Introduction
Anonymous agents must often agree on distinct roles before any higher-level distributed task can begin. This symmetry-breaking step appears as leader election, naming, rendezvous, and channel or resource allocation. The agents may be radio devices or identical software programs (for instance, copies of the same AI model started from the same prompt) that must accomplish a task jointly without a coordinator. These problems tend to be harder when every agent starts from the same prior information, executes the same algorithm, and observes only a private noisy response to the joint actions. Shared randomness does not by itself break the symmetry, since all users see the same random value. Independent private randomness creates different tentative identifiers, but two users may still draw the same one, and they have to find out through the channel whether or not this has happened.
Our main motivation is the decentralized multiplayer multi-armed bandit problem. In this model, J J J players repeatedly choose among K K K arms and observe private rewards. The players label the arms consistently, and they act in synchronized time slots, so that they share a common clock but no common identity. Simultaneous use of the same arm causes a collision and suppresses the corresponding rewards. Algorithms with strong regret guarantees use collisions to assign roles and exchange statistics [1 , 2 , 3 , 4 ] ; see [5 ] for a recent survey. State-of-the-art algorithms typically start with an initialization phase in which the players obtain distinct ranks, and the simplest rule to accomplish this is called musical chairs [6 ] : each player selects a random arm until they receive a nonzero reward, and then keeps selecting that arm, since without a nonzero reward proves that nobody else pulled it. Random ranks are also used to break the symmetry in [7 ] , and initialization without collision information is studied in [8 ] . The resulting channel model was formalized as the bandit channel in [9 ] . Unlike a conventional multiple-access channel, each player is simultaneously a transmitter and a receiver, its own action is known as receiver side information, and interference is useful because it carries information about the actions of the other players.
Two architectures come to mind for symmetry breaking over such a channel. In the first, each user draws a long random name and transmits the codeword of a common codebook indexed by its name; every user then decodes the names of the others, and ranking the names gives distinct identifiers unless two names coincide—an event of exponentially small probability in the number of bits used to encode the name—or decoding fails. This turns symmetry breaking into all-source multicast over a multiway channel. The two-user case is naturally related to Shannon’s two-way channel [10 ] , for which a complete single-letter capacity characterization is unavailable in general [11 ] , [12 ] . With more than two users, each receiver must recover every other user’s message. A tractable inner bound is nevertheless obtained by independent random coding and simultaneous decoding: at a fixed receiver, the other J − 1 J-1 J − 1 codewords form a multiple-access channel with the receiver’s own codeword as side information. In [9 ] , two players communicate over the bandit channel with a common codebook and inputs that do not depend on the outputs: the random-coding exponent and the achievable rate are derived, and Shannon’s outer bound for the two-way channel is evaluated as an upper bound on the capacity. In [13 ] , block coding with random names breaks the symmetry
R. Zhang is with INRIA, France. R. Combes and S. Yang are with CentraleSupélec, Paris-Saclay University, France. This paper is supported by the CHIST-ERA Wireless AI 2022 call MLDR project (ANR-23-CHR4-0005), partially funded by AEI and NCN under projects PCI2023-145958-2 and 2023/05/Y/ST7/00004, respectively. This paper was presented in part at the 2025 IEEE Information Theory Workshop (ITW), Sydney, Australia, September–October 2025, and at the 2026 IEEE International Symposium on Information Theory (ISIT), Guangzhou, China, June–July 2026. between two players on that channel, with codebooks that time-share between input laws; its error exponent is derived under maximum-likelihood and universal decoding, and the regret of the resulting initialization phase of a bandit algorithm is bounded, with a minimax lower bound. The two-user, noiseless case of our block-coding results with a single input law is therefore in [9 ] , [13 ] (see Section VI-C), and so is the value of our name-decoding ceiling for two users and arms with equal means (Proposition 15). The change-of-measure converses, block coding and the cut-set bound for J ≥ 3 J \ge3 J ≥ 3 users, the ceiling on the name-decoding exponent, the interactive role controllers and noisy collisions are not in [9 ] , [13 ] , and universal decoding and regret are not treated here. In the second architecture, the users contend directly for the identifiers in [ J ] : = { 1 , … , J } [J] := \{1,\ldots,J\} [ J ] := { 1 , … , J } . A user keeps its tentative identifier as long as its outputs suggest that nobody else holds it, and draws a new one otherwise, as in musical chairs. No name is decoded: the channel itself resolves the conflicts.
In this paper, we consider the problem of determining how fast the probability of error of symmetry breaking can decay with the number of channel uses, for every strategy. For those two architectures we quantify their error exponents, and we show that, on the noisy bandit channel below, the second architecture can exceed every name-decoding exponent of the first, that is, every exponent of the probability that two names coincide or are decoded wrongly: provably with noiseless collisions, and according to numerical estimates with noisy ones.
We consider a memoryless channel whose law is unchanged when the users are permuted: permuting the inputs of the users permutes their outputs in the same way, and the outputs may be arbitrarily correlated given the inputs. All users apply the same strategy: the input of a user may depend on its own past outputs, the users may use common and private randomness, and each user outputs a number in [ J ] [J] [ J ] . The symmetry is broken if the J J J numbers are distinct. Section II gives the formal definition.
We then specialize to a noisy bandit channel , in which arm x x x yields a Bernoulli reward with probability μ x \mu_x μ x when it is pulled by a single user and with probability ε x \varepsilon_x ε x when it is shared. The case ε x = 0 \varepsilon_x = 0 ε x = 0 for every arm is the bandit channel of [9 ] , where a reward certifies that the user was alone on its arm. Allowing ε x > 0 \varepsilon_x > 0 ε x > 0 is not meant as a detailed model of collisions, although a collision does not need to suppress every reward, for instance with capture in wireless systems or imperfect collision detection, and nonzero rewards on collisions are also considered in multiplayer bandits [14 ] . It is the simplest departure from certified outputs: the outputs stay binary and conditionally independent, the channel stays symmetric, and only a collision can change the law of an output—but a reward no longer proves that the user was alone. For ε x = 0 \varepsilon_x = 0 ε x = 0 , a single observation can certify a role, musical chairs and its generalization apply, and, for two users, the converse can be tight. For 0 < μ x < 1 0 < \mu_x < 1 0 < μ x < 1 and 0 < ε x < 1 0 < \varepsilon_x < 1 0 < ε x < 1 , every output has positive probability, no observation is conclusive, the users have to accumulate evidence over many uses, and a gap remains between the converses and the schemes. When ε x = μ x \varepsilon_x = \mu_x ε x = μ x for every x x x , collisions are invisible and symmetry breaking is impossible.
The objective of assigning distinct names to initially indistinguishable processes is classical in distributed computing [15 ] , [16 ] . Under the names naming and initialization , randomized wireless networking protocols assign identifiers to anonymous stations in single- and multichannel collision models [17 ] , [18 ] . More recent work treats anonymous beeping networks, where a listening node learns only whether at least one neighbor beeped [19 ] , [20 ] , also with noisy beeps [21 ] . That literature primarily optimizes rounds, energy, random bits, and success probability for particular distributed models. Indistinguishability arguments give lower bounds for contention resolution over collision channels [22 ] , [23 ] , and optimal protocols are known for two parties [24 ] ; the collision channel without feedback [25 ] is an information-theoretic model of uncoordinated access in which the users have distinct protocol sequences. Our symmetry-breaking criterion is the same underlying task, but the object of study is its reliability as a function of the number of channel uses n n n over a general noisy multiway channel.
The closely related leader-election task asks each user to output a bit, with success when exactly one user outputs 1 and becomes the leader while other users become followers; stronger formulations additionally require identifying or communicating with the elected leader [26 ] . A symmetry-breaking protocol also solves the task of leader election if the user whose identifier is 1 elects itself as the leader by outputting 1, so that an election failure is contained in a symmetry-breaking failure. For J = 2 J = 2 J = 2 the tasks are exactly equivalent, since an election bit L j L_j L j is mapped back to the identifier 2 − L j 2 - L_j 2 − L j without additional observations. For J > 2 J > 2 J > 2 , symmetry breaking requires all J J J identifiers to be different, whereas election may leave the followers indistinguishable. Neither criterion additionally requires a certificate or common knowledge of success.
Both tasks are special cases of a population version of symmetry breaking. Given L L L states and a population vector N = ( N 1 , … , N L ) \mathbf{N} = (N_1,\ldots,N_L) N = ( N 1 , … , N L ) with N 1 + ⋯ + N L = J N_1+\cdots+N_L = J N 1 + ⋯ + N L = J , the users succeed when exactly N l N_l N l of them choose state l l l , for every l ∈ [ L ] l \in[L] l ∈ [ L ] . Our task is N = ( 1 , … , 1 ) \mathbf{N} = (1,\ldots,1) N = ( 1 , … , 1 ) , and leader election is N = ( 1 , J − 1 ) \mathbf{N} = (1,J-1) N = ( 1 , J − 1 ) . Our task is the finest of these: if every user maps its identifier through a fixed map from [ J ] [J] [ J ] to [ L ] [L] [ L ] that sends exactly N l N_l N l identifiers to state l l l , for instance one that splits the ranks into consecutive blocks of sizes N 1 , … , N L N_1,\ldots,N_L N 1 , … , N L , a protocol for N = ( 1 , … , 1 ) \mathbf{N} = (1,\ldots,1) N = ( 1 , … , 1 ) becomes a protocol for N \mathbf{N} N with no larger error probability. Every achievability result of this paper therefore holds for every population vector, and so do the converses of Section III, for every population vector with at least two positive entries (Proposition 2). Apart from this, the rest of the paper treats N = ( 1 , … , 1 ) \mathbf{N} = (1,\ldots,1) N = ( 1 , … , 1 ) only.
The differences between symmetry breaking and two related information-theoretic problems are worth mentioning. Coordination capacity includes task assignment without duplication, for which the target law is uniform over the permutations of the tasks [27 ] . Its terminals, however, have fixed labels and use terminal-dependent mappings over prescribed communication links. Unsourced random access instead gives statistically identical transmitters a common codebook and asks one central receiver to recover an unordered message list [28 ] . Here there is no central observer: every anonymous user is simultaneously a transmitter and a receiver, sees a different private output, and must infer its own place in a common ordering. The distinction is already visible before communication. If C C C is common randomness and Π C \Pi_C Π C is a random permutation, labeled terminals may set their values to Π C ( j ) \Pi_C(j) Π C ( j ) , which is collision-free, whereas identical anonymous maps of C C C produce the same value for every user. Private randomness can create different tentative names, but the channel must let the users determine their relative order consistently.
Our converse is a change-of-measure argument in the spirit of the meta-converse [29 ] and of its Rényi form [30 ] ; with noninteracting references, its exponent has the form of Haroutunian’s sphere-packing bound for channels with feedback [31 ] . Change-of-measure lower bounds that hold for adaptive strategies are also standard in bandit problems [32 , 33 ] .
Finally, when there are more arms than players, the goal of the players should be to cycle through the arms in a fair fashion while avoiding collisions. This task is sometimes called orthogonalization [1 ] and can be reduced to symmetry breaking as well.
To summarize, the symmetry-breaking objective is ubiquitous in distributed systems; our emphasis is its information-theoretic reliability. To the best of our knowledge, the present work is the first Shannon-theoretic treatment of anonymous symmetry breaking over a general noisy symmetric multiway channel, with private receiver observations, identical causal strategies, general converses, reliability exponents, and both coding and interactive achievability.
Contributions
The main contributions are summarized as follows.
Formulation. We cast anonymous symmetry breaking as a Shannon-theoretic problem for J J J users over a general symmetric memoryless channel, with the error probability of the best strategy and its reliability exponent as the quantities of interest (Section II).
A converse for every strategy. We prove finite-blocklength converses on the probability of error for every symmetric causal strategy, including adaptive ones and those using arbitrary common and private randomness (Section III). We compare the channel with a reference under which the identifiers of m m m users are conditionally i.i.d. given a latent variable, so that they collide with probability at least that of i.i.d. uniform draws. A Rényi change of measure transfers this probability to the true channel, at a cost that is the n n n -th power of a one-step quantity of a nonnegative kernel defining the reference. The bound is optimized over the Rényi order and the kernel. Two instances are explicit: a noninteracting reference, whose exponent reduces to a KL-radius formula, and a common-output reference, which keeps only the trajectories on which the selected users receive identical outputs and is effective when collisions are detected without noise. The converses extend to every population vector with at least two positive entries. In the examples we evaluate kernels with m = 2 m=2 m = 2 selected users only, and the exponents of the resulting bounds do not depend on J J J beyond the distinction between J = 2 J=2 J = 2 and J ≥ 3 J \ge3 J ≥ 3 .
Block coding and the limit of name decoding. A shared random codebook and independent private names reduce achievability to all-source multicast, and the error probability is at most the probability that two names coincide, plus the multicast decoding error. For arbitrary fixed J J J , we prove a Gallager random-coding exponent for the error probability of symmetry breaking, which for i.i.d. inputs is positive if and only if I ( X − j ; Y j ∣ X j ) > 0 I(X_{-j};Y_j \mid X_j)>0 I ( X − j ; Y j ∣ X j ) > 0 , and is at most 1 J I ( X − j ; Y j ∣ X j ) \frac{1}{J}I(X_{-j};Y_j \mid X_j) J 1 I ( X − j ; Y j ∣ X j ) . A multiway cut-set bound shows that no name-decoding scheme, block or interactive, has a name-decoding exponent above log J / ( J − 1 ) \log J/(J-1) log J / ( J − 1 ) . Symmetry breaking itself is not subject to this limit: for J = 2 J=2 J = 2 , a channel whose output pair is uniform on { ( 0 , 1 ) , ( 1 , 0 ) } \{(0,1),(1,0)\} {( 0 , 1 ) , ( 1 , 0 )} for any input breaks the symmetry in a single use, although it carries no information (Section IV).
Interactive schemes. We introduce role controllers (§ 5 ), in which each user runs a probabilistic automaton, in the spirit of the finite-memory automata studied for hypothesis testing and the two-armed bandit [34 , 35 , 36 ] : it holds a tentative identifier and a state updated from its own outputs. Their error probability is the exact probability of a failure state of a Markov chain on the joint states, and every symmetric causal strategy, block coding in particular, is a role controller. When some output certifies that a role is not shared, a finite-state controller achieves an error exponent of at least − log -\log − log of a spectral radius, with equality under a condition that holds in the cases we evaluate. When every output has positive probability, no time-invariant finite-state controller has a vanishing error probability, as time-invariant finite memory keeps the error of hypothesis testing away from zero [35 ] . On the noisy bandit channel with J J J arms of common parameters, a state set of bounded size with maps that vary with time does give a vanishing error probability, by a device of Cover [34 ] , but whether it can give a positive exponent is open. We analyze confidence controllers, which accumulate evidence in a counter whose range grows with n n n . We also introduce generalized musical chairs , a finite-state scheme in which a user avoids the roles on which it recently failed. It contains musical chairs as the special case of no exclusion.
The noisy bandit channel. We define a noisy bandit channel with one pair of reward probabilities per arm (μ x \mu_x μ x for a user alone on arm x x x , ε x \varepsilon_x ε x for each user sharing it), for which we obtain explicit converses, in closed form for arms with equal parameters, the Gallager function of block coding, and the exponents of the interactive schemes (Section VI). For two users, one informative arm and a silent arm (an arm that never yields a reward, μ 2 = ε 2 = 0 \mu_2=\varepsilon_2=0 μ 2 = ε 2 = 0 ), and noiseless collisions ε 1 = 0 \varepsilon_1=0 ε 1 = 0 , the converse and one-arm musical chairs match, and the error exponent is − log ( 1 − μ 1 / 2 ) -\log(1-\mu_1/2) − log ( 1 − μ 1 /2 ) ; for two users and two noiseless arms of mean one, musical chairs matches the converse as well. With noiseless collisions, musical chairs
provably exceeds every name-decoding exponent, for every number J J J of users and J J J arms of a common mean below one. With noisy collisions, our numerical estimates indicate that the confidence controllers also exceed it, but we have no proof that their exponent is positive (Table I). Generalized musical chairs goes further: for three users and three noiseless arms of mean μ = 1 \mu= 1 μ = 1 , its exponent is more than twice that of musical chairs. How the optimal exponent depends on J J J is left open. The remainder of the paper is organized as follows. Section II introduces the general channel and the symmetry-breaking problem, and Section III proves the converses. Section IV develops block coding and the limits of name decoding, and Section V the interactive schemes. Section VI applies the results to the noisy bandit channel and compares the schemes with the converses, and Section VII concludes. Proofs that would interrupt the main flow are presented in the appendices.
Symmetric Channels and the Symmetry-Breaking Problem
Notation
For a positive integer m m m , let [ m ] : = { 1 , … , m } [m] := \{1,\ldots,m\} [ m ] := { 1 , … , m } . For S ⊆ [ J ] S \subseteq[J] S ⊆ [ J ] , write X S : = ( X j : j ∈ S ) X_S := (X_j : j \in S) X S := ( X j : j ∈ S ) and S c : = [ J ] ∖ S S^c := [J] \setminus S S c := [ J ] ∖ S . For j ∈ [ J ] j \in[J] j ∈ [ J ] , write X − j : = X [ J ] ∖ { j } ∈ X J − 1 X_{-j} := X_{[J]\setminus\{j\}} \in\mathcal{X}^{J-1} X − j := X [ J ] ∖ { j } ∈ X J − 1 , with the indices in increasing order. For a finite set A \mathcal{A} A , S ⊆ [ J ] S \subseteq[J] S ⊆ [ J ] , u ∈ A ∣ S ∣ u \in\mathcal{A}^{|S|} u ∈ A ∣ S ∣ and v ∈ A ∣ S c ∣ v \in\mathcal{A}^{|S^c|} v ∈ A ∣ S c ∣ , let u ⊔ S v ∈ A J u \mathbin{\sqcup_S} v \in\mathcal{A}^J u ⊔ S v ∈ A J denote the vector z z z with z S = u z_S = u z S = u and z S c = v z_{S^c} = v z S c = v , the components of u u u and v v v being read in increasing index order. Thus x [ J ] = x S ⊔ S x S c x_{[J]} = x_S \mathbin{\sqcup_S} x_{S^c} x [ J ] = x S ⊔ S x S c for every x [ J ] ∈ X J x_{[J]} \in\mathcal{X}^J x [ J ] ∈ X J . For a singleton, we abbreviate ⊔ { j } \sqcup_{\{j\}} ⊔ { j } as ⊔ j \sqcup_j ⊔ j , so that u ⊔ j v = ( v 1 , … , v j − 1 , u , v j , … , v J − 1 ) u \sqcup_j v = (v_1,\ldots,v_{j-1},u,v_j,\ldots,v_{J-1}) u ⊔ j v = ( v 1 , … , v j − 1 , u , v j , … , v J − 1 ) for u ∈ A u \in\mathcal{A} u ∈ A and v ∈ A J − 1 v \in\mathcal{A}^{J-1} v ∈ A J − 1 , x [ J ] = x j ⊔ j x − j x_{[J]} = x_j \sqcup_j x_{-j} x [ J ] = x j ⊔ j x − j , and ( u , v ) ↦ u ⊔ j v (u,v) \mapsto u \sqcup_j v ( u , v ) ↦ u ⊔ j v is a bijection A × A J − 1 → A J \mathcal{A} \times\mathcal{A}^{J-1} \to\mathcal{A}^J A × A J − 1 → A J . A tuple written with a comma, such as ( u , v ) (u,v) ( u , v ) or ( x j , x − j ) (x_j,x_{-j}) ( x j , x − j ) , always denotes the plain concatenation of its entries. A superscript n n n denotes a length-n n n sequence and X n [ i ] X^n[i] X n [ i ] denotes its i i i -th entry. All alphabets are finite, all logarithms are natural, and rates are measured in nats per channel use. For a finite set A \mathcal{A} A , P ( A ) \mathcal{P}(\mathcal{A}) P ( A ) denotes the set of probability distributions on A \mathcal{A} A . The binary entropy is denoted
h 2 ( u ) : = − u log u − ( 1 − u ) log ( 1 − u ) , u ∈ [ 0 , 1 ] . h_2(u) := -u \log u - (1-u)\log(1-u), \qquad u \in[0,1]. h 2 ( u ) := − u log u − ( 1 − u ) log ( 1 − u ) , u ∈ [ 0 , 1 ] .
For integers 1 ≤ k ≤ N 1 \leq k \leq N 1 ≤ k ≤ N , let ( N ) k : = N ( N − 1 ) ⋯ ( N − k + 1 ) (N)_k := N(N-1)\cdots(N-k+1) ( N ) k := N ( N − 1 ) ⋯ ( N − k + 1 ) denote the falling factorial.
Symmetric Memoryless Channel
At time i i i , user j j j selects X j [ i ] ∈ X X_j[i] \in\mathcal{X} X j [ i ] ∈ X and observes Y j [ i ] ∈ Y Y_j[i] \in\mathcal{Y} Y j [ i ] ∈ Y at the end of the channel use. The memoryless channel has the joint transition law
P Y [ J ] ∣ X [ J ] ( y [ J ] ∣ x [ J ] ) . (1) P_{Y_{[J]}|X_{[J]}}(y_{[J]} \mid x_{[J]}). \tag*{(1)} P Y [ J ] ∣ X [ J ] ( y [ J ] ∣ x [ J ] ) . (1)
Conditioned on the input chosen in each slot, successive output vectors are generated independently according to this law. When no confusion arises we abbreviate P Y [ J ] ∣ X [ J ] P_{Y_{[J]}|X_{[J]}} P Y [ J ] ∣ X [ J ] as P P P .
The users are anonymous, and the physical channel does not depend on how they are labeled.
Definition 1 (Symmetric channel). For a permutation σ \sigma σ of [ J ] [J] [ J ] , write x σ : = ( x σ ( 1 ) , … , x σ ( J ) ) x_\sigma:= (x_{\sigma(1)},\ldots,x_{\sigma(J)}) x σ := ( x σ ( 1 ) , … , x σ ( J ) ) , and define y σ y_\sigma y σ analogously. The memoryless channel P P P is symmetric (or permutation invariant ) if, for every permutation σ \sigma σ of [ J ] [J] [ J ] , every x [ J ] ∈ X J x_{[J]} \in\mathcal{X}^J x [ J ] ∈ X J and every y [ J ] ∈ Y J y_{[J]} \in\mathcal{Y}^J y [ J ] ∈ Y J ,
P ( y σ ∣ x σ ) = P ( y [ J ] ∣ x [ J ] ) . (2) P(y_\sigma\mid x_\sigma) = P(y_{[J]} \mid x_{[J]}). \tag*{(2)} P ( y σ ∣ x σ ) = P ( y [ J ] ∣ x [ J ] ) . (2)
Throughout, the channel is symmetric. Note that this is a symmetry under relabeling of the users, not the symmetry of a single-user channel in the sense of [37 ] : no symmetry is imposed on the input or the output alphabet. For j ∈ [ J ] j \in[J] j ∈ [ J ] , the channel observed by receiver j j j is the marginal
W j ( u ∣ x [ J ] ) : = ∑ v ∈ Y J − 1 P ( u ⊔ j v ∣ x [ J ] ) , u ∈ Y . (3) W_j(u \mid x_{[J]}) := \sum_{v \in\mathcal{Y}^{J-1}} P(u \sqcup_j v \mid x_{[J]}), \qquad u \in\mathcal{Y}. \tag*{(3)} W j ( u ∣ x [ J ] ) := v ∈ Y J − 1 ∑ P ( u ⊔ j v ∣ x [ J ] ) , u ∈ Y . (3)
obtained by summing over the outputs y − j = v y_{-j} = v y − j = v of the other users while the output y j = u y_j = u y j = u of user j j j is held fixed. The symmetry (2) implies that the marginals yield a single local law. For b ∈ X J − 1 b \in\mathcal{X}^{J-1} b ∈ X J − 1 and a permutation τ \tau τ of [ J − 1 ] [J-1] [ J − 1 ] , write b τ : = ( b τ ( 1 ) , … , b τ ( J − 1 ) ) . b_\tau:= (b_{\tau(1)},\ldots,b_{\tau(J-1)}). b τ := ( b τ ( 1 ) , … , b τ ( J − 1 ) ) . Define
W ( y ∣ a , b ) : = W 1 ( y ∣ a ⊔ 1 b ) , a ∈ X , b ∈ X J − 1 , y ∈ Y . (4) W(y \mid a,b) := W_1(y \mid a \sqcup_1 b), \qquad a \in\mathcal{X},\ b \in\mathcal{X}^{J-1},\ y \in\mathcal{Y}. \tag*{(4)} W ( y ∣ a , b ) := W 1 ( y ∣ a ⊔ 1 b ) , a ∈ X , b ∈ X J − 1 , y ∈ Y . (4)
The next lemma shows that every user sees the same local law W W W , which is invariant under permutations of the other users’ inputs.
Lemma 1 (Single local law). Under (2), for every j ∈ [ J ] j \in[J] j ∈ [ J ] , x [ J ] ∈ X J x_{[J]} \in\mathcal{X}^J x [ J ] ∈ X J , y ∈ Y y \in\mathcal{Y} y ∈ Y , and every permutation σ \sigma σ of [ J ] [J] [ J ] and τ \tau τ of [ J − 1 ] [J-1] [ J − 1 ] , (i) W j ( y ∣ x σ ) = W σ ( j ) ( y ∣ x [ J ] ) W_j(y \mid x_\sigma) = W_{\sigma(j)}(y \mid x_{[J]}) W j ( y ∣ x σ ) = W σ ( j ) ( y ∣ x [ J ] ) ; (ii) W j ( y ∣ x [ J ] ) = W ( y ∣ x j , x − j ) W_j(y \mid x_{[J]}) = W(y \mid x_j,x_{-j}) W j ( y ∣ x [ J ] ) = W ( y ∣ x j , x − j ) ; (iii) W ( y ∣ x j , ( x − j ) τ ) = W ( y ∣ x j , x − j ) W(y \mid x_j,(x_{-j})_\tau) = W(y \mid x_j,x_{-j}) W ( y ∣ x j , ( x − j ) τ ) = W ( y ∣ x j , x − j ) .
Proof. (i) Since y ↦ y σ y \mapsto y_\sigma y ↦ y σ is a bijection of Y J \mathcal{Y}^J Y J with ( y σ ) j = y σ ( j ) (y_\sigma)_j = y_{\sigma(j)} ( y σ ) j = y σ ( j ) , summing (2) over the y y y with y σ ( j ) = u y_{\sigma(j)} = u y σ ( j ) = u gives W j ( u ∣ x σ ) = W σ ( j ) ( u ∣ x [ J ] ) W_j(u \mid x_\sigma) = W_{\sigma(j)}(u \mid x_{[J]}) W j ( u ∣ x σ ) = W σ ( j ) ( u ∣ x [ J ] ) , by (3). (ii) follows from (i) and (4) for the permutation σ \sigma σ with σ ( 1 ) = j \sigma(1) = j σ ( 1 ) = j that keeps the order of the other indices, for which x σ = x j ⊔ x − j x_{\sigma}=x_j \mathbin{\sqcup} x_{-j} x σ = x j ⊔ x − j . (iii) follows from (i) for the permutation that fixes 1 and permutes the other positions by τ \tau τ .
Definition 2 (Anonymous symmetry breaking). Let Π n \Pi_n Π n denote the class of all admissible n n n -round anonymous symmetric causal strategies (defined below), where n n n is the number of channel uses. We call n n n the blocklength even though no block code is involved in general. The private seeds S 1 , … , S J S_1,\ldots,S_J S 1 , … , S J are i.i.d. with common law P S P_S P S and are independent of the common randomness C C C , whose law is P C P_C P C . The laws P C P_C P C and P S P_S P S are part of the strategy, and, given the inputs, the channel outputs are independent of ( C , S [ J ] ) (C,S_{[J]}) ( C , S [ J ] ) . Under our convention, all private randomization is absorbed into the private seeds, and a strategy π n ∈ Π n \pi_n \in\Pi_n π n ∈ Π n is specified by a sequence of common causal deterministic encoding maps { f i } i = 1 n \{f_i\}_{i=1}^n { f i } i = 1 n such that
X j [ i ] = f i ( C , S j , Y j i − 1 ) , i ∈ [ n ] , j ∈ [ J ] , (5) X_j[i] = f_i(C,S_j,Y_j^{i-1}), \qquad i \in[n], j \in[J], \tag*{(5)} X j [ i ] = f i ( C , S j , Y j i − 1 ) , i ∈ [ n ] , j ∈ [ J ] , (5) and a common terminal map g n g_n g n that produces the identifier Z j ∈ [ J ] Z_j \in[J] Z j ∈ [ J ] of user j j j ,
Z j = g n ( C , S j , Y j n ) ∈ [ J ] , j ∈ [ J ] . (6) Z_j = g_n(C,S_j,Y_j^n) \in[J], \qquad j \in[J]. \tag*{(6)} Z j = g n ( C , S j , Y j n ) ∈ [ J ] , j ∈ [ J ] . (6) We call a strategy nonadaptive if the encoding maps f i f_i f i do not depend on the past channel outputs Y j i − 1 Y_j^{i-1} Y j i − 1 , i.e., X j [ i ] = f i ( C , S j ) X_j[i] = f_i(C,S_j) X j [ i ] = f i ( C , S j ) for all i ∈ [ n ] i \in[n] i ∈ [ n ] and j ∈ [ J ] j \in[J] j ∈ [ J ] . A strategy is also called a protocol . We define the symmetry-breaking error event as
E : = { ∃ j ≠ k : Z j = Z k } , (7) \mathcal{E} := \{\exists j \ne k : Z_j = Z_k\}, \tag*{(7)} E := { ∃ j = k : Z j = Z k } , (7) and its probability under strategy π n \pi_n π n and channel P P P is denoted by
P b ( n ) : = P b ( n ) ( π n ; P ) . (8) P_{\mathrm{b}}^{(n)} := P_{\mathrm{b}}^{(n)}(\pi_n;P). \tag*{(8)} P b ( n ) := P b ( n ) ( π n ; P ) . (8) Symmetry breaking is reliable if there exists a sequence { π n } n ≥ 1 \{\pi_n\}_{n\ge1} { π n } n ≥ 1 of strategies with π n ∈ Π n \pi_n \in\Pi_n π n ∈ Π n such that P b ( n ) → 0 P_{\mathrm{b}}^{(n)} \to0 P b ( n ) → 0 .
Common randomness cannot lower the optimal error probability, since P b ( n ) P_{\mathrm{b}}^{(n)} P b ( n ) averages the conditional error probabilities over C C C ; we keep C C C because shared random codebooks are naturally described with it.
Collisions of i.i.d. identifiers: Any strategy whose terminal map ignores the channel outputs cannot be reliable, because under such a strategy the identifiers Z 1 , … , Z J Z_1,\ldots,Z_J Z 1 , … , Z J are i.i.d. given C C C . For integers 2 ≤ m ≤ N 2 \le m \le N 2 ≤ m ≤ N , let
γ m , N : = 1 − ( N ) m N m (9) \gamma_{m,N} := 1-\frac{(N)_m}{N^m} \tag*{(9)} γ m , N := 1 − N m ( N ) m (9)
be the probability that m m m i.i.d. uniform draws from [ N ] [N] [ N ] are not all distinct.
Lemma 2 (Collision among i.i.d. identifiers). Let 2 ≤ m ≤ J 2 \le m \le J 2 ≤ m ≤ J and let Z 1 , … , Z m Z_1,\ldots,Z_m Z 1 , … , Z m be i.i.d. random variables taking values in [ J ] [J] [ J ] . Then
P { ∃ j ≠ k : Z j = Z k } ≥ γ m , J . (10) \mathbb{P}\{\exists j \ne k : Z_j = Z_k\} \ge\gamma_{m,J}. \tag*{(10)} P { ∃ j = k : Z j = Z k } ≥ γ m , J . (10)
Proof. If Z 1 , … , Z m Z_1,\ldots,Z_m Z 1 , … , Z m are i.i.d. with distribution ( p z ) z ∈ [ J ] (p_z)_{z\in[J]} ( p z ) z ∈ [ J ] , then
P { Z 1 , … , Z m all distinct } = m ! e m ( p 1 , … , p J ) ≤ ( J ) m J m , (11) \mathbb{P}\{Z_1,\ldots,Z_m\text{ all distinct}\}=m!e_m(p_1,\ldots,p_J)\le\frac{(J)_m}{J^m}, \tag*{(11)} P { Z 1 , … , Z m all distinct } = m ! e m ( p 1 , … , p J ) ≤ J m ( J ) m , (11)
where e m ( p 1 , … , p J ) : = ∑ i 1 < ⋯ < i m p i 1 ⋯ p i m e_m(p_1,\ldots,p_J):=\sum_{i_1<\cdots<i_m}p_{i_1}\cdots p_{i_m} e m ( p 1 , … , p J ) := ∑ i 1 < ⋯ < i m p i 1 ⋯ p i m is the m m m -th elementary symmetric polynomial and the inequality is Maclaurin’s inequality [38 Th. 52] . The uniform law attains equality. Taking the complement proves the claim.
Thus γ m , J \gamma_{m,J} γ m , J is the smallest collision probability among m m m i.i.d. draws from [ J ] [J] [ J ] . In particular,
γ J : = γ J , J = 1 − J ! J J , γ 2 , J = 1 J . (12) \gamma_J:=\gamma_{J,J}=1-\frac{J!}{J^J}, \qquad\gamma_{2,J}=\frac{1}{J}. \tag*{(12)} γ J := γ J , J = 1 − J J J ! , γ 2 , J = J 1 . (12)
If the identifiers Z 1 , … , Z J Z_1,\ldots,Z_J Z 1 , … , Z J were i.i.d. given the common randomness, for instance drawn from the private seeds without using the channel, then P b ( n ) ≥ γ J P_{\mathrm{b}}^{(n)}\ge\gamma_J P b ( n ) ≥ γ J for every n n n . Reliable symmetry breaking therefore requires the channel to make the users’ observations differ. The converse of Section III bounds how fast the error probability can decay, and the collision term P c o l ( M , J ) = γ J , M P_{\mathrm{col}}(M,J)=\gamma_{J,M} P col ( M , J ) = γ J , M of the ranking scheme below, (17), is another instance of (9).
Private randomness is essential for the coding construction below: users independently draw tentative names. The codebook is common but can be fixed deterministically after the random-coding analysis.
Random Names, Multicast Decoding, and Ranking
We propose a nonadaptive scheme that uses a common codebook and private randomization to generate tentative names. It is one particular strategy in Π n \Pi_n Π n ; many other strategies exist, including the interactive ones of § 5 , and the converse of Section III applies to all of them. Fix a codebook size M M M and a common codebook
C : = ( x n ( m ) ) m ∈ [ M ] ∈ ( X n ) M . (13) \mathcal{C} := \left(x^n(m)\right)_{m\in[M]} \in\left(\mathcal{X}^n\right)^M. \tag*{(13)} C := ( x n ( m ) ) m ∈ [ M ] ∈ ( X n ) M . (13)
Each user independently draws
S j = M j ∼ Unif [ M ] (14) S_j = M_j \sim\operatorname{Unif}[M] \tag*{(14)} S j = M j ∼ Unif [ M ] (14)
and transmits X j n = x n ( M j ) X_j^n=x^n(M_j) X j n = x n ( M j ) . We call M j M_j M j the name of user j j j . It plays the role of the message in the multicast code, and R : = 1 n log M R := \frac{1}{n}\log M R := n 1 log M is the name rate . User j j j knows ( M j , X j n , Y j n ) (M_j,X_j^n,Y_j^n) ( M j , X j n , Y j n ) and decodes an unordered list (a multiset) M ^ − j \widehat{\mathcal{M}}_{-j} M − j of J − 1 J-1 J − 1 foreign names. The unordered formulation is important: anonymous users need not attach physical user labels to the decoded codewords.
User j j j forms the multiset { M j } ∪ M ^ − j \{M_j\}\cup\widehat{\mathcal{M}}_{-j} { M j } ∪ M − j and chooses its rank, the number of decoded names below its own plus one,
Z j : = 1 + ∑ m ∈ M ^ − j 1 { m < M j } ∈ [ J ] . (15) Z_j := 1+\sum_{m\in\widehat{\mathcal{M}}_{-j}}\boldsymbol{1}\{m<M_j\}\in[J]. \tag*{(15)} Z j := 1 + m ∈ M − j ∑ 1 { m < M j } ∈ [ J ] . (15)
If all names are distinct and every decoder is correct, the J J J ranks Z 1 , … , Z J Z_1,\ldots,Z_J Z 1 , … , Z J are a permutation of [ J ] [J] [ J ] , so the symmetry-breaking error event E \mathcal{E} E does not occur. Define the decoding error probability given that the names are distinct,
P e , j ( n ) : = P { M ^ − j ≠ { M k : k ≠ j } ∣ M 1 , … , M J distinct } , (16) P_{e,j}^{(n)} := \mathbb{P}\{\widehat{\mathcal{M}}_{-j}\ne\{M_k:k\ne j\}\mid M_1,\ldots,M_J\text{ distinct}\}, \tag*{(16)} P e , j ( n ) := P { M − j = { M k : k = j } ∣ M 1 , … , M J distinct } , (16)
and the collision probability, i.e., the probability that the names are not distinct,
P c o l ( M , J ) : = 1 − ( M ) J M J = γ J , M , M ≥ J . (17) P_{\mathrm{col}}(M,J):=1-\frac{(M)_J}{M^J}=\gamma_{J,M},\qquad M\ge J. \tag*{(17)} P col ( M , J ) := 1 − M J ( M ) J = γ J , M , M ≥ J . (17)
By the union bound over the J J J receivers, the ranking scheme (15) gives
P b ( n ) ≤ P c o l ( M , J ) + ∑ j = 1 J P e , j ( n ) . (18) P_{\mathrm{b}}^{(n)}\le P_{\mathrm{col}}(M,J)+\sum_{j=1}^{J}P_{e,j}^{(n)}. \tag*{(18)} P b ( n ) ≤ P col ( M , J ) + j = 1 ∑ J P e , j ( n ) . (18)
By the union bound over the ( J 2 ) \binom{J}{2} ( 2 J ) pairs of names, P c o l ( M , J ) ≤ ( J 2 ) 1 M P_{\mathrm{col}}(M,J)\le\binom{J}{2}\frac{1}{M} P col ( M , J ) ≤ ( 2 J ) M 1 , and P c o l ( M , J ) = ( J 2 ) 1 M + O ( M − 2 ) P_{\mathrm{col}}(M,J)=\binom{J}{2}\frac{1}{M}+O(M^{-2}) P col ( M , J ) = ( 2 J ) M 1 + O ( M − 2 ) as M → ∞ M\to\infty M → ∞ . For fixed J J J and R R R , and M = ⌈ e n R ⌉ M=\lceil e^{nR}\rceil M = ⌈ e n R ⌉ , the tentative-name collision term P c o l ( M , J ) P_{\mathrm{col}}(M,J) P col ( M , J ) has exponent R R R . Section IV bounds the decoding error probability P e , j ( n ) P_{e,j}^{(n)} P e , j ( n ) in (18) when the common codebook is drawn from the i.i.d. random-coding ensemble: independently for every name m m m , generate
P C { X n ( m ) = x n } = ∏ i = 1 n Q X ( x [ i ] ) . (19) \mathbb{P}_{\mathcal{C}}\{X^n(m)=x^n\}=\prod_{i=1}^{n}Q_X(x[i]). \tag*{(19)} P C { X n ( m ) = x n } = i = 1 ∏ n Q X ( x [ i ]) . (19)
Conditioned on distinct names, the J J J transmitted codewords are independent draws from Q X n Q_X^n Q X n , exactly as required by the random-coding analysis.
Change-of-Measure Converses
In this section, we derive converses, i.e., lower bounds on the error probability of every anonymous symmetric causal strategy in Π n \Pi_n Π n , independently of any particular coding architecture. To do so, we compare the physical experiment with references under which the identifiers of m m m users are conditionally i.i.d. given a latent variable, using a change-of-measure argument. For a channel P P P , define the optimal error probability and the upper reliability exponent, respectively,
ϵ n ∗ ( P ) : = inf π n ∈ Π n P b ( n ) ( π n ; P ) , (20) \epsilon_n^*(P):=\inf_{\pi_n\in\Pi_n}P_{\mathrm{b}}^{(n)}(\pi_n;P), \tag*{(20)} ϵ n ∗ ( P ) := π n ∈ Π n inf P b ( n ) ( π n ; P ) , (20)
E ‾ b ( P ) : = lim sup n → ∞ − 1 n log ϵ n ∗ ( P ) . (21) \overline{E}_{\mathrm{b}}(P):=\limsup_{n\to\infty}-\frac{1}{n}\log\epsilon_n^*(P). \tag*{(21)} E b ( P ) := n → ∞ lim sup − n 1 log ϵ n ∗ ( P ) . (21)
The lower reliability exponent E b ( P ) E_{\mathrm{b}}(P) E b ( P ) is defined with lim inf \liminf lim inf in place of lim sup \limsup lim sup ; converses bound E ‾ b ( P ) \overline{E}_{\mathrm{b}}(P) E b ( P ) from above, and schemes bound E b ( P ) E_{\mathrm{b}}(P) E b ( P ) from below. Unlike the cut-set bounds of Section IV, which constrain only name decoding, these converses hold for every strategy.
Law of the Complete Local Views
By definition, given a protocol π n ∈ Π n \pi_n \in\Pi_n π n ∈ Π n and a realization C = c C=c C = c of the common randomness, the identifier chosen by user j j j , Z j = g n ( c , S j , Y j n ) Z_j=g_n(c,S_j,Y_j^n) Z j = g n ( c , S j , Y j n ) , is a function of
H j : = ( S j , Y j n ) , (22) H_j := (S_j,Y_j^n), \tag*{(22)} H j := ( S j , Y j n ) , (22)
which we call the complete local view of user j j j , and our first step is to derive its distribution. The private seeds S 1 , … , S J S_1,\ldots,S_J S 1 , … , S J are i.i.d. with law P S P_S P S , hence the distribution of the complete local views of all users H [ J ] H_{[J]} H [ J ] is
P π n , c π n , c ( d s [ J ] , d y [ J ] n ) : = ∏ j = 1 J P S ( d s j ) ∏ i = 1 n P ( y [ J ] [ i ] ∣ x [ J ] [ i ] ) , x j [ i ] = f i ( c , s j , y j i − 1 ) , (23) P_{\pi_n,c}^{\pi_n,c}(ds_{[J]},dy_{[J]}^n) := \prod_{j=1}^{J} P_S(ds_j)\prod_{i=1}^{n}P(y_{[J]}[i]\mid x_{[J]}[i]), \qquad x_j[i]=f_i(c,s_j,y_j^{i-1}), \tag*{(23)} P π n , c π n , c ( d s [ J ] , d y [ J ] n ) := j = 1 ∏ J P S ( d s j ) i = 1 ∏ n P ( y [ J ] [ i ] ∣ x [ J ] [ i ]) , x j [ i ] = f i ( c , s j , y j i − 1 ) , (23)
and this distribution depends on the protocol π n \pi_n π n through f i f_i f i . We also write P π n , c : = R π n , c π n , c P^{\pi_n,c}:=R_{\pi_n,c}^{\pi_n,c} P π n , c := R π n , c π n , c .
Symmetric References
Consider 2 ≤ m ≤ J 2\le m\le J 2 ≤ m ≤ J , and an alternative distribution Q Q Q for local views H [ J ] H_{[J]} H [ J ] where the local views of the first m m m users H [ m ] H_{[m]} H [ m ] are i.i.d. Under this alternative distribution, the identifiers Z [ m ] Z_{[m]} Z [ m ] are i.i.d. as well, so that they collide with probability at least γ m , J \gamma_{m,J} γ m , J , as shown in Lemma 2. By linearity, this is also true when considering a mixture of such Q Q Q , which leads to the following definition and lemma.
Definition 3 (Symmetric reference). A probability law Q Q Q on the complete local views H [ J ] H_{[J]} H [ J ] is an m m m -symmetric reference for ( π n , c ) (\pi_n,c) ( π n , c ) if Q ≪ P π n , c Q\ll P^{\pi_n,c} Q ≪ P π n , c and there exists a latent random variable Λ \Lambda Λ such that, under Q Q Q , the views H 1 , … , H m H_1,\ldots,H_m H 1 , … , H m are conditionally i.i.d. given Λ \Lambda Λ ; formally, Q Q Q is the marginal on H [ J ] H_{[J]} H [ J ] of a law of ( Λ , H [ J ] ) (\Lambda,H_{[J]}) ( Λ , H [ J ] ) under which H 1 , … , H m H_1,\ldots,H_m H 1 , … , H m are conditionally i.i.d. given Λ \Lambda Λ . The set of such laws is denoted by L m , J ( n ) ( π n , c ; P ) \mathcal{L}_{m,J}^{(n)}(\pi_n,c;P) L m , J ( n ) ( π n , c ; P ) .
Note that exchangeability of H 1 , … , H m H_1,\ldots,H_m H 1 , … , H m would not suffice: on the channel of Section IV-C whose output pair is uniform on { ( 0 , 1 ) , ( 1 , 0 ) } \{(0,1),(1,0)\} {( 0 , 1 ) , ( 1 , 0 )} , the two views are exchangeable, and yet the identifiers 1 + Y j [ 1 ] 1+Y_j[1] 1 + Y j [ 1 ] never collide.
Lemma 3 (Failure under a symmetric reference). Every Q ∈ L m , J ( n ) ( π n , c ; P ) Q\in\mathcal{L}_{m,J}^{(n)}(\pi_n,c;P) Q ∈ L m , J ( n ) ( π n , c ; P ) satisfies
Q ( E ) ≥ γ m , J , E : = { ∃ j ≠ k : Z j = Z k } . (24) Q(\mathcal{E})\ge\gamma_{m,J}, \qquad\mathcal{E}:=\{\exists j\ne k: Z_j=Z_k\}. \tag*{(24)} Q ( E ) ≥ γ m , J , E := { ∃ j = k : Z j = Z k } . (24)
Proof. Conditionally on Λ \Lambda Λ , the identifiers Z j = g n ( c , S j , Y j n ) Z_j=g_n(c,S_j,Y_j^n) Z j = g n ( c , S j , Y j n ) , j ∈ [ m ] j\in[m] j ∈ [ m ] , are i.i.d., since all users apply the same map g n g_n g n . By Lemma 2 they collide with conditional probability at least γ m , J \gamma_{m,J} γ m , J . A collision among the selected users is a failure; averaging over Λ \Lambda Λ proves the claim.
Rényi Divergences
Before performing a change of measure, we first recall a few notions about Rényi divergences, which are critical for our analysis. For probability laws Q ≪ P Q\ll P Q ≪ P with likelihood ratio L = d Q / d P L=dQ/dP L = d Q / d P and for α > 1 \alpha>1 α > 1 , the Rényi divergence is
D α ( Q ∥ P ) : = 1 α − 1 log E P [ L α ] , (25) D_\alpha(Q\|P):=\frac{1}{\alpha-1}\log\mathbb{E}_P[L^\alpha], \tag*{(25)} D α ( Q ∥ P ) := α − 1 1 log E P [ L α ] , (25)
and D α ( Q ∥ P ) : = + ∞ D_\alpha(Q\|P):=+\infty D α ( Q ∥ P ) := + ∞ if Q ≪̸ P Q\not\ll P Q ≪ P . On a finite set, E P [ L α ] = ∑ ω Q ( ω ) α P ( ω ) 1 − α \mathbb{E}_P[L^\alpha]=\sum_\omega Q(\omega)^\alpha P(\omega)^{1-\alpha} E P [ L α ] = ∑ ω Q ( ω ) α P ( ω ) 1 − α . Moreover,
lim α ↓ 1 D α ( Q ∥ P ) = D K L ( Q ∥ P ) , (26) \lim_{\alpha\downarrow1}D_\alpha(Q\|P)=D_{\mathrm{KL}}(Q\|P), \tag*{(26)} α ↓ 1 lim D α ( Q ∥ P ) = D KL ( Q ∥ P ) , (26)
D ∞ ( Q ∥ P ) : = log ess sup P L ; (27) D_\infty(Q\|P):=\log\operatorname*{ess\,sup}_{P} L; \tag*{(27)} D ∞ ( Q ∥ P ) := log P ess sup L ; (27)
see [39 Ths. 5 and 6] ; (26) holds since D α ( Q ∥ P ) < ∞ D_\alpha(Q\|P)<\infty D α ( Q ∥ P ) < ∞ on a finite set with Q ≪ P Q\ll P Q ≪ P . Set b α : = α / ( α − 1 ) b_\alpha:=\alpha/(\alpha-1) b α := α / ( α − 1 ) for 1 < α < ∞ 1<\alpha<\infty 1 < α < ∞ and b ∞ : = 1 b_\infty:=1 b ∞ := 1 .
Lemma 4 (Rényi change of measure). For every event A \mathcal{A} A and every α ∈ ( 1 , ∞ ] \alpha\in(1,\infty] α ∈ ( 1 , ∞ ] ,
P ( A ) ≥ Q ( A ) b α exp { − D α ( Q ∥ P ) } . (28) P(\mathcal{A})\ge Q(\mathcal{A})^{b_\alpha}\exp\{-D_\alpha(Q\|P)\}. \tag*{(28)} P ( A ) ≥ Q ( A ) b α exp { − D α ( Q ∥ P )} . (28)
Proof. For α < ∞ \alpha<\infty α < ∞ , Hölder’s inequality gives
Q ( A ) = E P [ L 1 A ] ≤ ( E P [ L α ] ) 1 / α P ( A ) ( α − 1 ) / α = e α − 1 α D α ( Q ∥ P ) P ( A ) ( α − 1 ) / α , (29) Q(\mathcal{A})=\mathbb{E}_P[L\mathbf{1}_{\mathcal{A}}]\le\bigl(\mathbb{E}_P[L^\alpha]\bigr)^{1/\alpha}P(\mathcal{A})^{(\alpha-1)/\alpha}=e^{\frac{\alpha-1}{\alpha}D_\alpha(Q\|P)}P(\mathcal{A})^{(\alpha-1)/\alpha}, \tag*{(29)} Q ( A ) = E P [ L 1 A ] ≤ ( E P [ L α ] ) 1/ α P ( A ) ( α − 1 ) / α = e α α − 1 D α ( Q ∥ P ) P ( A ) ( α − 1 ) / α , (29)
and rearranging proves the claim. For α = ∞ \alpha=\infty α = ∞ , Q ( A ) = ∫ A L d P ≤ e D ∞ ( Q ∥ P ) P ( A ) Q(\mathcal{A})=\int_{\mathcal{A}}L\,dP\le e^{D_\infty(Q\|P)}P(\mathcal{A}) Q ( A ) = ∫ A L d P ≤ e D ∞ ( Q ∥ P ) P ( A ) . If Q ≪̸ P Q\not\ll P Q ≪ P the bound is trivial. The bound also follows from the data-processing inequality for Rényi divergences, as in the Arimoto converse of [30 ] .
Change of Measure
Equipped with the above, we then use a change-of-measure argument between the distribution P π n , c P^{\pi_n,c} P π n , c and the closest symmetric reference distribution Q ∈ L m , J ( n ) ( π n , c ; P ) Q \in\mathcal{L}^{(n)}_{m,J}(\pi_n,c;P) Q ∈ L m , J ( n ) ( π n , c ; P ) to lower bound the error probability, where the closeness between P π n , c P^{\pi_n,c} P π n , c and Q Q Q is quantified by the Rényi divergence. Namely, combining Lemma 3 with Lemma 4 for A = E \mathcal{A}=\mathcal{E} A = E gives the basic converse of Proposition 1.
Proposition 1 (Trajectory converse). For every protocol π n \pi_n π n , every realization c c c of the common randomness, every 2 ≤ m ≤ J 2 \le m \le J 2 ≤ m ≤ J and every α ∈ ( 1 , ∞ ] \alpha\in(1,\infty] α ∈ ( 1 , ∞ ] ,
P b ( n ) ( π n ; P ∣ C = c ) ≥ γ m , J b α e − D α , m ( n ) ( π n , c ; P ) , (30) P^{(n)}_{\mathrm{b}}(\pi_n;P\mid C=c)\ge\gamma^{b_\alpha}_{m,J}e^{-D^{(n)}_{\alpha,m}(\pi_n,c;P)}, \tag*{(30)} P b ( n ) ( π n ; P ∣ C = c ) ≥ γ m , J b α e − D α , m ( n ) ( π n , c ; P ) , (30)
where the trajectory radius is
D α , m ( n ) ( π n , c ; P ) : = inf Q ∈ L m , J ( n ) ( π n , c ; P ) D α ( Q ∥ P π n , c ) . (31) D^{(n)}_{\alpha,m}(\pi_n,c;P):=\inf_{Q\in\mathcal{L}^{(n)}_{m,J}(\pi_n,c;P)}D_\alpha(Q\|P^{\pi_n,c}). \tag*{(31)} D α , m ( n ) ( π n , c ; P ) := Q ∈ L m , J ( n ) ( π n , c ; P ) inf D α ( Q ∥ P π n , c ) . (31)
The trajectory radius D α , m ( n ) ( π n , c ; P ) D^{(n)}_{\alpha,m}(\pi_n,c;P) D α , m ( n ) ( π n , c ; P ) quantifies the closeness between the law P π n , c P^{\pi_n,c} P π n , c and the closest symmetric reference Q ∈ L m , J ( n ) ( π n , c ; P ) Q\in\mathcal{L}^{(n)}_{m,J}(\pi_n,c;P) Q ∈ L m , J ( n ) ( π n , c ; P ) , and it depends on the blocklength n n n , the channel P P P and the protocol π n \pi_n π n . For a fixed protocol π n \pi_n π n , computing D α , m ( n ) ( π n , c ; P ) D^{(n)}_{\alpha,m}(\pi_n,c;P) D α , m ( n ) ( π n , c ; P ) does not seem feasible, mostly because of the complex structure of L m , J ( n ) ( π n , c ; P ) \mathcal{L}^{(n)}_{m,J}(\pi_n,c;P) L m , J ( n ) ( π n , c ; P ) . Even worse, a converse that holds for all protocols would require maximizing D α , m ( n ) ( π n , c ; P ) D^{(n)}_{\alpha,m}(\pi_n,c;P) D α , m ( n ) ( π n , c ; P ) over the protocols, which seems even more involved. We now provide a tractable bound that involves only single-letter quantities.
To obtain a tractable converse, we consider a subset of the set of symmetric references L m , J ( n ) ( π n , c ; P ) \mathcal{L}^{(n)}_{m,J}(\pi_n,c;P) L m , J ( n ) ( π n , c ; P ) which have a special product form, which allows us to derive a single-letter bound. Consider distributions Q B π n , c Q^{\pi_n,c}_B Q B π n , c of the form
Q B π n , c : = R B π n , c / β B π n , c , (32) Q^{\pi_n,c}_B:=R^{\pi_n,c}_B/\beta^{\pi_n,c}_B, \tag*{(32)} Q B π n , c := R B π n , c / β B π n , c , (32)
where
R B π n , c ( d s [ J ] , d y [ J ] n ) : = ∏ j = 1 J P S ( d s j ) ∏ i = 1 n B ( y [ J ] [ i ] ∣ x [ J ] [ i ] ) , x j [ i ] = f i ( c , s j , y j i − 1 ) , (33) R^{\pi_n,c}_B(\mathrm{d}s_{[J]},\mathrm{d}y^n_{[J]}):=\prod_{j=1}^{J}P_S(\mathrm{d}s_j)\prod_{i=1}^{n}B(y_{[J]}[i]\mid x_{[J]}[i]),\qquad x_j[i]=f_i(c,s_j,y_j^{i-1}), \tag*{(33)} R B π n , c ( d s [ J ] , d y [ J ] n ) := j = 1 ∏ J P S ( d s j ) i = 1 ∏ n B ( y [ J ] [ i ] ∣ x [ J ] [ i ]) , x j [ i ] = f i ( c , s j , y j i − 1 ) , (33)
and β B π n , c \beta^{\pi_n,c}_B β B π n , c is the total mass of R B π n , c R^{\pi_n,c}_B R B π n , c , which depends on the kernel1 B B B , of the form
B ( y [ J ] ∣ x [ J ] ) = ∑ r ∈ R [ ∏ j = 1 m a r ( x j , y j ) ] c r ( y m + 1 : J ∣ x [ J ] , y [ m ] ) , (34) B(y_{[J]}\mid x_{[J]})=\sum_{r\in\mathcal{R}}\left[\prod_{j=1}^{m}a_r(x_j,y_j)\right]c_r(y_{m+1:J}\mid x_{[J]},y_{[m]}), \tag*{(34)} B ( y [ J ] ∣ x [ J ] ) = r ∈ R ∑ [ j = 1 ∏ m a r ( x j , y j ) ] c r ( y m + 1 : J ∣ x [ J ] , y [ m ] ) , (34)
with R \mathcal{R} R a finite set of component indices and all factors a r , c r a_r,c_r a r , c r chosen nonnegative, and with B ≪ P B\ll P B ≪ P , which means that B ( y ∣ x ) = 0 B(y\mid x)=0 B ( y ∣ x ) = 0 whenever P ( y ∣ x ) = 0 P(y\mid x)=0 P ( y ∣ x ) = 0 . The same one-user factor a r a_r a r is used by every selected user. For m = J m=J m = J the last factor is absent. We say that a kernel B B B is admissible if the corresponding a r a_r a r and c r c_r c r satisfy the following.
Definition 4 (Admissible kernels). A kernel of the form (34) is admissible if one of the following holds.
(I) Independent completion: every c r ( y m + 1 : J ∣ x [ J ] , y [ m ] ) c_r(y_{m+1:J}\mid x_{[J]},y_{[m]}) c r ( y m + 1 : J ∣ x [ J ] , y [ m ] ) does not depend on the inputs and outputs ( x [ m ] , y [ m ] ) (x_{[m]},y_{[m]}) ( x [ m ] , y [ m ] ) of the selected users; we then write it c r ( x m + 1 : J , y m + 1 : J ) c_r(x_{m+1:J},y_{m+1:J}) c r ( x m + 1 : J , y m + 1 : J ) .
(N) Normalized completion: for every r r r , the sum ∑ y m + 1 : J c r ( y m + 1 : J ∣ x [ J ] , y [ m ] ) \sum_{y_{m+1:J}}c_r(y_{m+1:J}\mid x_{[J]},y_{[m]}) ∑ y m + 1 : J c r ( y m + 1 : J ∣ x [ J ] , y [ m ] ) is a constant, independent of x [ J ] x_{[J]} x [ J ] and y [ m ] y_{[m]} y [ m ] . Since the constant can be absorbed into a r a_r a r , we normalize it to 1 without loss of generality.
The set of admissible kernels is denoted by A m , J ( P ) \mathcal{A}_{m,J}(P) A m , J ( P ) . For m = J m=J m = J both conditions are void.
Expanding the product in (33) writes R B π n , c R^{\pi_n,c}_B R B π n , c as a sum over component sequences r n r^n r n . For a fixed r n r^n r n , the selected users run the protocol with the same one-user weights a r i a_{r_i} a r i , so they are i.i.d. unless the factor c r i c_{r_i} c r i couples them. Under (I), c r c_r c r ignores the selected users; under (N), it may depend on them but has total mass one, so that summing out the unselected outputs, from the last use backward, removes it without favoring any selected trajectory. The following lemma makes this precise.
Lemma 5 (Product-form references). Let B ∈ A m , J ( P ) B\in\mathcal{A}_{m,J}(P) B ∈ A m , J ( P ) . For every protocol π n \pi_n π n and every c c c such that β B π n , c > 0 \beta^{\pi_n,c}_B>0 β B π n , c > 0 : Q B π n , c ∈ L m , J ( n ) ( π n , c ; P ) Q^{\pi_n,c}_B\in\mathcal{L}^{(n)}_{m,J}(\pi_n,c;P) Q B π n , c ∈ L m , J ( n ) ( π n , c ; P ) .
Proof sketch. The full proof is in Appendix A. Fix a component sequence r n r^n r n and sum out the unselected users. Under (I) they factor off. Under (N), each factor c r i c_{r_i} c r i sums to one, working backward from i = n i=n i = n . What remains is a product of m m m identical one-user measures.
Definition 5 (Normalized kernels). A kernel of the form (34) is normalized if ∑ y [ J ] B ( y [ J ] ∣ x [ J ] ) = 1 \sum_{y_{[J]}} B(y_{[J]} \mid x_{[J]})=1 ∑ y [ J ] B ( y [ J ] ∣ x [ J ] ) = 1 for all x [ J ] x_{[J]} x [ J ] .
Two remarks are in order: i) We do not assume that B B B is normalized, which allows for a better converse as we shall see. In turn, when B B B is not normalized, Q B π n , c Q_{B}^{\pi_n,c} Q B π n , c is in general not the law of the protocol run over a channel, so that computing β B π n , c \beta_{B}^{\pi_n,c} β B π n , c in closed form seems difficult. We will show how to obtain bounds on β B π n , c \beta_{B}^{\pi_n,c} β B π n , c . ii) In the special case where B B B is normalized, one may readily check that β B π n , c = 1 \beta_{B}^{\pi_n,c}=1 β B π n , c = 1 , and B B B can be interpreted as an alternative channel to the actual channel P P P ; furthermore, the Rényi divergence quantifies the difficulty of distinguishing P P P from B B B based on the local views H [ J ] H_{[J]} H [ J ] .
Rényi Divergence of Product-Form References
We now show how to bound the Rényi divergence between Q B π n , c Q_{B}^{\pi_n,c} Q B π n , c and P π n , c P^{\pi_n,c} P π n , c by exploiting the product form, to evaluate the converse. We first define the following quantities. In the remainder of this section, to simplify notation, we write u : = x m + 1 : J u:=x_{m+1:J} u := x m + 1 : J and v : = y m + 1 : J v:=y_{m+1:J} v := y m + 1 : J for the inputs and outputs of the unselected users. The likelihood ratio between Q B π n , c Q_{B}^{\pi_n,c} Q B π n , c and P π n , c P^{\pi_n,c} P π n , c is given by the product
d Q B π n , c d P π n , c = 1 β B π n , c ∏ i = 1 n B ( Y [ J ] [ i ] ∣ X [ J ] [ i ] ) P ( Y [ J ] [ i ] ∣ X [ J ] [ i ] ) . (35) \frac{\mathrm{d}Q_{B}^{\pi_n,c}}{\mathrm{d}P^{\pi_n,c}}=\frac{1}{\beta_{B}^{\pi_n,c}}\prod_{i=1}^{n}\frac{B(Y_{[J]}[i]\mid X_{[J]}[i])}{P(Y_{[J]}[i]\mid X_{[J]}[i])}. \tag*{(35)} d P π n , c d Q B π n , c = β B π n , c 1 i = 1 ∏ n P ( Y [ J ] [ i ] ∣ X [ J ] [ i ]) B ( Y [ J ] [ i ] ∣ X [ J ] [ i ]) . (35)
a) Persistence coefficient: Define A r ( x ) : = ∑ y a r ( x , y ) A_r(x):=\sum_y a_r(x,y) A r ( x ) := ∑ y a r ( x , y ) , and C r ( u ) : = ∑ v c r ( u , v ) C_r(u):=\sum_v c_r(u,v) C r ( u ) := ∑ v c r ( u , v ) under (I), or C r ≡ 1 C_r\equiv1 C r ≡ 1 under (N). The one-step mass of B B B is then
b B ( x [ J ] ) : = ∑ y [ J ] B ( y [ J ] ∣ x [ J ] ) = ∑ r [ ∏ j ≤ m A r ( x j ) ] C r ( u ) . (36) b_B(x_{[J]}):=\sum_{y_{[J]}}B(y_{[J]}\mid x_{[J]})=\sum_r\left[\prod_{j\le m}A_r(x_j)\right]C_r(u). \tag*{(36)} b B ( x [ J ] ) := y [ J ] ∑ B ( y [ J ] ∣ x [ J ] ) = r ∑ [ j ≤ m ∏ A r ( x j ) ] C r ( u ) . (36)
In turn, define the minimal expected one-step mass when the selected users draw their inputs i.i.d. and the unselected users play arbitrary inputs, namely
κ m ( B ) : = min p ∈ P ( X ) u ∈ X J − m ∑ x [ m ] [ ∏ j = 1 m p ( x j ) ] b B ( x [ m ] , u ) = min p , u ∑ r [ ∑ x p ( x ) A r ( x ) ] m C r ( u ) , (37) \kappa_m(B):=\min_{\substack{p\in\mathcal{P}(\mathcal{X})\\u\in\mathcal{X}^{J-m}}}\sum_{x_{[m]}}\left[\prod_{j=1}^{m}p(x_j)\right]b_B(x_{[m]},u)=\min_{p,u}\sum_r\left[\sum_xp(x)A_r(x)\right]^m C_r(u), \tag*{(37)} κ m ( B ) := p ∈ P ( X ) u ∈ X J − m min x [ m ] ∑ [ j = 1 ∏ m p ( x j ) ] b B ( x [ m ] , u ) = p , u min r ∑ [ x ∑ p ( x ) A r ( x ) ] m C r ( u ) , (37)
omitting u u u when m = J m=J m = J . It is noted that κ m ( B ) \kappa_m(B) κ m ( B ) only depends on single-letter quantities. We will use it to lower bound the normalization constant β B π n , c \beta_{B}^{\pi_n,c} β B π n , c . Further, κ m ( B ) \kappa_m(B) κ m ( B ) is readily computable by convex minimization over the simplex, since p ↦ ∑ r [ ∑ x p ( x ) A r ( x ) ] m C r ( u ) p\mapsto\sum_r\left[\sum_xp(x)A_r(x)\right]^m C_r(u) p ↦ ∑ r [ ∑ x p ( x ) A r ( x ) ] m C r ( u ) is convex for any fixed u u u , and u u u ranges over a finite set.
b) Likelihood-ratio moment and reference radius: For 1 < α < ∞ 1<\alpha<\infty 1 < α < ∞ and κ m ( B ) > 0 \kappa_m(B)>0 κ m ( B ) > 0 , define the worst-case likelihood-ratio moment and the one-step reference radius
Ψ α , m ( B ; P ) : = max x ∈ X J ∑ y : P ( y ∣ x ) > 0 B ( y ∣ x ) α P ( y ∣ x ) 1 − α , (38) \Psi_{\alpha,m}(B;P):=\max_{x\in\mathcal{X}^{J}}\sum_{y:P(y\mid x)>0}B(y\mid x)^\alpha P(y\mid x)^{1-\alpha}, \tag*{(38)} Ψ α , m ( B ; P ) := x ∈ X J max y : P ( y ∣ x ) > 0 ∑ B ( y ∣ x ) α P ( y ∣ x ) 1 − α , (38)
d α , m r e f ( B ∥ P ) : = log Ψ α , m ( B ; P ) − α log κ m ( B ) α − 1 , (39) d_{\alpha,m}^{\mathrm{ref}}(B\mathbin{\|}P):=\frac{\log\Psi_{\alpha,m}(B;P)-\alpha\log\kappa_m(B)}{\alpha-1}, \tag*{(39)} d α , m ref ( B ∥ P ) := α − 1 log Ψ α , m ( B ; P ) − α log κ m ( B ) , (39)
and the effective persistence
K α , m ( B ; P ) : = e − d α , m r e f ( B ∥ P ) = [ κ m ( B ) α Ψ α , m ( B ; P ) ] 1 / ( α − 1 ) . (40) K_{\alpha,m}(B;P):=e^{-d_{\alpha,m}^{\mathrm{ref}}(B\mathbin{\|}P)}=\left[\frac{\kappa_m(B)^\alpha}{\Psi_{\alpha,m}(B;P)}\right]^{1/(\alpha-1)}. \tag*{(40)} K α , m ( B ; P ) := e − d α , m ref ( B ∥ P ) = [ Ψ α , m ( B ; P ) κ m ( B ) α ] 1/ ( α − 1 ) . (40)
At order infinity put
ρ ∞ , m ( B ; P ) : = max x , y : B ( y ∣ x ) > 0 B ( y ∣ x ) P ( y ∣ x ) , \rho_{\infty,m}(B;P):=\max_{\substack{x,y:B(y\mid x)>0}}\frac{B(y\mid x)}{P(y\mid x)}, ρ ∞ , m ( B ; P ) := x , y : B ( y ∣ x ) > 0 max P ( y ∣ x ) B ( y ∣ x ) ,
d ∞ , m r e f ( B ∥ P ) : = log ρ ∞ , m ( B ; P ) κ m ( B ) , K ∞ , m ( B ; P ) : = κ m ( B ) ρ ∞ , m ( B ; P ) . (41) d_{\infty,m}^{\mathrm{ref}}(B\mathbin{\|}P):=\log\frac{\rho_{\infty,m}(B;P)}{\kappa_m(B)},\qquad K_{\infty,m}(B;P):=\frac{\kappa_m(B)}{\rho_{\infty,m}(B;P)}. \tag*{(41)} d ∞ , m ref ( B ∥ P ) := log κ m ( B ) ρ ∞ , m ( B ; P ) , K ∞ , m ( B ; P ) := ρ ∞ , m ( B ; P ) κ m ( B ) . (41)
In the special case where B B B is normalized we have b B ( x [ J ] ) = 1 b_B(x_{[J]})=1 b B ( x [ J ] ) = 1 , so that κ m ( B ) = 1 \kappa_m(B)=1 κ m ( B ) = 1 , and 1 α − 1 log Ψ α , m ( B ; P ) \frac{1}{\alpha-1}\log\Psi_{\alpha,m}(B;P) α − 1 1 log Ψ α , m ( B ; P ) is the maximal Rényi divergence between the conditional distributions B ( ⋅ ∣ x [ J ] ) B(\cdot\mid x_{[J]}) B ( ⋅ ∣ x [ J ] ) and P ( ⋅ ∣ x [ J ] ) P(\cdot\mid x_{[J]}) P ( ⋅ ∣ x [ J ] ) , when maximizing over the input x [ J ] x_{[J]} x [ J ] . In turn, the maximizer can be interpreted as the “most informative” input to differentiate between the channels P P P and B B B . We now provide a tractable bound on the Rényi divergence D α ( Q B π n , c ∥ P π n , c ) D_\alpha(Q_{B}^{\pi_n,c}\mathbin{\|}P^{\pi_n,c}) D α ( Q B π n , c ∥ P π n , c ) for any given protocol π n \pi_n π n , which makes the converse of Proposition 1 tractable for every protocol π n \pi_n π n and every product-form reference Q B π n , c Q_{B}^{\pi_n,c} Q B π n , c .
Lemma 6 (Kernel references). Let B ∈ M m , J ( P ) B\in\mathcal{M}_{m,J}(P) B ∈ M m , J ( P ) with κ m ( B ) > 0 \kappa_m(B)>0 κ m ( B ) > 0 . For every protocol π n \pi_n π n and every c c c :
β B π n , c ≥ κ m ( B ) n \beta_{B}^{\pi_n,c}\ge\kappa_m(B)^n β B π n , c ≥ κ m ( B ) n ;
D α ( Q B π n , c ∥ P π n , c ) ≤ n d α , m r e f ( B ∥ P ) D_\alpha(Q_{B}^{\pi_n,c}\mathbin{\|}P^{\pi_n,c})\le n\,d_{\alpha,m}^{\mathrm{ref}}(B\mathbin{\|}P) D α ( Q B π n , c ∥ P π n , c ) ≤ n d α , m ref ( B ∥ P ) for every α ∈ ( 1 , ∞ ] \alpha\in(1,\infty] α ∈ ( 1 , ∞ ] .
Proof sketch. The full proof is in Appendix A.
Summing out the unselected users as in the proof of Lemma 5, now on prefixes of the component sequence, shows that the branches that extend a given branch by one channel use have a total mass equal to the mass of that branch times the factor ∑ r [ E p A r ] m E q C r \sum_{r} [\mathbb{E}_{p} A_r]^m \mathbb{E}_{q} C_r ∑ r [ E p A r ] m E q C r . Here p p p is the next-input law of a selected user and q q q that of the unselected users on this branch, so the factor is at least κ m ( B ) \kappa_m(B) κ m ( B ) .
Conditioning on the past, the α \alpha α -th moment of the product in (35), that is, of d R B π n , c d P π n , c \frac{dR_{B}^{\pi_n,c}}{dP^{\pi_n,c}} d P π n , c d R B π n , c , grows by at most Ψ α , m ( B ; P ) \Psi_{\alpha,m}(B;P) Ψ α , m ( B ; P ) per channel use. Dividing by ( β B π n , c ) α (\beta_{B}^{\pi_n,c})^\alpha ( β B π n , c ) α and using the first item gives the claim.
The Kernel Converse
The above provides a lower bound on the error probability for any fixed protocol π n \pi_n π n and any value of c c c ; however, this bound depends on π n \pi_n π n and c c c . To obtain a lower bound which applies to all protocols, we average over C C C , and combine Proposition 1 with Lemma 6, which gives the main converse. This converse only involves single-letter quantities defined above.
Theorem 1 (Kernel converse). For every 2 ≤ m ≤ J 2 \le m \le J 2 ≤ m ≤ J , every B ∈ A m , J ( P ) B \in\mathcal{A}_{m,J}(P) B ∈ A m , J ( P ) with κ m ( B ) > 0 \kappa_m(B)>0 κ m ( B ) > 0 , every symmetric causal protocol π n \pi_n π n , and every α ∈ ( 1 , ∞ ] \alpha\in(1,\infty] α ∈ ( 1 , ∞ ] ,
P b ( n ) ( π n ; P ) ≥ γ m , J b α K α , m ( B ; P ) n . (42) P_{\mathrm{b}}^{(n)}(\pi_n;P) \ge\gamma_{m,J}^{b_\alpha}K_{\alpha,m}(B;P)^n. \tag*{(42)} P b ( n ) ( π n ; P ) ≥ γ m , J b α K α , m ( B ; P ) n . (42)
Corollary 1 (Optimized kernel converse). Let
K α , m r e f ( P ) : = sup B ∈ A m , J ( P ) κ m ( B ) > 0 K α , m ( B ; P ) , (43) K_{\alpha,m}^{\mathrm{ref}}(P) := \sup_{\substack{B\in\mathcal{A}_{m,J}(P)\\ \kappa_m(B)>0}} K_{\alpha,m}(B;P), \tag*{(43)} K α , m ref ( P ) := B ∈ A m , J ( P ) κ m ( B ) > 0 sup K α , m ( B ; P ) , (43) with the convention that an empty supremum is zero. Then
ϵ n ⋆ ( P ) ≥ sup 2 ≤ m ≤ J α ∈ ( 1 , ∞ ] γ m , J b α [ K α , m r e f ( P ) ] n , (44) \epsilon_n^\star(P) \ge\sup_{\substack{2\le m\le J\\ \alpha\in(1,\infty]}} \gamma_{m,J}^{b_\alpha}[K_{\alpha,m}^{\mathrm{ref}}(P)]^n, \tag*{(44)} ϵ n ⋆ ( P ) ≥ 2 ≤ m ≤ J α ∈ ( 1 , ∞ ] sup γ m , J b α [ K α , m ref ( P ) ] n , (44) E ‾ b ( P ) ≤ inf 2 ≤ m ≤ J α ∈ ( 1 , ∞ ] [ − log K α , m r e f ( P ) ] . (45) \overline{E}_{\mathrm{b}}(P) \le\inf_{\substack{2\le m\le J\\ \alpha\in(1,\infty]}}[-\log K_{\alpha,m}^{\mathrm{ref}}(P)]. \tag*{(45)} E b ( P ) ≤ 2 ≤ m ≤ J α ∈ ( 1 , ∞ ] inf [ − log K α , m ref ( P )] . (45)
Proof of Theorem 1 and Corollary 1. Fix c c c . By the first item of Lemma 6, β B π n , c ≥ κ m ( B ) n > 0 \beta_{B}^{\pi_n,c}\ge\kappa_m(B)^n>0 β B π n , c ≥ κ m ( B ) n > 0 , so that Q B π n , c Q_{B}^{\pi_n,c} Q B π n , c is defined, and it belongs to L m , J ( π n , c ; P ) \mathcal{L}_{m,J}(\pi_n,c;P) L m , J ( π n , c ; P ) by Lemma 5. Proposition 1 and the second item of Lemma 6 give P b ( n ) ( π n ; P ∣ C = c ) ≥ γ m , J b α e − n d α , m r e f ( B ∥ P ) = γ m , J b α K α , m ( B ; P ) n P_{\mathrm{b}}^{(n)}(\pi_n;P\mid C=c)\ge\gamma_{m,J}^{b_\alpha}e^{-nd_{\alpha,m}^{\mathrm{ref}}(B\Vert P)}=\gamma_{m,J}^{b_\alpha}K_{\alpha,m}(B;P)^n P b ( n ) ( π n ; P ∣ C = c ) ≥ γ m , J b α e − n d α , m ref ( B ∥ P ) = γ m , J b α K α , m ( B ; P ) n , and averaging over C C C proves (42). Optimizing over B B B , m m m and α \alpha α proves Corollary 1.
Note that the symmetry (2) of the channel is not used in this section. Therefore, the converses hold for every memoryless channel, provided that all users apply the same strategy.
a) Computing the coefficient: For s > 0 s>0 s > 0 , κ m ( s B ) = s κ m ( B ) \kappa_m(sB)=s\kappa_m(B) κ m ( s B ) = s κ m ( B ) and Ψ α , m ( s B ; P ) = s α Ψ α , m ( B ; P ) \Psi_{\alpha,m}(sB;P)=s^\alpha\Psi_{\alpha,m}(B;P) Ψ α , m ( s B ; P ) = s α Ψ α , m ( B ; P ) , so that K α , m ( B ; P ) K_{\alpha,m}(B;P) K α , m ( B ; P ) is scale invariant, and, for 1 < α < ∞ 1<\alpha<\infty 1 < α < ∞ , K α , m r e f ( P ) = ( Ψ ⋆ ) − 1 / ( α − 1 ) K_{\alpha,m}^{\mathrm{ref}}(P)=(\Psi^\star)^{-1/(\alpha-1)} K α , m ref ( P ) = ( Ψ ⋆ ) − 1/ ( α − 1 ) , where Ψ ⋆ \Psi^\star Ψ ⋆ is the infimum of Ψ α , m ( B ; P ) \Psi_{\alpha,m}(B;P) Ψ α , m ( B ; P ) over B ∈ A m , J ( P ) B\in\mathcal{A}_{m,J}(P) B ∈ A m , J ( P ) with κ m ( B ) ≥ 1 \kappa_m(B)\ge1 κ m ( B ) ≥ 1 . Under each of the completions (I) and (N), the admissible kernels form a convex cone (a sum merges the components), Ψ α , m ( ⋅ ; P ) \Psi_{\alpha,m}(\cdot;P) Ψ α , m ( ⋅ ; P ) is convex and κ m \kappa_m κ m is concave, so that the minimization is a convex program over each of the two cones, although neither has an explicit description in terms of the ∣ X ∣ J ∣ Y ∣ J |\mathcal{X}|^J|\mathcal{Y}|^J ∣ X ∣ J ∣ Y ∣ J entries of B B B . For m < J m<J m < J one solves it over both cones and keeps the better value; for m = J m=J m = J there are no unselected users, and the two cones coincide. Convexity does not make it easy: for m = J = 2 m=J=2 m = J = 2 , the product form says that B B B , as a matrix indexed by ( x 1 , y 1 ) (x_1,y_1) ( x 1 , y 1 ) and ( x 2 , y 2 ) (x_2,y_2) ( x 2 , y 2 ) , is completely positive, and membership in that cone is NP-hard [40 ] , as the order of the matrices grows. In practice one can optimize over a fixed number of components, a nonconvex problem; this affects only optimality, since every admissible B B B with κ m ( B ) > 0 \kappa_m(B)>0 κ m ( B ) > 0 gives a valid bound. Also, at order infinity,
K ∞ , m r e f ( P ) = sup { κ m ( B ) : B ∈ A m , J ( P ) , B ⪯ P } . (46) K_{\infty,m}^{\mathrm{ref}}(P)=\sup\{\kappa_m(B):B\in\mathcal{A}_{m,J}(P),\ B\preceq P\}. \tag*{(46)} K ∞ , m ref ( P ) = sup { κ m ( B ) : B ∈ A m , J ( P ) , B ⪯ P } . (46)
b) Choice of the order: For a normalized kernel, d α , m r e f ( B ∥ P ) d_{\alpha,m}^{\mathrm{ref}}(B\Vert P) d α , m ref ( B ∥ P ) is a maximal Rényi divergence, nondecreasing in α \alpha α [39 ] , so that α ↓ 1 \alpha\downarrow1 α ↓ 1 gives the best exponent, while at finite blocklength the prefactor γ m , J b α \gamma_{m,J}^{b_\alpha} γ m , J b α favors larger orders. If b B b_B b B is not constant, however, Ψ α , m ( B ; P ) → max x b B ( x ) > κ m ( B ) \Psi_{\alpha,m}(B;P)\to\max_x b_B(x)>\kappa_m(B) Ψ α , m ( B ; P ) → max x b B ( x ) > κ m ( B ) as α ↓ 1 \alpha\downarrow1 α ↓ 1 , so that d α , m r e f ( B ∥ P ) → + ∞ d_{\alpha,m}^{\mathrm{ref}}(B\Vert P)\to+\infty d α , m ref ( B ∥ P ) → + ∞ ; the best order is then finite or ∞ \infty ∞ , as for the common-output bounds and the optimized references of Table 1 .
Table 1. Exponents (nats per use) for μ = ( 0.8 , 0.8 , 0 ) \boldsymbol{\mu}=(0.8,0.8,0) μ = ( 0.8 , 0.8 , 0 ) , ε = ( ε , ε , 0 ) \boldsymbol{\varepsilon}=(\varepsilon,\varepsilon,0) ε = ( ε , ε , 0 ) , J = 2 J=2 J = 2 . Bold: the smallest converse, and the largest exponent of a scheme (proved for ε = 0 \varepsilon=0 ε = 0 , estimated for ε > 0 \varepsilon>0 ε > 0 ), in each column. †: valid bound found by a numerical search, which a better search can only improve; ‡: numerical estimate, not a proven bound (Appendix G). The name-decoding ceiling bounds the name-decoding exponent of every scheme (Corollary 3); it is neither a converse nor an achievable exponent.
c) Support bound: The effective persistence of a kernel is at most the physical probability of its support. Let S B ( x ) : = ∑ y : B ( y ∣ x ) > 0 P ( y ∣ x ) S_B(x):=\sum_{y:B(y\mid x)>0}P(y\mid x) S B ( x ) := ∑ y : B ( y ∣ x ) > 0 P ( y ∣ x ) be the physical probability of the support of B B B and t = ( α − 1 ) / α t=(\alpha-1)/\alpha t = ( α − 1 ) / α . Hölder’s inequality at each input gives b B ( x ) ≤ Ψ α , m ( B ; P ) 1 / α S B ( x ) t b_B(x)\le\Psi_{\alpha,m}(B;P)^{1/\alpha}S_B(x)^t b B ( x ) ≤ Ψ α , m ( B ; P ) 1/ α S B ( x ) t , and b B ( x ) ≤ ρ ∞ , m ( B ; P ) S B ( x ) b_B(x)\le\rho_{\infty,m}(B;P)S_B(x) b B ( x ) ≤ ρ ∞ , m ( B ; P ) S B ( x ) at order infinity. Averaging, minimizing, and using the concavity of s ↦ s t s\mapsto s^t s ↦ s t yields
K α , m ( B ; P ) ≤ [ min p , u E p ⊗ m S B ( X [ m ] , u ) t ] 1 / t ≤ min p , u E p ⊗ m S B ( X [ m ] , u ) , 1 < α < ∞ , K_{\alpha,m}(B;P)\le\left[\min_{p,u}\mathbb{E}_{p\otimes m}S_B(X_{[m]},u)^t\right]^{1/t}\le\min_{p,u}\mathbb{E}_{p\otimes m}S_B(X_{[m]},u),\qquad1<\alpha<\infty, K α , m ( B ; P ) ≤ [ p , u min E p ⊗ m S B ( X [ m ] , u ) t ] 1/ t ≤ p , u min E p ⊗ m S B ( X [ m ] , u ) , 1 < α < ∞ ,
K ∞ , m ( B ; P ) ≤ min p , u E p ⊗ m S B ( X [ m ] , u ) . (47) K_{\infty,m}(B;P) \leq\min_{p,u} \mathbb{E}_{p\otimes m} S_B(X_{[m]},u). \tag*{(47)} K ∞ , m ( B ; P ) ≤ p , u min E p ⊗ m S B ( X [ m ] , u ) . (47)
Thus a reference cannot retain more mass than the channel places on the outputs it keeps. At order infinity, the bound holds with equality for every admissible restriction of the channel, B ( y ∣ x ) = P ( y ∣ x ) 1 { y ∈ S ( x ) } B(y\mid x)=P(y\mid x)1\{y\in S(x)\} B ( y ∣ x ) = P ( y ∣ x ) 1 { y ∈ S ( x )} with S ( x ) ⊆ Y J S(x)\subseteq\mathcal{Y}^J S ( x ) ⊆ Y J , since then ρ ∞ , m ( B ; P ) = 1 \rho_{\infty,m}(B;P)=1 ρ ∞ , m ( B ; P ) = 1 and b B = S B b_B=S_B b B = S B .
Noninteracting and Common-Output References
We now evaluate Theorem 1 for two explicit families of admissible kernels: noninteracting references, under which the output of each selected user depends only on its own input, and common-output references, under which the selected users always receive the same output.
a) Noninteracting references: Let V V V be a one-user channel from X \mathcal{X} X to Y \mathcal{Y} Y and write
q m , V ( y [ m ] ∣ x [ m ] ) : = ∏ j = 1 m V ( y j ∣ x j ) , P m ( y [ m ] ∣ x [ J ] ) : = ∑ v P ( y [ m ] , v ∣ x [ J ] ) . (48) q_{m,V}(y_{[m]}\mid x_{[m]}) := \prod_{j=1}^{m} V(y_j\mid x_j), \qquad P_m(y_{[m]}\mid x_{[J]}) := \sum_v P(y_{[m]},v\mid x_{[J]}). \tag*{(48)} q m , V ( y [ m ] ∣ x [ m ] ) := j = 1 ∏ m V ( y j ∣ x j ) , P m ( y [ m ] ∣ x [ J ] ) := v ∑ P ( y [ m ] , v ∣ x [ J ] ) . (48)
Complete the selected outputs with the physical conditional law of the unselected outputs:
B V ( y [ J ] ∣ x [ J ] ) : = q m , V ( y [ m ] ∣ x [ m ] ) P ( v ∣ x [ J ] , y [ m ] ) , (49) B_V(y_{[J]}\mid x_{[J]}) := q_{m,V}(y_{[m]}\mid x_{[m]})P(v\mid x_{[J]},y_{[m]}), \tag*{(49)} B V ( y [ J ] ∣ x [ J ] ) := q m , V ( y [ m ] ∣ x [ m ] ) P ( v ∣ x [ J ] , y [ m ] ) , (49)
with an arbitrary probability law in place of P ( ⋅ ∣ x [ J ] , y [ m ] ) P(\cdot\mid x_{[J]},y_{[m]}) P ( ⋅ ∣ x [ J ] , y [ m ] ) where P m ( y [ m ] ∣ x [ J ] ) = 0 P_m(y_{[m]}\mid x_{[J]})=0 P m ( y [ m ] ∣ x [ J ] ) = 0 . Under B V B_V B V each selected user behaves as if it were alone on the channel V V V , so the selected users do not interact. This kernel has one component (∣ R ∣ = 1 |\mathcal{R}|=1 ∣ R ∣ = 1 , a = V a=V a = V ) and a normalized completion, so B V B_V B V satisfies (N) and b B V ≡ 1 b_{B_V}\equiv1 b B V ≡ 1 , hence κ m ( B V ) = 1 \kappa_m(B_V)=1 κ m ( B V ) = 1 . It satisfies B V ≪ P B_V\ll P B V ≪ P if and only if q m , V ( ⋅ ∣ x [ m ] ) ≪ P m ( ⋅ ∣ x [ J ] ) q_{m,V}(\cdot\mid x_{[m]})\ll P_m(\cdot\mid x_{[J]}) q m , V ( ⋅ ∣ x [ m ] ) ≪ P m ( ⋅ ∣ x [ J ] ) for every x [ J ] x_{[J]} x [ J ] . In that case the completion cancels in B V / P B_V/P B V / P , and Ψ α , m ( B V ; P ) = e ( α − 1 ) d α , m ( V ∥ P ) \Psi_{\alpha,m}(B_V;P)=e^{(\alpha-1)d_{\alpha,m}(V\Vert P)} Ψ α , m ( B V ; P ) = e ( α − 1 ) d α , m ( V ∥ P ) with the one-step radius
d α , m ( V ∥ P ) : = max x [ J ] D α ( q m , V ( ⋅ ∣ x [ m ] ) ∥ P m ( ⋅ ∣ x [ J ] ) ) . (50) d_{\alpha,m}(V\Vert P) := \max_{x_{[J]}} D_\alpha\left(q_{m,V}(\cdot\mid x_{[m]})\middle\Vert P_m(\cdot\mid x_{[J]})\right). \tag*{(50)} d α , m ( V ∥ P ) := x [ J ] max D α ( q m , V ( ⋅ ∣ x [ m ] ) P m ( ⋅ ∣ x [ J ] ) ) . (50)
At α = 1 \alpha=1 α = 1 the Rényi divergence is the KL divergence, and at α = ∞ \alpha=\infty α = ∞ it is the logarithm of the largest likelihood ratio. We set d α , m ( V ∥ P ) = ∞ d_{\alpha,m}(V\Vert P)=\infty d α , m ( V ∥ P ) = ∞ if absolute continuity fails.
Corollary 2 (Noninteraction bound). For every one-user channel V V V , every 2 ≤ m ≤ J 2\leq m\leq J 2 ≤ m ≤ J , every symmetric causal protocol π n \pi_n π n , and every α ∈ ( 1 , ∞ ] \alpha\in(1,\infty] α ∈ ( 1 , ∞ ] ,
P b ( n ) ( π n ; P ) ≥ γ m , J b α e − n d α , m ( V ∥ P ) , (51) P_b^{(n)}(\pi_n;P) \geq\gamma_{m,J}^{b_\alpha}e^{-nd_{\alpha,m}(V\Vert P)}, \tag*{(51)} P b ( n ) ( π n ; P ) ≥ γ m , J b α e − n d α , m ( V ∥ P ) , (51) and hence
E ‾ b ( P ) ≤ inf V d 1 , 2 ( V ∥ P ) . (52) \overline{E}_b(P) \leq\inf_V d_{1,2}(V\Vert P). \tag*{(52)} E b ( P ) ≤ V inf d 1 , 2 ( V ∥ P ) . (52) Proof. The finite-blocklength bound is Theorem 1 for B = B V B=B_V B = B V , since K α , m ( B V ; P ) = e − d α , m ( V ∥ P ) K_{\alpha,m}(B_V;P)=e^{-d_{\alpha,m}(V\Vert P)} K α , m ( B V ; P ) = e − d α , m ( V ∥ P ) . For fixed V V V and finite alphabets, d α , m ( V ∥ P ) d_{\alpha,m}(V\Vert P) d α , m ( V ∥ P ) is nondecreasing in α \alpha α and tends to d 1 , m ( V ∥ P ) d_{1,m}(V\Vert P) d 1 , m ( V ∥ P ) as α ↓ 1 \alpha\downarrow1 α ↓ 1 . Marginalizing one selected output gives d α , m ≤ d α , m + 1 d_{\alpha,m}\leq d_{\alpha,m+1} d α , m ≤ d α , m + 1 by data processing. Hence m = 2 m=2 m = 2 and α ↓ 1 \alpha\downarrow1 α ↓ 1 give the best exponent. At finite blocklength, m m m still trades the cost against γ m , J \gamma_{m,J} γ m , J .
For noninteracting references the best exponent is obtained as α ↓ 1 \alpha\downarrow1 α ↓ 1 , and at finite blocklength the order can be optimized for each n n n : if d α , m ( V ∥ P ) = d 1 , m ( V ∥ P ) + ( α − 1 ) ϑ + o ( α − 1 ) d_{\alpha,m}(V\Vert P)=d_{1,m}(V\Vert P)+(\alpha-1)\vartheta+o(\alpha-1) d α , m ( V ∥ P ) = d 1 , m ( V ∥ P ) + ( α − 1 ) ϑ + o ( α − 1 ) with ϑ > 0 \vartheta>0 ϑ > 0 , the order α ∗ = 1 + log ( 1 / γ m , J ) / ( n ϑ ) \alpha^*=1+\sqrt{\log(1/\gamma_{m,J})/(n\vartheta)} α ∗ = 1 + log ( 1/ γ m , J ) / ( n ϑ ) gives P b ( n ) ≥ exp { − n d 1 , m ( V ∥ P ) − 2 n ϑ log ( 1 / γ m , J ) + o ( n ) } P_b^{(n)}\geq\exp\{-n d_{1,m}(V\Vert P)-2\sqrt{n\vartheta\log(1/\gamma_{m,J})}+o(\sqrt{n})\} P b ( n ) ≥ exp { − n d 1 , m ( V ∥ P ) − 2 n ϑ log ( 1/ γ m , J ) + o ( n )} .
Remark 1 (Other normalized completions). The physical completion in (49) can be combined with any selected factor: the kernel B = ∑ r ∈ R ∏ j ≤ m a r ( x j , y j ) P ( v ∣ x [ J ] , y [ m ] ) B=\sum_{r\in\mathcal{R}}\prod_{j\leq m}a_r(x_j,y_j)P(v\mid x_{[J]},y_{[m]}) B = ∑ r ∈ R ∏ j ≤ m a r ( x j , y j ) P ( v ∣ x [ J ] , y [ m ] ) satisfies condition (N), κ m ( B ) = min p ∑ r [ ∑ x p ( x ) A r ( x ) ] m \kappa_m(B)=\min_p\sum_r\left[\sum_xp(x)A_r(x)\right]^m κ m ( B ) = min p ∑ r [ ∑ x p ( x ) A r ( x ) ] m , and, whenever B ≪ P B\ll P B ≪ P , Ψ α , m ( B ; P ) \Psi_{\alpha,m}(B;P) Ψ α , m ( B ; P ) involves only the m m m -user marginal P m P_m P m . Noninteracting references are the normalized one-component case.
b) Common-output references: Let O m : = { y ∈ Y J : y 1 = ⋯ = y m } \mathcal{O}_m:=\{y\in\mathcal{Y}^J:y_1=\cdots=y_m\} O m := { y ∈ Y J : y 1 = ⋯ = y m } be the set of output vectors on which the selected users agree. Let C m , J ( P ) \mathcal{C}_{m,J}(P) C m , J ( P ) be the set of kernels B B B that satisfy
B ( y [ J ] ∣ x [ J ] ) = ∑ r ∈ R 1 { y [ J ] = y ˉ r } [ ∏ j = 1 m a r ( x j ) ] c r ( x m + 1 : J ) , (53) B(y_{[J]}\mid x_{[J]})=\sum_{r\in\mathcal{R}}1\{y_{[J]}=\bar y_r\}\left[\prod_{j=1}^{m}a_r(x_j)\right]c_r(x_{m+1:J}), \tag*{(53)} B ( y [ J ] ∣ x [ J ] ) = r ∈ R ∑ 1 { y [ J ] = y ˉ r } [ j = 1 ∏ m a r ( x j ) ] c r ( x m + 1 : J ) , (53)
with R \mathcal{R} R finite, an output vector y ˉ r ∈ O m \bar y_r\in\mathcal{O}_m y ˉ r ∈ O m attached to each component r r r , nonnegative factors a r a_r a r on X \mathcal{X} X and c r c_r c r on X J − m \mathcal{X}^{J-m} X J − m , and B ≪ P B\ll P B ≪ P . Such a kernel is supported on O m \mathcal{O}_m O m . Writing y ˉ r = ( y ˉ r , 1 , … , y ˉ r , J ) \bar y_r=(\bar y_{r,1},\ldots,\bar y_{r,J}) y ˉ r = ( y ˉ r , 1 , … , y ˉ r , J ) , it has the form (34) with selected factor a r ( x ) 1 { y = y ˉ r , 1 } a_r(x)1\{y=\bar y_{r,1}\} a r ( x ) 1 { y = y ˉ r , 1 } and unselected factor c r ( u ) 1 { v = y ˉ r , m + 1 : J } c_r(u)1\{v=\bar y_{r,m+1:J}\} c r ( u ) 1 { v = y ˉ r , m + 1 : J } . The unselected factor depends only on the unselected users, so the kernel satisfies (I). The physical completion of Remark 1 gives a second family, with the same selected factors and the unselected factor P ( v ∣ x [ J ] , y [ m ] ) P(v\mid x_{[J]},y_{[m]}) P ( v ∣ x [ J ] , y [ m ] ) , which satisfies (N); the kernel of Proposition 12 is of this kind. For m = J m=J m = J the two families coincide, and we call the kernels of both families common-output kernels. Under such a reference the selected users always receive identical outputs, so their views differ only through their private seeds. The reference retains only trajectories on which the channel gives the selected users no distinguishing information, and κ m \kappa_m κ m measures how much physical probability these trajectories keep, uniformly over protocols.
Common-output kernels are admissible, so that they give at most K α , m r e f ( P ) K^{\mathrm{ref}}_{\alpha,m}(P) K α , m ref ( P ) , and the inequality can be strict: for a noninteracting channel, P ( y [ J ] ∣ x [ J ] ) = ∏ j V ( y j ∣ x j ) P(y_{[J]} \mid x_{[J]}) = \prod_{j} V(y_j \mid x_j) P ( y [ J ] ∣ x [ J ] ) = ∏ j V ( y j ∣ x j ) with no V ( ⋅ ∣ x ) V(\cdot\mid x) V ( ⋅ ∣ x ) a point mass, the reference B V = P B_V = P B V = P gives K α , m r e f ( P ) = 1 K^{\mathrm{ref}}_{\alpha,m}(P) = 1 K α , m ref ( P ) = 1 , whereas by (47) every common-output kernel has K α , m ( B ; P ) ≤ max x [ J ] P ( Y 1 = ⋯ = Y m ∣ x [ J ] ) < 1 K_{\alpha,m}(B;P) \le\max_{x_{[J]}} P(Y_1 = \cdots= Y_m \mid x_{[J]}) < 1 K α , m ( B ; P ) ≤ max x [ J ] P ( Y 1 = ⋯ = Y m ∣ x [ J ] ) < 1 .
Remark 2 (Extension to leader election). Let each user output a bit L j L_j L j , and let E L E : = { ∑ j L j ≠ 1 } \mathcal{E}_{\mathrm{LE}} := \{\sum_j L_j \ne1\} E LE := { ∑ j L j = 1 } , whose probability under π n \pi_n π n is P L E ( n ) ( π n ; P ) P_{\mathrm{LE}}^{(n)}(\pi_n;P) P LE ( n ) ( π n ; P ) ; let E ‾ L E ( P ) \overline{E}_{\mathrm{LE}}(P) E LE ( P ) be the upper exponent of inf π n P L E ( n ) ( π n ; P ) \inf_{\pi_n} P_{\mathrm{LE}}^{(n)}(\pi_n;P) inf π n P LE ( n ) ( π n ; P ) . Take m = J m = J m = J . Under any J J J -symmetric reference, the terminal bits are conditionally i.i.d. Bernoulli given Λ \Lambda Λ . Hence
Q ( E L E ) ≥ η J : = 1 − max p ∈ [ 0 , 1 ] J p ( 1 − p ) J − 1 = 1 − ( 1 − 1 J ) J − 1 > 0. (54) Q(\mathcal{E}_{\mathrm{LE}}) \ge\eta_J := 1 - \max_{p \in[0,1]} Jp(1-p)^{J-1} = 1 - \left(1-\frac{1}{J}\right)^{J-1} > 0. \tag*{(54)} Q ( E LE ) ≥ η J := 1 − p ∈ [ 0 , 1 ] max J p ( 1 − p ) J − 1 = 1 − ( 1 − J 1 ) J − 1 > 0. (54)
The same proof gives, for every B ∈ A J , J ( P ) B \in\mathcal{A}_{J,J}(P) B ∈ A J , J ( P ) with κ J ( B ) > 0 \kappa_J(B) > 0 κ J ( B ) > 0 ,
P L E ( n ) ( π n ; P ) ≥ η J K α , J ( B ; P ) n , E ‾ L E ( P ) ≤ inf α ∈ ( 1 , ∞ ] [ − log K α , J r e f ( P ) ] . (55) P_{\mathrm{LE}}^{(n)}(\pi_n;P) \ge\eta_J K_{\alpha,J}(B;P)^n,\qquad\overline{E}_{\mathrm{LE}}(P) \le\inf_{\alpha\in(1,\infty]}[-\log K^{\mathrm{ref}}_{\alpha,J}(P)]. \tag*{(55)} P LE ( n ) ( π n ; P ) ≥ η J K α , J ( B ; P ) n , E LE ( P ) ≤ α ∈ ( 1 , ∞ ] inf [ − log K α , J ref ( P )] . (55)
Unlike for symmetry breaking, where any m ≥ 2 m \ge2 m ≥ 2 selected users suffice, the reference must make all J J J users conditionally i.i.d.: a proper subset can consist entirely of followers while an unselected user is the leader. For J = 2 J = 2 J = 2 , η 2 = γ 2 = 1 / 2 \eta_2 = \gamma_2 = 1/2 η 2 = γ 2 = 1/2 , and symmetry breaking and basic leader election are equivalent.
Proposition 2 (General population vectors). Consider the population version of the task mentioned in Section I-A: each user outputs Z j ∈ [ L ] Z_j \in[L] Z j ∈ [ L ] through a common terminal map, and the task fails unless exactly N l N_l N l users output l l l , for every l ∈ [ L ] l \in[L] l ∈ [ L ] . Let P N ( n ) ( π n ; P ) P_N^{(n)}(\pi_n;P) P N ( n ) ( π n ; P ) be the probability of failure, N max : = max l N l N_{\max} := \max_l N_l N m a x := max l N l , and, for N max < m ≤ J N_{\max} < m \le J N m a x < m ≤ J ,
γ m ( N ) : = min p ∈ P ( [ L ] ) P p ⊗ m { ∃ l : ∣ { j ≤ m : Z j = l } ∣ > N l } ≥ L 1 − m . (56) \gamma_m(\mathbf{N}) := \min_{p \in\mathcal{P}([L])} \mathbb{P}_{p^{\otimes m}}\left\{\exists l : \left|\{j \le m : Z_j = l\}\right| > N_l\right\} \ge L^{1-m}. \tag*{(56)} γ m ( N ) := p ∈ P ([ L ]) min P p ⊗ m { ∃ l : ∣ { j ≤ m : Z j = l } ∣ > N l } ≥ L 1 − m . (56) Then, for every such m m m , every B ∈ A m , J ( P ) B \in\mathcal{A}_{m,J}(P) B ∈ A m , J ( P ) with κ m ( B ) > 0 \kappa_m(B) > 0 κ m ( B ) > 0 , every strategy π n \pi_n π n and every α ∈ ( 1 , ∞ ] \alpha\in(1,\infty] α ∈ ( 1 , ∞ ] ,
P N ( n ) ( π n ; P ) ≥ γ m ( N ) b α K α , m ( B ; P ) n . (57) P_N^{(n)}(\pi_n;P) \ge\gamma_m(\mathbf{N})^{b_\alpha} K_{\alpha,m}(B;P)^n. \tag*{(57)} P N ( n ) ( π n ; P ) ≥ γ m ( N ) b α K α , m ( B ; P ) n . (57) Proof. Under an m m m -symmetric reference, Z 1 , … , Z m Z_1,\ldots,Z_m Z 1 , … , Z m are conditionally i.i.d. given Λ \Lambda Λ , with some law p p p on [ L ] [L] [ L ] . Success requires that at most N l N_l N l of them output l l l , for every l l l , so that the reference fails with probability at least γ m ( N ) \gamma_m(\mathbf{N}) γ m ( N ) . The event Z 1 = ⋯ = Z m Z_1 = \cdots= Z_m Z 1 = ⋯ = Z m has conditional probability ∑ l p l m ≥ L 1 − m \sum_l p_l^m \ge L^{1-m} ∑ l p l m ≥ L 1 − m and implies failure, since m > N max m > N_{\max} m > N m a x , which proves the bound in (56) . The proofs of Proposition 1 and Theorem 1 then apply verbatim, with γ m ( N ) \gamma_m(\mathbf{N}) γ m ( N ) in place of γ m , J \gamma_{m,J} γ m , J .
The range of m m m is sharp: for m ≤ N max m \le N_{\max} m ≤ N m a x , the law that puts all its mass on a state l l l with N l = N max N_l = N_{\max} N l = N m a x gives zero in (56) , so the bound is vacuous. For N = ( 1 , … , 1 ) \mathbf{N} = (1,\ldots,1) N = ( 1 , … , 1 ) , γ m ( N ) = γ m , J \gamma_m(\mathbf{N}) = \gamma_{m,J} γ m ( N ) = γ m , J , which is Lemma 3, and for N = ( 1 , J − 1 ) \mathbf{N} = (1,J-1) N = ( 1 , J − 1 ) , only m = J m = J m = J is allowed and γ J ( N ) = η J \gamma_J(\mathbf{N}) = \eta_J γ J ( N ) = η J , which is Remark 2. The exponents of these bounds are those of symmetry breaking with m > N max m > N_{\max} m > N m a x selected users, so that N \mathbf{N} N affects only the prefactor and the range of m m m ; with noninteracting references, for instance, the best exponent bound is inf V d 1 , N max + 1 ( V ∥ P ) \inf_V d_{1,N_{\max}+1}(V\Vert P) inf V d 1 , N m a x + 1 ( V ∥ P ) , since d 1 , m d_{1,m} d 1 , m is nondecreasing in m m m .
Achievability with Block Coding
In a block code , each user transmits the codeword of the common codebook indexed by its private name and ignores its outputs until the end of the block. The outputs are used only in the terminal decision: user j j j decodes the unordered set of the other names from its own codeword and its outputs, and ranks its own name as in (15) .
The Block-Coding Exponent
The codebook is drawn from the i.i.d. ensemble (19) with input law Q X Q_X Q X . For x ∈ X k x \in\mathcal{X}^k x ∈ X k write Q X k ( x ) : = ∏ i = 1 k Q X ( x i ) Q_X^k(x) := \prod_{i=1}^k Q_X(x_i) Q X k ( x ) := ∏ i = 1 k Q X ( x i ) .
a) Decoder: Receiver j j j knows its name M j M_j M j , hence its codeword X j n = X n ( M j ) X_j^n = X^n(M_j) X j n = X n ( M j ) . For an unordered set A ⊆ [ M ] ∖ { M j } \mathcal{A} \subseteq[M] \setminus\{M_j\} A ⊆ [ M ] ∖ { M j } of J − 1 J-1 J − 1 names, listed as a 1 , … , a J − 1 a_1,\ldots,a_{J-1} a 1 , … , a J − 1 , let
L j ( A ) : = ∏ i = 1 n W ( Y j [ i ] ∣ X j [ i ] , ( X [ i ] ( a 1 ) , … , X [ i ] ( a J − 1 ) ) ) . (58) L_j(\mathcal{A}) := \prod_{i=1}^n W(Y_j[i] \mid X_j[i],(X[i](a_1),\ldots,X[i](a_{J-1}))). \tag*{(58)} L j ( A ) := i = 1 ∏ n W ( Y j [ i ] ∣ X j [ i ] , ( X [ i ] ( a 1 ) , … , X [ i ] ( a J − 1 ))) . (58)
By Lemma 1, L j ( A ) L_j(\mathcal{A}) L j ( A ) does not depend on the order in which the elements of A \mathcal{A} A are listed, so it is a function of the unordered set. The receiver outputs M ^ − j ∈ arg max A L j ( A ) \widehat{\mathcal{M}}_{-j} \in\arg\max_{\mathcal{A}} L_j(\mathcal{A}) M − j ∈ arg max A L j ( A ) , chosen by a fixed rule if the maximizer is not unique. The bounds below are for the event that some set other than the true one has a likelihood at least that of the true set, which contains the decoding error event. No physical user labels are attached to the decoded names.
b) Error events: A decoding error replaces the names of some nonempty set of foreign users while keeping the others (Appendix B shows that every error is of this form). For a receiver j j j , a subset S ⊆ [ J ] ∖ { j } S \subseteq[J] \setminus\{j\} S ⊆ [ J ] ∖ { j } of size s s s , and ρ ∈ [ 0 , 1 ] \rho\in[0,1] ρ ∈ [ 0 , 1 ] , define
E 0 , s ( ρ , Q X ) : = − log ∑ x S c , y j Q X J − s ( x S c ) [ ∑ x S Q X s ( x S ) W j ( y j ∣ x [ J ] ) 1 1 + ρ ] 1 + ρ (59) E_{0,s}(\rho,Q_X) := -\log\sum_{x_{S^c},y_j} Q_X^{J-s}(x_{S^c}) \left[\sum_{x_S} Q_X^s(x_S) W_j(y_j \mid x_{[J]})^{\frac{1}{1+\rho}}\right]^{1+\rho} \tag*{(59)} E 0 , s ( ρ , Q X ) := − log x S c , y j ∑ Q X J − s ( x S c ) [ x S ∑ Q X s ( x S ) W j ( y j ∣ x [ J ] ) 1 + ρ 1 ] 1 + ρ (59)
where S c = [ J ] ∖ S S^c=[J]\setminus S S c = [ J ] ∖ S contains the receiver j j j . By Lemma 1, the right-hand side depends on s s s but not on j j j or the particular subset. It is Gallager’s function for the channel from X S X_S X S to ( X S c , Y j ) (X_{S^c},Y_j) ( X S c , Y j ) with i.i.d. inputs: the receiver’s own codeword and the correctly decoded names act as side information.
Define
E r , s ( R , Q X ) : = max 0 ≤ ρ ≤ 1 { E 0 , s ( ρ , Q X ) − ρ s R } . (60) E_{r,s}(R,Q_X) := \max_{0\leq\rho\leq1}\{E_{0,s}(\rho,Q_X)-\rho sR\}. \tag*{(60)} E r , s ( R , Q X ) := 0 ≤ ρ ≤ 1 max { E 0 , s ( ρ , Q X ) − ρ s R } . (60)
Theorem 2 (Block-coding exponent). Fix J ≥ 2 J \geq2 J ≥ 2 and Q X Q_X Q X .
(i) For every receiver j j j , the ensemble-average decoding error probability, conditioned on distinct names, satisfies
E C [ P e , j ( n ) ( C ) ] ≤ ∑ s = 1 J − 1 ( J − 1 s ) inf 0 ≤ ρ ≤ 1 ( M − J ) s ρ e − n E 0 , s ( ρ , Q X ) . (61) \mathbb{E}_{\mathsf C}[P_{e,j}^{(n)}(\mathsf C)] \leq\sum_{s=1}^{J-1} \binom{J-1}{s}\inf_{0\leq\rho\leq1}(M-J)^{s\rho}e^{-nE_{0,s}(\rho,Q_X)}. \tag*{(61)} E C [ P e , j ( n ) ( C )] ≤ s = 1 ∑ J − 1 ( s J − 1 ) 0 ≤ ρ ≤ 1 inf ( M − J ) s ρ e − n E 0 , s ( ρ , Q X ) . (61) (ii) For M = ⌈ e n R ⌉ M=\lceil e^{nR}\rceil M = ⌈ e n R ⌉ , there exists a deterministic common codebook such that
P b ( n ) ≤ ( J 2 ) e − n R + o ( n ) + J ∑ s = 1 J − 1 ( J − 1 s ) e − n E r , s ( R , Q X ) + o ( n ) . (62) P_b^{(n)} \leq\binom{J}{2}e^{-nR+o(n)}+J\sum_{s=1}^{J-1}\binom{J-1}{s}e^{-nE_{r,s}(R,Q_X)+o(n)}. \tag*{(62)} P b ( n ) ≤ ( 2 J ) e − n R + o ( n ) + J s = 1 ∑ J − 1 ( s J − 1 ) e − n E r , s ( R , Q X ) + o ( n ) . (62) (iii) Consequently, E b ( P ) ≥ E b c ( Q X ; P ) E_b(P)\geq E_{\mathrm{bc}}(Q_X;P) E b ( P ) ≥ E bc ( Q X ; P ) , where
E b c ( Q X ; P ) : = sup R > 0 min { R , min 1 ≤ s ≤ J − 1 E r , s ( R , Q X ) } . (63) E_{\mathrm{bc}}(Q_X;P):=\sup_{R>0}\min\left\{R,\min_{1\leq s\leq J-1}E_{r,s}(R,Q_X)\right\}. \tag*{(63)} E bc ( Q X ; P ) := R > 0 sup min { R , 1 ≤ s ≤ J − 1 min E r , s ( R , Q X ) } . (63)
The proof is given in Appendix B. In (61), the factor ( J − 1 s ) \binom{J-1}{s} ( s J − 1 ) chooses the users whose names are replaced, while ( M − J ) s (M-J)^s ( M − J ) s upper bounds the number of competing unordered sets. The exponent (63) is a minimum of two competing terms. Duplicate names occur with probability of order e − n R e^{-nR} e − n R , which favors a large name rate. Longer names are harder to decode, since E r , s ( R , Q X ) E_{r,s}(R,Q_X) E r , s ( R , Q X ) decreases in R R R , which favors a small one. The name rate is therefore a design parameter to be optimized. As in the random-coding exponents of the multiple-access channel [41 ] , [42 ] , (61) is a union bound over the sets of wrongly decoded users; here the receiver decodes an unordered set of names, with its own codeword as side information. The exponent E b c ( Q X ; P ) E_{\mathrm{bc}}(Q_X;P) E bc ( Q X ; P ) is achieved by one ensemble with one decoder; it is not an upper bound for block codes.
For J = 2 J=2 J = 2 the only error event is s = 1 s=1 s = 1 , and with { j , j ′ } = [ 2 ] \{j,j'\}=[2] { j , j ′ } = [ 2 ] , part (i) is the usual Gallager bound E C [ P e , j ( n ) ( C ) ] ≤ ( M − 2 ) ρ e − n E 0 , 1 ( ρ , Q X ) \mathbb{E}_{\mathsf C}[P_{e,j}^{(n)}(\mathsf C)]\leq(M-2)^\rho e^{-nE_{0,1}(\rho,Q_X)} E C [ P e , j ( n ) ( C )] ≤ ( M − 2 ) ρ e − n E 0 , 1 ( ρ , Q X ) , with the receiver’s own input as side information; the derivative of E 0 , 1 E_{0,1} E 0 , 1 at ρ = 0 \rho=0 ρ = 0 is I Q ( X j ′ ; Y j ∣ X j ) I_{Q}(X_{j'};Y_j\mid X_j) I Q ( X j ′ ; Y j ∣ X j ) , the rate of Shannon’s independent-input inner bound for the two-way channel [10 ] .
Condition for a Positive Block-Coding Exponent
For a nonempty subset S ⊆ [ J ] ∖ { j } S\subseteq[J]\setminus\{j\} S ⊆ [ J ] ∖ { j } with ∣ S ∣ = s |S|=s ∣ S ∣ = s and i.i.d. inputs X 1 , … , X J X_1,\ldots,X_J X 1 , … , X J with law Q X Q_X Q X , define
I s ( Q X ) : = I Q ( X S ; Y j ∣ X S c ) , (64) I_s(Q_X):=I_Q(X_S;Y_j\mid X_{S^c}), \tag*{(64)} I s ( Q X ) := I Q ( X S ; Y j ∣ X S c ) , (64)
where $S^c:=[J]\setminus S includes the receiver input X j X_j X j . By user symmetry, the value depends only on s s s .
Differentiating (59) gives
∂ E 0 , s ( ρ , Q X ) ∂ ρ ∣ ρ = 0 = I s ( Q X ) , (65) \left.\frac{\partial E_{0,s}(\rho,Q_X)}{\partial\rho}\right|_{\rho=0}=I_s(Q_X), \tag*{(65)} ∂ ρ ∂ E 0 , s ( ρ , Q X ) ρ = 0 = I s ( Q X ) , (65)
and E 0 , s ( ⋅ , Q X ) E_{0,s}(\cdot,Q_X) E 0 , s ( ⋅ , Q X ) is concave with E 0 , s ( 0 , Q X ) = 0 E_{0,s}(0,Q_X)=0 E 0 , s ( 0 , Q X ) = 0 [37 ] . Hence E r , s ( R , Q X ) > 0 E_{r,s}(R,Q_X)>0 E r , s ( R , Q X ) > 0 if and only if s R < I s ( Q X ) sR<I_s(Q_X) s R < I s ( Q X ) . The name rates at which every decoding exponent is positive are therefore those below
R t h ( Q X ) : = min 1 ≤ s ≤ J − 1 I s ( Q X ) s . (66) R_{\mathrm{th}}(Q_X):=\min_{1\leq s\leq J-1}\frac{I_s(Q_X)}{s}. \tag*{(66)} R th ( Q X ) := 1 ≤ s ≤ J − 1 min s I s ( Q X ) . (66)
This is also the symmetric rate of the all-source multicast inner bound: the constraints s R < I s ( Q X ) sR<I_s(Q_X) s R < I s ( Q X ) are those of a multiple-access channel with the receiver’s input as side information. The following lemma eliminates the minimization over subset sizes.
Lemma 7 (Normalized subset information is nonincreasing). Suppose X 1 , … , X m X_1,\ldots,X_m X 1 , … , X m are i.i.d. and independent of T T T , and the conditional law of Y Y Y given ( X 1 , … , X m , T ) (X_1,\ldots,X_m,T) ( X 1 , … , X m , T ) is invariant under permutations of the m m m inputs. Then
1 s I ( X [ s ] ; Y ∣ X [ m ] ∖ [ s ] , T ) (67) \frac{1}{s}I(X_{[s]};Y\mid X_{[m]\setminus[s]},T) \tag*{(67)} s 1 I ( X [ s ] ; Y ∣ X [ m ] ∖ [ s ] , T ) (67)
is nonincreasing in s = 1 , … , m s=1,\ldots,m s = 1 , … , m .
A complete entropy proof is given in Appendix C.
Proposition 3 (Rate threshold of block coding). For every symmetric channel and every Q X Q_X Q X ,
R t h ( Q X ) = 1 J − 1 I Q ( X − j ; Y j ∣ X j ) , (68) R_{\mathrm{th}}(Q_X)=\frac{1}{J-1}I_Q(X_{-j};Y_j\mid X_j), \tag*{(68)} R th ( Q X ) = J − 1 1 I Q ( X − j ; Y j ∣ X j ) , (68) and
E b c ( Q X ; P ) = min 1 ≤ s ≤ J − 1 max 0 ≤ ρ ≤ 1 E 0 , s ( ρ , Q X ) 1 + s ρ ≤ 1 J I Q ( X − j ; Y j ∣ X j ) . (69) E_{\mathrm{bc}}(Q_X;P)=\min_{1\le s\le J-1}\max_{0\le\rho\le1}\frac{E_{0,s}(\rho,Q_X)}{1+s\rho}\le\frac{1}{J}I_Q(X_{-j};Y_j\mid X_j). \tag*{(69)} E bc ( Q X ; P ) = 1 ≤ s ≤ J − 1 min 0 ≤ ρ ≤ 1 max 1 + s ρ E 0 , s ( ρ , Q X ) ≤ J 1 I Q ( X − j ; Y j ∣ X j ) . (69) Moreover, E b c ( Q X ; P ) > 0 E_{\mathrm{bc}}(Q_X;P)>0 E bc ( Q X ; P ) > 0 if and only if I Q ( X − j ; Y j ∣ X j ) > 0 I_Q(X_{-j};Y_j\mid X_j)>0 I Q ( X − j ; Y j ∣ X j ) > 0 .
Proof. The equality (68) is Lemma 7 with m = J − 1 m=J-1 m = J − 1 and T = X j T=X_j T = X j . For each s s s , E r , s ( R , Q X ) ≥ R E_{r,s}(R,Q_X)\ge R E r , s ( R , Q X ) ≥ R if and only if R ( 1 + s ρ ) ≤ E 0 , s ( ρ , Q X ) R(1+s\rho)\le E_{0,s}(\rho,Q_X) R ( 1 + s ρ ) ≤ E 0 , s ( ρ , Q X ) for some ρ ∈ [ 0 , 1 ] \rho\in[0,1] ρ ∈ [ 0 , 1 ] , that is, if and only if R ≤ R s : = max ρ E 0 , s ( ρ , Q X ) / ( 1 + s ρ ) R\le R_s:=\max_\rho E_{0,s}(\rho,Q_X)/(1+s\rho) R ≤ R s := max ρ E 0 , s ( ρ , Q X ) / ( 1 + s ρ ) ; moreover E r , s ( R s , Q X ) = R s E_{r,s}(R_s,Q_X)=R_s E r , s ( R s , Q X ) = R s by continuity, and E r , s ( ⋅ , Q X ) E_{r,s}(\cdot,Q_X) E r , s ( ⋅ , Q X ) is nonincreasing. Hence min { R , min s E r , s ( R , Q X ) } \min\{R,\min_s E_{r,s}(R,Q_X)\} min { R , min s E r , s ( R , Q X )} equals R R R for R ≤ min s R s R\le\min_s R_s R ≤ min s R s and is at most min s R s \min_s R_s min s R s otherwise, which gives the equality in (69) . Concavity gives E 0 , s ( ρ , Q X ) ≤ ρ I s ( Q X ) E_{0,s}(\rho,Q_X)\le\rho I_s(Q_X) E 0 , s ( ρ , Q X ) ≤ ρ I s ( Q X ) , so that R s ≤ I s ( Q X ) / ( 1 + s ) R_s\le I_s(Q_X)/(1+s) R s ≤ I s ( Q X ) / ( 1 + s ) , and s = J − 1 s=J-1 s = J − 1 gives the inequality. Conversely, suppose R t h ( Q X ) > 0 R_{\mathrm{th}}(Q_X)>0 R th ( Q X ) > 0 and fix 0 < R < R t h ( Q X ) 0<R<R_{\mathrm{th}}(Q_X) 0 < R < R th ( Q X ) . Every function ρ ↦ E 0 , s ( ρ , Q X ) − ρ s R \rho\mapsto E_{0,s}(\rho,Q_X)-\rho sR ρ ↦ E 0 , s ( ρ , Q X ) − ρ s R vanishes at ρ = 0 \rho=0 ρ = 0 with slope I s ( Q X ) − s R > 0 I_s(Q_X)-sR>0 I s ( Q X ) − s R > 0 . Hence every E r , s ( R , Q X ) E_{r,s}(R,Q_X) E r , s ( R , Q X ) is positive, and so is the minimum.
This is the role of the symmetric rate in symmetry breaking: it decides whether block coding with random codes breaks the symmetry at an exponential rate, and it caps the resulting exponent, which is even at most J − 1 J R t h ( Q X ) \frac{J-1}{J}R_{\mathrm{th}}(Q_X) J J − 1 R th ( Q X ) . Since I Q ( X − j ; Y j ∣ X j ) ≤ log ∣ Y ∣ I_Q(X_{-j};Y_j\mid X_j)\le\log|\mathcal{Y}| I Q ( X − j ; Y j ∣ X j ) ≤ log ∣ Y ∣ , the random-coding exponent is at most log ∣ Y ∣ / J \log|\mathcal{Y}|/J log ∣ Y ∣/ J ; on the bandit channel of Section VI with two users, two arms of mean one and ϵ = 0 \epsilon=0 ϵ = 0 , both sides equal 1 2 log 2 \frac{1}{2}\log2 2 1 log 2 . The full set dominates the rate threshold, but not necessarily the exponent at the optimized rate: the minimizing s s s in (69) can be smaller than J − 1 J-1 J − 1 ; for three users and three arms of mean one with the uniform law, s = 1 s=1 s = 1 gives 0.176 and s = 2 s=2 s = 2 gives 0.227 (the former is the entry of Table II).
The proof also applies to unequal message rates and independent user-specific input laws. The resulting familiar multiple-access constraints are ∑ k ∈ S R k < I ( X S ; Y j ∣ X S c ) \sum_{k\in S}R_k<I(X_S;Y_j\mid X_{S^c}) ∑ k ∈ S R k < I ( X S ; Y j ∣ X S c ) for every j j j and every nonempty S ⊆ [ J ] ∖ { j } S\subseteq[J]\setminus\{j\} S ⊆ [ J ] ∖ { j } . In the anonymous shared-codebook construction, anonymity, that is, the use of identical strategies, requires every user to employ the same name set and the same codeword ensemble. We therefore specialize to R k = R R_k=R R k = R and a common marginal input law Q X k = Q X Q_{X_k}=Q_X Q X k = Q X for all k k k . This is a restriction of the present random-coding construction, not a claim that every anonymous protocol must have independent or identically distributed inputs; more general exchangeable input processes remain possible.
Limits of Name Decoding
The ranking scheme (15), as implemented in this section, has every user recover all other names. A cut-set argument bounds the name rate of any such scheme, even an adaptive one.
Proposition 4 (Multiway cut-set bound). Suppose each user draws an independent uniform name from a set of size ⌈ e n R ⌉ \lceil e^{nR}\rceil ⌈ e n R ⌉ and, using an arbitrary causal protocol, every user recovers the multiset of all names with an error probability that tends to zero as n → ∞ n\to\infty n → ∞ . Then there is a joint input law P X [ J ] P_{X_{[J]}} P X [ J ] on X J \mathcal{X}^J X J such that, for every nonempty proper subset S ⊂ [ J ] S\subset[J] S ⊂ [ J ] ,
∣ S ∣ R ≤ I ( X S ; Y S c ∣ X S c ) , (70) |S|R\le I(X_S;Y_{S^c}\mid X_{S^c}), \tag*{(70)} ∣ S ∣ R ≤ I ( X S ; Y S c ∣ X S c ) , (70) where Y [ J ] Y_{[J]} Y [ J ] is the output of the channel for the input X [ J ] ∼ P X [ J ] X_{[J]}\sim P_{X_{[J]}} X [ J ] ∼ P X [ J ] . In particular, choosing S = [ J ] ∖ { j } S=[J]\setminus\{j\} S = [ J ] ∖ { j } yields
( J − 1 ) R ≤ I ( X − j ; Y j ∣ X j ) ≤ H ( Y j ) . (71) (J-1)R\le I(X_{-j};Y_j\mid X_j)\le H(Y_j). \tag*{(71)} ( J − 1 ) R ≤ I ( X − j ; Y j ∣ X j ) ≤ H ( Y j ) . (71) Since H ( Y j ) ≤ log ∣ Y ∣ H(Y_j)\le\log|\mathcal{Y}| H ( Y j ) ≤ log ∣ Y ∣ ,
R ≤ log ∣ Y ∣ J − 1 . (72) R\le\frac{\log|\mathcal{Y}|}{J-1}. \tag*{(72)} R ≤ J − 1 log ∣ Y ∣ . (72)
The bound (70) is the standard cut-set bound for memoryless networks [43 ] ; a proof specialized to the present model, with common and private randomness and the decoding of a multiset, is included in Appendix D. For binary outputs, (72) is R ≤ log 2 / ( J − 1 ) R\le\log2/(J-1) R ≤ log 2/ ( J − 1 ) .
Call a scheme name-decoding at rate R R R if every user draws an independent uniform name from a set of size M : = ⌈ e n R ⌉ \mathcal{M}:=\lceil e^{nR}\rceil M := ⌈ e n R ⌉ , forms by an arbitrary causal protocol an estimate of the multiset of the J − 1 J-1 J − 1 other names, and outputs one plus the number of names in this estimate that are smaller than its own, as in (15). Let E c o l \mathcal{E}_{\mathrm{col}} E col be the event that two names coincide, E d e c \mathcal{E}_{\mathrm{dec}} E dec the event that the estimate of some user is wrong, and N n : = E c o l ∪ E d e c \mathcal{N}_n:=\mathcal{E}_{\mathrm{col}}\cup\mathcal{E}_{\mathrm{dec}} N n := E col ∪ E dec . When every estimate is correct, the ranks are a permutation of [ J ] [J] [ J ] if the names are distinct, and two users with the same name obtain the same rank otherwise. Hence the symmetry-breaking error event E \mathcal{E} E satisfies
E c o l ∩ E d e c c ⊆ E ⊆ E c o l ∪ E d e c = N n , (73) \mathcal{E}_{\mathrm{col}}\cap\mathcal{E}_{\mathrm{dec}}^c\subseteq\mathcal{E}\subseteq\mathcal{E}_{\mathrm{col}}\cup\mathcal{E}_{\mathrm{dec}}=\mathcal{N}_n, \tag*{(73)} E col ∩ E dec c ⊆ E ⊆ E col ∪ E dec = N n , (73)
and in particular P b ( n ) ≤ P ( N n ) P_{\mathrm b}^{(n)} \leq\mathbb{P}(\mathcal{N}_n) P b ( n ) ≤ P ( N n ) . We call lim inf n → ∞ − 1 n log P ( N n ) \liminf_{n\to\infty}-\frac{1}{n}\log\mathbb{P}(\mathcal{N}_n) lim inf n → ∞ − n 1 log P ( N n ) the name-decoding exponent of the scheme. It is a lower bound on the error exponent of the scheme, and it is the exponent that an analysis through N n \mathcal{N}_n N n can prove; for block coding, the right-hand side of (18) is an upper bound on P ( N n ) \mathbb{P}(\mathcal{N}_n) P ( N n ) .
Corollary 3 (Ceiling on the name-decoding exponent). For every sequence of name-decoding schemes at rate R R R ,
lim sup n → ∞ − 1 n log P ( N n ) ≤ 1 J − 1 max P X [ J ] I ( X − j ; Y j ∣ X j ) ≤ log ∣ Y ∣ J − 1 . (74) \limsup_{n\to\infty}-\frac{1}{n}\log\mathbb{P}(\mathcal{N}_n) \leq\frac{1}{J-1}\max_{P_{X_{[J]}}} I(X_{-j};Y_j\mid X_j) \leq\frac{\log|\mathcal{Y}|}{J-1}. \tag*{(74)} n → ∞ lim sup − n 1 log P ( N n ) ≤ J − 1 1 P X [ J ] max I ( X − j ; Y j ∣ X j ) ≤ J − 1 log ∣ Y ∣ . (74)
Proof. Two names coincide with probability P c o l ( M , J ) = γ J , M ≥ 1 / M P_{\mathrm{col}}(M,J)=\gamma_{J,M}\geq1/M P col ( M , J ) = γ J , M ≥ 1/ M , so that − 1 n log P ( N n ) ≤ 1 n log ⌈ e n R ⌉ → R -\frac{1}{n}\log\mathbb{P}(\mathcal{N}_n)\leq\frac{1}{n}\log\lceil e^{nR}\rceil\to R − n 1 log P ( N n ) ≤ n 1 log ⌈ e n R ⌉ → R . If the left-hand side of (74) is positive, then P ( N n ) → 0 \mathbb{P}(\mathcal{N}_n)\to0 P ( N n ) → 0 along a subsequence, along which every user recovers the multiset of all names with vanishing error probability; (71), applied to this subsequence, then gives ( J − 1 ) R ≤ max P X [ J ] I ( X − j ; Y j ∣ X j ) (J-1)R\leq\max_{P_{X_{[J]}}} I(X_{-j};Y_j\mid X_j) ( J − 1 ) R ≤ max P X [ J ] I ( X − j ; Y j ∣ X j ) , and the left-hand side is at most R R R .
The two inclusions in (73) can be far apart: the error probability P b ( n ) P_{\mathrm b}^{(n)} P b ( n ) of a name-decoding scheme can be much smaller than P ( N n ) \mathbb{P}(\mathcal{N}_n) P ( N n ) , since two users with the same name need not receive the same rank when their estimates are wrong. However, for channels with conditionally independent outputs, we have
Proposition 5 (Nonadaptive strategies with conditionally independent outputs). Suppose that the outputs are conditionally independent given the inputs,
P ( y [ J ] ∣ x [ J ] ) = ∏ j = 1 J W ( y j ∣ x j , x − j ) . (75) P(y_{[J]}\mid x_{[J]})=\prod_{j=1}^{J}W(y_j\mid x_j,x_{-j}). \tag*{(75)} P ( y [ J ] ∣ x [ J ] ) = j = 1 ∏ J W ( y j ∣ x j , x − j ) . (75) Then every nonadaptive strategy satisfies
P b ( n ) ≥ γ 2 , J P { ∃ j ≠ k : X j n = X k n } . (76) P_{\mathrm b}^{(n)}\geq\gamma_{2,J}\mathbb{P}\{\exists j\ne k:X_j^n=X_k^n\}. \tag*{(76)} P b ( n ) ≥ γ 2 , J P { ∃ j = k : X j n = X k n } . (76)
In particular, every block code with M M M uniform names, with a fixed or random codebook and any terminal map, satisfies P b ( n ) ≥ P c o l ( M , J ) / J ≥ 1 / ( J M ) P_{\mathrm b}^{(n)}\geq P_{\mathrm{col}}(M,J)/J\geq1/(JM) P b ( n ) ≥ P col ( M , J ) / J ≥ 1/ ( J M ) , and, for M = ⌈ e n R ⌉ M=\lceil e^{nR}\rceil M = ⌈ e n R ⌉ , lim sup n → ∞ − 1 n log P b ( n ) ≤ R \limsup_{n\to\infty}-\frac{1}{n}\log P_{\mathrm b}^{(n)}\leq R lim sup n → ∞ − n 1 log P b ( n ) ≤ R .
Proof sketch. Two users that send the same codeword have equal inputs in every slot. Under (75) they then receive i.i.d. outputs, by Lemma 1, and their seeds are i.i.d. given the codeword. Their identifiers are thus i.i.d., and they coincide with probability at least γ 2 , J \gamma_{2,J} γ 2 , J by Lemma 2. Equal names give equal codewords, and two names coincide with probability P c o l ( M , J ) ≥ 1 / M P_{\mathrm{col}}(M,J)\geq1/M P col ( M , J ) ≥ 1/ M , see (17).
The noisy bandit channel of Definition 10 satisfies (75), so that on it the error exponent of a block code at rate R R R is at most R R R , whatever its decoder, and at most the right-hand side of (74) if P ( N n ) → 0 \mathbb{P}(\mathcal{N}_n)\to0 P ( N n ) → 0 , by Proposition 4.
Symmetry breaking itself is not limited in this way. For J = 2 J=2 J = 2 , a channel whose output pair is uniform on { ( 0 , 1 ) , ( 1 , 0 ) } \{(0,1),(1,0)\} {( 0 , 1 ) , ( 1 , 0 )} , whatever the inputs, breaks the symmetry in a single use although it carries no information, even if all users send the same codeword, so that Proposition 5 needs conditional independence.
Remark 3 (Relation to noisy network coding). For the corresponding labeled network, noisy network coding can exploit causal relaying and may yield a stronger interactive inner bound [44 ] , but its standard construction uses node-specific codebooks and compression indices, whereas our users share one encoding rule and decode unordered sets of names. Adapting it to the anonymous setting is left for future work. Also note that Proposition 5 only holds for nonadaptive codes.
Achievability with Interactive Role Controllers
Block coding gives every user a name from a very large space at time zero. Conflicts between names are never detected; they are only made rare, at the price of the collision term e − n R e^{-nR} e − n R , and the names are then ordered by decoding. An alternative approach lets the users contend directly for the identifiers in [ J ] [J] [ J ] and resolves conflicts through the channel. A user keeps its tentative identifier while its observations suggest that nobody else holds it, and draws a new one otherwise; musical chairs is the simplest rule of this kind. This section develops the second approach: Sections V-A and V-B define role controllers and confidence controllers, Section V-C bounds the exponent of finite-state controllers that use certifying outputs, and Section V-D introduces generalized musical chairs.
Role Controllers
Definition 6 (Role controller). A role controller consists of a countable state set T \mathcal{T} T , an initial state t 0 ∈ T t_0\in\mathcal{T} t 0 ∈ T , a law ω \omega ω on a finite set U \mathcal{U} U , maps ϕ i : T × U → X \phi_i:\mathcal{T}\times\mathcal{U}\to\mathcal{X} ϕ i : T × U → X and ψ i : T × U × Y → T \psi_i:\mathcal{T}\times\mathcal{U}\times\mathcal{Y}\to\mathcal{T} ψ i : T × U × Y → T for i ∈ [ n ] i\in[n] i ∈ [ n ] , and a terminal map d n : T → [ J ] d_n:\mathcal{T}\to[J] d n : T → [ J ] . User j j j draws U j [ 1 ] , … , U j [ n ] U_j[1],\ldots,U_j[n] U j [ 1 ] , … , U j [ n ] i.i.d. with law ω \omega ω , independently of the other users; that is, its private seed is S j = ( U j [ 1 ] , … , U j [ n ] ) S_j=(U_j[1],\ldots,U_j[n]) S j = ( U j [ 1 ] , … , U j [ n ]) . It sets T j [ 0 ] = t 0 T_j[0]=t_0 T j [ 0 ] = t 0 and applies
X j [ i ] = ϕ i ( T j [ i − 1 ] , U j [ i ] ) , T j [ i ] = ψ i ( T j [ i − 1 ] , U j [ i ] , Y j [ i ] ) , Z j = d n ( T j [ n ] ) . (77) X_j[i]=\phi_i(T_j[i-1],U_j[i]),\qquad T_j[i]=\psi_i(T_j[i-1],U_j[i],Y_j[i]),\qquad Z_j=d_n(T_j[n]). \tag*{(77)} X j [ i ] = ϕ i ( T j [ i − 1 ] , U j [ i ]) , T j [ i ] = ψ i ( T j [ i − 1 ] , U j [ i ] , Y j [ i ]) , Z j = d n ( T j [ n ]) . (77)
All the elements of the definition may depend on n n n . More generally, every user may draw its initial state from a law π 0 \pi_0 π 0 on T \mathcal{T} T with its private randomness; then e t 0 T \mathrm{e}_{t_0}^{\mathsf{T}} e t 0 T is replaced by π 0 ⊗ J \pi_0^{\otimes J} π 0 ⊗ J below. Since the randomization is private and every user applies the same maps, a role controller is an anonymous symmetric causal strategy in Π n \Pi_n Π n . We call d n ( T j [ i ] ) ∈ [ J ] d_n(T_j[i]) \in[J] d n ( T j [ i ]) ∈ [ J ] the tentative role of user j j j at time i i i : the identifier it would output if the block ended then. The state contains the tentative role together with whatever else the user remembers. The final identifiers collide exactly when two users hold the same role at time n n n . For example, musical chairs, which is treated in Section VI, is a role controller in which a user that has not yet found a free role draws a role uniformly at random in every channel use and keeps it for good once it finds it free; the identifier of a user that is still searching at time n n n is the last role it drew.
Because each U j [ i ] U_j[i] U j [ i ] is used only at time i i i and the channel is memoryless, the joint state T [ i ] = ( T 1 [ i ] , … , T J [ i ] ) T[i]=(T_1[i],\ldots,T_J[i]) T [ i ] = ( T 1 [ i ] , … , T J [ i ]) is a Markov chain with transition matrix
K i ( t , t ′ ) = ∑ u [ J ] ( ∏ j = 1 J ω ( u j ) ) ∑ y [ J ] P ( y [ J ] ∣ ϕ i ( t 1 , u 1 ) , … , ϕ i ( t J , u J ) ) ∏ j = 1 J 1 { ψ i ( t j , u j , y j ) = t j ′ } . (78) K_i(t,t')=\sum_{u^{[J]}}\left(\prod_{j=1}^{J}\omega(u_j)\right)\sum_{y^{[J]}}P\left(y^{[J]}\mid\phi_i(t_1,u_1),\ldots,\phi_i(t_J,u_J)\right)\prod_{j=1}^{J}\mathbf{1}\{\psi_i(t_j,u_j,y_j)=t'_j\}. \tag*{(78)} K i ( t , t ′ ) = u [ J ] ∑ ( j = 1 ∏ J ω ( u j ) ) y [ J ] ∑ P ( y [ J ] ∣ ϕ i ( t 1 , u 1 ) , … , ϕ i ( t J , u J ) ) j = 1 ∏ J 1 { ψ i ( t j , u j , y j ) = t j ′ } . (78)
Let
F : = { t ∈ T J : d n ( t j ) = d n ( t k ) for some j ≠ k } (79) \mathcal{F}:=\{t\in\mathcal{T}^{J}:d_n(t_j)=d_n(t_k)\text{ for some }j\ne k\} \tag*{(79)} F := { t ∈ T J : d n ( t j ) = d n ( t k ) for some j = k } (79)
be the set of failure states.
Proposition 6 (Exact error probability). For every role controller,
P b ( n ) = e t 0 T K 1 K 2 ⋯ K n 1 F , (80) P_{\mathrm{b}}^{(n)}=\mathrm{e}_{t_0}^{\mathsf{T}}K_1K_2\cdots K_n\mathbf{1}_{\mathcal{F}}, \tag*{(80)} P b ( n ) = e t 0 T K 1 K 2 ⋯ K n 1 F , (80) where t 0 = ( t 0 , … , t 0 ) t_0=(t_0,\ldots,t_0) t 0 = ( t 0 , … , t 0 ) and 1 F \mathbf{1}_{\mathcal{F}} 1 F is the indicator vector of F \mathcal{F} F .
Proposition 7 (Symmetric causal strategies are role controllers). For every strategy π n ∈ Π n \pi_n\in\Pi_n π n ∈ Π n and every value c c c of the common randomness, there is a role controller whose inputs, outputs and identifiers have the same joint law as those of π n \pi_n π n given C = c C=c C = c . In particular, the infimum of P b ( n ) P_{\mathrm{b}}^{(n)} P b ( n ) over role controllers is ϵ n ∗ ( P ) \epsilon_n^*(P) ϵ n ∗ ( P ) , and every block code of Section IV is a role controller, so that, for every Q X Q_X Q X , role controllers achieve the exponent E b c ( Q X ; P ) E_{\mathrm{bc}}(Q_X;P) E bc ( Q X ; P ) of (63), and this exponent is positive whenever I Q ( X − j ; Y j ∣ X j ) > 0 I_Q(X_{-j};Y_j\mid X_j)>0 I Q ( X − j ; Y j ∣ X j ) > 0 .
Proof. Given C = c C=c C = c , a seed s s s enters only through the maps f 1 ( c , s , ⋅ ) , … , f n ( c , s , ⋅ ) f_1(c,s,\cdot),\ldots,f_n(c,s,\cdot) f 1 ( c , s , ⋅ ) , … , f n ( c , s , ⋅ ) and g n ( c , s , ⋅ ) g_n(c,s,\cdot) g n ( c , s , ⋅ ) , which range over a finite set since the alphabets are finite, so that we may assume that the seed takes values in a finite set S S S . Take U = S U=S U = S with ω = P S \omega=P_S ω = P S , of which only U [ 1 ] U[1] U [ 1 ] is used, as the seed of the user, and T = { t 0 } ∪ ⋃ i < n ( S × Y i ) \mathcal{T}=\{t_0\}\cup\bigcup_{i<n}(S\times\mathcal{Y}^i) T = { t 0 } ∪ ⋃ i < n ( S × Y i ) . Put ϕ 1 ( t 0 , u ) = f 1 ( c , u ) \phi_1(t_0,u)=f_1(c,u) ϕ 1 ( t 0 , u ) = f 1 ( c , u ) and ψ 1 ( t 0 , u , y ) = ( u , y ) \psi_1(t_0,u,y)=(u,y) ψ 1 ( t 0 , u , y ) = ( u , y ) ; for i ≥ 2 i\ge2 i ≥ 2 , put ϕ i ( ( s , y i − 1 ) , u ) = f i ( c , s , y i − 1 ) \phi_i((s,y^{i-1}),u)=f_i(c,s,y^{i-1}) ϕ i (( s , y i − 1 ) , u ) = f i ( c , s , y i − 1 ) and ψ i ( ( s , y i − 1 ) , u , y ) = ( s , ( y i − 1 , y ) ) \psi_i((s,y^{i-1}),u,y)=(s,(y^{i-1},y)) ψ i (( s , y i − 1 ) , u , y ) = ( s , ( y i − 1 , y )) ; and let d n ( s , y n ) = g n ( c , s , y n ) d_n(s,y^n)=g_n(c,s,y^n) d n ( s , y n ) = g n ( c , s , y n ) . The joint law is then that of π n \pi_n π n given C = c C=c C = c , and since P b ( n ) ( π n ; P ) P_{\mathrm{b}}^{(n)}(\pi_n;P) P b ( n ) ( π n ; P ) averages over C C C , some c c c does at least as well as π n \pi_n π n . A block code is the strategy whose seed is the name m m m , with f i ( c , m , y i − 1 ) = x [ i ] ( m ) f_i(c,m,y^{i-1})=x[i](m) f i ( c , m , y i − 1 ) = x [ i ] ( m ) and the rank rule (15) as terminal map. The exponent is Theorem 2, which provides a deterministic common codebook, and its positivity is Proposition 3.
The state of this role controller contains the whole output sequence, and by Proposition 8 below, a finite-state controller cannot match it when every output has positive probability. The mutual-information condition in Proposition 7 is sufficient but not necessary for a positive exponent: on the channel of Section IV-C that carries no information, a role controller that outputs one plus its first output bit as its identifier has zero error.
Definition 7 (Stationary and finite-state controllers). A role controller is stationary if ϕ i = ϕ \phi_i=\phi ϕ i = ϕ and ψ i = ψ \psi_i=\psi ψ i = ψ for all i ∈ [ n ] i\in[n] i ∈ [ n ] . Its transition matrix is then a single matrix K K K , which may depend on n n n , and (80) reads P b ( n ) = e t 0 T K n 1 F P_{\mathrm{b}}^{(n)}=\mathrm{e}_{t_0}^{\mathsf{T}}K^n\mathbf{1}_{\mathcal{F}} P b ( n ) = e t 0 T K n 1 F . A finite-state controller is a stationary controller with a finite state set T \mathcal{T} T such that T \mathcal{T} T , t 0 t_0 t 0 (or the initial law π 0 \pi_0 π 0 ), ω \omega ω , ϕ \phi ϕ , ψ \psi ψ and d n = d d_n=d d n = d do not depend on n n n : the same controller, with the same matrix K K K , is run at every blocklength.
Remark 4 (Time in the state). Stationarity alone does not restrict the class: any role controller becomes stationary once the time index is stored in its state, with T × { 0 , … , n } \mathcal{T}\times\{0,\ldots,n\} T × { 0 , … , n } in place of T \mathcal{T} T . In block coding, for instance, the clock is already in the state ( m , y i ) (m,y^i) ( m , y i ) , since i i i is the length of y i y^i y i . What a finite-state controller excludes is this clock: a finite state set that does not depend on n n n cannot count the channel uses. For the same reason, the finite-memory theory of hypothesis testing and of the two-armed bandit is restricted to time-invariant machines [35 , 36 ] : with maps that vary with time, the time index serves as a free clock, and four states then suffice for an error probability that tends to zero in hypothesis testing [34 ] , as 4 J 4J 4 J states do on the noisy bandit channel (Appendix E); see [45 ] for a recent survey. The memory log ∣ T ∣ |\mathcal{T}| ∣ T ∣ of a user grows linearly in n n n for block coding, logarithmically for confidence controllers with thresholds of order n n n (§ 5 -B), and stays bounded for generalized musical chairs (§ 5 -D).
If every output has positive probability, or more generally if the possible outputs of a user do not depend on what the others play or observe, no finite-state controller has a vanishing error probability (Proposition 8). To drive the error probability to zero, a finite-state controller needs an output that breaks this condition, such as a certificate, an observation that is impossible when the role is shared (§ 5 -C).
Proposition 8 (Floor for finite-state controllers). Let the controller be a finite-state controller, let X 0 \mathcal{X}_0 X 0 be the set of inputs that it plays with positive probability, X 0 = { ϕ ( t , u ) : t ∈ T , ω ( u ) > 0 } \mathcal{X}_0 = \{\phi(t,u): t \in\mathcal{T},\ \omega(u) > 0\} X 0 = { ϕ ( t , u ) : t ∈ T , ω ( u ) > 0 } , and suppose that there are sets S ( x ) ⊆ Y S(x) \subseteq\mathcal{Y} S ( x ) ⊆ Y , x ∈ X 0 x \in\mathcal{X}_0 x ∈ X 0 , such that for all x [ J ] ∈ X 0 J x_{[J]} \in\mathcal{X}_0^J x [ J ] ∈ X 0 J , P ( y [ J ] ∣ x [ J ] ) > 0 P(y_{[J]} \mid x_{[J]}) > 0 P ( y [ J ] ∣ x [ J ] ) > 0 if and only if y j ∈ S ( x j ) y_j \in S(x_j) y j ∈ S ( x j ) for every j j j ; this holds in particular if P ( y [ J ] ∣ x [ J ] ) > 0 P(y_{[J]} \mid x_{[J]}) > 0 P ( y [ J ] ∣ x [ J ] ) > 0 for all x [ J ] ∈ X 0 J x_{[J]} \in\mathcal{X}_0^J x [ J ] ∈ X 0 J and all y [ J ] y_{[J]} y [ J ] . Then the controller satisfies lim inf n → ∞ P b ( n ) > 0 \liminf_{n \to\infty} P_{\mathrm{b}}^{(n)} > 0 lim inf n → ∞ P b ( n ) > 0 .
Proof. If the users draw their initial state from a law π 0 \pi_0 π 0 , they all start at a common state t 0 t_0 t 0 with π 0 ( t 0 ) > 0 \pi_0(t_0) > 0 π 0 ( t 0 ) > 0 with probability π 0 ( t 0 ) J > 0 \pi_0(t_0)^J > 0 π 0 ( t 0 ) J > 0 , and we argue from there. Call t ′ t' t ′ a successor of t t t if t ′ = ψ ( t , u , y ) t' = \psi(t,u,y) t ′ = ψ ( t , u , y ) with ω ( u ) > 0 \omega(u) > 0 ω ( u ) > 0 and y ∈ S ( ϕ ( t , u ) ) y \in S(\phi(t,u)) y ∈ S ( ϕ ( t , u )) . Since the seeds are independent and the supports have product form, a joint move has positive probability if and only if every user makes a successor step; hence the joint states reachable in k k k steps are the products of the local states reachable in k k k steps, and from ( t , … , t ) (t,\ldots,t) ( t , … , t ) all users make the same move with probability at least some η > 0 \eta> 0 η > 0 , which leads to a failure state. Let a a a be reachable from t 0 t_0 t 0 with the smallest set of reachable states, counting a a a itself, which is reachable from itself in zero steps; by minimality, one can show that a a a is reachable from every state reachable from a a a . If ( b 1 , … , b J ) (b_1,\ldots,b_J) ( b 1 , … , b J ) is reachable from ( a , … , a ) (a,\ldots,a) ( a , … , a ) in k k k steps and b i b_i b i reaches a a a in ℓ i \ell_i ℓ i steps, then every b j b_j b j reaches a a a in exactly L = ∑ i ℓ i + ( J − 1 ) k L = \sum_i \ell_i + (J-1)k L = ∑ i ℓ i + ( J − 1 ) k steps, through the loops a → b i → a a \to b_i \to a a → b i → a , i ≠ j i \ne j i = j . Hence, from every joint state reachable from ( a , … , a ) (a,\ldots,a) ( a , … , a ) , the users return to ( a , … , a ) (a,\ldots,a) ( a , … , a ) within some L ⋆ L^\star L ⋆ steps with probability at least some p > 0 p > 0 p > 0 .
Since a a a is reachable from t 0 t_0 t 0 , say in k 0 k_0 k 0 steps, the users are at ( a , … , a ) (a,\ldots,a) ( a , … , a ) at time k 0 k_0 k 0 with probability at least some q > 0 q > 0 q > 0 , by following this path together. Let n ≥ k 0 + L ⋆ n \ge k_0 + L^\star n ≥ k 0 + L ⋆ . On this event, the joint state at time n − L ⋆ ≥ k 0 n - L^\star\ge k_0 n − L ⋆ ≥ k 0 is reachable from ( a , … , a ) (a,\ldots,a) ( a , … , a ) . From there, the users return to ( a , … , a ) (a,\ldots,a) ( a , … , a ) within L ⋆ L^\star L ⋆ steps with probability at least p p p , and then make identical moves until time n n n , at most L ⋆ L^\star L ⋆ of them, so that the joint state at time n n n is a failure state with probability at least η L ⋆ \eta^{L^\star} η L ⋆ . Hence P b ( n ) ≥ q p η L ⋆ P_{\mathrm{b}}^{(n)} \ge qp\eta^{L^\star} P b ( n ) ≥ q p η L ⋆ for all n ≥ k 0 + L ⋆ n \ge k_0 + L^\star n ≥ k 0 + L ⋆ . □
Proposition 8 uses the finiteness of T \mathcal{T} T , stationarity, and the independence from n n n . Without stationarity, a state set of bounded size can suffice for a vanishing error probability: Appendix E gives such a controller with 4 J 4J 4 J states on the noisy bandit channel, by a device of Cover [34 ] , but its error probability decays only subexponentially. Whether such a controller can achieve a positive exponent when every output has positive probability is open.
Confidence Controllers
By Proposition 8, on a channel on which every output has positive probability, a controller with a vanishing error probability is not a finite-state controller: its state set is infinite, its maps vary with time, or it depends on n n n . Block coding stores the whole output sequence. A more economical summary of the observations is a user’s confidence in its current role: a counter that goes up when the observations suggest that the role is held alone, and down when they suggest that it is shared.
Definition 8 (Confidence controller). A confidence controller is specified by the following data.
For every role a ∈ [ J ] a \in[J] a ∈ [ J ] : an input law Q a Q_a Q a on X \mathcal{X} X for a user that is still searching, an input law Q a ⋆ Q_a^\star Q a ⋆ on X \mathcal{X} X for a user that has committed (by default Q a ⋆ = Q a Q_a^\star= Q_a Q a ⋆ = Q a ), a score s a ( x , y ) ∈ Z ∪ { ± ∞ } s_a(x,y) \in\mathbb{Z} \cup\{\pm\infty\} s a ( x , y ) ∈ Z ∪ { ± ∞ } of an input-output pair, and thresholds ξ − a ∈ { 0 , 1 , 2 , … } ∪ { ∞ } \xi_-^a \in\{0,1,2,\ldots\} \cup\{\infty\} ξ − a ∈ { 0 , 1 , 2 , … } ∪ { ∞ } and ξ + a ∈ { 1 , 2 , … } \xi_+^a \in\{1,2,\ldots\} ξ + a ∈ { 1 , 2 , … } .
A redraw kernel R ( a , ⋅ ) R(a,\cdot) R ( a , ⋅ ) on [ J ] [J] [ J ] for every a ∈ [ J ] a \in[J] a ∈ [ J ] , and an initial law π 0 \pi_0 π 0 on [ J ] [J] [ J ] .
The state of a user is its role a ∈ [ J ] a \in[J] a ∈ [ J ] together with an integer counter ξ ∈ { − ξ − a + 1 , … , ξ + a } \xi\in\{-\xi_-^a+1,\ldots,\xi_+^a\} ξ ∈ { − ξ − a + 1 , … , ξ + a } ; initially a ∼ π 0 a \sim\pi_0 a ∼ π 0 and ξ = 0 \xi= 0 ξ = 0 . In every channel use, a user in state ( a , ξ ) (a,\xi) ( a , ξ ) proceeds as follows, all draws being made with fresh private randomness.
It draws X ∼ Q a X \sim Q_a X ∼ Q a if it is not committed and X ∼ Q a ⋆ X \sim Q_a^\star X ∼ Q a ⋆ if it is, transmits X X X and observes Y Y Y .
If ξ < ξ + a \xi< \xi_+^a ξ < ξ + a , it updates ξ ← min { ξ + s a ( X , Y ) , ξ + a } \xi\gets\min\{\xi+s_a(X,Y),\xi_+^a\} ξ ← min { ξ + s a ( X , Y ) , ξ + a } , with ξ + ∞ : = ∞ \xi_+^\infty:= \infty ξ + ∞ := ∞ and ξ − ∞ : = − ∞ \xi_-^\infty:= -\infty ξ − ∞ := − ∞ . A user with ξ = ξ + a \xi= \xi_+^a ξ = ξ + a is committed : it keeps its state until time n n n and transmits with the law Q a ⋆ Q_a^\star Q a ⋆ .
If ξ ≤ − ξ − a \xi\le-\xi_-^a ξ ≤ − ξ − a , it releases its role: it draws a new role a ′ ∼ R ( a , ⋅ ) a' \sim R(a,\cdot) a ′ ∼ R ( a , ⋅ ) and resets ξ = 0 \xi= 0 ξ = 0 .
After n n n channel uses, the user outputs its role a a a as its identifier.
Since every draw uses fresh private randomness, a confidence controller is a role controller in the sense of Definition 6, with state ( a , ξ ) (a,\xi) ( a , ξ ) and tentative role d n ( a , ξ ) = a d_n(a,\xi) = a d n ( a , ξ ) = a , and with the initial role drawn from π 0 \pi_0 π 0 , as allowed after Definition 6. Let N ( ⋅ ∣ ( a , ξ ) , x , y ) N(\cdot\mid(a,\xi),x,y) N ( ⋅ ∣ ( a , ξ ) , x , y ) be the law of the next state of a user in state ( a , ξ ) (a,\xi) ( a , ξ ) that transmitted x x x and observed y y y . If ξ = ξ + a \xi= \xi_+^a ξ = ξ + a , the state does not change. Otherwise, put ξ + : = min { ξ + s a ( x , y ) , ξ + a } \xi^+ := \min\{\xi+s_a(x,y),\xi_+^a\} ξ + := min { ξ + s a ( x , y ) , ξ + a } ; the next state is ( a , ξ + ) (a,\xi^+) ( a , ξ + ) with probability R ( a , a ′ ) R(a,a') R ( a , a ′ ) if ξ + ≤ − ξ − a \xi^+ \le-\xi_-^a ξ + ≤ − ξ − a . A user in state ( a , ξ ) (a,\xi) ( a , ξ ) transmits with the law Q a ⟨ ξ ⟩ Q_a^{\langle\xi\rangle} Q a ⟨ ξ ⟩ , where Q a ⟨ ξ ⟩ : = Q a Q_a^{\langle\xi\rangle} := Q_a Q a ⟨ ξ ⟩ := Q a if ξ < ξ + a \xi< \xi_+^a ξ < ξ + a and Q a ⟨ ξ ⟩ : = Q a ⋆ Q_a^{\langle\xi\rangle} := Q_a^\star Q a ⟨ ξ ⟩ := Q a ⋆ if ξ = ξ + a \xi= \xi_+^a ξ = ξ + a .
Proposition 9 (Exact error probability of a confidence controller). The joint state t = ( t 1 , … , t J ) t = (t_1,\ldots,t_J) t = ( t 1 , … , t J ) , t j = ( a j , ξ j ) t_j = (a_j,\xi_j) t j = ( a j , ξ j ) , of the J J J users is a Markov chain with initial law μ 0 = ( π 0 ⊗ δ 0 ) ⊗ J \mu_0 = (\pi_0 \otimes\delta_0)^{\otimes J} μ 0 = ( π 0 ⊗ δ 0 ) ⊗ J and transition matrix
K ( t , t ′ ) = ∑ x [ J ] , y [ J ] ∏ j = 1 J Q a j ⟨ ξ j ⟩ ( x j ) P ( y [ J ] ∣ x [ J ] ) ∏ j = 1 J N ( t j ′ ∣ t j , x j , y j ) . (81) K(t,t') = \sum_{x_{[J]},y_{[J]}} \prod_{j=1}^{J} Q_{a_j}^{\langle\xi_j\rangle}(x_j)\,P(y_{[J]} \mid x_{[J]}) \prod_{j=1}^{J} N(t_j' \mid t_j,x_j,y_j). \tag*{(81)} K ( t , t ′ ) = x [ J ] , y [ J ] ∑ j = 1 ∏ J Q a j ⟨ ξ j ⟩ ( x j ) P ( y [ J ] ∣ x [ J ] ) j = 1 ∏ J N ( t j ′ ∣ t j , x j , y j ) . (81)
With F \mathcal{F} F the set of joint states in which two users hold the same role,
P b ( n ) = μ 0 K n 1 F . (82) P_{\mathrm{b}}^{(n)}=\mu_0K^n\mathbf{1}_{\mathcal{F}}. \tag*{(82)} P b ( n ) = μ 0 K n 1 F . (82)
Proof. This is Proposition 6 for the stationary role controller with the states ( a , ξ ) (a,\xi) ( a , ξ ) , with the initial state drawn from π 0 ⊗ δ 0 \pi_0\otimes\delta_0 π 0 ⊗ δ 0 .
For J = 2 J=2 J = 2 and thresholds of order n n n , the chain has O ( n 2 ) O(n^2) O ( n 2 ) states, and the transitions of the counters are sparse, so (82) can be evaluated exactly for moderate n n n . Its transitions are simple: the counters move by the scores, and a release redraws a role. The state of a user takes O ( log n ) O(\log n) O ( log n ) bits whenever the thresholds are O ( n ) O(n) O ( n ) .
A score + ∞ +\infty + ∞ commits the user at once. It is meant for outputs that cannot occur when the role is shared, that is, for a certificate; combined with finite scores on the other outputs, it lets a controller use a certificate and soft evidence at the same time. The score is meant to be positive on average when the role is held alone and negative on average when it is shared, so that the counter of a user in a shared role drifts down until it releases, and the counter of a user in an unshared role drifts up until it commits. A typical score is a quantized log-likelihood ratio between the two situations. The input law Q a Q_a Q a makes the role informative: a user in role a a a must send something whose effect on its own output depends on whether another user holds the same role.
The class contains the following schemes.
Uniform confidence controllers. Fixed inputs Q a = δ χ a Q_a=\delta_{\chi_a} Q a = δ χ a with χ a ∈ X \chi_a\in\mathcal{X} χ a ∈ X , the same parameters for every role, the uniform kernel R ( a , ⋅ ) = Unif ( [ J ] ) R(a,\cdot)=\operatorname{Unif}([J]) R ( a , ⋅ ) = Unif ([ J ]) and π 0 = Unif ( [ J ] ) \pi_0=\operatorname{Unif}([J]) π 0 = Unif ([ J ]) , with large thresholds; Section VI evaluates such controllers, together with optimized ones that use other kernels and initial laws.
Musical chairs and its one-arm variant are also confidence controllers (Section VI).
The thresholds determine the behavior.
With ξ + a \xi_+^a ξ + a finite and independent of n n n , this is a finite-state controller, subject to Proposition 8. If some output certifies that a role is not shared, a finite-state controller can nevertheless work, and Proposition 10 gives an achievable exponent, namely that of the survival probability; musical chairs is of this kind.
With ξ + a \xi_+^a ξ + a growing linearly in n n n , the controller is stationary but not finite-state: the counter needs only O ( log n ) O(\log n) O ( log n ) bits and is not subject to the floor; Section VI evaluates such controllers on the noisy bandit channel.
With ξ + a = ∞ \xi_+^a=\infty ξ + a = ∞ , the controller is stationary and does not depend on n n n , but its state set is infinite, so that Proposition 8 does not apply either.
Certified Completion
Some channels have certifying outputs, which are impossible when the role is shared. A user that receives one knows that it holds its role alone and can keep it for good. A finite-state controller then completes with an error exponent of at least − log -\log − log of a spectral radius.
Proposition 10 (Certified completion). Let the controller be a finite-state controller; with an initial law π 0 \pi_0 π 0 , read π 0 ⊗ J \pi_0^{\otimes J} π 0 ⊗ J for E t 0 ⊗ J \mathsf{E}_{t_0}^{\otimes J} E t 0 ⊗ J and the support of π 0 ⊗ J \pi_0^{\otimes J} π 0 ⊗ J for t 0 t_0 t 0 below. Let G ⊆ T J ∖ F \mathcal{G}\subseteq\mathcal{T}^J\setminus\mathcal{F} G ⊆ T J ∖ F be absorbing, that is, K ( t , G ) = 1 K(t,\mathcal{G})=1 K ( t , G ) = 1 for t ∈ G t\in\mathcal{G} t ∈ G , with t 0 ∉ G t_0\notin\mathcal{G} t 0 ∈ / G . Let B \mathcal{B} B be the set of states reachable from t 0 t_0 t 0 that are not in G \mathcal{G} G , and let K B K_{\mathcal{B}} K B be the corresponding substochastic submatrix. Then
P b ( n ) ≤ e t 0 T K B n 1. (83) P_{\mathrm{b}}^{(n)}\leq e_{t_0}^{\mathsf{T}}K_{\mathcal{B}}^n\mathbf{1}. \tag*{(83)} P b ( n ) ≤ e t 0 T K B n 1 . (83) If 0 < ϱ : = ρ ( K B ) < 1 0<\varrho:=\rho(K_{\mathcal{B}})<1 0 < ϱ := ρ ( K B ) < 1 , where ρ \rho ρ denotes the spectral radius, the right-hand side decays with exponent exactly − log ϱ -\log\varrho − log ϱ :
lim n → ∞ − 1 n log ( e t 0 T K B n 1 ) = − log ϱ . (84) \lim_{n\to\infty}-\frac{1}{n}\log\left(e_{t_0}^{\mathsf{T}}K_{\mathcal{B}}^n\mathbf{1}\right)=-\log\varrho. \tag*{(84)} n → ∞ lim − n 1 log ( e t 0 T K B n 1 ) = − log ϱ . (84) If, moreover, a strongly connected class of B \mathcal{B} B (a maximal set of joint states that all reach one another inside B \mathcal{B} B ) whose submatrix has spectral radius ϱ \varrho ϱ contains a failure state in each of its cyclic classes, then P b ( n ) ≥ c ϱ n P_{\mathrm{b}}^{(n)}\geq c\varrho^n P b ( n ) ≥ c ϱ n for some c > 0 c>0 c > 0 and all large n n n , so that P b ( n ) P_{\mathrm{b}}^{(n)} P b ( n ) itself decays with exponent exactly − log ϱ -\log\varrho − log ϱ .
Proof. Since G \mathcal{G} G is absorbing and disjoint from F \mathcal{F} F , failure at time n n n implies that G \mathcal{G} G has not been entered, whose probability is the right-hand side of (83). Write s n s_n s n for this right-hand side. Since s n ≤ ∥ K B n ∥ ∞ s_n\leq\lVert K_{\mathcal{B}}^n\rVert_\infty s n ≤ ∥ K B n ∥ ∞ , the largest row sum of K B n K_{\mathcal{B}}^n K B n , and ∥ K B n ∥ ∞ 1 / n → ϱ \lVert K_{\mathcal{B}}^n\rVert_\infty^{1/n}\to\varrho ∥ K B n ∥ ∞ 1/ n → ϱ by Gelfand’s formula, lim inf n → ∞ − 1 n log s n ≥ − log ϱ \liminf_{n\to\infty}-\frac{1}{n}\log s_n\geq-\log\varrho lim inf n → ∞ − n 1 log s n ≥ − log ϱ . More precisely, the Jordan form of K B K_{\mathcal{B}} K B gives s n ≤ C n k ϱ n s_n\leq Cn^k\varrho^n s n ≤ C n k ϱ n for a constant C C C , where k + 1 k+1 k + 1 is the size of the largest Jordan block of an eigenvalue of modulus ϱ \varrho ϱ , so that k ≤ ∣ B ∣ − 1 k\leq|\mathcal{B}|-1 k ≤ ∣ B ∣ − 1 . For the lower bound, a strongly connected class of B \mathcal{B} B attaining ϱ \varrho ϱ is reachable from t 0 t_0 t 0 by definition of B \mathcal{B} B . Following a fixed path to it and then using a positive Perron eigenvector of its irreducible submatrix bounds s n s_n s n below by a positive constant times ϱ n \varrho^n ϱ n , without an aperiodicity assumption. For the last claim, let B 0 ⊆ B \mathcal{B}_0\subseteq\mathcal{B} B 0 ⊆ B be such a class, of period d d d , entered at a state e e e with positive probability. By the Perron–Frobenius theorem for irreducible matrices, for every f ∈ B 0 f\in\mathcal{B}_0 f ∈ B 0 , ( K B 0 k ) e f (K_{\mathcal{B}_0}^k)_{ef} ( K B 0 k ) e f is at least a positive constant times ϱ k \varrho^k ϱ k for all k k k in the residue class modulo d d d of the cyclic class of f f f . Taking f f f among the failure states of the d d d cyclic classes shows that the probability of a failure state at time n n n is at least c ϱ n c\varrho^n c ϱ n for all large $n. □ \square □
When ϱ < 1 \varrho<1 ϱ < 1 , the error probability vanishes, so that, by Proposition 8, the possible outputs of a user must depend on what the others play or observe, as they do with certificates. Generalized musical chairs, below, is a role controller of this kind.
Generalized Musical Chairs
Musical chairs redraws its role uniformly after a failure, although a failure may mean that the role is held by a committed user. Generalized musical chairs lets a user remember the roles on which it failed and avoid them. It applies when every role a ∈ [ J ] a \in[J] a ∈ [ J ] has an input χ a ∈ X \chi_a \in\mathcal{X} χ a ∈ X , the inputs being distinct, and a set C a ⊆ Y \mathcal{C}_a \subseteq\mathcal{Y} C a ⊆ Y of certifying outputs, which have probability zero whenever another user also plays χ a \chi_a χ a .
Definition 9 (Generalized musical chairs). Fix an exclusion probability β ∈ [ 0 , 1 ] \beta\in[0,1] β ∈ [ 0 , 1 ] . A user is either committed to a role or searching with an excluded set E ⊆ [ J ] E \subseteq[J] E ⊆ [ J ] , initially E = ∅ E = \varnothing E = ∅ . In every channel use, a searching user proceeds as follows.
If E = [ J ] E = [J] E = [ J ] , it sets E = ∅ E = \varnothing E = ∅ .
It draws a role a a a uniformly in [ J ] ∖ E [J] \setminus E [ J ] ∖ E , transmits χ a \chi_a χ a and observes y y y .
If y ∈ C a y \in\mathcal{C}_a y ∈ C a , it commits to a a a . Otherwise, with probability β \beta β , it sets E ← E ∪ { a } E \gets E \cup\{a\} E ← E ∪ { a } .
A committed user transmits χ a \chi_a χ a until time n n n . The identifier of a user is its committed role, or, if it is still searching, the last role it drew (the bound below counts every such case as a failure).
For β = 0 \beta= 0 β = 0 the excluded set stays empty, and this is musical chairs: a user commits at its first certificate and otherwise redraws a role uniformly. For β > 0 \beta> 0 β > 0 a searching user avoids the roles on which it recently failed. A failure is not conclusive: the role may be held by a committed user, but it may also be a collision with another searching user or, when the certifying outputs are not certain under a lone user, a missed reward. The exclusion is therefore a heuristic, and β \beta β sets how much to trust it; the reset of E E E when it covers [ J ] [J] [ J ] prevents a permanent lock-out, even for β = 1 \beta= 1 β = 1 . The exponent is discontinuous at β = 0 \beta= 0 β = 0 , as Section VI shows.
Corollary 4 (Exponent of generalized musical chairs). Generalized musical chairs is a finite-state controller. Let G \mathcal{G} G be the set of joint states in which all users are committed to distinct roles, and K B K_B K B the substochastic matrix of Proposition 10. Then
P b ( n ) ≤ e t 0 K B n 1 , P_b^{(n)} \le\mathbf{e}_{t_0} K_B^n \mathbf{1}, P b ( n ) ≤ e t 0 K B n 1 , and the error exponent is at least − log ρ ( K B ) -\log\rho(K_B) − log ρ ( K B ) , with equality under the last condition of Proposition 10. Maximized over β \beta β , this bound is at least the exponent of musical chairs, which is the case β = 0 \beta= 0 β = 0 .
Proof. The state of a searching user is its excluded set together with the last role it drew, which determines its identifier, and that of a committed user is its role; this state set is finite, the maps do not depend on n n n , and the randomization is private. A user commits at a channel use only if it received a certifying output, which means that no other user played the same input at that use. Committed users keep playing their input, so a later user cannot obtain a certificate for their roles. The committed roles are therefore distinct at all times, so every reachable joint state in which all users are committed lies in G \mathcal{G} G . The set G \mathcal{G} G is disjoint from the failure states, and it is absorbing because committed users do not change. Proposition 10 gives the claims. □
The exponent is positive, that is, ρ ( K B ) < 1 \rho(K_B) < 1 ρ ( K B ) < 1 , when, for every a ∈ [ J ] a \in[J] a ∈ [ J ] , a user alone on χ a \chi_a χ a receives an output in C a \mathcal{C}_a C a with positive probability, whatever the others play. Indeed, since commitments are permanent, it suffices that some searching user j j j commits within J + 1 J + 1 J + 1 uses with positive probability. Failing and excluding the roles it draws, j j j empties its excluded set within J J J uses unless it commits. It then draws, with positive probability, a role that no committed user holds and that no other searching user is forced to draw (one exists, since the other searching users are fewer than such roles), while the others draw other roles, and it receives a certificate with positive probability. Hence from every reachable joint state in which some user is still searching, all users are committed within some m m m steps with probability at least some p > 0 p > 0 p > 0 . This implies that every row sum of K B m K_B^m K B m is at most 1 − p 1-p 1 − p , and hence ρ ( K B ) ≤ ( 1 − p ) 1 / m < 1 \rho(K_B) \le(1-p)^{1/m} < 1 ρ ( K B ) ≤ ( 1 − p ) 1/ m < 1 .
The survival probability does not depend on the last roles drawn and is invariant under permutations of the users, so that the chain that computes it has one state per assignment of committed roles and excluded sets, up to a permutation of the users, and the spectral radius is computed numerically in Section VI. With β > 0 \beta> 0 β > 0 , generalized musical chairs is not a confidence controller, since the excluded set is needed, and it requires certifying outputs (Proposition 8).
Application: The Noisy Bandit Channel
We now apply the general results to the channel that motivated the work: J J J players choose among K K K arms and observe a Bernoulli reward, but a collision suppresses the reward of the players involved. We define a noisy bandit channel with one pair of parameters per arm, so that the channel of [9 ] is the noiseless case.
The Channel
Definition 10 (Noisy bandit channel). Let X = [ K ] \mathcal{X} = [K] X = [ K ] and Y = { 0 , 1 } \mathcal{Y} = \{0,1\} Y = { 0 , 1 } ; the inputs are the arms. Each arm has a solo reward probability μ x ∈ [ 0 , 1 ] \mu_x \in[0,1] μ x ∈ [ 0 , 1 ] and a collision reward probability ε x ∈ [ 0 , 1 ] \varepsilon_x \in[0,1] ε x ∈ [ 0 , 1 ] . Given the inputs x [ J ] x_{[J]} x [ J ] , the outputs are independent, and user j j j observes Y j = 1 Y_j = 1 Y j = 1 with probability
W ( 1 ∣ x j , x − j ) = { μ x j if x k ≠ x j for all k ≠ j , ε x j if x k = x j for some k ≠ j , W ( 0 ∣ ⋅ ) = 1 − W ( 1 ∣ ⋅ ) . (85) W(1 \mid x_j,x_{-j}) =
\begin{cases}
\mu_{x_j} & \text{if } x_k \ne x_j \text{ for all } k \ne j,\\
\varepsilon_{x_j} & \text{if } x_k = x_j \text{ for some } k \ne j,
\end{cases}
\qquad
W(0 \mid\cdot) = 1 - W(1 \mid\cdot). \tag*{(85)} W ( 1 ∣ x j , x − j ) = { μ x j ε x j if x k = x j for all k = j , if x k = x j for some k = j , W ( 0 ∣ ⋅ ) = 1 − W ( 1 ∣ ⋅ ) . (85)
A user that is alone on arm x x x is rewarded with probability μ x \mu_x μ x . Users that share an arm are each rewarded, independently, with probability ε x \varepsilon_x ε x , whatever the number of colliders. An arm with μ x = ε x = 0 \mu_x = \varepsilon_x = 0 μ x = ε x = 0 is never rewarded, whether or not it is shared, so it is a silent input . The channel is symmetric (Definition 1), and the outputs are conditionally independent, so that the per-receiver statements of Section IV apply.
Below are some notable special cases.
ε = 0 \varepsilon= 0 ε = 0 is the original bandit channel: a collision is an erasure whose output is indistinguishable from a failed reward. A positive output is then a certificate that the user was alone, and no collision is ever reported as a reward.
If 0 < μ x < 1 0 < \mu_x < 1 0 < μ x < 1 and 0 < ε x < 1 0 < \varepsilon_x < 1 0 < ε x < 1 for the arms in use, every output has positive probability, no output certifies a role, and by Proposition 8 no finite-state controller has a vanishing error probability, even if silent arms are also used. With maps that vary with time and J J J arms with the same parameters ε < μ \varepsilon< \mu ε < μ , a bounded state set suffices for a vanishing error probability (cf. Proposition 16).
If ε x = μ x \varepsilon_x = \mu_x ε x = μ x for every x x x , collisions are invisible, the channel is noninteracting, and symmetry breaking is impossible (cf. Corollary 7).
The Converse
The converse of Section III bounds the error probability of every symmetric causal protocol. We apply Theorem 1 with m = 2 m = 2 m = 2 : users 1 and 2 are selected, and the other J − 2 J - 2 J − 2 users are unselected. Since the outputs are independent given the inputs, the selected pair sees the channel through the two local laws W ( ⋅ ∣ x 1 , x − 1 ) W(\cdot\mid x_1,x_{-1}) W ( ⋅ ∣ x 1 , x − 1 ) and W ( ⋅ ∣ x 2 , x − 2 ) W(\cdot\mid x_2,x_{-2}) W ( ⋅ ∣ x 2 , x − 2 ) . For p , q ∈ [ 0 , 1 ] p, q \in[0,1] p , q ∈ [ 0 , 1 ] write D α ( p ∥ q ) : = D α ( Bern ( p ) ∥ Bern ( q ) ) D_\alpha(p\Vert q) := D_\alpha(\operatorname{Bern}(p)\Vert\operatorname{Bern}(q)) D α ( p ∥ q ) := D α ( Bern ( p ) ∥ Bern ( q )) , with D 1 D_1 D 1 the Kullback–Leibler divergence.
a) A noninteracting reference: A noninteracting reference gives every selected user a private reward law v x ∈ [ 0 , 1 ] v_x \in[0,1] v x ∈ [ 0 , 1 ] for arm x x x , whatever the other users do. In each channel use, a selected user on arm x x x contributes to the cost of the change of measure the Rényi divergence between Bern ( v x ) \operatorname{Bern}(v_x) Bern ( v x ) and its true reward law, which is Bern ( μ x ) \operatorname{Bern}(\mu_x) Bern ( μ x ) if it is alone on the arm and Bern ( ε x ) \operatorname{Bern}(\varepsilon_x) Bern ( ε x ) if it shares it:
a x : = D α ( v x ∥ μ x ) , s x : = D α ( v x ∥ ε x ) . (86) a_x := D_\alpha(v_x\Vert\mu_x), \qquad s_x := D_\alpha(v_x\Vert\varepsilon_x). \tag*{(86)} a x := D α ( v x ∥ μ x ) , s x := D α ( v x ∥ ε x ) . (86)
On a silent arm both laws are δ 0 \delta_0 δ 0 , so that a x = s x = 0 a_x = s_x = 0 a x = s x = 0 if v x = 0 v_x = 0 v x = 0 and a x = s x = ∞ a_x = s_x = \infty a x = s x = ∞ otherwise. For J ≥ 3 J \ge3 J ≥ 3 , the unselected users can make a selected user alone or in collision. Let a ˉ x : = a x \bar a_x := a_x a ˉ x := a x if J = 2 J = 2 J = 2 and a ˉ x : = max { a x , s x } \bar a_x := \max\{a_x,s_x\} a ˉ x := max { a x , s x } if J ≥ 3 J \ge3 J ≥ 3 , and put
Δ α ( v ) : = max { max x ∈ [ K ] 2 s x , max 1 ≤ x < x ′ ≤ K ( a ˉ x + a ˉ x ′ ) } . (87) \Delta_\alpha(v) := \max\left\{\max_{x \in[K]} 2s_x,\ \max_{1 \le x < x' \le K}(\bar a_x+\bar a_{x'})\right\}. \tag*{(87)} Δ α ( v ) := max { x ∈ [ K ] max 2 s x , 1 ≤ x < x ′ ≤ K max ( a ˉ x + a ˉ x ′ ) } . (87)
Proposition 11 (Noninteraction converse). For every v ∈ [ 0 , 1 ] K v \in[0,1]^K v ∈ [ 0 , 1 ] K , every α ∈ ( 1 , ∞ ] \alpha\in(1,\infty] α ∈ ( 1 , ∞ ] , every blocklength n n n and every symmetric causal protocol,
P b ( n ) ≥ γ 2 , J α e − n Δ α ( v ) . (88) P_{\mathrm{b}}^{(n)} \ge\gamma_{2,J}^{\alpha}e^{-n\Delta_\alpha(v)}. \tag*{(88)} P b ( n ) ≥ γ 2 , J α e − n Δ α ( v ) . (88)
Consequently E ‾ b ( P ) ≤ inf v Δ 1 ( v ) \overline{E}_{\mathrm{b}}(P) \le\inf_v \Delta_1(v) E b ( P ) ≤ inf v Δ 1 ( v ) .
Proof. Let V ( 1 ∣ x ) = v x V(1 \mid x) = v_x V ( 1 ∣ x ) = v x . Since P 2 P_2 P 2 is a product of the two local laws and D α D_\alpha D α is additive on products, the radius (50) is
d α , 2 ( V ∥ P ) = max x [ J ] ∑ j = 1 2 D α ( V ( ⋅ ∣ x j ) ∥ W ( ⋅ ∣ x j , x − j ) ) . d_{\alpha,2}(V\Vert P) = \max_{x_{[J]}} \sum_{j=1}^{2}D_\alpha\bigl(V(\cdot\mid x_j)\Vert W(\cdot\mid x_j,x_{-j})\bigr). d α , 2 ( V ∥ P ) = x [ J ] max j = 1 ∑ 2 D α ( V ( ⋅ ∣ x j ) ∥ W ( ⋅ ∣ x j , x − j ) ) .
The term of user j j j is a x j a_{x_j} a x j if it is alone and s x j s_{x_j} s x j if it shares its arm. If x 1 = x 2 = x x_1 = x_2 = x x 1 = x 2 = x , both users share their arm and the sum is 2 s x 2s_x 2 s x . If x 1 ≠ x 2 x_1 \ne x_2 x 1 = x 2 , then for J = 2 J = 2 J = 2 both users are alone, and for J ≥ 3 J \ge3 J ≥ 3 the unselected users can make each selected user alone or in collision, except that for J = 3 J = 3 J = 3 the two selected users cannot both be in collision, and for K = 2 K = 2 K = 2 they cannot both be alone. Hence d α , 2 ( V ∥ P ) ≤ Δ α ( v ) d_{\alpha,2}(V\Vert P) \le\Delta_\alpha(v) d α , 2 ( V ∥ P ) ≤ Δ α ( v ) , with equality if J = 2 J = 2 J = 2 or K ≥ 3 K \ge3 K ≥ 3 , since s x + s x ′ ≤ max { 2 s x , 2 s x ′ } s_x+s_{x'} \le\max\{2s_x,2s_{x'}\} s x + s x ′ ≤ max { 2 s x , 2 s x ′ } . Corollary 2 gives the bound, and its second statement the exponent. □ \square □
Since D 1 ( v ∥ q ) D_1(v\Vert q) D 1 ( v ∥ q ) is convex in v v v , so is Δ 1 \Delta_1 Δ 1 , and the infimum in Proposition 11 is a convex program. For arms with the same parameters the optimal reference is explicit. Let
Λ ( μ , ε ) : = min v ∈ [ 0 , 1 ] max { D 1 ( v ∥ μ ) , D 1 ( v ∥ ε ) } = sup 0 < λ < 1 − log [ μ λ ε 1 − λ + ( 1 − μ ) λ ( 1 − ε ) 1 − λ ] (89) \Lambda(\mu,\varepsilon) := \min_{v \in[0,1]}\max\{D_1(v\Vert\mu),D_1(v\Vert\varepsilon)\} = \sup_{0<\lambda<1}-\log\left[\mu^\lambda\varepsilon^{1-\lambda}+(1-\mu)^\lambda(1-\varepsilon)^{1-\lambda}\right] \tag*{(89)} Λ ( μ , ε ) := v ∈ [ 0 , 1 ] min max { D 1 ( v ∥ μ ) , D 1 ( v ∥ ε )} = 0 < λ < 1 sup − log [ μ λ ε 1 − λ + ( 1 − μ ) λ ( 1 − ε ) 1 − λ ] (89)
be the Chernoff information between Bern ( μ ) \operatorname{Bern}(\mu) Bern ( μ ) and Bern ( ε ) \operatorname{Bern}(\varepsilon) Bern ( ε ) , where the equality is from the minimax theorem.
Corollary 5 (Closed-form noninteraction bounds). If μ x = μ \mu_x = \mu μ x = μ and ε x = ε \varepsilon_x = \varepsilon ε x = ε for all x ∈ [ K ] x \in[K] x ∈ [ K ] and K ≥ 2 K \ge2 K ≥ 2 , then E ‾ b ( P ) ≤ 2 Λ ( μ , ε ) \overline{E}_{\mathrm{b}}(P) \le2\Lambda(\mu,\varepsilon) E b ( P ) ≤ 2Λ ( μ , ε ) . If ε = 0 \varepsilon= 0 ε = 0 and K ≥ 2 K \ge2 K ≥ 2 , then E ‾ b ( P ) ≤ − log ( 1 − μ ( 1 ) ) − log ( 1 − μ ( 2 ) ) \overline{E}_{\mathrm{b}}(P) \le-\log(1-\mu_{(1)})-\log(1-\mu_{(2)}) E b ( P ) ≤ − log ( 1 − μ ( 1 ) ) − log ( 1 − μ ( 2 ) ) , where μ ( 1 ) ≥ μ ( 2 ) \mu_{(1)} \ge\mu_{(2)} μ ( 1 ) ≥ μ ( 2 ) are the two largest means.
Proof. With v x = v v_x = v v x = v , (87) is 2 max { D 1 ( v ∥ μ ) , D 1 ( v ∥ ε ) } 2\max\{D_1(v\Vert\mu),D_1(v\Vert\varepsilon)\} 2 max { D 1 ( v ∥ μ ) , D 1 ( v ∥ ε )} for every J ≥ 2 J \ge2 J ≥ 2 , and minimizing over v v v gives (89). If ε x = 0 \varepsilon_x = 0 ε x = 0 , then s x = ∞ s_x = \infty s x = ∞ unless v x = 0 v_x = 0 v x = 0 , and v = 0 v = 0 v = 0 gives a x = − log ( 1 − μ x ) a_x = -\log(1-\mu_x) a x = − log ( 1 − μ x ) and s x = 0 s_x = 0 s x = 0 , so that (87) is the sum of the two largest values of − log ( 1 − μ x ) -\log(1-\mu_x) − log ( 1 − μ x ) . □ \square □
The reference ν \nu ν is the reward law of a lone player that cannot tell whether it shares its arm. The radius is small when collisions are hard to detect, ε x \varepsilon_x ε x close to μ x \mu_x μ x , and it vanishes when ε x = μ x \varepsilon_x=\mu_x ε x = μ x for all x x x . For ε = 0 \varepsilon=0 ε = 0 the noninteraction bound is − log ( 1 − μ ( 1 ) ) − log ( 1 − μ ( 2 ) ) -\log(1-\mu_{(1)})-\log(1-\mu_{(2)}) − log ( 1 − μ ( 1 ) ) − log ( 1 − μ ( 2 ) ) , which diverges as μ ( 1 ) → 1 \mu_{(1)}\to1 μ ( 1 ) → 1 , although a perfect collision detector has a finite exponent (Corollary 6). The following reference does not have this defect.
b) A common-output reference: Assume that a collision does not increase the reward probability, i.e., ε x ≤ μ x \varepsilon_x\le\mu_x ε x ≤ μ x . The idea of the reference is to keep only the trajectories on which the two selected users see the same outputs, so that their views differ only through their private seeds. In the notation of Section III, this is a kernel B B B of the form (34) with m = 2 m=2 m = 2 and the normalized completion (N), with the physical completion of Remark 1. For a kernel with B ≤ P B\le P B ≤ P , the order-infinity case of Theorem 1 gives P b ( n ) ≥ γ 2 , J κ 2 ( B ) n P_b^{(n)}\ge\gamma_{2,J}\kappa_2(B)^n P b ( n ) ≥ γ 2 , J κ 2 ( B ) n , where κ 2 ( B ) \kappa_2(B) κ 2 ( B ) is the persistence coefficient (37); see (46) . We therefore look for a kernel B ≤ P B\le P B ≤ P with the largest possible persistence, and compute κ 2 ( B ) \kappa_2(B) κ 2 ( B ) for the kernel below. Put
μ ˉ x : = 1 − μ x , τ x : = ( 1 − ε x ) 2 − ( 1 − μ x ) 2 ≥ 0 , x ∈ X , (90) \bar{\mu}_x:=1-\mu_x,\qquad\tau_x:=(1-\varepsilon_x)^2-(1-\mu_x)^2\ge0,\qquad x\in\mathcal{X}, \tag*{(90)} μ ˉ x := 1 − μ x , τ x := ( 1 − ε x ) 2 − ( 1 − μ x ) 2 ≥ 0 , x ∈ X , (90)
so that a silent arm has μ ˉ x = 1 \bar{\mu}_x=1 μ ˉ x = 1 and τ x = 0 \tau_x=0 τ x = 0 , and for a law p p p on X \mathcal{X} X let
κ ( p ) : = ( ∑ x p x μ ˉ x ) 2 + ( ∑ x p x ε x ) 2 + ∑ x τ x p x 2 , κ C O : = min p ∈ P ( X ) κ ( p ) > 0. (91) \kappa(p):=\left(\sum_x p_x\bar{\mu}_x\right)^2+\left(\sum_x p_x\varepsilon_x\right)^2+\sum_x\tau_xp_x^2,\qquad\kappa_{\mathrm{CO}}:=\min_{p\in\mathcal{P}(\mathcal{X})}\kappa(p)>0. \tag*{(91)} κ ( p ) := ( x ∑ p x μ ˉ x ) 2 + ( x ∑ p x ε x ) 2 + x ∑ τ x p x 2 , κ CO := p ∈ P ( X ) min κ ( p ) > 0. (91)
The function κ \kappa κ is a convex quadratic on the simplex.
Proposition 12 (Common-output converse). If ε x ≤ μ x \varepsilon_x\le\mu_x ε x ≤ μ x for all x ∈ [ K ] x\in[K] x ∈ [ K ] , then for every J ≥ 2 J\ge2 J ≥ 2 , every n n n and every symmetric causal protocol (an instance of Theorem 1 with m = 2 m=2 m = 2 and α = ∞ \alpha=\infty α = ∞ ),
P b ( n ) ≥ γ 2 , J κ C O n , E ‾ b ( P ) ≤ − log κ C O . (92) P_b^{(n)}\ge\gamma_{2,J}\kappa_{\mathrm{CO}}^n,\qquad\overline{E}_b(P)\le-\log\kappa_{\mathrm{CO}}. \tag*{(92)} P b ( n ) ≥ γ 2 , J κ CO n , E b ( P ) ≤ − log κ CO . (92)
Proof. Define for a pair of selected users
B 2 ( y 1 , y 2 ∣ x 1 , x 2 ) : = 1 { y 1 = y 2 = 0 } [ μ ˉ x 1 μ ˉ x 2 + 1 { x 1 = x 2 } τ x 1 ] + 1 { y 1 = y 2 = 1 } ε x 1 ε x 2 , B^2(y_1,y_2\mid x_1,x_2):=\mathbb{1}\{y_1=y_2=0\}\left[\bar{\mu}_{x_1}\bar{\mu}_{x_2}+\mathbb{1}\{x_1=x_2\}\tau_{x_1}\right]+\mathbb{1}\{y_1=y_2=1\}\varepsilon_{x_1}\varepsilon_{x_2}, B 2 ( y 1 , y 2 ∣ x 1 , x 2 ) := 1 { y 1 = y 2 = 0 } [ μ ˉ x 1 μ ˉ x 2 + 1 { x 1 = x 2 } τ x 1 ] + 1 { y 1 = y 2 = 1 } ε x 1 ε x 2 ,
and B ( y [ J ] ∣ x [ J ] ) : = B 2 ( y 1 , y 2 ∣ x 1 , x 2 ) P ( y 3 : J ∣ x [ J ] ) B(y_{[J]}\mid x_{[J]}):=B^2(y_1,y_2\mid x_1,x_2)P(y_{3:J}\mid x_{[J]}) B ( y [ J ] ∣ x [ J ] ) := B 2 ( y 1 , y 2 ∣ x 1 , x 2 ) P ( y 3 : J ∣ x [ J ] ) , where the last factor is the physical law of the unselected outputs, which are independent of y 1 , y 2 y_1,y_2 y 1 , y 2 . This kernel has a normalized completion, so it satisfies condition (N). Its selected part is
∑ r ∈ { u } ∪ [ K ] ∪ { e } a r ( x 1 , y 1 ) a r ( x 2 , y 2 ) , \sum_{r\in\{u\}\cup[K]\cup\{e\}}a_r(x_1,y_1)a_r(x_2,y_2), r ∈ { u } ∪ [ K ] ∪ { e } ∑ a r ( x 1 , y 1 ) a r ( x 2 , y 2 ) ,
with the components a u ( x , y ) = μ ˉ x 1 { y = 0 } a_u(x,y)=\bar{\mu}_x\mathbb{1}\{y=0\} a u ( x , y ) = μ ˉ x 1 { y = 0 } , a x ′ ( x , y ) = τ x ′ 1 { x = x ′ , y = 0 } a_{x'}(x,y)=\sqrt{\tau_{x'}}\mathbb{1}\{x=x',y=0\} a x ′ ( x , y ) = τ x ′ 1 { x = x ′ , y = 0 } for x ′ ∈ [ K ] x'\in[K] x ′ ∈ [ K ] , and a e ( x , y ) = ε x 1 { y = 1 } a_e(x,y)=\varepsilon_x\mathbb{1}\{y=1\} a e ( x , y ) = ε x 1 { y = 1 } .
We show B ≤ P B\le P B ≤ P . If x 1 = x 2 = x x_1=x_2=x x 1 = x 2 = x , both selected users share their arm, and B 2 ( 0 , 0 ∣ x , x ) = μ ˉ x 2 + τ x = ( 1 − ε x ) 2 B^2(0,0\mid x,x)=\bar{\mu}_x^2+\tau_x=(1-\varepsilon_x)^2 B 2 ( 0 , 0 ∣ x , x ) = μ ˉ x 2 + τ x = ( 1 − ε x ) 2 and B 2 ( 1 , 1 ∣ x , x ) = ε x 2 B^2(1,1\mid x,x)=\varepsilon_x^2 B 2 ( 1 , 1 ∣ x , x ) = ε x 2 are the physical probabilities. Otherwise, in every placement of the unselected users, each selected user is rewarded with probability μ x j \mu_{x_j} μ x j or ε x j \varepsilon_{x_j} ε x j , both at least ε x j \varepsilon_{x_j} ε x j and at most μ x j \mu_{x_j} μ x j since ε x j ≤ μ x j \varepsilon_{x_j}\le\mu_{x_j} ε x j ≤ μ x j . Hence its probability of a zero is at least μ ˉ x j \bar{\mu}_{x_j} μ ˉ x j and of a one at least ε x j \varepsilon_{x_j} ε x j , and B 2 ≤ P 2 B^2\le P_2 B 2 ≤ P 2 for all placements. Therefore ρ ∞ , 2 ( B ; P ) ≤ 1 \rho_{\infty,2}(B;P)\le1 ρ ∞ , 2 ( B ; P ) ≤ 1 , and K ∞ , 2 ( B ; P ) ≥ κ 2 ( B ) \mathcal{K}_{\infty,2}(B;P)\ge\kappa_2(B) K ∞ , 2 ( B ; P ) ≥ κ 2 ( B ) . Finally, A r ( x ) = ∑ y a r ( x , y ) A_r(x)=\sum_y a_r(x,y) A r ( x ) = ∑ y a r ( x , y ) equals μ ˉ x \bar{\mu}_x μ ˉ x , τ x ′ 1 { x = x ′ } \sqrt{\tau_{x'}}\mathbb{1}\{x=x'\} τ x ′ 1 { x = x ′ } and ε x \varepsilon_x ε x for the three kinds of component, and (37) gives κ 2 ( B ) = min p κ ( p ) \kappa_2(B)=\min_p\kappa(p) κ 2 ( B ) = min p κ ( p ) . Theorem 1 with α = ∞ \alpha=\infty α = ∞ and b ∞ = 1 b_\infty=1 b ∞ = 1 concludes. □ \square □
The minimization in (91) is a quadratic program. For equal parameters it has a closed form.
Corollary 6 (Equal parameters). Let k ≥ 1 k\ge1 k ≥ 1 arms have the parameters ( μ , ε ) (\mu,\varepsilon) ( μ , ε ) with ε ≤ μ \varepsilon\le\mu ε ≤ μ , suppose that the other K − k ≥ 0 K-k\ge0 K − k ≥ 0 arms are silent, and put τ : = ( 1 − ε ) 2 − ( 1 − μ ) 2 \tau:=(1-\varepsilon)^2-(1-\mu)^2 τ := ( 1 − ε ) 2 − ( 1 − μ ) 2 and a ∗ : = μ / ( μ 2 + ε 2 + τ / k ) a^*:=\mu/(\mu^2+\varepsilon^2+\tau/k) a ∗ := μ / ( μ 2 + ε 2 + τ / k ) . If some arm is silent and a ∗ ≤ 1 a^*\le1 a ∗ ≤ 1 , the optimal law puts mass a ∗ a^* a ∗ uniformly on the k k k arms and the rest on a silent arm, and κ C O = 1 − μ a ∗ \kappa_{\mathrm{CO}}=1-\mu a^* κ CO = 1 − μ a ∗ . Otherwise the optimal law is uniform on the k k k arms, and κ C O = ( 1 − μ ) 2 + ε 2 + τ / k \kappa_{\mathrm{CO}}=(1-\mu)^2+\varepsilon^2+\tau/k κ CO = ( 1 − μ ) 2 + ε 2 + τ / k . For ε = 0 \varepsilon=0 ε = 0 and a silent arm, a ∗ ≤ 1 a^*\le1 a ∗ ≤ 1 if and only if k − 2 ≤ ( k − 1 ) μ k-2\le(k-1)\mu k − 2 ≤ ( k − 1 ) μ , and then
κ C O = 2 − μ 2 + ( k − 1 ) μ , E ‾ b ( P ) ≤ log 2 + ( k − 1 ) μ 2 − μ ; (93) \kappa_{\mathrm{CO}}=\frac{2-\mu}{2+(k-1)\mu},\qquad\overline{E}_b(P)\le\log\frac{2+(k-1)\mu}{2-\mu}; \tag*{(93)} κ CO = 2 + ( k − 1 ) μ 2 − μ , E b ( P ) ≤ log 2 − μ 2 + ( k − 1 ) μ ; (93) for ε = 0 \varepsilon=0 ε = 0 and no silent arm, κ C O = ( 1 − μ ) 2 + ( 2 μ − μ 2 ) / k \kappa_{\mathrm{CO}}=(1-\mu)^2+(2\mu-\mu^2)/k κ CO = ( 1 − μ ) 2 + ( 2 μ − μ 2 ) / k ; and for k = 1 k=1 k = 1 and a silent arm, κ C O = 1 − μ 2 / [ 2 ( μ − ε ( 1 − ε ) ) ] \kappa_{\mathrm{CO}}=1-\mu^2/[2(\mu-\varepsilon(1-\varepsilon))] κ CO = 1 − μ 2 / [ 2 ( μ − ε ( 1 − ε ))] if 2 ε ( 1 − ε ) ≤ μ 2\varepsilon(1-\varepsilon)\le\mu 2 ε ( 1 − ε ) ≤ μ , and κ C O = ( 1 − ε ) 2 + ε 2 \kappa_{\mathrm{CO}}=(1-\varepsilon)^2+\varepsilon^2 κ CO = ( 1 − ε ) 2 + ε 2 otherwise.
Proof. Since κ \kappa κ is convex and invariant under permutations of the k k k arms, and of the silent arms, an optimal law spreads its mass uniformly on the k k k arms, and only its total mass on the silent arms matters. With a a a the mass on the k k k arms, κ ( a ) = ( 1 − μ a ) 2 + ε 2 a 2 + τ a 2 / k \kappa(a)=(1-\mu a)^2+\varepsilon^2a^2+\tau a^2/k κ ( a ) = ( 1 − μ a ) 2 + ε 2 a 2 + τ a 2 / k , a convex quadratic minimized over the real line at a = a ∗ a=a^* a = a ∗ , where κ ( a ∗ ) = 1 − μ a ∗ \kappa(a^*)=1-\mu a^* κ ( a ∗ ) = 1 − μ a ∗ by the first-order condition; a = 1 a=1 a = 1 is forced if no arm is silent, and is optimal if a ∗ > 1 a^*>1 a ∗ > 1 . For ε = 0 \varepsilon=0 ε = 0 , τ = 2 μ − μ 2 \tau=2\mu-\mu^2 τ = 2 μ − μ 2 and a ∗ = k / ( 2 + ( k − 1 ) μ ) a^*=k/(2+(k-1)\mu) a ∗ = k / ( 2 + ( k − 1 ) μ ) , and substitution gives (93); for k = 1 k=1 k = 1 , μ 2 + ε 2 + τ = 2 ( μ − ε ( 1 − ε ) ) \mu^2+\varepsilon^2+\tau=2(\mu-\varepsilon(1-\varepsilon)) μ 2 + ε 2 + τ = 2 ( μ − ε ( 1 − ε )) . □ \square □
For μ → 1 \mu\to1 μ → 1 the exponent bound tends to log ( k + 1 ) \log(k+1) log ( k + 1 ) , its value at μ = 1 \mu=1 μ = 1 ; for μ < 1 \mu<1 μ < 1 it is smaller. Unlike the noninteraction bound of Corollary 5, it stays finite as μ → 1 \mu\to1 μ → 1 .
c) Combined statement, optimized references, and the dependence on J J J : Both propositions hold at once, so that P ‾ b ( n ) ≥ max { γ 2 , J b α e − n Δ α ( v ) , γ 2 , J κ C O n } \overline{P}_{b}^{(n)} \ge\max\{\gamma_{2,J}^{b_\alpha}e^{-n\Delta_\alpha(v)},\gamma_{2,J}\kappa_{\mathrm{CO}}^n\} P b ( n ) ≥ max { γ 2 , J b α e − n Δ α ( v ) , γ 2 , J κ CO n } for every v v v and α \alpha α , and E ‾ b ( P ) ≤ min { inf v Δ 1 ( v ) , − log κ C O } \overline{E}_b(P) \le\min\{\inf_v\Delta_1(v),-\log\kappa_{\mathrm{CO}}\} E b ( P ) ≤ min { inf v Δ 1 ( v ) , − log κ CO } (the second terms are present if ε x ≤ μ x \varepsilon_x \le\mu_x ε x ≤ μ x for all x x x ). The noninteraction bound is the best for noisy collisions, where the two laws to be told apart are close, and the common-output bound for erasures, where the noninteraction bound is large. Those are two kernels of the family of Corollary 1, and neither is optimal in it. A reference B 2 = ∑ r a r ⊗ a r B^2=\sum_r a_r\otimes a_r B 2 = ∑ r a r ⊗ a r with several components, optimized numerically under the constraint B ≤ P B\le P B ≤ P or, for a finite α \alpha α , by minimizing Ψ α , 2 ( B ; P ) \Psi_{\alpha,2}(B;P) Ψ α , 2 ( B ; P ) under κ 2 ( B ) ≥ 1 \kappa_2(B)\ge1 κ 2 ( B ) ≥ 1 , is a valid reference as soon as it is feasible, and the numerical values in Table 1 are of this kind. They improve on the smaller of the two closed forms by up to 12% in the exponent, at the intermediate noise levels where the two closed forms cross, and coincide with them at ε = 0 \varepsilon=0 ε = 0 and ε = 0.2 \varepsilon=0.2 ε = 0.2 for μ = 0.8 \mu=0.8 μ = 0.8 . How they are computed, and how precise they are, is described in Appendix § 3.7 .
Both propositions are instances of the kernel converse with m = 2 m=2 m = 2 . For noninteracting references m = 2 m=2 m = 2 is optimal, by data processing; for general kernels there is no such monotonicity, and we have not evaluated kernels with m > 2 m>2 m > 2 . As they stand, the exponents of the bounds depend on the number of users only through the distinction between J = 2 J=2 J = 2 and J ≥ 3 J\ge3 J ≥ 3 in Proposition 11, and not at all in Proposition 12: only the two selected users are charged, and the unselected users are adversarial. Section VI-E discusses the dependence on J J J .
Block Coding
For block coding the Gallager function of Theorem 2 is explicit. Fix an input law Q Q Q on X \mathcal{X} X with q x = Q ( x ) q_x=Q(x) q x = Q ( x ) , an error subset of size s s s , and c = J − 1 − s c=J-1-s c = J − 1 − s correct foreign codewords. Let
G ρ , s ( J ) ( Q ; μ , ε ) : = ∑ x ∈ X q x { 1 − ( 1 − q x ) c + ( 1 − q x ) c ∑ y ∈ { 0 , 1 } [ ( 1 − q x ) s W μ x ( y ) 1 1 + ρ + ( 1 − ( 1 − q x ) s ) W ε x ( y ) 1 1 + ρ ] 1 + ρ } . (94) G_{\rho,s}^{(J)}(Q;\mu,\varepsilon):=\sum_{x\in\mathcal{X}}q_x\left\{1-(1-q_x)^c+(1-q_x)^c\sum_{y\in\{0,1\}}\left[(1-q_x)^sW_{\mu_x}(y)^{\frac{1}{1+\rho}}+\left(1-(1-q_x)^s\right)W_{\varepsilon_x}(y)^{\frac{1}{1+\rho}}\right]^{1+\rho}\right\}. \tag*{(94)} G ρ , s ( J ) ( Q ; μ , ε ) := x ∈ X ∑ q x ⎩ ⎨ ⎧ 1 − ( 1 − q x ) c + ( 1 − q x ) c y ∈ { 0 , 1 } ∑ [ ( 1 − q x ) s W μ x ( y ) 1 + ρ 1 + ( 1 − ( 1 − q x ) s ) W ε x ( y ) 1 + ρ 1 ] 1 + ρ ⎭ ⎬ ⎫ . (94)
where W ν ( 1 ) : = ν W_\nu(1):=\nu W ν ( 1 ) := ν and W ν ( 0 ) : = 1 − ν W_\nu(0):=1-\nu W ν ( 0 ) := 1 − ν .
Proposition 13 (Noisy bandit Gallager function). For the channel of Definition 10, E 0 , s ( ρ , Q ) = − log G ρ , s ( J ) ( Q ; μ , ε ) E_{0,s}(\rho,Q)=-\log G_{\rho,s}^{(J)}(Q;\mu,\varepsilon) E 0 , s ( ρ , Q ) = − log G ρ , s ( J ) ( Q ; μ , ε ) , and hence
E r , s ( R , Q ) = max 0 ≤ ρ ≤ 1 { − log G ρ , s ( J ) ( Q ; μ , ε ) − ρ s R } . (95) E_{r,s}(R,Q)=\max_{0\le\rho\le1}\{-\log G_{\rho,s}^{(J)}(Q;\mu,\varepsilon)-\rho sR\}. \tag*{(95)} E r , s ( R , Q ) = 0 ≤ ρ ≤ 1 max { − log G ρ , s ( J ) ( Q ; μ , ε ) − ρ s R } . (95)
The proof is in Appendix § 3.6 . Together with Theorem 2 this gives an explicit error bound for J J J users. The two terms in the braces of (94) are the following: if a correct foreign codeword collides with the receiver, the output law does not depend on the candidate and contributes 1; if no correct codeword collides, the candidate tuple avoids the receiver’s arm with probability ( 1 − q x ) s (1-q_x)^s ( 1 − q x ) s , and the likelihood is W μ x W_{\mu_x} W μ x or W ε x W_{\varepsilon_x} W ε x accordingly. Proposition 13 does not require ε x ≤ μ x \varepsilon_x\le\mu_x ε x ≤ μ x . At ε = 0 \varepsilon=0 ε = 0 it is the function of the noiseless bandit channel.
a) Two users: For J = 2 J=2 J = 2 the only error subset is s = 1 s=1 s = 1 , c = 0 c=0 c = 0 . With { j , j ′ } = [ 2 ] \{j,j'\}=[2] { j , j ′ } = [ 2 ] , the receiver’s output depends on X j ′ X_{j'} X j ′ only through the indicator of a collision, which has probability q x q_x q x given X j ′ = x X_{j'}=x X j ′ = x , so that
I Q ( X j ′ ; Y j ∣ X j ) = ∑ x ∈ X q x [ h 2 ( ( 1 − q x ) μ x + q x ε x ) − ( 1 − q x ) h 2 ( μ x ) − q x h 2 ( ε x ) ] . (96) I_Q(X_{j'};Y_j\mid X_j)=\sum_{x\in\mathcal{X}}q_x\left[h_2\left((1-q_x)\mu_x+q_x\varepsilon_x\right)-(1-q_x)h_2(\mu_x)-q_xh_2(\varepsilon_x)\right]. \tag*{(96)} I Q ( X j ′ ; Y j ∣ X j ) = x ∈ X ∑ q x [ h 2 ( ( 1 − q x ) μ x + q x ε x ) − ( 1 − q x ) h 2 ( μ x ) − q x h 2 ( ε x ) ] . (96)
E 0 , 1 ( ρ , Q ) = − log ∑ x ∈ X q x ∑ y ∈ { 0 , 1 } [ ( 1 − q x ) W μ x ( y ) 1 1 + ρ + q x W ε x ( y ) 1 1 + ρ ] 1 + ρ . (97) E_{0,1}(\rho,Q)=-\log\sum_{x\in\mathcal{X}}q_x\sum_{y\in\{0,1\}}\left[(1-q_x)W_{\mu_x}(y)^{\frac{1}{1+\rho}}+q_xW_{\varepsilon_x}(y)^{\frac{1}{1+\rho}}\right]^{1+\rho}. \tag*{(97)} E 0 , 1 ( ρ , Q ) = − log x ∈ X ∑ q x y ∈ { 0 , 1 } ∑ [ ( 1 − q x ) W μ x ( y ) 1 + ρ 1 + q x W ε x ( y ) 1 + ρ 1 ] 1 + ρ . (97)
For ε = 0 \varepsilon=0 ε = 0 , (97) and (96) are the exponent function of [9 Cor. 1] and the rate of [9 Th. 4] , and those of [13 Ths. 1 and 2] with a single block. With R R R the name rate, the collision of the two names contributes e − n R e^{-nR} e − n R , so that for J = 2 J=2 J = 2 the exponent (63) is
E b c ( Q ) = sup R > 0 min { R , E r , 1 ( R , Q ) } = max 0 ≤ ρ ≤ 1 E 0 , 1 ( ρ , Q ) 1 + ρ . (98) E_{\mathrm{bc}}(Q)=\sup_{R>0}\min\{R,E_{r,1}(R,Q)\}=\max_{0\le\rho\le1}\frac{E_{0,1}(\rho,Q)}{1+\rho}. \tag*{(98)} E bc ( Q ) = R > 0 sup min { R , E r , 1 ( R , Q )} = 0 ≤ ρ ≤ 1 max 1 + ρ E 0 , 1 ( ρ , Q ) . (98)
because E r , 1 ( R , Q ) ≥ R E_{r,1}(R,Q)\ge R E r , 1 ( R , Q ) ≥ R exactly when R ≤ E 0 , 1 ( ρ , Q ) / ( 1 + ρ ) R\le E_{0,1}(\rho,Q)/(1+\rho) R ≤ E 0 , 1 ( ρ , Q ) / ( 1 + ρ ) for some ρ \rho ρ .
For J J J users the full-set information, which includes the rate threshold in Proposition 3, is obtained in the same way, with θ x : = ( 1 − q x ) J − 1 \theta_x:=(1-q_x)^{J-1} θ x := ( 1 − q x ) J − 1 the probability that no foreign input equals x x x :
I Q ( X − j ; Y j ∣ X j ) = ∑ x ∈ X q x [ h 2 ( θ x μ x + ( 1 − θ x ) ε x ) − θ x h 2 ( μ x ) − ( 1 − θ x ) h 2 ( ε x ) ] . (99) I_Q(X_{-j};Y_j\mid X_j)=\sum_{x\in\mathcal{X}}q_x\left[h_2\left(\theta_x\mu_x+(1-\theta_x)\varepsilon_x\right)-\theta_xh_2(\mu_x)-(1-\theta_x)h_2(\varepsilon_x)\right]. \tag*{(99)} I Q ( X − j ; Y j ∣ X j ) = x ∈ X ∑ q x [ h 2 ( θ x μ x + ( 1 − θ x ) ε x ) − θ x h 2 ( μ x ) − ( 1 − θ x ) h 2 ( ε x ) ] . (99)
It is positive if and only if μ x ≠ ε x \mu_x\ne\varepsilon_x μ x = ε x for some x x x with q x ∈ ( 0 , 1 ) q_x\in(0,1) q x ∈ ( 0 , 1 ) , by the strict concavity of h 2 h_2 h 2 .
Corollary 7 (Dichotomy). For every J ≥ 2 J\ge2 J ≥ 2 , E b ( P ) > 0 E_b(P)>0 E b ( P ) > 0 if K ≥ 2 K\ge2 K ≥ 2 and μ x ≠ ε x \mu_x\ne\varepsilon_x μ x = ε x for some x ∈ [ K ] x\in[K] x ∈ [ K ] , and ε n ∗ ( P ) ≥ γ J = 1 − J ! / J J \varepsilon_n^*(P)\ge\gamma_J=1-J!/J^J ε n ∗ ( P ) ≥ γ J = 1 − J ! / J J for every n n n if μ x = ε x \mu_x=\varepsilon_x μ x = ε x for all x ∈ [ K ] x\in[K] x ∈ [ K ] or if $K=1.
Proof. If K ≥ 2 K \ge2 K ≥ 2 and μ x ≠ ϵ x \mu_x \ne\epsilon_x μ x = ϵ x , take Q Q Q supported on { x , x ′ } \{x,x'\} { x , x ′ } for some x ′ ≠ x x' \ne x x ′ = x , with q x ∈ ( 0 , 1 ) q_x \in(0,1) q x ∈ ( 0 , 1 ) , so that the term of arm x x x of (99) is positive and the others are nonnegative, and apply Proposition 3 and Theorem 2. If μ x = ϵ x \mu_x = \epsilon_x μ x = ϵ x for all x x x , or if K = 1 K=1 K = 1 , so that the users always share the single arm, the output of each user depends only on its own input: P ( y [ J ] ∣ x [ J ] ) = ∏ j W ( y j ∣ x j ) P(y_{[J]} \mid x_{[J]}) = \prod_j W(y_j \mid x_j) P ( y [ J ] ∣ x [ J ] ) = ∏ j W ( y j ∣ x j ) for a one-user channel W W W . The complete local views are then i.i.d. given C C C , and so are the identifiers, and Lemma 2 with m = J m=J m = J gives the claim. □
Two arms therefore suffice to break the symmetry, one of which may be silent, as soon as a collision changes the reward law of an arm, however little.
Interactive Schemes
a) Generalized musical chairs: Take ϵ = 0 \epsilon=0 ϵ = 0 and use J J J of the arms as the roles, with χ a = a \chi_a=a χ a = a and C a = { 1 } C_a=\{1\} C a = { 1 } : a rewarded user is certain to have been alone. Definition 9 then gives generalized musical chairs with exclusion probability β \beta β , and Corollary 4 gives a lower bound on its exponent as a spectral radius, which we compute exactly for small J J J (Tables I and Table II ); in all these cases it is the exact exponent (Appendix § 3.7 ). The parameter β = 0 \beta=0 β = 0 is musical chairs. Musical chairs is also a confidence controller (Definition 8), with Q a = δ a Q_a=\delta_a Q a = δ a , score + 1 +1 + 1 on a positive output and − 1 -1 − 1 on a zero, ξ + = ξ − = 1 \xi^+=\xi^-=1 ξ + = ξ − = 1 , and the uniform redraw kernel and initial law: a user commits at its first reward and draws a new role after every zero. The two schemes differ only in the identifier of a user that is still searching at time n n n : the last arm it played in musical chairs, and a fresh draw in the confidence controller; this does not change the exponent. For β = 0 \beta=0 β = 0 the exponent has a closed form.
Table II. Exponents (nats per use) for J = K J = K J = K users and arms, all with the same mean μ \mu μ , and ϵ = 0 \epsilon= 0 ϵ = 0 ; maximizing β \beta β of generalized musical chairs in parentheses. Bold: the largest exponent of a scheme in each column. †: valid bound found by a numerical search over β \beta β , exact at the stated β \beta β (Appendix § 3.7 ). The name-decoding ceiling is as in Table I.
Proposition 14 (Musical chairs, β = 0 \beta=0 β = 0 ). On the noiseless bandit channel with J J J arms of means μ 1 , … , μ J > 0 \mu_1,\ldots,\mu_J>0 μ 1 , … , μ J > 0 used as roles, musical chairs satisfies
lim n → ∞ − 1 n log P b ( n ) = − log ( 1 − min x μ x J ) . (100) \lim_{n\to\infty}-\frac{1}{n}\log P_b^{(n)}=-\log\left(1-\frac{\min_x\mu_x}{J}\right). \tag*{(100)} n → ∞ lim − n 1 log P b ( n ) = − log ( 1 − J min x μ x ) . (100)
Proof. Let S S S be the set of committed arms. Uncommitted users redraw uniformly, so the transition of S S S depends only on S S S , and S S S is a Markov chain that only grows. Let s S s_S s S be the probability that the state S S S is not left in one slot, that is, that no uncommitted user is rewarded. The substochastic matrix of the chain on the states S ≠ [ J ] S\ne[J] S = [ J ] , ordered by size, is triangular with diagonal entries s S s_S s S , so that P ( S n ≠ [ J ] ) P(S_n\ne[J]) P ( S n = [ J ]) is at most a constant times n J − 1 max S ≠ [ J ] s S n − J + 1 n^{J-1}\max_{S\ne[J]}s_S^{n-J+1} n J − 1 max S = [ J ] s S n − J + 1 , since a path that has not reached [ J ] [J] [ J ] at time n n n leaves states at most J − 1 J-1 J − 1 times, hence has probability at most max S ≠ [ J ] s S n − J + 1 \max_{S\ne[J]}s_S^{n-J+1} max S = [ J ] s S n − J + 1 , and there are at most a constant times n J − 1 n^{J-1} n J − 1 such paths, and, since every state is reachable, at least a constant times max S ≠ [ J ] s S n \max_{S\ne[J]}s_S^n max S = [ J ] s S n . The maximum is 1 − min x μ x / J 1-\min_x\mu_x/J 1 − min x μ x / J . It is attained by S = [ J ] ∖ { x } S=[J]\setminus\{x\} S = [ J ] ∖ { x } with a single uncommitted user, which can be rewarded only if it draws x x x , an event of probability 1 / J 1/J 1/ J , and then is rewarded with probability μ x \mu_x μ x , so that s S = 1 − μ x / J s_S=1-\mu_x/J s S = 1 − μ x / J . For any other S S S with w ≤ J w\le J w ≤ J uncommitted users and any free arm x x x , the probability that exactly one of them draws x x x is ( w 1 ) 1 J ( 1 − 1 J ) w − 1 ≥ 1 J \binom{w}{1}\frac{1}{J}(1-\frac{1}{J})^{w-1}\ge\frac{1}{J} ( 1 w ) J 1 ( 1 − J 1 ) w − 1 ≥ J 1 , since w ≤ J w\le J w ≤ J and ( w 1 ) ( 1 − 1 J ) w − 1 \binom{w}{1}(1-\frac{1}{J})^{w-1} ( 1 w ) ( 1 − J 1 ) w − 1 is nondecreasing in w ≤ J w\le J w ≤ J and equals 1 1 1 at w = 1 w=1 w = 1 ; that user is rewarded with probability μ x \mu_x μ x . Hence s S ≤ 1 − μ x / J s_S\le1-\mu_x/J s S ≤ 1 − μ x / J for every free arm x x x , and in particular s S ≤ 1 − min x μ x / J s_S\le1-\min_x\mu_x/J s S ≤ 1 − min x μ x / J .
This gives the exponent of the survival probability, which upper-bounds P b ( n ) P_b^{(n)} P b ( n ) by Proposition 10. For the lower bound, let x x x be an arm of smallest mean and S ∗ = [ J ] ∖ { x } S^*=[J]\setminus\{x\} S ∗ = [ J ] ∖ { x } . The chain reaches S ∗ S^* S ∗ and stays there until the end of slot n − 1 n-1 n − 1 with probability at least a constant times s S ∗ n − 1 s_{S^*}^{n-1} s S ∗ n − 1 . In slot n n n the single uncommitted user draws the arm of a committed user with probability ( J − 1 ) / J (J-1)/J ( J − 1 ) / J , and its identifier, the last role it drew, then coincides with that of a committed user. Hence P b ( n ) ≥ J − 1 J s S ∗ n − 1 P_b^{(n)}\ge\frac{J-1}{J}s_{S^*}^{n-1} P b ( n ) ≥ J J − 1 s S ∗ n − 1 . □
The exponent is that of the last uncommitted user, which must find its own arm, and it decays like μ μ / J \mu_\mu/J μ μ / J .
For β > 0 \beta>0 β > 0 the last uncommitted user no longer searches blindly. In musical chairs it draws among all J J J arms although J − 1 J-1 J − 1 of them are held by committed users and a failure reveals it, whereas an excluding user searches the remaining arms. The exponent is not monotone in β \beta β , and it is discontinuous at β = 0 \beta=0 β = 0 . Indeed, let β > 0 \beta>0 β > 0 . For J = 2 J=2 J = 2 , the joint state in which both users search with the same excluded arm is reachable (both draw this arm, collide, and exclude it); it is a failure state, which persists with probability ( 1 − β ) 2 (1-\beta)^2 ( 1 − β ) 2 , since both users then draw the other arm and collide. For J ≥ 3 J\ge3 J ≥ 3 , the joint state in which J − 1 J-1 J − 1 users are committed and the last user searches with all arms excluded but one, held by a committed user, is reachable; it is a failure state, which persists with probability 1 − β 1-\beta 1 − β . Hence the exponent of P b ( n ) P_b^{(n)} P b ( n ) is at most − 2 log ( 1 − β ) -2\log(1-\beta) − 2 log ( 1 − β ) for J = 2 J=2 J = 2 and − log ( 1 − β ) -\log(1-\beta) − log ( 1 − β ) for J ≥ 3 J\ge3 J ≥ 3 , which tends to zero as β ↓ 0 \beta\downarrow0 β ↓ 0 , whereas the exponent for β = 0 \beta=0 β = 0 is (100). For μ = 1 \mu=1 μ = 1 and β = 0.05 \beta=0.05 β = 0.05 these upper bounds, 0.103 0.103 0.103 for J = 2 J=2 J = 2 and 0.051 0.051 0.051 for J = 3 J=3 J = 3 , coincide with the lower bounds of Corollary 4, against 0.693 0.693 0.693 and $0.405 for musical chairs, so that β \beta β has to be tuned (the best values found are in Table II ). With the best β \beta β found on a grid, generalized musical chairs improves on musical chairs for every J ≥ 2 J\ge2 J ≥ 2 in Table II , except for μ = 1 \mu=1 μ = 1 and J = 2 J=2 J = 2 , where musical chairs already meets the converse, so that β = 0 \beta=0 β = 0 is optimal; β = 1 / 2 \beta=1/2 β = 1/2 attains the same exponent.
b) One-arm musical chairs: For J = 2 J=2 J = 2 , one informative arm together with a silent arm is enough to break symmetry. Call them arm 1, with mean μ 1 \mu_1 μ 1 , and arm 2, which is silent. Role 1 is a claimer that pulls the arm at every use and never releases; role 2 is a prober that pulls arm 1 with probability 1 / 2 1/2 1/2 and plays the silent arm otherwise. Both users start as probers. A prober that pulls and is rewarded knows that it was alone on the arm; it becomes a claimer and takes the arm for good, after which the other prober is never rewarded again. As a confidence controller, the claimer has ξ + 1 = 0 \xi^1_+=0 ξ + 1 = 0 , so that it is committed on entry, ξ − 1 = 1 \xi^1_-=1 ξ − 1 = 1 and score s 1 = 0 s_1=0 s 1 = 0 ; the prober has Q 2 Q_2 Q 2 uniform on { 1 , 2 } \{1,2\} { 1 , 2 } , ξ + 2 = ∞ \xi^2_+=\infty ξ + 2 = ∞ and ξ − 2 = 1 \xi^2_-=1 ξ − 2 = 1 , with score − 1 -1 − 1 if it played arm 1 and was rewarded, and 0 0 0 otherwise; the kernel is R ( 2 , ⋅ ) = δ 1 R(2,\cdot)=\delta_1 R ( 2 , ⋅ ) = δ 1 and π 0 = δ 2 \pi_0=\delta_2 π 0 = δ 2 .
Theorem 3 (One arm and a silent arm). Let J = 2 J=2 J = 2 , K = 2 K=2 K = 2 , ε = 0 \varepsilon=0 ε = 0 , and let arm 2 be silent, i.e., μ 2 = 0 \mu_2=0 μ 2 = 0 . One-arm musical chairs has P b ( n ) = ( 1 − μ 1 / 2 ) n P_{\mathrm{b}}^{(n)}=(1-\mu_1/2)^n P b ( n ) = ( 1 − μ 1 /2 ) n , and every symmetric causal protocol satisfies P b ( n ) ≥ 1 2 ( 1 − μ 1 / 2 ) n P_{\mathrm{b}}^{(n)}\ge\frac{1}{2}(1-\mu_1/2)^n P b ( n ) ≥ 2 1 ( 1 − μ 1 /2 ) n . Hence
1 2 ( 1 − μ 1 2 ) n ≤ ϵ n ∗ ( P ) ≤ ( 1 − μ 1 2 ) n , E ‾ b ( P ) = − log ( 1 − μ 1 2 ) . \frac{1}{2}\left(1-\frac{\mu_1}{2}\right)^n \le\epsilon_n^*(P) \le\left(1-\frac{\mu_1}{2}\right)^n,\qquad\overline{E}_{\mathrm{b}}(P)=-\log\left(1-\frac{\mu_1}{2}\right). 2 1 ( 1 − 2 μ 1 ) n ≤ ϵ n ∗ ( P ) ≤ ( 1 − 2 μ 1 ) n , E b ( P ) = − log ( 1 − 2 μ 1 ) .
Proof. The controller fails exactly when both users are probers at time n n n , since a claimer and a prober have distinct roles, and the users leave the state (prober, prober) in a slot only if exactly one pulls, probability 1 / 2 1/2 1/2 , and is rewarded, probability μ 1 \mu_1 μ 1 . Thus P b ( n ) = ( 1 − μ 1 / 2 ) n P_{\mathrm{b}}^{(n)}=(1-\mu_1/2)^n P b ( n ) = ( 1 − μ 1 /2 ) n . The converse is Proposition 12 with K = 2 K=2 K = 2 , where κ C O = 1 − μ 1 / 2 \kappa_{\mathrm{CO}}=1-\mu_1/2 κ CO = 1 − μ 1 /2 by Corollary 6 with k = 1 k=1 k = 1 , and γ 2 , 2 = 1 2 \gamma_{2,2}=\frac{1}{2} γ 2 , 2 = 2 1 .
With exactly one informative arm and a silent arm (K = 2 K=2 K = 2 ), this matches the converse. For J = 2 J=2 J = 2 , ε = 0 \varepsilon=0 ε = 0 and any channel with a silent arm, using only the best arm, one-arm musical chairs achieves − log ( 1 − μ ( 1 ) / 2 ) -\log(1-\mu_{(1)}/2) − log ( 1 − μ ( 1 ) /2 ) ; with more informative arms this is only achievability, and the converse can be larger. The controller relies on the certificate, and the silent arm is what lets a prober be alone.
c) Confidence controllers with noisy collisions: Take J = 2 J=2 J = 2 , two arms as roles with ε a < μ a \varepsilon_a<\mu_a ε a < μ a , and Q a = δ a Q_a=\delta_a Q a = δ a . A user in role a a a has score + u a +u_a + u a on a reward and − d a -d_a − d a otherwise, with integers u a , d a ≥ 1 u_a,d_a\ge1 u a , d a ≥ 1 ; its thresholds are ξ + = ⌈ ν n ⌉ \xi^+ = \lceil\nu n\rceil ξ + = ⌈ ν n ⌉ and ξ − = ℓ \xi^-=\ell ξ − = ℓ ; it redraws its role with the kernel R ( a , a ˉ ) = w R(a,\bar a)=w R ( a , a ˉ ) = w . When the roles coincide, each user is rewarded with probability ε a \varepsilon_a ε a , so that the counter drifts down if ε a u a < ( 1 − ε a ) d a \varepsilon_a u_a<(1-\varepsilon_a)d_a ε a u a < ( 1 − ε a ) d a ; when they are distinct, it drifts up if μ a u a > ( 1 − μ a ) d a \mu_a u_a>(1-\mu_a)d_a μ a u a > ( 1 − μ a ) d a . Integers u a , d a u_a,d_a u a , d a satisfying both inequalities exist precisely when ε a < μ a \varepsilon_a<\mu_a ε a < μ a , since they require ( 1 − μ a ) / μ a < u a / d a < ( 1 − ε a ) / ε a (1-\mu_a)/\mu_a<u_a/d_a<(1-\varepsilon_a)/\varepsilon_a ( 1 − μ a ) / μ a < u a / d a < ( 1 − ε a ) / ε a . We do not have a proof of a positive exponent for these controllers. We compute the error probability exactly (Proposition 9) for a chain of O ( n 2 ) O(n^2) O ( n 2 ) states, up to n = 960 n=960 n = 960 (1280 1280 1280 for ε = 0.2 \varepsilon=0.2 ε = 0.2 ), and read the exponent off the slope of − log P b ( n ) -\log P_{\mathrm{b}}^{(n)} − log P b ( n ) . Since the threshold ξ + \xi^+ ξ + grows with n n n , this exponent is a limit and not a spectral radius, and the values we report are estimates, marked ‡ in Table 1 . A truncated chain gives a rigorous upper bound on it, which exceeds the extrapolated value by less than 5 ⋅ 10 − 4 5\cdot10^{-4} 5 ⋅ 1 0 − 4 for ε ≤ 0.1 \varepsilon\le0.1 ε ≤ 0.1 and by less than 3 ⋅ 10 − 3 3\cdot10^{-3} 3 ⋅ 1 0 − 3 for ε = 0.2 \varepsilon=0.2 ε = 0.2 (Appendix G). A controller of this class is called uniform if both roles use the same parameters and the kernel and initial law are uniform (w = 1 2 w=\frac{1}{2} w = 2 1 , π 0 = Unif ( [ 2 ] ) \pi_0=\operatorname{Unif}([2]) π 0 = Unif ([ 2 ]) ); the optimized controller also optimizes the redraw probability w w w , and the score and threshold parameters. The optimized controller at ε = 0.2 \varepsilon=0.2 ε = 0.2 has w = 1 w=1 w = 1 ; its error probability is evaluated exactly in the same way.
In Table 1 the parameters ( u , d , ℓ , ν ) (u,d,\ell,\nu) ( u , d , ℓ , ν ) of the uniform controller are ( 11 , 6 , 5 , 2 ) (11,6,5,2) ( 11 , 6 , 5 , 2 ) , ( 10 , 7 , 3 , 2 ) (10,7,3,2) ( 10 , 7 , 3 , 2 ) and ( 4 , 4 , 1 , 1 ) (4,4,1,1) ( 4 , 4 , 1 , 1 ) for ε = 0.05 \varepsilon=0.05 ε = 0.05 , 0.1 0.1 0.1 and 0.2 0.2 0.2 , and those ( u , d , ℓ , ν , w , π 0 ) (u,d,\ell,\nu,w,\pi_0) ( u , d , ℓ , ν , w , π 0 ) of the optimized controller are ( 12 , 6 , 5 , 2 ; 0.7 , Unif ( [ 2 ] ) ) (12,6,5,2;0.7,\operatorname{Unif}([2])) ( 12 , 6 , 5 , 2 ; 0.7 , Unif ([ 2 ])) , ( 10 , 7 , 4 , 2 ; 0.8 , Unif ( [ 2 ] ) ) (10,7,4,2;0.8,\operatorname{Unif}([2])) ( 10 , 7 , 4 , 2 ; 0.8 , Unif ([ 2 ])) and ( 4 , 4 , 1 , 1 ; 1 , δ 1 ) (4,4,1,1;1,\delta_1) ( 4 , 4 , 1 , 1 ; 1 , δ 1 ) . The initial law changes only the constant, not the exponent: replacing δ 1 \delta_1 δ 1 by the uniform law at ε = 0.2 \varepsilon=0.2 ε = 0.2 multiplies P b ( n ) P_{\mathrm{b}}^{(n)} P b ( n ) by about 0.6 0.6 0.6 and changes the slopes by less than 10 − 4 10^{-4} 1 0 − 4 , so the gain comes from w = 1 w=1 w = 1 .
Comparison
Table 1 compares the converse of Section VI-B, block coding, and the interactive schemes for J = 2 J=2 J = 2 and K = 3 K=3 K = 3 arms, with μ = ( 0.8 , 0.8 , 0 ) \boldsymbol{\mu}=(0.8,0.8,0) μ = ( 0.8 , 0.8 , 0 ) and ε = ( ε , ε , 0 ) \boldsymbol{\varepsilon}=(\varepsilon,\varepsilon,0) ε = ( ε , ε , 0 ) , ε ∈ { 0 , 0.05 , 0.1 , 0.2 } \varepsilon\in\{0,0.05,0.1,0.2\} ε ∈ { 0 , 0.05 , 0.1 , 0.2 } , so that arm 3 is silent. The converse is for the channel with the three arms; the schemes use only the two positive arms, which can only reduce their exponents. Table II adds more users, with as many arms as users and no silent arm.
a) Interaction beats name decoding: For a name-decoding scheme (Section IV-C), Corollary 3 caps the name-decoding exponent, the exponent of P ( N n ) \mathbb{P}(\mathcal{N}_n) P ( N n ) , at I ( X − j ; Y j ∣ X j ) / ( J − 1 ) I(X_{-j};Y_j\mid X_j)/(J-1) I ( X − j ; Y j ∣ X j ) / ( J − 1 ) maximized over joint input laws, whether the scheme is block or interactive. For the noisy bandit channel this maximum is explicit.
Proposition 15 (Name-decoding ceiling). Let C ( μ , ε ) : = max c ∈ [ 0 , 1 ] [ h 2 ( c ε + ( 1 − c ) μ ) − c h 2 ( ε ) − ( 1 − c ) h 2 ( μ ) ] C(\mu,\varepsilon):=\max_{c\in[0,1]}[h_2(c\varepsilon+(1-c)\mu)-ch_2(\varepsilon)-(1-c)h_2(\mu)] C ( μ , ε ) := max c ∈ [ 0 , 1 ] [ h 2 ( c ε + ( 1 − c ) μ ) − c h 2 ( ε ) − ( 1 − c ) h 2 ( μ )] be the capacity of the binary-input channel with output laws Bern ( ε ) \operatorname{Bern}(\varepsilon) Bern ( ε ) and Bern ( μ ) \operatorname{Bern}(\mu) Bern ( μ ) . For every sequence of name-decoding schemes, the exponent of P ( N n ) \mathbb{P}(\mathcal{N}_n) P ( N n ) is at most max x ∈ [ K ] C ( μ x , ϵ x ) / ( J − 1 ) \max_{x \in[K]} C(\mu_x,\epsilon_x)/(J-1) max x ∈ [ K ] C ( μ x , ϵ x ) / ( J − 1 ) . For ϵ ≠ μ \epsilon\ne\mu ϵ = μ , C ( μ , ϵ ) = log ( 1 + e − ς ) + μ ς − h 2 ( μ ) C(\mu,\epsilon) = \log(1+e^{-\varsigma})+\mu\varsigma-h_2(\mu) C ( μ , ϵ ) = log ( 1 + e − ς ) + μ ς − h 2 ( μ ) , where ς : = ( h 2 ( ϵ ) − h 2 ( μ ) ) / ( ϵ − μ ) \varsigma:= (h_2(\epsilon)-h_2(\mu))/(\epsilon-\mu) ς := ( h 2 ( ϵ ) − h 2 ( μ )) / ( ϵ − μ ) ; in particular, C ( μ , 0 ) = log ( 1 + e − h 2 ( μ ) / μ ) C(\mu,0)=\log(1+e^{-h_2(\mu)/\mu}) C ( μ , 0 ) = log ( 1 + e − h 2 ( μ ) / μ ) , the capacity of a Z Z Z -channel; for J = 2 J=2 J = 2 and arms with equal means, this is the upper bound on the capacity of the bandit channel in [9 Cor. 2] .
Proof. Given X j = x X_j=x X j = x , the output Y j Y_j Y j depends on X − j X_{-j} X − j only through the indicator A A A of a collision, and the conditional law of Y j Y_j Y j is Bern ( ϵ x ) \operatorname{Bern}(\epsilon_x) Bern ( ϵ x ) if A = 1 A=1 A = 1 and Bern ( μ x ) \operatorname{Bern}(\mu_x) Bern ( μ x ) otherwise. Hence I ( X − j ; Y j ∣ X j ) = I ( A ; Y j ∣ X j ) ≤ max x C ( μ x , ϵ x ) I(X_{-j};Y_j\mid X_j)=I(A;Y_j\mid X_j)\le\max_x C(\mu_x,\epsilon_x) I ( X − j ; Y j ∣ X j ) = I ( A ; Y j ∣ X j ) ≤ max x C ( μ x , ϵ x ) , since given X j = x X_j=x X j = x the channel from A A A to Y j Y_j Y j has capacity C ( μ x , ϵ x ) C(\mu_x,\epsilon_x) C ( μ x , ϵ x ) , and the claim follows from Corollary 3. The closed form follows by setting to zero the derivative of the concave objective, which gives c ϵ + ( 1 − c ) μ = 1 / ( 1 + e ς ) c\epsilon+(1-c)\mu=1/(1+e^{\varsigma}) cϵ + ( 1 − c ) μ = 1/ ( 1 + e ς ) .
The ceiling is governed by the best arm alone, and it does not depend on any other arm. It bounds the name-decoding exponent of every scheme, and hence the random-coding exponent of block coding, which by (69) is even at most max x C ( μ x , ϵ x ) / J \max_x C(\mu_x,\epsilon_x)/J max x C ( μ x , ϵ x ) / J . In Table 1 the interactive schemes exceed it. For ϵ = 0 \epsilon=0 ϵ = 0 the comparison is proved: the exponent of musical chairs is a closed form, and that of generalized musical chairs is computed exactly. For J J J arms of mean μ ∈ ( 0 , 1 ) \mu\in(0,1) μ ∈ ( 0 , 1 ) , musical chairs exceeds the ceiling for every J ≥ 2 J\ge2 J ≥ 2 . Indeed, ( J − 1 ) [ − log ( 1 − μ / J ) ] (J-1)[-\log(1-\mu/J)] ( J − 1 ) [ − log ( 1 − μ / J )] is nondecreasing in J J J , since its derivative in J J J is at least μ ( 1 − μ ) / ( J ( J − μ ) ) ≥ 0 \mu(1-\mu)/(J(J-\mu))\ge0 μ ( 1 − μ ) / ( J ( J − μ )) ≥ 0 by − log ( 1 − u ) ≥ u -\log(1-u)\ge u − log ( 1 − u ) ≥ u ; and with t = 1 − μ t=1-\mu t = 1 − μ , the inequality − log ( 1 − μ / 2 ) > C ( μ , 0 ) -\log(1-\mu/2)>C(\mu,0) − log ( 1 − μ /2 ) > C ( μ , 0 ) reads ϕ ( t ) : = t log t + ( 1 − t ) log ( 1 + t ) < 0 \phi(t):=t\log t+(1-t)\log(1+t)<0 ϕ ( t ) := t log t + ( 1 − t ) log ( 1 + t ) < 0 , which holds on ( 0 , 1 ) (0,1) ( 0 , 1 ) because ϕ ( 0 + ) = ϕ ( 1 ) = 0 \phi(0+)=\phi(1)=0 ϕ ( 0 + ) = ϕ ( 1 ) = 0 and ϕ ′ ′ ( t ) = ( 1 − t ) / ( t ( 1 + t ) 2 ) > 0 \phi''(t)=(1-t)/(t(1+t)^2)>0 ϕ ′′ ( t ) = ( 1 − t ) / ( t ( 1 + t ) 2 ) > 0 . For μ = 1 \mu=1 μ = 1 the two exponents are equal for J = 2 J=2 J = 2 , and musical chairs is larger for J ≥ 3 J\ge3 J ≥ 3 . For ϵ > 0 \epsilon>0 ϵ > 0 the exponents of the controllers are numerical estimates; if they are accurate, the optimized controller exceeds the name-decoding ceiling by 14% to 22%. The comparison is least certain at ϵ = 0.2 \epsilon=0.2 ϵ = 0.2 , where the margin is smallest and only two digits of the estimate are reliable. Directly contending for roles through the channel is not a form of name decoding, and it does better.
b) Gap to the converse: For ϵ > 0 \epsilon>0 ϵ > 0 the converse is about twice the exponent of the optimized controller. Both sides are numerical at ϵ ∈ { 0.05 , 0.1 } \epsilon\in\{0.05,0.1\} ϵ ∈ { 0.05 , 0.1 } : a better search for the reference could lower the converse, and better parameters could raise the exponent of the controller, so the gap may be smaller. The optimized controller improves on the uniform one by about 10%, through the redraw probability and the score and threshold parameters. For ϵ = 0 \epsilon=0 ϵ = 0 and J = 2 J=2 J = 2 , Theorem 3 closes the gap when there is exactly one positive arm and a silent arm. The gap is also closed for two arms of mean μ = 1 \mu=1 μ = 1 without a silent arm (first column of Table II): musical chairs has P b ( n ) = 2 − n P_b^{(n)}=2^{-n} P b ( n ) = 2 − n , since the users leave their initial state exactly when they draw distinct arms, and Proposition 12 gives ϵ n ∗ ( P ) ≥ 1 2 2 − n \epsilon_n^*(P)\ge\frac{1}{2}2^{-n} ϵ n ∗ ( P ) ≥ 2 1 2 − n , so that the optimal error probability is again determined within a factor of two. With two positive arms and a silent one a gap remains between the converse and musical chairs, one-arm musical chairs, which has the same exponent at equal means, and generalized musical chairs on the two positive arms. Without the silent arm, the converse for μ = 0.8 \mu=0.8 μ = 0.8 is 0.654, against 0.631 for generalized musical chairs (Table II), so that most of the gap at ϵ = 0 \epsilon=0 ϵ = 0 in Table I is due to the silent arm, which raises the converse but which the schemes on the two positive arms do not use. Whether schemes that use both positive arms and the silent arm close the gap is open.
c) Scaling with the number of users: The m = 2 m=2 m = 2 converse bounds of Section VI-B do not depend on J J J beyond J ≥ 3 J\ge3 J ≥ 3 . The exponent of musical chairs is − log ( 1 − μ / J ) ∼ μ / J -\log(1-\mu/J)\sim\mu/J − log ( 1 − μ / J ) ∼ μ / J , the name-decoding ceiling is at most log 2 / ( J − 1 ) \log2/(J-1) log 2/ ( J − 1 ) , and the random-coding exponent of block coding is at most log 2 / J \log2/J log 2/ J . These three exponents decay like 1 / J 1/J 1/ J , but the decay is not intrinsic to interactive schemes: generalized musical chairs, which does not decode names, stays above the ceiling, and far above musical chairs, at J = 3 J=3 J = 3 and J = 4 J=4 J = 4 in Table II . Against the m = 2 m=2 m = 2 converse the gap is much smaller than the decay of musical chairs suggests, though it still widens from J = 3 J=3 J = 3 to J = 4 J=4 J = 4 as the converse grows with the number of arms. Whether the true exponent decays with J J J , and how the kernels with m > 2 m>2 m > 2 show it, is open.
Conclusion
We studied how fast the error probability of anonymous symmetry breaking can decay with the blocklength over symmetric multiway channels. A Rényi change of measure gives a converse for every symmetric causal strategy, and for every population vector with at least two positive entries. The name-decoding exponent, that is, the exponent of the probability that two names coincide or are decoded wrongly, is limited by a cut-set bound, and so is the random-coding exponent of block coding; controllers that contend for identifiers without decoding names do not face this limit. On the noisy bandit channel, interaction pays: musical chairs already provably exceeds the limit of name decoding when collisions are noiseless and the arms have a common mean below one, and generalized musical chairs, introduced here, improves substantially on musical chairs. For two users, noiseless collisions, one informative arm and a silent arm, the converse and one-arm musical chairs determine the optimal error probability up to a factor of two, and so do the converse and musical chairs for two arms of mean one. Several questions remain open.
Gap between converse and schemes. With noisy collisions the converse is about twice the exponent of the best controller we found, and with two positive arms and a silent arm it exceeds the exponent of generalized musical chairs. It is unknown whether the converse or the schemes are loose. We have no proof that the confidence controllers have a positive exponent, and it is open whether a controller with a bounded state set and time-varying maps, which can drive the error probability to zero on the noisy bandit channel, can achieve a positive exponent when every output has positive probability.
Dependence on the number of users. The converse bounds we evaluated involve m = 2 m=2 m = 2 selected users and do not depend on J J J beyond J ≥ 3 J \ge3 J ≥ 3 . The exponents of musical chairs and block coding decay like 1 / J 1/J 1/ J , and so does the ceiling on the name-decoding exponent, but the exponents of generalized musical chairs that we evaluated, for J = 3 J=3 J = 3 and J = 4 J=4 J = 4 , are much larger and too few to show a rate. How the optimal exponent depends on J J J is open, and so is whether kernels that involve m > 2 m>2 m > 2 users, which we did not evaluate, capture it.
Sharper converses. The references used here are built from kernels with a product structure over the selected users. Posterior-dependent references, and kernels with several users, may give tighter bounds.
Population vectors. Every achievability result holds for any population vector, in particular for leader election, and so do the converses, with the same kernels and a range of m m m that depends on the vector, when it has at least two positive entries (Proposition 2); otherwise the task is trivial. Whether the intermediate tasks, and leader election in particular, have strictly larger optimal exponents than the task N = ( 1 , … , 1 ) N=(1,\ldots,1) N = ( 1 , … , 1 ) is open.
Unknown channels and complexity. We assumed that the channel and the number of users are known, and that the users are synchronized. Schemes that learn the channel parameters while breaking the symmetry, universal decoding for block coding, and low-complexity implementations are of practical interest, as is the use of symmetry breaking as the initialization phase of multiplayer bandit algorithms: with horizon T T T , an initialization of length about ( log T ) / E (\log T)/E ( log T ) / E , for a scheme of exponent E E E , keeps the probability of a failed initialization of order 1 / T 1/T 1/ T , so that 1 / E 1/E 1/ E multiplies log T \log T log T in the regret.
Appendix A: Proof of the Kernel-Reference Lemmas
We prove Lemma 5 and Lemma 6. Fix π n , c \pi_{n,c} π n , c , 2 ≤ m ≤ J 2 \le m \le J 2 ≤ m ≤ J , and B ∈ A m , J ( P ) B \in\mathcal{A}_{m,J}(P) B ∈ A m , J ( P ) ; for Lemma 5 assume β B π n , c > 0 \beta_{B}^{\pi_{n,c}}>0 β B π n , c > 0 , and for Lemma 6 assume κ m ( B ) > 0 \kappa_m(B)>0 κ m ( B ) > 0 . Throughout, x j [ i ] = f i ( c , s j , y j i − 1 ) x_j[i]=f_i(c,s_j,y_j^{i-1}) x j [ i ] = f i ( c , s j , y j i − 1 ) , u [ i ] = x m + 1 : J [ i ] u[i]=x_{m+1:J}[i] u [ i ] = x m + 1 : J [ i ] and v [ i ] = y m + 1 : J [ i ] v[i]=y_{m+1:J}[i] v [ i ] = y m + 1 : J [ i ] . The seeds may take values in a general space; the outputs use counting measure.
a) Branch decomposition: For 0 ≤ t ≤ n 0 \le t \le n 0 ≤ t ≤ n , let R ( t ) R^{(t)} R ( t ) denote the measure (33) with n n n replaced by t t t , so that R ( n ) = R B π n , c R^{(n)}=R_{B}^{\pi_{n,c}} R ( n ) = R B π n , c and R ( 0 ) = P S ⊗ J R^{(0)}=P_{S}^{\otimes J} R ( 0 ) = P S ⊗ J . Expanding each factor B [ y [ i ] ∣ x [ i ] ] B[y[i]\mid x[i]] B [ y [ i ] ∣ x [ i ]] according to (34) gives R ( t ) = ∑ r t ∈ R t R r t R^{(t)}=\sum_{r^t\in\mathcal{R}^t}R_{r^t} R ( t ) = ∑ r t ∈ R t R r t with
R r t ( d s [ J ] , d y [ J ] t ) : = ∏ j = 1 J P S ( d s j ) ∏ i = 1 t [ ∏ j = 1 m a r i ( x j [ i ] , y j [ i ] ) ] c r i ( v [ i ] ∣ x [ i ] , y [ m ] [ i ] ) . (101) R_{r^t}(ds_{[J]},dy_{[J]}^t):=\prod_{j=1}^{J}P_S(ds_j)\prod_{i=1}^{t}\left[\prod_{j=1}^{m}a_{r_i}(x_j[i],y_j[i])\right]c_{r_i}(v[i]\mid x[i],y_{[m]}[i]). \tag*{(101)} R r t ( d s [ J ] , d y [ J ] t ) := j = 1 ∏ J P S ( d s j ) i = 1 ∏ t [ j = 1 ∏ m a r i ( x j [ i ] , y j [ i ]) ] c r i ( v [ i ] ∣ x [ i ] , y [ m ] [ i ]) . (101)
Define the one-user measure
λ r t ( d s , d y t ) : = P S ( d s ) ∏ i = 1 t a r i ( f i ( c , s , y i − 1 ) , y [ i ] ) , ℓ ( r t ) : = λ r t ( t o t a l ) . (102) \lambda_{r^t}(ds,dy^t):=P_S(ds)\prod_{i=1}^{t}a_{r_i}(f_i(c,s,y^{i-1}),y[i]),\qquad\ell(r^t):=\lambda_{r^t}(\mathrm{total}). \tag*{(102)} λ r t ( d s , d y t ) := P S ( d s ) i = 1 ∏ t a r i ( f i ( c , s , y i − 1 ) , y [ i ]) , ℓ ( r t ) := λ r t ( total ) . (102)
Claim. The marginal of R r t R_{r^t} R r t on the selected views ( s [ m ] , y [ m ] t ) (s_{[m]},y_{[m]}^t) ( s [ m ] , y [ m ] t ) equals ζ ( r t ) ∏ j = 1 m λ r t ( d s j , d y j t ) \zeta(r^t)\prod_{j=1}^{m}\lambda_{r^t}(ds_j,dy_j^t) ζ ( r t ) ∏ j = 1 m λ r t ( d s j , d y j t ) for a constant ζ ( r t ) ≥ 0 \zeta(r^t)\ge0 ζ ( r t ) ≥ 0 , and ζ ≡ 1 \zeta\equiv1 ζ ≡ 1 under (N).
Under (I). The factors c r i ( u [ i ] , v [ i ] ) c_{r_i}(u[i],v[i]) c r i ( u [ i ] , v [ i ]) involve only unselected seeds, inputs and outputs. The unselected inputs are functions of the unselected seeds and outputs, and the selected inputs are functions of the selected seeds and outputs. Hence R r t R_{r^t} R r t is exactly the product of ∏ j ≤ m λ r t ( d s j , d y j t ) \prod_{j\le m}\lambda_{r^t}(ds_j,dy_j^t) ∏ j ≤ m λ r t ( d s j , d y j t ) and
η r t ( d s m + 1 : J , d v t ) : = ∏ j > m P S ( d s j ) ∏ i = 1 t c r i ( u [ i ] , v [ i ] ) , \eta_{r^t}(ds_{m+1:J},dv^t):=\prod_{j>m}P_S(ds_j)\prod_{i=1}^{t}c_{r_i}(u[i],v[i]), η r t ( d s m + 1 : J , d v t ) := j > m ∏ P S ( d s j ) i = 1 ∏ t c r i ( u [ i ] , v [ i ]) ,
and the claim holds with ζ ( r t ) : = η r t ( t o t a l ) \zeta(r^t):=\eta_{r^t}(\mathrm{total}) ζ ( r t ) := η r t ( total ) .
Under (N). Integrate out the unselected outputs in the order v [ t ] , v [ t − 1 ] , … , v [ 1 ] v[t],v[t-1],\ldots,v[1] v [ t ] , v [ t − 1 ] , … , v [ 1 ] , then the unselected seeds. Suppose v [ k ] v[k] v [ k ] has already been integrated out for every k > i k>i k > i . The factors c r k c_{r_k} c r k for k > i k>i k > i have then disappeared, and they were the only factors containing the unselected inputs u [ k ] u[k] u [ k ] for k > i k>i k > i . These inputs are the only other places where v [ i ] v[i] v [ i ] appears. The one remaining factor containing v [ i ] v[i] v [ i ] is therefore c r i ( v [ i ] ∣ x [ i ] , y [ m ] [ i ] ) c_{r_i}(v[i]\mid x[i],y_{[m]}[i]) c r i ( v [ i ] ∣ x [ i ] , y [ m ] [ i ]) , which sums to one by (N). The selected factors never involve unselected variables. After all unselected variables are integrated out, what remains is ∏ j ≤ m λ r t ( d s j , d y j t ) \prod_{j\le m}\lambda_{r^t}(ds_j,dy_j^t) ∏ j ≤ m λ r t ( d s j , d y j t ) .
The claim does not assert that the unselected views are independent of the selected ones under (N), and this is not needed. In both cases the total mass of R r t R_{r^t} R r t is
w ( r t ) : = ζ ( r t ) ℓ ( r t ) m . (103) w(r^t) := \zeta(r^t)\ell(r^t)^m. \tag*{(103)} w ( r t ) := ζ ( r t ) ℓ ( r t ) m . (103)
b) Proof of Lemma 5: Let Q ^ \widehat{Q} Q be the probability law on ( r n , h [ J ] ) (r^n,h_{[J]}) ( r n , h [ J ] ) proportional to R r n ( d h [ J ] ) R_{r^n}(dh_{[J]}) R r n ( d h [ J ] ) . Its h h h -marginal is Q B π n , c Q_B^{\pi_n,c} Q B π n , c . Conditionally on r n r^n r n with w ( r n ) > 0 w(r^n)>0 w ( r n ) > 0 , the claim shows that the selected views have law ∏ j ≤ m λ r n / ℓ ( r n ) \prod_{j\le m}\lambda_{r^n}/\ell(r^n) ∏ j ≤ m λ r n / ℓ ( r n ) , so they are i.i.d. Moreover, since the seed law and the causal maps are common to both measures, B ≪ P B\ll P B ≪ P gives d R B π n , c / d P π n , c = ∏ i B ( Y [ i ] ∣ X [ i ] ) / P ( Y [ i ] ∣ X [ i ] ) dR_B^{\pi_n,c}/dP^{\pi_n,c}=\prod_i B(Y[i]\mid X[i])/P(Y[i]\mid X[i]) d R B π n , c / d P π n , c = ∏ i B ( Y [ i ] ∣ X [ i ]) / P ( Y [ i ] ∣ X [ i ]) , hence (35), and in particular Q B π n , c ≪ P π n , c Q_B^{\pi_n,c}\ll P^{\pi_n,c} Q B π n , c ≪ P π n , c . Hence Q B π n , c ∈ L ( m ) ( π n , c ; P ) Q_B^{\pi_n,c}\in\mathcal{L}^{(m)}(\pi_n,c;P) Q B π n , c ∈ L ( m ) ( π n , c ; P ) with Λ = r n \Lambda=r^n Λ = r n .
c) Proof of Lemma 6, first item: Let M t : = ∑ r t w ( r t ) M_t:=\sum_{r^t}w(r^t) M t := ∑ r t w ( r t ) , so M 0 = 1 M_0=1 M 0 = 1 and M n = β B π n , c M_n=\beta_B^{\pi_n,c} M n = β B π n , c . These are masses of unnormalized prefix measures, not marginals of Q B π n , c Q_B^{\pi_n,c} Q B π n , c . Fix a prefix r t r^t r t with w ( r t ) > 0 w(r^t)>0 w ( r t ) > 0 . Let p r t p_{r^t} p r t be the law of the next input f t + 1 ( c , S , Y t ) f_{t+1}(c,S,Y^t) f t + 1 ( c , S , Y t ) of one selected user under λ r t / ℓ ( r t ) \lambda_{r^t}/\ell(r^t) λ r t / ℓ ( r t ) . Summing a r ( x , y [ t + 1 ] ) a_{r}(x,y[t+1]) a r ( x , y [ t + 1 ]) over y [ t + 1 ] y[t+1] y [ t + 1 ] gives
ℓ ( r t , r ) = ℓ ( r t ) ∑ x p r t ( x ) A r ( x ) . (104) \ell(r^t,r)=\ell(r^t)\sum_x p_{r^t}(x)A_r(x). \tag*{(104)} ℓ ( r t , r ) = ℓ ( r t ) x ∑ p r t ( x ) A r ( x ) . (104)
Under (N), ζ ( r t , r ) = 1 = ζ ( r t ) \zeta(r^t,r)=1=\zeta(r^t) ζ ( r t , r ) = 1 = ζ ( r t ) . Under (I), let q r t q_{r^t} q r t be the law of the next unselected input tuple under η r t / ζ ( r t ) \eta_{r^t}/\zeta(r^t) η r t / ζ ( r t ) ; then ζ ( r t , r ) = ζ ( r t ) ∑ u q r t ( u ) C r ( u ) \zeta(r^t,r)=\zeta(r^t)\sum_u q_{r^t}(u)C_r(u) ζ ( r t , r ) = ζ ( r t ) ∑ u q r t ( u ) C r ( u ) . Setting C r ≡ 1 C_r\equiv1 C r ≡ 1 and letting q r t q_{r^t} q r t be arbitrary under (N), both cases give
∑ r w ( r t , r ) = w ( r t ) ∑ r [ ∑ x p r t ( x ) A r ( x ) ] m [ ∑ u q r t ( u ) C r ( u ) ] = w ( r t ) E p r t ⊗ m q r t [ b B ( X [ J ] ) ] ≥ w ( r t ) κ m ( B ) . (105) \begin{aligned}
\sum_r w(r^t,r)
&=w(r^t)\sum_r\left[\sum_x p_{r^t}(x)A_r(x)\right]^m
\left[\sum_u q_{r^t}(u)C_r(u)\right]\\
&=w(r^t)\mathbb{E}_{p_{r^t}^{\otimes m}q_{r^t}}\left[b_B(X_{[J]})\right]\ge w(r^t)\kappa_m(B). \tag*{(105)}
\end{aligned} r ∑ w ( r t , r ) = w ( r t ) r ∑ [ x ∑ p r t ( x ) A r ( x ) ] m [ u ∑ q r t ( u ) C r ( u ) ] = w ( r t ) E p r t ⊗ m q r t [ b B ( X [ J ] ) ] ≥ w ( r t ) κ m ( B ) . (105)
The last step uses (37) and the linearity of the expectation in q r t q_{r^t} q r t . The inequality (105) also holds when w ( r t ) = 0 w(r^t)=0 w ( r t ) = 0 : then R r t R_{r^t} R r t is the zero measure, so every extension ( r t , r ) (r^t,r) ( r t , r ) has w ( r t , r ) = 0 w(r^t,r)=0 w ( r t , r ) = 0 . Summing over r t r^t r t gives M t + 1 ≥ κ m ( B ) M t M_{t+1}\ge\kappa_m(B)M_t M t + 1 ≥ κ m ( B ) M t , and iterating proves β B π n , c ≥ κ m ( B ) n \beta_B^{\pi_n,c}\ge\kappa_m(B)^n β B π n , c ≥ κ m ( B ) n .
d) Proof of Lemma 6, second item: Let g i : = B ( Y [ i ] ∣ X [ i ] ) / P ( Y [ i ] ∣ X [ i ] ) g_i:=B(Y[i]\mid X[i])/P(Y[i]\mid X[i]) g i := B ( Y [ i ] ∣ X [ i ]) / P ( Y [ i ] ∣ X [ i ]) and F n : = ∏ i = 1 n g i F_n:=\prod_{i=1}^n g_i F n := ∏ i = 1 n g i , which is d R B π n , c / d P π n , c dR_B^{\pi_n,c}/dP^{\pi_n,c} d R B π n , c / d P π n , c , so that d Q B π n , c / d P π n , c = F n / β B π n , c dQ_B^{\pi_n,c}/dP^{\pi_n,c}=F_n/\beta_B^{\pi_n,c} d Q B π n , c / d P π n , c = F n / β B π n , c by (35). Under P π n , c P^{\pi_n,c} P π n , c , given the seeds and all past outputs, the current input tuple X [ i ] X[i] X [ i ] is fixed and Y [ i ] ∼ P ( ⋅ ∣ X [ i ] ) Y[i]\sim P(\cdot\mid X[i]) Y [ i ] ∼ P ( ⋅ ∣ X [ i ]) . Hence, for 1 < α < ∞ 1<\alpha<\infty 1 < α < ∞ ,
E [ g i α ∣ S [ J ] , Y [ J ] i − 1 ] = ∑ y : P ( y ∣ X [ i ] ) > 0 B ( y ∣ X [ i ] ) α P ( y ∣ X [ i ] ) 1 − α ≤ Ψ α , m ( B ; P ) . (106) \mathbb{E}\left[g_i^\alpha\mid S_{[J]},Y^{i-1}_{[J]}\right]
=\sum_{y:P(y\mid X[i])>0}B(y\mid X[i])^\alpha P(y\mid X[i])^{1-\alpha}
\le\Psi_{\alpha,m}(B;P). \tag*{(106)} E [ g i α ∣ S [ J ] , Y [ J ] i − 1 ] = y : P ( y ∣ X [ i ]) > 0 ∑ B ( y ∣ X [ i ] ) α P ( y ∣ X [ i ] ) 1 − α ≤ Ψ α , m ( B ; P ) . (106)
By the tower property, E [ F n α ] ≤ Ψ α , m ( B ; P ) n \mathbb{E}[F_n^\alpha]\le\Psi_{\alpha,m}(B;P)^n E [ F n α ] ≤ Ψ α , m ( B ; P ) n . The first item then gives
D α ( Q B π n , c ∥ P π n , c ) = log E [ F n α ] − α log β B π n , c α − 1 ≤ n log Ψ α , m ( B ; P ) − α n log κ m ( B ) α − 1 = n d α , m r e f ( B ∥ P ) . (107) D_\alpha(Q_B^{\pi_n,c}\mathbin\Vert P^{\pi_n,c})
=\frac{\log\mathbb{E}[F_n^\alpha]-\alpha\log\beta_B^{\pi_n,c}}{\alpha-1}
\le\frac{n\log\Psi_{\alpha,m}(B;P)-\alpha n\log\kappa_m(B)}{\alpha-1}
=n\,d^{\mathrm{ref}}_{\alpha,m}(B\mathbin\Vert P). \tag*{(107)} D α ( Q B π n , c ∥ P π n , c ) = α − 1 log E [ F n α ] − α log β B π n , c ≤ α − 1 n log Ψ α , m ( B ; P ) − α n log κ m ( B ) = n d α , m ref ( B ∥ P ) . (107)
At order infinity, F n ≤ ρ ∞ , m ( B ; P ) n F_n\le\rho_{\infty,m}(B;P)^n F n ≤ ρ ∞ , m ( B ; P ) n , and hence
d Q B π n , c d P π n , c ≤ [ ρ ∞ , m ( B ; P ) κ m ( B ) ] n . \frac{dQ_B^{\pi_n,c}}{dP^{\pi_n,c}}
\le\left[\frac{\rho_{\infty,m}(B;P)}{\kappa_m(B)}\right]^n. d P π n , c d Q B π n , c ≤ [ κ m ( B ) ρ ∞ , m ( B ; P ) ] n .
The bounds E [ F n α ] ≤ Ψ α , m ( B ; P ) n \mathbb{E}[F_n^\alpha]\le\Psi_{\alpha,m}(B;P)^n E [ F n α ] ≤ Ψ α , m ( B ; P ) n and F n ≤ ρ ∞ , m ( B ; P ) n F_n\le\rho_{\infty,m}(B;P)^n F n ≤ ρ ∞ , m ( B ; P ) n use neither (I) nor (N); admissibility enters only through Lemma 5 and the first item.
Appendix B: Proof of the Block-Coding Exponent
We prove Theorem 2, conditioning on distinct names M 1 , … , M J M_1,\ldots,M_J M 1 , … , M J . If receiver j j j errs, some unordered set A ≠ { M k : k ≠ j } \mathcal{A}\ne\{M_k:k\ne j\} A = { M k : k = j } satisfies L j ( A ) ≥ L j ( { M k : k ≠ j } ) L_j(\mathcal{A})\ge L_j(\{M_k:k\ne j\}) L j ( A ) ≥ L j ({ M k : k = j }) in (58). The users whose names are missing from A \mathcal{A} A form a nonempty set S ⊆ [ J ] ∖ { j } \mathcal{S}\subseteq[J]\setminus\{j\} S ⊆ [ J ] ∖ { j } , and A \mathcal{A} A consists of the names of the users in S c ∖ { j } \mathcal{S}^c\setminus\{j\} S c ∖ { j } together with s = ∣ S ∣ s=|\mathcal{S}| s = ∣ S ∣ names outside { M 1 , … , M J } \{M_1,\ldots,M_J\} { M 1 , … , M J } . Every error is therefore of the following form.
Fix receiver j j j , a nonempty set S ⊆ [ J ] ∖ { j } \mathcal{S}\subseteq[J]\setminus\{j\} S ⊆ [ J ] ∖ { j } of users whose names are replaced, that is, wrongly decoded, and s = ∣ S ∣ s=|\mathcal{S}| s = ∣ S ∣ . The receiver input and the J − 1 − s J-1-s J − 1 − s correctly retained foreign codewords form X S c n X_{\mathcal{S}^c}^n X S c n . The number of unordered competing sets differing in exactly these s s s names is
( M − J s ) ≤ ( M − J ) s . (108) \binom{M-J}{s}\le(M-J)^s. \tag*{(108)} ( s M − J ) ≤ ( M − J ) s . (108)
Channel symmetry (Lemma 1) makes any assignment of the new codewords to the positions in S \mathcal{S} S equivalent. With the likelihood
W j n ( y j n ∣ x [ J ] n ) = ∏ i = 1 n W j ( y j [ i ] ∣ x [ J ] [ i ] ) . (109) W_j^n(y_j^n\mid x_{[J]}^n)=\prod_{i=1}^n W_j(y_j[i]\mid x_{[J]}[i]). \tag*{(109)} W j n ( y j n ∣ x [ J ] n ) = i = 1 ∏ n W j ( y j [ i ] ∣ x [ J ] [ i ]) . (109)
and 0 ≤ ρ ≤ 1 0 \le\rho\le1 0 ≤ ρ ≤ 1 , the indicator of an ML error on S \mathcal{S} S , the event that some candidate set that differs from the true one exactly in the names of S \mathcal{S} S has a likelihood at least that of the true set, satisfies
1 { M L e r r o r o n S } ≤ ( ∑ m ~ S [ W j n ( Y j n ∣ X S n ( m ~ S ) ⊔ S X S c n ) W j n ( Y j n ∣ X [ J ] n ) ] 1 1 + ρ ) ρ . (110) 1\{\mathrm{ML\ error\ on\ }\mathcal{S}\}\leq\left(\sum_{\widetilde{m}_{\mathcal{S}}}\left[\frac{W_j^n(Y_j^n\mid X_{\mathcal{S}}^n(\widetilde{m}_{\mathcal{S}})\sqcup_{\mathcal{S}}X_{\mathcal{S}^c}^n)}{W_j^n(Y_j^n\mid X_{[J]}^n)}\right]^{\frac{1}{1+\rho}}\right)^\rho. \tag*{(110)} 1 { ML error on S } ≤ m S ∑ [ W j n ( Y j n ∣ X [ J ] n ) W j n ( Y j n ∣ X S n ( m S ) ⊔ S X S c n ) ] 1 + ρ 1 ρ . (110)
where the sum is taken over one representative tuple m ~ S \widetilde{m}_{\mathcal{S}} m S of s s s distinct unused names per candidate set. Indeed, on the error event at least one summand is at least one; off the event the right side is nonnegative.
Concavity of t ↦ t ρ t\mapsto t^\rho t ↦ t ρ gives the factor ( M − J ) s ρ (M-J)^{s\rho} ( M − J ) s ρ : conditioned on the transmitted codewords and on Y j n Y_j^n Y j n , the competing codewords are i.i.d. with law Q X n Q_X^n Q X n , because distinct names index independently generated codewords, so Jensen’s inequality and (108) move the expectation inside ( ⋅ ) ρ (\cdot)^\rho ( ⋅ ) ρ . Averaging the candidate tuple, the true tuple, the conditioned correct codewords, and the output yields the one-letter quantity
∑ x S c , y j ( ∏ k ∈ S c Q X ( x k ) ) [ ∑ x ~ S ( ∏ k ∈ S Q X ( x ~ k ) ) W j ( y j ∣ x ~ S ⊔ S x S c ) 1 1 + ρ ] 1 + ρ . (111) \sum_{x_{\mathcal{S}^c},y_j}\left(\prod_{k\in\mathcal{S}^c}Q_X(x_k)\right)\left[\sum_{\widetilde{x}_{\mathcal{S}}}\left(\prod_{k\in\mathcal{S}}Q_X(\widetilde{x}_k)\right)W_j(y_j\mid\widetilde{x}_{\mathcal{S}}\sqcup_{\mathcal{S}}x_{\mathcal{S}^c})^{\frac{1}{1+\rho}}\right]^{1+\rho}. \tag*{(111)} x S c , y j ∑ ( k ∈ S c ∏ Q X ( x k ) ) [ x S ∑ ( k ∈ S ∏ Q X ( x k ) ) W j ( y j ∣ x S ⊔ S x S c ) 1 + ρ 1 ] 1 + ρ . (111)
This is e − E 0 , s ( ρ , Q X ) e^{-E_{0,s}(\rho,Q_X)} e − E 0 , s ( ρ , Q X ) . Product coding and memorylessness raise it to the n n n -th power. There are ( J − 1 s ) \binom{J-1}{s} ( s J − 1 ) choices of S \mathcal{S} S , so the union bound over S \mathcal{S} S , with ρ \rho ρ optimized separately for each s s s , proves (61).
For M = ⌈ e n R ⌉ M=\lceil e^{nR}\rceil M = ⌈ e n R ⌉ , optimize each term over ρ \rho ρ , sum over the J J J receivers, and choose a deterministic codebook no worse than the ensemble average. Adding the collision term from (17) proves (62). Taking − 1 n log -\frac{1}{n}\log − n 1 log and optimizing over R R R gives part (iii).
For completeness, differentiating (59) at zero can be done by writing its inner sum as an exponential moment of log W j ( Y j ∣ X [ J ] ) \log W_j(Y_j\mid X_{[J]}) log W j ( Y j ∣ X [ J ] ) . At ρ = 0 \rho=0 ρ = 0 the normalizing expression is one; the derivative of its negative logarithm is
∑ x [ J ] , y j Q X J ( x [ J ] ) W j ( y j ∣ x [ J ] ) log W j ( y j ∣ x [ J ] ) ∑ x ~ S Q X s ( x ~ S ) W j ( y j ∣ x ~ S ⊔ S x S c ) = I ( X S ; Y j ∣ X S c ) , (112) \sum_{x_{[J]},y_j}Q_X^J(x_{[J]})W_j(y_j\mid x_{[J]})\log\frac{W_j(y_j\mid x_{[J]})}{\sum_{\widetilde{x}_{\mathcal{S}}}Q_X^s(\widetilde{x}_{\mathcal{S}})W_j(y_j\mid\widetilde{x}_{\mathcal{S}}\sqcup_{\mathcal{S}}x_{\mathcal{S}^c})}
=I(X_{\mathcal{S}};Y_j\mid X_{\mathcal{S}^c}), \tag*{(112)} x [ J ] , y j ∑ Q X J ( x [ J ] ) W j ( y j ∣ x [ J ] ) log ∑ x S Q X s ( x S ) W j ( y j ∣ x S ⊔ S x S c ) W j ( y j ∣ x [ J ] ) = I ( X S ; Y j ∣ X S c ) , (112)
which verifies (65).
We prove Lemma 7. Let
I s = I ( X [ s ] ; Y ∣ X [ m ] ∖ [ s ] , T ) , 1 ≤ s ≤ m . (113) I_s=I(X_{[s]};Y\mid X_{[m]\setminus[s]},T),\qquad1\leq s\leq m. \tag*{(113)} I s = I ( X [ s ] ; Y ∣ X [ m ] ∖ [ s ] , T ) , 1 ≤ s ≤ m . (113)
Fix s < m s<m s < m and put
U = ( Y , T , X s + 2 , … , X m ) . (114) U=(Y,T,X_{s+2},\ldots,X_m). \tag*{(114)} U = ( Y , T , X s + 2 , … , X m ) . (114)
Because the inputs are i.i.d. and independent of T T T ,
I s = s H ( X 1 ) − H ( X 1 , … , X s ∣ X s + 1 , U ) , (115) I_s=sH(X_1)-H(X_1,\ldots,X_s\mid X_{s+1},U), \tag*{(115)} I s = sH ( X 1 ) − H ( X 1 , … , X s ∣ X s + 1 , U ) , (115)
I s + 1 = ( s + 1 ) H ( X 1 ) − H ( X 1 , … , X s + 1 ∣ U ) . (116) I_{s+1}=(s+1)H(X_1)-H(X_1,\ldots,X_{s+1}\mid U). \tag*{(116)} I s + 1 = ( s + 1 ) H ( X 1 ) − H ( X 1 , … , X s + 1 ∣ U ) . (116)
Let A = H ( X 1 , … , X s ∣ X s + 1 , U ) A=H(X_1,\ldots,X_s\mid X_{s+1},U) A = H ( X 1 , … , X s ∣ X s + 1 , U ) . By the chain rule,
H ( X 1 , … , X s + 1 ∣ U ) = H ( X s + 1 ∣ U ) + A . (117) H(X_1,\ldots,X_{s+1}\mid U)=H(X_{s+1}\mid U)+A. \tag*{(117)} H ( X 1 , … , X s + 1 ∣ U ) = H ( X s + 1 ∣ U ) + A . (117)
Moreover,
A ≤ ∑ i = 1 s H ( X i ∣ X s + 1 , U ) ≤ ∑ i = 1 s H ( X i ∣ U ) = s H ( X s + 1 ∣ U ) . (118) \begin{aligned}
A\leq\sum_{i=1}^{s}H(X_i\mid X_{s+1},U) \\
&\leq\sum_{i=1}^{s}H(X_i\mid U)=sH(X_{s+1}\mid U). \tag*{(118)}
\end{aligned} A ≤ i = 1 ∑ s H ( X i ∣ X s + 1 , U ) ≤ i = 1 ∑ s H ( X i ∣ U ) = sH ( X s + 1 ∣ U ) . (118)
The final equality in (118) follows from exchangeability. Indeed, the inputs are i.i.d., and the conditional law of Y Y Y is invariant under permutations of X 1 , … , X s + 1 X_1,\ldots,X_{s+1} X 1 , … , X s + 1 . Combining (117) and (118) yields
H ( X 1 , … , X s + 1 ∣ U ) s + 1 ≥ A s . (119) \frac{H(X_1,\ldots,X_{s+1}\mid U)}{s+1}\geq\frac{A}{s}. \tag*{(119)} s + 1 H ( X 1 , … , X s + 1 ∣ U ) ≥ s A . (119)
Subtract the two sides from H ( X 1 ) H(X_1) H ( X 1 ) using (115)–(116) to obtain I s / s ≥ I s + 1 / ( s + 1 ) I_s/s\geq I_{s+1}/(s+1) I s / s ≥ I s + 1 / ( s + 1 ) .
Proof of the Multiway Cut-Set Bound
We prove Proposition 4. Let M 1 , … , M J M_1,\ldots,M_J M 1 , … , M J be the independent uniform names, each on a set of size ⌈ e n R ⌉ \lceil e^{nR}\rceil ⌈ e n R ⌉ , and Γ j = ( M j , Θ j ) \Gamma_j=(M_j,\Theta_j) Γ j = ( M j , Θ j ) the name together with the private randomness of user j j j . Then Γ 1 , … , Γ J \Gamma_1,\ldots,\Gamma_J Γ 1 , … , Γ J are independent, and independent of the common randomness C C C , and X j [ i ] X_j[i] X j [ i ] is a function of ( C , Γ j , Y i − 1 ) (C,\Gamma_j,Y^{i-1}) ( C , Γ j , Y i − 1 ) . Until (124), all entropies and mutual informations are conditional on C C C , which we omit from the notation. Fix a nonempty proper subset S ⊂ [ J ] S\subset[J] S ⊂ [ J ] , and let M S \mathcal{M}_S M S be the multiset of the names of the users in S S S . Every terminal knows its own name and estimates the multiset of all names, so that the terminals of S c S^c S c , viewed collectively, know M S c \mathcal{M}_{S^c} M S c and can recover M S \mathcal{M}_S M S from ( Γ S c , Y S c n ) (\Gamma_{S^c},Y^n_{S^c}) ( Γ S c , Y S c n ) with vanishing error probability. The multiset loses at most the order of the ∣ S ∣ |S| ∣ S ∣ names, so that
H ( M S ∣ Γ S c ) ≥ ∣ S ∣ log ⌈ e n R ⌉ − log ∣ S ∣ ! ≥ ∣ S ∣ n R − log ∣ S ∣ ! . (120) H(\mathcal{M}_S\mid\Gamma_{S^c})\ge|S|\log\lceil e^{nR}\rceil-\log|S|!\ge|S|nR-\log|S|!. \tag*{(120)} H ( M S ∣ Γ S c ) ≥ ∣ S ∣ log ⌈ e n R ⌉ − log ∣ S ∣ ! ≥ ∣ S ∣ n R − log ∣ S ∣ ! . (120)
Fano’s inequality gives
H ( M S ∣ Γ S c , Y S c n ) ≤ n ϵ n , ϵ n → 0. (121) H(\mathcal{M}_S\mid\Gamma_{S^c},Y^n_{S^c})\le n\epsilon_n,\qquad\epsilon_n\to0. \tag*{(121)} H ( M S ∣ Γ S c , Y S c n ) ≤ n ϵ n , ϵ n → 0. (121)
Since M S \mathcal{M}_S M S is a function of Γ S \Gamma_S Γ S ,
n ∣ S ∣ R − log ∣ S ∣ ! ≤ I ( Γ S ; Y S c n ∣ Γ S c ) + n ϵ n = ∑ i = 1 n I ( Γ S ; Y S c [ i ] ∣ Γ S c , Y S c i − 1 ) + n ϵ n . (122) \begin{aligned}
n|S|R-\log|S|!\le I(\Gamma_S;Y^n_{S^c}\mid\Gamma_{S^c})+n\epsilon_n \\
&=\sum_{i=1}^{n}I(\Gamma_S;Y_{S^c}[i]\mid\Gamma_{S^c},Y_{S^c}^{i-1})+n\epsilon_n. \tag*{(122)}
\end{aligned} n ∣ S ∣ R − log ∣ S ∣ ! ≤ I ( Γ S ; Y S c n ∣ Γ S c ) + n ϵ n = i = 1 ∑ n I ( Γ S ; Y S c [ i ] ∣ Γ S c , Y S c i − 1 ) + n ϵ n . (122)
We bound each term by a difference of entropies. The input X S c [ i ] X_{S^c}[i] X S c [ i ] is a function of ( Γ S c , Y S c i − 1 ) (\Gamma_{S^c},Y_{S^c}^{i-1}) ( Γ S c , Y S c i − 1 ) , and conditioning reduces entropy, so that
H ( Y S c [ i ] ∣ Γ S c , Y S c i − 1 ) ≤ H ( Y S c [ i ] ∣ X S c [ i ] ) . H(Y_{S^c}[i]\mid\Gamma_{S^c},Y_{S^c}^{i-1})\le H(Y_{S^c}[i]\mid X_{S^c}[i]). H ( Y S c [ i ] ∣ Γ S c , Y S c i − 1 ) ≤ H ( Y S c [ i ] ∣ X S c [ i ]) .
Also X [ J ] [ i ] X_{[J]}[i] X [ J ] [ i ] is a function of ( Γ [ J ] , Y [ J ] i − 1 ) (\Gamma_{[J]},Y_{[J]}^{i-1}) ( Γ [ J ] , Y [ J ] i − 1 ) , and the channel is memoryless, so that
H ( Y S c [ i ] ∣ Γ [ J ] , Y S c i − 1 ) ≥ H ( Y S c [ i ] ∣ Γ [ J ] , Y [ J ] i − 1 ) = H ( Y S c [ i ] ∣ X [ J ] [ i ] ) . H(Y_{S^c}[i]\mid\Gamma_{[J]},Y_{S^c}^{i-1})\ge H(Y_{S^c}[i]\mid\Gamma_{[J]},Y_{[J]}^{i-1})=H(Y_{S^c}[i]\mid X_{[J]}[i]). H ( Y S c [ i ] ∣ Γ [ J ] , Y S c i − 1 ) ≥ H ( Y S c [ i ] ∣ Γ [ J ] , Y [ J ] i − 1 ) = H ( Y S c [ i ] ∣ X [ J ] [ i ]) .
Subtracting,
I ( Γ S ; Y S c [ i ] ∣ Γ S c , Y S c i − 1 ) ≤ I ( X S [ i ] ; Y S c [ i ] ∣ X S c [ i ] ) , (123) I(\Gamma_S;Y_{S^c}[i]\mid\Gamma_{S^c},Y_{S^c}^{i-1})\le I(X_S[i];Y_{S^c}[i]\mid X_{S^c}[i]), \tag*{(123)} I ( Γ S ; Y S c [ i ] ∣ Γ S c , Y S c i − 1 ) ≤ I ( X S [ i ] ; Y S c [ i ] ∣ X S c [ i ]) , (123)
where the right-hand side is evaluated under the joint law of X [ J ] [ i ] X_{[J]}[i] X [ J ] [ i ] induced by the protocol; this law is arbitrary, since the inputs may be dependent through the past outputs.
Let U U U be uniform on [ n ] [n] [ n ] , independent of the experiment, and define ( X [ J ] , Y [ J ] ) = ( X [ J ] [ U ] , Y [ J ] [ U ] ) (X_{[J]},Y_{[J]})=(X_{[J]}[U],Y_{[J]}[U]) ( X [ J ] , Y [ J ] ) = ( X [ J ] [ U ] , Y [ J ] [ U ]) . Divide (122) by n n n , use (123), and make the conditioning on C C C explicit again, to obtain, for the scheme of blocklength n n n ,
∣ S ∣ R ≤ I ( X S ; Y S c ∣ X S c , U , C ) + ϵ n + log ∣ S ∣ ! n . (124) |S|R\le I(X_S;Y_{S^c}\mid X_{S^c},U,C)+\epsilon_n+\frac{\log|S|!}{n}. \tag*{(124)} ∣ S ∣ R ≤ I ( X S ; Y S c ∣ X S c , U , C ) + ϵ n + n log ∣ S ∣ ! . (124)
Since ( U , C ) − X [ J ] − Y [ J ] (U,C)-X_{[J]}-Y_{[J]} ( U , C ) − X [ J ] − Y [ J ] is a Markov chain, I ( X S ; Y S c ∣ X S c , U , C ) = H ( Y S c ∣ X S c , U , C ) − H ( Y S c ∣ X [ J ] ) ≤ I ( X S ; Y S c ∣ X S c ) I(X_S;Y_{S^c}\mid X_{S^c},U,C)=H(Y_{S^c}\mid X_{S^c},U,C)-H(Y_{S^c}\mid X_{[J]})\le I(X_S;Y_{S^c}\mid X_{S^c}) I ( X S ; Y S c ∣ X S c , U , C ) = H ( Y S c ∣ X S c , U , C ) − H ( Y S c ∣ X [ J ] ) ≤ I ( X S ; Y S c ∣ X S c ) , where the right-hand side is evaluated under the law P ‾ n \overline{P}_n P n of X [ J ] X_{[J]} X [ J ] , averaged over U U U and C C C . Hence ∣ S ∣ R ≤ I P ‾ n ( X S ; Y S c ∣ X S c ) + η n |S|R\le I_{\overline{P}_n}(X_S;Y_{S^c}\mid X_{S^c})+\eta_n ∣ S ∣ R ≤ I P n ( X S ; Y S c ∣ X S c ) + η n for every S S S , with η n → 0 \eta_n\to0 η n → 0 . The laws on X X X form a compact set, and for a fixed channel the mutual information is continuous in the input law, so that along a subsequence P ‾ n \overline{P}_n P n converges to a law P X [ J ] P_{X_{[J]}} P X [ J ] that satisfies (70) for every S S S . For S = [ J ] ∖ { j } S=[J]\setminus\{j\} S = [ J ] ∖ { j } , (71) follows, and I ( X − j ; Y j ∣ X j ) ≤ H ( Y j ) ≤ log ∣ Y ∣ I(X_{-j};Y_j\mid X_j)\le H(Y_j)\le\log|\mathcal{Y}| I ( X − j ; Y j ∣ X j ) ≤ H ( Y j ) ≤ log ∣ Y ∣ .
Bounded Memory with Time-Varying Maps
The floor of Proposition 8 needs stationarity. The following controller uses the device that Cover used for hypothesis testing between two Bernoulli laws with finite memory [34 ] .
Proposition 16 (Bounded memory with time-varying maps). On the noisy bandit channel of Definition 10 with J J J arms of common parameters 0 ≤ ε < μ ≤ 1 0\le\varepsilon<\mu\le1 0 ≤ ε < μ ≤ 1 , in particular with 0 < ε < μ < 1 0<\varepsilon<\mu<1 0 < ε < μ < 1 , where every output has positive probability, there is a role controller with 4 J 4J 4 J states for every n n n , whose maps vary with time, and whose error probability tends to zero as n → ∞ n\to\infty n → ∞ .
Proof. The state of a user is ( a , A , F ) ∈ [ J ] × { 0 , 1 } 2 (a,A,F)\in[J]\times\{0,1\}^2 ( a , A , F ) ∈ [ J ] × { 0 , 1 } 2 , initially ( 1 , 1 , 0 ) (1,1,0) ( 1 , 1 , 0 ) ; the user plays arm a a a , and its identifier is a a a . Time is divided into epochs, and epoch k k k consists of N k : = ⌈ k μ − k ⌉ N_k:=\lceil k\mu^{-k}\rceil N k := ⌈ k μ − k ⌉ windows of k k k channel uses. Within a window, the flag A A A (for all ) records whether all outputs so far equal 1, and the flag F F F (for found ) records whether a window of ones has been found in the current epoch; at the end of an epoch, F ← max { F , A } F\gets\max\{F,A\} F ← max { F , A } and A ← 1 A\gets1 A ← 1 ; at the end of an epoch, a user with F = 0 F=0 F = 0 draws a new role uniformly in [ J ] [J] [ J ] , and F ← 0 F\gets0 F ← 0 . The maps depend on the time index only through this schedule, and the roles do not change during an epoch. A user alone on its arm finds no window of ones with probability ( 1 − μ k ) N k ≤ e − k (1-\mu^k)^{N_k}\le e^{-k} ( 1 − μ k ) N k ≤ e − k , and a user that shares its arm finds one with probability at most δ k : = N k ε k ≤ k ( ε / μ ) k + ε k \delta_k:=N_k\varepsilon^k\le k(\varepsilon/\mu)^k+\varepsilon^k δ k := N k ε k ≤ k ( ε / μ ) k + ε k . If the roles of epoch k k k are distinct, they remain unchanged with probability at least 1 − J e − k 1-Je^{-k} 1 − J e − k . Otherwise, with probability at least 1 − J ( δ k + e − k ) 1-J(\delta_k+e^{-k}) 1 − J ( δ k + e − k ) exactly the users that share a role redraw, and the c c c users that redraw take the c c c free roles with probability c ! / J c ≥ J ! / J J c!/J^c \ge J!/J^J c ! / J c ≥ J ! / J J . Hence the probability p k p_k p k that the roles of epoch k k k are not distinct satisfies p k + 1 ≤ p k [ 1 − J ! J J ( 1 − J δ k − J e − k ) ] + J e − k p_{k+1} \le p_k\left[1-\frac{J!}{J^J}(1-J\delta_k-Je^{-k})\right]+Je^{-k} p k + 1 ≤ p k [ 1 − J J J ! ( 1 − J δ k − J e − k ) ] + J e − k , so that p k → 0 p_k \to0 p k → 0 , and P b ( n ) ≤ max { p k , p k + 1 } P_b^{(n)} \le\max\{p_k,p_{k+1}\} P b ( n ) ≤ max { p k , p k + 1 } when time n n n lies in epoch k k k .
In the construction the error probability decays only subexponentially in n n n , polynomially when μ < 1 \mu<1 μ < 1 . Indeed, for μ < 1 \mu<1 μ < 1 a user alone on its arm misses every window of epoch k k k with probability of order e − k e^{-k} e − k and then collides with probability at least 1 / J 1/J 1/ J , while epoch k k k ends at a time of order k 2 μ − k k^2\mu^{-k} k 2 μ − k ; for μ = 1 \mu=1 μ = 1 all users start on the same role, a redraw of c ≥ 2 c\ge2 c ≥ 2 users fails with probability at least 1 − c ! / J c ≥ 1 / 2 1-c!/J^c\ge1/2 1 − c ! / J c ≥ 1/2 , and epoch k k k ends at a time of order k 3 k^3 k 3 . The floor of Proposition 8 is therefore a property of stationary finite-state controllers: a controller with a bounded state set may use the time index as a free clock, which Remark 4 counts as memory.
Appendix F: Evaluation of the Noisy Bandit Gallager Function
We prove Proposition 13. Fix the receiver j j j , an error subset of size s s s , and a realization X j = x X_j=x X j = x . Let c = J − 1 − s c=J-1-s c = J − 1 − s correct foreign codewords remain fixed. By (85) , the local law W ( ⋅ ∣ x , x − j ) W(\cdot\mid x,x_{-j}) W ( ⋅ ∣ x , x − j ) depends on the foreign inputs only through whether at least one of them equals x x x : it is W μ x W_{\mu_x} W μ x if none does and W ε x W_{\varepsilon_x} W ε x otherwise, where W ν ( 1 ) = ν W_\nu(1)=\nu W ν ( 1 ) = ν and W ν ( 0 ) = 1 − ν W_\nu(0)=1-\nu W ν ( 0 ) = 1 − ν ; for a silent arm both laws are δ 0 \delta_0 δ 0 .
Each of the c c c correct inputs equals x x x with probability q x q_x q x . If at least one does, which has probability
1 − ( 1 − q x ) c (125) 1-(1-q_x)^c \tag*{(125)} 1 − ( 1 − q x ) c (125)
the likelihood is W ε x ( y ) W_{\varepsilon_x}(y) W ε x ( y ) for every candidate tuple, so that the inner sum in (59) is W ε x ( y ) 1 / ( 1 + ρ ) W_{\varepsilon_x}(y)^{1/(1+\rho)} W ε x ( y ) 1/ ( 1 + ρ ) , its ( 1 + ρ ) (1+\rho) ( 1 + ρ ) -th power summed over y ∈ { 0 , 1 } y\in\{0,1\} y ∈ { 0 , 1 } is W ε x ( 0 ) + W ε x ( 1 ) = 1 W_{\varepsilon_x}(0)+W_{\varepsilon_x}(1)=1 W ε x ( 0 ) + W ε x ( 1 ) = 1 , and the outer average contributes (125) .
Suppose none of the correct inputs equals x x x , which has probability ( 1 − q x ) c (1-q_x)^c ( 1 − q x ) c . A random candidate tuple avoids x x x with probability
θ x = ( 1 − q x ) s . (126) \theta_x=(1-q_x)^s. \tag*{(126)} θ x = ( 1 − q x ) s . (126)
If it does, the likelihood of the output y y y is W μ x ( y ) W_{\mu_x}(y) W μ x ( y ) ; otherwise it is W ε x ( y ) W_{\varepsilon_x}(y) W ε x ( y ) . The inner sum for the output y y y is therefore
θ x W μ x ( y ) 1 1 + ρ + ( 1 − θ x ) W ε x ( y ) 1 1 + ρ . (127) \theta_x W_{\mu_x}(y)^{\frac{1}{1+\rho}}+(1-\theta_x)W_{\varepsilon_x}(y)^{\frac{1}{1+\rho}}. \tag*{(127)} θ x W μ x ( y ) 1 + ρ 1 + ( 1 − θ x ) W ε x ( y ) 1 + ρ 1 . (127)
Raise (127) to the power 1 + ρ 1+\rho 1 + ρ , sum over y ∈ { 0 , 1 } y\in\{0,1\} y ∈ { 0 , 1 } , multiply by ( 1 − q x ) c (1-q_x)^c ( 1 − q x ) c , add (125) , and finally average over x ∼ Q x\sim Q x ∼ Q . The result is exactly G ρ , s ( J ) ( Q ; μ , ε ) G_{\rho,s}^{(J)}(Q;\mu,\varepsilon) G ρ , s ( J ) ( Q ; μ , ε ) in (94) . Taking the negative logarithm proves Proposition 13. For J = 2 J=2 J = 2 , s = 1 s=1 s = 1 and c = 0 c=0 c = 0 it is (97) . Differentiating at ρ = 0 \rho=0 ρ = 0 and using (65) gives (96) and (99) : for s = J − 1 s=J-1 s = J − 1 and c = 0 c=0 c = 0 , (127) is the mixture, with weight θ x \theta_x θ x on W μ x W_{\mu_x} W μ x , of the two output laws of a receiver that is alone or in collision, and the derivative is the mutual information between the collision indicator and the output given X j = x X_j=x X j = x , averaged over x ∼ Q x\sim Q x ∼ Q .
Appendix G: Numerical Evaluation of the Tables
The entries of Tables I and II are of three kinds.
a ) a) a ) Exact entries (unmarked): The common-output converse, musical chairs, and the converse and the name-decoding ceiling of Table II are closed forms; the noninteraction converse and the ceiling of Table 1 are one-dimensional convex problems, solved to 10 − 12 10^{-12} 1 0 − 12 . Block coding is maximized over ρ \rho ρ and over the laws on the two opposite arms in Table 1 , where the maximum is at the uniform law and ρ = 1 \rho=1 ρ = 1 , and restricted to uniform laws in Table II . At a given β \beta β , the spectral radius of generalized musical chairs is computed to about 10 − 10 10^{-10} 1 0 − 10 on a chain of at most 6560 states, and − log ρ ( K β ) -\log\rho(K_\beta) − log ρ ( K β ) is then the exponent of P b ( n ) P_b^{(n)} P b ( n ) itself. For μ = 1 \mu=1 μ = 1 , a user that is still searching after the first channel use collided at its last draw, so that every joint state outside G \mathcal{G} G is a failure state. For μ = 0.8 \mu=0.8 μ = 0.8 , the class that attains the spectral radius contains failure states in each of its cyclic classes, as we checked on a chain whose states include the last roles drawn, and Proposition 10 applies.
b ) b) b ) Bounds found by a numerical search († \dagger † ): These are valid bounds, and a better search can only improve them. Each optimized reference is an explicit feasible kernel: of order ∞ \infty ∞ with 8 components at ε = 0.05 \varepsilon=0.05 ε = 0.05 , rescaled so that B ≤ P B\le P B ≤ P holds exactly, and of order 1.25 with 6 components at ε = 0.1 \varepsilon=0.1 ε = 0.1 ; Ψ α , 2 ( B ; P ) \Psi_{\alpha,2}(B;P) Ψ α , 2 ( B ; P ) and κ 2 ( B ) \kappa_2(B) κ 2 ( B ) are recomputed from the kernel, κ 2 ( B ) \kappa_2(B) κ 2 ( B ) by an exact minimization over the simplex. The search is not convex. Its restart agrees to 10 − 8 10^{-8} 1 0 − 8 at the orders that define the entries, and refining the grid of orders lowered the entry at ε = 0.1 \varepsilon=0.1 ε = 0.1 from 0.653 to 0.652. At ε = 0 \varepsilon=0 ε = 0 and ε = 0.2 \varepsilon=0.2 ε = 0.2 the search does not improve on the closed forms, which are the entries. The best β \beta β of generalized musical chairs is searched on successively finer grids; refining them to 10 − 4 10^{-4} 1 0 − 4 changes no digit shown, and for μ = 0.8 \mu=0.8 μ = 0.8 and J ∈ { 3 , 4 } J\in\{3,4\} J ∈ { 3 , 4 } the exponent increases up to β = 1 \beta=1 β = 1 on the grid.
c) Estimates (‡): The commit threshold ⌈ ν n ⌉ \lceil\nu n\rceil ⌈ ν n ⌉ of the confidence controllers grows with n n n , so that Proposition 10 does not apply, and we have no proof that their exponent exists. Proposition 9 gives P b ( n ) P_b^{(n)} P b ( n ) exactly, which we compute up to n = 960 n=960 n = 960 ($1280$ for ε = 0.2 \varepsilon=0.2 ε = 0.2 ); it agrees with Monte Carlo simulation (4 ⋅ 10 6 4\cdot10^{6} 4 ⋅ 1 0 6 runs at n = 4 , 8 , … , 40 n=4,8,\ldots,40 n = 4 , 8 , … , 40 ) within 2.72 binomial standard deviations. The slopes of − log P b ( n ) -\log P_b^{(n)} − log P b ( n ) decrease beyond n = 200 n=200 n = 200 , and least-squares fits of − log P b ( n ) = E n + c log n + a + b / n -\log P_b^{(n)}=En+c\log n+a+b/n − log P b ( n ) = E n + c log n + a + b / n extrapolate them; these fits assume the form of the correction, so they give estimates. For an upper bound, let K L ∘ K_L^\circ K L ∘ be the transition matrix of the controller in which no user commits and a path is killed when a counter exceeds L L L , restricted to the joint states reachable from the initial ones. Both rewards have probabilities in ( 0 , 1 ) (0,1) ( 0 , 1 ) , so that the support graph of K L ∘ K_L^\circ K L ∘ is the direct product of two copies of that of one user, which is strongly connected, and every component of this product is closed and strongly connected. For every L L L we used, the joint states reachable from the initial ones form one such component, so that K L ∘ K_L^\circ K L ∘ is irreducible (with period 2 for the optimized controller at ε = 0.2 \varepsilon=0.2 ε = 0.2 , aperiodic otherwise), and it has failure states. For n > L / ν n>L/\nu n > L / ν the paths on which no counter exceeds L L L have the same probability in both chains, so that lim inf n → ∞ − 1 n log P b ( n ) ≤ − log ρ ( K L ∘ ) \liminf_{n\to\infty}-\frac{1}{n}\log P_b^{(n)}\leq-\log\rho(K_L^\circ) lim inf n → ∞ − n 1 log P b ( n ) ≤ − log ρ ( K L ∘ ) by the Perron–Frobenius theorem. We bracket ρ ( K L ∘ ) \rho(K_L^\circ) ρ ( K L ∘ ) by the Collatz–Wielandt inequalities, with a positive vector obtained by power iteration, to a relative width below 10 − 6 10^{-6} 1 0 − 6 ; standard eigenvalue solvers fail on this far-from-normal chain. Table I shows the extrapolation, rounded, and the upper bound exceeds it by less than 5 ⋅ 10 − 4 5\cdot10^{-4} 5 ⋅ 1 0 − 4 for ε ≤ 0.1 \varepsilon\leq0.1 ε ≤ 0.1 and by less than 3 ⋅ 10 − 3 3\cdot10^{-3} 3 ⋅ 1 0 − 3 for $\varepsilon=0.2 (Table III ), where both still decrease at the largest n n n and L L L . The parameters of the controllers were tuned numerically, so that a better controller of this form may exist.
Table III. Confidence controllers of Table I: slope of − log P b ( n ) -\log P_b^{(n)} − log P b ( n ) on [ 240 , 320 ] [240,320] [ 240 , 320 ] and on the last window ([ 800 , 960 ] [800,960] [ 800 , 960 ] , or [ 960 , 1280 ] [960,1280] [ 960 , 1280 ] for ε = 0.2 \varepsilon=0.2 ε = 0.2 ), range of the extrapolations, upper bound − log ρ ( K L ∘ ) -\log\rho(K_L^\circ) − log ρ ( K L ∘ ) at the largest L L L , and entry of Table I.
Acknowledgment
The authors thank Henrique Miyamoto for helpful discussions on the problem formulation and the error-exponent analysis, and for his careful reading of the manuscript and his constructive suggestions.
References