Cellular Networks With Finite Precision CSIT: GDoF Optimality of Multi-Cell TIN and Extremal Gains of Multi-Cell Cooperation
CCellular Networks With Finite Precision CSIT: GDoF Optimalityof Multi-Cell TIN and Extremal Gains of Multi-Cell Cooperation
Hamdi Joudeh and Giuseppe Caire
Abstract
We study the generalized degrees-of-freedom (GDoF) of cellular networks under finite pre-cision channel state information at the transmitters (CSIT). We consider downlink settingsmodeled by the interfering broadcast channel (IBC) under no multi-cell cooperation, and theoverloaded multiple-input-single-output broadcast channel (MISO-BC) under full multi-cell co-operation. We focus on three regimes of interest: the mc-TIN regime, where a scheme based ontreating inter-cell interference as noise (mc-TIN) was shown to be GDoF optimal for the IBC;the mc-CTIN regime, where the GDoF region achievable by mc-TIN is convex without the needfor time-sharing; and the mc-SLS regime which extends a previously identified regime, wherea simple layered superposition (SLS) scheme is optimal for the 3-transmitter-3-user MISO-BC,to overloaded cellular-type networks with more users than transmitters. We first show that theoptimality of mc-TIN for the IBC extends to the entire mc-CTIN regime when CSIT is lim-ited to finite precision. The converse proof of this result relies on a new application of alignedimages bounds. We then extend the IBC converse proof to the counterpart overloaded MISO-BC, obtained by enabling full transmitter cooperation. This, in turn, is utilized to show thata multi-cell variant of the SLS scheme is optimal in the mc-SLS regime under full multi-cellcooperation, albeit only for 2-cell networks. The overwhelming combinatorial complexity of theGDoF region stands in the way of extending this result to larger networks. Alternatively, weappeal to extremal network analysis, recently introduced by Chan et al., and study the GDoFgain of multi-cell cooperation over mc-TIN in the three regimes of interest. We show that thisextremal GDoF gain is bounded by small constants in the mc-TIN and mc-CTIN regimes, yetscales logarithmically with the number of cells in the mc-SLS regime.
The authors are with the Faculty of Electrical Engineering and Computer Science, Technische Universit¨at Berlin,10587 Berlin, Germany (e-mail: [email protected]; [email protected]). This work was partially supported bythe European Research Council (ERC) under the ERC Advanced Grant N. 789190 (CARENET). This article waspresented in part at the 2020 IEEE International Symposium on Information Theory (ISIT) [1], and will be presentedin part at the 2020 IEEE Global Communications Conference (GLOBECOM) [2]. a r X i v : . [ c s . I T ] A ug Introduction
For over a decade, generalized degrees-of-freedom (GDoF) studies have continued to contributenew insights into the fundamental limits of wireless communication networks [3–10]. The GDoFis defined by proportionally scaling all individual link capacities in a network to infinity, whilenormalizing the total network capacity by the same scaling factor. From an analytic standpoint,this scaling eliminates intricacies related to additive noise, unavoidable in exact capacity analysis,and focuses the attention on aspects related to interference management. This is accomplishedwhile preserving the character of the original network, carried through to the GDoF limit via thealmost constant normalized link capacities (or channel strength parameters), rendering the GDoF amore comprehensive figure of merit compared to the more commonly employed degrees-of-freedom(DoF) metric. The validity of the GDoF metric as a good approximator is confirmed by severalconstant gap capacity characterizations, obtained from their GDoF counterparts (e.g. [3, 11, 12]).Amongst the various insights brought to light by GDoF studies, perhaps the most striking arethose related to the fundamental role of channel state information at the transmitters (CSIT) inwireless interference networks. Initial GDoF studies considered somewhat idealized settings, inwhich the availability of perfect CSIT is assumed; and gave rise to novel schemes that challengedconventional wisdom at the time, especially those based on interference alignment (IA) and the useof structured codes as opposed to random codes [12–14]. In practical wireless networks, however,uncertainty in channel knowledge is inevitable due to fading and limited feedback resources [15];and robustness against such uncertainty is of paramount importance to avoid potential catastrophicfailures. This has recently motivated a surge of interest in robust GDoF characterizations, obtainedby limiting channel knowledge at the transmitters to finite precision.While the study of wireless networks with finite precision CSIT is not new [16], progress inthis area was long hindered by the lack of tight information-theoretic outer bounds that matchthe best known inner bounds. This, however, has changed in recent years with a breakthrough byDavoodi and Jafar [17] known as the aligned images (AI) approach: a combinatorial argument forbounding the number of codewords that align at one receiver and remain resolvable at another.This approach has been instrumental in deriving tight GDoF outer bounds under finite precisionCSIT for a variety of interference and multi-antenna networks [7–10]. A common theme emergingfrom these robust characterizations is that GDoF gains due to sophisticated schemes that relyon precise CSIT—as signal-space and signal-scale IA [12, 13], zero forcing (ZF) and dirty papercoding (DPC) [18–20]—are entirely lost under finite precision CSIT. The new insights into thefragility of those once-promising elegant schemes have brought back to the forefront simple androbust alternatives that require only coarse channel knowledge—as power control and treatinginterference as noise (TIN) [21–30], rate-splitting and simple layered superposition (SLS) [31–34]—which, perhaps surprisingly, have turned out to be GDoF optimal in a number of settings.In the present paper, we extend the state-of-the-art in robust GDoF studies to cellular networkswith finite precision CSIT. A basic cellular setting operating in the downlink mode is modeled bya Gaussian interfering broadcast channel (IBC), comprising K mutually interfering Gaussian BCswith L users each. When all transmitters are allowed to cooperate (i.e. full multi-cell cooperation),the setting becomes an overloaded multiple-input-single-output BC (MISO-BC), with K transmitantennas and KL users. We will refer to the above settings as K × KL networks, where thenumber of cells K and the number of users per-cell L may be arbitrary. A major challenge inGDoF studies of large networks, including cellular networks, is the richness of the underlying This also includes early DoF results, which can be seen as a special case of GDoF results (see Remark 2). While the AI approach was initially introduced to bound the DoF under partial CSIT [17], it is now well-understood that DoF bounds under partial CSIT translate to GDoF bounds under finite precision CSIT [8, 9]. weak interference regimes. Prime examples are the fairly broad regimes in which simpleschemes based on TIN turn out to be GDoF optimal, which have been identified for a variety ofinterference networks [21–30]. This approach has also proven useful for more practical purposes—the inherent simplicity of TIN combined with its (approximate) information-theoretic optimalityin some regimes have inspired the design of new high-performing resource allocation algorithmsfor device-to-device (D2D) networks [24, 35, 36]. As far as cellular networks are concerned, it isoften the case in practice that a user equipment (UE) connects to the base station (BS) with thestrongest signal within its vicinity; and hence a direct link between a BS and an associated UEis typically no weaker than any of the inter-cell interference links to the same UE. This intrinsicproperty of cellular networks, combined with the goal of making progress on GDoF problems thatremain open in their generality, naturally motivate us to study weak inter-cell interference regimes.Against this background, we focus our attention on two main questions in this paper.
Q1. Robust GDoF Optimality of Multi-Cell TIN : The first question we set out to answer relatesto the GDoF optimality of treating inter-cell interference as noise (multi-cell TIN, or mc-TIN) forthe IBC with finite precision CSIT. In the mc-TIN scheme, single-cell transmissions are used—thisincludes power control, superposition coding and successive decoding in each cell—while all inter-cell interference is treated as additional Gaussian noise at the receivers [30]. This is arguably the simplest multi-cell scheme for which GDoF optimality results can be show in fairly broad parameterregimes, and it is also especially suitable for weak inter-cell interference. Two relevant weak inter-cell interference regimes have been identified in [30]: a mc-TIN regime, in which mc-TIN achievesthe entire GDoF region of the IBC; and a multi-cell convex TIN (mc-CTIN) regime, a strictly largerregime in which the GDoF region achieved through mc-TIN is a convex polyhedron without theneed for time-sharing. Note that the mc-TIN and mc-CTIN regimes respectively generalize theTIN regime of Geng et al. [21] and CTIN regime of Yi and Caire [24], identified for the K × K interference channel (IC), to K × KL cellular networks (see Section 3.1). Perfect CSIT is assumedin [30], and the optimality of mc-TIN remains unexplored under finite precision CSIT. Q2. Robust GDoF Gain of Multi-Cell Cooperation:
The second question relates to under-standing the extent to which multi-cell cooperation (mc-Co) is beneficial when CSIT is limited tofinite precision, especially in weak inter-cell interference regimes where the simple non-cooperativemc-TIN scheme is known (or expected) to be quite powerful. While cooperation in cellular net-works has been subject to intensive research for many years (see [37–39] and references therein),the fundamental limits and benefits of robust mc-Co under finite precision CSIT remain largelyunexplored, apart from GDoF results for K × K networks [8, 9, 33, 40]. These results, however, donot extend directly to overloaded K × KL networks, which better resemble cellular settings.In principle, the gain from mc-Co can be studied by comparing the GDoF of the K × KL IBC and Similar results also hold in uplink networks modeled by the interfering multiple access channel (IMAC) [29]. K × KL MISO-BC. Nevertheless, such direct comparisons are currentlyinfeasible in large asymmetric settings, mainly due to the difficulty of obtaining explicit GDoFcharacterizations—for instance, while it remains plausible that the simple layer superposition (SLS)scheme is GDoF optimal for the K × K MISO-BC in the so-called SLS regime, this has been shownonly for K ≤ K [9,33]. This hurdle hasbeen recently circumvented through extremal network analysis: a novel approach proposed by Chan et al. [40], where the focus is shifted towards studying a class of networks in a regime of interestwhich maximize the gain of one scheme over another. This approach was successfully applied to K × K networks with finite precision CSIT in [40] to characterize the maximum multiplicative GDoFgains from transmitter cooperation over TIN in three weak interference regimes of interest: the TIN,CTIN and SLS regimes. The prospect of leveraging extremal network analysis to understand thegain of mc-Co over mc-TIN in K × KL cellular networks is an intriguing one, and it is yet to beexplored. Next, we summarize the main findings and contributions of this work. In the first result of this paper (Theorem 1), we show that when CSIT is limited to finite precisionin the IBC, the GDoF optimality of mc-TIN extends to the entire mc-CTIN regime. This stands insharp contrast to the IBC with perfect CSIT, where IA-based schemes achieve strict GDoF gainsover mc-TIN in some parts of the mc-CTIN regime. The converse proof of Theorem 1 relies on anew application of AI bounds to the K × KL IBC, which departs from previous applications thatfocus on the K × K interference channel (IC) and its cooperative K × K MISO-BC counterpart[7–9, 33, 40]. In particular, the mixed BC-IC nature of the IBC, which captures both intra-cell andinter-cell interference in cellular networks, requires a careful selection of auxiliary random variablesto obtain a tight GDoF outer bound in this case (see Section 5).Theorem 1 generalizes a previous result for the K × K IC, obtained by Chan et al. in [40, Th.4.1], to the K × KL IBC; and further reveals key insights into cellular networks which cannot beseen from studying the IC alone. For instance, in the mc-CTIN regime, apart from the strongest user with the highest signal-to-interference ratio (SIR) in each cell, all weaker users turn out tobe redundant in the sense that they do not contribute to increasing the overall GDoF of the IBCwith finite precision CSIT. This again is in sharp contrast to the the IBC with perfect CSIT, whereweaker users are still very useful for increasing the overall GDoF through IA. This point and otherkey insights are discussed in greater detail further on in Section 4.1.Next, we consider the GDoF gain from mc-Co under finite precision CSIT, focusing on the mc-TIN and mc-CTIN regimes, and a strictly larger regime which we call the multi-cell SLS (mc-SLS)regime, which naturally extends the SLS regime of [33] to K × KL settings. To this end, we startby deriving an AI-based GDoF outer bound for the K × KL MISO-BC with finite precision CSIT(Theorem 2). We further show that this outer bound is achievable in the 2-cell case using a multi-cell variant of the SLS scheme, hence settling the GDoF region question for the 2 × L MISO-BCwith finite precision CSIT in the mc-SLS regime (Theorem 3). Remarkably, the general structureof the SLS scheme greatly simplifies in 2 × L cellular networks in the mc-SLS regime—instead ofexponentially many encoded sub-messages, a linear number of sub-messages (in L ) suffices. Thisreduction is key to the achievability proof—it enables an efficient elimination of the optimizationvariables used in describing the GDoF region achievable through the SLS scheme, from which wecan match it to the outer bound of Theorem 2 (see Section 6). This has been shown using a signal-scale IA scheme for fixed channels in [30], and can also be shown for varying(generic) channels using the signal-space IA scheme in [41], as we will see in Section 4.1. K × KL MISO-BC is verychallenging due to the explosion in the number of parameters and optimization variables, renderinga direct comparison between mc-Co and mc-TIN infeasible in general cellular networks. Inspiredby [40], we circumvent this complexity barrier by appealing to extremal network analysis andfocusing on the extremal GDoF gain from mc-Co over mc-TIN. The cooperative GDoF outerbound in Theorem 2 proves very useful for this purpose—it turns out to be tight for the classesof extremal networks, which maximize the GDoF gain of mc-Co over mc-TIN, in the regimes ofinterest. We show that the extremal GDoF gain from mc-Co over mc-TIN is given by: 3 / − /K in the mc-CTIN regime; and scales as Θ (cid:0) log( K ) (cid:1) in the mc-SLS regime.Interestingly, these gains do not depend on the number of users per-cell L , and they are exactlyequal to their counterpart extremal gains in K × K networks [40]. In other words, in the threeweak interference regimes of interest, additional (weaker) users in each cell have no influence onextremal GDoF gains, governed by the underlying K × K networks of single-user cells. For positive integers z and z where z ≤ z , the sets { , , . . . , z } and { z , z + 1 , . . . , z } aredenoted by (cid:104) z (cid:105) and (cid:104) z : z (cid:105) , respectively. For any real number a ∈ R , we have ( a ) + = max { , a } .Bold symbols denote tuples, e.g. a = ( a , . . . , a Z ) and A = ( A , . . . , A Z ), while calligraphic symbolsdenote sets, e.g. A = { a , . . . , a Z } . For any pair of sets A , B ⊆ R K , their Minkowski sum A ⊕ B isa set in R K defined as A ⊕ B (cid:44) { a + b : a ∈ A , b ∈ B} . The indicator function with condition A is denoted by ( A ), which is 1 when A holds and 0 otherwise. For functions f ( K ) and g ( K ), wehave f ( K ) = Θ (cid:0) g ( K ) (cid:1) if lim K →∞ f ( K ) /g ( K ) = c , where c > Consider a K -cell cellular network in which each cell k , where k ∈ (cid:104) K (cid:105) , comprises a base stationdenoted by BS- k and L user equipments, each denoted by UE-( l k , k ), where l k ∈ (cid:104) L (cid:105) . The set oftuples corresponding to all UEs in the network is given by U (cid:44) { ( l k , k ) : l k ∈ (cid:104) L (cid:105) , k ∈ (cid:104) K (cid:105)} .We focus on the downlink mode, where an independent message denoted by W [ l k ] k is communi-cated to each UE-( l k , k ). The input-output relationship at the t -th channel use is given by Y [ l k ] k ( t ) = K (cid:88) i =1 ¯ P α [ lk ] ki G [ l k ] ki ( t ) X i ( t ) + Z [ l k ] k ( t ) . (1)In the above, X i ( t ) , Y [ l k ] k ( t ) , Z [ l k ] k ( t ) ∈ C are, respectively, the symbol transmitted by BS- i , thesymbol received by UE-( l k , k ), and the zero-mean unit-variance additive white Gaussian noise(AWGN) at UE-( l k , k ). The signal transmitted by BS- i is subject to the unit average powerconstraint T (cid:80) Tt =1 E (cid:2)(cid:12)(cid:12) X i ( t ) (cid:12)(cid:12) (cid:3) ≤
1, where T is the communication duration in channel uses. The T -channel-use-long signal (or codeword) of BS- i is given by X i (cid:44) (cid:0) X i ( t ) : t ∈ (cid:104) T (cid:105) (cid:1) .For GDoF purposes, we define ¯ P (cid:44) √ P , where P is a nominal power parameter that approachesinfinity in the GDoF limit. The exponent α [ l k ] ki > i and UE-( l k , k ), while G [ l k ] ki ( t ) ∈ C is the corresponding channel fading coefficient. Thearray of all fading coefficients is given by G (cid:44) (cid:0) G [ l k ] ki ( t ) : ( l k , k ) ∈ U , i ∈ (cid:104) K (cid:105) , t ∈ (cid:104) T (cid:105) (cid:1) . We assumethat UEs have access to both channel strength parameters and channel fading coefficients, whileBSs know channel strength parameters (and the probability distribution of fading coefficients) only.5hese assumptions, which are discussed in greater detail further on, lead to robust GDoF charac-terizations, that are far less sensitive to perturbations in the channel state as seen by transmitters,compared to GDoF results that assume perfect channel knowledge at the transmitters.We define α ∈ R K × K × L + as the 3-dimensional array comprising all channel strength parametersof a given network, where the ( i, j, l i )-th element of α is given by α ( i, j, l i ) = α [ l i ] ij . Note that α describes the topology of a K × KL network, specifying the strengths of connections betweendifferent BS-UE pairs; and hence we will often refer to α as a network. As seen further on, wefocus on special regimes (i.e. subsets of networks) specified by imposing conditions on α . Remark 1.
For ease of exposition, we assume that we have an equal number of users given by L in each cell. As far as GDoF results are concerned, this symmetrization in the number of usersper-cell incurs no loss of generality—for any cell i with L i < L users, we may add L − L i trivialusers, with strengths α [ l i ] ij = 0 for all j ∈ (cid:104) K (cid:105) , which have no influence on the GDoF. ♦ Remark 2.
In the considered model, the capacity of a link connecting BS- j and UE-( l i , i ) isgiven by C [ l i ] ij ( P ) ≈ α [ l i ] ij log( P ). While individual link capacities, and hence the network capacity,are scaled up to infinity in the GDoF limit, defined by taking P → ∞ while normalizing bylog( P ), their ratios remain approximately fixed, i.e. C [ l i ] ij ( P ) /C [ l k ] kl ( P ) ≈ α [ l i ] ij /α [ l k ] kl . This is themain advantage of the GDoF model [3, 4] compared to the less refined (yet more intuitive) DoFmodel, recovered by setting α [ l i ] ij = 1 for all links [13, 41]. Indeed, the GDoF model allows us to de-emphasize the effects of additive noise, hence focusing on the interaction between different signals,while maintaining the disparity amongst the strengths of different links, which in turn allows us todistinguish between different regimes (e.g. weak and strong interference regimes). A caveat is theoverwhelming complexity incurred by the many parameters in α . ♦ We assume that channel strength parameters are perfectly known to both the transmitters (BSs)and the receivers (UEs). Channel fading coefficients, however, are perfectly known to the receiversbut only available to finite precision at the transmitters (finite precision CSIT). This implies that:a) the joint and conditional probability density functions of the channel coefficients (entries of G )exist and the peak values of these distributions are bounded by a constant which is independent of P ; and b) the transmitters only know the distributions of channel coefficients, but not their actualrealizations [7, 17]. Therefore, the codewords X , . . . , X K are independent of the realizations ofchannel coefficients, yet may depend on their distributions as well as channel strength parameters. Under no cooperation, each BS has access to the set of messages intended to UEs in the samecell only, and the network is modeled by an IBC comprising K mutually interfering Gaussian BCs(cells)—see Fig. 1(left). For instance, BS- k has messages W k = (cid:0) W [ l k ] k : l k ∈ (cid:104) L (cid:105) (cid:1) and encodesthem into the codeword X k , independently of all other BSs. On the other end, UE-( l k , k ) in cell k sees the contributions from all codewords X i with i ∈ (cid:104) K (cid:105) \ { k } as inter-cell interference.For any given P , achievable rate tuples R ( P ) = (cid:0) R [ l k ] k ( P ) : ( l k , k ) ∈ U (cid:1) and the capacityregion C IBC ( P ) are defined in a standard manner, see, e.g. [29, 30]. A GDoF tuple is denote by6 𝑊 ![!] !𝑊 ![$] 𝑊 ! ! 𝑊 ! $ !𝑊 $[!] !𝑊 $[$] !𝑊 %[!] !𝑊 %[$] 𝑊 $! 𝑊 $$ 𝑊 %! 𝑊 %$ No Cooperation (IBC) Full Cooperation (MISO-BC) !𝑊 ![!] !𝑊 ![$] 𝑊 ! ! 𝑊 ! $ 𝑊 $! 𝑊 $$ 𝑊 %! 𝑊 %$ !𝑊 $[!] !𝑊 $[$] !𝑊 %[!] !𝑊 %[$] →→→→ Figure 1: 3-cell network with 2 users per-cell under different levels of cooperation. Left: Nomulti-cell cooperation (IBC). Right: Full multi-cell cooperation (MISO-BC). d = (cid:0) d [ l k ] k : ( l k , k ) ∈ U (cid:1) , and the GDoF region is defined in a standard fashion as D IBC (cid:44) (cid:40) d ∈ R |U| + : d [ l k ] k = lim P →∞ R [ l k ] k ( P )log( P ) , R ( P ) ∈ C IBC ( P ) (cid:41) . Under full cooperation, all BSs have access to all messages (cid:0) W , . . . , W K (cid:1) , jointly encoded intothe vector codeword X (cid:44) ( X , . . . , X K ), comprising K scalar codewords, and of which the k -thcomponent X k is transmitted through BS- k . In this cooperative setting, each BS is viewed as anantenna in a large multi-antenna transmitter, and the network is modeled by a MISO-BC with a K -antenna transmitter and KL single-antenna receivers—see Fig. 1(right). The capacity region ofthe MISO-BC is denoted by C MBC ( P ), and the GDoF region is defined as D MBC (cid:44) (cid:40) d ∈ R |U| + : d [ l k ] k = lim P →∞ R [ l k ] k ( P )log( P ) , R ( P ) ∈ C MBC ( P ) (cid:41) . The finite precision CSIT assumption eliminates the GDoF benefits of schemes that rely on preciseCSIT [17], such as IA for the IBC [41], and ZF and DPC for the MISO-BC [19], which naturallygives way to classes of robust schemes based on power control, TIN, rate-splitting and SLS [21–34],which only require coarse CSIT (i.e. channel strength parameters).
Perhaps the simplest robust multi-cell scheme, for which somewhat general optimality results canbe shown, is the mc-TIN scheme [30]. In mc-TIN, no BS cooperation is assumed and each BS This setting, with more receivers than transmit antennas, is also called an overloaded
MISO-BC [32]. i is composed as X i ( t ) = (cid:88) l i ∈(cid:104) L (cid:105) (cid:113) q [ l i ] i X [ l i ] i ( t ) (2)where each message W [ l i ] i is encoded into the signal X [ l i ] i using an independent Gaussian codebookwith unit average power, and q [ l i ] i ≥ l i , i ) such that the per-BSpower constraint (cid:80) l i ∈(cid:104) L (cid:105) q [ l i ] i ≤ p , . . . , p K , each UE- (cid:0) p ( l ) , i (cid:1) successively decodes and cancels the signals X [ p i (1)] i , . . . , X [ p i ( l − i , in this specific order, before decoding its own signal X [ p i ( l )] i , while treatingall other signals (i.e. both intra-cell and inter-cell interference) as noise.The set of GDoF tuples achieved through all feasible power allocation policies and successivedecoding orders constitute the mc-TIN achievable GDoF region (i.e. TINA region), denoted by D TINA in this paper. An explicit characterizations of the TINA region is obtained in [30].
Under full BS cooperation, i.e. the MISO-BC with finite precision CSIT, there is generally no non-trivial regime for which mc-TIN is GDoF optimal. The closest robust alternative in this case is theSLS scheme [33], which we call mc-SLS in multi-cell settings. In the most general form of mc-SLS,the K BSs jointly transmit a superposition of independent Gaussian codewords ( X S : S ⊆ U ),where each X S is decoded by all users in the subset S , and treated as additional Gaussian noiseby users not in S . The signal transmitted by BS- i is hence composed as X i ( t ) = (cid:88) S⊆U √ q S ,i X S ( t ) (3)where q S ,i ≥ X S by BS- i , such that (cid:80) S⊆U q S ,i ≤
1. UEsemploy successive decoding with arbitrary orders, optimized alongside power control variables.The above scheme is essentially a multiple multicast transmission, where a distinct message iscommunicated to each subset of users, and the GDoF achieved by different messages comprise thetuple ( d S : S ⊆ U ). To go from the original unicast messages (cid:0) W [ l k ] k : ( l k , k ) ∈ U (cid:1) to multiplemulticast transmission, rate-splitting is used. Each message W [ l k ] k is split into sub-messages givenby (cid:0) W [ l k ] k, S : S ⊆ U , ( l k , k ) ∈ S (cid:1) , and then each set of sub-messages indexed by the same S ⊆ U , i.e. W S = (cid:0) W [ l k ] k, S : ( l i , i ) ∈ U (cid:1) , is encoded into the codeword X S . It follows that d S = (cid:88) ( l i ,i ) ∈S d [ l i ] i, S (4)where d [ l i ] i, S is the GDoF carried by sub-message W [ l i ] i, S , i.e. the portion of d S assigned to UE-( l i , i ).It follows that the SLS region, denoted by D SLS , is given by all GDoF tuples d ∈ R |U| + that satisfy d [ l i ] i ≤ (cid:88) S⊆U :( l i ,i ) ∈S d [ l i ] i, S (5) Schemes similar to SLS have been investigated in tandem with precoder design under the name of rate-splitting,see, e.g., [42, 43]. These works focus on optimization aspects, with less emphasis on fundamental limits. d S : S ⊆ U ), with components given by (4), achievedthrough a feasible power control policy and successive decoding orders. It is readily seen that theSLS scheme depends on a high number of auxiliary design variables—a major barrier in the faceof obtaining an explicit characterization of D SLS and proving optimality results [33]. However, insome cases, the representation of D SLS can be simplified as we will see further on in Section 6.
In this section, we present a number of definitions which are essential for the formulation and proofsof the main results, presented in subsequent sections. We start by highlighting that, without lossof generality, we may assume that in each cell i , direct link strengths are ordered as α [1] ii ≤ α [2] ii ≤ · · · ≤ α [ L ] ii . (6)That is, same-cell users are in an ascending order with respect to their SNRs. Note that in theabsence of inter-cell interference, (6) determines the degradedness order in each cell (or BC).Throughout this work, we focus on three regimes of channel parameters (or networks). Eachregime is a subset of R K × K × L + , and can be thought of as a collection of networks that share certain(desirable) properties. Before introducing the regimes of interest, it is worthwhile highlighting thatall three regimes are included in a larger weak inter-cell interference regime, described by the setof all networks α ∈ R K × K × L + that satisfy the SIR order0 ≤ α [1] ii − α [1] ij ≤ α [2] ii − α [2] ij ≤ · · · ≤ α [ L ] ii − α [ L ] ij (7)for all i, j ∈ (cid:104) K (cid:105) . We find it quite instructive to think of (7), as well as the three regimes of interestintroduced further on, in terms of two types of conditions: 1) inter-cell conditions, which controlinterference levels between distinct cells; and 2) intra-cell conditions, which govern (and preserve)the order amongst same-cell users under inter-cell interference.The first type of conditions is captured by the left-most inequality in (7), which (alongside theother inequalities) implies that for all ( l i , i ) ∈ U and j ∈ (cid:104) K (cid:105) , we have α [ l i ] ii ≥ α [ l i ] ij . (8)That is, a direct link between a BS and any of its associated UEs must be no weaker than interfering(or cross) links to the same UE. This places the cellular network in the weak inter-cell interferenceregime, an analog of the IC’s weak interference regime [3]. As we will see in the following subsection,all three regimes of interest fall within this weak inter-cell interference regime.The second type of conditions is captured by the right-most L − i , and against interference from any other cell j , the SIR order ofsame-cell users should follow their SNR order in (6). In other words, a stronger user in the SNRsense must also be stronger in the SIR sense, that is α [ l i ] ii ≥ α [ l i − ii = ⇒ α [ l i ] ii − α [ l i ] ij ≥ α [ l i − ii − α [ l i − ij . (9)As we will see, the above SIR order also holds in all three regimes of interest. Remark 3.
The SIR order in (7) greatly simplifies the mc-TIN scheme described in Section 2.3.1.In particular, under this order it is optimum from the TINA region perspective to use the optimuminterference-free successive decoding order (natural order) in each cell, i.e. p i ( l ) = l for all l ∈ (cid:104) L (cid:105) and i ∈ (cid:104) K (cid:105) . This is not necessarily this case when (7) does not hold, see [29, 30]. Moreover, theorder in (7) may be seen as quite natural—cell-centre users with high long-term SNRs are alsoexpected to have higher long-term SIRs compared to cell-edge users. ♦ .1 Multi-Cell TIN, CTIN and SLS regimes We now present the three regimes of interest, starting with the mc-TIN regime [29, 30].
Definition 1. (mc-TIN Regime).
This regime is denoted by A TIN , and is given by all networks α ∈ R K × K × L + that satisfy α [ l i ] ii ≥ α [ l i ] ij + α [ l k ] ki (10) α [ l i ] ii ≥ α [ l i ] ij + α [ l i − ii − (cid:0) α [ l i − ij − α [ l i ] ij (cid:1) + (11)for all cells i, j, k ∈ (cid:104) K (cid:105) such that i / ∈ { j, k } , and for all users l i ∈ (cid:104) L (cid:105) and l k ∈ (cid:104) L (cid:105) . ♦ It can be verified that the inequalities in (10) and (11) are stricter versions of their counterpartsin (8) and (9), respectively. The mc-TIN regime derives its significance from the fact that mc-TINachieves the entire GDoF region of the IBC in this regime [30]. Hence, this regime can be thoughtof as a very weak inter-cell interference regime. The mc-TIN regime generalizes the TIN regimeintroduced for K × K interference networks by Geng et al. [21] to K × KL cellular networks. Next,we present a strictly larger regime called the mc-CTIN regime [29, 30]. Definition 2. (mc-CTIN Regime).
This regime is denoted by A CTIN , and is given by allnetworks α ∈ R K × K × L + that satisfy α [ l i ] ii ≥ max (cid:0) α [ l i ] ij + α [ l j ] ji , α [ l i ] ik + α [ l j ] ji − α [ l j ] jk (cid:1) (12) α [ l i ] ii ≥ α [ l i ] ij + α [ l i − ii − α [ l i − ij (13)for all cells i, j, k ∈ (cid:104) K (cid:105) such that i / ∈ { j, k } , and for all users l i ∈ (cid:104) L (cid:105) and l j ∈ (cid:104) L (cid:105) . ♦ Note that (12) is a stricter version of (8), while (13) is identical to (9). In the mc-CTIN regime,the TINA GDoF region D TINA is a convex polyhedron without the need for time-sharing, which isnot necessarily the case outside the mc-CTIN regime [29, 30]. The mc-CTIN regime generalizes theCTIN regime introduced for K × K interference networks by Yi and Caire [24].While the TINA region is convex in the mc-CTIN regime, the GDoF region of the IBC remainsunknown in this regime. In Section 4.1, we settle this question under the assumption of finiteprecision CSIT, and we show that mc-TIN is GDoF optimal for the IBC in the mc-CTIN regime.Next, we introduce the largest of the three regimes, which we call the mc-SLS regime. Definition 3. (mc-SLS Regime).
This regime is denoted by A SLS , and is given by all networks α ∈ R K × K × L + that satisfy α [ l i ] ii ≥ max (cid:0) α [ l i ] ij , α [ l k ] ki , α [ l i ] ik + α [ l j ] ji − α [ l j ] jk (cid:1) (14) α [ l i ] ii ≥ α [ l i ] ij + α [ l i − ii − α [ l i − ij (15)for all cells i, j, k ∈ (cid:104) K (cid:105) such that i / ∈ { j, k } , and for all users l i ∈ (cid:104) L (cid:105) and l k ∈ (cid:104) L (cid:105) . ♦ Note that (14) is a stricter version of (8), while (15) is identical to (9). Moreover, we have A TIN ⊆ A
CTIN ⊆ A
SLS . The mc-SLS regime extends the SLS regime introduced by Davoodi and Jafar [33] to K × KL cellular networks. The significance of the SLS regime is due to the fact that the SLS scheme is10DoF optimal for K × K MISO-BCs with K ≤ K >
3, no proof is heretofore available, mainly due to the overwhelming complexity of the GDoFregion when K is arbitrary. Nevertheless, the SLS regime is still very useful in the sense that itlends itself to extremal network analysis as recently shown by Chan et al. [40], where extremalgains of transmitter cooperation over TIN are explicitly obtained.In Sections 4.2 and 4.3, we show that K × KL cellular networks enjoy desirable properties inthe mc-SLS regime. In particular, we derive a malleable outer bound for the K × KL MISO-BCin the mc-SLS regime, which lends itself directly to extremal network analysis, through which webound the gain of mc-Co over mc-TIN in all three regimes of interest.
Definition 4. (Cycles).
A cycle π of length | π | = M is an ordered sequence of M users fromdistinct cells, given by π = (cid:0) ( l i , i ) → ( l i , i ) → · · · → ( l i M , i M ) (cid:1) . (16)We define { π } (cid:44) (cid:8) ( l i , i ) , ( l i , i ) , · · · , ( l i M , i M ) (cid:9) as the set of users involved in cycle π . The m -thuser in a cycle π is also denoted by π ( m ) = ( l i m , i m ), from which (16) is equivalently expressed as π = (cid:0) π (1) → · · · → π ( M ) (cid:1) . The set of all cycles (of all lengths) is denoted by Π. Each cycle π isassociated with an implicit cycle encompassing BS indices, given by σ = ( i → i → · · · → i M ) (17)which is also written as σ = (cid:0) σ (1) → · · · → σ ( M ) (cid:1) . For any cycle of length M , indices areinterpreted modulo M , e.g. π ( M + m ) = π ( m ) and σ ( M + m ) = σ ( m ), for all integers m . ♦ Next, we define cycle bounds for the IBC and MISO-BC associated with cycles defined above.For each cycle π ∈ Π, a cycle bound is a bound on the sum-GDoF of users in the set (cid:8) ( s k , k ) : s k ∈ (cid:104) l k (cid:105) , ( l k , k ) ∈ { π } (cid:9) . That is, the set comprising each participating user ( l k , k ) ∈ { π } , as well as same-cell users thatprecede user ( l k , k ) in the SNR (or SIR) order. Cycle bounds are defined as follows. Definition 5. (IBC Cycle Bounds).
The IBC cycle bound associated with π ∈ Π is given by (cid:88) ( l k ,k ) ∈{ π } (cid:88) s k ∈(cid:104) l k (cid:105) d [ s k ] k ≤ ∆ π (18)where the quantity ∆ π is defined as∆ π (cid:44) α [ l i ] i i , if M = 1 (cid:80) Mm =1 (cid:16) α [ l im ] i m i m − α [ l im +1 ] i m +1 i m (cid:17) , if M > . (19)As shown in [30], the TINA region D TINA can be expressed as a union of polyhedra (polyhedralTIN regions), each described in terms of the above IBC cycle bounds. These cycle bounds will alsoconstitute outer bounds for the IBC in the mc-CTIN regime, as we will show further on. ♦ Definition 6. (MISO-BC Cycle Bounds).
Each π ∈ Π gives rise to M = | π | MISO-BC cyclebounds. The m -th bound associated with π is given by (cid:88) ( l k ,k ) ∈{ π } (cid:88) s k ∈(cid:104) l k (cid:105) d [ s k ] k ≤ ∆ + π,m (20)11here the quantity ∆ + π ( m ) is defined as∆ + π,m (cid:44) (cid:40) ∆ π , if M = 1 , ∆ π + α [ l im +1 ] i m +1 i m , if M > . (21)The above cycle bounds will constitute outer bounds for the MISO-BC in the mc-SLS regime, aswe will show further on. ♦ In this section, we present the main results of this work, alongside some observations and insights.It is worthwhile highlighting that our main results are all under the assumption of finite precisionCSIT. Therefore, this is always assumed (implicitly), unless mentioned otherwise.
From the IBC cycle bounds in Definition 5, we construct what is known as the polyhedral-TIN(PTIN) region, denoted as D PTIN . This is given by all tuples d ∈ R KL + that satisfy (cid:88) ( l k ,k ) ∈{ π } ¯ d [ l k ] k ≤ ∆ π , ∀ π ∈ Π (22)where ¯ d [ l k ] k is a shorthand notation for (cid:80) s k ∈(cid:104) l k (cid:105) d [ s k ] k . From [30], we know that D TINA = D PTIN inthe mc-CTIN regime, and D IBC = D TINA = D PTIN in the mc-TIN regime. Next, we show thatthe latter also holds in the mc-CTIN regime under finite precision CSIT.
Theorem 1.
In the mc-CTIN regime, mc-TIN is GDoF optimal for the K × KL IBC. That is, α ∈ A CTIN = ⇒ D IBC = D TINA = D PTIN . The direct part (achievability) of Theorem 1 follows from [30], as highlighted above. Theconverse relies on a new application of AI bounds [7, 17] to cellular networks, and is presented inSection 5. Next, we draw some insights from the results in Theorem 1.We first observe that under prefect CSIT, mc-TIN is not necessarily GDoF optimal for the IBCin the mc-CTIN regime, which is in sharp contrast to the finite precision CSIR result in Theorem 1.This is seen through the example in Fig. 2, particularly in the case where α = 0 . < β ≤ . is achieved through the signal-space downlink IA scheme proposed by Suh and Tsein [41]. On the other hand, the sum-GDoF achieved using mc-TIN is at most 2(1 − β ) < .The observation that signal-space IA seizes to be useful for the IBC under finite precisionCSIT from the GDoF perspective is perhaps not very surprising in light of previous results [7, 17].It is, however, not immediately clear (without Theorem 1) whether a form of robust inter-cellinterference management that is more sophisticated than mc-TIN, e.g. based on rate-splitting andpartial interference decoding [3,7], is required to achieve the optimal GDoF in the mc-CTIN regime.Theorem 1 settles this issue, and shows that the simple mc-TIN scheme is indeed GDoF optimal forthe IBC in the mc-CTIN regime. Moreover, we observe in this regime that the sum-GDoF of the K × KL IBC is equal to the sum-GDoF of the underlying K × K IC, comprising only the strongest These statements hold regardless of the CSIT assumptions. Moreover, outside the mc-CTIN regime, D TINA is aunion of PTIN regions, each defined by selecting a subset of active users and deactivating remaining users. E − UE − BS-1 UE − UE − BS-2 𝛼𝛽𝛽𝛼𝛼𝛽1111 𝛼 = 𝛽 𝛼 = 2𝛽
Figure 2: Left: Symmetric 2-cell network with 2 users per-cell. Right: The mc-CTIN regime ingrey and the mc-TIN regime in striped blue, assuming α ≥ β .user in the SNR (or SIR) sense from each cell, i.e. UE-( L, i ) for all i ∈ (cid:104) K (cid:105) . This is seen fromthe fact that in the mc-TIN scheme, UE-( L, i ) decodes all L messages in cell i , hence limiting thetotal GDoF achieved in cell i to the GDoF achievable by UE-( L, i ). As mc-TIN is GDoF optimalin the mc-CTIN regime under finite precision CSIT, it follows that in each cell i , all users otherthan UE-( L, i ) are redundant from a standpoint of maximizing the sum-GDoF.It is also worthwhile noting that the redundancy of weaker users in the mc-CTIN regime doesnot necessarily hold under perfect CSIT. This is clearly seen through the above example of Fig.2, where lower SIR users are necessary to achieve IA gains—the sum-GDoF is limited to thatachieved via TIN in the absence of these weaker users [3]. Finite precision CSIT, on the otherhand, eliminates all GDoF gains of IA (in both signal-space and signal-scale), and brings classicalschemes, based on Gaussian random codes, superposition and power control, back to the forefront.Under these schemes, inter-cell interference is Gaussian and unstructured, and the SIR order in (7)takes on the interference-free role of the SNR order in (6), i.e. cells resemble their interference-freedegraded BC counterparts, in which weaker users are redundant.
Similar to the region D PTIN constructed from IBC cycle bounds, the MISO-BC cycle bounds inDefinition 6 can be used to construct a region given by all tuples d ∈ R KL + that satisfy (cid:88) ( l k ,k ) ∈{ π } ¯ d [ l k ] k ≤ ∆ + π,m , ∀ m ∈ (cid:104)| π |(cid:105) , π ∈ Π . (23)We denote this region by D SLSout , a notational choice justified by the following theorem.
Theorem 2.
In the mc-SLS regime, the GDoF region of the K × KL MISO-BC is included in theouter bound D SLSout . That is, α ∈ A SLS = ⇒ D MBC ⊆ D
SLSout . The proof of Theorem 2 also relies on AI bounds, and it is presented in Section 5.3. A naturalquestion that follows is whether D SLSout is tight for the K × KL MISO-BC in the mc-SLS regime. Inthe following result, we show that this is indeed the case in 2-cell networks.
Theorem 3.
In the mc-SLS regime, mc-SLS is GDoF optimal for the × L MISO-BC. In thiscase, we have α ∈ A SLS = ⇒ D MBC = D SLS = D SLSout . D SLS , described in its generality in Section 2.3.2, greatly simplifies for2-cell networks in the mc-SLS regime. Instead of the 2 L − L + 1 codewords in this case: 2 L non-cooperative codewords from the mc-TIN scheme; and a superimposed cooperative commoncodeword decoded by all users. This simplification allows for an efficient elimination of auxiliaryvariables used in describing D SLS , which turns out to coincide with D SLSout .Beyond 2-cell networks, the outer bound in Theorem 2 is not tight in general. This is seenfrom the 3 × D MBC = D SLS in theSLS regime (i.e. SLS is optimal here), the achievable region D SLS and the outer bound D SLSout donot coincide in this settings. In particular, additional bounds, which are not implied by the cyclebounds in (23), are generally required to fully describe D SLS in the 3 × D SLSout is not tight in general, this outer bound remains very useful for studying thepotential benefits of full mc-Co over mc-TIN under finite precision CSIT. In particular, in each ofthe three regimes on interest, D SLSout is tight for a subset of extremal networks, in which sum-GDoFgains of full multi-cell cooperation over mc-TIN are maximized. This allows us to obtain sharpcharacterizations of these extremal GDoF gains, as we see next.
Equipped with the results in Theorem 1 and Theorem 2, we are ready to conduct extremal networkanalysis to understand the robust GDoF benefits of mc-Co over mc-TIN in the weak inter-cellinterference regimes of interest. To this end, we define the sum-GDoF achievable by mc-TIN as d TINΣ ( α ) (cid:44) max d ∈D TINA ( α ) (cid:88) ( l k ,k ) ∈U d [ l k ] k . (24)In a similar manner, the sum-GDoF achieved through mc-Co is defined as d MBCΣ ( α ) (cid:44) max d ∈D MBC ( α ) (cid:88) ( l k ,k ) ∈U d [ l k ] k . (25)Note that the dependency of the GDoF on α is made explicit in this part. It is also worth notingthat d MBCΣ ( α ) is the optimal sum-GDoF of the underlying K × KL MISO-BC, with no restrictionon the employed scheme, hence representing the ultimate performance of mc-Co schemes. On theother hand, d TINΣ ( α ) is the maximum sum-GDoF achieved while restricting to the simple non-cooperative scheme of mc-TIN. This is optimal for the underlying K × KL IBC in the mc-CTINregime, yet its optimality is not guaranteed outside the mc-CTIN regime.We are interested in the extremal sum-GDoF gain from mc-Co relative to mc-TIN in the threeregimes of interest. For a regime
A ⊂ R K × K × L + , the extremal gain is defined as η K,L ( A ) (cid:44) max α ∈A d MBCΣ ( α ) d TINΣ ( α ) (26)which is parametrized by the network dimensions K, L . As elaborated by Chan et al. [40], theextremal gain η K,L ( A ) captures the potential benefits of mc-Co over mc-TIN in A , and does notnecessarily reflect typical or average performance gains. Nevertheless, studying η K,L ( A ) can still bevery useful for bringing closure to regimes in which potential gains are small; as well as identifyingregimes that warrant further investigation due to large potential gains. The two types of conclusions14re seen through the following theorem, which generalizes the results in [40, Th. 5.1, 6.1 and 7.1]to K × KL networks. We implicitly assume that K ≥
2, as the single-cell case is degenerate.
Theorem 4.
The extremal sum-GDoF gain of mc-Co over mc-TIN in K × KL networks underfinite precision CSIT in the three regimes of interest is as follows: η K,L ( A ) = , A = A TIN − K , A = A CTIN Θ (cid:0) log( K ) (cid:1) , A = A SLS . (27)It is readily seen that in each of the three regimes of interest, the extremal gain η K,L ( A ) isindependent of the number of users per-cell L , and therefore we have η K,L ( A ) = η K, ( A ). This isa consequence of the SIR order in (7), which under finite precision CSIT and from a sum-GDoFstandpoint, renders the first L − redundant with respect to user L (i.e. thestrongest user). This redundancy has already been highlighted in the context of mc-TIN and theIBC in Section 4.1. The same type of redundancy is exhibited by the MISO-BC in the mc-SLSregime, as seen from the outer bound D SLSout (see Section 7). The redundancy of weaker users in theregimes of interest is exploited in the proof of Theorem 4, presented in Section 7.Theorem 4 quantifies the intuition that in regimes where inter-cell interference is sufficientlyweak to the extent that it is GDoF optimal to treat it as noise for the IBC, gains due to BScooperation are limited. In particular, mc-Co schemes provide at most constant factor GDoF gainsover mc-TIN in the mc-TIN and mc-CTIN regimes, bounded above by 1 . K and L . On theother hand, there is far greater potential in the mc-SLS regime, specifically the part not includedin the mc-CTIN regime. Here the extremal GDoF gain of mc-Co over mc-TIN may scale as log( K )in large networks, rendering this regime more interesting for further investigation. In this section we present proofs for the outer bounds in Theorem 1 and Theorem 2. For thispurpose, we work with a deterministic approximation of the channel model in (1) given by¯ Y [ l k ] k ( t ) = K (cid:88) i =1 (cid:4) ¯ P α [ lk ] ki − α max ,i G [ l k ] ki ( t ) ¯ X i ( t ) (cid:5) . (28)In (28), we have α max ,i (cid:44) max ( l j ,j ) ∈K α [ l j ] ji , and both the real and imaginary components of ¯ X i ( t ) aredrawn from the integer alphabet (cid:10) (cid:100) ¯ P α max ,i (cid:101) (cid:11) . It can be easily checked that in all three regimesof interest, we have α max ,i = α [ L i ] ii . As shown in [17], the GDoF region of the above deterministicchannel model contains its counterpart GDoF of the original Gaussian model.We now recall a key lemma from [40] (see [7] for the proof). To this end, we consider the outputs¯ Y k ( t ) = K (cid:88) i =1 (cid:4) ¯ P λ i − α max ,i G ki ( t ) ¯ X i ( t ) (cid:5) (29)¯ Y j ( t ) = K (cid:88) i =1 (cid:4) ¯ P ν i − α max ,i G ji ( t ) ¯ X i ( t ) (cid:5) (30)where λ i , ν i ∈ [0 , α max ,i ], for all i ∈ (cid:104) K (cid:105) , are the corresponding channel strengths. We use our stan-dard notation ¯ X i , ¯ Y k and G for codewords, received signals and channel coefficients, respectively.15 emma 1. (Aligned Images Bounds [40, Lemma 2.1], [7, Lemma 1]) Let U be an auxiliaryrandom variable and assume that ( U, ¯ X , . . . , ¯ X K ) are independent of G . We have H (cid:0) ¯ Y k | G , U (cid:1) − H (cid:0) ¯ Y j | G , U (cid:1) ≤ max i ∈(cid:104) K (cid:105) ( λ i − ν i ) + T log( P ) + T o (log( P )) . (31)Lemma 1 bounds the maximum difference of entropies (in the GDoF sense) between the tworeceived signals ¯ Y k and ¯ Y j , that can be created by any set of codewords ¯ X , . . . , ¯ X K , which areindependent of the exact realizations of channel coefficients in G . The bound in (31) tells us thata maximum difference of entropies is created through ¯ X i , where i is the index yielding a maximumdifference in strengths ( λ i − ν i ) + . An example of Lemma 1 is shown in Fig. 3. Tx-1 𝜆 ! Tx-2Tx-3 Rx- 𝑘 Rx- 𝑗 𝜆 " 𝜆 𝜈 ! 𝜈 " 𝜈 𝜆 ! 𝜆 " 𝜆 𝜈 ! 𝜈 " 𝜈 Figure 3: Left: 3-transmitter, 2-receiver network. Right: The corresponding received power levels.The entropy difference between receivers k and j is maximized through transmitter 3.Since o (log( P )) terms are inconsequential for GDoF results, they will be dropped henceforthfor brevity. Moreover, we adopt the compact notation of [40] to represent differences of entropiesof the type in Lemma 1. Using this compact notion, inequality (31) is expressed as H (cid:0) λ | U (cid:1) − H (cid:0) ν | U (cid:1) ≤ max i ∈(cid:104) K (cid:105) ( λ i − ν i ) + T log( P ) (32)where λ (cid:44) [ λ · · · λ K ] and ν (cid:44) [ ν · · · ν K ]. The expression in (32) succinctly captures theessential parts in (31), i.e. channel strength levels, and will be employed further on. Next, we employ Lemma 1 to bound the sum-rate of users in cell i , for any i ∈ (cid:104) K (cid:105) , underfinite precision CSIT. This single-cell bound holds in the general weak inter-cell interference regimespecified by the order in (7), and hence holds in all three regimes of interest. Moreover, the boundis applicable to both the IBC and MISO-BC, as we will see further on. Lemma 2.
In the weak inter-cell interference regime specified by (7) , the sum-rate of users asso-ciated with cell i , where i ∈ (cid:104) K (cid:105) , is bounded above as T (cid:88) l ∈(cid:104) L (cid:105) R [ l ] i ≤ (cid:0) α [ L ] ii − α [1] ii (cid:1) T log( P ) + H (cid:0) ¯ Y [1] i | G , U i (cid:1) − H (cid:0) ¯ Y [ L ] i | G , W i , U i (cid:1) (33) where U i is a side information random variable, independent of G and W i . roof. In what follows, we use W [1: l ] i to denoted (cid:0) W [1] i , . . . , W [ l ] i (cid:1) , where ( l, i ) ∈ U . From Fano’sinequality, and the independence of W [ l ] i and (cid:0) W [1: l − i , U i (cid:1) , we obtain T L (cid:88) l =1 R [ l ] i ≤ L (cid:88) l =1 I (cid:16) W [ l ] i ; ¯ Y [ l ] i | G , W [1: l − i , U i (cid:17) = L (cid:88) l =1 H (cid:16) ¯ Y [ l ] i | G , W [1: l − i , U i (cid:17) − H (cid:16) ¯ Y [ l ] i | G , W [1: l ] i , U i (cid:17) = L (cid:88) l =2 (cid:104) H (cid:16) ¯ Y [ l ] i | G , W [1: l − i , U i (cid:17) − H (cid:16) ¯ Y [ l − i | G , W [1: l − i , U i (cid:17)(cid:105) + H (cid:16) ¯ Y [1] i | G , U i (cid:17) − H (cid:16) ¯ Y [ L ] i | G , W i , U i (cid:17) = L (cid:88) l =2 (cid:104) H (cid:16) α [ l ] i | W [1: l − i , U i (cid:17) − H (cid:16) α [ l − i | W [1: l − i , U i (cid:17)(cid:105) + H (cid:16) ¯ Y [1] i | G , U i (cid:17) − H (cid:16) ¯ Y [ L ] i | G , W i , U i (cid:17) . (34)Next, we focus on bounding the sum of differences of entropies described in brief notation in (34).By employing Lemma 1, we obtain L (cid:88) l =2 H (cid:16) α [ l ] i | W [1: l − i , U i (cid:17) − H (cid:16) α [ l − i | W [1: l − i , U i (cid:17) ≤ L (cid:88) l =2 max j ∈(cid:104) K (cid:105) (cid:0) α [ l ] ij − α [ l − ij (cid:1) + T log( P ) ≤ L (cid:88) s =2 (cid:0) α [ l ] ii − α [ l − ii (cid:1) T log( P ) (35)= (cid:0) α [ L ] ii − α [1] ii (cid:1) T log( P ) . (36)(35) follows from the SIR order α [ l ] ii − α [ l ] ij ≥ α [ l − ii − α [ l − ij in (7), which implies α [ l ] ii − α [ l − ii ≥ α [ l ] ij − α [ l − ij , ∀ l ∈ (cid:104) L (cid:105) , as well as the SNR order α [ l ] ii ≥ α [ l − ii in (6), which allows us to drop ( · ) + . By combining (34) and(36), we obtain the desired inequality in (31), which holds for any cell i ∈ (cid:104) K (cid:105) .Note that Lemma 2 is derived without making explicit assumptions on transmitter cooperation,and it will be adapted to prove the outer bounds in Theorem 1 and Theorem 2 by adjusting thetransmitter cooperation assumptions and selecting suitable auxiliary variables U i . Remark 4.
Before we prove the multi-cell converse bounds, it is worthwhile noting that whenusing Lemma 2 in the context of the IBC, the negative entropy term in (33) can be written as H (cid:0) ¯ Y [ L ] i | G , W i , U i (cid:1) = H (cid:0) (cid:101) Y [ L ] i | G , U i (cid:1) where (cid:101) Y [ L ] i is obtained from ¯ Y [ L ] i by subtracting the contribution of the input signal ¯ X i . This holdssince the signal transmitted from BS- i depends only on the set of messages in the same cell underno transmitter cooperation, i.e. ¯ X i is fully determined by W i (see Section 2.2.1). Therefore, thecontribution of ¯ X i to ¯ Y [ L ] i can be subtracted, after which the conditioning on ¯ X i (or W i ) can bedropped, since ¯ X i is independent of all remaining signals from BSs other than BS- i . This, however,does not hold for the MISO-BC, where messages to all users in the network are jointly encodedby all BSs (see Section 2.2.2). In this case, choosing the right side information variables becomescritical to obtain the desired outer bound, as we will see in Section 5.3. ♦
17e are now equipped to prove the multi-cell converse bounds in Theorem 1 and Theorem 2.We follow in the footsteps of the proofs of [40, Th. 4.1] and [40, Lem. B.4], where similar conversebounds are derived for K × K networks. In the following proofs, we generalize the bounds in [40]to K × KL cellular networks, with the aid of single-cell bounds derived in Lemma 2. In this part, we prove that each of the inequalities in (22) is a valid outer bound for the IBC in themc-CTIN regime. Single-cell bounds in (22), associated with cycles of length | π | = 1, are directlyobtained from the capacity region of the degraded Gaussian BC (see, e.g., [44]). We hence focuson multi-cell bounds in (22), associated with cycles of length | π | ≥ π ∈ Π of length M = | π | ≥
2, and eliminate all non-participating users (andtheir message), i.e. users not in the set (cid:8) ( s k , k ) : s k ∈ (cid:104) l k (cid:105) , ( l k , k ) ∈ { π } (cid:9) . This cannot hurtparticipating users. For any participating cell i and its l i participating users, where ( l i , i ) ∈ { π } ,the corresponding sum-rate ¯ R [ l i ] i (cid:44) (cid:80) l i s i =1 R [ s i ] i is bounded above as¯ R [ l i ] i ≤ (cid:0) α [ l i ] ii − α [1] ii (cid:1) T log( P ) + H (cid:16) ¯ Y [1] i | G (cid:17) − H (cid:16) ¯ Y [ l i ] i | G , W i (cid:17) . (37)which follows directly from Lemma 2. Note that after eliminating non-participating users, W i nowcorresponds to (cid:0) W [1] i , . . . , W [ l i ] i (cid:1) . By summing over single-cell bounds obtained from (37) of all M participating cells, and after rearranging entropy terms, we obtain a cycle sum-rate bound as T (cid:88) ( l i ,i ) ∈{ π } ¯ R [ l i ] i ≤ M (cid:88) m =1 (cid:0) α [ l im ] i m i m − α [1] i m i m (cid:1) T log( P ) + M (cid:88) m =1 H (cid:16) ¯ Y [1] i m | G (cid:17) − H (cid:16) ¯ Y [ l im +1 ] i m +1 | G , W i m +1 (cid:17) . (38)Focusing on the the sum of differences of entropies in (38), we obtain M (cid:88) m =1 H (cid:16) ¯ Y [1] i m | G (cid:17) − H (cid:16) ¯ Y [ l im +1 ] i m +1 | G , W i m +1 (cid:17) = M (cid:88) m =1 H (cid:0) α [1] i m (cid:1) − H (cid:0) α [ l im +1 ] i m +1 | ¯ X i m +1 (cid:1) (39)= M (cid:88) m =1 H (cid:0) α [1] i m (cid:1) − H (cid:0) α [ l im +1 ] i m +1 ◦ ¯ e i m +1 (cid:1) . (40)The equality in (39) is obtained by invoking (32), while noting that ¯ X i m +1 depends only on W i m +1 in the IBC. In (40), ¯ e i m +1 is an appropriately-sized vector of all ones except for a single zero at the( i m +1 )-th element, while ◦ denotes the element-wise product. Therefore α [ l im +1 ] i m +1 ◦ ¯ e i m +1 = (cid:2) α [ l im +1 ] i m +1 · · · ( i m +1 )-th entry (cid:122)(cid:125)(cid:124)(cid:123) · · · α [ l im +1 ] i m +1 K (cid:3) . For each m ∈ (cid:104) M (cid:105) in (40), we subtracted the contribution of ¯ X i m +1 to ¯ Y [ l im +1 ] i m +1 , after which theconditioning on ¯ X i m +1 can be dropped (see Remark 4). From a GDoF perspective, this is equivalentto replacing α [ l im +1 ] i m +1 i m +1 in α [ l im +1 ] i m +1 with a zero, as in (40). By applying Lemma 1 to (40), we obtain M (cid:88) m =1 H (cid:0) α [1] i m (cid:1) − H (cid:0) α [ l im +1 ] i m +1 ◦ ¯ e i m +1 (cid:1) M (cid:88) m =1 max (cid:16) α [1] i m i m +1 , max k ∈(cid:104) K (cid:105) ,k (cid:54) = i m +1 (cid:16) α [1] i m k − α [ l im +1 ] i m +1 k (cid:17) + (cid:17) T log( P ) (41) ≤ M (cid:88) m =1 (cid:0) α [1] i m i m − α [ l im +1 ] i m +1 i m (cid:1) T log( P ) . (42)The bound in (42) holds due to the mc-CTIN condition in (12), which implies that α [1] i m i m − α [ l im +1 ] i m +1 i m ≥ α [1] i m k − α [ l im +1 ] i m +1 k α [1] i m i m − α [ l im +1 ] i m +1 i m ≥ α [1] i m i m +1 . By combining the bounds in (42) and (38), we obtain the desired cycle bound as T (cid:88) ( l i ,i ) ∈{ π } ¯ R [ l i ] i ≤ M (cid:88) m =1 (cid:16) α [ l im ] i m i m − α [ l im +1 ] i m +1 i m (cid:17) T log( P )= ∆ π T log( P ) . (43)The above applies to any cycle π of length | π | ≥
2, which concludes the converse proof for the IBC.
We now move on to proving that each of the inequalities in (23) is a valid outer bound for theMISO-BC in the mc-SLS regime. As in the previous part, we consider a cycle π ∈ Π of length M = | π | ≥
2. Unlike the IBC, however, such cycle is associated with M bounds for the MISO-BC.We first focus on the M -th of such bounds, i.e. (cid:80) ( l i ,i ) ∈{ π } (cid:80) s i ∈(cid:104) l i (cid:105) d [ s i ] k ≤ ∆ + π,M , and we addressremaining bounds further on. Following similar steps leading to (38), we obtain T (cid:88) ( l i ,i ) ∈{ π } ¯ R [ l i ] i ≤ M (cid:88) m =1 (cid:0) α [ l im ] i m i m − α [1] i m i m (cid:1) T log( P )+ M (cid:88) m =1 H (cid:16) ¯ Y [1] i m | G , U i m (cid:17) − H (cid:16) ¯ Y [ l im +1 ] i m +1 | G , W i m +1 , U i m +1 (cid:17) . (44)For m ∈ (cid:104) M (cid:105) , we select the side information variable U i m as U i m = (cid:0) W i m +1 , . . . , W i M (cid:1) . (45)Note that U i M is empty, and hence users in cell i M are given no side information about messagesintended to users in other cells. The sum of entropy differences in (44) leads to M (cid:88) m =1 H (cid:16) ¯ Y [1] i m | G , U i m (cid:17) − H (cid:16) ¯ Y [ l im +1 ] i m +1 | G , W i m +1 , U i m +1 (cid:17) ≤ M − (cid:88) m =1 (cid:20) H (cid:16) ¯ Y [1] i m | G , U i m (cid:17) − H (cid:16) ¯ Y [ l im +1 ] i m +1 | G , U i m (cid:17)(cid:21) + H (cid:16) ¯ Y [1] i M | G (cid:17) (46) It is worth noting that modulo M is not used for cell indices in (45). M − (cid:88) m =1 (cid:20) H (cid:16) α [1] i m | U i m (cid:17) − H (cid:16) α [ l im +1 ] i m +1 | U i m (cid:17)(cid:21) + α [1] i M i M T log( P ) (47) ≤ M − (cid:88) m =1 max k ∈(cid:104) K (cid:105) (cid:16) α [1] i m k − α [ l im +1 ] i m +1 k (cid:17) + T log( P ) + α [1] i M i M T log( P ) (48) ≤ M − (cid:88) m =1 (cid:16) α [1] i m i m − α [ l im +1 ] i m +1 i m (cid:17) T log( P ) + α [1] i M i M T log( P ) . (49)In (46), we used the fact ( W i m +1 , U i m +1 ) = U i m , while the upper bound in (48) is obtained usingLemma 1. On the other hand, (49) holds due to the mc-SLS condition in (14), which implies α [1] i m i m − α [ l im +1 ] i m +1 i m ≥ α [1] i m k − α [ l im +1 ] i m +1 k α [1] i m i m − α [ l im +1 ] i m +1 i m ≥ . By combining the bounds in (49) and (44), we obtain T (cid:88) ( l i ,i ) ∈{ π } ¯ R [ l i ] i ≤ M − (cid:88) m =1 (cid:0) α [ l im ] i m i m − α [ l im +1 ] i m +1 i m (cid:1) T log( P ) + α [ l iM ] i M i M T log( P )= M (cid:88) m =1 (cid:0) α [ l im ] i m i m − α [ l im +1 ] i m +1 i m (cid:1) T log( P ) + α [ l i ] i i M T log( P ) (50)= ∆ π T log( P ) + α [ l i ] i i M T log( P ) . (51)This proves the M -th outer bound associated with cycle π . To obtain the remaining M − π , we follow the same steps while replacing π with a shifted cycle π (cid:48) , where π (cid:48) ( m ) = π ( m + j ) for some j ∈ (cid:104) M − (cid:105) . We obtain the bound T (cid:88) ( l i ,i ) ∈{ π (cid:48) } ¯ R [ l i ] i ≤ ∆ π (cid:48) T log( P ) + α [ l i ] σ (cid:48) (1) σ (cid:48) ( M ) T log( P )= ∆ π T log( P ) + α [ l i ] σ ( j +1) σ ( j + M ) T log( P ) (52)= ∆ π T log( P ) + α [ l i ] i j +1 i j T log( P ) (53)where(52) holds since ∆ π (cid:48) = ∆ π and σ (cid:48) ( m ) = σ ( j + m ). Repeating the same steps for all j , weobtain all M bounds associated with cycle π . The same can be done for all cycles π ∈ Π with | π | ≥
2, from which we obtain the outer bound in Theorem 2. This concludes the proof.
Here we show that the MISO-BC outer bound in Theorem 2 is tight for 2-cell networks in themc-SLS regime, hence proving the result in Theorem 3. In the 2-cell case, the outer bound D SLSout isdescribed by all GDoF tuples d ∈ R K + with components satisfying¯ d [ l ]1 ≤ α [ l ]11 (54)¯ d [ l ]2 ≤ α [ l ]22 (55)20 d [ l ]1 + ¯ d [ l ]2 ≤ α [ l ]11 − α [ l ]21 + α [ l ]22 (56)¯ d [ l ]1 + ¯ d [ l ]2 ≤ α [ l ]11 + α [ l ]22 − α [ l ]12 (57)for all l , l ∈ (cid:104) L (cid:105) . Recall that ¯ d [ l i ] i denotes (cid:80) l i s i =1 d [ s i ] i . To show that the outer bound D SLSout isachievable in the 2-cell case, we consider a simplified SLS scheme and show that the correspondingachievable GDoF region coincides with the one described in (54)–(57). Before we proceed, wehighlight that the set of all UEs, given by U = (cid:8) ( l i , i ) : l i ∈ (cid:104) L (cid:105) , i ∈ (cid:104) (cid:105) (cid:9) in this part, is partitionedas U ∪ U , where U i denotes the set of UEs in cell i , for i ∈ (cid:104) (cid:105) . In the considered scheme, the message intended to UE-( l i , i ) is split as W [ l i ] i = (cid:0) W [ l i ]s ,i , W [ l i ]c ,i (cid:1) ,where W [ l i ]s ,i is a single-cell sub-message, transmitted from BS- i only; while W c ,i is a multi-cellsub-message, transmitted in a cooperative fashion from both BS-1 and BS-2. Each single-cell sub-message W [ l i ]s ,i is encoded into a codeword X [ l i ]s ,i . On the other hand, the 2 L multi-cell sub-messages (cid:0) W [ l i ]c ,i : ( l i , i ) ∈ U (cid:1) are jointly encoded into the common codeword X c . Codewords are independent,and each is drawn from a Gaussian codebook with unit average power. Moreover, sub-messages W [ l i ]s ,i and W [ l i ]c ,i carry GDoFs of d [ l i ]s ,i and d [ l i ]c ,i respectively; and hence, the total GDoF achieved byUE-( l i , i ) is given by a sum of two contributions as d [ l i ] i = d [ l i ]s ,i + d [ l i ]c ,i . GDoF tuples of single-celland multi-cell sub-messages are given by d s (cid:44) (cid:0) d [ l i ]s ,i : ( l i , i ) ∈ U (cid:1) and d c (cid:44) (cid:0) d [ l i ]c ,i : ( l i , i ) ∈ U (cid:1) respectively, from which a GDoF tuple of user messages is given by d = d s + d c .The transmit signal of BS- i is compose as a superposition of the L designated single-cell code-words and the common codeword. For a single use of the channel, this is given by: X i = √ q c X c + (cid:88) l i ∈(cid:104) L (cid:105) (cid:113) q [ l i ] i X [ l i ]s ,i (58)where q c , q [1] i , . . . , q [ L ] i are power control variables that satisfy q c + (cid:80) l i ∈(cid:104) L (cid:105) q [ l i ] i ≤
1. For the purposeof GDoF analysis, we set the power control variables as follows: q c = 1 L + 1 and q [ l i ] i = 1 L + 1 · P r [ li ] i (59)where the exponents r [ l i ] i ≤
0, for all ( l i , i ) ∈ U , are power control variables on the GDoF scale.The tuple of (GDoF scale) power control variables is given by r (cid:44) (cid:0) r [ l i ] i : ( l i , i ) ∈ U (cid:1) .At the other end of the channel, each UE-( l i , i ) successively decodes X c , X [1]s ,i , . . . , X [ l i ]s ,i , in thisspecific order, while treating (intra- and inter-cell) interference from all remaining codewords as ad-ditional Gaussian noise. Note that in successive decoding, each decoded codeword is reconstructed,and its contribution is removed from the received signal before decoding the following codewords.We further note that X c is decoded by all UEs in the network and it may be useful (in part) foreach of them, depending on the GDoF allocation d c . On the other hand, each codeword X [ l i ]s ,i isdecoded by UEs ( l i , i ) , . . . , ( L, i ) in cell i , and it is intended to UE-( l i , i ) only.For a given power control policy r , the multi-cell sub-messages, carried through the commoncodeword X c , achieve any GDoF tuple d c ∈ R K + with components satisfying (cid:88) ( l i ,i ) ∈U d [ l i ]c ,i ≤ min ( l i ,i ) ∈U (cid:18) α [ l i ] ii − max (cid:26) α [ l i ] ii + max s i ∈(cid:104) L (cid:105) (cid:8) r [ s i ] i (cid:9) , α [ l i ] ij + max s j ∈(cid:104) L (cid:105) (cid:8) r [ s j ] j (cid:9) , (cid:27)(cid:19) + . (60)21n the other hand, single-cell sub-messages achieve d s ∈ R K + with components satisfying d [ l i ]s ,i ≤ min s i ≥ l i (cid:18) α [ s i ] ii + r [ l i ] i − max (cid:26) α [ s i ] ii + max l (cid:48) i >l i (cid:8) r [ l (cid:48) i ] i (cid:9) , α [ s i ] ij + max l j ∈(cid:104) L (cid:105) (cid:8) r [ l j ] j (cid:9) , (cid:27)(cid:19) + . (61)Note that the successive decoding order specified earlier is reflected in (60) and (61). Remark 5.
The simplified SLS scheme presented above is a superposition of a mc-TIN schemeand an overlaying multicast codeword (also known as a common, or public, message). As seen from(59), single-cell codewords, constituting the mc-TIN part, occupy lower power levels, and they aretreated as noise by other-cell users. On the other hand, the multi-cell multicast codeword occupieshigher power levels, and it is decoded by all users in the network. This scheme is a special case ofthe more general mc-SLS scheme, described in Section 2.3.2. On the other hand, the scheme mayalso be seen as a multi-cell extension (or generalization) of the rate-splitting schemes in [9, 31]. Inwhat follows, we refer to d s as the TIN contribution, and to d c as the multicast contribution. ♦ For a fixed power control policy r , the set of all tuples d c that satisfy (60) is denoted by D Mul ( r );while D TIN ( r ) denotes the set of all tuples d s that satisfy (61). The GDoF region achievable throughthe proposed scheme is hence given by all tuples d , where d = d c + d s for some d c ∈ D Mul ( r ), d s ∈ D TIN ( r ), and r ≤ . Denoting this achievable region by D SLS , it follows that D SLS = (cid:91) r ≤ D Mul ( r ) ⊕ D TIN ( r ) (62)where ⊕ is the Minkowski sum (or vector sum) operation (see Section 1.2). Achievable GDoF tuples in (62) are highly coupled with auxiliary design variables. For instance,each d ∈ D SLS is determined by a multicast contribution d c and a TIN contribution d s , which inturn depend on the power control tuple r . Together, these auxiliary variables are jointly optimizedto achieve different GDoF trade-offs (or tuples d ). On the other hand, the outer bound D MBCout is described in terms of fixed channel parameters only (i.e. α ) with no auxiliary variables, asseen in (54)–(57). A direct comparison between D MBCout and D SLS can be facilitated by eliminatingthe auxiliary variables in the latter, which is typically accomplished by means of Fourier-Motzkin(FM) elimination. Nevertheless, the intricate structure of D SLS and the high number of auxiliaryvariables prohibit a direct application of the FM procedure in this case. To circumvent this technicalchallenge, we devise a sequence of reductions, which are summarized as follows: • We first restrict the space of admissible power control variables r , which in turn allows usto obtain simpler inner bounds for D Mul ( r ) and D TIN ( r ), denoted by D Mul ( a ) and D PTIN ( a ),respectively. These inner bounds depend on a single power control variable a only. As we willsee, D PTIN ( a ) is obtained through a new application of the potential graph approach [21], astructured FM procedure tailored for GDoF regions achievable through TIN. • We then evaluate the Minkowski sum D Mul ( a ) ⊕ D PTIN ( a ), for any feasible a , in terms ofinequalities that bound d only by eliminating the auxiliary GDoF variables d c and d s . Thisis accomplished by exploiting polymatroid properties of these inner bounds, which allows usto utilize a result on the Minkowski sum of polymatroids [45, Th. 44.6]. At this point, weobtain an inner bound for D SLS with a as the only remaining auxiliary variable. With a slight abuse of notation, we denote the achievable region here by D SLS . Strictly speaking, however, theachievable region on the right-hand-side of (62) is included in the SLS region D SLS , described in Section 2.3.2. We then eliminate the remaining variable a via as second round of FM elimination. Remark-ably, the region obtained after this step matches the outer bound in (54)–(57).The remainder of this section is dedicated to explaining the above steps in more detail. As a first simplifying step, we restrict the space of feasible power control variables r such that forboth i ∈ (cid:104) (cid:105) , we have the following order r [ L ] i ≤ r [ L − i ≤ · · · ≤ r [1] i ≤ − a (63)where 0 ≤ a ≤ max i,j,l i : i (cid:54) = j (cid:8) α [ l i ] ij (cid:9) . (64)As it turns out, imposing the power allocation order in (63) incurs no loss of generality in the regimesof interest. This is due to the SIR order in (7), which holds in the mc-SLS regime, requiring higherpower levels for weaker users. On the other hand, it is sufficient to consider values of a which areno greater than the strongest interfering link as in (64), since this level of attenuation guaranteesthat all inter-cell interference from the mc-TIN scheme is received below the noise level.Given the power allocation order in (63), we may bound the right-hand-side of (60) below asmin ( l i ,i ) ∈U (cid:18) α [ l i ] ii − max (cid:26) α [ l i ] ii + max s i ∈(cid:104) L (cid:105) (cid:8) r [ s i ] i (cid:9) , α [ l i ] ij + max s j ∈(cid:104) L (cid:105) (cid:8) r [ s j ] j (cid:9) , (cid:27)(cid:19) + ≥ min ( l i ,i ) ∈U (cid:16) α [ l i ] ii − max (cid:110) α [ l i ] ii − a, α [ l i ] ij − a, (cid:111)(cid:17) + (65)= min ( l i ,i ) ∈U (cid:16) min (cid:110) a, α [ l i ] ii − α [ l i ] ij + a, α [ l i ] ii (cid:111)(cid:17) + (66)= a (67)where the equality in (67) holds due to a ≥ α [ l i ] ii ≥ a and α [ l i ] ii ≥ α [ l i ] ij . It follows that D Mul ( r )includes the set of GDoF tuples d c ∈ R K + with components satisfying (cid:88) ( l i ,i ) ∈U d [ l i ]c ,i ≤ a. (68)For any a satisfying (64), and with a slight abuse of notation, we denote the set of GDoF tuplessatisfying (68) as D Mul ( a ). This leads to an inner bound for D SLS in (62), given by D SLS ⊇ (cid:91) a (cid:91) r ∈R ( a ) D Mul ( a ) ⊕ D TIN ( r )= (cid:91) a D Mul ( a ) ⊕ (cid:91) r ∈R ( a ) D TIN ( r ) (69)where R ( a ) is the set of power control tuples r satisfying (63) for fixed a , while the union withrespect to a is taken over the interval specified in (64). The equality in (69) holds due to the factthat ( A ⊕ B ) ∪ ( A ⊕ B ) = A ⊕ ( B ∪ B ). Next, we turn to obtaining a simplified inner bound forthe TIN region ∪ r ∈R ( a ) D TIN ( r ), for any fixed a in the designated interval (64).The TIN region D TIN ( r ) includes a smaller region known as the polyhedral-TIN region, denotedby D PTIN ( r ), obtained by relaxing the ( · ) + operation on the right-hand-side of (61), hence posing23urther restrictions on admissible power control tuples—see, e.g., [21, 30]. Under the order in (63), D PTIN ( r ) is described by the set of GDoF tuples d s ∈ R K + with components satisfying d [ l i ]s ,i ≤ min s i ≥ l i (cid:26) α [ s i ] ii + r [ l i ] i − max (cid:26) α [ s i ] ii + max l (cid:48) i >l i (cid:8) r [ l (cid:48) i ] i (cid:9) , α [ s i ] ij + max l j ∈(cid:104) L (cid:105) (cid:8) r [ l j ] j (cid:9) , (cid:27)(cid:27) (70)= min s i ≥ l i (cid:26) r [ l i ] i − max l (cid:48) i >l i (cid:8) r [ l (cid:48) i ] i (cid:9) , α [ s i ] ii − α [ s i ] ij + r [ l i ] i − max l j ∈(cid:104) L (cid:105) (cid:8) r [ l j ] j (cid:9) , α [ s i ] ii + r [ l i ] i (cid:27) (71)= min s i ≥ l i (cid:110) r [ l i ] i − r [ l i +1] i , α [ s i ] ii − α [ s i ] ij + r [ l i ] i − r [1] j , α [ s i ] ii + r [ l i ] i (cid:111) (72)= min (cid:110) r [ l i ] i − r [ l i +1] i , α [ l i ] ii − α [ l i ] ij + r [ l i ] i − r [1] j , α [ l i ] ii + r [ l i ] i (cid:111) (73)where in the above, we set r [ L +1] i = −∞ . Note that (72) is obtained from (71) by invoking thepower allocation order in (63), while the equality in (73) holds due to the SIR order in (7) and theSNR order in (6), which imply α [ l i ] ii − α [ l i ] ij ≤ α [ s i ] ii − α [ s i ] ij and α [ l i ] ii ≤ α [ s i ] ii , for all s i ∈ (cid:104) l i : L (cid:105) . Fora given a satisfying (64), and with another slight abuse of notation, we use D PTIN ( a ) to denote ∪ r ∈R ( a ) D PTIN ( r ). As it turns out, the region D PTIN ( a ) lends itself to an efficient FM eliminationprocedure using the potential graph approach [21], yielding the following result. Lemma 3.
For any ≤ a ≤ max i,j,l i : i (cid:54) = j (cid:8) α [ l i ] ij (cid:9) , the polyehdral TIN region D PTIN ( a ) is equal tothe region described by all tuples d s ∈ R K + with components satisfying ¯ d [ l i ]s ,i ≤ α [ l i ] ii − a (74)¯ d [ l i ]s ,i + ¯ d [ l j ]s ,j ≤ α [ l i ] ii + α [ l j ] jj − max (cid:110) α [ l i ] ij + a, α [ l j ] ji + a, a, α [ l i ] ij + α [ l j ] ji (cid:111) (75) for all i, j ∈ (cid:104) (cid:105) , i (cid:54) = j , and l i , l j ∈ (cid:104) L (cid:105) . A detailed proof of Lemma 3 is relegated to Appendix A. After the elimination of r in Lemma3, we now have an inner bound for D SLS given by D SLS ⊇ (cid:91) a D Mul ( a ) ⊕ D PTIN ( a ) . (76)This inner bound does not depend on r , yet it is still characterized in terms of a , d c and d s .Next, we eliminate the auxiliary GDoF tuples d c and d s by characterizing the Minkowski sum D Mul ( a ) ⊕ D PTIN ( a ) in terms of inequalities that bound d = d c + d s , for any fixed a . To this end,it is easier to work with an inner bound for D PTIN ( a ), denoted by D PTIN (cid:48) ( a ), given by¯ d [ l i ]s ,i ≤ α [ l i ] ii − max (cid:110) α [ l i ] ij , a (cid:111) (77)¯ d [ l i ]s ,i + ¯ d [ l j ]s ,j ≤ α [ l i ] ii + α [ l j ] jj − max (cid:110) α [ l i ] ij + a, α [ l j ] ji + a, a, α [ l i ] ij + α [ l j ] ji (cid:111) . (78)Note that D PTIN ( a ) and D PTIN (cid:48) ( a ) are almost identical, with the exception that (77) in the latter istighter than the corresponding inequality (74) in the former. The rationale behind this tighteningwill become clear further on. We now observe that D Mul ( a ) may be described as¯ d [ l i ]c ,i ≤ a (79)¯ d [ l i ]c ,i + ¯ d [ l j ]c ,j ≤ a (80)for all i, j ∈ (cid:104) (cid:105) , i (cid:54) = j , and l i , l j ∈ (cid:104) L (cid:105) . This is obtained directly from (68) by including redundantbounds so that the linear inequalities describing D Mul ( a ) in (79) and (80) are of the same type asthose used to describe D PTIN (cid:48) ( a ) in (77) and (78). This leads us to the following result.24 emma 4. For any ≤ a ≤ max i,j,l i : i (cid:54) = j (cid:8) α [ l i ] ij (cid:9) , the Minkowski sum D Mul ( a ) ⊕ D PTIN (cid:48) ( a ) ischaracterized by all tuples d ∈ R K + with components satisfying ¯ d [ l i ] i ≤ α [ l i ] ii − max (cid:110) α [ l i ] ij − a, (cid:111) (81)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii + α [ l j ] jj − max (cid:110) α [ l i ] ij , α [ l j ] ji , a, α [ l i ] ij + α [ l j ] ji − a (cid:111) (82) for all i, j ∈ (cid:104) (cid:105) , i (cid:54) = j , and l i , l j ∈ (cid:104) L (cid:105) . It is perhaps clear from Lemma 4 that the linear inequalities that describe D Mul ( a ) ⊕ D PTIN (cid:48) ( a )are simply the direct sums of the corresponding inequalities describing the constituent polyhedra D Mul ( a ) and D PTIN (cid:48) ( a ). That is, (81) is obtained by adding (77) and (79); and (82) is obtained byadding (78) and (80). Since summing inequalities loosens them in general, one can directly concludethat D Mul ( a ) ⊕ D PTIN (cid:48) ( a ) is included in, yet not necessarily equal to, the region described by (81)and (82) in Lemma 4. Remarkably, it turns out that in this special case D Mul ( a ) ⊕ D PTIN (cid:48) ( a ) isequal to the polyhedron described by (81) and (82). This holds since both D Mul ( a ) and D PTIN (cid:48) ( a )are polymatroids (see Appendix B); and polymatroids have the nice property that their Minkowskisums are given by the direct sums of their corresponding linear inequalities [45, Th. 44.6] (seealso [46, Th. 3]). This point, and the proof of Lemma 4, are discussed in detail in Appendix B.In light of the above, going from D PTIN ( a ) to the smaller region D PTIN (cid:48) ( a ) can be explained. Inparticular, this guarantees that the set function associated with the polyhedron in (77) and (78),defined over the ground set U , is non-decreasing. This monotonicity, alongside submodularity,imply that the region D PTIN (cid:48) ( a ) is a polymatroid, as shown in Appendix B.Building upon the result in Lemma 4, it follows hat the inner bound given by (cid:91) a D Mul ( a ) ⊕ D PTIN (cid:48) ( a ) (83)is described by all tuples d ∈ R K + that satisfy¯ d [ l i ] i ≤ α [ l i ] ii − max (cid:110) α [ l i ] ij − a, (cid:111) (84)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii + α [ l j ] jj − max (cid:110) α [ l i ] ij , α [ l j ] ji , a, α [ l i ] ij + α [ l j ] ji − a (cid:111) (85)0 ≤ a ≤ max i,j,l i : i (cid:54) = j α [ l i ] ij . (86)for all i, j ∈ (cid:104) (cid:105) , i (cid:54) = j , and l i , l j ∈ (cid:104) L (cid:105) . Now it remains to eliminate the last auxiliary variable a .This is accomplished by a standard application of the FM procedure [47, Appendix D].To this end, we classify the inequalities in (84)–(86) with respect to the presence and sign ofthe variable a on the right-hand-side. We have the three following classes: • Inequalities with no a : ¯ d [ l i ] i ≤ α [ l i ] ii (87)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii + α [ l j ] jj − max (cid:110) α [ l i ] ij , α [ l j ] ji (cid:111) . (88) • Inequalities with + a : ¯ d [ l i ] i ≤ α [ l i ] ii − α [ l i ] ij + a (89)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii + α [ l j ] jj − (cid:16) α [ l i ] ij + α [ l j ] ji (cid:17) + a (90)0 ≤ a. (91)25 Inequalities with − a : ¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii + α [ l j ] jj − a (92)0 ≤ max i,j,l i : i (cid:54) = j α [ l i ] ij − a. (93)To eliminate a , we add each inequality with + a to every inequality with − a . We start by adding anarbitrary inequality from (92), given by ¯ d [ l (cid:48) i ] i + ¯ d [ l (cid:48) j ] j ≤ α [ l (cid:48) i ] ii + α [ l (cid:48) j ] jj − a , to all inequalities in (89)–(91).We obtain the following set of inequalities¯ d [ l (cid:48) i ] i + ¯ d [ l (cid:48) j ] j + ¯ d [ l i ] i ≤ α [ l (cid:48) i ] ii + α [ l (cid:48) j ] jj + α [ l i ] ii − α [ l i ] ij (94)¯ d [ l (cid:48) i ] i + ¯ d [ l (cid:48) j ] j + ¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l (cid:48) i ] ii + α [ l (cid:48) j ] jj + α [ l i ] ii + α [ l j ] jj − (cid:16) α [ l i ] ij + α [ l j ] ji (cid:17) (95)¯ d [ l (cid:48) i ] i + ¯ d [ l (cid:48) j ] j ≤ α [ l (cid:48) i ] ii + α [ l (cid:48) j ] jj . (96)Next, we show that these resulting inequalities in (94)–(96) are all redundant, as they are impliedby the set of inequalities in (87) and (88). In particular, from (87) and (88), we obtain¯ d [ l (cid:48) i ] i ≤ α [ l (cid:48) i ] ii (97)¯ d [ l (cid:48) j ] j ≤ α [ l (cid:48) j ] ii (98)¯ d [ l i ] i + ¯ d [ l (cid:48) j ] j ≤ α [ l i ] ii + α [ l (cid:48) j ] jj − α [ l i ] ij (99)¯ d [ l (cid:48) i ] i + ¯ d [ l j ] j ≤ α [ l (cid:48) i ] ii + α [ l j ] jj − α [ l j ] ji (100)from which it is clear that (94) is implied by (97) and (99), i.e. by summing them; (95) is impliedby (99) and (100); and (96) is implied by (97) and (98).We move on to the next step of the FM elimination, where we add the inequality in (93) to allinequalities in (89)–(91), from which we obtain¯ d [ l i ] i ≤ α [ l i ] ii − α [ l i ] ij + max i,j,l i : i (cid:54) = j α [ l i ] ij (101)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii + α [ l j ] jj − (cid:16) α [ l i ] ij + α [ l j ] ji (cid:17) + max i,j,l i : i (cid:54) = j α [ l i ] ij (102)0 ≤ max i,j,l i : i (cid:54) = j α [ l i ] ij . (103)It is evident that (101) and (102) are redundant, as they are implied by (87) and (88) respectively.Moreover, (103) holds by definition of channel strength parameters.From the above, it follows that all inequalities obtained by eliminating a in (89)–(93) areredundant with respect to (87) and (88). Therefore, the inner bound in (83) is equal to the regionspecified by all GDoF tuples d ∈ R K + that satisfy¯ d [ l i ] i ≤ α [ l i ] ii (104)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii + α [ l j ] jj − max (cid:110) α [ l i ] ij , α [ l j ] ji (cid:111) (105)for all i, j ∈ (cid:104) (cid:105) , i (cid:54) = j , and l i , l j ∈ (cid:104) L (cid:105) . This exactly matches the 2-cell outer bound in (54)–(57),which in turn concludes the proof of Theorem 3.26 Extremal Gains of Multi-Cell Cooperation over Multi-Cell TIN
In this section, we present a proof for Theorem 4. The main idea of the proof is to exploit theredundancy of weaker UEs in each cell of the network, exhibited in the regimes of interest due tothe SIR order in (7), which allows us to form a direct relationship between K × KL networks and K × K networks. Thins in turn enables us to utilize previous extremal results in [40].Similar to the sum-GDoF definitions in (24) and (25), we define a sum-GDoF outer bound forthe MISO-BC in the mc-SLS regime from the outer bound region in Theorem 2 as d MBCout , Σ ( α ) (cid:44) max d ∈D SLSout ( α ) (cid:88) ( l k ,k ) ∈U d [ l k ] k . (106)Moreover, for an arbitrary network α ∈ R K × K × L + , we use α [ L ] ∈ R K × K + to denote the sub-networkobtained by keeping the strongest UE (in the SNR or SIR sense) in each cell, and eliminating allremaining UEs. Therefore, the ( i, j )-th element of α [ L ] is given by α [ L ] ij . To avoid confusion, weuse α exclusively to denote K × KL networks, while K × K networks are denoted by α [ L ] . Witha slight abuse of notation, a regime A is automatically adjusted to the dimensions of the networkunder consideration, i.e. α [ L ] ∈ A implies A ⊂ R K × K + , while α ∈ A implies A ⊂ R K × KL + .We now recall the main results in [40], which are analog to Theorem 4 but for K × K networks. Theorem 5. ( [40, Th. 5.1, 6.1 and 7.1] )
The extremal GDoF gains of transmitter cooperationover TIN in the three regimes of interest for K × K networks are as follows: max α [ L ] ∈A d MBCΣ ( α [ L ] ) d TINΣ ( α [ L ] ) = η K, ( A ) = , A = A TIN − K , A = A CTIN Θ (cid:0) log( K ) (cid:1) , A = A SLS . (107)In each regime of interest A , drawn from {A TIN , A CTIN , A SLS } , the authors in [40] identify anetwork α (cid:63) [ L ] ∈ A (or a class of networks) for which the following lower bound holds d MBCΣ ( α (cid:63) [ L ] ) ≥ η K, ( A ) d TINΣ ( α (cid:63) [ L ] ) (108)where η K, ( A ) takes on values as given in (107). This lower bound is combined with a matchingupper bound, derived using an analog of the sum-GDoF outer bound in (106), but specialized to K × K networks. This matching upper bound is given by d MBCout , Σ ( α [ L ] ) ≤ η K, ( A ) d TINΣ ( α [ L ] ) (109)which is shown to hold for all networks α [ L ] ∈ A , in each of the regimes of interest. Next, thesebounds for K × K networks are utilized to derive similar bounds for K × KL networks. In whatfollows, we fix an arbitrary regime A drawn from {A TIN , A CTIN , A SLS } . Remark 6.
It is worthwhile highlighting that the upper bound in (109) does not appear in thisexplicit form in [40], yet it can be inferred from the proofs of Theorems 5.1, 6.1 and 7.1, presentedrespectively in Sections V.A, VI.A and VII.A of the same paper. It is shown that the inequality d MBCΣ ( α [ L ] ) ≤ η K, ( A ) d TINΣ ( α [ L ] ) holds for the three regimes of interest by bounding d MBCΣ ( α [ L ] )above using analogs of the MISO-BC cycle bounds in Definition 6, specialized to K × K networks.This is equivalent to bounding d MBCΣ ( α [ L ] ) above by d MBCout , Σ ( α [ L ] ), obtained from the outer boundregion D SLSout ( α [ L ] ) in Theorem 2. This in turn allows us to write the upper bound in (109).27 .1 Lower Bound We first show that a lower bound on the extremal gain in Theorem 4 give bymax α ∈A d MBCΣ ( α ) d TINΣ ( α ) ≥ η K, ( A ) (110)is easily obtained from Theorem 5. Let α (cid:63) [ L ] be a K × K network that attains η K, ( A ), that is α (cid:63) [ L ] = arg max α [ L ] ∈A d MBCΣ ( α [ L ] ) d TINΣ ( α [ L ] ) . (111)Now let α (cid:63) be a K × KL network that includes α (cid:63) [ L ] as a sub-network, and which is obtainedby adding L − d MBCΣ ( α (cid:63) ) = d MBCΣ ( α (cid:63) [ L ] ) and d TINΣ ( α (cid:63) ) = d TINΣ ( α (cid:63) [ L ] ). This directly leads tomax α ∈A d MBCΣ ( α ) d TINΣ ( α ) ≥ d MBCΣ ( α (cid:63) ) d TINΣ ( α (cid:63) ) = d MBCΣ ( α (cid:63) [ L ] ) d TINΣ ( α (cid:63) [ L ] ) = η K, ( A ) . (112)Having obtained a lower bound, it now remains to prove a matching upper bound. We first observe that for any network α , we have d TINΣ ( α ) = d TINΣ ( α [ L ] ) . (113)In other words, starting from a network α [ L ] with single-user cells, including additional (weaker)users in each cell does not increase the sum-GDoF achieved using TIN. To see this, we first notethat d TINΣ ( α ) ≥ d TINΣ ( α [ L ] ) evidently holds, since α [ L ] is a sub-network of α , and the TIN schemeis a special case of mc-TIN scheme. The other direction, i.e. d TINΣ ( α ) ≤ d TINΣ ( α [ L ] ), holds byconstruction of the mc-TIN scheme. All L messages of cell k are decoded by UE-( L, k ), which inturn bounds the achievable GDoF in each cell k by the achievable GDoF of UE-( L, k ).Next, we make a somewhat similar observation to the one above, but in the context of theMISO-BC GDoF outer bound. In particular, we observe that for any network α , we have d MBCout , Σ ( α ) ≤ d MBCout , Σ ( α [ L ] ) (114)To see this, we first observe that the outer bound D SLSout ( α ) in Theorem 2 is included in the region¯ D SLSout ( α ), described by all tuples d ∈ R KL + that satisfy (cid:88) k ∈{ σ } ¯ d k ≤ ∆ + π,m , ∀ m ∈ (cid:104)| π |(cid:105) , π ∈ Π , { π } = { ( L, σ (1)) , . . . , ( L, σ ( | π | )) } . (115)This holds as (115) is obtained from (23) by removing some of the cycle bounds in the latter, andonly keeping the bounds that involve all L users in each participating cell of any cycle π . Denotingthe sum-GDoF outer bound obtained from ¯ D MBCout ( α ) in (115) as ¯ d MBCout , Σ ( α ), we clearly have d MBCout , Σ ( α ) ≤ ¯ d MBCout , Σ ( α ) . (116)28ote that ¯ D MBCout ( α ) in (115) may be viewed as a GDoF region for a K × K network, with GDoFtuples given by ¯ d (cid:44) (cid:0) ¯ d k : k ∈ (cid:104) K (cid:105) (cid:1) ∈ R K + . From this observation, it can be easily verified that¯ D MBCout ( α ) in (115) coincides with D MBCout ( α [ L ] ), and therefore we have¯ d MBCout , Σ ( α ) = d MBCout , Σ ( α [ L ] ) . (117)By combining (116) and (117), the inequality in (114) is obtained.Next, we employ (113) and (114) to obtain the followingmax α ∈A d MBCΣ ( α ) d TINΣ ( α ) ≤ max α ∈A d MBCΣ , out ( α ) d TINΣ ( α ) (118) ≤ max α ∈A d MBCΣ , out ( α [ L ] ) d TINΣ ( α [ L ] ) . (119)Finally, combining (119) with the upper bound in (109), we obtain the desired upper boundmax α ∈A d MBCΣ ( α ) d TINΣ ( α ) ≤ η K, ( A ) . (120)This completes the proof of Theorem 4. We studied the GDoF of downlink cellular networks with finite precision CSIT, modeled by the IBCunder no BS cooperation and the overloaded MISO-BC under full BS cooperation; while focusingon three weak inter-cell interference regimes of interest, given (in a monotonically increasing order)by: the mc-TIN regime, mc-CTIN regime, and mc-SLS regime. Through a new application of AIbounds to K × KL cellular-type networks, we derived outer bounds for the K × KL IBC and the K × KL MISO-BC under finite precision CSIT. The former outer bound is utilized to show thatthe mc-TIN scheme is GDoF optimal for the K × KL IBC under finite precision CSIT in the entiremc-CTIN regime; while the latter is combined with a non-trivial achievability argument to showthat the mc-SLS scheme is GDoF optimal for the 2 × L MISO-BC in the mc-SLS regime. Wethen leveraged the recently introduced extremal network analysis framework to study the GDoFgain from mc-Co over mc-TIN in large asymmetric cellular networks. Our analysis reveals that theextremal GDoF gain from mc-Co over mc-TIN is limited to (small) constant factors in the mc-TINand mc-CTIN regimes, and scales logarithmically with the number of cells K in the mc-SLS regime.These results mirror and generalize recent results for K × K networks by Chan et al. [40].The results presented in this paper give rise to a number of interesting questions, which remainunanswered. For instance, it is not clear whether the logarithmic extremal GDoF gain in the mc-SLS regime is a fundamental gain of the MISO-BC over the IBC, or rather an artefact of limitingthe IBC to mc-TIN schemes. Will this gain diminish if we replace the mc-TIN scheme with aGDoF-optimal robust scheme for the IBC in the mc-SLS? Such robust scheme for the IBC inthe mc-SLS regime will most likely rely on layered superposition and inter-cell rate-splitting withoverwhelmingly many parameters and optimization variables, rendering extremal network analysisall the more essential in this case. Another intriguing direction is to study extremal GDoF gainsof the MISO-BC over the IBC beyond the mc-SLS regime, e.g. in the general weak inter-cellinterference regime specified in (7). Moreover, while the extremal network analysis carried out inthis paper focuses on the sum-GDoF viewpoint, it is of interest to investigate extremal gains from29he perspective of fair utility functions, such as the symmetric-GDoF. In this case, switching offweaker users and reducing K × KL networks to K × K networks may not be possible any more;and the cell-edge benefits of robust BS cooperation may be better reflected. A Proof of Lemma 3
Here we show that the polyhedral TIN region D PTIN ( a ), given by all tuples d s ∈ R K + that satisfy(73) for some r ∈ R ( a ), is identical to the region specified by (74) and (75). To this end, we firstnote from (63) and (73) that D PTIN ( a ) may be described by all tuples d ∈ R K + that satisfy: d [ l i ] i ≤ r [ l i ] i − r [ l i +1] i (121) d [ l i ] i − (cid:0) α [ l i ] ii − α [ l i ] ij (cid:1) ≤ r [ l i ] i − r [1] j (122) d [ l i ] i − α [ l i ] ii ≤ r [ l i ] i (123)for all i, j ∈ (cid:104) (cid:105) , i (cid:54) = j , and l i , l j ∈ (cid:104) L (cid:105) , where0 ≤ r [ l i ] i − r [ l i +1] i (124) a ≤ − r [1] i . (125)In the above, we set r [ L +1] i = −∞ as in (73). Note that in (121)–(123) and the remainder of thisappendix, we drop the subscript in d s , and we simply used d instead (as no confusion may arisedue to this). Moreover, it is worthwhile noting that inequalities as the one in (124) are in factredundant, as they are implied by (121) and the non-negativity of GDoF components.Next, we construct a directed graph (digraph) known as the potential graph [21], through whichwe can carry out an efficient FM elimination of the power control variables r in (121)–(125). A.1 Potential Graph
We define the potential graph associated with the 2-cell network of interest as G = ( V , E ), where V is a set of 2 L + 1 vertices, while E is a set of directed edges. The set of vertices is given by V (cid:44) (cid:8) u (cid:9) ∪ (cid:8) v [ l i ] i : ( l i , i ) ∈ U (cid:9) comprising a vertex v [ l i ] i for each UE-( l i , i ), and a ground vertex u . The set of directed edges E isgiven by the union E ∪ E ∪ E ∪ E , where constituent subsets are defined as follows: E = (cid:110) ( v [ l i ] i , v [ l i +1] i ) : i ∈ (cid:104) (cid:105) , l i ∈ (cid:104) L − (cid:105) (cid:111) E = (cid:110) ( v [ l i ] i , v [1] j ) : ( l i , i ) ∈ U , j ∈ (cid:104) (cid:105) , i (cid:54) = j (cid:111) E = (cid:110) ( v [ l i ] i , u ) : ( l i , i ) ∈ U (cid:111) E = (cid:110) ( u, v [1] i ) : i ∈ (cid:104) (cid:105) (cid:111) . It is worthwhile noting that G is not a complete digraph, e.g. vertex v [ l i +1] i with l i ∈ (cid:104) L − (cid:105) mayonly be reached through its preceding vertex v [ l i ] i of the same cell—see Fig. 4. This reflects thesuccessive decoding order of the mc-TIN scheme as described in Section 6.1, as well as the power30 !! 𝑣 !" 𝑣 "! 𝑣 "" 𝑢 UE − UE − BS-1 UE − UE − BS-2
Figure 4: Left: 2-cell network with 2 users per-cell. Right: The corresponding potential graph G ,with subsets of edges given by E in red, E in blue, E in orange, and E in green.allocation order in (63). Note that this incompleteness of the potential graph in the multi-cellsetting is a key difference to the potential graph in [21], constructed for the interference channel.Next, we assign the following lengths to edges from the four subsets defined above: l ( v [ l i ] i , v [ l i +1] i ) = − d [ l i ] i (126) l ( v [ l i ] i , v [1] j ) = α [ l i ] ii − α [ l i ] ij − d [ l i ] i (127) l ( v [ l i ] i , u ) = α [ l i ] ii − d [ l i ] i (128) l ( u, v [1] i ) = − a. (129)As these lengths clearly depend on d , we parametrize G by d in what follows, and we write G ( d ).A potential associated with G ( d ) is a function p : V → R such that for any v (cid:48) , v (cid:48)(cid:48) ∈ V , we have l ( v (cid:48) , v (cid:48)(cid:48) ) ≥ p ( v (cid:48)(cid:48) ) − p ( v (cid:48) ) . These inequalities only depend on differences of potentials, and hence we may set the potential ofthe ground node as p ( u ) = 0, without any loss in generality. It follows that p must satisfy − d [ l i ] i ≥ p ( v [ l i +1] i ) − p ( v [ l i ] i ) (130) α [ l i ] ii − α [ l i ] ij − d [ l i ] i ≥ p ( v [1] j ) − p ( v [ l i ] i ) (131) α [ l i ] ii − d [ l i ] i ≥ − p ( v [ l i ] i ) (132) − a ≥ p ( v [1] i ) (133)for all i, j ∈ (cid:104) (cid:105) , i (cid:54) = j , and l i , l j ∈ (cid:104) L (cid:105) . Setting p ( v [ l i ] i ) = r [ l i ] i , for all ( l i , i ) ∈ U , we notice that theset of inequalities for potentials in (130)–(133) is identical to the set of inequalities involving powercontrol variables in (121)–(125). Therefore, it holds that: for any GDoF tuple d ∈ D PTIN ( a ) , apower control tuple r is feasible and achieves d , i.e. by satisfying the set of inequalities in (121) – (125) , if an only if it is a valid potential for the digraph G ( d ) . We are now ready to invoke the potential theorem [45, Th. 8.2], which states that: there existsa potential function for G ( d ) if and only if each directed circuit in G ( d ) has a non-negative length .Therefore, we may conclude that for any d ∈ D PTIN ( a ), there exists a feasible power control tuple Recall that (124) in the set of inequalities (121)–(125) is redundant. that achieves d if and only if each directed circuit in G ( d ) has a non-negative length. It remainsto interpret non-negative lengths of directed circuits in terms of GDoF inequalities. To this end, adirected circuit is represented as ( v → v → · · · → v M ), where v m ∈ V such that ( v m , v m +1 ) ∈ E ,for all m ∈ (cid:104) M (cid:105) and v M +1 = v . The length of such circuit is given by (cid:80) Mm =1 l ( v m , v m +1 ). A.2 From Directed Circuits to GDoF Inequalities
We start with single-cell circuits, involving users from one cell only. These circuits take the form (cid:0) v [1] i → v [2] i → · · · → v [ l i ] i → u (cid:1) (134)for some ( l i , i ) ∈ U . From the non-negative circuit length condition, it can be easily checked usingthe lengths in (126)–(129) that a circuit of the form in (134) gives rise to a GDoF bound given by¯ d [ l i ] i ≤ α [ l i ] ii − a. (135)Recall that ¯ d [ l i ] i is a shorthand notation for the sum (cid:80) l i s i =1 d [ s i ] i .Next, we move on to multi-cell circuits that involve users from both cells. For any pair of users( l i , i ) , ( l j , j ) ∈ U , where i (cid:54) = j , we have the four following types of circuits: (cid:0) v [1] i → v [2] i · · · → v [ l i ] i → u → v [1] j → v [2] j · · · → v [ l j ] j (cid:1)(cid:0) v [1] i → v [2] i · · · → v [ l i ] i → v [1] j → v [2] j · · · → v [ l j ] j → u (cid:1)(cid:0) v [1] i → v [2] i · · · → v [ l i ] i → u → v [1] j → v [2] j · · · → v [ l j ] j → u (cid:1)(cid:0) v [1] i → v [2] i · · · → v [ l i ] i → v [1] j → v [2] j · · · → v [ l j ] j (cid:1) . It can be verified that the above circuits give rise o the following GDoF inequalities:¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii − a + α [ l j ] jj − α [ l j ] ji (136)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii − α [ l i ] ij + α [ l j ] jj − a (137)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii − a + α [ l j ] jj − a (138)¯ d [ l i ] i + ¯ d [ l j ] j ≤ α [ l i ] ii − α [ l i ] ij + α [ l j ] jj − α [ l j ] ji . (139)It is evident that together with the non-negativity of the GDoF, the inequalities in (135) and (136)–(139) define a GDoF region identical to the one defined in (74) and (75) in Lemma 3. Therefore,after eliminating the power control variables r , the polyhedral TIN region D PTIN ( a ) does indeedreduce to the one in Lemma 3, which concludes the proof. B Proof of Lemma 4
The result in Lemma 4 is a consequence of D Mul ( a ) and D PTIN (cid:48) ( a ) both being polymatroids, andthe direct summability property of polymatroids given in [45, Th. 44.6] (see also [46, Th. 3]), asshown in detail next. Before we proceed, it is worthwhile recalling that the set of UEs in the 2-cellsetting of interest is given by U = U ∪ U , where U i (cid:44) { ( l i , i ) : l i ∈ (cid:104) L (cid:105)} for all i ∈ (cid:104) (cid:105) .First, we observe that D Mul ( a ), described in (79) and (80), can be equivalently expressed in apolymatroid-like fashion as all tuples d c ∈ R K + that satisfy d c ( S ) ≤ a · ( S (cid:54) = ∅ ) , ∀S ⊆ U (140)32here d c ( S ) is just a shorthand notation for the sum (cid:80) ( l i ,i ) ∈S d [ l i ]c ,i . This follows as the additionalinequalities in (140), not included in (79) and (80), are redundant. It can be easily checked that(140) is a polymatroid—note that the associated set function is normalized (since d c ( ∅ ) = 0),non-decreasing and submodular (since it is constant for all S (cid:54) = ∅ ).Next, we move on to the region D PTIN (cid:48) ( a ), described in (77) and (78), and we express it in asimilar polymatroid-like fashion as all tuples d s ∈ R K + that satisfy d s ( S ) ≤ f ( S ) , ∀S ⊆ U . (141)In the above, f : 2 U → R + is a set function associated with the polyhedron D PTIN (cid:48) ( a ), where 2 U denotes the power set over U . This set function f is given by f ( S ) (cid:44) , S = ∅ α [ s ]11 − max (cid:8) α [ s ]12 , a (cid:9) , S (cid:54) = ∅ , S ∩ U = ∅ α [ s ]22 − max (cid:8) α [ s ]21 , a (cid:9) , S (cid:54) = ∅ , S ∩ U = ∅ α [ s ]11 + α [ s ]22 − max (cid:8) α [ s ]12 + a, a + α [ s ]21 , a, α [ s ]12 + α [ s ]21 (cid:9) , S ∩ U (cid:54) = ∅ , S ∩ U (cid:54) = ∅ , (142)where we define s i (cid:44) max ( l i ,i ) ∈U i ∩S l i , ∀ i ∈ (cid:104) (cid:105) . It is readily seen from (142) that for any
S ⊆ U , the value of f ( S ) depends on the pair of UEs ( s , s ,
2) only, or more specifically, on the user indices s and s as defined in (142). Therefore,and with a slight abuse of notation, f ( S ) is written as f ( s , s ) whenever convenient.As mentioned above, the region D PTIN (cid:48) ( a ), described in (77) and (78), is equivalently representedby (141). This holds since the inequalities in (141) include those in (77) and (78), while theremaining inequalities in (141), which are not included in (77) and (78), are redundant. This isseen by observing that any inequality d s ( S ) ≤ f ( S ) = f ( s , s ), where S = S ∪ S and S i ⊆ U i for all i ∈ (cid:104) (cid:105) , is implied by the inequality d s ( S (cid:48) ) ≤ f ( S (cid:48) ) = f ( s , s ), where S (cid:48) = S (cid:48) ∪ S (cid:48) and S (cid:48) i = { ( l i , i ) : l i ∈ (cid:104) s i (cid:105)} for all i ∈ (cid:104) (cid:105) . The latter inequality is in turn included in (77) and (78).In addition to the above, it turns out that the region described (141) is a polymatroid, as the setfunction f is normalized (by definition), non-decreasing and submodular. This is shown in detailfurther on in part B.1. With D Mul ( a ) and D PTIN (cid:48) ( a ) both being polymatroids, we are now readyto invoke [45, Th. 44.6]. This theorem states that: the Minkowski sum of a pair of polymatroids,associated with the set functions f ( S ) and g ( S ) , is a polymatroid associated with the set function f ( S ) + g ( S ) . Through a direct application of this theorem, it follows that D Mul ( a ) ⊕ D PTIN (cid:48) ( a ) isgiven by all tuples d ∈ R K + that satisfy d ( S ) ≤ f ( S ) + a · ( S (cid:54) = ∅ ) , ∀S ⊆ U . (143)As for the region in (141), it can be seen that an inequality in (143) associated with a set S = S ∪S ,where S i ⊆ U i for all i ∈ (cid:104) (cid:105) , is redundant unless each S i takes the form { ( l i , i ) : l i ∈ (cid:104) s i (cid:105)} , for some s i ∈ (cid:104) L (cid:105) . After removing redundant inequalities, the region in (143) reduces to the one in Lemma4, hence completing the proof. It remains to show that D PTIN (cid:48) ( a ) is indeed a polymatroid. B.1 Polymatroidality of (141) D PTIN (cid:48) ( a ) is a polymatroids if the associated set function f must satisfy the following conditions: • Normalized: f ( ∅ ) = 0 33 Non-decreasing: f ( S ) ≤ f ( T ), for all S ⊆ T ⊆ U . • Submodular : f (cid:0) S ∪ T (cid:1) + f (cid:0) S ∩ T (cid:1) ≤ f (cid:0) S (cid:1) + f (cid:0) T (cid:1) , for all S , T ⊆ U .As the first condition holds by definition, we have to show that the second and third conditionshold as well. Let us start by establishing some notation. whenever we consider a subset
S ⊆ U , weassume that S = S ∪S , where S ⊆ U and S ⊆ U . Moreover, s and s denote max ( l i ,i ) ∈S i l i andmax ( l i ,i ) ∈S i l i , respectively. Note that whenever S i = ∅ for some i ∈ (cid:104) (cid:105) , we set the corresponding s i to 0. Similar notation is used for a subset T ⊆ U , given by T ∪ T , and with t and t denotingthe corresponding maximum user indices in T and T respectively.Now let us revisit the shorthand notation for f ( S ), given by f ( s , s ), which was introduced inthe previous part. It can be verified that f ( s i , s j ), defined in (142), is equivalently given as f ( s i , s j ) = min ( α [ s i ] ii − α [ s i ] ij ) ( s i (cid:54) = 0) + ( α [ s j ] jj − a ) ( s j (cid:54) = 0) , ( α [ s i ] ii − a ) ( s i (cid:54) = 0) + ( α [ s j ] jj − α [ s j ] ji ) ( s j (cid:54) = 0) , ( α [ s i ] ii − a ) ( s i (cid:54) = 0) + ( α [ s j ] jj − a ) ( s j (cid:54) = 0) , ( α [ s i ] ii − α [ s i ] ij ) ( s i (cid:54) = 0) + ( α [ s j ] jj − α [ s j ] ji ) ( s j (cid:54) = 0) . Moreover, it is readily seen from the above that f ( s , s ) can be written more compactly as f ( s , s ) (cid:44) min m ,m ∈(cid:104) (cid:105) (cid:8) δ m ( s ) + δ m ( s ) (cid:9) where we define δ i ( s i ) (cid:44) (cid:40) α [ s i ] ii − α [ s i ] ij , s i ∈ (cid:104) L (cid:105) , s i = 0 δ i ( s i ) (cid:44) (cid:40) α [ s i ] ii − a, s i ∈ (cid:104) L (cid:105) , s i = 0 . Note that both δ i ( s i ) and δ i ( s i ) are non-decreasing in their arguments, i.e. δ m i i ( s i − ≤ δ m i i ( s i ) , ∀ s i ∈ (cid:104) L (cid:105) . (144)This holds for m i = 1 due to the SIR order α [ s i ] ii − α [ s i ] ij ≥ α [ s i − ii − α [ s i − ij ; and for m i = 2 due tothe SNR order α [ s i ] ii ≥ α [ s i − ii . Recall that both SIR and SNR orders hold in the mc-SLS regime.Next, we prove that f is monotonic. Consider the subsets S and T , where S ⊆ T . We have f ( T ) = f ( t , t ) (145)= min m ,m ∈(cid:104) (cid:105) (cid:8) δ m ( t ) + δ m ( t ) (cid:9) (146)= δ m (cid:63) ( t ) + δ m (cid:63) ( t ) (147)where m (cid:63) and m (cid:63) in (147) are the coefficients that attain the minimum in (146). For S , we have f ( S ) = f ( s , s ) (148)= min m ,m ∈(cid:104) (cid:105) (cid:8) δ m ( s ) + δ m ( s ) (cid:9) (149)34 δ m (cid:63) ( s ) + δ m (cid:63) ( s ) (150) ≤ δ m (cid:63) ( t ) + δ m (cid:63) ( t ) (151)= f ( T ) . (152)The inequality in (150) holds since m (cid:63) and m (cid:63) , which attain the minimum in (146), do not nec-essarily minimize (149). On the other hand, (151) holds since δ m (cid:63)i i ( s i ) is non-decreasing in s i , ashighlighted in (144). Next, we move on to showing that f is submodular.Consider an arbitrary pair of subsets S , T ⊆ U . We consider the two following cases:1. s ≥ t and s ≥ t : In this case, we have f ( S ∪ T ) + f ( S ∩ T ) = f (cid:0) s , s (cid:1) + f ( S ∩ T ) (153) ≤ f (cid:0) s , s (cid:1) + f (cid:0) t , t (cid:1) (154)= f ( S ) + f ( T ) . (155)In the above, (153) holds since here users with maximal indices in S ∪ T are also in S . Onthe other hand, the inequality in (154) holds as users with maximal indices in S ∩ T mustalso be in T , and the fact that f (cid:0) l , l (cid:1) ≤ f (cid:0) t , t (cid:1) , for all ( l , , ( l , ∈ T .2. s ≥ t and s ≤ t : We may express f ( S ) + f ( T ) as f ( S ) + f ( T ) = f (cid:0) s , s (cid:1) + f ( t , t ) (156)= min m ,n ∈(cid:104) (cid:105) (cid:8) δ m ( s ) + δ n ( s ) (cid:9) + min n ,m ∈(cid:104) (cid:105) (cid:8) δ n ( t ) + δ m ( t ) (cid:9) (157)= δ m (cid:63) ( s ) + δ n (cid:63) ( s ) + δ n (cid:63) ( t ) + δ m (cid:63) ( t ) (158)where m (cid:63) , m (cid:63) , n (cid:63) and n (cid:63) are the coefficients that attain the minimum in (157). On the otherhand, in this case the sum f ( S ∪ T ) + f ( S ∩ T ) is bounded above as f ( S ∪ T ) + f ( S ∩ T ) = f (cid:0) s , t (cid:1) + f ( S ∩ T ) (159) ≤ f (cid:0) s , t (cid:1) + f (cid:0) min { s , t } , min { s , t } (cid:1) (160)= f (cid:0) s , t (cid:1) + f (cid:0) t , s (cid:1) (161)= min m ,m ∈(cid:104) (cid:105) (cid:8) δ m ( s ) + δ m ( t ) (cid:9) + min n ,n ∈(cid:104) (cid:105) (cid:8) δ n ( t ) + δ n ( s ) (cid:9) (162) ≤ δ m (cid:63) ( s ) + δ m (cid:63) ( t ) + δ n (cid:63) ( t ) + δ n (cid:63) ( s ) (163)= f ( S ) + f ( T ) . (164)The inequality in (160) holds as the maximal indices of users in S ∩ T are at most min { s , t } and min { s , t } . On the other hand, the inequality in (163) holds since m (cid:63) , m (cid:63) , n (cid:63) and n (cid:63) ,the minimizers in (157), are not necessarily minimizers for (162).The remaining two cases, i.e. t ≥ s and t ≤ s , and t ≥ s and t ≤ s , can be addressed in asimilar manner by swapping indices. This proves that f is submodular, and hence completes theproof of polymatroidality for the region described in (141). References [1] H. Joudeh and G. Caire, “Optimality of treating inter-cell interference as noise under finiteprecision CSIT,” in
Proc. IEEE Int. Symp. Inf. Theory , 2020.352] ——, “Extremal network theory and robust GDoF gain of multi-cell cooperation over multi-cellTIN,” in
Proc. IEEE Global Commun. Conf. (GLOBECOM) , 2020.[3] R. H. Etkin, D. N. C. Tse, and H. Wang, “Gaussian interference channel capacity to withinone bit,”
IEEE Trans. Inf. Theory , vol. 54, no. 12, pp. 5534–5562, 2008.[4] S. A. Jafar and S. Vishwanath, “Generalized degrees of freedom of the symmetric Gaussian K user interference channel,” IEEE Trans. Inf. Theory , vol. 56, no. 7, pp. 3297–3303, 2010.[5] C. Huang, V. R. Cadambe, and S. A. Jafar, “Interference alignment and the generalized degreesof freedom of the X channel,” IEEE Trans. Inf. Theory , vol. 58, no. 8, pp. 5130–5150, 2012.[6] S. Karmakar and M. K. Varanasi, “The generalized degrees of freedom region of the MIMOinterference channel and its achievability,”
IEEE Trans. Inf. Theory , vol. 58, no. 12, pp. 7188–7203, 2012.[7] A. G. Davoodi and S. A. Jafar, “Generalized degrees of freedom of the symmetric K userinterference channel under finite precision CSIT,” IEEE Trans. Inf. Theory , vol. 63, no. 10,pp. 6561–6572, 2017.[8] ——, “Transmitter cooperation under finite precision CSIT: A GDoF perspective,”
IEEETrans. Inf. Theory , vol. 63, no. 9, pp. 6020–6030, 2017.[9] A. G. Davoodi, B. Yuan, and S. A. Jafar, “GDoF region of the MISO BC: Bridging the gapbetween finite precision and perfect CSIT,”
IEEE Trans. Inf. Theory , vol. 64, no. 11, pp.7208–7217, 2018.[10] A. G. Davoodi and S. A. Jafar, “ K -user symmetric M × N MIMO interference channel underfinite precision CSIT: A GDoF perspective,”
IEEE Trans. Inf. Theory , vol. 65, no. 2, pp.1126–1136, 2019.[11] S. Karmakar and M. K. Varanasi, “The capacity region of the MIMO interference channeland its reciprocity to within a constant gap,”
IEEE Trans. Inf. Theory , vol. 59, no. 8, pp.4781–4797, 2013.[12] G. Bresler, A. Parekh, and D. N. C. Tse, “The approximate capacity of the many-to-one andone-to-many gaussian interference channels,”
IEEE Trans. Inf. Theory , vol. 56, no. 9, pp.4566–4592, 2010.[13] V. R. Cadambe and S. A. Jafar, “Interference alignment and degrees of freedom of the K -userinterference channel,” IEEE Trans. Inf. Theory , vol. 54, no. 8, pp. 3425–3441, 2008.[14] S. A. Jafar, “Interference alignment: A new look at signal dimensions in a communicationnet-work,”
Found. Trends Commun. Inf. Theory , vol. 7, no. 1, pp. 1–134, 2011.[15] D. J. Love, R. W. Heath, V. K. N. Lau, D. Gesbert, B. D. Rao, and M. Andrews, “An overviewof limited feedback in wireless communication systems,”
IEEE J. Sel. Areas Commun. , vol. 26,no. 8, pp. 1341–1365, 2008.[16] A. Lapidoth, S. Shamai, and M. Wigger, “On the capacity of fading MIMO broadcast channelswith imperfect transmitter side-information,” in
Proc. 43rd Annu. Allerton Conf. Commun.,Control Comput. , 2005. 3617] A. G. Davoodi and S. A. Jafar, “Aligned image sets under channel uncertainty: Settlingconjectures on the collapse of degrees of freedom under finite precision CSIT,”
IEEE Trans.Inf. Theory , vol. 62, no. 10, pp. 5603–5618, 2016.[18] S. Shamai and B. M. Zaidel, “Enhancing the cellular downlink capacity via co-processing atthe transmitting end,” in
Proc. IEEE Vehicular Technology Conf. (VTC) Spring , 2001.[19] G. Caire and S. Shamai, “On the achievable throughput of a multiantenna Gaussian broadcastchannel,”
IEEE Trans. Inf. Theory , vol. 49, no. 7, pp. 1691–1706, 2003.[20] H. Weingarten, Y. Steinberg, and S. Shamai, “The capacity region of the Gaussian multiple-input multiple-output broadcast channel,”
IEEE Trans. Inf. Theory , vol. 52, no. 9, pp. 3936–3964, 2006.[21] C. Geng, N. Naderializadeh, A. S. Avestimehr, and S. A. Jafar, “On the optimality of treatinginterference as noise,”
IEEE Trans. Inf. Theory , vol. 61, no. 4, pp. 1753–1767, 2015.[22] C. Geng and S. A. Jafar, “On the optimality of treating interference as noise: Compoundinterference networks,”
IEEE Trans. Inf. Theory , vol. 62, no. 8, pp. 4630–4653, 2016.[23] H. Sun and S. A. Jafar, “On the optimality of treating interference as noise for K -user parallelGaussian interference networks,” IEEE Trans. Inf. Theory , vol. 62, no. 4, pp. 1911–1930, 2016.[24] X. Yi and G. Caire, “Optimality of treating interference as noise: A combinatorial perspective,”
IEEE Trans. Inf. Theory , vol. 62, no. 8, pp. 4654–4673, 2016.[25] C. Geng, H. Sun, and S. A. Jafar, “On the optimality of treating interference as noise: Generalmessage sets,”
IEEE Trans. Inf. Theory , vol. 61, no. 7, pp. 3722–3736, 2015.[26] S. Gherekhloo, A. Chaaban, and A. Sezgin, “Expanded GDoF-optimality regime of treatinginterference as noise in the M × IEEE Trans. Inf. Theory , vol. 63, no. 1, pp.355–376, 2017.[27] X. Yi and H. Sun, “Opportunistic treating interference as noise,”
IEEE Trans. Inf. Theory ,vol. 66, no. 1, pp. 520–533, 2020.[28] S. Gherekhloo, A. Chaaban, C. Di, and A. Sezgin, “(Sub-)optimality of treating interferenceas noise in the cellular uplink with weak interference,”
IEEE Trans. Inf. Theory , vol. 62, no. 1,pp. 322–356, 2016.[29] H. Joudeh and B. Clerckx, “On the optimality of treating inter-cell interference as noise inuplink cellular networks,”
IEEE Trans. Inf. Theory , vol. 65, no. 11, pp. 7208–7232, 2019.[30] H. Joudeh, X. Yi, B. Clerckx, and G. Caire, “On the optimality of treating inter-cell inter-ference as noise: Downlink cellular networks and uplink-downlink duality,”
IEEE Trans. Inf.Theory, to appear .[31] H. Joudeh and B. Clerckx, “Sum-rate maximization for linearly precoded downlink multiuserMISO systems with partial CSIT: A rate-splitting approach,”
IEEE Trans. Commun. , vol. 64,no. 11, pp. 4847–4861, 2016.[32] E. Piovano, H. Joudeh, and B. Clerckx, “Overloaded multiuser MISO transmission with im-perfect CSIT,” in
Proc. Asilomar Conf. Signals, Syst. Comput. , Nov. 2016, pp. 34–38.3733] A. G. Davoodi and S. A. Jafar, “Optimality of simple layered superposition coding in the3 user MISO BC with finite precision CSIT,”
IEEE Trans. Inf. Theory , vol. 65, no. 11, pp.7181–7207, 2019.[34] H. Joudeh and B. Clerckx, “On the separability of parallel MISO broadcast channels underpartial CSIT: A degrees of freedom region perspective,”
IEEE Trans. Inf. Theory , vol. 66,no. 7, pp. 4513–4529, 2020.[35] N. Naderializadeh and A. S. Avestimehr, “ITLinQ: A new approach for spectrum sharing indevice-to-device communication systems,”
IEEE J. Sel. Areas Commun. , vol. 32, no. 6, pp.1139–1151, 2014.[36] X. Yi and G. Caire, “ITLinQ+: An improved spectrum sharing mechanism for device-to-devicecommunications,” in
Proc. Asilomar Conf. Signals, Syst. Comput. , 2015.[37] D. Gesbert, S. Hanly, H. Huang, S. Shamai Shitz, O. Simeone, and W. Yu, “Multi-cell MIMOcooperative networks: A new look at interference,”
IEEE J. Sel. Areas Commun. , vol. 28,no. 9, pp. 1380–1408, 2010.[38] O. Simeone, N. Levy, A. Sanderovich, O. Somekh, B. M. Zaidel, H. V. Poor, and S. Shamai,“Cooperative wireless cellular systems: An information-theoretic view,”
Found. Trends Com-mun. Inf. Theory , vol. 8, no. 1-2, pp. 1–177, 2012.[39] A. Lozano, R. W. Heath, and J. G. Andrews, “Fundamental limits of cooperation,”
IEEETrans. Inf. Theory , vol. 59, no. 9, pp. 5213–5226, 2013.[40] Y. Chan, J. Wang, and S. A. Jafar, “Toward an extremal network theory—robust GDoF gainof transmitter cooperation over TIN,”
IEEE Trans. Inf. Theory , vol. 66, no. 6, pp. 3827–3845,2020.[41] C. Suh, M. Ho, and D. N. C. Tse, “Downlink interference alignment,”
IEEE Trans. Commun. ,vol. 59, no. 9, pp. 2616–2626, 2011.[42] Y. Mao, B. Clerckx, and V. O. Li, “Rate-splitting multiple access for downlink communica-tion systems: Bridging, generalizing, and outperforming SDMA and NOMA,”
EURASIP J.Wireless Commun. Network. , vol. 2018, no. 1, p. 133, 2018.[43] Z. Li, C. Ye, Y. Cui, S. Yang, and S. Shamai, “Rate splitting for multi-antenna downlink:Precoder design and practical implementation,”
IEEE J. Sel. Areas Commun., to appear .[44] T. M. Cover and J. A. Thomas,
Elements of information theory . John Wiley & Sons, 2012.[45] A. Schrijver,
Combinatorial optimization: Polyhedra and efficiency . Springer, 2003, vol. 24.[46] C. J. McDiarmid, “Rado’s theorem for polymatroids,” in
Math. Proc. Cambridge Philos. Soc. ,vol. 78, no. 02. Cambridge Univ Press, 1975, pp. 263–281.[47] A. El Gamal and Y.-H. Kim,