approximate solutions with provable worst-case performance. convex, as well. Associate dual variable with the power con-
Note that Shannon’s Gaussian channel capacity formula straint, and with each of the positivity constraints on , the
is used here, and a capacity-achieving Gaussian codebook Lagrangian is then
is assumed. In reality, where practical codes and modulation
methods are used, the achievable rate can be computed by
the same formula with the noise variance increased by a
constant factor “SNR gap,” which denotes the amount of extra
coding gain needed to achieve Shannon capacity [8], [2]. (SNR (12)
gap is called SNR in [9].) Without loss of generality, the
SNR gap is assumed to be 0 dB for the rest of the letter, unless The dual objective function is the infimum of the La-
otherwise stated. grangian over primal variables . At the infimum, the partial
derivative of the Lagrangian with respect to must be zero
B. Duality Gap
(13)
The optimization problem (2) belongs to the class of convex
optimization problems, where a convex objective function is to from which the classical waterfilling condition is obtained
be minimized subject to a convex constraint set. A general form
of a convex optimization problem is the following: (14)

minimize (5) This condition, together with the constraints of the original
subject to (6) primal problem, the positivity constraints on the dual vari-
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 54, NO. 1, JANUARY 2006 25

ables, and the complementary slackness constraints, form the Assuming that the approximate waterfilling algorithm satisfies
Karush–Kuhn–Tucker (KKT) condition, which is sufficient and the power constraint with equality,3 the above
necessary in this case. More specifically, the complementary gives the following:
slackness condition states that the constraint for the original
primal problem is satisfied with equality if and only if the dual
variable associated with the inequality is strictly greater than (20)
zero. In the waterfilling problem, this translates to the condition
that is greater than zero if and only if is zero. Thus, when
a positive power is allocated in a subchannel, (i.e., , The preceding development is summarized in the following the-
), the sum of the signal power and the normalized orem.
noise power in each subchannel must be a constant, Theorem 1: For the optimization problem (2), if is
otherwise (i.e., when and ), the normalized noise a power-allocation strategy that satisfies the power constraint
power must exceed the water level. The waterfilling condition with equality, then the achievable data rate using is at most
gives the following optimal adaptation strategy: b/s/Hz away from the optimal waterfilling solution, where
is expressed in (20).
if This bound applies to all approximate waterfilling algorithms
(15)
if in general. For example, it can be used to bound the performance
of power-allocation strategies with an integer-bit constraint.4 It
where the cutoff point is determined by the average power is clear that if exact waterfilling is used, i.e., when
constraint and the fading distribution. is a constant whenever , the gap reduces to zero. There-
Substituting the waterfilling condition (14) into (12) gives the fore, the cost of not doing waterfilling is in the decrease of the
dual objective denominator in the second term. The simplicity of the above ex-
pression makes it quite useful in deriving new results, as it shall
soon be seen.

D. Constant Power Adaptation

(16) We now turn our attention to the particular class of constant-
power adaptation algorithms. As mentioned before,
The dual objective is always convex, and it is a lower bound to SNR is more sensitive to SNR when SNR is low. So, it makes
the primal objective2 for all nonnegative and . The lower sense that the critical task in waterfilling should be to ensure
bound is, in fact, tight when and achieve the optimum of that low SNR subchannels are allocated the correct amount of
the dual program. Finding the tightest and is equivalent power. In particular, those subchannels that would be allocated
to solving the original optimization problem, which is compli- zero power in exact waterfilling should not receive a positive
cated. However, if instead, the dual variables associated with power in an approximate waterfilling algorithm, for otherwise,
the primal variables are chosen via (14), then a simple bound the power is mostly wasted. This intuition allowed Chow [1]
emerges. In this case, the duality gap, defined as the difference to observe that a constant-power allocation strategy, where the
between the primal objective and the dual objective, and denoted transmitter allocates zero power to subchannels that would re-
as , has the following form: ceiver zero power in exact waterfilling, but allocates constant
power in subchannels that would receive positive power in exact
waterfilling, is often close to the optimal. In this section, this in-
(17) tuition will be made precise using the gap bound derived before.
Consider the following class of constant-power allocation
To express the gap exclusively in primal variables , a suitable strategies, where beyond a cutoff point , all subchannels are
needs to be found. A small is desirable, because it makes allocated the same power
the duality gap small. Since
if
(21)
if
(18)
Here, the subchannels are assumed to be ordered so that
whenever . If the same cutoff point is used as in exact
and recall that and need to be nonnegative, the smallest
waterfilling, we have
nonnegative is then
(22)

 must be used in
3When the power constraint is not satisfied with equality, S

(19) the second sum in (20) instead of p S .

4Equation (20) can be used to show that integer-bit restriction costs at most
2g (;  ) is a lower bound to the minimization problem. So 0 g (;  ) is 1= ln 2 b/s/Hz by noticing that an integer bit-allocation algorithm essentially
an upper bound to the rate-maximization problem. To avoid notational inconve- doubles S +  = in allocating each additional bit. Unfortunately, this bound
nience, the rest of the letter will use the term “duality gap” only. is rather loose.
26 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 54, NO. 1, JANUARY 2006

The inequality holds because in the transmission band (i.e.,

when ), the constant-power allocation is a suboptimal
strategy, therefore, the minimal sum of power and (normalized)
noise is less than the water level, which is . Equation (22)
ensures that

(23)

In this case, (20) becomes

Fig. 1. Constant-power waterfilling.

channels. In that case, will be nearly 1 for many subchan-

nels, and the duality gap becomes large (although still bounded
by the constant 1.44 b/s/Hz.) It is therefore of interest to use as
few subchannels as possible without violating the condition, so
(24) as to simultaneously make the number of terms in the summa-
tion small, and make each individual term small (since fewer
where denotes the number of channel states with positive subchannels implies larger , which, in turn, implies smaller
power allocation. Note that an immediate constant bound can ). This suggests that a simple power-allocation strategy
be obtained by replacing with 1. In which sets the cutoff point to be the largest that satisfies
this case, b/s/Hz is an upper bound to is close to the optimal. Graphically,
the maximum capacity loss for constant-power allocation algo- an algorithm that tries to find the smallest so that level A is
rithms. But this bound is usually too loose to be of practical less than level B has the smallest duality gap. This fact is used
interest. Instead, we can simplify the notation using the fact to devise the following algorithm.
that the number of bits allocated in each subchannel is given Algorithm 1: Assume that the channel gain ’s are ordered
by . In this case, can be written in a partic- so that . Let be the cutoff point, so that a
ularly simple form constant power is allocated for all . Let be the
largest , such that . The following steps find the
(25) with the smallest duality gap.
1) Set .
where is the number of bits allocated in each subchannel. 2) Compute .
Note that are not restricted to integer values in the above 3) If , set , and
bound. Also note that the crucial assumption for the bound to repeat step 2. Otherwise, set and go to the
hold is . Having the next step.
same cutoff point as in exact waterfilling is sufficient, but not 4) Compute for .
necessary. Thus, we have the following theorem. Set .
Theorem 2: For a constant-power allocation strategy of the Theorem 3: The low-complexity constant-power algorithm
form (21) that satisfies the power constraint with equality, if produces a rate that is at most b/s/Hz
, then it is at most from capacity.
b/s/Hz away from the waterfilling op- The proof of the theorem follows directly from the develop-
timal, where the sum is over all subchannels that are allo- ment in Section II. Two properties of this algorithm make it at-
cated amount of power, and is the number of bits allocated tractive. First, unlike most previous low-complexity bit-loading
in subchannel , i.e., . methods (e.g., [1]), where the boundary point is found by finding
Fig. 1 illustrates the theorem graphically. As long as the level the cutoff point that gives the highest data rate, this algorithm
A in Fig. 1 is lower than the level B, the achievable rate is finds the optimal cutoff point without actually computing the
bounded by (25). Note that subchannels with low SNRs (and data rate achieved in each step, and is therefore free of loga-
hence, low bit allocations) are precisely those contributing most rithmic operations. The most expensive operation in this algo-
to the bound, thus confirming the intuition that low SNR sub- rithm is the single division in each step, thus making its com-
channels are the most sensitive to power misallocation. plexity very low. Second, this algorithm has a provable worst-
case performance bound, as given by Theorem 3. Note that the
III. LOW-COMPLEXITY ADAPTATION algorithm is designed to minimize the duality-gap bound, which
The crucial condition in Theorem 2 is is not the same as maximizing the actual data rate. There are ex-
. This condition states that the bound is valid amples where the two criteria give quite different cutoff points.
only if not too few subchannels are used. The condition is triv- The presentation of the above algorithm has been simplified
ially satisfied, for example, by putting equal power in all sub- using a linear search. In practice, a binary search for can
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 54, NO. 1, JANUARY 2006 27

easily be implemented, thus further increasing its efficiency.

However, the asymptotic algorithmic efficiency is bounded by
the sorting of the subchannels, which is .
The simplicity of the algorithm also points to the possibility
of easy adaptive implementation when the channel distribution
is not known in advance. Since the power allocation is parame-
terized by the single cutoff point , an adaptive algorithm can
set an initial cutoff point, then adjust based on the resulting
power consumption. Hence, it is possible to approximately wa-
terfill without estimating the exact channel distribution in ad-
vance, and therefore, without the sorting operation.

IV. RAYLEIGH CHANNEL

The bound developed previously can be explicitly com-
puted if channel-fading statistics are known. In particular,
for a Rayleigh fading channel, it can be shown that the con-
stant-power adaptation strategy is only a small fraction of one
bit away from capacity. Fig. 2. Spectral efficiencies of exact waterfilling and constant-power
In a wireless channel where a large number of scatterers allocation on a Rayleigh channel.
contribute to the signal at the receiver, application of the central
limit theorem leads to a (zero-mean) complex Gaussian model this worst case, the average data rate is 1.3631 b/s/Hz, and the
for the channel response. The envelope of the channel response duality gap is b/s/Hz away from capacity.
at a given time instant has a Rayleigh distribution, i.e., the The following theorem summarizes the result.
square magnitude of the channel gain is exponentially dis- Theorem 4: For a flat i.i.d. Rayleigh fading channel with
tributed: , where , the average channel perfect side information at the transmitter and the receiver,
gain, parameterizes all Rayleigh distributions. assuming infinite granularity on the channel state, a con-
Fixing , the constant-power control strategy is determined stant-power adaptation method should allocate to all
by the average power constraint, or alternatively, by the cutoff subchannels that could support at least one bit, where is
value . The low-complexity power-allocation algorithm states determined from the power constraint. In this case, the resulting
that the constant power allocated in each state should be such spectral efficiency is at most 0.266 b/s/Hz away from capacity.
that
V. SIMULATION RESULTS
(26)
A. Wireless Rayleigh Channels
The Rayleigh distribution has a nonzero probability for arbi- The performance of the power-adaptation algorithm is
trarily large amplitudes of , so the above reduces to simulated on a Rayleigh channel. The low-complexity con-
. Interestingly, the constant-power allocation algorithm stant-power adaptation is used. The power transmission level
allocates a constant power to all subchannels that can sup- is determined by the instantaneous channel gain. The average
port at least 1 b/s/Hz with . channel gain ( ) is chosen to be dB. In Fig. 2, the av-
Now, using the gap bound (24), the spectral efficiency for erage spectral efficiencies of the exact waterfilling and the
an optimal constant-power allocation with cutoff is bounded low-complexity constant-power allocation are plotted against
within the following constant from capacity: the average power constraint, together with the duality-gap
bound. The average power constraint shown in the figure is the
(27) normalized value with noise power spectral density level set
to dB. It is seen from Fig. 2 that the true waterfilling
By a change of variable (and also ), define and the low-complexity algorithm give indistinguishable re-
sults. The capacity upper bound for the low-complexity power
(28) allocation as computed by (29) is also plotted in Fig. 2. As it
can be seen, the true capacity lies well within the upper bound.
the duality gap can be expressed as However, while the capacity bound may be loose, the con-
stant-power allocation method designed using the performance
(29) bound nevertheless works very well.
The authors are not aware of a closed-form expression for the
integral (28). Numerical evaluation reveals that it has a single B. Digital Subscriber Line Channels
maximum occurring at about 0.39, and the value of the The performance of the proposed low-complexity con-
maximum is about 0.1840. The duality gap is largest when the stant-power waterfilling algorithm is also simulated for the
power constraint is such that the cutoff point . In very-high-speed digital subscriber line (VDSL) application.
28 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 54, NO. 1, JANUARY 2006

the wireline ISI channel. Our main contribution is a rigorous

performance bound for the constant-power waterfilling algo-
rithm based on the duality-gap analysis in convex optimization.
Furthermore, the duality-gap analysis allows a very-low-com-
plexity constant-power adaptation method to be developed. The
low-complexity algorithm has the desirable properties of having
a provable worst-case performance and being logarithm-free.
The performance bound is applied to Rayleigh fading channels.
It is shown that constant-power adaptive modulation is at most
0.266 b/s/Hz away from capacity. Simulation results suggest
that the actual gap is even smaller.

REFERENCES
[1] P. S. Chow, “Bandwidth optimized digital transmission techniques for
spectrally shaped channels with impulse noise,” Ph.D. dissertation, Stan-
ford Univ., Stanford, CA, 1993.
Fig. 3. Achievable rates for VDSL lines at various line lengths. [2] A. J. Goldsmith and S. Chua, “Variable-rate variable-power MQAM for
fading channels,” IEEE Trans. Commun., vol. 45, no. 10, pp. 1218–1230,
Oct. 1997.
VDSL 26-gauge transmission lines are simulated at various [3] J. T. Aslanis, “Coding for communication channels with memory,” Ph.D.
distances. VDSL twisted-pair cables are severe ISI channels. dissertation, Stanford Univ., Stanford, CA, 1989.
[4] B. Schein and M. Trott, “Sub-optimal power spectra for colored
VDSL transmission can potentially use 4069 tones, with each Gaussian channels,” in Proc. IEEE Int. Symp. Inf. Theory, 1997, p. 340.
tone occupying 4.3125 kHz bandwidth. A mix of alien crosstalk [5] G. Caire, G. Taricco, and E. Biglieri, “Optimum power control over
signals are also included. A moderate combined SNR gap and fading channels,” IEEE Trans. Inf. Theory, vol. 45, no. 7, pp. 1468–1489,
margin of 12 dB is assumed. The data rates achievable with the Jul. 1999.
[6] D. Dardari, “Ordered subcarrier selection algorithm for OFDM-based
low-complexity constant-power waterfilling algorithm and that high-speed WLANs,” IEEE Trans. Wireless Commun., vol. 3, no. 5, pp.
with true waterfilling are plotted as a function of line length in 1452–1458, Sep. 2004.
Fig. 3. At distances shown in Fig. 3, between 700–3000 tones [7] A. Goldsmith and P. Varaiya, “Capacity of fading channel with channel
side information,” IEEE Trans. Inf. Theory, vol. 43, no. 11, pp.
are typically used. Clearly, constant-power waterfilling has 1986–1992, Nov. 1997.
negligible rate loss, compared with true waterfilling. [8] J. M. Cioffi, “Digital communications,” Stanford Univ., Stanford, CA,
Course Notes, EE379A.
[9] D. Forney and G. Ungerboeck, “Modulation and coding for linear
VI. CONCLUSIONS Gaussian channels,” IEEE Trans. Inf. Theory, vol. 44, no. 10, pp.
2384–2415, Oct. 1998.
In this letter, we investigate low-complexity power-adap- [10] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
tation algorithms for both the wireless fading channel and U.K.: Cambridge Univ. Press, 2004.