ebook img

Approximate Capacity of a Class of Partially Connected Interference Channels PDF

1.6 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Approximate Capacity of a Class of Partially Connected Interference Channels

1 Approximate Capacity of a Class of Partially Connected Interference Channels Muryong Kim, Yitao Chen, and Sriram Vishwanath, Senior Member, IEEE Abstract—We derive inner and outer bounds on the capacity based on rate-splitting, lattice alignment, and successive de- region for a class of three-user partially connected interference coding. channels. We focus on the impact of topology, interference alignment, and interplay between interference and noise. The 7 representative channels we consider are the ones that have clear B. Related Work interference alignment gain. For these channels, Z-channel type 1 Lattice coding based on nested lattices is shown to achieve outerboundsaretighttowithinaconstantgapfromcapacity.We 0 present near-optimal achievable schemes based on rate-splitting the capacity of the single user Gaussian channel in [12], [27]. 2 and lattice alignment. The idea of lattice-based interference alignment by decoding b the sum of lattice codewords appeared in the conference e Index Terms—Interference channel, interference alignment, F nested lattice code, side information graph, topological interfer- version of [4]. This lattice alignment technique is used to ence management. derive capacity bounds for three-user interference channel in 3 [2], [3]. The idea of decoding the sum of lattice codewords is 1 also used in [13]–[15] to derive the approximate capacity of I. INTRODUCTION ] the two-way relay channel. An extended approach, compute- T A. Motivation and-forward [16], [17] enables to first decode some linear I . ThecapacityoftheInterferencechannelremainsoneofthe combinations of lattice codewords and then solve the lattice s most challenging open problems in the domain of network equation to recover the desired messages. This approach is c [ information theory. The capacity region is not known in alsousedin[7]tocharacterizeapproximatesum-ratecapacity general, except for a specific range of channel parameters. of the fully connected K-user interference channel. 2 For the two-user scalar Gaussian interference channel, where The idea of sending multiple copies of the same sub- v 6 the interference alignment is not required, the approximate message at different signal levels, so-called Zigzag decoding, 7 capacity region to within one bit is known [1]. For the appeared in [5] where receivers collect side information and 5 channels where interference alignment is required such as the use them for interference cancellation. 7 K-user Gaussian interference channel [2]–[5], [7], [11] and The K-user cyclic Gaussian interference channel is con- 0 the Gaussian X-channel [9]–[11], a tight characterization of sidered in [6] where an approximate capacity for the weak . 1 the capacity region is not known, even for symmetric channel interference regime (SNR INR for all k) and the exact 0 k ≥ k cases. capacityforthestronginterferenceregime(SNR INR for 7 k k ≤ 1 Atractableapproachtothecapacityofinterferencechannels allk)arederived.Ourtype4and5channelsareK =3cases : is to consider partial connectivity of interference links and in mixed interference regimes, which were not considered in v analyze the impact of topology on the capacity. Topological [6]. i X interference management [8] approach gives important in- r sightson thedegrees-of-freedom (DoF)of partiallyconnected a C. Main Results interference channels and their connection to index coding problems [18]–[25]. It is shown that the symmetric DoF of We consider five channel types defined in Table I and a partially connected interference channel can be found by described in Fig. 1 (a)–(e). Each channel type is a partially solving the corresponding index coding problem. connected three-user Gaussian interference channel. Each conInnetchtiesdpinatpeerrf,erwenececocnhsaindneerlsaacnldascshaorfacttherreiez-euaseprprpoaxritmiaalltye tLraentsumsidtteenroties tshuebjneocitsetovaproiawnecrecboynsNtrkai=ntEE[[ZXk2k2].]W≤itPhoku=tloPss. capacity regions at finite SNR. We focus on the impact of of generality, we assume that N1 N2 N3. ≤ ≤ interferencetopology,interferencealignment,andinterplaybe- Definition1(sideinformationgraph):Thesideinformation tween interference and noise. We choose a few representative graph representation of an interference channel satisfies the topologies where we can achieve clear interference alignment following. gain. For these topologies, Z-channel type outer bounds are A node represents a transmitter-receiver pair, or equiva- • tight to within a constant gap from the corresponding inner lently, the message. bound. For each topology, we present an achievable scheme There is a directed edge from node i to node j if • transmitter i does not interfere at receiver j. The authors are with the University of Texas at Austin, Austin, The side information graphs for five channel types are de- TX 78701 USA (e-mail: [email protected], [email protected], sri- [email protected]). scribed in Fig. 1 (f)–(j). We state the main results in the 2 Type Channelmodel II. CAPACITYOUTERBOUNDS Y1=X1+X2+Z1 1 Y2=X1+X2+X3+Z2 We prove the capacity outer bound in Theorem 1 for each Y3=X2+X3+Z3 Y1=X1+X2+X3+Z1 channel type. The result is summarized in Table II. The shape 2 Y2=X1+X2+Z2 oftheouterboundregionisillustratedinFig.2.Forallchannel Y3=X1+X3+Z3 types, we assume P =P =P =P and N N N . Y1=X1+X3+Z1 1 2 3 1 ≤ 2 ≤ 3 3 Y2=X2+X3+Z2 Y3=X1+X2+X3+Z3 Y1=X1+X3+Z1 A. Channel Type 1 4 Y2=X1+X2+Z2 Y3=X2+X3+Z3 In this section, we present an outer bound on the capacity Y1=X1+X2+Z1 region of Type 1 channel defined by 5 Y2=X2+X3+Z2 Y3=X1+X3+Z3        TABLEI Y 1 1 0 X Z 1 1 1 FIVECHANNELTYPES  Y2 = 1 1 1  X2 + Z2 . Y 0 1 1 X Z 3 3 3 We state the outer bound in the following theorem. following two theorems, of which the proofs will be given Theorem 3: The capacity region of Type 1 channel is in the main body of the paper. contained in the following outer bound region: Theorem 1 (Capacity region outer bound): For the five channel types, if (R1,R2,R3) is achievable, it must satisfy Rk ≤Ck, k(cid:18)=1,2,3(cid:19) (cid:18) (cid:19) 1 P 1 2P +N (cid:88) 1 (cid:18) P (cid:19) R1+R2 log 1+ + log 2 Rj log 1+ |K| (1) ≤ 2 N1 2 P +N2 j∈K ≤ 2 minj∈K{Nj} R2+R3 1log(cid:18)1+ P (cid:19)+ 1log(cid:18)2P +N3(cid:19). ≤ 2 N 2 P +N foreverysubset ofthenodes 1,2,3 thatdoesnotinclude 2 3 K { } a directed cycle in the side information graph over the subset. Proof: The individual rate bounds are obvious. We pro- Theorem 2 (Capacity region to within one bit): ceed to sum-rate bounds. For any rate triple (R ,R ,R ) on the boundary of the outer 1 2 3 boundregion,thepoint(R 1,R 1,R 1)isachievable. n(R +R (cid:15)) 1− 2− 3− 1 2− I(Xn;Yn)+I(Xn;Yn) ≤ 1 1 2 2 I(Xn;Yn Xn)+I(Xn;Yn Xn) D. Paper Organization and Notation ≤ 1 1 | 2 2 2 | 3 =h(Yn Xn) h(Yn Xn,Xn) The capacity outer bounds are derived in Section II. The 1 | 2 − 1 | 1 2 +h(Yn Xn) h(Yn Xn,Xn) innerboundsforeachchanneltypeandthecorrespondinggap 2 | 3 − 2 | 2 3 analysis are given in Section III, IV, V, VI, VII, respectively. =h(Xn+Zn) h(Zn) 1 1 − 1 Section VIII concludes the paper. While lattice coding-based +h(Xn+Xn+Zn) h(Xn+Zn) achievablerateregionsforchanneltypes4and5arepresented (cid:18)1 2 (cid:19) 2 − (cid:18) 1 2 (cid:19) n P +N n 2P +N 1 2 in Section VI and VII, random coding achievability is given log + log ≤ 2 N 2 P +N in Appendix. 1 2 Signal x is a coded version of message M with code where the first inequality is by Fano’s inequality, the second ij ij rate R unless otherwise stated. The single user capacity at inequality due to the independence of X ,X ,X . The third ij (cid:16) (cid:17) 1 2 3 receiver k is denoted by C = 1log 1+ P . Let denote inequalityholdsfromthefactthatGaussiandistributionmaxi- k 2 Nk C mizesdifferentialentropyandthath(Xn+Zn) h(Xn+Zn) the capacity region of an interference channel. Also, let i 1 1 − 1 2 R is also maximized by Gaussian distribution. Similarly, and denotethecapacityinnerboundandthecapacityouter o R bound, respectively. Thus, . Let δ denote the i o k R ⊂ C ⊂ R n(R +R (cid:15)) gapontherateRk between i and o.Letδjk denotethegap 2 3− on the sum-rate Rj +Rk bRetweenRRi and Ro. For example, ≤I(X2n;Y2n)+I(X3n;Y3n) if I(Xn;Yn Xn,Xn)+I(Xn;Yn) ≤ 2 2 | 1 3 3 3 =h(Yn Xn,Xn) h(Yn Xn,Xn,Xn) i = (Rj,Rk):Rk Lk,Rj +Rk Ljk (2) 2 | 1 3 − 2 | 1 2 3 R { ≤ ≤ } +h(Yn) h(Yn Xn) o = (Rj,Rk):Rk Uk,Rj +Rk Ujk , (3) 3 − 3 | 3 R { ≤ ≤ } =h(Xn+Zn) h(Zn) 2 2 − 2 thenδk =Uk Lk andδjk =Ujk Ljk.Forsideinformation +h(Xn+Xn+Zn) h(Xn+Zn) graph, we us−e graph notation of−[23]. For example, = (cid:18)2 3 (cid:19) 3 − (cid:18) 2 3 (cid:19) 1 n P +N n 2P +N G 2 3 (13),(2),(31) means that node 1 has an incoming edge log + log . f{rom| node3,t|hat}node2hasnoincomingedge,andthatnode ≤ 2 N2 2 P +N3 3 has an incoming edge from node 1. 3 1 1 1 1 11 11 11 11 1 1 2 2 2 2 22 22 22 22 2 2 3 3 3 3 33 33 33 33 3 3 (a) Type1 (b) Type2 (c) Type3 (d) Type4 (e) Type5 1 1 1 1 11 11 1 1 2 1 2 2 1 2 22 1 22 2 1 2 1 3 3 3 3 33 33 3 3 1 1 1 1 2 3 2 3 2 3 2 3 2 3 2 2 2 2 (f) G1 3 (g) G2 3 3 (h) G3 3 (i) G4 (j) G5 Fig. 1. Five channel types and their side information graphs: G1 = {(1|3),(2),(3|1)}, G2 = {(1),(2|3),(3|2)}, G3 = {(1|2),(2|1),(3)}, G4 = {(1|2),(2|3),(3|1)},andG5={(1|3),(2|1),(3|2)}. B. Channel Type 2 n(R +R (cid:15)) 1 3 − I(Xn;Yn)+I(Xn;Yn) ≤ 1 1 3 3 In this section, we present an outer bound on the capacity I(Xn;Yn Xn,Xn)+I(Xn;Yn) ≤ 1 1 | 2 3 3 3 region of Type 2 channel defined by =h(Yn Xn,Xn) h(Yn Xn,Xn,Xn) 1 | 2 3 − 1 | 1 2 3 +h(Yn) h(Yn Xn) 3 − 3 | 3        =h(Xn+Zn) h(Zn) Y1 1 1 1 X1 Z1 1 1 − 1  Y2 = 1 1 0  X2 + Z2 . +h(X(cid:18)1n+X3n(cid:19)+Z3n)−h(cid:18)(X1n+Z3n)(cid:19) Y 1 0 1 X Z n P +N n 2P +N 3 3 3 1 3 log + log . ≤ 2 N 2 P +N 1 3 We state the outer bound in the following theorem. Theorem 4: The capacity region of Type 2 channel is contained in the following outer bound region: Rk Ck, k =1,2,3 C. Channel Type 3 ≤ (cid:18) (cid:19) (cid:18) (cid:19) 1 P 1 2P +N 2 R +R log 1+ + log 1 2 ≤ 2 N1 2 P +N2 In this section, we present an outer bound on the capacity (cid:18) (cid:19) (cid:18) (cid:19) 1 P 1 2P +N3 region of Type 3 channel defined by R +R log 1+ + log . 1 3 ≤ 2 N 2 P +N 1 3        Y 1 0 1 X Z 1 1 1 Proof:  Y2 = 0 1 1  X2 + Z2 . Y 1 1 1 X Z 3 3 3 n(R +R (cid:15)) We state the outer bound in the following theorem. 1 2 − I(Xn;Yn)+I(Xn;Yn) Theorem 5: The capacity region of Type 3 channel is ≤ 1 1 2 2 I(Xn;Yn Xn,Xn)+I(Xn;Yn) contained in the following outer bound region: ≤ 1 1 | 2 3 2 2 =h(Yn Xn,Xn) h(Yn Xn,Xn,Xn) 1 | 2 3 − 1 | 1 2 3 +h(Yn) h(Yn Xn) 2 − 2 | 2 Rk Ck, k =1,2,3 =h(Xn+Zn) h(Zn) ≤ 1 (cid:18) P (cid:19) 1 (cid:18)2P +N (cid:19) 1 1 − 1 R +R log 1+ + log 3 +h(Xn+Xn+Zn) h(Xn+Zn) 1 3 ≤ 2 N 2 P +N (cid:18)1 2 (cid:19) 2 − (cid:18) 1 2 (cid:19) (cid:18) 1(cid:19) (cid:18) 3 (cid:19) n P +N n 2P +N 1 P 1 2P +N 1 2 3 log + log . R +R log 1+ + log . 2 3 ≤ 2 N 2 P +N ≤ 2 N 2 P +N 1 2 2 3 4 where Y = 1 Y , Z = 1 Z , N = E[Z 2] = 1, R3 R3 E[X2] k(cid:48)P =√PNkandk k(cid:48) √Nk k 0 k(cid:48) k ≤ k  h h h   1 0 1  11 12 13 √N1 √N1 R1 R1  h21 h22 h23 = √1N2 √1N2 0 . R2 R2 h31 h32 h33 0 √1N3 √1N3 (a) Channeltype1 (b) Channeltypes4and5 With the usual definitions of SNRk = h2kNk0Pk and Fig.2. Theshapeoftheouterboundregion.Theregionsforchanneltypes INR = h2jkPk for j =k as in [1], [6], 2and3looksimilartotheoneforchanneltype1(withchangeofaxis). k N0 (cid:54) P P SNR = INR = (4) 1 1 N ≥ N Proof: 1 2 P P SNR = INR = (5) 2 2 N ≥ N n(R +R (cid:15)) 2 3 1 3 − P P ≤II((XX1nn;;YY1nn)X+nI)(+XI3n(;XY3nn;)Yn Xn) SNR3 = N3 ≤INR3 = N1. (6) ≤ 1 1 | 3 3 3 | 2 =h(Yn Xn) h(Yn Xn,Xn) We state the outer bound in the following theorem. 1 | 3 − 1 | 1 3 +h(Yn Xn) h(Yn Xn,Xn) Theorem 6: The capacity region of Type 4 channel is 3 | 2 − 3 | 2 3 contained in the following outer bound region: =h(Xn+Zn) h(Zn) 1 1 − 1 +h(Xn+Xn+Zn) h(Xn+Zn) (cid:18)1 3 (cid:19) 3 − (cid:18) 1 3 (cid:19) Rk Ck, k =1,2,3 nlog P +N1 + nlog 2P +N3 . ≤ 1 (cid:18) P (cid:19) 1 (cid:18)2P +N (cid:19) 2 ≤ 2 N1 2 P +N3 R1+R2 ≤ 2log 1+ N + 2log P +N 1 2 (cid:18) (cid:19) 1 2P R +R log 1+ n(≤R2I(+XR2n;3Y−2n(cid:15)))+I(X3n;Y3n) R12+R33 ≤ 21log(cid:18)1+ NP1(cid:19)+ 1log(cid:18)2P +N3(cid:19). I(Xn;Yn Xn)+I(Xn;Yn Xn) ≤ 2 N2 2 P +N3 ≤ 2 2 | 3 3 3 | 1 =h(Yn Xn) h(Yn Xn,Xn) 2 | 3 − 2 | 2 3 Proof: +h(Yn Xn) h(Yn Xn,Xn) 3 | 1 − 3 | 1 3 =h(Xn+Zn) h(Zn) 2 2 − 2 n(R1+R2 (cid:15)) n+h(X(cid:18)2Pn++XN3n(cid:19)+Z3nn)−h(cid:18)(X2P2n++ZN3n)(cid:19) ≤I(X1n;Y−1n)+I(X2n;Y2n) log 2 + log 3 . I(Xn;Yn Xn)+I(Xn;Yn) ≤ 2 N 2 P +N ≤ 1 1 | 3 2 2 2 3 =h(Yn Xn) h(Yn Xn,Xn) 1 | 3 − 1 | 1 3 +h(Yn) h(Yn Xn) 2 − 2 | 2 =h(Xn+Zn) h(Zn) 1 1 − 1 D. Channel Type 4 +h(X(cid:18)1n+X2n(cid:19)+Z2n)−h(cid:18)(X1n+Z2n)(cid:19) n P +N n 2P +N 1 2 In this section, we present an outer bound on the capacity log + log . ≤ 2 N 2 P +N region of Type 4 channel defined by 1 2        Y 1 0 1 X Z 1 1 1  Y2 = 1 1 0  X2 + Z2 . n(R2+R3−(cid:15)) Y3 0 1 1 X3 Z3 I(Xn;Yn)+I(Xn;Yn) ≤ 2 2 3 3 I(Xn;Yn Xn)+I(Xn;Yn) This is a cyclic Gaussian interference channel [6]. We first ≤ 2 2 | 1 3 3 show that channel type 4 is in the mixed interference regime. =h(Yn Xn) h(Yn Xn,Xn) 2 | 1 − 2 | 1 2 By normalizing the noise variances, we get the equivalent +h(Yn) h(Yn Xn) channel given by =h(Xn+3Zn−) h3(Z|n)3 2 2 − 2  Y1(cid:48)   h11 h12 h13  X1   Z1(cid:48)  +h(X(cid:18)2n+X3n(cid:19)+Z3n)−h(cid:18)(X2n+Z3n)(cid:19)  Y2(cid:48) = h21 h22 h23  X2 + Z2(cid:48)  nlog P +N2 + nlog 2P +N3 . Y3(cid:48) h31 h32 h33 X3 Z3(cid:48) ≤ 2 N2 2 P +N3 5 n(R +R (cid:15)) Proof: 1 3 − I(Xn;Yn)+I(Xn;Yn) ≤≤≤==≤hhIII(((((XXXXY11111n1nnnn),;;(cid:18)+YYX−111nn3Xnh));3(nY++Y1+1nnII)|((ZXXX1n(cid:19)1n333nn),;;−XYY133h3nnn()||ZXX1n12nn))) n(≤≤≤≤=R1IIIIh(((((+XXXXYR1111nnnnn);;;,2YYYX−111nnn2nh(cid:15))));()Y+++Y1nnIII)(((XXXXn222nnn,;;;XYYY221nnnn))||XX31nn)) n 2P +N1 1 − 1 | 1 2 ≤ 2 log N1 =h(X1n(cid:18)+X2n+Z1n(cid:19))−h(Z1n) n 2P +N 1 log where we used the fact that I(Xn;Yn Xn) = I(Xn;Xn + ≤ 2 N1 3 3 | 2 3 3 Zn) I(Xn;Xn+Zn)=I(Xn;Yn Xn). 3 ≤ 3 3 1 3 1 | 1 where we used the fact that I(Xn;Yn Xn) = I(Xn;Xn + 2 2 | 3 2 2 Zn) I(Xn;Xn+Zn)=I(Xn;Yn Xn). 2 ≤ 2 2 1 2 1 | 1 E. Channel Type 5 n(R2+R3 (cid:15)) − I(Xn;Yn)+I(Xn;Yn) In this section, we present an outer bound on the capacity ≤I(X2n;Y2n)+I(X3n;Y3n Xn) region of Type 5 channel defined by ≤ 2 2 3 3 | 1 I(Xn;Yn)+I(Xn;Yn Xn) ≤ 2 2 3 2 | 2        I(Xn,Xn;Yn) Y1 1 1 0 X1 Z1 ≤ 2 3 2  Y2 = 0 1 1  X2 + Z2 . =h(Y2n)−h(Y2n|X2n,X3n) Y3 1 0 1 X3 Z3 =h(X2n(cid:18)+X3n+Z2n(cid:19))−h(Z2n) n 2P +N 2 log This is a cyclic Gaussian interference channel [6]. We first ≤ 2 N 2 show that channel type 5 is in the mixed interference regime. By normalizing the noise variances, we get the equivalent where we used the fact that I(Xn;Yn Xn) = I(Xn;Xn + 3 3 | 1 3 3 channel given by Zn) I(Xn;Xn+Zn)=I(Xn;Yn Xn). 3 ≤ 3 3 2 3 2 | 2  YYY123(cid:48)(cid:48)(cid:48) = √√011NN13 √√011NN12 √√011NN23  XXX123 + ZZZ123(cid:48)(cid:48)(cid:48) . n(≤≤=R1IIh(((+XXYR11nnn;;3XYY−11nnn)(cid:15))|X)+2nhI)((+YXnI3n(;XXY3n3nn,;)XY3nn)) 1 | 2 − 1 | 1 2 We can see that +h(Yn) h(Yn Xn) 3 − 3 | 3 =h(Xn+Zn) h(Zn) SNR = P INR = P (7) +h1(Xn+1 X−n+Z1n) h(Xn+Zn) 1 N1 ≥ 1 N3 n (cid:18)1P +N3 (cid:19) 3n − (cid:18)2P1 +N3 (cid:19) P P log 1 + log 3 SNR = INR = (8) ≤ 2 N 2 P +N 2 N ≤ 2 N 1 3 2 1 P P SNR = INR = . (9) 3 3 N ≤ N 3 2 We state the outer bound in the following theorem. F. Relaxed Outer Bounds Theorem 7: The capacity region of Type 5 channel is contained in the following outer bound region: For ease of gap calculation, we also derive relaxed outer bounds. First, we can see that for N N , j k ≤ R C , k =1,2,3 (cid:18) (cid:19) (cid:18) (cid:19) (cid:18) (cid:19) k k 1 P 1 2P +N 1 2P ≤ 1 (cid:18) 2P(cid:19) log 1+ + log k log 1+ . R1+R2 log 1+ 2 Nj 2 P +Nk ≤ 2 Nj ≤ 2 N 1 (cid:18) (cid:19) 1 2P Five outer bound theorems in this section, together with this R +R log 1+ 2 3 ≤ 2 N inequality, give the sum-rate bound expression in Theorem 1. 2 R +R 1log(cid:18)1+ P (cid:19)+ 1log(cid:18)2P +N3(cid:19). Next, we can assume that P ≥ 3Nj for j = 1,2,3. 1 3 Otherwise, showing one-bit gap capacity is trivial as the ≤ 2 N 2 P +N 1 3 6 Type OuterboundregionRo RelaxedouterboundregionR(cid:48)o Two-dimensionalcross-sectionofR(cid:48)o (cid:16) (cid:17) 1 RR21++RRRk32≤≤≤C2211kll,ooggk(cid:16)(cid:16)=PPNN++1,NN12212,3·· 22PPPP++++NNNN2323(cid:17)(cid:17) RR12++RRRk23≤≤≤121212llloooggg(cid:16)(cid:16)(cid:16)NNNPPPk12 ···437733(cid:17)(cid:17)(cid:17) RRAt13s≤≤ommmeiiRnn2(cid:110)(cid:110)∈1122ll[oo0gg,C(cid:16)(cid:16)2NNPP],12 ·· 3377(cid:17)(cid:17)−−RR22,,1212lloogg(cid:16)(cid:16)NNPP13 ·· 4343(cid:17)(cid:17)(cid:111)(cid:111) 2 RR11++RRRk32≤≤≤C2211kll,ooggk(cid:16)(cid:16)=PPNN++1,NN11211,3·· 22PPPP++++NNNN2323(cid:17)(cid:17) RR11++RRRk23≤≤≤121212llloooggg(cid:16)(cid:16)(cid:16)NNNPPPk11 ···437733(cid:17)(cid:17)(cid:17) RRAt23s≤≤ommmeiiRnn1(cid:110)(cid:110)∈1122ll[oo0gg,C(cid:16)(cid:16)1NNPP],11 ·· 3377(cid:17)(cid:17)−−RR11,,1212lloogg(cid:16)(cid:16)NNPP23 ·· 4343(cid:17)(cid:17)(cid:111)(cid:111) 3 RR21++RRRk33≤≤≤C2211kll,ooggk(cid:16)(cid:16)=PPNN++1,NN12212,3·· 22PPPP++++NNNN3333(cid:17)(cid:17) RR12++RRRk33≤≤≤121212llloooggg(cid:16)(cid:16)(cid:16)NNNPPPk12 ···437733(cid:17)(cid:17)(cid:17) RRAt12s≤≤ommmeiiRnn3(cid:110)(cid:110)∈1122ll[oo0gg,C(cid:16)(cid:16)3NNPP],12 ·· 3377(cid:17)(cid:17)−−RR33,,1212lloogg(cid:16)(cid:16)NNPP12 ·· 4343(cid:17)(cid:17)(cid:111)(cid:111) 4 RRR121+++RRRRk332≤≤≤≤C222111klll,ooogggk(cid:16)(cid:16)(cid:16)=P2PPNN++1N+,NN1212N12,13··(cid:17)22PPPP++++NNNN2323(cid:17)(cid:17) RRR112+++RRRRk233≤≤≤≤12121212lllloooogggg(cid:16)(cid:16)(cid:16)(cid:16)NNNNPPPPk112 ····43777333(cid:17)(cid:17)(cid:17)(cid:17) RRRAt232s+≤≤omRmme3iiRnn≤1(cid:110)(cid:110)∈121122lll[ooo0ggg,C(cid:16)(cid:16)(cid:16)1NNNPPP],211 ···733377(cid:17)(cid:17)(cid:17)−−RR11,,1212lloogg(cid:16)(cid:16)NNPP23 ·· 4343(cid:17)(cid:17)(cid:111)(cid:111) 5 RRR211+++RRRRk332≤≤≤≤C222111klll,ooogggk(cid:16)(cid:16)(cid:16)=22PPPN+1NN++,N1122NN1,123·(cid:17)(cid:17)2PP++NN33(cid:17) RRR121+++RRRRk233≤≤≤≤12121212lllloooogggg(cid:16)(cid:16)(cid:16)NNNNPPPPk121 ····43777333(cid:17)(cid:17)(cid:17) RRRAt131s+≤≤omRmme3iiRnn≤2(cid:110)(cid:110)∈121122lll[ooo0ggg,C(cid:16)(cid:16)(cid:16)2NNNPPP],112 ···733377(cid:17)(cid:17)(cid:17)−−RR22,,1212lloogg(cid:16)(cid:16)NNPP13 ·· 4343(cid:17)(cid:17)(cid:111)(cid:111) TABLEII CAPACITYOUTERBOUNDS capacity region is included in the unit hypercube, i.e., R A. Preliminaries: Lattice Coding j (cid:16) (cid:17) ≤ 1log 1+ P <1. For P 3N , 2 Nj ≥ j LatticeΛisadiscretesubgroupofRn,Λ= t=Gu:u 1 (cid:18) 2P(cid:19) 1 (cid:18) P (cid:19) 1 (cid:18)Nj (cid:19) Zn whereG Rn n isarealgeneratormatri{x.Quantizatio∈n log 1+ = log + log +2 × } ∈ 2 Nj 2 Nj 2 P with respect to Λ is QΛ(x) = argminλ Λ x λ . Modulo 1 (cid:18) P (cid:19) 1 (cid:18)7(cid:19) operation with respect to Λ is M (x) =∈ [x(cid:107)] m−od(cid:107)Λ = x log + log Λ − ≤ 2 Nj 2 3 QΛ(x). For convenience, we use both notations MΛ(·) and (cid:18) (cid:19) (cid:18) (cid:19) (cid:18) (cid:19) [] mod Λ interchangeably. Fundamental Voronoi region of Λ 1 P 1 P 1 4 log 1+ log + log . i·s (Λ) = x : Q (x) = 0 . Volume of the Voronoi region 2 Nj ≤ 2 Nj 2 3 V { Λ(cid:82) } of Λ is V(Λ) = dx. Normalized second moment of (Λ) The resulting relaxed outer bounds R(cid:48)o are summarized in Λ is G(Λ) = σ2(ΛV) where σ2(Λ) = 1 (cid:82) x 2dx. Table II. V(Λ)2/n nV(Λ) (Λ)(cid:107) (cid:107) Lattices Λ , Λ and Λ are said to be nested if ΛV Λ Λ . 1 2 2 1 ⊆ ⊆ For nested lattices Λ Λ , Λ /Λ =Λ (Λ ). 2 1 1 2 1 2 III. INNERBOUND:CHANNELTYPE1 ⊂ ∩V Webrieflyreviewthelatticedecodingprocedurein[12].We RαThiseodreefimne8d: bGyiven α = (α0,α2) ∈ [0,1]2, the rate region Vus(eΛn)e=ste(d2lπaettSic)en2s.ΛT⊆heΛtrtawnsimthitσte2r(Λse)n=dsSx,=G([Λt)+=d]2π1meo,danΛd 1 (cid:18)1 α (1 α )P (cid:19) over the point-to-point Gaussian channel y=x+z where the R1 ≤ 2log+ 2−−α00 + (α0+−α(cid:18)2)P0 +N2(cid:19) tchoedetrwaonrsdmtit∈poΛwte∩r 1V(xΛ)2, t=heSdiatnhderthsiegnnaolisde z∼ Unif((0V,(NΛI))),. +1log 1+ α0P The code rate is givnen(cid:107)b(cid:107)y R= 1 log(cid:16)V(Λ)(cid:17). ∼N 2 N1 n V(Λt) 1 (cid:18) α P (cid:19) After linear scaling, dither removal, and mod-Λ operation, 2 R log 1+ 2 ≤ 2 α P +N we get 0 2 (cid:18) (cid:19) 1 1 P R log+ + 3 ≤ 2 2 α (α +α )P +N − 0 0 2 3 y(cid:48) =[βy d] mod Λ=[t+ze] mod Λ (10) − where log+()=max 0,log() . And, · { · } (cid:32) (cid:33) (cid:91) where the effective noise is ze = (β 1)x + βz1 and its R= CONV α Rα vMaMriaSnEcescσae2lin=g fn1acEt[o(cid:107)rzeβ(cid:107)2=] =S(β −plu1g)g2−eSd+inβ, 2wNe.gWetitσh2th=e S+N e is achievable where CONV() is convex hull operator. βN = SN .Thecapacityofthemod-Λchannel[12]between · S+N 7 t and y is where t Λ (Λ ) and t Λ (Λ ) are lattice 11 c 1 3 c 3 ∈ ∩V ∈ ∩V codewords. The dither signals d and d are uniformly 1 1 1 11 3 I(t;y) = h(y) h(yt) distributed over (Λ ) and (Λ ), respectively. To satisfy n n1 − n1 | power constraintsV, we1chooseVE[ 3x11 2] = nσ2(Λ1) = (1 = h(y) h(z mod Λ) α )nP, E[ x 2] = α nP, E[ (cid:107)x 2(cid:107)] = α nP, E[ x 2] −= n − n 1 (cid:107) 10(cid:107) 1 (cid:107) 2(cid:107) 2 (cid:107) 3(cid:107) 1 1 nσ2(Λ3)=nP. h(y) h(z) Withthechoiceoftransmitsignals,thereceivedsignalsare ≥ n − n 1 1 given by = logV(Λ) h(z) n − n (cid:18) (cid:19) y =x +x +x +z 1 S 1 11 2 10 1 = log 2 βN y2 =[x11+x3]+x2+z(cid:48)2 (cid:18) (cid:19) = 1log 1+ S y3 =x3+z(cid:48)3. 2 N where x = [x +x ] is the sum of interference, and z = = C f 11 3 (cid:48)2 x +z and z = x +z are the effective Gaussian noise. 10 2 (cid:48)3 2 3 where I() and h() are mutual information and differential The signal scale diagram at each receiver is shown in Fig. 3 · · entropy, respectively. For reliable decoding of t, we have (a). the code rate constraint R C. With the choice of lattice At the receivers, successive decoding is performed in the pσa2r(aΛmt)eter2s,πeσβ2N(Λ,t) ≥ βN,≤G(Λt) = 2π1e and V(Λt)n2 = froeclleoiwveirng2,oardnedr:rexc1e1iv→er 3x2on→lyxd1e0coadtersecxe3i.ver 1, xf → x2 at G(Λt) ≥ Note that the aligned lattice codewords t + t Λ , 1 (cid:18)V(Λ)(cid:19) and t = [t + t ] mod Λ Λ (Λ )1.1We 3sta∈te thce R = log f 11 3 1 ∈ c ∩ V 1 n V(Λt) relationship between xf and tf in the following lemmas. 1 (cid:18) (2πeS)n2 (cid:19) Lemma 1: The following holds. log ≤ n (2πeβN)n2 1 (cid:18) S (cid:19) [xf −df] mod Λ1 =tf = log . 2 βN where d =d +d . f 11 3 Thus, the constraint R C can be satisfied. By lattice Proof: ≤ decoding [12], we can recover t, i.e., [x d ] mod Λ f f 1 − QΛt(y(cid:48))=t, (11) =[MΛ1(t11+d11)+MΛ3(t3+d3)−df] mod Λ1 with probability 1−Pe where =[MΛ1(t11+d11)+MΛ1(t3+d3)−df] mod Λ1 =[t +d +t +d d ] mod Λ Pe =Pr[QΛt(y(cid:48))(cid:54)=t] (12) =[t11+t1]1mod3Λ 3− f 1 11 3 1 is the probability of decoding error. If we choose Λ to be =t f Poltyrev-good [27], then P 0 as n . e → →∞ Thesecondandthirdequalitiesareduetodistributivelawand B. Achievable Scheme the identity in the following lemma. Lemma 2: For any nested lattices Λ Λ and We present an achievable scheme for the proof of any x Rn, it holds that 3 ⊂ 1 Theorem 8. The achievable scheme is based on rate- ∈ splitting, lattice coding, and interference alignment. Message [M (x)] mod Λ =[x] mod Λ . M1 1,2,...,2nR1 is split into two parts: M11 Λ3 1 1 ∈ { } ∈ 1,2,...,2nR11 and M10 1,2,...,2nR10 , so R1 = Proof: { } ∈ { } R + R . Transmitter 1 sends x = x + x where 11 10 1 11 10 x and x are coded signals of M and M , respec- [M (x)] mod Λ 11 10 11 10 Λ3 1 tively. Transmitters 2 and 3 send x2 and x3, coded signals =[x λ3] mod Λ1 poafrtMicu2la∈r, x{111,2a,n.d..x,32naRre2}latatnicde-Mco3ded∈si{g1n,a2l,s....,2nR3}. In =[M−Λ1(x)−MΛ1(λ3)] mod Λ1 We use the lattice construction of [14], [15] with the lattice =[MΛ1(x)−λ3+QΛ1(λ3)] mod Λ1 partition chain Λ /Λ /Λ , so Λ Λ Λ are nested =[M (x)] mod Λ c 1 3 3 ⊂ 1 ⊂ c Λ1 1 lattices. Λc is the coding lattice for both x11 and x3. Λ1 and =[x] mod Λ1 Λ areshapinglatticesforx andx ,respectively.Thelattice 3 11 3 signals are formed by where λ =Q (x) Λ , thus Q (λ )=λ . 3 Λ3 ∈ 1 Λ1 3 3 Lemma 3: The following holds. x =[t +d ] mod Λ (13) 11 11 11 1 x =[t +d ] mod Λ (14) [t +d ] mod Λ =[x ] mod Λ . 3 3 3 3 f f 1 f 1 8 (α¯0+1)PNe2 . The capacity of the mod-Λ channel between (1−α0)P x x +x x P t(αf¯0a+n1d)Py+2(cid:48)Nies2 1 11 11 3 3 α P 1 2 I(t ;y ) x x x n f 2(cid:48) 2 2 2 α P !"#$$%&’()*%’+’ 1 (cid:18)V(Λ )(cid:19) 0 1 log x10 x10 ≥ n 2h(ze2) (cid:18) (cid:19) 1 α¯ P 0 RX1 RX2 RX3 = log 2 β N 2 e2 (a) Channeltype1 (cid:18) (cid:19) 1 α¯ (α¯ +1)P +α¯ N 0 0 0 e2 = log PP 2 (α¯0+1)Ne2 (cid:18) (cid:19) x2+xx3 xx2 xx3 = 1log α¯0 + α¯0P α1PP !"#$$%&’()*%’,’ 2 α¯0+1 Ne2 xx1 xx1 xx1 1 (cid:18) α¯0 α¯0P (cid:19) = log + 2 α¯ +1 (α +α )P +N RRXX11 RRXX22 RRXX33 0 0 2 2 =C f (b) Channeltype2 For reliable decoding of t at receiver 2, we have the PP xxx3 xxx3 xxx3 PPP icmodpeliersattehactonRstra=int1Rlo1g1(cid:16)=V(n1Λf2l)o(cid:17)g(cid:16)VVC((ΛΛ1c))+(cid:17) 1≤loCgf(cid:16).VT(hΛi2s)(cid:17)als=o αP xx xx xx (cid:16) (cid:17)3 n (cid:16) V(Λc) ≤ f n (cid:17) V(Λ1) x1 x2 x1+x2 12log β2PNe2 = 12log α¯01+1 + (α0+αP2)P+N2 . By lattice RX1 RX2 RX3 decoding, we can recover the modulo sum of interference RX1 RX2 RX3 RX1 RX2 RX3 codewords t from y . Then, we can recover the real sum f 2(cid:48) (c) Channeltype3 xf in the following way. Recover M (x ) by calculating [t + d ] mod Λ Fig.3. Signalscalediagram. • Λ1 f f f 1 (lemma 3). Subtract it from the received signal, • Proof: y2−MΛ1(xf)=QΛ1(xf)+z(cid:48)2(cid:48) (18) [tf +df] mod Λ1 where z(cid:48)2(cid:48) =x2+x10+z2. Quantize it to recover Q (x ), =[MΛ1(t11+t3)+df] mod Λ1 • Λ1 f =[t11+t3+df] mod Λ1 QΛ1(QΛ1(xf)+z(cid:48)2(cid:48))=QΛ1(xf) (19) =[MΛ1(t11+d11)+MΛ1(t3+d3)] mod Λ1 with probability 1−Pe where =[M (t +d )+M (t +d )] mod Λ =[x Λ1+x11] mo1d1Λ Λ3 3 3 1 Pe =Pr[QΛ1(QΛ1(xf)+z(cid:48)2(cid:48))(cid:54)=QΛ1(xf)] (20) 11 3 1 is the probability of decoding error. If we choose Λ to =[x ] mod Λ 1 f 1 be simultaneously Rogers-good and Poltyrev-good [27] with V(Λ ) V(Λ ), then P 0 as n . 1 c e ≥ → →∞ Receiver 2 does not need to recover the codewords t and Recover xf by adding two vectors, 11 • t but the real sum x to remove the interference from y . 3 f 2 M (x )+Q (x )=x . (21) Sincex =M (x )+Q (x ),wefirstrecoverthemodulo Λ1 f Λ1 f f f Λ1 f Λ1 f part and then the quantized part to cancel out xf. This idea Wenowproceedtodecodingx2fromy2−xf =x2+z(cid:48)2.Since appearedin[17]asanachievableschemeforthemany-to-one x2 isacodewordfromani.i.d.randomcodeforpoint-to-point interference channel. channel, we can achieve rate up to The mod-Λ1 channel between tf and y2(cid:48) is given by R 1log(cid:18) α2P (cid:19). (22) 2 ≤ 2 α P +N y =[β y d ] mod Λ (15) 0 2 2(cid:48) 2 2− f 1 At receiver 1, we first decode x while treating other =[x d +z ] mod Λ (16) 11 f f e2 1 − signals x + x + z as noise. The effective noise in the =[t +z ] mod Λ (17) 2 10 1 f e2 1 mod-Λ channel is z =(β 1)2x +β (x +x +z ) 1 e1 1 11 1 2 10 1 NwohteereththaeteEff[ecxtfive2n]o=ise(zα¯e02+=1()βn2P−,1)axndf+thβe2(exff2e+ctxiv1e0+nozi2s)e. wwhitehrevaNriea1nc=e σ(αe210=+αn12E)[P(cid:107)z+e1−(cid:107)N21].=Fo(rβ1re−lia1b)le2α¯d0ePco+dinβg12,Nthe1e variance σe22(cid:107)= n1(cid:107)E[(cid:107)ze2(cid:107)2] = (β2−1)2(α¯0+1)P +β22Ne2 rate R11 must satisfy where Ne2 = (α0 + α2)P + N2. With the MMSE scaling 1 (cid:18)σ2(Λ1)(cid:19) 1 (cid:18) α¯0P (cid:19) factor β2 = (α¯0(+α¯01+)P1+)PNe2 plugged in, we get σe22 =β2Ne2 = R11 ≤ 2log β1σe21 = 2log 1+ (α0+α2)P +N1 9 wheretheMMSEscalingparameterβ = α¯0P .Similarly, C. The Gap 1 α¯0P+Ne1 we have the other rate constraints at receiver 1: We choose the parameter α = N2, which is suboptimal (cid:18) (cid:19) 0 P 1 α P R log 1+ 2 (23) but good enough to achieve a constant gap. This choice of 2 ≤ 2 α0P +N1 parameter, inspired by [1], ensures making efficient use of (cid:18) (cid:19) 1 α P signalscaledifferencebetweenN andN atreceiver1,while R log 1+ 0 . (24) 1 2 10 ≤ 2 N1 keepingtheinterferenceofx10atthenoiselevelN2atreceiver 2. By substitution, we get At receiver 3, the signal x is decoded with the effective 3 noise x2+z3. For reliable decoding, R3 must satisfy 1 (cid:18) P N2 (cid:19) T = log c + − 1 (cid:18) P (cid:19) 1 2 11 α2P +2N2 R3 ≤ 2log 1+ α2P +N3 . (25) +1log(cid:18)1+ N2(cid:19) (33) 2 N In summary, 1 (cid:18) (cid:19) 1 α P • x11 decoded at receivers 1 and 2 T2 = 2log 1+ 2N2 (34) (cid:18) (cid:19) 2 1 (1 α )P (cid:18) (cid:19) R11 ≤T1(cid:48)1 = 2log 1+ (α0+−α2)P0 +N1 T3 ≥ 12log c3+ α P +PN +N . (35) (cid:18) (cid:19) 2 2 3 1 (1 α )P 0 wxherdeeRcco111d1e=≤d a(T1t(−1(cid:48)1r(cid:48)1−αec0=α)e0Pi)v2P+ePlro1g= 21c−−11αα+00. (α0+−α2)P +N2 SanindScteca3rαti=n0g=2−frNNoP1m22/P∈≥(cid:2)o0,f12r31.o(cid:3)m, itTafobllleowIIs, wtheatcca1n1e=xp12r−−esNNs22t//hPPe≥two25-, 10 R • dimensional outer bound region at R as (cid:18) (cid:19) 2 1 α P R10 ≤T10 = 2log 1+ N0 (26) (cid:26)1 (cid:18) 2P(cid:19) (cid:27) 1 R min log 1+ R ,C 1 2 1 ≤ 2 N − x decoded at receivers 1 and 2 1 • 2 (cid:26)1 (cid:18) P 7(cid:19) 1 (cid:18) P 4(cid:19)(cid:27) R2 ≤T2(cid:48) = 12log(cid:18)1+ α0Pα2+PN1(cid:19) (27) ≤min(cid:26)21log(cid:18)N1 ·23P(cid:19)−R2,2lo(cid:27)g N1 · 3 R2 ≤T2(cid:48)(cid:48) = 12log(cid:18)1+ α0Pα2+PN2(cid:19) (28) R3 ≤min(cid:26)12log(cid:18)1P+ N72(cid:19) −R2,1C3 (cid:18) P 4(cid:19)(cid:27) min log R , log . • x3 decoded at receivers 2 and 3 ≤ 2 N2 · 3 − 2 2 N3 · 3 (cid:18) (cid:19) 1 P R T = log c + Depending on the bottleneck of min , expressions, there 3 ≤ 3(cid:48) 2 3 (α +α )P +N {· ·} 0 2 2 are three cases: (cid:18) (cid:19) 1 P R3 ≤T3(cid:48)(cid:48) = 2log 1+ α2P +N3 (29) • R12lo≤g(cid:0)217l(cid:1)og(cid:0)R74(cid:1) 1log(cid:16)N3 7(cid:17) Notewthhaetre0c3 =c(1−α0P)P1,+Pc =+2−c1α0=. 1, and 1 c 1. •• R22 ≥ 124lo≤g(cid:16)NN232≤· 472(cid:17). N2 · 4 ≤ 11 ≤ 2 11 3 2 ≤ 3 ≤ (cid:16) (cid:17) Putting together, we can see that the following rate region is At R = 1log α2P 7 , the outer bound region is achievable. 2 2 N2 · 4 (cid:26) (cid:18) (cid:19) (cid:18) (cid:19)(cid:27) 1 P N 4 1 P 4 R1 ≤T1 =min{T1(cid:48)1,T1(cid:48)(cid:48)1}+T10 =T1(cid:48)(cid:48)1+T10 R1 min log 2 , log ≤ 2 α P · N · 3 2 N · 3 R T =min T ,T =T 2 1 1 2 ≤ 2 { 2(cid:48) 2(cid:48)(cid:48)} 2(cid:48)(cid:48) (cid:26)1 (cid:18) P 4(cid:19) 1 (cid:18) P 4(cid:19)(cid:27) R3 ≤T3 =min{T3(cid:48),T3(cid:48)(cid:48)} R3 ≤min 2log α P · 3 ,2log N · 3 . 2 3 where (cid:18) (cid:19) Depending on the bottleneck of min , expressions, we T = 1log c + (1−α0)P consider the following three cases: {· ·} 1 11 2 (α +α )P +N 0 2 2 +1log(cid:18)1+ α0P(cid:19) (30) • αN2P ≥αNP3 N 2 N • 2 ≤ 2 ≤ 3 1 α P N . 1 (cid:18) α P (cid:19) • 2 ≤ 2 T2 = 2log(cid:18)1+ α0P2+N2 (cid:19) (31) 1lCogas(cid:16)eα2i)Pα27P(cid:17) i≥s N3: The outer bound region at R2 = 1 P 2 N2 · 4 T log c + . (32) 3 3 ≥ 2 (α0+α2)P +N3 (cid:18) (cid:19) (cid:18) (cid:19) 1 P N 4 1 P 4 Thus, Theorem 8 is proved. R1 log 2 ,R3 log . (36) ≤ 2 α P · N · 3 ≤ 2 α P · 3 2 1 2 10 For comparison, let us take a look at the achievable rate Let us take a look at the achievable rate region. The first region. The first term of T is lower bounded by term of T is lower bounded by 1 1 T1(cid:48)(cid:48)1 = 121lloogg(cid:18)(cid:18)c211++Pα−2PPα−+2PN2(cid:19)N2 2(cid:19) ((3378)) T1(cid:48)(cid:48)1 = 121lloogg(cid:18)(cid:18)c211++Pα−2PPN−+2N(cid:19)2N2 2(cid:19) ((5523)) ≥ 12 (cid:18)5P (cid:19)3α2P ≥ 12 (cid:18)5P (cid:19)3N2 > log . (39) > log . (54) 2 3α2P 2 3N2 We get the lower bounds: We get the lower bounds: T =T +T (55) 1 1(cid:48)(cid:48)1 10 T1 =T1(cid:48)(cid:48)1+T10 (40) 1 (cid:18) P (cid:19) 1 (cid:18) N (cid:19) 1 (cid:18) P (cid:19) 1 (cid:18) N (cid:19) > log + log 1+ 2 (56) > log + log 1+ 2 (41) 2 3N2 2 N1 2 3α2P 2 N1 1 (cid:18) P (cid:19) 1 (cid:18) P N (cid:19) > log (57) > log 2 (42) 2 3N1 2 3α2P · N1 1 (cid:18)1 P (cid:19) 1 (cid:18)1 P (cid:19) T3 log + (58) T3 log + (43) ≥ 2 2 α2P +N2+N3 ≥ 2 2 α2P +N2+N3 1 (cid:18) P (cid:19) 1 (cid:18) P (cid:19) > log . (59) > log . (44) 2 3N3 2 3α P 2 (cid:16) (cid:17) For fixed α and R = 1log α2P , the following two- For fixed α2 and R2 = 12log(cid:16)α2N2P2(cid:17), the two-dimensional dimensional2rate regio2n is a2chievab2lNe.2 achievable rate region is given by (cid:18) (cid:19) (cid:18) (cid:19) 1 P 1 P R log , R log . (60) (cid:18) (cid:19) (cid:18) (cid:19) 1 3 1 P N 1 P ≤ 2 3N ≤ 2 3N R log 2 , R log . (45) 1 3 1 3 ≤ 2 3α2P · N1 ≤ 2 3α2P In all three cases above, by comparing the inner and outer bound regions, we can see that δ 1log(cid:0)3 4(cid:1) = 1, δ Cas(cid:16)eii)N2 (cid:17)≤α2P ≤N3:TheouterboundregionatR2 = 1log(cid:0)2 7(cid:1) = 0.91 and δ 11lo≤g(cid:0)23 4(cid:1) =· 31. Theref2or≤e, 1log α2P 7 is 2 · 4 3 ≤ 2 · 3 2 N2 · 4 wecanconcludethatthegapistowithinonebitpermessage. (cid:18) (cid:19) (cid:18) (cid:19) 1 P N 4 1 P 4 R log 2 , R log . (46) IV. INNERBOUND:CHANNELTYPE2 1 3 ≤ 2 α P · N · 3 ≤ 2 N · 3 2 1 3 Theorem 9: Given α [0,1], the region is defined by 1 α ∈ R Now, let us take a look at the achievable rate region. We (cid:18) (cid:19) 1 α P 1 have the lower bounds: R log 1+ 1 ≤ 2 N 1 T1 > 21log(cid:18)3αP2P · NN21(cid:19) (47) R2 ≤ 12log+(cid:18)12 + α1PP+N2(cid:19) 1 (cid:18)1 P (cid:19) 1 (cid:18)1 P (cid:19) T3 ≥ 2log 2 + α2P +N2+N3 (48) R3 ≤ 2log+ 2 + α1P +N3 , (cid:18) (cid:19) 1 P (cid:0)(cid:83) (cid:1) > 2log 3N . (49) and R= CONV α1Rα is achievable. 3 For fixed α and R = 1log(cid:16)α2P(cid:17), the two-dimensional A. Achievable Scheme achievable ra2te region2 is gi2ven by2N2 For this channel type, rate splitting is not necessary. Trans- R 1log(cid:18) P N2(cid:19), R 1log(cid:18) P (cid:19). (50) m1,i2t,s3ig.nIanlpxakrtiiscualacro,dxe2dasnigdnxal3oafreMlkat∈tic{e1-c,o2d,e.d..s,i2gnnRalks}u,skin=g 1 3 ≤ 2 3α P · N ≤ 2 3N the same pair of coding and shaping lattices. As a result, 2 1 3 the sum x +x is a dithered lattice codeword. The power Case iii) α P N : The outer bound region at R = 2 3 12log(cid:16)αN22P · 742(cid:17) is≤ 2 2 Eal[loxca3tio2n]=santiPsfi.eTshEe[(cid:107)rexc1e(cid:107)iv2e]d=siαgn1anlPs ,arEe[(cid:107)x2(cid:107)2] = nP, and (cid:107) (cid:107) 1 (cid:18) P 4(cid:19) 1 (cid:18) P 4(cid:19) y1 =[x2+x3]+x1+z1 R log , R log . (51) 1 3 ≤ 2 N · 3 ≤ 2 N · 3 y =x +x +z 1 3 2 2 1 2 For this range of α , the rate R is small, i.e., R = y3 =x3+x1+z3. 12log(cid:16)αN22P · 74(cid:17) ≤ 12l2og(cid:0)74(cid:1) < 12, a2nd R1 and R3 are c2lose The signal scale diagram at each receiver is shown in Fig. 3 to single user capacities C and C , respectively. (b). Decoding is performed in the following way. 1 3

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.