1274

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 3, MARCH 2011

To Decode the Interference or to Consider It as Noise

Abolfazl S. Motahari, Student Member, IEEE, and Amir K. Khandani, Member, IEEE
AbstractIn this paper, the impact of noncoordinated interfering signals on a point to point communication is addressed. While the transmitter has no information about the other users messages, the receiver has full knowledge of the codebooks of the interfering users and can potentially decode some part of the interference. A simple coding strategy is proposed for this channel. Assuming its own data is decoded successfully, the receiver partitions the set of interfering users into two disjoint subsets, namely the set of decodable users and the set of nondecodable users. Then the transmitters rate is chosen such that the intended signal can be jointly decoded with the set of decodable users. It is proved that the proposed strategy achieves the capacity of the additive Gaussian channel with Gaussian interfering users. A polynomial time algorithm is proposed to compute the achievable rate of the scheme. This algorithm is based a subroutine which separates the set of interfering users into decodable and nondecodable users in polynomial time. The proposed scheme is also applied to the case -user interference channel and some achievable points are of characterized by successive maximization of users rates.

Index TermsCapacity region, combinatorial optimization, interference channel, submodular functions.

I. INTRODUCTION

NTERFERENCE has structure as it is caused by data transmission of active users in a network. Treating interference as noise ignores the structure of interference which results in inefcient usage of available resources in most cases. Decoding some part of interference may increase the throughput of a network. For some networks such as the strong two-user Gaussian Interference Channel (IC), cf. [1], [2], the capacity region is attained by decoding the interference at receivers. The objective of this article is neither to present methods to design new networks, nor to increase the throughout of existing networks. The work is motivated by the existence of decentralized legacy networks which are based on single layer Gaussian code-books without power allocation. As technology evolves, some more advanced nodes will be added to the network which have capabilities like interference cancelation. This paper aims to nd ways such that these more advanced nodes operate in the best possible way in conjunction with the legacy nodes. Issues of this nature, including backward compatibility is a major effort in any standardization process. Focusing on a particular

Manuscript received November 12, 2007; revised July 16, 2010; accepted September 27, 2010. Date of current version February 18, 2011. The material in this paper was presented at the IEEE International Symposium on Information Theory, Nice, France, June 2007. A. S. Motahari is with the Department of Electrical Engineering and Computer Science, University of California, Berkeley, Berkeley, CA 94720 USA (e-mail: motahari@berkeley.edu). A. K. Khandani is with the Coding and Signal Transmission Laboratory, University of Waterloo, Waterloo, Ontario, Canada N2L 3G1 (e-mail: khandani@cst.uwaterloo.ca). Communicated by H. Yamamoto, Associate Editor for Shannon Theory. Digital Object Identier 10.1109/TIT.2010.2103756

transmit-receive pair in these systems, one obtains a point to point channel where the received signal incurs interference from a number of interfering users. The problem of nding an achievable rate for this channel in an efcient manner is the subject of this paper. In fact, the receiver can look for some part of interference that can be decoded which in turn allows higher data transmission in the channel. If all users use single codebooks for data transmission, then decoding a part of interference corresponds to decoding of the messages of a set of users. This in turn means that the signal transmitted from any user is either decoded or considered as noise. Therefore, the main task of the receiver is to partition the set of interfering users into decodable and nondecodable users. We propose the following method to obtain an achievable rate for the channel. Assuming its own data is decoded successfully, the receiver nds the maximum decodable subset of interfering users. By a maximum decodable subset, we mean a set of users that are decodable at the receiver, regarding the rest as noise and any decodable set is a proper subset of it. It is shown that this task can be accomplished by using a polynomial time algorithm. Once the receiver obtains the maximum decodable subset, it can partition the interfering users into decodable and nondecodable users. Then, the transmitters rate is chosen such that the intended signal can be jointly decoded with the set of decodable users. We also propose a polynomial time algorithm to nd the maximum achievable rate obtainable by this method. To show the strength of this method, we prove that for the additive Gaussian channel with Gaussian interfering users, the Gaussian distribution is optimal and the achievable rate is the capacity of this channel. As an application, we use this algorithm to characterize some achievable points for the -user Gaussian IC. There are several works related to this paper. Considering interference as noise in an -user Gaussian IC is studied in the literature [4][12]. For example, in [5] the -user Gaussian IC is studied where transmitters are allowed to allocate different powers in different bandwidths and receivers treat interference as noise. Recently, successive interference cancelation is also considered, cf. [13][16]. For instance, in [13] the optimal order of decoding at all receivers that maximizes the minimum rate among all users is obtained. The organization of this paper is as follows. In Section II, we introduce the system model and some background materials. In Section III, we consider a discrete memoryless channel contransmitters and one receiver. We assume that the sisting of users rate vector is not necessarily inside the capacity region of the Multiple Access Channel (MAC) seen at the receiver side which results in failure of the receiver to reliable decoding of all data streams. The receivers task, however, is to nd a maximum decodable subset of transmitters so that their data can be decoded from the received signal. We propose a polynomial-time algorithm which nds the maximum decodable subset of users.

0018-9448/$26.00 2011 IEEE

MOTAHARI AND KHANDANI: DECODE THE INTERFERENCE

1275

We also consider the continuous Gaussian case modeled by (1) and denote transmitted and received symbols, rewhere , is the input symbol correspectively. sponding to the th interfering user that uses a single Gaussian and rate . is the additive white codebook with power Gaussian noise with variance . The transmitter is subject to and tries to send data at the maximum the average power . rate B. Submodular Functions Denition 1: Let be a nite nonempty set. A function is called a submodular function if it satises (2) . A function is called suppermodular if for any is submodular. A modular function is a function which is both submodular and supermodular. Submodular functions are one of the most important objects in discrete optimization. In fact, they play the same role in discrete optimization as convex functions do in the continuous case [17]. In particular, minimization of submodular functions is of great interest to the eld of discrete optimization. Besides having a polynomial-time algorithm based on the ellipsoid method [18], there are combinatorial algorithms for minimizing submodular functions in strongly polynomial time, cf. [17] and [19]. If a submodular function is nondecreasing, i.e., if , and , then the associated polyhedron (3) is a polymatroid. Likewise, if a supermodular function is non, then the associated polyhedron decreasing and (4) is a contra-polymatroid. C. Properties of Mutual Information for Independent Random Variables independent random variables . Moreover, let denote the set of random variables indices. For any random variable , we have the following properties. 1) Chain Rule: For any disjoint subsets and , we have the following: Consider (5) 2) Independent Conditioning Inequality: For any disjoint subsets and , the following inequality holds: (6)

Fig. 1. Single user in an interfering medium. x is the input letter from the intended user and x ; i ; ; ; M , is the input symbol corresponding to the ith interfering user.

= 2 3 ...

In Section IV, we consider single-user data transmission over a channel with interfering users. We rst obtain a lower bound and an upper bound on the capacity of this channel. Then, we propose a method that characterizes an achievable rate for the channel. This achievable rate is a function of other users rates. We then prove that this function is piecewise linear. In Section V, we consider additive channels where the interference caused by other users is Gaussian. We prove that for this case, the Gaussian codebook achieves the capacity where each interfering user is either decoded or treated as noise by the receiver. In Section VI, we investigate an application of the proposed algorithms to the -user Gaussian IC. We develop a polynomial time algorithm that characterizes points obtainable from successive maximization of users rates. In Section VII, we conclude the paper. Notation: Throughout this paper, we use the following notations. Vectors are represented by bold faced letters. Random variables and sets are denoted by capital letters where the difference is clear from the context. The difference, union, and inter, and section of two sets and are represented by , respectively. The complement of a subset is denoted . is always by . The cardinality of a set is denoted by the number of users in the system. We use to denote the set . denotes the power set of which is the collection of all subsets of . For any set and any vector , we use the compact notations and to denote and , respectively. In particular, . is the set of denotes a -dimensional Euclidean space. real numbers and II. PRELIMINARIES A. System Model We consider single-user data transmission over a channel with interfering users. is specied by the transition probability function where is the input letter to the channel from the th user and is the output letter received by the receiver, see Fig. 1. The set is the input letter from the of users indices is denoted by . intended user and for are input letters from interfering users. We assume that the interfering users transmit by using single data at the rate vector codebooks generated randomly from the joint probability dis. We are interested in tribution characterizing the capacity of this channel.

1276

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 3, MARCH 2011

3) Polymatroidal Property: In [20], it is shown that the set is submodular and nondefunction creasing, i.e.

is the transmitted symbol where is the received symbol, of user , and is the additive white Gaussian noise with zero mean and variance . User is also subject to an average power constraint . The capacity region of the -user Gaussian MAC can be stated as (16) where . III. MAXIMUM DECODABLE SUBSET In this section, we consider a discrete memoryless channel transmitters with input alphabet for the consisting of th transmitter and one receiver with output alphabet where each transmitter uses a single codebook for data transmission. This channel is specied by the transition probability where is the input function letter to the channel from the th transmitter and is the output letter received by the receiver. The random codebooks used for data transmission at the rate vector are generated by using the joint probfor random ability distribution . variables The rate vector may fall outside of the capacity region of the MAC seen at the receiver side which results in failure to reliable decoding of all data streams. The receivers task, however, is to nd a decodable subset of transmitters so that their data can be decoded from the received signal. To this end, the receiver partitions the set of transmitters into two disjoint parts and tries to jointly decode the data sent by the transmitters within the rst partition, while considering the signals of transmitters in the second partition as noise. In what follows, we compute the complexity of nding a decodable subset of transmitters by an exhaustive search. Let denote the set of transmitters indices. ways to partition into two subsets; and to verify There are that a subset with cardinality is decodable, inequalities must be checked due to (14). Hence, in general, the total number of inequalities to be veried is

(7) Hence, its associated polyhedron is a polymatroid. 4) Contra-Polymatroidal Property: We claim that the set function dened as is a supermodular function. To this end, x any arbitrarily subsets and . Let . From the chain rule, we have (8) which can equivalently be written as (9) From Independent Conditioning Property, we . Hence have (10) Adding to both sides, we obtain

(11) Since , we have (12) as claimed. It is easy to show that is nondecreasing and hence its associated polyhedron is a contra-polymatroid. D. Multiple Access Capacity Region The capacity region of the MAC [21], [22] can be represented as follows, cf., [21] and [22]. We dene as the collection of all probability distributions which can be written as , where is the channel transition probability function. Now, the capacity region of a MAC is (13) where dened as denotes convex hull operation, and is

which is exponential in the number of users. Denition 2 (Maximum Decodable Subset): A set of transmitters is a maximal decodable subset if all transmitters in the subset are jointly decodable by the receiver, and is not a proper subset of any other decodable subset. If the maximal decodable subset is unique, we call it the maximum decodable subset. Lemma 1: For any channel, there is a maximum decodable subset. Proof: Suppose the receiver is able to decode two subsets of transmitters, namely and , such that none of them is a subset of the other. and are proper subsets of their union . Besides, their union is decodable by the receiver. This contradicts the fact that both subsets are maximal.

(14) Using the polymatroidal property of the mutual information, it is easy to show that is a polymatroid. It is worth is the union of polymatroids, it is noting that even though not necessarily a polymatroid. However, is a polymatroid for the -user Gaussian MAC modeled by (15)

MOTAHARI AND KHANDANI: DECODE THE INTERFERENCE

1277

We rst describe some properties of the maximum decodable subset. There are two cases of special interest. The rst case occurs when all transmitters are decodable by the receiver, i.e., the maximum decodable subset is the set . In this case, the transmitters rates must satisfy the inequalities given in (14). In the second case, however, none of the transmitters is decodable by the receiver, i.e., the maximum decodable subset is empty. The following Lemma shows that for the second case the rate vector must be in a certain contra-polymatroid. Lemma 2: None of the signals is decodable by the receiver if and only if transmitters rates satisfy (17) Moreover, the region of the rate vectors satisfying above inequalities forms a contra-polymatroid. Proof: We rst prove that if a rate vector satises (17), then none of the signals are decodable. To this end, we assume . Since that is the maximum decodable subset and is a decodable subset, we have the following constraints on the rates of the members of . (18) By substituting in the above equation, we have (19) which is a contradiction and this completes the if part of the proof. Next, we need to prove that if the inequalities in (17) are not satised, there is at least one transmitter which is decodable. Suppose there are some subsets that do not satisfy (17). Assume has the minimum cardinality among all and satises (20) If , then the transmitter in is decodable by considering everything else as noise which is the desired result. Hence, . If all members of are jointly decodable, we assume then we have found a decodable subset. Otherwise, there must be a subset of , say , satisfying (21) By decomposing the mutual information in (20), we obtain (22) From the minimality of , we have (23) By combining the two inequalities (21) and (23) and considering the fact that , we conclude that (24) (25) (28) is a polyhedron because it is the intersection of nitely consists of all rate vecmany half spaces. By Theorem 1, tors with the same maximum decodable subset . Since for any rate vector there is an associated maximum decodable subset, . This means that is represented we have as the union of nitely many polyhedral sets. An example for the case of the additive two-user Gaussian channel is given in Fig. 2.
Fig. 2. Decision regions used for determining the maximum decodable subset , the receiver for a two-user additive Gaussian Channel. For any rate in can decode both signals. For rates in and , the receiver is able to decode transmitters 1 and 2, respectively. Finally, the receiver can decode neither . 1 nor 2 for any rate in

which is a contradiction. This completes the only if part of the proof. is a In Section II.C, it is shown that supermodular function and monotone, hence the region formed by rates satisfying (17) is a contra-polymatroid. In the following theorem, the characterization of the maximum decodable subset is presented. Theorem 1: A subset is a maximum decodable subset if and only if the transmitters rates satisfy the following inequalities (26) (27) Proof: Inequality (26) corresponds to the capacity region of the MAC for members of considering members of as noise. Hence, the members of are decodable iff the inequalities in (26) are satised. The set is a maximum decodable subset if no other transmitters in is decodable by the receiver. Now, by applying Lemma 2 and considering that all members of are decoded, we conclude that none of the transmitters in is decodable iff the inequalities in (27) are satised. This completes the proof. For a given maximum decodable subset as , we dene

1278

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 3, MARCH 2011

The result of this section can be directly extended to continuous channels. In the following example, we apply the result of Theorem 1 to a two-user additive Gaussian channel. Example 1: Consider a two-user additive Gaussian channel where the received signal can be written as . In this case, has four subsets, namely , and . By applying Theorem 1, we obtain the following conditions for the subsets of to be the maximum decodable subset. is the maximum decodable subset if and only if 1) , and . is the maximum decodable subset if and only if 2) and . 3) is the maximum decodable subset if and only if and . is the maximum decodable subset if and only if 4) , and . The set of conditions described above partitions into four regions, as illustrated in Fig. 2. It can be seen from the gure that is a polymatroid corresponding to the capacity region of is a contra-polymatroid according to a two-user MAC and Lemma 2. The above example shows that nding the maximum decodable subset is equivalent to nding the region where the transmitters rate vector belongs to. Since the number of regions grows exponentially with the number of transmitters, nding a polynomial-time algorithm for solving the problem is desired. as follows: To this end, we rst dene the function (29) where . dened in (29) is a submodular Substituting (34) into (33) and using we obtain

Fig. 3. E is the ground set. minimizer of f in (30).

S is the maximum decodable subset. W

. Since

is the

Hence and (26), we have

is a subset of , from (31)

The inclusion imply

and independence of random variables . Hence, (32)

From the denition of

in (29), we have (33)

From the chain rule and the fact that and two disjoint subsets, we have the following:

partition

into

(34) , (35) Using (32), we conclude that (36) If , then it contradicts the optimality of , and , then it contradicts the fact that has if minimum cardinality among all minimizers. This completes the proof. By applying Theorem 2 and using the well-known submodular function minimization algorithms as a subroutine, cf. [19] and [17], we propose the following polynomial-time algorithm to nd the maximum decodable subset. Algorithm 1 (Finding the Maximum Decodable Subset): . 1) Set such that 2) Find

Lemma 3: The function function.

Proof: The result directly follows from the modularity of and the submodularity of . Since there are polynomial-time algorithms for minimizing any submodular functions, cf., [17] and [19], the following optimization problem can be solved efciently: (30) If the minimum of in (30) is zero, then all transmitters are decodable by the receiver due to (14). Otherwise, there is at which is not decodable. least one transmitter from the set In the following theorem, we prove that indeed all members of the minimizer of are not decodable, and they need to be considered as noise. Theorem 2: No member of the subset that minimizes in (30) is decodable by the receiver, provided that the minimum in (30) is not zero and the minimum cardinal minimizer is used. In must be considered as noise, i.e., if is fact, all members of . the maximum decodable subset then Proof: We rst partition the minimizer subset into two and , see disjoint sets and where Fig. 3. We need to show that . Suppose is nonempty.

where

is (37)

3) If STOP. Otherwise,

is the maximal decodable subset. .

MOTAHARI AND KHANDANI: DECODE THE INTERFERENCE

1279

4) If STOP. No subset of GO TO step 2.

is decodable. Otherwise,

Theorem 3: Algorithm 1 converges to the maximum decodable subset in polynomial time. is a nonempty set (otherProof: Since in each iteration wise, the algorithm stops), this algorithm converges at most in iterations. Furthermore, in each iteration, we need to minimize a submodular function which can be done in polynomial time [17]. Hence, the total running time of the algorithm is polynomial in time. IV. AN ACHIEVABLE RATE In this section, we propose a method to obtain an achievable rate for the channel . We also provide a polynomial time algorithm to characterize this achievable rate. A lower bound for the capacity of can be obtained by considering interfering users in as noise and optimizing over all input distributions. Hence, we have (38) where denotes the capacity of . Now, we assume that regardless of the input distribution, the receiver is able to decode all interfering users considering its own signal as noise. By this assumption, an upper bound on the capacity can be obtained as follows: (39) Let us assume that the transmitter uses to generate a single random codebook. We need to nd the maximum achiev. If is an achievable rate, then the receiver can able rate successfully decode its intended data. After decoding its own for the maximum signal, the receiver can search in the set . This procedure can be done efdecodable subset . ciently using Algorithm 1. Let us dene is the set of users that the receiver treats as noise. From (26), we have (40) To nd , we consider the MAC consisting of user 1 and the users in , while the users in are considered as noise. From (14), the rate vector is achievable if (41) Since half of the inequalities in (41) are satised by (40) and the only unknown parameter is , we can maximize the users rate based on the following optimization problem:

Algorithm 2 (Finding an Achievable Rate): , nd the maximum decodable subset among 1) Given interfering users by using Algorithm 1 and assuming that the users data is decoded. 2) Solve the submodular function minimization in (42). As a byproduct of the above algorithm, we can nd the subset of interfering users that can be rst decoded at the receiver and its effect can be removed. Proposition 1: If minimizes (42), then the receiver is capable of decoding all users in by considering everything else as noise. Proof: At the rst step, one needs to decode quires . This re(43) Suppose there is a subset that does not satisfy (43), that is (44) Hence

(45) where (a) follows from the chain rule and the fact that and , (b) follows from the denition of and minimality of , and (c) follows form (44). The last inequality contradicts the fact that is the solution for the minimization problem in (42). This completes the proof. In light of Proposition 1, the set is decomposable into three , and . is the compledisjoint subsets, namely ment of , namely the union of the maximum decodable subset and the intended user. Therefore, the receiver is not able to decode the interfering users in and considers them as is the part of that the receiver can decode by connoise. sidering everything else as noise. is the subset of users that need to be decoded jointly after removing the effect of . As indicated in (42), the achievable rate is a function of interfering users rates. In order to derive some properties of this function, we need the following denition. Denition 3 (Piecewise Linear Functions [23]): A function is piecewise linear if rstly its domain can be represented as the union of nitely many polyhedral sets, and secondly is afne within each polyhedral set, i.e., for some vector and scalar . In the following theorem, we summarize some properties of as a function of . Theorem 4: The function wise linear. More precisely, afne functions. dened in (42) is piececonsists of at most

(42) The optimization problem (42) is again a submodular function minimization and can be solved by polynomial-time algorithms. In the following, we summarize the above procedure.

1280

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 3, MARCH 2011

sequences with show that

. In this case, it is easy to

(48) where if fact that the function components. and 0 otherwise. This reects the may have less than

V. CHANNELS CAPACITY FOR THE GAUSSIAN CASE In this section, we prove that provided using Gaussian distribution for random codebook generation, the achievable rate obtained in the previous section is indeed the maximum rate that one can hope for in a decentralized system. In other words, Gaussian codes are saddle points for a noncoordinated interference channel system even if one allows decoding the interference at the receivers. (the output of Algorithm To show that any rate above is Gaussian) is not achievable, we construct a 2 where degraded broadcast channel and show that if a rate is achievable, then one can communicate reliably outside the capacity region of this channel which is a contradiction. The following lemma assists us in constructing such a degraded channel. Lemma 4: For any set of independent Gaussian codebooks and rate vector with power vector , there is a -user Gaussian broadcast channel with the following properties: . 1) The transmitters total power is . 2) There are noise levels: 3) Users are partitioned into disjoint subsets, that is, . All users in have the same noise level , for . 4) The rate vector lies on the boundary of the capacity is achievable using Gaussian codebooks with region. powers in one to one correspondence with the components of . Proof: We aim at building a Gaussian broadcast channel with as input and as outputs, where and is additive white Gaussian noise with variance . To this end, assume there is a -user Gaussian MAC with noise level and transmit power vector . Then, is achievable if (49) By monotonicity of , it is always possible to nd an such is achievable. Indeed, is achievable that the rate vector for any , where corresponds to the case that at least one of the inequalities in for any noise level above denote the set of users for which (49) turns to equality. Let the corresponding inequality in (49) turns to equality with noise (in case of having more than one equality we choose level the maximum cardinal subset), i.e., (50)

Fig. 4. The function R (

) for a channel with two interfering users.

Proof: Likewise (28), let us dene the region

as

(46) where . Due to (42), the function is afne functions over dened as the pointwise minimum of . As a result, is piecewise linear, the polyhedral set , cf., [23, Theorem 2.49]. continuous, and concave over is a piecewise linear function over each Since and , it is a piecewise linear function over . Moreover, each polyhedron is divided into at most subpolyhedra in each of which is an afne function. Hence, the total number of components is not more than (47) This completes the proof. It is worth noting that, although is a concave function over , it is not a concave function over . each Example 2: Consider an additive channel where all users use Gaussian codebooks for data transmission. In this case, the maximum decodable subset of in. Hence, there are four reterfering users is a subset of , and where is a congions cave function over each of them. For instance, over and over where is either or 0. In Fig. 4, an example for this channel is illustrated. As deof the function picted in the gure, is a piecewise linear and continuous function. It also consists of 9 components, i.e., for . Example 3: In this example, we consider a binary adder channel with interfering users. The channel model can . We further assume be written as that users codebooks are randomly chosen from Bernouli

MOTAHARI AND KHANDANI: DECODE THE INTERFERENCE

1281

By plugging in (49), we have (51) to the output of the Gaussian Now, we correspond users in channel , where is additive Gaussian noise with . variance We can apply the same procedure to (51), i.e., we increase until one of the inequalities turns to equality. Let denote the maximum noise level satisfying (51) with equality. Clearly, . If denotes the set of users satisfying (51) with equality, then we have (52) By plugging in (51), we obtain

fully characterized and there is an explicit expression for the boundary points, cf. [24]. The equalities in (54) guarantee that lies on the boundary of the capacity region. the rate vector This completes the proof. Theorem 5: The rate , the output of Algorithm 2, is the capacity of the channel described in (1). Proof: We rewrite the achievable rate given in Algorithm 2 by using the Gaussian distribution as the codebook generator. As discussed earlier, the set of users can be partitioned into three , and . subsets is the subset of interfering users that the receiver can decode considering everything else as noise. We remove users in from the active users. This operation only increases the capacity region (one can also argue that the Gaussian noise is the is worst noise for additive channels, cf. [25] and [26], and decodable when other users are considered as Gaussian noise, is decodable for any arbitrary distribution for intended user). is the complement of the maximum decodable subset and must be considered as noise. From (27), we have (57) is the solution to the minimization problem in (42). Hence, we have (58)

(53) to the output of the Gaussian Now, we correspond users in , where is additive Gaussian noise with channel variance . By repeating the above procedure, we can construct a set of and associate channels with noise levels with such that set of users (54)

(55)

We apply Lemma 4 to the set of users in with associated and rate vector . Let power vector denote the noise levels and denote the collection of subsets of users associated to each level of noise for the Gaussian broadcast channel with the properties given in . To verify this, we substitute Lemma 4. We claim that into (54) and (57). Hence, we obtain (59) which results in . as a set of new users to the Next, we add and Gaussian broadcast channel with noise level . It is easy increase the transmitters total power by to verify that the conditions in (54), (55), and (56) are still satised with new broadcast channel. Consequently, the rate vector lies on the boundary of the capacity region. Therefore, reliable results in reliable data data transmission at any rate above transmission outside the capacity region which is a contradiction. This completes the proof. VI. APPLICATION FOR THE -USER GAUSSIAN IC

(56) Now, assume that the transmitter with total power uses -level Gaussian codebooks for data broadcasting. The transmitted signal can be written as , where is a Gaussian codeword with power and rate and contains information for th user. The received signal at noise level can be written as . The set of inequalities in (56) implies that all users at noise level can decode data streams considering everything else as noise. By of users in from the received removing the effect of users in signal, the set of inequalities in (55) implies that all users in can decode their own signal considering users in as noise. In other words, all users at the same level of noise can decode their signals by rst decoding the users at upper levels and removing their effect and considering users at lower levels as Gaussian noise. Hence, we obtain a Gaussian broadcast channel in which the rate vector is achievable and Gaussian codebooks are constructed according to the power vector . It remains to show that is on the boundary of the capacity region. The capacity region of the Gaussian broadcast channel is

In this section, we apply the proposed algorithms to the -user Gaussian IC modeled by (60)

1282

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 3, MARCH 2011

where is the transmitted symbol of user and denotes the links gain between the th transmitter and the th receiver. is additive white Gaussian noise with zero mean and variance . User is also subject to an average power constraint . . The capacity region of this channel is denoted by Let us assume each transmitter is allowed to transmit data by using a single Gaussian codebook and each receiver is allowed to decode any subset of interfering users. Let denote the set of decoding strategies. By a decoding strategy , we mean that the receiver tries to deand . Since code all users data in . Clearly, possible choices for each , we have there are possible strategies in total. Hence, . Given a strategy, a rate vector is achievable with respect to that strategy if every receiver can reliably decode its associated can be dened users. Therefore, an achievable rate region as a set of all rate vectors that are achievable with respect to the . Clearly, and it strategy . Let is not convex in general. can be shown that Given an ordering of users, we aim at maximizing users orderings rates in accordance with . In general, there are not necessarily distinct achievable of users which result in rates in the capacity region. Due to the polymatroidal property of the capacity region of the Gaussian MAC, every permutais not tion leads to a distinct achievable rate vector; whereas, a polymatroid and hence there may be some permutations that lead to the same achievable rate vector. Without loss of generality, we may assume the order is the same as that of users indices, i.e., permutation matrix is identity. Setting the rst users rate to its maximum value imposes some constraints on the other users rates as they must is achievbe decoded by the rst receiver. The reason is that able if the rst receiver can decode all the interfering users by considering its own signal as noise and eliminating their effects from the received signal. Maximization of the second users rate is more delicate, since its transmission should not affect the rst users data rate. However, we have the choice of lowering other users rates as much are as needed. Hence, we assume users in the set decoded at the rst and second receivers and their effects are remust be chosen such that both receivers can decode moved. by it. The maximum decodable subset at the rst receiver is the assumption. For the second user, we can nd the maximum decodable subset of interfering users which in this case is either or . Now, we can run Algorithm 2 at both receivers to nd an achievable rate for each receiver. Clearly, the minimum of to this the two achievable rates are achievable and we set in value. Besides, we obtain the strategy which and are achievable, where and . To maximize the rate of user , we proceed as follows. We treat users above index as they do not exist, i.e., we put constraints on their rates in such a way that all the receivers with can decode them rst and remove their effects. indices in From maximization of users rates in the previous steps, we have and its corresponding achievable strategy , where . must be chosen such that all receivers in can decode

it. The maximum decodable subset of interfering users is given for all receivers in . We can also nd the maxby imum decodable subset of interfering users at receiver by running Algorithm 1. Let denote this subset. From (42), is achievable at the receiver , if it is less than which is dened as

(61) Hence, can be chosen as the minimum of all s. For the next step, we need a new achievable strategy. It is easy to see is that the proper strategy at step . Now, we can iterate until the last user. Algorithm 3 (Successive Maximization of Users Rates): 1) Set and . , do: 2) For a) Find the maximum decodable subset of interfering users in the set for receiver assuming are decoded and their that users in the set effects are removed. b) Solve the following optimization problem (62) where is dened in (61). , for all . c) For the sake of completeness, in the following theorem, we state that the above algorithm nishes in polynomial time. Theorem 6: Algorithm 3 converges to an extreme point of in polynomial time. Proof: At the th iteration, we need to solve submodular function minimizations. Hence, in total, a submodular function minimization subroutine is invoked times. Moreover, at each step, we need to nd the maximum decodable subset which can be accomplished in polynomial time based on Theorem 1. Hence, Algorithm 3 converges to an extreme point in polynomial time. of It is worth noting that for the two-user Gaussian IC in the case of strong and very strong interference [28], the output of Algorithm 3 is a point on the boundary of the capacity region. Recently, it is shown that for the mixed interference case the algorithm, depending on the ordering of users, outputs one of the corner points of the capacity region, [30], [31]. In the case of weak interference, however, the output of Algorithm 3 coincides with Costas result in [29]. However, it was pointed in [28] that the proof in [29] for the corner points was awed (see [28, item 6 in the concluding remarks (Section 6)]). VII. CONCLUSION We investigated data transmission over a channel with interfering users. By establishing certain properties of the maximum decodable subset, we proposed a polynomial time algorithm that separates the interfering users into two

MOTAHARI AND KHANDANI: DECODE THE INTERFERENCE

1283

disjoint parts: the users that the receiver is able to jointly decode their messages and its complement. We introduced an optimization problem that gives an achievable rate for this channel. We proposed a polynomial time algorithm for solving this optimization problem. We also established the capacity of the additive Gaussian channel with Gaussian interfering users and showed that the Gaussian distribution is optimal and the proposed achievable rate is the capacity of this channel. As an application of this method, we investigated data transmission for the case of -user interference channel when transmitters use single codebooks for data transmission, and receivers are allowed to decode other users messages. We then . We obtained some introduced an achievable rate region by successive maximization of users extreme points of rates. REFERENCES
[1] H. Sato, The capacity of the Gaussian interference channel under strong interference, IEEE Trans. Inf. Theory, vol. IT-27, pp. 786788, Nov. 1981. [2] A. B. Carleial, A case where interference does not reduce capacity, IEEE Trans. Inf. Theory, vol. IT-21, pp. 569570, Sep. 1975. [3] T. S. Han and K. Kobayashi, A new achievable rate region for the interference channel, IEEE Trans. Inf. Theory, vol. IT-27, pp. 4960, Jan. 1981. [4] P. Gupta and P. R. Kumar, The capacity of wireless networks, IEEE Trans. Inf. Theory, vol. 46, no. 2, pp. 388404, Mar. 2000. [5] R. Etkin, A. Parekh, and D. Tse, Spectrum sharing for unlicensed bands, IEEE J. Sel. Areas Commun., vol. 52, pp. 18131827, Apr. 2007. [6] S. Webere, J. G. Andrews, and N. Jindal, The effect of fading, channel inversion, and threshold scheduling on ad hoc networks, IEEE Trans. Inf. Theory, vol. 43, no. 11, pp. 4127 4149 , Nov. 2007. [7] M. Ebrahimi, M. A. Maddah-Ali, and A. K. Khandani, Throughput scaling laws for wireless networks with fading channels, IEEE Trans. Inf. Theory, vol. 53, no. 11, pp. 42504254, Nov. 2007. [8] D. Gesbert and M. Kountouris, Joint power control and user scheduling in multicell wireless networks: Capacity scaling laws, IEEE Trans. Inf. Theory, to be published. [9] O. Lvque and E. Telatar, Information theoretic upper bounds on the capacity of large extended ad hoc wireless networks, IEEE Trans. Inf. Theory, vol. 51, no. 3, pp. 858865, Mar. 2005. [10] L. A. Imhof and R. Mathar, Capacity regions and optimal power allocation for CDMA cellular radio, IEEE Trans. Inf. Theory, vol. 51, no. 6, pp. 20112019, June 2005. [11] H. Boche and S. Stanczak, Convexity of some feasible QoS regions and asymptotic behavior of the minimum total power in CDMA systems, IEEE Trans. Commun., vol. 52, no. 12, pp. 21902197, Dec. 2004. [12] H. Boche and S. Stanczak, Iterative algorithm for nding optimal resource allocations in symbol-asynchronous CDMA channels with different SIR requirements, in Ptoc. 36th Asilomar Conf. Signals, Syst. Comput., Nov. 2002, vol. 2, no. 36, pp. 19091913. [13] M. A. Maddah-Ali, H. Mahdavi-Doost, and A. K. Khandani, Optimal order of decoding for max-min fairness in k -user memoryless interference channels, Jan. 2007 [Online]. Available: http://arxiv.org/ abs/cs.it/0701118v1, Available at: [14] S. P. Weber, J. G. Andrews, X. Yang, and G. de Veciana, Transmission capacity of wireless ad hoc networks with successive interference cancellation, IEEE Trans. Inf. Theory, vol. 53, no. 8, pp. 27992814, Aug. 2007. [15] N. Prasad and X. Wang, Outage minimization and rate allocation for the multiuser Gaussian interference channels with successive group decoding, IEEE Trans. Inf. Theory, vol. 55, pp. 55405557, Dec. 2009.

[16] A. Tajer, N. Prasad, and X. Wang, Outage minimization and rate allocation for the multiuser Gaussian interference channels with successive group decoding, Aug. 2009 [Online]. Available: http://arxiv.org/abs/ 0908.1077 [17] A. Schrijver, A combinatorial algorithm minimizing sub-modular functions in strongly polynomial time, J. Combinator. Theory, vol. 80, pp. 346355, 2000. [18] M. Grtschel, L. Lovsz, and A. Schrijver, Geometric Algorithms and Combinatorial Optimization. Berlin, Germany: Springer-Verlag, 1988. [19] S. Iwata, L. Fleischer, and S. Fujishige, A combinatorial strongly polynomial algorithm for minimizing sub-modular functions, J. ACM, vol. 48, pp. 761777, 2001. [20] T. S. Han, The capacity region of general multiple-access channel with certain correlated sources., Inf. Control, vol. 40, pp. 3760, Jan. 1979. [21] H. Liao, Multiple access channels, Ph.D. dissertation, Dep. Elec. Eng., Univ. Hawaii, Honolulu, 1972. [22] R. Ahlswede, The capacity region of a channel with two senders and two receivers, Annals Probabil., vol. 2, no. 5, pp. 805814, 1974. [23] R. T. Rockafellar and R. J.-B. Wets, Variational Analysis. Berlin Heidelberg, Germany: Springer-Verlag, 1998. [24] T. Cover and J. Thomas, Elements of Information Theory. New York: Wiley, 1991. [25] S. Ihara, On the capacity of channels with additive non-Gaussian noise., Inf. Control, vol. 37, no. 1, pp. 3439, Apr. 1978. [26] S. N. Diggavi and T. Cover, The worst additive noise under a covariance constraint, IEEE Trans. Inf. Theory, vol. 47, no. 7, pp. 30723081, Nov. 2001. [27] A. B. Carleial, Interference channels, IEEE Trans. Inf. Theory, vol. IT-24, pp. 6070, Jan. 1978. [28] I. Sason, On achievable rate regions for the Gaussian interference channel, IEEE Trans. Inf. Theory, vol. 50, pp. 13451356, Jun. 2004. [29] M. H. M. Costa, On the Gaussian interference channel, IEEE Trans. Inf. Theory, vol. IT-31, pp. 607615, Sep. 1985. [30] A. S. Motahari and A. K. Khandani, Capacity bounds for the Gaussian interference channel, IEEE Trans. Inf. Theory, vol. IT-55, pp. 620643, Feb. 2009. [31] X. Shang, G. Kramer, and B. Chen, A new outer bound and the noisyinterference SumRate capacity for Gaussian interference channels, IEEE Trans. Inf. Theory, vol. IT-55, pp. 689699, Feb. 2009.

Abolfazl S. Motahari (S05) received the B.Sc. degree from the Iran University of Science and Technology (IUST), Tehran, in 1999, the M.Sc. degree from Sharif University of Technology, Tehran, in 2001, and the Ph.D. degree from University of Waterloo, Waterloo, Canada, in 2009, all in electrical engineering. From August 2000 to August 2001, he was a Research Scientist with the Advanced Communication Science Research Laboratory, Iran Telecommunication Research Center (ITRC), Tehran. From October 2009 to September 2010, he was a Postdoctoral Fellow with the University of Waterloo, Waterloo. Currently, he is a Postdoctoral Fellow with the Department of Electrical Engineering and Computer Science, University of California at Berkeley. His research interests include multiuser information theory, joint source-channel coding, spread-spectrum systems, and optical communication systems. Dr. Motahari received several awards including Natural Science and Engineering Research Council of Canada (NSERC) Post-Doctoral Fellowship.

Amir K. Khandani (A93M02) received the B.Sc. and M.A.Sc. degrees from the University of Tehran, Tehran, Iran, in 1985 and the Ph.D. degree from McGill University, Montreal, QC, Canada, in 1992. From 1985 to 1988, he was a Design Engineer with Iran Communication Industries, Ltd., where he worked on circuit design for telecommunication systems. From 1992 to 1993, he was a Research Associate with INRS Telecommunications (Quebec University), Montreal, Canada, and then joined the Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON, Canada, in 1993, where he is currently a Professor and holds an NSERC Industrial Research Chair and a Tier I Canada Research Chair.