Content-Type Coding. arxiv: v2 [cs.it] 3 Jun Linqi Song UCLA

Transcription

1 Content-Type Coding Linqi Song UCLA Eail: Christina Fragouli UCLA and EPFL Eail: arxiv: v [cs.it] 3 Jun 015 Abstract This paper is otivated by the observation that, in any cases, we do not need to serve specific essages, but rather, any essage within a content-type. Content-type traffic pervades a host of applications today, ranging fro search engines and recoender networks to newsfeeds and advertiseent networks. The paper asks a novel question: if there are benefits in designing network and channel codes specifically tailored to content-type requests. It provides three exaples of content-type forulations to argue that, indeed in soe cases we can have significant such benefits. I. INTRODUCTION Coding traditionally ais to securely and efficiently convey specific inforation essages to one or ore receivers. This broad ai encopasses ost of the work in the field, fro the channel coding theore of Shannon [1], to recent breakthroughs in channel coding [], [3], network coding [4], and index coding [5]. However, counication networks today are increasingly used to serve a fundaentally different traffic, that delivers type of content rather than specific essages. As a siple exaple, when we use the Internet to access our bank account, we ask and want to see very specific inforation. However, if we search for a photo of a huing bird, we do not care which specific huing bird photos we receive - we do not even know what huing bird pictures are available - we only care about the content type, that it is a huing bird and not an owl photo. Content-type traffic pervades a host of applications today. For instance, content-delivery networks, such as the Akaai network, in any cases do not need to satisfy essagespecific requests, but instead content-type requests (eg., latest news on Greece, popular sites on CNN.co, etc; sart search engines and recoendation systes (eg. Google, Pandora generate in the ajority content-type traffic; advertiseent networks (eg. serving adds on hotels or cars, and newsfeeds on social networks (eg., cultural trends, following celebrities also fall in the content-type category. The fact that content fors a significant percentage of the Internet traffic has been well recognized especially in the networking counity; however, ost of the work looks at what to replicate, where and how to store and fro where to retrieve specific data. We ask a very different question: are there benefits in designing network and channel codes specifically tailored to content-type traffic? We ake the case that, indeed, if we realize that we need to serve content-type rather than specific essages, we can in soe cases achieve significant benefits. The fundaental reason content type coding helps, is that it offers an additional diension of freedo: we can select which, within the content type, specific essage to serve, to optiize different criteria. For instance, we can create ore coding opportunities, transfor ultiple unicast to a ulticast session, and adapt to rando erasure patterns, all of which can lead to higher throughput. For different content-type coding probles, there are different atheatical representation. It is non-trivial to find a consistent atheatical odel for a general contenttype coding proble and it is beyond the scope of this paper. In this paper, we study the content-type coding probles by providing several exaples to show the significant benefits of content-type coding over conventional essage-specific coding. First, we provide an exaple of benefits over networks: we consider a classical exaple in the network coding literature, the cobination network, and show the benefits in a content-type setup. Second, we provide an exaple of benefits over lossy networks: we consider a broadcast erasure channel with feedback setting, where a source wants to send essages to two receivers. Third, we review an exaple within the index coding setup, tered pliable coding, that we presented in a previous work, introduce an algebraic fraework for pliable coding and use this fraework to provide a novel lower bound. The paper is organized as follows. Section II considers content-type coding over the cobination network; Section III looks at broadcast erasure networks; Section IV introduces an algebraic fraework for pliable index coding; and Section V concludes the paper with a short discussion. II. CONTENT-TYPE CODING OVER NETWORKS Motivating exaple: Fig. 1 provides a siple exaple of content-type coding benefits. We start fro the classical butterfly network, with two sources, s 1 and s, two receivers, r 1 and r, and unit capacity directed links, where each source has essages of a given type. In the exaple, source s 1 (say advertising hotels has unit rate essages, {b 11, b 1 }, and source s (say advertising car rentals also has unit rate essages {b 1, b }. The in-cut to each receiver is two, thus she used all the network resources by herself, we could send to her two specific essages, for instance, to r 1 : b 11 and b 1 and to r : b 1 and b. However, if we want to send to both receivers these specific requests, it is clearly not possible, as there are no coding

2 b 11, b 1 b 1, b s 1 s r 1 r b 11, b 1 b 1, b (a Requesting specific essages. b 11, b 1 b 1, b s 1 s r 1 r b 11 or b 1 b 1 or b b 11 or b 1 b 1 or b (b Requesting one essage fro each type. Fig. 1: Exaple where each source has essages of a given type. opportunities; we can satisfy one receiver but not both. In contrast, if each receiver requires one essage fro each type, then we can ulticast say b 11 and b 1 to both receivers. Proble forulation: We consider a network represented as a directed graph G = (V, E, a set of sources {s 1, s,, s } and a set of n receivers {r 1, r,, r n }. Each source has u essages of the sae type, and different sources have different types of essages. We denote by B i = {b i1, b i,, b iu } the set of type-i essages fro source s i. Perforance etrics: Given a transission schee, we use the rate towards a receiver to easure the efficiency of the transission schee. In the conventional essagespecific coding proble, each receiver r j requests specific essages, one fro each type, denoted by a fixed eleent (b 1, b,, b B 1 B. For T transissions, the essages requested by receiver r j are denoted by (b T 1, b T,, b T, where each eleent is a vector b T i = (b i (1,, b i (t,, b i (T. We denote by R j the set of essages that the receiver r j can decode after T transissions, and we say that the transission rate towards r j is the nuber of requested essages that can be decoded by the receiver r j per transission, i.e., R j = 1 T T i=1 I {bi(t R j}. (1 In the content-type forulation, each receiver r j requests to receive (any one essage fro each type, and does not care which specific one. We denote the requested contenttype essages by an arbitrary eleent x 1, x,, x B 1 B. For T transissions, the essages requested by r j are denoted by (x T 1, xt,, xt, where each eleent is a vector x T i = (x i (t,, x i (t,, x i (T. Siilarly, the rate towards r j is defined as the nuber of requested essages that can be decoded by the receiver r j per transission, i.e., R c j = 1 T T i=1 I { xi(t R j, such that x i(t B i}. ( We would like to study the rate averaged aong all receivers for essage-specific coding, denoted by R, and the that for content type coding, denoted by R c.clearly, for essage specific coding, the rate depends upon the specific essage requests. We denote by R w the worstcase rate (iniizing aong all possible sets of requests, and by R a the average rate (averaged over all possible sets of requests. We define the worst case and average case gains, respectively, as: G w = Rc R w, A. Cobination-like network G a = Rc R a. (3 We consider the cobination network structure B(, k, u, where k. shown in Fig.. The network has four layers: the first layer is the sources and each source connects to every node in the second layer; the second layer has k interediate A nodes and each A node connects to a B node in the third layer, which also contains k interediate B nodes; the fourth layer contains n = ( k u receivers, where we have u receivers connected to each subset of size of the B nodes. Theore 1: In a B(, k, u network, G w = u, li k G a = u. (4 The proof of this theore follows fro Leas 1-3. Lea 1: In the network B(, k, u, content type coding achieves R c =. Proof : Use network coding to ulticast one specific essage fro each type to all receivers. Lea : In the network B(, k, u, the worst case essage-specific coding rate is R w = /u. Proof : We construct the following receiver requireents: for every u receivers that are connected to the sae B nodes, each request is an eleent of the set B 1 B, and no two requests are the sae. Since all u essages are requested, we need to use the sae set of A-B edges, at least u ties. Using network coding we can ensure that at the end of u transissions, each receiver gets all u essages and thus also the essages required by her; thus the transission rate is /u. Note that receiving all u essages by each receiver is a worst case scenario in ters of rate. Lea 3: In a network B(, k, u, the average rate of the essage-specific coding proble is bounded by R a u + u +1 + (1+ ln u u +1 + (+1 ln u. ( u k +1 Proof : We here provide a short outline and give the coplete proof in the appendix. We consider out of the k edges that connect A to B nodes, and the u receivers that can be reached only through these edges. We call these edges and u receivers a basic structure. We argue that, through each such structure, we need to send alost surely all essages to have a good average - with high probability all essages are required by less than (5

3 s 1 s s b 1 b b 3 A nodes A 1 A A k 1 A k B nodes B 1 B B k 1 B k c 1 c u c n u +1 c n Fig. : Cobination network structure. c 1 c c 3 c 4 c 5 c 6 c 7 c 8 Fig. 3: A coplete pliable coding instance. u 1 (1 δ 1 receivers. Thus the rate we can get through each basic structure converges to /u. III. CONTENT-TYPE CODING OVER ERASURE NETWORKS We here ake the case that, over erasure networks with feedback, we can realize benefits by allowing the rando erasures to dictate the specific essages within a content type that the receivers get - essentially we shape what we serve to the rando channel realizations. We consider the following content-type coding setup. A server, s, has k 1 essages of content-type 1, K 1, and k essages of content-type, K (eg. an ad serving broadcasting station in a all has k 1 sale coupons for clothes and k sale coupons for gadgets. Receiver c 1 wants to receive all the k 1 sale coupons for clothes and soe (a fraction α, any αk of the sale coupons for gadgets; receiver c wants the reverse, all the k coupons for gadgets and soe (any αk 1 of the coupons for clothes. The server is connected to the two receivers through a broadcast erasure channel with state feedback; in particular, when the server broadcasts a essage, each receiver c i receives it correctly with probability 1 ǫ i, independently across tie and across receivers. Each receiver causally acknowledges whether she received the essage or not. The corresponding essage-specific schee is as follows [6]. The server wants to send to c 1 all the essages in K 1 and in a specific subset of K, K 1 K, of size αk ; and to c, all the essages in K and in a specific subset of K 1, K1 K 1, of size αk. We again have independent broadcast erasure channels with feedback. Rate region: We say that rates (r 1, r, with r 1 = k 1 + αk T, r = k + αk 1, (6 T are achievable, if with T transissions by s both c 1 and c receive all they require. Strategy for essage-specific coding: The work in [6] proposes the following achievability strategy and proves it is capacity achieving. Recall that we use the subscript to indicate the content type, and the superscript for the receiver. Phase 1: The source repeatedly transits each of the essages in K 1 \K1 and K \K, 1 until one (any one of the two receivers acknowledges it has received it. Phase : The source transits linear cobinations of the essages in K, 1 K1, those in K 1 \K1 that were not received by c 1, and those in K \K 1 that were not received by c. The intuition behind the algorith is that, in Phase 1, each private essage (that only one receiver requests is either received by its legitiate receiver, or, it becoes side inforation for the other receiver. In Phase, each receiver either wants to receive each essage in a linear cobination or already has it as side inforation and can thus subtract it. The source creates linear cobinations that are innovative (bring new inforation to each receiver (eg., through unifor at rando cobining over a high enough field [6]. The strategy achieves the rate region: 1 ǫ 0 r 1 in{1 ǫ 1, } 1 (1 φ (1 α 1 ǫ 0 r in{1 ǫ, 1 } 1 (1 φ 1 (1 α r 1 ( 1 ǫ 1 1 αφ 1 1+α + r 1 1 (7 1 ǫ 1 1+α r ( r 1 ǫ 1 1+α 1 ǫ 1 αφ 1+α 1 where ǫ 1 = ǫ 1 ǫ, and φ i = (1 ǫ i /(1 ǫ 1 for i = 1,. Strategy for content-type coding: We propose the following strategy. Phase 1: For the essages in K 1, denote by K 1r the essages not yet transitted. Initially K 1r = K 1. 1 The server repeatedly broadcasts a essage in K 1r, until (at least one of the receivers acknowledges it has successfully received it. The essage is reoved fro K 1r. If c 1 receives the essages, she puts it into a queue Q 1 1. If c receives the essage, she puts it into a queue Q 1. The server continues with transitting a next essage in K 1r, until either K 1r becoe epty, or K 1r + Q 1 = αk 1. (8 The server follows the sae procedure for essage set K. Phase : The source transits linear cobinations of the essages in the three sets: K 1r K r, Q 1 \Q1 1, and Q1 \Q until both receivers are satisfied.

4 0.7 r =1 ε,r for content type coding 0.7 r =1 ε,r for content type coding 0.7 r =1 ε,r for content type coding Rate for receiver (r Gain,r for essage specific coding r 1 =1 ε 1 Rate for receiver (r ,r for essage specific coding Gain r =1 ε 1 1 Rate for receiver (r ,r for essage specific coding Gain r =1 ε Rate for receiver 1 (r Rate for receiver 1 (r Rate for receiver 1 (r 1 (a Case 1: α < in{φ 1, φ }; α = 0.5. (b Case : φ 1 < α < φ ; α = 0.7. (c Case 3: α > ax{φ 1, φ }; α = Fig. 4: Coparison of rate region, as defined in (6, by essage-specific and content-type coding, across three cases. The shaded regions show the gains of content-type over essage-specific coding. The channel paraeters are ǫ 1 = 0.4 and ǫ = 0.3, which give φ 1 = 0.68 and φ = The intuition behind the strategy is that, during Phase 1 we transit essages fro K 1 until we either run out of essages, or both receivers want to receive the reaining K 1r : c 1 because she wants all the essages in K 1 and c because, on top of the Q 1 she has already received, she also needs the K 1r to coplete the fraction αk 1. Note that originally, K 1r = k 1 and Q 1 = 0; at every step, the quantity in (8 either reains the sae (if c receives the essage, or reduces by one (if she does not. Siilarly for K. In the second phase, the source again transits linear cobinations of essages that either a receiver wants, or already has and can subtract to solve for what she wants. Using the above ethod, we can show the achievable rate of content-type broadcasting schee in the following theore. Theore : The rate region of the 1- content-type broadcasting counication with erasures is 0 r 1 in{1 ǫ 1, 1 ǫ } α 0 r in{1 ǫ, 1 ǫ 1 } α r 1 1 ǫ 1 [1 α(φ 1 α + ] + r 1 α (φ 1 α + (9 1 ǫ α r 1 ǫ [1 α(φ α + ] + r 1 α 1 (φ α + 1 ǫ 1 1 α where (x + = ax{x, 0}. In fact, this achievable schee achieves the capacity for -1 content-type broadcasting counication with erasures. The proof of achievability and converse of Theore is shown in Appendix B. Fig. 4 copares the rate regions for the content-type and essage coding. For content-type, we have three distinct cases, depending on the relative values of α and φ i. Note that φ i expresses the fraction of essages that c i receives during Phase 1. Thus, if α < in{φ 1, φ } (Fig. 4 (a, c 1 and c already receive αk 1 and αk essages during Phase 1; essentially broadcasting content-type essages coes for free, has not additional rate cost to providing c 1 with K 1 and c with K. If in{φ 1, φ } < α < ax{φ 1, φ }, say for instance φ 1 < α < φ (Fig. 4 (b, c receives the content-type essages for free, but for c 1 we need additional transissions in Phase. In α > ax{φ 1, φ } (Fig. 4 (c, c 1 and c require large percentages of essages fro another type; interestingly, when we have ax{φ 1, φ } < α < in{φ 1 /φ, φ /φ 1 }, we can achieve the point (1 ǫ 1, 1 ǫ, which iplies that, all transissions by s are useful for both receivers. Message-specific coding in general does not achieve this point. IV. CONTENT-TYPE CODING IN THE INDEX CODING FRAMEWORK Previous work has investigated a specific type of contenttype coding within the fraework of index coding, tered pliable index coding [7], [8]. In index coding we have a server with essages, and n clients. Each client has as side-inforation a subset of essages, and requires one specific essage. The goal is to iniize the nuber of broadcast transissions so that each client receives the essage she requested. In pliable index coding, the clients still have side inforation, but they no longer require a specific essage: they are happy to receive any essage they do not already have. The work in [7] [8] has shown that, although in the worst case, for index coding, we ay require O(n transissions, for pliable index coding we require at ost O(log n, i.e., we can be exponentially ore efficient. This result was derived by using a bipartite representation of the pliable index coding proble and a randoized construction. In this paper, apart fro drawing attention to the fact that pliable index coding is a for of content-type coding, we also ake two contributions: we derive an algebraic condition for clients to be satisfied in pliable index coding, and use this condition to prove a new lower bound: we show that there exist pliable index coding instances where Ω(log n transissions are necessary. Bipartite graph representation: We can represent a pliable index coding instance using an undirected bipartite graph, where on one side we have a vertex corresponding to each of the essages, say b 1,..., b and on the other side one vertex corresponding to each client, c 1,..., c n.

5 We connect with edges clients to the essages they do not have. For instance, in Fig. 3, client c 5 does not have (and would be happy to receive any of b 1 and b. A. An algebraic criterion for pliable index coding Assue that the source broadcasts K linear cobinations of the essages over a finite field F q ; that is, the k-th transission is α k,1 b 1 + α k, b + + α k, b, with a i,j F q. We can collect the K transissions into a K coding atrix A, where row k contains the linear coefficients (α k,1, α k,,, α k, used for the k- th transission. We also denote by α i the i-th colun vector of the atrix A. Then each client receives Ab = c, where b is the 1 vector that contains the essages and c is a constant vector, and needs using this and his side inforation to recover one essage he does not have. For client c j, let us denote by N[j] the set of indices of essages that c j does not have, and by N C [j] the set of indices of c j s side inforation. For exaple, N[] = {1} and N C [] = {, 3} for client c in the exaple in Fig. 3. Clearly, client c j can use his side inforation to reove fro the source transissions the essages in N C [j]; it thus can recover fro atrix A the subatrix A N[j] which only contains the coluns of A with indices in N[j]. That is, if b N[j] contains the essages he does not have, and c is a constant vector, he needs to solve A N[j] b N[j] = c, (10 to retrieve any one essage he does not have. Lea 4: In pliable index coding, client c j is satisfied by the coding atrix A if and only if there exists a colun α i, with α i / span{a N[j]\{i} }, for soe i N[j], (11 where span{a N[j]\{i} } = { l N[j]\{i} λ lα l λ l F q, α l F K q } is the linear space spanned by coluns of A N[j] other than α i. Proof : We are going to argue that, if such a colun α i exists, then client c j can uniquely decode b i he does not have fro his observations, and thus is satisfied. The condition α i / span{a N[j]\{i} } iplies that any vector in the null-space N (A N[j] has a zero value in position i; indeed, since colun α i is not in the span of the other coluns, then for every vector x N (A N[j], the only way we can have l N[j] x lα l = 0 is for x i = 0. But the solution space of any linear equation can be expressed as a specific solution plus any vector in the nullspace; and thus fro (10 we can get a unique solution for b i if and only if any vector x in N (A N[j] has a zero value in the eleent corresponding to i, as is our case. We can retrieve b i by colun and row transforations in (10. B. A Lower Bound Theore 3: There exist pliable index coding instances that require Ω(log(n broadcast transissions. Proof : We constructively prove this theore by providing a specific such instance, that we ter the coplete instance. In the coplete instance, we have a client for every possible side inforation set corresponding to a client. In this case, the client vertex set A corresponds to the power set B of the essage vertex set B, and we have = log (n (note that we can have an arbitrary nuber of additional essages and assue that all clients already have these. We define W d (A to be the set of client vertices that have a degree d (d = 0, 1,,, i.e., the vertices that need d essages. An exaple of the coplete instance with = 3 is show in Fig. 3. Obviously, we can trivially satisfy all clients with = log (n transissions, where each b i is sequentially transitted once. We next argue that we cannot do better. We will use induction to prove that the rank of the coding atrix A needs to be at least for the receivers to be satisfied according to Lea 4. Let J denote an index subset of the coluns; in the coplete odel, Lea 4 needs to hold for any subset J. For J = 1, i.e., to satisfy the clients who iss only one essage, no colun of the coding atrix A can be zero. Otherwise, if for exaple, colun i 1 is zero, then the client who requests essage b i1 cannot be satisfied. So rank(a J = 1 for J = 1. Siilarly, for J =, any two coluns of the coding atrix ust be linearly independent. Otherwise, if for exaple, coluns i 1 and i are linearly dependent, then α i1 span{α i } and α i span{α ii }, and the clients who only iss essages b i1 and b i cannot be satisfied. So rank(a J =. Suppose we have rank(a J = l for J = l. For J = l + 1, we can see that if all clients who only iss l + 1 essages can be satisfied, then for soe i J, we have α i / span{a J\{i} }. Therefore, rank(a J = rank(α i + rank(a J\{i} = 1 + l. Therefore, to satisfy all the clients, the rank of the coding atrix α is, resulting in K, fro which the result follows. V. CONCLUSIONS AND SHORT DISCUSSION This paper introduced a new proble forulation, that we tered content-type coding. Although there is significant work in content-distribution networks, the work still considers essage-specific requests, where now ultiple requests are interested in the sae essage. The research questions are focused ostly on where to store the content, how uch to replicate, how uch to copress, and what networks to use to upload and download content. There is also a rich literature in network coding and index coding, yet as far as we know, apart fro the pliable index coding forulation, there has not been work that has looked at content-type coding. We discussed in this paper three exaples, where if we realize that we need to serve content-type rather than specific essages, we can have significant benefits. We believe that there are any ore scenarios where we can realize benefits, as, downloading content-type rather than essage-specific content, can help all aspects of content

6 distribution networks, ranging fro storage to coding to delivery. Even in the specific forulations we proposed in this paper, there are several followup directions and extensions, for instance looking at ore than two users and ultihop counication over erasure networks. REFERENCES [1] C. E. Shannon, A atheatical theory of counication, ACM SIGMOBILE Mobile Coputing and Counications Review, vol. 5, no. 1, pp. 3 55, 001. [] E. Arikan, Channel polarization: A ethod for constructing capacityachieving codes for syetric binary-input eoryless channels, Inforation Theory, IEEE Transactions on, vol. 55, no. 7, pp , 009. [3] S. Kudekar, T. J. Richardson, and R. L. Urbanke, Threshold saturation via spatial coupling: Why convolutional LDPC ensebles perfor so well over the BEC, Inforation Theory, IEEE Transactions on, vol. 57, no., pp , 011. [4] R. Ahlswede, N. Cai, S.-Y. Li, and R. W. Yeung, Network inforation flow, Inforation Theory, IEEE Transactions on, vol. 46, no. 4, pp , 000. [5] Y. Birk and T. Kol, Infored-source coding-on-deand (ISCOD over broadcast channels, in INFOCOM 98. Seventeenth Annual Joint Conference of the IEEE Coputer and Counications Societies., vol. 3, pp vol.3, Mar [6] L. Georgiadis and L. Tassiulas, Broadcast erasure channel with feedback - capacity and algoriths, in Network Coding, Theory, and Applications, 009. NetCod 09. Workshop on, pp , June 009. [7] S. Braha and C. Fragouli, Pliable index coding, in Inforation Theory Proceedings (ISIT, 01 IEEE International Syposiu on, pp , 01. [8] S. Braha and C. Fragouli, Pliable index coding: The ultiple requests case, in Inforation Theory Proceedings (ISIT, 013 IEEE International Syposiu on, pp , 013. APPENDIX A PROOF OF LEMMA 3 Let us denote by E = {e 1, e,, e k } the set of k edges connecting A nodes and B nodes. We refer to different edges in E and the u receivers that are only connected to the, as a basic structure. Fro construction, there are ( k such structures. For each structure, it is straightforward to see that: For any basic structure, the axiu transission rate through these edges to the receivers in this structure is (since the in-cut is. Denote by s ij the nuber of requireents of essage b ij for receivers in a basic structure (i.e., the nuber of receivers requesting this essage. Then the axiu transission rate is l=1 ax l s ij /u, where ax l s ij represents the l-th axiu nuber aong {s i,j 1 i, 1 j u}. This eans that the axiu transission rate is achieved by transitting the ost popular essages through these edges. Consider any given basic structure. We first observe that E[s ij ] = u 1. We define the event E δ1 i,j = {s ij > E[s ij ](1 + δ 1 } as an abnoral event with respect to the essage b ij. We also define the event E δ1 = 1 i,1 j u E δ1 i,j as an abnoral event for this basic structure. Fro the Chernoff bound, we have that Pr{E δ1 i,j } = Pr{s ij > E[s ij ](1 + δ 1 } e u 1δ 1 3, (1 and p 1 = Pr{E δ1 } = u Pr{E δ1 i,j } u 1 δ 1 ue 3. (13 Here we denote by p 1 the probability that an abnoral event happens. When an abnoral event happens for this basic structure, the rate for this structure is (at ost. If an abnoral event does not happen for this basic structure, the rate for this structure is (at ost u 1 (1+δ 1 /u = (1 + δ 1 /u. Next, we consider the v = ( k structures. Let us denote by T E the nuber of structures with an abnoral event happening. The expected value of T E is vp 1. The probability that {T E > vp 1 (1 + δ } happens, denoted by p, can be bounded using the Chernoff bound: p = Pr{T E > vp 1 (1 + δ } e vp 1 δ 3, (14 Hence, if the event {T E > vp 1 (1+δ } does not happen, the nuber of structures with an abnoral event is at ost vp 1 (1 + δ. Therefore, the average rate can be bounded by R a p + (1 p [ vp1(1+δ+(v vp1(1+δ (1+δ 1 u v ] p + p 1 (1 + δ + (1+δ1 Let us set δ 1 = Then we have u. 3 (+3 ln u u 1 p 1 u +1, p u +1. and δ = (15 3 (+1 ln u vp1. (16 Plugging e.q. (16 into e.q. (15, we can have an upperbound for the average rate: R a p + p 1 (1 + δ + (1+δ1 u u + u +1 + (1+ ln u u +1 + (+1 ln u. (17 ( u k +1 Setting k = h 1, and = h u, where h 1 and h are constants, we have, as k. A. Achievability R a u, (18 APPENDIX B PROOF OF THEOREM To prove this theore, we assue that k 1 and k are large. Recall that we define ǫ 1 = ǫ 1 ǫ, and φ i = (1 ǫ i /(1 ǫ 1 for i = 1,. Let us denote by k 1 and k the nuber of essages transitted fro sets K 1 and K at the end of phase 1. Therefore, the average nuber of transissions needed to coplete phase 1 is: N 1 = k 1 + k 1 ǫ 1. (19

7 W W k1 1 k Enc n X S 1t, py (, y x 1 n Y1 n Y Dec1 Dec ˆk1 k W 1 W ˆ ˆ k W1 1 ˆk W py (, y x py ( xpy ( x 1 1 S t, py ( x i x, e, wp 1- wp i i On average, the nuber of essages fro set K j (j = 1, received by receiver i (i = 1, is: Fig. 5: Content type coding odel for the broadcast erasure channel with feedback. M i j = k j φ i. (0 Fro the algorith, we know that k 1 k 1 = (αk 1 M 1 + and k k = (αk M 1 +. Therefore, we have k i = [1 (α φ i + 1 φ i ]k i. (1 In phase, the required nuber of essages for receiver i (i = 1, is then kr i = (k i M i i + (k 1 k 1 + (k k, ( where the first ter is the nuber of erased essages fro the set K i that are received by another receiver, the second and third ters are the reaining essages to be transitted. For phase, the average nuber of transissions needed is N = ax{ k1 r kr, }. (3 1 ǫ 1 1 ǫ Then, the rate region can be calculated as: { (r1, r : r 1 0, r 0, r 1 = k1+αk T, r = k+αk1 T, N 1 + N T } (4, where T is an auxiliary variable and can be cancelled out. Plugging (19 and (3 into (4, we get the theore. Note that for ax{φ 1, φ } < α < in{φ 1 /φ, φ /φ 1 }, the conditions are siplified as r 1 1 ǫ 1 and r 1 ǫ, iplying that the axiu rates can be achieved. B. Converse To prove the converse of the theore, we use an inforation theory ethod to show that this rate region is tight. We first depict the syste odel in Fig. 5. Recall that we denote by s the transitter, and K 1 and K the sets of two types of essages. We denote by k 1 and k the cardinalities of K 1 and K. Two receivers request inforation essages fro the server s. Receiver 1 requests all inforation essages fro K 1 and α percent of the inforation essages fro K. In contrast, receiver requests all inforation essages fro K and α percent of the inforation essages fro K 1. Note that in our content type coding proble, receiver 1 is satisfied as long as she successfully receives αk essages fro K, disregarding of which ones, so is the sae for receiver. We consider broadcasting over an erasure channel with feedback. In this channel, when a essage x is transitted, the receiver i has a probability of 1 ǫ i to receive the essage correctly and has a probability of ǫ i to receive nothing (i.e., the essage is erased for receiver i. A essage is received or erased independently for receivers 1 and. We ai to find the capacity region (r 1, r of for this broadcast channel. Without loss of generality, let us assue W 1 = W = X = Y 1 1 = Y 1 =. To prove the converse, we just need to show the first and the third equations in (9, and then according to the syetry, we get the whole set of equations. For the first equation, it is equivalent to point-to-point counication, so we can directly get it. For the third equation, we consider two parts: k 1 + αk 1, 1 ǫ 1 1 ǫ 1 First, we consider n n H(X t n H(X t Y t 1 1, S t 1 = n [H(X t Y1 t 1, Y t 1, S t 1 k 1 1 ǫ 1 + k 1 ǫ 1, + I(X t ; Y t 1 Y1 t 1, S t 1 = n [H(X t W k1 1, W k, Y 1 t 1, Y t 1, S t 1 + I(X t ; Y t 1 Y1 t 1, S t 1 + I(X t ; W k1 1, W k t 1 Y1, Y t 1, S t 1 ] and the following two conditions: (5 (k 1 + k nε n I(W k1 1, W k ; Y 1 n, Y n, Sn = n I(Y 1,t, Y,t, S t ; W k1 1, W k t 1 Y1, Y t 1, S t 1 = n I(Y 1,t, Y,t ; W k1 1, W k t 1 Y1, Y t 1, S t 1, S t = n I(X t ; W k1 1, W k t 1 Y1, Y t 1, S t 1 Pr(S t 0 = (1 ǫ 1 n I(X t ; W k1 1, W k t 1 Y1, Y t 1, S t 1, (6

8 where the above conditions hold due to the Fano s inequality and the incidence property of S t, and k 1 nε n 1 ǫ 1 I(W k1 1 ; X t Y1 t 1, S t 1 n k 1 1 ǫ 1 [I(X t ; W k1 1 Y t 1 1, Y t 1, S t 1 + I(X t ; Y t 1 Y1 t 1, S t 1 ] + n [I(X t ; Y t 1 Y1 t 1, S t 1 ], (7 where the last inequality follows fro (using the sae idea as (6 k 1 I(W k1 1 ; Y 1 n, Y n, Sn = (1 ǫ 1 n I(X t ; W k1 t 1 1 Y1, Y t 1, S t 1. (8 Plugging (6 and (7 into (5, we get the first part of the third equation in (9. Siilarly, we can get the second part of the third equation in (9 using (k 1 + αk nε n I(W k1 1, W αk ; Y1 n, S n = (1 ǫ 1 n I(X t ; W k1 1, W αk Y1 t 1, S t 1. (1 ǫ 1 n H(X t (1 ǫ 1 n. (9