IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO.

2, MARCH 2003 375

Optimum Power Control for Successive Interference

Cancellation With Imperfect Channel Estimation
Jeffrey G. Andrews, Member, IEEE, and Teresa H. Meng, Fellow, IEEE

Abstract—Successive interference cancellation, in conjunction Despite the abundance of academic work on MUD, in-
with orthogonal convolutional codes, has been shown to approach dustry implementations still predominantly use the single-user
the Shannon capacity for an additive white Gaussian noise channel matched filter. There are several explanations for industry’s
(Viterbi: 1990). However, this requires highly accurate estimates
reluctance to use the results from MUD, and they usually
for the amplitude and phase of each user’s signal. In this paper,
we derive an optimal power control strategy specifically designed center around continuing questions about the complexity of
to maximize the overall capacity under the constraint of a high even the reduced-complexity suboptimal techniques and the
degree of estimation error. This power control strategy presents robustness of such techniques to the difficulties of the multicell
a general formula of which other power control algorithms wireless channel. Successive interference cancellation (SIC) as
are special cases. Even with estimation error as high as 50%, proposed in [1] is a MUD technique that is different from much
capacity can be approximately doubled relative to not using of the MUD research in that it does not rely on dimensional
interference cancellation. In addition, when properly applied to
multicell mobile networks, this power control scheme can reduce separation or short-period spreading sequences in order to
the handset transmit power, and therefore other-cell interference, distinguish users from one another. Further, its entire design
by more than an order of magnitude. as presented in this paper is based on an extremely strong
error-correcting code. For these reasons, it is well-suited to an
Index Terms—Code-division multiple access (CDMA), inter-
ference cancellation, multiuser detection (MUD), power control, uncoordinated, noisy, asynchronous environment such as the
superorthogonal codes. uplink in a cellular system.
There are some serious challenges in making a SIC system
feasible in practice. First, the decoding time increases linearly
I. INTRODUCTION with the number of users. This is because users are decoded suc-
cessively, as implied in the name of the technique. However, a
F OR A number of reasons, code-division multiple access
(CDMA) continues to be a dominant air-interface tech-
nology for personal wireless communication systems. While
latency increase that is linear with the number of users is gener-
ally considered palatable because processor speed is increasing
systems based on the widely available commercial standards exponentially. Second, relative to conventional CDMA, a more
such as IS-95 and the newer third-generation (3G) standards complicated power control distribution is required to make full
wideband code-divison multiple access (W-CDMA) and use of SIC, because the users must be received with differing
CDMA2000 have proven reasonably robust for low bandwidth powers, dependent on the order of decoding. Third, the ampli-
applications, it is generally believed that significant increases tude and phase of each user must be accurately estimated. If
in capacity and performance are attainable for future CDMA the estimates are inaccurate, residual interference remains in
systems. An abundance of theoretical and practical research the composite signal, and the system capacity rapidly erodes.
has been undertaken with this goal in mind. Fundamental work Fourth, as is true of all realistic MUD systems, other-cell inter-
done by Verdu [2] showed the remarkable extent to which the ference (OCI) is uncancelable and, thus, proposes a particular
single-user matched filter present in IS-95 systems could be problem.
improved upon by using more sophisticated receiver design In this paper, we focus on the latter two problems. In the at-
and signal processing. While the optimal implementations of tempt to relax the requirement on accurate amplitude and phase
this work are prohibitively complex for even a modest numbers estimation, a novel and general power control algorithm is de-
of users, an assortment of suboptimal methods have been veloped that is shown to be optimal for all CDMA systems. For
developed which reduce the complexity drastically while still the sake of receiver simplicity, a channel with only one path
providing large gains over the conventional single-user detector. from transmitter to receiver is assumed, but the power control
This field has come to be known as multiuser detection (MUD), results apply to a multipath channel as well, since the power
and an accessible summary of the field can be found in [3]. control distribution only depends on the total amount of power
received per user. It will be shown in Section IV that if the es-
timation error is considered when developing a power control
Manuscript received May 4, 2001; revised October 30, 2001; accepted distribution, a sizeable amount of estimation error can be toler-
November 2, 2001. The editor coordinating the review of this paper and ated while still maintaining robust bit-error rate (BER) perfor-
approving it for publication is J. K. Cavers.
J. G. Andrews was with the Electrical Engineering Department, Stanford Uni-
mance at an increased spectral efficiency. In addition, extending
versity, Stanford, CA 94305 USA. He is now with the Department of Electrical the work of [4] and [5], it will be shown in Section V that if the
and Computer Engineering, University of Texas, Austin, TX 78712–1084 USA users’ relative distances from the cell are considered when as-
(e-mail: [email protected]). signing powers, OCI can be reduced by approximately an order
T. H. Meng is with the Electrical Engineering Department, Stanford Univer-
sity, Stanford, CA 94305 USA (e-mail: [email protected]). of magnitude over equal power CDMA systems such as IS-95
Digital Object Identifier 10.1109/TWC.2003.809123 and 3G CDMA. In Section VI, the spectral efficiency of a SIC
1536-1276/03$17.00 © 2003 IEEE
376 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 2, MARCH 2003

Fig. 1. System block diagram. (a) Transmitter. (b) Channel. (c) Receiver.

where is the gain factor (due to power control), is the

binary serial sequence of encoder output, and are
the binary in-phase and quadrature scrambling sequences,
is the carrier frequency in radians per second, and is the
impulse response of the pulse-shaping filter. and
Fig. 2. Superorthogonal encoder may have an arbitrarily long period and, thus, are modeled as
psuedorandom Bernoulli sequences.
system with optimal power control will be demonstrated and To simplify the analysis and relate it directly to simulated re-
compared with other approaches. sults, we consider the discrete-time baseband transmitted signal
as
II. SYSTEM MODEL
(2)
A. Transmitter
(3)
The transmitter, channel, and receiver models are shown in
Fig. 1. At the transmitter, each user’s data bits are encoded by Perfect separation between the in-phase and quadrature chan-
a superorthogonal convolutional encoder. This powerful code nels is assumed, so all digital-domain analysis can be considered
is proposed and described in [6] and the encoder is shown in for uncorrelated and .
Fig. 2. The spreading gain achieved by a superorthogonal code
is , where is the constraint length of the code. The benefits B. Channel
of the proposed system and power control distribution can also The channel is modeled as an asynchronous fading channel
be achieved with other low-rate convolutional codes, such as with additive white Gaussian noise (AWGN). Each user’s signal
those recently proposed in [7], which also has slightly superior experiences an independent delay during transmission, but
performance to the superorthogonal code and can accommodate it is assumed that the receiver can learn the value of this delay
spreading factors that are not powers of two. through the usual methods employed in commercial systems.
After the data has been encoded, it is split into and Thus, the asynchronicity is relevant in that it demonstrates
branches and scrambled by independent binary sequences, in that no alignment of the users is required for the system.
order to ensure that other-user interference produces a random Fast closed-loop power control must be employed in practical
component whose variance is independent of the relative phases CDMA systems in order to mitigate the effects of rapid changes
between users [1]. The resulting and signals are converted in the received signal strength. In this work, it is assumed that
to analog signals and then run through a pulse shaping low- the power control helps neutralize the fading, but some residual
pass filter before being quadrature modulated by the carrier fre- power control error (PCE) remains, which can be thought of as
quency. The resulting transmitted signal for user is unmitigated fading.
The received signal

(4)
(1)
ANDREWS AND MENG: OPTIMUM POWER CONTROL FOR SUCCESSIVE INTERFERENCE CANCELLATION 377

is the sum of the transmitted signals delayed by their respective III. OPTIMUM POWER CONTROL
propagation times (assumed here to be an integer multiple of the
sample interval), plus additive noise , which has noise power Power control is required for all realistic CDMA systems be-
. Path loss is neglected throughout the paper. The power con- cause of what is known as the near-far problem: users far from
trol error arises from the imperfect mitigation of fading the base station experience far greater path loss than users that
and will be quantified in detail when the optimum power con- are near the base station. Optimum power control is achieved
trol distribution is presented. when all users are decoded with the same signal-to-interference
ratio (SIR) [8]. Otherwise, a user with a low SIR dominates the
C. Interference Cancelling Receiver BER performance of the system, which is defined as the average
BER over all users.
As implied by the name of the technique, in a SIC system, In commercial CDMA systems, the near–far problem is mit-
users’ signals are extracted from the composite received signal igated by controlling the output power of the mobile units with
successively, rather than in parallel. SIC attempts to remove the a tight feedback loop, so that the users’ signals all arrive at the
interference of the th user (the most recently decoded user) base station with approximately the same power, which results
from the current composite received signal , by re-en- in a consistent quality of service, as each user experiences an
coding the decoded bit sequence for user , modulating it with
approximate SIR of
the appropriate amplitude and phase adjustment, and subtracting
it out from . This process is illustrated in Fig. 1(c). The
forward path is similar to that of a typical CDMA matched filter SIR (10)
receiver: The down converted and sampled signal is despread
with the synchronized pseudonoise (PN) sequences for user
where is the received power of each user, is the
and then combined and decoded by an appropriately modified
number of users, and is the power of the background AWGN,
Viterbi decoder for superorthogonal codes. It is assumed that
which can also include OCI, assuming such interference appears
synchronization with each user is achieved through the usual
as noncoherent additive noise.
methods, namely, an overhead channel with training, and then
a phase locked loop. As stated previously however, no cooper- When SIC is used, the situation is significantly different. In
ation between users is assumed. this case, it is also desirable that each user experiences the same
Once reliably decoded, user ’s decoded bits can then be used SIR at the time of decoding. However, interference is being sub-
to cancel the interference that its signal would cause to later tracted out of the received signal after each user, so the first user
users. The estimated bits for user are reencoded, and estimates to be decoded sees the most interference, the last user the least.
of the amplitude and phase, or equivalently the amplitude of the Heuristically, the first user to be decoded should be the strongest
and branches are formed user, the weakest user should be decoded last.
If the successive cancellation scheme proceeds with no
channel estimation error or bit errors, then finding the optimum
(5) power control scheme is straightforward as described in [9]. Of
course, the amplitude and phase estimation are never perfect
and, thus, it is desirable to know the optimum power solution in
(6) the presence of imperfect cancellation. If there is cancellation
error, the following equations describe the SIRs for each
user:
where is the number of symbols in a frame, and and
are the amplitude estimates of user ’s in-phase and quadrature
branches, respectively.
Using these values, an estimate of the received signal from
user can be obtained as
(11)
(7)

The stored composite signal may then be updated

where is again the number of users and is the frac-
(8) tion of the th user’s power not cancelled. We desire
such that ,
since the will directly determine the BER.
(9)
In (11), there are equations and unknown relative
power weightings, since one of the can be set to an arbitrary
Thus, it is intuitive that the first user is exposed to the most value depending on the desired receiver sensitivity. These equa-
multiple access interference (MAI), while the final user sees a tions can be solved in terms of the SIR as in [5] and [10], but
composite signal with a large amount of MAI removed from it. ideally one would like to equalize the users’ SIRs in the pres-
This motivates the discussion of the next session on optimum ence of interference without knowing the target SIR. Hence, a
power control. recursive approach was adopted and the derivation is shown in
378 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 2, MARCH 2003

Fig. 3. Algorithm for computing fP g.

the Appendix. The resulting power control distribution for user

can be expressed as

(12) Fig. 4. Sample optimal power distributions for SIC.

where is the fractional residual cancellation error for user , It should also be noted that the above reasoning assumes
and is the total remaining multiple-access interference (MAI) that the fractional cancellation error is known for each user.
for user plus their own power: Clearly, the amount of cancellation error is rarely known and
must instead be estimated or guessed. This will be addressed in
Section IV.
(13) A few sample power distributions are shown in Fig. 4, as a
function of the amount of uncancelable noise and cancellation
Note that this is the general optimal solution for CDMA power error. As can be seen, the relative distribution of powers amongst
control, and that although it is possible that if many bit users is highly dependent on the amount of cancellation error
errors are made, all cases of interest are for . For perfect in- and somewhat dependent on the relative amount of noise power.
terference cancellation and (12) is shown in Appendix II Note that when the cancellation error and to a lesser extent the
to be identical in this case to the distribution derived in [9]. For noise are kept low, the power differential between earlier and
no interference cancellation as in a typical equal power com- later users is much greater than if those quantities are high, be-
mercial CDMA system, and it can be easily seen that cause more successful interference cancellation will take place
(12) reduces to the familiar equal power solution. for earlier users and, thus, the later users will require less power
These equations still cannot be solved analytically for all the to achieve the same signal to interference ratio. As can be seen
given , due to the introduced variable , but in Fig. 4, the dynamic range in received power is less than 10 dB
they can be solved quickly by iteration to arbitrary accuracy. We for most cases of practical interest.
provide a simple algorithm in Fig. 3 for computing the optimal
power weightings , where is some chosen step size. IV. ESTIMATION AND POWER CONTROL ERROR (PCE)
Following the above steps, the will converge to the power MODELING AND ANALYSIS
distribution given in (12) given a total power constraint as A. PCE
. Unlike other uplink power control schemes [11], there
PCE results when a user is received with a power level that is
are no convergence conditions regarding a target SIR in this al-
different than that assigned by the base station. This occurs due
gorithm. The proposed algorithm simply equalizes the received
to the processing and propagation delay between the transmitter
SIRs and hence always converges for . Convergence is
and receiver that makes it difficult to track fast changes in the
defined as the ability to make arbitrarily
channel, and also because typically just one bit “up” or “down”
small. The argument that the algorithm presented in Fig. 3 con-
commands are sent to the mobile. Nevertheless, fast power con-
verges to (12) is as follows.
trol has proven effective for large-scale commercial cellular sys-
Initially, . It is known from simple inspection of
tems such as IS-95 and WCDMA.
(12) that with equality iff . Thus, by
As can be seen in Fig. 1, PCE is applied to the received signal
using (12), it can be seen initially that . Hence,
in order to realistically model the received power over a fading
all initial power estimates are conservative, and the initial total
channel. PCE has been found to closely follow a lognormal dis-
power . Thus, by increasing by , the
tribution [6], [12] and is defined to have the following normal-
instant . Thus, by letting and
ized variance:
following the algorithm in Fig. 3, . In practice, is a
finite value and hence , but can be made as small as
(14)
desired by lowering at the cost of increased iterations.
ANDREWS AND MENG: OPTIMUM POWER CONTROL FOR SUCCESSIVE INTERFERENCE CANCELLATION 379

where is the actual received power for user , and is the

assigned (optimal) power for user . Thus, the received power
is modeled as

(15)
where
(16)

For simplicity, it is assumed that the fading is uncorrelated from

frame to frame, but constant over a frame. Thus, takes on
a different value over each frame.

B. Estimation Error Modeling

As seen in the power control distribution of Section III, in
order to optimally assign powers, the amount of cancellation
error per user must be known. Cancellation error has two
sources: incorrect bit decisions and imperfect amplitude and
phase estimation. Because the BER is assumed to be low, Fig. 5. BER versus estimation error for various values of "^.
virtually all of the cancellation error comes from amplitude
and phase estimation error. Naturally, estimation error is not at the cancellation error. Thus, we have three different quan-
typically known either, so in this section, a model for estimation tities related to estimation error and for clarity it is important to
error will be presented. Using this model, it will be shown distinguish between them.
that despite the lack of knowledge of the exact amount of
1) the amount of cancellation error , which is unknown and
estimation error, a conservative estimate of the estimation error
expressed as a fraction of the received power;
will produce robust performance.
2) the standard deviation of the amplitude and phase estima-
In the AWGN channel model presented, the amplitude and
tion error . This is also unknown but modeled in our
phase estimation in (5) and (6) will produce an estimate of the
simulations;
received power over each of the and branches of the system.
3) a guess at or estimate of the cancellation error , which in
In any realistic system, this estimate will not be perfect for a
general will be quite close to . This guess is called and is
number of reasons including finite frame length and a dynamic
the value used for computing the power control distribution
channel, and this estimation error will cause imperfect cancel-
in (12).
lation of the estimated signal. The amount of residual interfer-
The system is simulated using parameters in Table I and the
ence is expressed as a fraction of the user’s total power, and
results are shown in Fig. 5. While this plot may at first seem
this is the fractional residual estimation error , as introduced
confusing, it is easy to understand if two underlying principles
in Section III. In order to create a realistic model for this esti-
are kept in mind.
mation error for simulation, errors are induced in the estimates
of (5) and (6). The estimation error is assumed to follow a log- 1) Due to error propagation in SIC, the system is less sensitive
normal distribution like the PCE and is, thus, modeled similarly to being overestimated than underestimated.
2) Because relates to a probability distribution, the rare
(17) instances when the estimation error is large dominate the
BER performance. This is the rationale for choosing
i.i.d for (18) conservatively.
For example, as can be seen from Fig. 1, a BER of 10
Thus, is the standard deviation of the estimation error for
can be achieved even if the standard deviation of the estimation
the amplitude estimates of the in-phase and quadrature branches
error is as high as about 0.30, as long as is simply chosen to
, and is approximately equal to the total fractional can-
be 0.27. Thus, from this example, no actual knowledge of the
cellation error since the amplitude and phase are completely
cancellation error is required as long as it remains under about
determined by the and amplitudes. We use the notation
30% of the received signal power.
suggestively for both the variance of the estimation error and
for the cancellation error because the majority of the cancella-
tion error derives from inaccuracies in the channel estimation, V. OCI REDUCTION
as shall be seen in the next section. In order for any interference cancellation system to op-
erate effectively, uncancelable interference must be kept to a
C. Estimation Error Analysis minimum. In most MUD systems, SIC included, OCI cannot
Using the models of the preceding two subsections, we will be cancelled because the signatures and timing of users in the
now analyze the performance of SIC with the proposed power neighboring cells are unknown to the base station in question.
control algorithm. The BER is plotted as a function of the esti- Thus, a method for the reduction of OCI is highly desirable.
mation error in Fig. 5. Each curve represents a different “guess” It will be shown in this section that SIC provides a method
380 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 2, MARCH 2003

TABLE I a SIC system than a conventional system. Thus, they will cause
SIMULATION PARAMETERS less interference to neighboring cells, and this reduction will be
quantified in the next section. Second, because far-away users
can now lower their power levels, the dynamic range required
for accurate power control will be reduced. Third, maintaining
an accurate power control distribution is important for system
capacity, as will be shown in Section VI. Thus, it is preferred
to tightly control the users’ power levels, relative to simply es-
timating the power levels and then ordering the decoding.

B. OCI Reduction
In order to quantify the OCI reduction, the OCI in a SIC
system with power levels as in (12) shall be compared with the
OCI in a conventional equal power system. The total average
OCI reduction is the average OCI reduction in each cell times
the number of neighboring cells

OCI
for drastically reducing OCI relative to that of a commercial
CDMA system. (20)
OCI
A. Power Assignment Strategy
A well-known model for path loss in cellular systems is given
by (21)

(19)

where is the received power, is the transmit power, is (22)

some reference distance (typically one meter), is the path loss
at a distance , is the separation distance of the transmitter
and receiver, and is the “path loss exponent,” usually between
2.5 and 6, often taken for macrocells to be four in the absence
of empirical data. (23)
Because of the typically large path loss exponent, users that
are far away from the base station must transmit at a much higher
power level than those close to the base station, if their power
levels are to be comparable at the receiver. This is known as the where is the number of neighboring cells, is the distance
“near–far problem” and necessitates power control in any prac- to a neighboring base station, is the distance to the desired
tical CDMA system. However, in a CDMA system using SIC base station, is the received power for user
as described in this paper, disparate powers amongst users are from (12), is the received power for user in a con-
actually preferable. Some work has argued that this relaxes the ventional system, and OCI is the interference power received
need for accurate power control: Simply decode first the users at the neighboring base station from user . It is assumed that
with the strongest receive powers [14]. In this paper, we propose the users are uniformly positioned throughout cells with a cir-
a different approach. The users far away from the base station cular coverage area of radius and that the neighboring base
can be assigned the lower power levels (and, thus, be decoded station is a distance from the desired base station.
later), while the users close to the base station can be assigned For fairness, and are constrained such that the
the higher power levels. While this strategy does not relax the conventional system and a SIC system have an equivalent SIR
need for accurate power control, it has several beneficial effects for each user at the time of decoding, that is, by setting
on OCI and capacity that more than compensate for the increase
in complexity. SIR SIR (24)
First and most importantly, the users closest to the neigh-
boring cells are the users who are typically farthest from the SIR (25)
desired base station. They must raise their powers to reach the
base station but in doing so cause increased interference to the SIR (26)
neighboring cell. By assigning these users the lower received
power levels, they will transmit with significantly less power in
ANDREWS AND MENG: OPTIMUM POWER CONTROL FOR SUCCESSIVE INTERFERENCE CANCELLATION 381

Fig. 6. OCI reduction. Fig. 7. SIC OCI versus conventional OCI.

Constraining SIR SIR is actually conservative since

at a given SIR and spreading factor, a CDMA system based
on superorthogonal codes will outperform a system with con-
catenated convolutional codes, Walsh modulation, and repeti-
tion codes, as used in IS-95 [6].
The conventional power distribution is easily
found empirically since

Constant (27)

However, the optimal SIC power distribution is

found recursively, so closed-form results for the OCI reduction
are not presented.
Note that since the OCI reduction considers a ratio of average
OCI reduction over all users and that the users are uniformly
distributed throughout the cell, neither the radius of the cell, Fig. 8. Spectral efficiencies for SIC at BER = 10 .

nor the number of users affect the results. As can be seen in

Fig. 6, the OCI reduction is an inverse function of two quanti- VI. CAPACITY
ties: the uncancelable interference (OCI and thermal noise) and
In the previous two sections, it has been demonstrated that the
the amount of estimation error . It is an inverse function of
proposed SIC system can be designed to be robust to estimation
these quantities because as noise and estimation error increase,
the power distribution becomes tighter because less interference error while simultaneously reducing the amount of OCI occur-
cancellation is possible. Since this makes the SIC system ap- ring throughout the system. Capacity, or equivalently spectral
proach an equal power system, clearly the gain from SIC de- efficiency, is defined in this paper as the amount of traffic that
creases in this case. Contrary to what intuition might predict, can be accommodated in a fixed bandwidth at a specified BER.
the OCI reduction is only very weakly dependent on the path The BER specification is taken to be 10 . Commercial CDMA
loss exponent . It can be seen that an improvement of about systems typically deploy sectorized antennas, which further in-
10 dB can be made even at high noise levels, when half of all crease capacity by approximately the number of sectors. For
interference comes from AWGN and users in neighboring cells, generality, in this work, just the capacity per sector is consid-
as is typical in equal-power commercial systems [14]. ered, so the achievable capacity per cell would be an integer
It is also of interest to know what a certain percentage of OCI (often three) multiple of the capacity presented here [15].
in an equal power system translates into in a SIC system. This is Simulation of the proposed SIC system in a low OCI envi-
shown in Fig. 7. As can be seen, a normal CDMA system with ronment ( 10 dB for all users) resulted in spectral
as much as 50% of its total interference coming from users in efficiencies summarized in Fig. 8 as a function of the amount of
other cells is reduced to under 10% in a SIC system using the estimation error. The top curve is for the proposed system, using
power assignment strategy proposed in this paper. the power control scheme shown in Section III and assuming
382 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 2, MARCH 2003

that the variance of the estimation error is known, i.e., . APPENDIX I

Agreeing with intuition and Fig. 5, the and DERIVATION OF OPTIMAL POWER ALLOCATION
curves show that there is a capacity penalty for choosing con- Defining the total power and cancellation
servatively when there is a small amount of estimation error but efficiency , the first equation for SIR
that the system capacity remains far more robust as estimation becomes
error increases.
If it is assumed that there is no estimation error when com- (28)
puting the power control distribution, as is generally done in the
literature, it is seen in Fig. 8 that the capacity is greatly reduced Solving for gives the power for the second user as a function
if estimation error does occur. At (estimation error of of the power of the first user and the efficiency of cancellation
approximately 50%), a system with optimum power control still
(29)
has about twice the capacity of the same system without SIC. On
the other hand, a system which falsely assumes that there is no The subsequent equations are less straightforward. We will find
estimation error when designing the power control distribution as an example, and then, the general result for will be
is far worse off with SIC than with no interference cancellation presented by induction. The second equation for SIR is
at all.
The two “curves” which do not change as a function of esti- (30)
mation error are provided for reference. The higher value (
b/s/Hz) consists of the transmitter in Fig. 1, but the receiver Introducing a convenient notation for the total remaining MAI
does not perform any interference cancellation. This line serves to user plus their own power
as a comparison basis to show how much SIC improves the
system performance. As can be seen, with optimum power con- (31)
trol, SIC adds significantly even as the estimation error grows
large. On the other hand, if the estimation error is incorrectly as- Substituting into (30) gives
sumed to be zero, for , the system is better off without
SIC. (32)
The spectral efficiency for a commercial IS-95 system
is shown as a comparison. A typical IS-95 system with which results in a solution for the third user’s power which is de-
three-sector antennas is reported by Qualcomm [16] to allow pendent on the second user’s power and cancellation efficiency
around 85 users per cell with an average data rate of 4 kb/s :
in a bandwidth of 1.25 MHz, at an approximate BER of
10 to 10 . This corresponds to a spectral efficiency of (33)
0.09 b/s/Hz/sector. It is important to note that this number was
quoted for a real-world channel (not flat fading), but it shows The remaining equations can be solved in an identical fashion,
the extent to which sophisticated signal processing may be able resulting in a recursive relation
to improve the capacity of CDMA systems.
(34)

VII. CONCLUSION (35)

In order for SIC to work properly, a power control algorithm

which takes inevitable channel estimation error into account APPENDIX II
is required. It is our contention that fast and appropriately de- PROOF THAT (12) IS EQUIVALENT TO PREVIOUS RESULT
FOR PERFECT CANCELLATION
signed power control is a key element in allowing a SIC system
to achieve high performance in practice. While this introduces Here, it is proven that (12) converges to the solution in [9]
complexity into the system, the potential rewards for doing so when : Letting , which represents the perfect
are considerable. A general formula for the optimum power con- cancellation case, (12) becomes
trol distribution for SIC and conventional CDMA was derived
in this paper. Using this distribution, it was shown that channel (36)
estimation error up to 50% can be tolerated, while still at least
doubling the capacity of a system without SIC. On the other where is a target SIR, and , with
hand, using suboptimal power control results in greatly reduced
(37)
capacity. In addition to this large gain in capacity, OCI can be si-
multaneously reduced by around an order of magnitude if users
are assigned power levels based on their distance from the base (38)
station.
ANDREWS AND MENG: OPTIMUM POWER CONTROL FOR SUCCESSIVE INTERFERENCE CANCELLATION 383

Using this result, (36) becomes [13] P. Patel and J. Holtzman, “Analysis of a simple successive cancellation
scheme in a DS/CDMA system,” IEEE J. Select. Areas Commun., vol.
12, pp. 796–807, June 1994.
(39) [14] A. J. Viterbi, A. M. Viterbi, and E. Zehavi, “Other-cell interference in
cellular power-controlled CDMA,” IEEE Trans. Commun., vol. 42, pp.
1501–1504, Feb.-Apr. 1994.
[15] K. S. Gilhousen et al., “On the capacity of a cellular CDMA system,”
By definition of the target SIR IEEE Trans. Veh. Technol., vol. 40, pp. 303–312, May 1991.
[16] S. Vembu and A. J. Viterbi, “Two different philosophies in CDMA—A
comparison,” in Proc. 46th IEEE Vehicular Technology Conf., vol. 2,
(40) May 1996, pp. 869–873.
[17] D. Divsalar, M. K. Simon, and D. Raphaeli, “Improved parallel inter-
ference cancellation for CDMA,” IEEE Trans. Commun., vol. 46, pp.
258–268, Feb. 1998.
By using the expression in (39) and summing the series, with [18] J. Andrews and T. Meng, “Amplitude and phase estimation consider-
some additional algebra it can be shown that ations for asynchronous CDMA with successive interference cancella-
tion,” in Proc. IEEE Vehicular Technology Conf., Boston, MA, Sept.
2000, pp. 1211–1215.
(41) [19] , “Transmit power and other-cell interference reduction via suc-
cessive interference cancellation with imperfect channel estimation,” in
Inserting (41) into (39) results in the perfect interference can- Proc. IEEE Int. Conf. Communications, Helsinki, Finland, June 2001,
pp. 1940–1944.
cellation case derived in [9]

(42)
Jeffrey G. Andrews (S’98–M’02) received the B.S. degree in engineering (with
high distinction) from Harvey Mudd College, Claremont, CA, in 1995 and the
M.S. and Ph.D. degrees in electrical engineering from Stanford University, Stan-
REFERENCES ford, CA, in 1999 and 2002, respectively.
[1] A. Agrawal, J. Andrews, J. Cioffi, and T. Meng, “Power control for He joined the Department of Electrical and Computer Engineering, Uni-
successive interference cancellation with imperfect cancellation,” in versity of Texas, Austin, in the summer of 2002 as an Assistant Professor
Proc. IEEE Int. Conf. Communications, New York, NY, Apr. 2002, pp. in the Wireless Networking and Communications Group (WNCG). From
356–360. 1995 to 1997, he was an Engineer at Qualcomm, San Diego, CA, developing
[2] S. Verdu, “Minimum probability of error for asynchronous Gaussian satellite-based CDMA systems. He has served as a frequent Consultant on com-
multiple-access channels,” IEEE Trans. Inform. Theory, vol. IT-32, pp. munication systems to numerous clients in both industry and the government,
85–96, Jan. 1986. including Microsoft, Ricoh, Telogy, and NASA. His research interests are in
[3] M. L. Honig and H. V. Poor, “Adaptive interference suppres- all areas of wireless systems, and are currently focused on multiuser wireless
sion,” in Wireless Communications: Signal Processing Perspec- communications and interference management in unlicensed spectrum.
tives. Englewood Cliffs, NJ: Prentice-Hall, 1998, ch. 2.
[4] P. Hatrack and J. M. Holtzman, “Reduction of other-cell interference
with integrated interference cancellation/power control,” in Proc. 47th
IEEE Vehicular Technology Conf., vol. 3, 1997, pp. 1842–1846.
[5] R. Muller and J. Huber, “Capacity of cellular CDMA systems applying Teresa H. Meng (S’82–M’83–SM’93–F’99) received the Ph.D. degree in
interference cancellation and channel coding,” in Proc. IEEE Global electrical engineering and computer science from the University of California,
Telecommunications Conf., Nov. 1997, pp. 179–184. Berkeley, in 1988.
[6] A. J. Viterbi, CDMA—Principles of Spread Spectrum Communica- She joined the faculty of the Electrical Engineering Department, Stanford
tion. Reading, MA: Addison-Wesley, 1995. University, Stanford, CA, in 1988, where she is now a Professor and the Robert
[7] P. Frenger, P. Orten, and T. Ottosson, “Code-spread CDMA using Bosch Faculty Fellow. In 1998, she took a leave from Stanford University and
maximum free distance low-rate convolutional codes,” IEEE Trans. founded Atheros Communications Inc., Sunnyvale, CA, which provides the core
Commun., vol. 48, pp. 135–144, Jan. 2000. technology for ubiquitous, high-performance wireless communications. She re-
[8] G. L. Stuber, Principles of Mobile Communication. Reading, MA: Ad- turned to Stanford University in 2000 to continue her research and teaching at
dison-Wesley, 1995. the University. Her current research activities include low-power circuit design,
[9] D. Warrier and U. Madhow, “On the capacity of cellular CDMA with wireless communication, and bio-signal processing. She has given many ple-
successive decoding and controlled power disparities,” in Proc. 48th nary talks at major conferences in the areas of signal processing and wireless
IEEE Vehicular Technology Conf., vol. 3, May 1998, pp. 1873–1877. communications. She is the author of one book, numerous book chapters, and
[10] A. Agrawal, J. Andrews, J. Cioffi, and T. Meng, “Iterative power con- over 200 technical articles in journals and conferences.
trol with successive intereference cancellation for DS-CDMA systems,” Dr. Meng’s received Awards and honors for her research work at Stanford
IEEE Trans. Commun., to be published. University which include an NSF Presidential Young Investigator Award, an
[11] R. D. Yates, “A framework for uplink power control in cellular radio ONR Young Investigator Award, an IBM Faculty Development Award, a Best
systems,” IEEE J. Select. Areas Commun., vol. 13, pp. 1341–1347, Sept. Paper Award from the IEEE Signal Processing Society, the Eli Jury Award from
1995. U.C. Berkeley, and awards from AT&T, TRW, and other industry and academic
[12] S. Ariyavisitakul and L. F. Chang, “Signal and interference statistics of organizations. She was named one of the Top 10 Entrepreneurs in 2001 by Red
a CDMA system with feedpack power control,” IEEE Trans. Commun., Herring and was one of the three finalists for the Innovator of the Year Award
vol. 41, pp. 1626–1634, Nov. 1993. in 2000.