Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Capacity-Maximizing Input Symbol Selection for Discrete Memoryless Channels

Maximilian Egger, Rawad Bitar, Antonia Wachter-Zeh, Deniz Gündüz and Nir Weinberger M.E., R.B. and A.W-Z. are with the Technical University of Munich. Emails: {maximilian.egger, rawad.bitar, antonia.wachter-zeh}@tum.de. D.G. is with Imperial College London. Email: d.gunduz@imperial.ac.uk. N.W. is at Technion — Israel Institute of Technology. Email: nirwein@technion.ac.il.This project has received funding from the German Research Foundation (DFG) under Grant Agreement Nos. BI 2492/1-1 and WA 3907/7-1, and from UKRI for project AI-R (ERC-Consolidator Grant, EP/X030806/1). The work of N.W. was partly supported by the Israel Science Foundation (ISF), grant no. 1782/22.
Abstract

Motivated by communication systems with constrained complexity, we consider the problem of input symbol selection for discrete memoryless channels (DMCs). Given a DMC, the goal is to find a subset of its input alphabet, so that the optimal input distribution that is only supported on these symbols maximizes the capacity among all other subsets of the same size (or smaller). We observe that the resulting optimization problem is non-concave and non-submodular, and so generic methods for such cases do not have theoretical guarantees. We derive an analytical upper bound on the capacity loss when selecting a subset of input symbols based only on the properties of the transition matrix of the channel. We propose a selection algorithm that is based on input-symbols clustering, and an appropriate choice of representatives for each cluster, which uses the theoretical bound as a surrogate objective function. We provide numerical experiments to support the findings.

I Introduction

We study the long-standing problem of reducing the input alphabet size of a Discrete Memoryless Channel (DMC) with input alphabet 𝒳𝒳\mathcal{X}caligraphic_X to a set of k<|𝒳|𝑘𝒳k<|\mathcal{X}|italic_k < | caligraphic_X | symbols (and possibly k|𝒳|much-less-than𝑘𝒳k\ll|\mathcal{X}|italic_k ≪ | caligraphic_X |), which are carefully selected to maximize the capacity of the resulting channel. A natural motivation for this problem is that an input alphabet of controlled cardinality allows to control the complexity of the transmitter and receiver. Furthermore, when the channel transition probability function is unknown, the restriction to a subset of the input symbols may reduce the cost of estimating the effective transition probability function. This possibility is outlined in, e.g. [1], where the goal was to identify the maximal-capacity channel among a set of candidate channels, through adaptive exploration. If it can be determined during the exploration phase that capacity can be achieved without using some of the input symbols (while not knowing the capacity exactly at this stage), then this reduces the cost of accurately estimating the capacity during the rest of the exploration phase.

The problem of input selection has been studied for the special case of conditionally Gaussian channels in [2], and in the context of Multiple-Input Multiple-Output (MIMO) channels, e.g., [3]. In the latter, the authors show submodularity for the problem of antenna subset selection for MIMO. This is useful, since the submodularity property leads to theoretical guarantees on the capacity achieved by greedy algorithms. Nonetheless, as we show, an analogous submodularity property does not hold for the inputs of DMCs, and so does not lead to direct performance guarantees on greedy algorithms. In [4] the binomial channel was considered, whose input alphabet is the continuous interval [0,1]01[0,1][ 0 , 1 ], and an efficient algorithm for finding the finitely supported capacity-achieving input distribution was proposed (called Dynamic Blahut-Arimoto). The algorithm was recently generalized to the multinomial channel in [5]).

The papers that consider DMCs, and hence, that are closest to ours, are [6] and [7]. The authors of [6] stress that among different formulations of the problem, the standard formulation of capacity through the maximization of the mutual information is the most interesting from an information-theoretic perspective, but conclude that it is challenging to efficiently solve this formulation. Hence, they instead focus on optimizing the symbol error rate or the cut-off rate. The papers [8, 7] considered capacity, but also argued that achieving theoretical guarantees is challenging, and so focused on numerical approaches based on the Blahut-Arimoto algorithm.

In this work, we revisit the problem of selecting input symbols for capacity maximization, and take a principled intermediate approach between generic greedy optimization methods, and high-complexity exhaustive-search optimal methods. Based on the properties of the transition matrix of a DMC, we first derive bounds on the loss in capacity incurred by only using a selected subset of the input symbols for transmission. We then use this bound as a surrogate measure in designing an algorithm for input symbol selection, and show the effectiveness of the proposed algorithm in various scenarios. Interestingly, our algorithm operates without computing the original channel’s capacity (with full use of input symbols). This is useful in case the input alphabet is very large and accurate computation of the capacity is computationally demanding.

Informally, our algorithm is based on clustering of similar rows of the channel transition matrix. The novelty is in the choice of cluster representatives: Our upper bound depends on the subset of the output alphabet’s probability simplex covered by the transition probabilities of the selected input symbols. Thus, the algorithm chooses the representatives to maximize this subset, and thus, to reduce the loss in capacity compared to the full usage of the input symbols. Such clustering points in the probability simplex have been studied, e.g., in [9], but our algorithm is tailored to maximize the mutual information, and hence differs from general-purpose choices.

II Problem Formulation

Notation

Random variables are denoted by capital script letters ZZ\mathrm{Z}roman_Z, their realizations by small letters, and sets by calligraphic letters 𝒵𝒵\mathcal{Z}caligraphic_Z. The entropy of random variable ZZ\mathrm{Z}roman_Z over alphabet 𝒵𝒵\mathcal{Z}caligraphic_Z is denoted by H(Z):=z𝒵Pr(Z=z)log(Pr(Z=z))assignHZsubscript𝑧𝒵PrZ𝑧PrZ𝑧\mathrm{H}\left(\mathrm{Z}\right):=-\sum_{z\in\mathcal{Z}}\Pr(\mathrm{Z}=z)% \log(\Pr(\mathrm{Z}=z))roman_H ( roman_Z ) := - ∑ start_POSTSUBSCRIPT italic_z ∈ caligraphic_Z end_POSTSUBSCRIPT roman_Pr ( roman_Z = italic_z ) roman_log ( roman_Pr ( roman_Z = italic_z ) ). All logarithms are taken to the natural base unless stated otherwise. The probability simplex over the alphabet 𝒵𝒵\mathcal{Z}caligraphic_Z is denoted by 𝒫(𝒵)𝒫𝒵\mathcal{P}(\mathcal{Z})caligraphic_P ( caligraphic_Z ). The KL-divergence between distributions P𝑃Pitalic_P and Q𝑄Qitalic_Q is denoted by D(PQ)Dconditional𝑃𝑄\mathrm{D}\left(P\|Q\right)roman_D ( italic_P ∥ italic_Q ), the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-divergence by χ2(P,Q)superscript𝜒2𝑃𝑄\chi^{2}\left(P,Q\right)italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_P , italic_Q ), and the Jensen-Shannon divergence by JSD(PQ):=12D(PM)+12D(QM),assignJSDconditional𝑃𝑄12Dconditional𝑃𝑀12Dconditional𝑄𝑀\text{JSD}(P\|Q):=\frac{1}{2}\mathrm{D}\left(P\|M\right)+\frac{1}{2}\mathrm{D}% \left(Q\|M\right),JSD ( italic_P ∥ italic_Q ) := divide start_ARG 1 end_ARG start_ARG 2 end_ARG roman_D ( italic_P ∥ italic_M ) + divide start_ARG 1 end_ARG start_ARG 2 end_ARG roman_D ( italic_Q ∥ italic_M ) , where M=P+Q2𝑀𝑃𝑄2M=\frac{P+Q}{2}italic_M = divide start_ARG italic_P + italic_Q end_ARG start_ARG 2 end_ARG. For an integer τ𝜏\tauitalic_τ, [τ]:={1,,τ}assigndelimited-[]𝜏1𝜏[\tau]:=\{1,\dots,\tau\}[ italic_τ ] := { 1 , … , italic_τ }.

We consider a DMC W𝑊Witalic_W with input alphabet 𝒳𝒳\mathcal{X}caligraphic_X and output alphabet 𝒴𝒴\mathcal{Y}caligraphic_Y to be an indexed set of its conditional probability mass functions W={WY|X=x}x𝒳𝑊subscriptsubscript𝑊conditionalYX𝑥𝑥𝒳W=\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{X}}italic_W = { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT, or, alternatively, the transition matrix WY|Xsubscript𝑊conditionalYXW_{\mathrm{Y}|\mathrm{X}}italic_W start_POSTSUBSCRIPT roman_Y | roman_X end_POSTSUBSCRIPT. When we use only a subset \mathcal{R}caligraphic_R of the input symbols, we conveniently refer to the channel as W:={WY|X=x}xassignsubscript𝑊subscriptsubscript𝑊conditionalYX𝑥𝑥W_{\mathcal{R}}:=\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT := { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_R end_POSTSUBSCRIPT, for some 𝒳𝒳\mathcal{R}\subset\mathcal{X}caligraphic_R ⊂ caligraphic_X. The mutual information between the input distribution PXsubscript𝑃XP_{\mathrm{X}}italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT and the output distribution PYsubscript𝑃𝑌P_{Y}italic_P start_POSTSUBSCRIPT italic_Y end_POSTSUBSCRIPT induced by the channel WY|Xsubscript𝑊conditionalYXW_{\mathrm{Y}|\mathrm{X}}italic_W start_POSTSUBSCRIPT roman_Y | roman_X end_POSTSUBSCRIPT is denoted by I(PX;WY|X)Isubscript𝑃Xsubscript𝑊conditionalYX\mathrm{I}\left(P_{\mathrm{X}};W_{\mathrm{Y}|\mathrm{X}}\right)roman_I ( italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ; italic_W start_POSTSUBSCRIPT roman_Y | roman_X end_POSTSUBSCRIPT ). The capacity of channel WY|Xsubscript𝑊conditionalYXW_{\mathrm{Y}|\mathrm{X}}italic_W start_POSTSUBSCRIPT roman_Y | roman_X end_POSTSUBSCRIPT can be written as the maximization of the mutual information over PXsubscript𝑃XP_{\mathrm{X}}italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT, i.e.,

C(W)=maxPXP(𝒳)I(PX;WY|X).𝐶𝑊subscriptsubscript𝑃X𝑃𝒳Isubscript𝑃Xsubscript𝑊conditionalYX\displaystyle C(W)=\max_{P_{\mathrm{X}}\in P(\mathcal{X})}\mathrm{I}\left(P_{% \mathrm{X}};W_{\mathrm{Y}|\mathrm{X}}\right).italic_C ( italic_W ) = roman_max start_POSTSUBSCRIPT italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ∈ italic_P ( caligraphic_X ) end_POSTSUBSCRIPT roman_I ( italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ; italic_W start_POSTSUBSCRIPT roman_Y | roman_X end_POSTSUBSCRIPT ) .

While the optimizer PXsuperscriptsubscript𝑃XP_{\mathrm{X}}^{\star}italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT to this optimization problem is not unique, any optimizer induces the unique capacity-achieving output distribution QYsuperscriptsubscript𝑄YQ_{\mathrm{Y}}^{\star}italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT [10, Corollary 2, Thm 4.5.1]. Throughout the paper, we additionally make use of the dual problem of this optimization problem, also known as the minimax capacity theorem [11, 12], which states that

C(W)=minQY𝒫(𝒴)maxx𝒳D(WY|X=xQY),C𝑊subscriptsubscript𝑄Y𝒫𝒴subscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥subscript𝑄Y\mathrm{C}(W)=\min_{Q_{\mathrm{Y}}\in{\cal P}({\cal Y})}\max_{x\in{\cal X}}% \mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}\right),roman_C ( italic_W ) = roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) , (1)

whose minimizer is the unique capacity-achieving output distribution QYsuperscriptsubscript𝑄YQ_{\mathrm{Y}}^{\star}italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT. Furthermore, by the convexity of the optimization problem, the KL-divergence D(WY|X=xQY)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) for all input symbols x𝒳𝑥𝒳x\in\mathcal{X}italic_x ∈ caligraphic_X equals C(W)𝐶𝑊C(W)italic_C ( italic_W ) if PX(x)>0superscriptsubscript𝑃X𝑥0P_{\mathrm{X}}^{\star}(x)>0italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_x ) > 0 and is at most C(W)𝐶𝑊C(W)italic_C ( italic_W ) if PX(x)=0superscriptsubscript𝑃X𝑥0P_{\mathrm{X}}^{\star}(x)=0italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_x ) = 0 [10, Thm 4.5.1]. Hence, the capacity is the information radius of the collection of conditional distributions.

We focus on the following questions: How to select a subset 𝒳superscript𝒳\mathcal{R}^{*}\subset\mathcal{X}caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT ⊂ caligraphic_X of input symbols, such that PXsuperscriptsubscript𝑃XP_{\mathrm{X}}^{\star}italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT is supported on superscript\mathcal{R}^{*}caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT, ||ksuperscript𝑘|\mathcal{R}^{*}|\leq k| caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT | ≤ italic_k and the capacity is maximized among all other choice of \mathcal{R}caligraphic_R, ||k𝑘|\mathcal{R}|\leq k| caligraphic_R | ≤ italic_k? How to quantify the capacity loss compared to the channel that uses all symbols in 𝒳𝒳\mathcal{X}caligraphic_X, i.e., C(W)C(W)𝐶𝑊𝐶subscript𝑊superscriptC(W)-C(W_{\mathcal{R}^{*}})italic_C ( italic_W ) - italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT )? In order to demonstrate the difficulty of the problem, we next present two approaches that are commonly used to lower the complexity of such optimization problems, yet we show that both fail to achieve that (at least not in the direct manner that we have considered).

First, we may consider a relaxation of the constraint. We note that the constraint in the primal formulation by limiting the input distribution PXsubscript𝑃XP_{\mathrm{X}}italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT to a support of size k𝑘kitalic_k, i.e.,

C(W)=maxPXP(𝒳):PX0=kI(PX;WY|X),𝐶𝑊subscript:subscript𝑃X𝑃𝒳subscriptnormsubscript𝑃X0𝑘Isubscript𝑃Xsubscript𝑊conditionalYX\displaystyle C(W)=\max_{P_{\mathrm{X}}\in P(\mathcal{X}):\|P_{\mathrm{X}}\|_{% 0}=k}\mathrm{I}\left(P_{\mathrm{X}};W_{\mathrm{Y}|\mathrm{X}}\right),italic_C ( italic_W ) = roman_max start_POSTSUBSCRIPT italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ∈ italic_P ( caligraphic_X ) : ∥ italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ∥ start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = italic_k end_POSTSUBSCRIPT roman_I ( italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ; italic_W start_POSTSUBSCRIPT roman_Y | roman_X end_POSTSUBSCRIPT ) ,

where x0:=limp0i|x|passignsubscriptnorm𝑥0subscript𝑝0subscript𝑖superscript𝑥𝑝\|x\|_{0}:=\lim_{p\rightarrow 0}\sum_{i}|x|^{p}∥ italic_x ∥ start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT := roman_lim start_POSTSUBSCRIPT italic_p → 0 end_POSTSUBSCRIPT ∑ start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT | italic_x | start_POSTSUPERSCRIPT italic_p end_POSTSUPERSCRIPT is known as the L0subscript𝐿0L_{0}italic_L start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT-pseudonorm. This formulation limits the support of PXsubscript𝑃XP_{\mathrm{X}}italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT to k𝑘kitalic_k, but is non-concave and NP-hard in general [13]. A common approach is to relax the L0subscript𝐿0L_{0}italic_L start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT constraint to an L1subscript𝐿1L_{1}italic_L start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT constraint (or higher order). However, even such relaxed constraint still results in a non-concave optimization problem.

Second, we may consider showing that this problem is submodular, since this facilitates various optimization tools with theoretical guarantees [14, 15], and specifically, guarantees on the loss of greedy algorithms. To that end, recall the following definition of a submodular set function:

Definition 1 (Submodular Set Functions).

Consider a set \mathcal{L}caligraphic_L, and let 2superscript22^{\mathcal{L}}2 start_POSTSUPERSCRIPT caligraphic_L end_POSTSUPERSCRIPT be its power set. A function f:2:𝑓superscript2f:2^{\mathcal{L}}\rightarrow\mathbb{R}italic_f : 2 start_POSTSUPERSCRIPT caligraphic_L end_POSTSUPERSCRIPT → blackboard_R is submodular if and only if for every 𝒥𝒦𝒥𝒦\mathcal{J}\subset\mathcal{K}\subset\mathcal{L}caligraphic_J ⊂ caligraphic_K ⊂ caligraphic_L and an element {𝒦}𝒦\ell\in\mathcal{L}\setminus\{\mathcal{K}\}roman_ℓ ∈ caligraphic_L ∖ { caligraphic_K }, it holds that

f(𝒥{})f(𝒥)f(𝒦{})f(𝒦).𝑓𝒥𝑓𝒥𝑓𝒦𝑓𝒦\displaystyle f(\mathcal{J}\cup\{\ell\})-f(\mathcal{J})\geq f(\mathcal{K}\cup% \{\ell\})-f(\mathcal{K}).italic_f ( caligraphic_J ∪ { roman_ℓ } ) - italic_f ( caligraphic_J ) ≥ italic_f ( caligraphic_K ∪ { roman_ℓ } ) - italic_f ( caligraphic_K ) .

This version of the definition of submodularity is based on the diminishing returns property. Informally, adding an element to a small set 𝒥𝒥\mathcal{J}caligraphic_J increases the value function by at least as much as adding the same element to a larger superset 𝒦𝒦\mathcal{K}caligraphic_K. However, we have the following:

Observation 1.

The capacity of a DMC is not submodular in the set of input symbols.

We show this observation via two counterexamples. First, trivially, the diminishing return property breaks down when 𝒥=𝒥\mathcal{J}=\emptysetcaligraphic_J = ∅, since the capacity can never increase by adding any symbol to an empty set. Second, assume that the conditional entropies H(WY|X=x)Hsubscript𝑊conditionalYX𝑥\mathrm{H}\left(W_{\mathrm{Y}|\mathrm{X}=x}\right)roman_H ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ) are equal for all x𝒳𝑥𝒳x\in\mathcal{X}italic_x ∈ caligraphic_X and, hence, capacity is achieved by maximizing H(Y)HY\mathrm{H}\left(\mathrm{Y}\right)roman_H ( roman_Y ) through balancing the output distribution. Indeed, suppose that there are 4444 input symbols, such that the first and second symbols (resp. third and fourth) are complementary, in the sense that a uniform mixture of their corresponding rows is a uniform distribution in 𝒫(𝒴)𝒫𝒴\mathcal{P}(\mathcal{Y})caligraphic_P ( caligraphic_Y ) (or some other high-entropy distribution). In this case, given the first two symbols, adding the third symbol will "unbalance" the output distribution and reduce its entropy, and thus the mutual information, while adding the fourth symbol will "re-balance" it, and again, increase that entropy. Evidently, this contradicts the submodularity condition.

Consequently, submodular properties and guarantees for greedy optimization algorithms cannot be directly exploited for the problem of channel input symbol selection. We note in passing that the capacity of a DMC may still fulfill approximate notions of submodularity [16, 17, 18], but we leave an investigation of this possibility for further research.

III Theoretical Guarantees on Input Symbol Selection

In this section, we present our main theoretical guarantees on the selection of input symbols \mathcal{R}caligraphic_R for DMCs that maximize the capacity of the resulting channel. When exploring the importance of input symbols for maximizing the capacity of a channel, it is natural to examine the interplay between the conditional distributions given by the rows of the channel’s transition matrix. We will concentrate our analysis on the convex hull spanned by a certain subset of the symbols’ conditional distributions, i.e., the transition matrix generated by \mathcal{R}caligraphic_R, and its relation to unused symbols’ distributions. Depending on the shape of the channel, such symbols 𝒳𝒳\mathcal{X}\setminus\mathcal{R}caligraphic_X ∖ caligraphic_R can potentially be pruned without or with only a minor loss in the capacity of the channel. We next formally define a convex hull of a channel based on a subset of the symbols 𝒳𝒳\mathcal{R}\subset\mathcal{X}caligraphic_R ⊂ caligraphic_X:

Definition 2 (Convex Hull of a Channel).

Let \mathcal{R}caligraphic_R be a set of symbols that form a channel W:={WY|X=r}rassignsubscript𝑊subscriptsubscript𝑊conditionalYX𝑟𝑟W_{\mathcal{R}}:=\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT := { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT. The convex hull of the channel is defined as

Conv(W):={rcrWY|X=r}𝐜𝒫()𝒫(𝒴).assignConvsubscript𝑊subscriptsubscript𝑟subscript𝑐𝑟subscript𝑊conditionalYX𝑟𝐜𝒫𝒫𝒴\displaystyle\operatorname{Conv}\left(W_{\mathcal{R}}\right):=\left\{\sum_{r% \in\mathcal{R}}c_{r}W_{\mathrm{Y}|\mathrm{X}=r}\right\}_{\mathbf{c}\in\mathcal% {P}(\mathcal{R})}\subseteq\mathcal{P}(\mathcal{Y}).roman_Conv ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) := { ∑ start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT italic_c start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT bold_c ∈ caligraphic_P ( caligraphic_R ) end_POSTSUBSCRIPT ⊆ caligraphic_P ( caligraphic_Y ) .

Having this definition at hand, we start with the special case of selecting input symbols when the input alphabet 𝒳𝒳\mathcal{X}caligraphic_X is large compared to the output alphabet 𝒴𝒴\mathcal{Y}caligraphic_Y. We later move to the cases in which the alphabet sizes are of the same order, and investigate when input selection can be done without a loss in capacity, and how to bound the loss in capacity otherwise.

III-A Symbol Selection Without Capacity Loss

When the input alphabet of a DMC is large compared to its output alphabet, the convex hull of the channel may be the entire output simplex. Indeed, it is well known that, due to Carathéodory’s theorem, the capacity-achieving output distribution QYConv(W)superscriptsubscript𝑄YConv𝑊Q_{\mathrm{Y}}^{\star}\in\operatorname{Conv}\left(W\right)italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ∈ roman_Conv ( italic_W ) can be written as a convex combination of at most |𝒴|𝒴|\mathcal{Y}|| caligraphic_Y | extreme points (conditional distributions) corresponding to inputs 𝒳𝒳\mathcal{E}\subset\mathcal{X}caligraphic_E ⊂ caligraphic_X. Thus, independently of the transition matrix, there exists a set of input symbols \mathcal{E}caligraphic_E with cardinality at most |𝒴|𝒴|\mathcal{Y}|| caligraphic_Y |, which achieves the capacity [10, Corollary 3, Thm 4.5.1].

Even if the number of input symbols |𝒳|𝒳|\mathcal{X}|| caligraphic_X | of a DMC is at most the size of the output alphabet |𝒴|𝒴|\mathcal{Y}|| caligraphic_Y |, we can potentially utilize the properties of the convex hull of a channel to prune some of the input symbols without losing in capacity. It can be found that this applies when the conditional distribution of a symbol x𝑥xitalic_x lies within the convex hull of the channel spanned by a subset of the symbols \mathcal{R}caligraphic_R. In general, symbol x𝒳𝑥𝒳x\in\mathcal{X}italic_x ∈ caligraphic_X can be removed from 𝒳𝒳\mathcal{X}caligraphic_X without loss in capacity when WY|X=xsubscript𝑊conditionalYX𝑥W_{\mathrm{Y}|\mathrm{X}=x}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT is a convex combination of the channels of the remaining symbols. We formalize and generalize this statement as follows:

Proposition 1.

Consider a channel W={WY|X=x}x𝒳𝑊subscriptsubscript𝑊conditionalYX𝑥𝑥𝒳W=\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{X}}italic_W = { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT, where the input symbols 𝒳𝒳\mathcal{X}caligraphic_X are partitioned into two disjoint sets 𝒰𝒰\mathcal{U}caligraphic_U and \mathcal{R}caligraphic_R, such that the conditional distributions WY|X=usubscript𝑊conditionalYX𝑢W_{\mathrm{Y}|\mathrm{X}=u}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_u end_POSTSUBSCRIPT of the symbols in u𝒰𝑢𝒰u\in\mathcal{U}italic_u ∈ caligraphic_U are contained in the convex hull of conditional distributions of the remaining symbols in r𝑟r\in\mathcal{R}italic_r ∈ caligraphic_R. Then, the symbols in 𝒰𝒰\mathcal{U}caligraphic_U can be removed from the input alphabet without incurring a loss in capacity, that is,

C({WY|X=x}x𝒳)=C({WY|X=r}r).𝐶subscriptsubscript𝑊conditionalYX𝑥𝑥𝒳𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟\displaystyle C(\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{X}})=C(\{W_{% \mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}).italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT ) = italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ) .

Hence, there exists a capacity-achieving input distribution for which PX(x)=0subscript𝑃X𝑥0P_{\mathrm{X}}(x)=0italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ( italic_x ) = 0 for x𝒳𝑥𝒳x\in\mathcal{X}\setminus\mathcal{R}italic_x ∈ caligraphic_X ∖ caligraphic_R and PX(x)0subscript𝑃X𝑥0P_{\mathrm{X}}(x)\geq 0italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ( italic_x ) ≥ 0 for x𝑥x\in\mathcal{R}italic_x ∈ caligraphic_R.

Consequently, keeping those symbols in the channel that form the convex hull suffices to achieve the capacity. Formally, let W𝑊Witalic_W be a channel over the input alphabet 𝒳𝒳\mathcal{X}caligraphic_X. There exists a capacity-achieving input distribution supported only on the input symbols that span the convex hull of W𝑊Witalic_W.

III-B Bounding the Capacity Loss

Removing symbols from the input alphabet that span the convex hull of a DMC will likely lead to a loss in capacity. Note that this might not be the case; e.g., in the case where the input alphabet is larger than the output alphabet, we can always write the capacity-achieving output distribution as convex combination of at most |𝒴|𝒴|\mathcal{Y}|| caligraphic_Y | extreme points, in this case the input symbols’ conditional distributions. Those points do not necessarily span the entire convex hull of the full channel. Hence, this special case is not covered by Proposition 1, yet does not lead to a capacity loss.

By knowing the distance from the conditional distributions associated with symbols in \mathcal{R}caligraphic_R to those in 𝒳𝒳\mathcal{X}\setminus\mathcal{R}caligraphic_X ∖ caligraphic_R, one can bound the capacity loss incurred by restricting the input distribution to be supported only on the symbols in \mathcal{R}caligraphic_R instead of the entire alphabet 𝒳𝒳\mathcal{X}caligraphic_X. This notion is captured by the following natural concept of the nearest neighbor of a symbol x𝑥xitalic_x to another set of symbols \mathcal{R}caligraphic_R, which we define using the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-divergence. The usage of χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-divergence stems from our bound in Theorem 1.

Definition 3 (Nearest Neighbor).

The nearest neighbor r(x)𝑟𝑥r(x)italic_r ( italic_x ) of a symbol x𝒳𝑥𝒳x\in\mathcal{X}\setminus\mathcal{R}italic_x ∈ caligraphic_X ∖ caligraphic_R is the symbol r𝑟r\in\mathcal{R}italic_r ∈ caligraphic_R closest to x𝑥xitalic_x in terms of the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-divergence, i.e.,

r(x):=argminrχ2(WY|X=x,WY|X=r).assign𝑟𝑥subscriptargmin𝑟superscript𝜒2subscript𝑊conditionalYX𝑥subscript𝑊conditionalYX𝑟r(x):=\operatorname*{arg\,min}_{r\in\mathcal{R}}\chi^{2}\left(W_{\mathrm{Y}|% \mathrm{X}=x},W_{\mathrm{Y}|\mathrm{X}=r}\right).italic_r ( italic_x ) := start_OPERATOR roman_arg roman_min end_OPERATOR start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ) .

Nonetheless, in the context of capacity maximization, it is sufficient to consider the distance between each of the removed symbols 𝒳𝒳\mathcal{X}\setminus\mathcal{R}caligraphic_X ∖ caligraphic_R and the convex hull Conv(W)Convsubscript𝑊\operatorname{Conv}\left(W_{\mathcal{R}}\right)roman_Conv ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ). In particular, the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT distance between the distributions will lead to theoretical guarantees for capacity. We formally define the distance in the following.

Definition 4 (Distance to the Convex Hull of a Channel).

Let W:={WY|X=r}rassignsubscript𝑊subscriptsubscript𝑊conditionalYX𝑟𝑟W_{\mathcal{R}}:=\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT := { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT be a channel that generates a convex hull on the probability simplex. Then, for a symbol x𝒳𝑥𝒳x\in\mathcal{X}~{}\setminus~{}\mathcal{R}italic_x ∈ caligraphic_X ∖ caligraphic_R whose conditional distribution WY|X=xsubscript𝑊conditionalYX𝑥W_{\mathrm{Y}|\mathrm{X}=x}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT is not contained in Conv(W)Convsubscript𝑊\operatorname{Conv}\left(W_{\mathcal{R}}\right)roman_Conv ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ), assuming the convex hull spans at least one distribution with the same support as WY|X=xsubscript𝑊conditionalYX𝑥W_{\mathrm{Y}|\mathrm{X}=x}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT, the distance from WY|X=xsubscript𝑊conditionalYX𝑥W_{\mathrm{Y}|\mathrm{X}=x}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT to the convex hull of Wsubscript𝑊W_{\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT is defined as

ε(x):=minWYConv(W)χ2(WY|X=x,WY).assignsubscript𝜀𝑥subscriptsubscript𝑊𝑌Convsubscript𝑊superscript𝜒2subscript𝑊conditionalYX𝑥subscript𝑊𝑌\displaystyle\varepsilon_{\mathcal{R}}(x):=\min_{W_{Y}\in\operatorname{Conv}% \left(W_{\mathcal{R}}\right)}\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x},W_{Y}% \right).italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x ) := roman_min start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT italic_Y end_POSTSUBSCRIPT ∈ roman_Conv ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) end_POSTSUBSCRIPT italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT italic_Y end_POSTSUBSCRIPT ) .

Let WY,xsuperscriptsubscript𝑊𝑌𝑥W_{Y,x}^{\star}italic_W start_POSTSUBSCRIPT italic_Y , italic_x end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT be the distribution that minimizes the above objective so that ε(x)=χ2(WY|X=x,WY,x).subscript𝜀𝑥superscript𝜒2subscript𝑊conditionalYX𝑥superscriptsubscript𝑊𝑌𝑥\varepsilon_{\mathcal{R}}(x)=\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x},W_{Y,x}% ^{\star}\right).italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x ) = italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT italic_Y , italic_x end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) .

With those definitions at hand, we can bound the expected loss in capacity by removing a set of symbols 𝒰𝒰\mathcal{U}caligraphic_U from the alphabet 𝒳𝒳\mathcal{X}caligraphic_X. Let =𝒳𝒰𝒳𝒰\mathcal{R}=\mathcal{X}\setminus\mathcal{U}caligraphic_R = caligraphic_X ∖ caligraphic_U be the selected (remaining) symbols. Then we can divide the set 𝒰𝒰\mathcal{U}caligraphic_U of removed (unused) symbols into symbols within the convex hull of the channel Wsubscript𝑊W_{\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT (referred to as \mathcal{I}caligraphic_I) and symbols outside the convex hull (referred to as 𝒩𝒩\mathcal{N}caligraphic_N), i.e., 𝒰=𝒩𝒰𝒩\mathcal{U}=\mathcal{I}\cup\mathcal{N}caligraphic_U = caligraphic_I ∪ caligraphic_N. From Proposition 1, we know that not using symbols from \mathcal{I}caligraphic_I will not decrease the channel’s capacity. What remains is to quantify the loss in capacity when removing symbols not contained in the convex hull of the remaining ones. We establish such a result in Theorem 1. For the proof and the theorem statement, we rely on the following concept of a pseudo-simplex and pseudo-capacity, as introduced in [1].

Definition 5 (Pseudo-Simplex and Pseudo-Capacity).

Let 𝒫η(𝒴)subscript𝒫𝜂𝒴\mathcal{P}_{\eta}(\mathcal{Y})caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) be a subset of the probability simplex 𝒫(𝒴)𝒫𝒴\mathcal{P}(\mathcal{Y})caligraphic_P ( caligraphic_Y ) over alphabet 𝒴𝒴\mathcal{Y}caligraphic_Y, where each probability mass is at least η𝜂\etaitalic_η. The pseudo-capacity of a DMC W𝑊Witalic_W with input symbols 𝒳𝒳\mathcal{X}caligraphic_X is

Cη(W)=minQY𝒫η(𝒴)maxx𝒳D(WY|X=xQY).subscript𝐶𝜂𝑊subscriptsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥subscript𝑄Y\displaystyle C_{\eta}(W)=\min_{Q_{\mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal{Y% })}\max_{x\in\mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{% \mathrm{Y}}\right).italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W ) = roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) .

Let QY,η(W):=argminQY𝒫η(𝒴)maxx𝒳D(WY|X=xQY)assignsuperscriptsubscript𝑄Y𝜂𝑊subscriptargminsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥subscript𝑄YQ_{\mathrm{Y},\eta}^{\star}(W):=\operatorname*{arg\,min}_{Q_{\mathrm{Y}}\in% \mathcal{P}_{\eta}(\mathcal{Y})}\max_{x\in\mathcal{X}}\mathrm{D}\left(W_{% \mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}\right)italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_W ) := start_OPERATOR roman_arg roman_min end_OPERATOR start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) be the capacity achieving output distribution that attains the pseudo-capacity; hence minimizes the above quantity.

For the statement of the theorem, we define for a symbol x𝒳𝑥𝒳x\in\mathcal{X}\setminus\mathcal{R}italic_x ∈ caligraphic_X ∖ caligraphic_R the quantity κ(x)𝜅𝑥\kappa(x)italic_κ ( italic_x ) as the smallest non-zero probability mass of the distance-minimizing distribution WY,xsuperscriptsubscript𝑊𝑌𝑥W_{Y,x}^{\star}italic_W start_POSTSUBSCRIPT italic_Y , italic_x end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT, i.e.,

κ(x)𝜅𝑥\displaystyle\kappa(x)italic_κ ( italic_x ) :=miny𝒴:WY,x(y)0WY,x(y).assignabsentsubscript:𝑦𝒴superscriptsubscript𝑊𝑌𝑥𝑦0superscriptsubscript𝑊𝑌𝑥𝑦\displaystyle:=\min_{y\in\mathcal{Y}:W_{Y,x}^{\star}(y)\neq 0}W_{Y,x}^{\star}(% y).:= roman_min start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y : italic_W start_POSTSUBSCRIPT italic_Y , italic_x end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) ≠ 0 end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT italic_Y , italic_x end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) .

It should be noted that κ(x)𝜅𝑥\kappa(x)italic_κ ( italic_x ) and WY,xsuperscriptsubscript𝑊𝑌𝑥W_{Y,x}^{\star}italic_W start_POSTSUBSCRIPT italic_Y , italic_x end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT depend on \mathcal{R}caligraphic_R, but we omit this from the notation for brevity.

Theorem 1.

For some 0<η12|𝒴|0𝜂12𝒴0<\eta\leq\frac{1}{2|\mathcal{Y}|}0 < italic_η ≤ divide start_ARG 1 end_ARG start_ARG 2 | caligraphic_Y | end_ARG, a set of chosen symbols 𝒳𝒳\mathcal{R}\subset\mathcal{X}caligraphic_R ⊂ caligraphic_X and any x𝒳𝑥𝒳x\in\mathcal{X}italic_x ∈ caligraphic_X that maximizes D(WY|X=xQY,η)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ), with ε(x)0,κ(x)>0formulae-sequencesubscript𝜀𝑥0𝜅𝑥0\varepsilon_{\mathcal{R}}(x)\geq 0,\kappa(x)>0italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x ) ≥ 0 , italic_κ ( italic_x ) > 0, the capacity loss due to only using inputs in \mathcal{R}caligraphic_R is bounded as

C(W)C(W)𝐶𝑊𝐶subscript𝑊\displaystyle C(W)\!-\!C(W_{\mathcal{R}})italic_C ( italic_W ) - italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) 4|𝒴|η+ε(x)absent4𝒴𝜂subscript𝜀𝑥\displaystyle\leq 4|\mathcal{Y}|\eta+\varepsilon_{\mathcal{R}}(x)≤ 4 | caligraphic_Y | italic_η + italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x )
+log(η)(C(W)log(κ(x)))ε(x).𝜂𝐶subscript𝑊𝜅𝑥subscript𝜀𝑥\displaystyle+\sqrt{-\log(\eta)\left(C(W_{\mathcal{R}})\!-\!\log\left(\kappa(x% )\right)\right)}\sqrt{\varepsilon_{\mathcal{R}}(x)}.+ square-root start_ARG - roman_log ( italic_η ) ( italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) - roman_log ( italic_κ ( italic_x ) ) ) end_ARG square-root start_ARG italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x ) end_ARG .

For multiple maximizers, x𝑥xitalic_x can be chosen to minimize the upper bound. The parameter η𝜂\etaitalic_η exhibits a bias-variance trade-off. With x𝒳superscript𝑥𝒳x^{\prime}\in\mathcal{X}italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∈ caligraphic_X being the maximizer of D(WY|X=xQY)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) among all x𝒳𝑥𝒳x\in\mathcal{X}italic_x ∈ caligraphic_X, a suitable choice of η𝜂\etaitalic_η is

η=(C(W)log(κ(x))ε(x)4|𝒴|+0.07)2.𝜂superscript𝐶subscript𝑊𝜅superscript𝑥subscript𝜀superscript𝑥4𝒴0.072\eta=\left(\frac{\sqrt{C(W_{\mathcal{R}})-\log\left(\kappa(x^{\prime})\right)}% \sqrt{\varepsilon_{\mathcal{R}}(x^{\prime})}}{4|\mathcal{Y}|}+0.07\right)^{2}.italic_η = ( divide start_ARG square-root start_ARG italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) - roman_log ( italic_κ ( italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) ) end_ARG square-root start_ARG italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) end_ARG end_ARG start_ARG 4 | caligraphic_Y | end_ARG + 0.07 ) start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT .

A simplified, yet looser, statement is obtained when considering for symbol x𝑥xitalic_x the nearest neighbor r(x)𝑟𝑥r(x)italic_r ( italic_x ) instead of ε(x)subscript𝜀𝑥\varepsilon_{\mathcal{R}}(x)italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x ), which avoids computing the distance to the convex hull.

Sketch of the Proof.

We apply Proposition 1 and the triangle inequality to show the equivalence of comparing the capacity of W𝑊Witalic_W to either (i) the capacity of Wsubscript𝑊W_{\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT or (ii) the capacity of Wsuperscriptsubscript𝑊W_{\mathcal{R}}^{\prime}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT, which is the channel that additionally contains all distributions WY,xsuperscriptsubscript𝑊𝑌𝑥W_{Y,x}^{\star}italic_W start_POSTSUBSCRIPT italic_Y , italic_x end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT in the convex hull of Wsubscript𝑊W_{\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT that represent the closest to each symbol x𝒳𝑥𝒳x\in\mathcal{X}\setminus\mathcal{R}italic_x ∈ caligraphic_X ∖ caligraphic_R in terms of χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-distance. Hence, we ensure that the nearest neighbor r(x)𝑟𝑥r(x)italic_r ( italic_x ) of each symbol x𝑥xitalic_x in W𝑊Witalic_W is the distribution in the convex hull of Wsubscript𝑊W_{\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT that represents the minimum distance ε(x)subscript𝜀𝑥\varepsilon_{\mathcal{R}}(x)italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_x ). Next, we use pseudo-capacity to bound the difference between Cη(W)subscript𝐶𝜂𝑊C_{\eta}(W)italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W ) and Cη(W)subscript𝐶𝜂subscript𝑊C_{\eta}(W_{\mathcal{R}})italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) based on the distance to the nearest neighbors (and due to the above, equivalently the distance to the convex hull) of certain symbols. Therefore, we make use of the following lemma, for which we restrict the capacity-achieving output distribution QYsubscript𝑄YQ_{\mathrm{Y}}italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT to the pseudo-simplex 𝒫η(𝒴)subscript𝒫𝜂𝒴\mathcal{P}_{\eta}(\mathcal{Y})caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ).

Lemma 1.

For any choice of 0<η12|𝒴|0𝜂12𝒴0<\eta\leq\frac{1}{2|\mathcal{Y}|}0 < italic_η ≤ divide start_ARG 1 end_ARG start_ARG 2 | caligraphic_Y | end_ARG, let QY,ηsuperscriptsubscript𝑄Y𝜂Q_{\mathrm{Y},\eta}^{\star}italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT be as defined above. Then, with x𝑥xitalic_x being the maximizer of D(WY|X=xQY,η)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) and r(x)𝑟𝑥r(x)\in\mathcal{R}italic_r ( italic_x ) ∈ caligraphic_R its nearest neighbor that minimizes α:=χ2(WY|X=x,WY|X=r(x))assign𝛼superscript𝜒2subscript𝑊conditionalYX𝑥subscript𝑊conditionalYX𝑟superscript𝑥\alpha:=\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x},W_{\mathrm{Y}|\mathrm{X}=r(x% ^{\star})}\right)italic_α := italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ), we have the following upper bound

Cη(W)Cη(W)subscript𝐶𝜂𝑊subscript𝐶𝜂subscript𝑊\displaystyle C_{\eta}\left(W\right)-C_{\eta}\left(W_{\mathcal{R}}\right)italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W ) - italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT )
α+αC(W)log(κ(x)η)+log(1η)log(1κ(x)).absent𝛼𝛼𝐶subscript𝑊𝜅𝑥𝜂1𝜂1𝜅𝑥\displaystyle\leq\alpha+\sqrt{\alpha}\sqrt{C(W_{\mathcal{R}})\log\left(\frac{% \kappa(x)}{\eta}\right)+\log\left(\frac{1}{\eta}\right)\log\left(\frac{1}{% \kappa(x)}\right)}.≤ italic_α + square-root start_ARG italic_α end_ARG square-root start_ARG italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) roman_log ( divide start_ARG italic_κ ( italic_x ) end_ARG start_ARG italic_η end_ARG ) + roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) roman_log ( divide start_ARG 1 end_ARG start_ARG italic_κ ( italic_x ) end_ARG ) end_ARG .

The gap between Cη(W)subscript𝐶𝜂subscript𝑊C_{\eta}(W_{\mathcal{R}})italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) to capacity C(W)𝐶subscript𝑊C(W_{\mathcal{R}})italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) (and between Cη(W)subscript𝐶𝜂𝑊C_{\eta}(W)italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W ) and C(W)𝐶𝑊C(W)italic_C ( italic_W ), respectively) can be bounded by 2η|𝒴|2𝜂𝒴2\eta|\mathcal{Y}|2 italic_η | caligraphic_Y | by the application of [19, Lemma 1]. With an appropriate choice of η𝜂\etaitalic_η to trade off the linear and non-linear terms, we obtain the statement in the theorem. Note that the computation of D(WY|X=xQY,η)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) determines the choice of x𝑥xitalic_x for which the bound applies; hence, it requires a choice of η𝜂\etaitalic_η. To find a value for η𝜂\etaitalic_η that is well suited for the maximum of D(WY|X=xQY,η)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ), we use the surrogate objective D(WY|X=xQY)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) for the calculation of η𝜂\etaitalic_η. ∎

IV Input Symbol Selection by Clustering

We now turn our attention to designing a practical algorithm for selecting a subset of k𝑘kitalic_k input symbols that minimize the loss in capacity compared to using all possible symbols. For this problem, exhaustive search over all possible subsets is typically computationally infeasible, especially in cases where k𝑘kitalic_k and the input alphabet are large. On the other hand, greedy algorithms require computing the channel’s capacity for many candidate sets of symbols, which can be computationally demanding, especially for large k𝑘kitalic_k, and might produce sub-optimal solutions. Further, as discussed, there is no simple theoretical guarantee on the loss in capacity incurred by the solutions of greedy algorithms compared to the optimal solution. We thus propose a clustering-based algorithm, which, in the first step, clusters symbols with similar conditional distributions, and in the second step, carefully chooses representatives of each cluster. Our proposed algorithm does not require any possibly expensive capacity calculation. We compare its performance to the optimal solution (whenever finding it with an exhaustive search is feasible).

Our strategy, summarized in Algorithm 1, first partitions the input symbols according to their conditional distributions into k𝑘kitalic_k clusters of similar symbols. Then, having identified the clusters, it determines a representative for each cluster. Inspired by Theorem 1, the representatives are chosen such that their resulting convex hull likely contains as many removed symbols as possible, while simultaneously minimizing the distance to the symbols outside the convex hull.

We use agglomerative (hierarchical) clustering due to its simplicity and flexibility in the usage of distance measures. The clustering first assigns one cluster to each symbol. Then, the pairwise Jensen-Shannon divergence between all symbols is computed. The pairwise distance between two clusters, called linkage, is computed as the maximum distance between their respective elements, i.e., for two clusters m𝑚mitalic_m and n𝑛nitalic_n with elements msubscript𝑚\mathcal{H}_{m}caligraphic_H start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT and nsubscript𝑛\mathcal{H}_{n}caligraphic_H start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT, respectively, we have Link(m,n):=maxxm,xnJSD(WY|X=xWY|X=x)assignLink𝑚𝑛subscriptformulae-sequence𝑥subscript𝑚superscript𝑥subscript𝑛JSDconditionalsubscript𝑊conditionalYX𝑥subscript𝑊conditionalYXsuperscript𝑥\text{Link}(m,n):=\max_{x\in\mathcal{H}_{m},x^{\prime}\in\mathcal{H}_{n}}\text% {JSD}(W_{\mathrm{Y}|\mathrm{X}=x}\|W_{\mathrm{Y}|\mathrm{X}=x^{\prime}})Link ( italic_m , italic_n ) := roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_H start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT , italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∈ caligraphic_H start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT end_POSTSUBSCRIPT JSD ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ). The two clusters with the minimum linkage are merged. This process is repeated until k𝑘kitalic_k clusters remain. While the bound in Theorem 1 suggests the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-divergence as a distance measure, we chose the Jensen-Shannon divergence and the linkage as the maximum distance between cluster elements since these achieved the best empirical results. It is of interest to close this gap and propose bounds based on JSD instead.

To choose a representative for each cluster, we select the symbol xmsubscript𝑥𝑚x_{m}italic_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT of cluster m𝑚mitalic_m that maximizes the average distance to all other symbols xxm𝒳𝑥subscript𝑥𝑚𝒳x\neq x_{m}\in\mathcal{X}italic_x ≠ italic_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ∈ caligraphic_X. We found that among various approaches, selecting the representative as the symbol in a cluster whose distribution maximizes the average distance to all other symbols (including those outside the cluster) provides a good trade-off between compute cost and the resulting performance. This aligns with Theorem 1, that advocates the selection of representatives that create a large convex hull.

Fig. 1 shows an illustrative example of the performance. It compares our hierarchical clustering algorithm with the bounds of Theorem 1, for different values of k{2,,10}𝑘210k\in\{2,\dots,10\}italic_k ∈ { 2 , … , 10 }, for a DMC with input and output alphabet sizes |𝒳|=|𝒴|=30𝒳𝒴30|\mathcal{X}|=|\mathcal{Y}|=30| caligraphic_X | = | caligraphic_Y | = 30. The DMC is generated according to a Dirichlet distribution, as explained in the sequel. The capacity of the full channel is computed by including all input symbols in 𝒳𝒳\mathcal{X}caligraphic_X. As a baseline for our clustering algorithm, we include the results from an exhaustive search over all sets of k𝑘kitalic_k symbols.

Algorithm 1 Input Symbol Selection.
DMC W𝑊Witalic_W, desired number k<|𝒳|𝑘𝒳k<|\mathcal{X}|italic_k < | caligraphic_X | of inputs
Compute JSD(WY|X=xWY|X=x)JSDconditionalsubscript𝑊conditionalYX𝑥subscript𝑊conditionalYXsuperscript𝑥\text{JSD}(W_{\mathrm{Y}|\mathrm{X}=x}\|W_{\mathrm{Y}|\mathrm{X}=x^{\prime}})JSD ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ) for all symbols x,x𝑥superscript𝑥x,x^{\prime}italic_x , italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT
Initialize clusters m,m[|𝒳|]subscript𝑚𝑚delimited-[]𝒳\mathcal{H}_{m},m\in[|\mathcal{X}|]caligraphic_H start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT , italic_m ∈ [ | caligraphic_X | ], one for each WY|X=xsubscript𝑊conditionalYX𝑥W_{\mathrm{Y}|\mathrm{X}=x}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT
while Number of clusters >kabsent𝑘>k> italic_k do
     For all clusters m,n𝑚𝑛m,nitalic_m , italic_n, compute the pairwise linkage
     Link(m,n)=maxxm,xnJSD(WY|X=xWY|X=x)Link𝑚𝑛subscriptformulae-sequence𝑥subscript𝑚superscript𝑥subscript𝑛JSDconditionalsubscript𝑊conditionalYX𝑥subscript𝑊conditionalYXsuperscript𝑥\displaystyle\text{Link}(m,n)=\max_{x\in\mathcal{H}_{m},x^{\prime}\in\mathcal{% H}_{n}}\text{JSD}(W_{\mathrm{Y}|\mathrm{X}=x}\|W_{\mathrm{Y}|\mathrm{X}=x^{% \prime}})Link ( italic_m , italic_n ) = roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_H start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT , italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∈ caligraphic_H start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT end_POSTSUBSCRIPT JSD ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT )
     Merge clusters m,nsuperscript𝑚superscript𝑛m^{\star},n^{\star}italic_m start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT , italic_n start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT with minimum Link(m,n)Link𝑚𝑛\text{Link}(m,n)Link ( italic_m , italic_n ), i.e.,
     msubscriptsuperscript𝑚\mathcal{H}_{m^{\star}}caligraphic_H start_POSTSUBSCRIPT italic_m start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT = mnsubscriptsuperscript𝑚subscriptsuperscript𝑛\mathcal{H}_{m^{\star}}\cup\mathcal{H}_{n^{\star}}caligraphic_H start_POSTSUBSCRIPT italic_m start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ∪ caligraphic_H start_POSTSUBSCRIPT italic_n start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT
end while
for Each remaining cluster m[k]𝑚delimited-[]𝑘m\in[k]italic_m ∈ [ italic_k ] do
     Select the symbol xmmsubscript𝑥𝑚subscript𝑚x_{m}\in\mathcal{H}_{m}italic_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ∈ caligraphic_H start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT of cluster m𝑚mitalic_m maximizing
     average distance to all other symbols xxm𝑥subscript𝑥𝑚x\neq x_{m}italic_x ≠ italic_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT in 𝒳𝒳\mathcal{X}caligraphic_X
end for
Output W={WY|X=xm}m[k]superscript𝑊subscriptsubscript𝑊conditionalYXsubscript𝑥𝑚𝑚delimited-[]𝑘W^{\prime}=\{W_{\mathrm{Y}|\mathrm{X}=x_{m}}\}_{m\in[k]}italic_W start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT = { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_m ∈ [ italic_k ] end_POSTSUBSCRIPT

For η𝜂\etaitalic_η, we use the value proposed in Theorem 1. For values k<5𝑘5k<5italic_k < 5, this choice of η𝜂\etaitalic_η exceeds the limit of 12|𝒴|12𝒴\frac{1}{2|\mathcal{Y}|}divide start_ARG 1 end_ARG start_ARG 2 | caligraphic_Y | end_ARG. Hence, obtaining a tight bound is not possible. We plot the optimal solution obtained through an exhaustive search for small values of k𝑘kitalic_k, and the capacity of the full channel that uses all the input symbols in 𝒳𝒳\mathcal{X}caligraphic_X. It can be seen that our clustering algorithm, on average, finds an optimal selection of the input symbols. Even without knowing the capacity of the full channel, with k=5𝑘5k=5italic_k = 5, the bound from Theorem 1 indicates that accounting for the remaining 25252525 symbols cannot improve capacity by more than 0.8absent0.8\approx 0.8≈ 0.8 bit. In this simple example the bound from Theorem 1 is conservative (as can be seen from the exact value of capacity). However, for very large input alphabet channels, the capacity is infeasible to compute, and this bound may be the only indication.

222233334444555566667777888899991010101000111122223333Numer of Symbols k𝑘\displaystyle kitalic_kC(W)𝐶𝑊\displaystyle C(W)italic_C ( italic_W ) in BitsUpper Bound (cf. Theorem 1)Full Channel (no symbol removal)Hierarchical Clustering (cf. Algorithm 1)Exhaustive Search
Figure 1: Input selection results for 50505050 DMCs with input and output alphabet size |𝒳|=|𝒴|=30𝒳𝒴30|\mathcal{X}|=|\mathcal{Y}|=30| caligraphic_X | = | caligraphic_Y | = 30, randomly sampled with ν=5𝜈5\nu=5italic_ν = 5, d1=0.005subscript𝑑10.005d_{1}=0.005italic_d start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 0.005 and d2=1010subscript𝑑2superscript1010d_{2}=10^{10}italic_d start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT = 10 start_POSTSUPERSCRIPT 10 end_POSTSUPERSCRIPT. Lines show average results, shaded areas the standard deviation.

Generating a DMC with the Dirichlet distribution: We introduce a random, yet structured, hierarchical sampling method for generating DMCs with input alphabet 𝒳𝒳\mathcal{X}caligraphic_X. We use the Dirichlet distribution parameterized by a vector α=(α1,,α|𝒴|)𝛼subscript𝛼1subscript𝛼𝒴\alpha=(\alpha_{1},\dots,\alpha_{|\mathcal{Y}|})italic_α = ( italic_α start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , … , italic_α start_POSTSUBSCRIPT | caligraphic_Y | end_POSTSUBSCRIPT ) which generates a probability distribution of the same dimension, i.e., over 𝒫(𝒴)𝒫𝒴\mathcal{P}(\mathcal{Y})caligraphic_P ( caligraphic_Y ). The relation αyy𝒴αysubscript𝛼𝑦subscript𝑦𝒴subscript𝛼𝑦\frac{\alpha_{y}}{\sum_{y\in\mathcal{Y}}\alpha_{y}}divide start_ARG italic_α start_POSTSUBSCRIPT italic_y end_POSTSUBSCRIPT end_ARG start_ARG ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT italic_α start_POSTSUBSCRIPT italic_y end_POSTSUBSCRIPT end_ARG determines the average probability mass of the symbols y𝒴𝑦𝒴y\in\mathcal{Y}italic_y ∈ caligraphic_Y. As opposed to uniform or Gaussian sampling of the conditional distributions, Dirichlet sampling allows to tune the expected capacity of the channels through the parameter choice. The larger the values αy,y𝒴subscript𝛼𝑦𝑦𝒴\alpha_{y},y\in\mathcal{Y}italic_α start_POSTSUBSCRIPT italic_y end_POSTSUBSCRIPT , italic_y ∈ caligraphic_Y, the more noisy the rows of the transition matrices will be; whereas smaller values of αysubscript𝛼𝑦\alpha_{y}italic_α start_POSTSUBSCRIPT italic_y end_POSTSUBSCRIPT will lead to cleaner rows of the transition matrices.111We add a small non-negative mass to each entry and normalize to ensure that the convex hull has the same support as all the removed rows, as assumed for Theorem 1. Meeting this assumption is likely in practice.

To model that the rows of the transition matrices might be dependent, we first sample ν<|𝒳|𝜈𝒳\nu<|\mathcal{X}|italic_ν < | caligraphic_X | of Dirichlet samples with α=d1(1,,1)𝛼subscript𝑑111\alpha=d_{1}(1,\dots,1)italic_α = italic_d start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ( 1 , … , 1 ), where the parameter d1subscript𝑑1d_{1}italic_d start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT will determine the variance of the sample, hence the capacity of the channel. Each sample is then used as a parameterization for a new Dirichlet distribution, scaled by d2subscript𝑑2d_{2}italic_d start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT. Hence, those distributions will be noisy samples around the previously sampled ones, which justifies the term hierarchical sampling. For our simulations, we used ν=5𝜈5\nu=5italic_ν = 5, d1=0.005subscript𝑑10.005d_{1}=0.005italic_d start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 0.005 and d2=1010subscript𝑑2superscript1010d_{2}=10^{10}italic_d start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT = 10 start_POSTSUPERSCRIPT 10 end_POSTSUPERSCRIPT.

V Conclusion

We investigated the problem of capacity-optimal input symbol selection for DMCs. Based on the channel’s transition matrix, we derived bounds for the capacity loss incurred by the removal of specific input symbols. We showed the dependency of the bounds on the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-distance of the removed symbols to the convex hull spanned by the selected inputs. We transferred our theoretical results to designing a clustering-based selection algorithm with a cluster representative choice tailored to maximizing the size of the resulting convex hull. With DMCs randomly sampled with a Dirichlet distribution, we compared our algorithm to an exhaustive search and observed the established theoretical bound on the capacity loss.

References

  • [1] M. Egger, R. Bitar, A. Wachter-Zeh, D. Gündüz, and N. Weinberger, “Maximal-capacity discrete memoryless channel identification,” in IEEE International Symposium on Information Theory (ISIT), 2023, pp. 2248–2253.
  • [2] T. Chan, S. Hranilovic, and F. Kschischang, “Capacity-achieving probability measure for conditionally Gaussian channels with bounded inputs,” IEEE Transactions on Information Theory, vol. 51, no. 6, pp. 2073–2088, 2005.
  • [3] A. Konar and N. D. Sidiropoulos, “A simple and effective approach for transmit antenna selection in multiuser massive mimo leveraging submodularity,” IEEE Transactions on Signal Processing, vol. 66, no. 18, pp. 4869–4883, 2018.
  • [4] R. D. Wesel, E. E. Wesel, L. Vandenberghe, C. Komninakis, and M. Medard, “Efficient binomial channel capacity computation with an application to molecular communication,” in Information Theory and Applications Workshop (ITA), 2018, pp. 1–5.
  • [5] A. Kobovich, E. Yaakobi, and N. Weinberger, “M-DAB: An input-distribution optimization algorithm for composite DNA storage by the multinomial channel,” arXiv preprint arXiv:2309.17193, 2023.
  • [6] A. Mezghani, M. T. Ivrlac, and J. A. Nossek, “Achieving near-capacity on large discrete memoryless channels with uniform distributed selected input,” in International Symposium on Information Theory and Its Applications, 2008, pp. 1–6.
  • [7] A. Schmeink and H. Zhang, “Capacity-achieving probability measure for a reduced number of signaling points,” Wireless Networks, vol. 17, pp. 987–999, 2011.
  • [8] A. Schmeink, R. Mathar, and H. Zhang, “Reducing the number of signaling points keeping capacity and cutoff rate high,” International Symposium on Wireless Communication Systems, pp. 932–936, 2010.
  • [9] F. Nielsen and K. Sun, “Clustering in hilbert simplex geometry,” arXiv preprint arXiv:1704.00454, 2017.
  • [10] R. G. Gallager, Information theory and reliable communication.   Springer, 1968, vol. 588.
  • [11] I. Csiszár, “A class of measures of informativity of observation channels,” Periodica Mathematica Hungarica, vol. 2, no. 1, pp. 191–213, Mar. 1972.
  • [12] J. H. B. Kemperman, “On the Shannon capacity of an arbitrary channel,” Indagationes Mathematicae (Proceedings), vol. 77, no. 2, pp. 101–115, 1974.
  • [13] M. Feng, J. E. Mitchell, J.-S. Pang, X. Shen, and A. Wächter, “Complementarity formulations of l0-norm optimization problems,” Industrial Engineering and Management Sciences. Technical Report. Northwestern University, Evanston, IL, USA, vol. 5, 2013.
  • [14] F. Bach et al., “Learning with submodular functions: A convex optimization perspective,” Foundations and Trends® in machine learning, vol. 6, no. 2-3, pp. 145–373, 2013.
  • [15] J. Bilmes, “Submodularity in machine learning and artificial intelligence,” arXiv preprint arXiv:2202.00132, 2022.
  • [16] U. Feige, M. Feldman, and I. Talgam-Cohen, “Approximate modularity revisited,” in Annual ACM SIGACT Symposium on Theory of Computing, 2017, pp. 1028–1041.
  • [17] A. Das and D. Kempe, “Approximate submodularity and its applications: Subset selection, sparse approximation and dictionary selection,” Journal of Machine Learning Research, vol. 19, no. 3, pp. 1–34, 2018.
  • [18] F. Chierichetti, A. Dasgupta, and R. Kumar, “On additive approximate submodularity,” Theoretical Computer Science, vol. 922, pp. 346–360, 2022.
  • [19] M. Egger, R. Bitar, A. Wachter-Zeh, D. Gündüz, and N. Weinberger, “Maximal-capacity discrete memoryless channel identification,” arXiv preprint arXiv:2401.10204, 2024.

-A Counter-Examples for Submodularity of a DMC

Let f𝑓fitalic_f be the capacity C𝐶Citalic_C that maps 2Wsuperscript2𝑊2^{W}2 start_POSTSUPERSCRIPT italic_W end_POSTSUPERSCRIPT to the capacity 0C(W)log(|W|)0𝐶𝑊𝑊0\leq C(W)\leq\log(|W|)0 ≤ italic_C ( italic_W ) ≤ roman_log ( | italic_W | ), where W𝑊Witalic_W refers to the collection of conditional distributions {WY|X=x}x𝒳subscriptsubscript𝑊conditionalYX𝑥𝑥𝒳\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{X}}{ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT. The definition of diminishing returns breaks when considering the empty set 𝒥=𝒥\mathcal{J}=\emptysetcaligraphic_J = ∅ and a single-element set 𝒦={WY|X=x}𝒦subscript𝑊conditionalYX𝑥\mathcal{K}=\{W_{\mathrm{Y}|\mathrm{X}=x}\}caligraphic_K = { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } such that =WY|X=xWY|X=xsubscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑥\ell=W_{\mathrm{Y}|\mathrm{X}=x^{\prime}}\neq W_{\mathrm{Y}|\mathrm{X}=x}roman_ℓ = italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ≠ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT for any two distinct elements x,x𝒳𝑥superscript𝑥𝒳x,x^{\prime}\in\mathcal{X}italic_x , italic_x start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ∈ caligraphic_X. In this case, we have

f(𝒥{})f(𝒥)𝑓𝒥𝑓𝒥\displaystyle f(\mathcal{J}\cup\{\ell\})-f(\mathcal{J})italic_f ( caligraphic_J ∪ { roman_ℓ } ) - italic_f ( caligraphic_J ) =00,absent00\displaystyle=0-0,= 0 - 0 ,
f(𝒦{})f(𝒦)𝑓𝒦𝑓𝒦\displaystyle f(\mathcal{K}\cup\{\ell\})-f(\mathcal{K})italic_f ( caligraphic_K ∪ { roman_ℓ } ) - italic_f ( caligraphic_K ) =C¯0,absent¯𝐶0\displaystyle=\bar{C}-0,= over¯ start_ARG italic_C end_ARG - 0 ,

for some 0<C¯log(2)0¯𝐶20<\bar{C}\leq\log(2)0 < over¯ start_ARG italic_C end_ARG ≤ roman_log ( 2 ), which contradicts the above definition of submodularity. While this result is straightforward, it remains to show that DMCs do not fulfill the diminishing return property when |𝒥|1𝒥1|\mathcal{J}|\geq 1| caligraphic_J | ≥ 1 and |𝒦|2𝒦2|\mathcal{K}|\geq 2| caligraphic_K | ≥ 2. We show this through a counterexample. Consider the channel W={WY|X=x}x[4]𝑊subscriptsubscript𝑊conditionalYX𝑥𝑥delimited-[]4W=\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in[4]}italic_W = { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ [ 4 ] end_POSTSUBSCRIPT:

WY|X=1subscript𝑊conditionalYX1\displaystyle W_{\mathrm{Y}|\mathrm{X}=1}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = 1 end_POSTSUBSCRIPT =[0.6,0.2,0.1,0.1],absent0.60.20.10.1\displaystyle=[0.6,0.2,0.1,0.1],= [ 0.6 , 0.2 , 0.1 , 0.1 ] , WY|X=2subscript𝑊conditionalYX2\displaystyle W_{\mathrm{Y}|\mathrm{X}=2}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = 2 end_POSTSUBSCRIPT =[0.6,0.1,0.1,0.2]absent0.60.10.10.2\displaystyle=[0.6,0.1,0.1,0.2]= [ 0.6 , 0.1 , 0.1 , 0.2 ]
WY|X=3subscript𝑊conditionalYX3\displaystyle W_{\mathrm{Y}|\mathrm{X}=3}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = 3 end_POSTSUBSCRIPT =[0.6,0.1,0.2,0.1],absent0.60.10.20.1\displaystyle=[0.6,0.1,0.2,0.1],= [ 0.6 , 0.1 , 0.2 , 0.1 ] , WY|X=4subscript𝑊conditionalYX4\displaystyle W_{\mathrm{Y}|\mathrm{X}=4}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = 4 end_POSTSUBSCRIPT =[0.1,0.6,0.1,0.2]absent0.10.60.10.2\displaystyle=[0.1,0.6,0.1,0.2]= [ 0.1 , 0.6 , 0.1 , 0.2 ]

With a slight abuse of notation, let C(𝒥)𝐶𝒥C(\mathcal{J})italic_C ( caligraphic_J ) denote the capacity of the channel WY|Xsubscript𝑊conditionalYXW_{\mathrm{Y}|\mathrm{X}}italic_W start_POSTSUBSCRIPT roman_Y | roman_X end_POSTSUBSCRIPT with input symbols 𝒥𝒳𝒥𝒳\mathcal{J}\subseteq\mathcal{X}caligraphic_J ⊆ caligraphic_X. Then, we have that

C({1,2,4}\displaystyle C(\{1,2,4\}italic_C ( { 1 , 2 , 4 } {3})C({1,2,4})\displaystyle\cup\{3\})-C(\{1,2,4\})∪ { 3 } ) - italic_C ( { 1 , 2 , 4 } )
>C({1,2}{3})C({1,2}),absent𝐶123𝐶12\displaystyle>C(\{1,2\}\cup\{3\})-C(\{1,2\}),> italic_C ( { 1 , 2 } ∪ { 3 } ) - italic_C ( { 1 , 2 } ) ,

which violates Definition 1. This counterexample is constructed by noting that the quantities H(WY|X=x)Hsubscript𝑊conditionalYX𝑥\mathrm{H}\left(W_{\mathrm{Y}|\mathrm{X}=x}\right)roman_H ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ) are equal for all x𝒳𝑥𝒳x\in\mathcal{X}italic_x ∈ caligraphic_X. Hence, capacity is obtained by maximizing the entropy of the output distribution, i.e.,

C(W)=H(Y|X=x)+maxPX𝒫(𝒳)H(Y).𝐶𝑊HconditionalYX𝑥subscriptsubscript𝑃X𝒫𝒳HYC(W)=-\mathrm{H}\left(\mathrm{Y}|\mathrm{X}=x\right)+\max_{P_{\mathrm{X}}\in% \mathcal{P}(\mathcal{X})}\mathrm{H}\left(\mathrm{Y}\right).italic_C ( italic_W ) = - roman_H ( roman_Y | roman_X = italic_x ) + roman_max start_POSTSUBSCRIPT italic_P start_POSTSUBSCRIPT roman_X end_POSTSUBSCRIPT ∈ caligraphic_P ( caligraphic_X ) end_POSTSUBSCRIPT roman_H ( roman_Y ) . (2)

For such channels, the question of diminishing returns can be answered by studying how well adding a certain input symbol to the sets 𝒥𝒥\mathcal{J}caligraphic_J and 𝒦𝒦\mathcal{K}caligraphic_K balances out the capacity achieving output distribution, reflected by an increase in its entropy. Our example shows that adding a specific symbol to a large alphabet balances the output distribution more positively than for a smaller alphabet, thus violating sub-additivity.

-B Proof of Proposition 1

Proof.

Let 𝒰𝒰\mathcal{U}caligraphic_U be symbols that lie in the convex hull of a set of symbols \mathcal{R}caligraphic_R, i.e. each symbol u𝒰𝑢𝒰u\in\mathcal{U}italic_u ∈ caligraphic_U can be described as a convex combination of symbols in \mathcal{R}caligraphic_R so that WY|X=u=rcu,rWY|X=rsubscript𝑊conditionalYX𝑢subscript𝑟subscript𝑐𝑢𝑟subscript𝑊conditionalYX𝑟W_{\mathrm{Y}|\mathrm{X}=u}=\sum_{r\in\mathcal{R}}c_{u,r}W_{\mathrm{Y}|\mathrm% {X}=r}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_u end_POSTSUBSCRIPT = ∑ start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT italic_c start_POSTSUBSCRIPT italic_u , italic_r end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT for some values crsubscript𝑐𝑟c_{r}italic_c start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT s.t. rcr=1subscript𝑟subscript𝑐𝑟1\sum_{r\in\mathcal{R}}c_{r}=1∑ start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT italic_c start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT = 1. We have by the strict convexity of KL-divergence that

D(WY|X=uQY)Dconditionalsubscript𝑊conditionalYX𝑢superscriptsubscript𝑄Y\displaystyle\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=u}\|Q_{\mathrm{Y}}^{% \star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_u end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) =D(rcrWY|X=uQY)absentDconditionalsubscript𝑟subscript𝑐𝑟subscript𝑊conditionalYX𝑢superscriptsubscript𝑄Y\displaystyle=\mathrm{D}\left(\sum_{r\in\mathcal{R}}c_{r}W_{\mathrm{Y}|\mathrm% {X}=u}\|Q_{\mathrm{Y}}^{\star}\right)= roman_D ( ∑ start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT italic_c start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_u end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )
rcrD(WY|X=uQY)Cabsentsubscript𝑟subscript𝑐𝑟Dconditionalsubscript𝑊conditionalYX𝑢superscriptsubscript𝑄Y𝐶\displaystyle\leq\sum_{r\in\mathcal{R}}c_{r}\mathrm{D}\left(W_{\mathrm{Y}|% \mathrm{X}=u}\|Q_{\mathrm{Y}}^{\star}\right)\leq C≤ ∑ start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT italic_c start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_u end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) ≤ italic_C

Hence, the symbols in 𝒴𝒴\mathcal{Y}caligraphic_Y did not contribute to the information radius of the channel, and can, hence, be removed without loss in capacity. Note that the capacity-achieving distribution still lies in the convex hull of Wsubscript𝑊W_{\mathcal{R}}italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT. This concludes the proof. ∎

-C Proof of Lemma 1

Proof.

Using the definitions of the pseudo-simplex and pseudo-capacity, we can bound the difference of the pseudo-capacities of the channels as follows. Let QY,ηsuperscriptsubscript𝑄Y𝜂Q_{\mathrm{Y},\eta}^{\star}italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT be the unique minimizer maxrD(WY|X=rQY)subscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟subscript𝑄Y\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{\mathrm{% Y}}\right)roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) over the pseudo simplex 𝒫η(𝒴)subscript𝒫𝜂𝒴\mathcal{P}_{\eta}(\mathcal{Y})caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ), i.e., QY,η:=argminQY𝒫η(𝒴)maxrD(WY|X=rQY)assignsuperscriptsubscript𝑄Y𝜂subscriptargminsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟subscript𝑄YQ_{\mathrm{Y},\eta}^{\star}:=\operatorname*{arg\,min}_{Q_{\mathrm{Y}}\in% \mathcal{P}_{\eta}(\mathcal{Y})}\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{% \mathrm{Y}|\mathrm{X}=r}\|Q_{\mathrm{Y}}\right)italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT := start_OPERATOR roman_arg roman_min end_OPERATOR start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ). Then, with xsuperscript𝑥x^{\star}italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT being any maximizer of D(WY|X=xQY,η)Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) and r(x)𝑟superscript𝑥r(x^{\star})italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) being its representative, assuming WY|X=r(x)subscript𝑊conditionalYX𝑟superscript𝑥W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT is at least supported where WY|X=xsubscript𝑊conditionalYXsuperscript𝑥W_{\mathrm{Y}|\mathrm{X}=x^{\star}}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT is, we have

Cη({WY|X=x}x𝒳)Cη({WY|X=r}r)subscript𝐶𝜂subscriptsubscript𝑊conditionalYX𝑥𝑥𝒳subscript𝐶𝜂subscriptsubscript𝑊conditionalYX𝑟𝑟\displaystyle C_{\eta}\left(\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{X}}% \right)-C_{\eta}\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\right)italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT ) - italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT )
minQY𝒫η(𝒴)maxx𝒳D(WY|X=xQY)minQY𝒫η(𝒴)maxrD(WY|X=rQY)subscriptsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥subscript𝑄Ysubscriptsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟subscript𝑄Y\displaystyle\min_{Q_{\mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal{Y})}\!\max_{x% \in\mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}% \right)\!-\!\!\!\!\!\!\!\min_{Q_{\mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal{Y})% }\!\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{% \mathrm{Y}}\right)roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) - roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT )
=minQY𝒫η(𝒴)maxx𝒳D(WY|X=xQY)maxrD(WY|X=rQY,η)absentsubscriptsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥subscript𝑄Ysubscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟superscriptsubscript𝑄Y𝜂\displaystyle=\min_{Q_{\mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal{Y})}\max_{x% \in\mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}% \right)-\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{% \mathrm{Y},\eta}^{\star}\right)= roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) - roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )
maxx𝒳D(WY|X=xQY,η)maxrD(WY|X=rQY,η)absentsubscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂subscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟superscriptsubscript𝑄Y𝜂\displaystyle\leq\max_{x\in\mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X% }=x}\|Q_{\mathrm{Y},\eta}^{\star}\right)-\max_{r\in\mathcal{R}}\mathrm{D}\left% (W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{\mathrm{Y},\eta}^{\star}\right)≤ roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) - roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )
=(aa)D(WY|X=xQY,η)D(WY|X=r(x)QY,η)𝑎𝑎Dconditionalsubscript𝑊conditionalYXsuperscript𝑥superscriptsubscript𝑄Y𝜂Dconditionalsubscript𝑊conditionalYX𝑟superscript𝑥superscriptsubscript𝑄Y𝜂\displaystyle\overset{(aa)}{=}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x^{% \star}}\|Q_{\mathrm{Y},\eta}^{\star}\right)-\mathrm{D}\left(W_{\mathrm{Y}|% \mathrm{X}=r(x^{\star})}\|Q_{\mathrm{Y},\eta}^{\star}\right)start_OVERACCENT ( italic_a italic_a ) end_OVERACCENT start_ARG = end_ARG roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) - roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )
=y𝒴WY|X=x(y)logWY|X=x(y)QY,η(y)WY|X=r(x)(y)logWY|X=r(x)(y)QY,η(y)absentmissing-subexpressionsubscript𝑦𝒴subscript𝑊conditionalYXsuperscript𝑥𝑦subscript𝑊conditionalYXsuperscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦missing-subexpressionsubscript𝑊conditionalYX𝑟superscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle=\begin{aligned} &\sum_{y\in\mathcal{Y}}W_{\mathrm{Y}|\mathrm{X}=% x^{\star}}(y)\log\frac{W_{\mathrm{Y}|\mathrm{X}=x^{\star}}(y)}{Q_{\mathrm{Y},% \eta}^{\star}(y)}\\ &-W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\log\frac{W_{\mathrm{Y}|\mathrm{X}=% r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}\end{aligned}= start_ROW start_CELL end_CELL start_CELL ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL - italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG end_CELL end_ROW
=(a)y𝒴WY|X=x(y)logWY|X=x(y)WY|X=r(x)(y)+logWY|X=r(x)(y)QY,η(y)(WY|X=x(y)WY|X=r(x)(y))𝑎missing-subexpressionsubscript𝑦𝒴subscript𝑊conditionalYXsuperscript𝑥𝑦subscript𝑊conditionalYXsuperscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦missing-subexpressionsubscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦subscript𝑊conditionalYXsuperscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦\displaystyle\overset{(a)}{=}\begin{aligned} &\sum_{y\in\mathcal{Y}}W_{\mathrm% {Y}|\mathrm{X}=x^{\star}}(y)\log\frac{W_{\mathrm{Y}|\mathrm{X}=x^{\star}}(y)}{% W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}\\ &+\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{% \star}(y)}\left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}}(y)-W_{\mathrm{Y}|\mathrm{X% }=r(x^{\star})}(y)\right)\end{aligned}start_OVERACCENT ( italic_a ) end_OVERACCENT start_ARG = end_ARG start_ROW start_CELL end_CELL start_CELL ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) - italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ) end_CELL end_ROW
=(b)D(WY|X=xWY|X=r(x))+y𝒴logWY|X=r(x)(y)QY,η(y)(WY|X=x(y)WY|X=r(x)(y))𝑏missing-subexpressionDconditionalsubscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥missing-subexpressionsubscript𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦subscript𝑊conditionalYXsuperscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦\displaystyle\overset{(b)}{=}\begin{aligned} &\mathrm{D}\left(W_{\mathrm{Y}|% \mathrm{X}=x^{\star}}\|W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}\right)\\ &+\sum_{y\in\mathcal{Y}}\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q% _{\mathrm{Y},\eta}^{\star}(y)}\left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}}(y)\!-% \!W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\right)\end{aligned}start_OVERACCENT ( italic_b ) end_OVERACCENT start_ARG = end_ARG start_ROW start_CELL end_CELL start_CELL roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ∥ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) - italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ) end_CELL end_ROW
=D(WY|X=xWY|X=r(x))+y𝒴WY|X=r(x)(y)logWY|X=r(x)(y)QY,η(y)(WY|X=x(y)WY|X=r(x)(y))WY|X=r(x)(y)absentmissing-subexpressionDconditionalsubscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥missing-subexpressionsubscript𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦missing-subexpressionabsentsubscript𝑊conditionalYXsuperscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦\displaystyle=\begin{aligned} &\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x^{% \star}}\|W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}\right)\\ &+\sum_{y\in\mathcal{Y}}\sqrt{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}\log% \frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y% )}\\ &\cdot\frac{\left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}}(y)-W_{\mathrm{Y}|\mathrm% {X}=r(x^{\star})}(y)\right)}{\sqrt{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}}% \end{aligned}= start_ROW start_CELL end_CELL start_CELL roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ∥ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT square-root start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL ⋅ divide start_ARG ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) - italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ) end_ARG start_ARG square-root start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG end_ARG end_CELL end_ROW
(c)D(WY|X=xWY|X=r(x))+y𝒴WY|X=r(x)(y)log2WY|X=r(x)(y)QY,η(y)y𝒴(WY|X=x(y)WY|X=r(x)(y))2WY|X=r(x)(y)𝑐missing-subexpressionDconditionalsubscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥missing-subexpressionsubscript𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦missing-subexpressionabsentsubscript𝑦𝒴superscriptsubscript𝑊conditionalYXsuperscript𝑥𝑦subscript𝑊conditionalYX𝑟superscript𝑥𝑦2subscript𝑊conditionalYX𝑟superscript𝑥𝑦\displaystyle\overset{(c)}{\leq}\begin{aligned} &\mathrm{D}\left(W_{\mathrm{Y}% |\mathrm{X}=x^{\star}}\|W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}\right)\\ &+\sqrt{\sum_{y\in\mathcal{Y}}W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\log^{2% }\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(% y)}}\\ &\cdot\sqrt{\sum_{y\in\mathcal{Y}}\frac{\left(W_{\mathrm{Y}|\mathrm{X}=x^{% \star}}(y)-W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\right)^{2}}{W_{\mathrm{Y}% |\mathrm{X}=r(x^{\star})}(y)}}\end{aligned}start_OVERACCENT ( italic_c ) end_OVERACCENT start_ARG ≤ end_ARG start_ROW start_CELL end_CELL start_CELL roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ∥ italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + square-root start_ARG ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG end_ARG end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL ⋅ square-root start_ARG ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT divide start_ARG ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ( italic_y ) - italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ) start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_ARG start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG end_ARG end_CELL end_ROW
(d)χ2(WY|X=x,WY|X=r(x))+𝔼WY|X=r(x)[log2WY|X=r(x)(y)QY,η(y)]χ2(WY|X=x,WY|X=r(x))𝑑missing-subexpressionsuperscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥missing-subexpressionsubscript𝔼subscript𝑊conditionalYX𝑟superscript𝑥delimited-[]superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦missing-subexpressionabsentsuperscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥\displaystyle\overset{(d)}{\leq}\begin{aligned} &\chi^{2}\left(W_{\mathrm{Y}|% \mathrm{X}=x^{\star}},W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}\right)\\ &+\sqrt{\mathbb{E}_{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}\left[\log^{2}\frac% {W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}% \right]}\\ &\cdot\sqrt{\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}},W_{\mathrm{Y}|% \mathrm{X}=r(x^{\star})}\right)}\end{aligned}start_OVERACCENT ( italic_d ) end_OVERACCENT start_ARG ≤ end_ARG start_ROW start_CELL end_CELL start_CELL italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + square-root start_ARG blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] end_ARG end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL ⋅ square-root start_ARG italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_ARG end_CELL end_ROW

where (aa)𝑎𝑎(aa)( italic_a italic_a ) is because for every r𝑟r\in\mathcal{R}italic_r ∈ caligraphic_R including r(x)𝑟superscript𝑥r(x^{\star})italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) we have D(WY|X=rQY,η)<maxrD(WY|X=rQY,η)Dconditionalsubscript𝑊conditionalYX𝑟superscriptsubscript𝑄Y𝜂subscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟superscriptsubscript𝑄Y𝜂\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{\mathrm{Y},\eta}^{\star}\right% )<\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{% \mathrm{Y},\eta}^{\star}\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) < roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ). Further (a)𝑎(a)( italic_a ) holds since abcd=abad+adcd=a(bd)+d(ac)𝑎𝑏𝑐𝑑𝑎𝑏𝑎𝑑𝑎𝑑𝑐𝑑𝑎𝑏𝑑𝑑𝑎𝑐ab-cd=ab-ad+ad-cd=a(b-d)+d(a-c)italic_a italic_b - italic_c italic_d = italic_a italic_b - italic_a italic_d + italic_a italic_d - italic_c italic_d = italic_a ( italic_b - italic_d ) + italic_d ( italic_a - italic_c ), (b)𝑏(b)( italic_b ) follows from rearranging the terms, and (c)𝑐(c)( italic_c ) holds by Cauchy–Schwarz, and (d)𝑑(d)( italic_d ) holds since χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divergence is an upper bound to KL-divergence and the definition of the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-divergence. It remains to bound the second moment of logWY|X=r(x)(y)QY,η(y)subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{% \star}(y)}roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG.

Let WY|X=r(x)subscriptsuperscript𝑊conditionalYX𝑟superscript𝑥W^{\prime}_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}italic_W start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT be the probability distribution over the subset of symbols in 𝒴𝒴\mathcal{Y}caligraphic_Y that correspond to non-zero entries in WY|X=r(x)subscript𝑊conditionalYX𝑟superscript𝑥W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT, using κ(x)miny𝒴:WY|X=r(x)(y)0(WY|X=r(x)(y))𝜅𝑥subscript:𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦0subscript𝑊conditionalYX𝑟superscript𝑥𝑦\kappa(x)\triangleq\min_{y\in\mathcal{Y}:W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})% }(y)\neq 0}\left(W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\right)italic_κ ( italic_x ) ≜ roman_min start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y : italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ≠ 0 end_POSTSUBSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ) as the smallest non-zero number in WY|X=r(x)subscript𝑊conditionalYX𝑟superscript𝑥W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT, we have

y𝒴WY|X=r(x)(y)log2WY|X=r(x)(y)QY,η(y)subscript𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\sum_{y\in\mathcal{Y}}W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)% \log^{2}\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{% \star}(y)}∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG
=𝔼WY|X=r(x)[log2WY|X=r(x)(y)QY,η(y)]absentsubscript𝔼subscript𝑊conditionalYX𝑟superscript𝑥delimited-[]superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle=\mathbb{E}_{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}\left[\log^{2% }\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(% y)}\right]= blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ]
(e)𝔼WY|X=r(x)[logWY|X=r(x)(y)QY,η(y)]2+VarWY|X=r(x)[logWY|X=r(x)(y)QY,η(y)]𝑒missing-subexpressionsubscript𝔼subscript𝑊conditionalYX𝑟superscript𝑥superscriptdelimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦2missing-subexpressionsubscriptVarsubscript𝑊conditionalYX𝑟superscript𝑥delimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\overset{(e)}{\leq}\begin{aligned} &\mathbb{E}_{W_{\mathrm{Y}|% \mathrm{X}=r(x^{\star})}}\left[\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})% }(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}\right]^{2}\\ &+\text{Var}_{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}\left[\log\frac{W_{% \mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}\right]% \end{aligned}start_OVERACCENT ( italic_e ) end_OVERACCENT start_ARG ≤ end_ARG start_ROW start_CELL end_CELL start_CELL blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + Var start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] end_CELL end_ROW
=(f)𝔼WY|X=r(x)[logWY|X=r(x)(y)QY,η(y)]2+VarWY|X=r(x)[logWY|X=r(x)(y)QY,η(y)]𝑓missing-subexpressionsubscript𝔼subscript𝑊conditionalYX𝑟superscript𝑥superscriptdelimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦2missing-subexpressionsubscriptVarsubscriptsuperscript𝑊conditionalYX𝑟superscript𝑥delimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\overset{(f)}{=}\begin{aligned} &\mathbb{E}_{W_{\mathrm{Y}|% \mathrm{X}=r(x^{\star})}}\left[\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})% }(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}\right]^{2}\\ &+\text{Var}_{W^{\prime}_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}\left[\log\frac{% W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}% \right]\end{aligned}start_OVERACCENT ( italic_f ) end_OVERACCENT start_ARG = end_ARG start_ROW start_CELL end_CELL start_CELL blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + Var start_POSTSUBSCRIPT italic_W start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] end_CELL end_ROW
(g)C(W)2+(log(1/η)C(W))(C(W)log(κ(x)))𝑔𝐶superscriptsubscript𝑊21𝜂𝐶subscript𝑊𝐶subscript𝑊𝜅𝑥\displaystyle\overset{(g)}{\leq}C(W_{\mathcal{R}})^{2}+\left(\log(1/\eta)-C(W_% {\mathcal{R}})\right)\left(C(W_{\mathcal{R}})-\log(\kappa(x))\right)start_OVERACCENT ( italic_g ) end_OVERACCENT start_ARG ≤ end_ARG italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT + ( roman_log ( 1 / italic_η ) - italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) ) ( italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) - roman_log ( italic_κ ( italic_x ) ) )
=C(W)log(κ(x)η)+log(1η)log(1κ(x)),absent𝐶subscript𝑊𝜅𝑥𝜂1𝜂1𝜅𝑥\displaystyle=C(W_{\mathcal{R}})\log\left(\frac{\kappa(x)}{\eta}\right)+\log% \left(\frac{1}{\eta}\right)\log\left(\frac{1}{\kappa(x)}\right),= italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) roman_log ( divide start_ARG italic_κ ( italic_x ) end_ARG start_ARG italic_η end_ARG ) + roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) roman_log ( divide start_ARG 1 end_ARG start_ARG italic_κ ( italic_x ) end_ARG ) ,

where (e)𝑒(e)( italic_e ) holds from the decomposition of second moments in terms of first moment and variance. (f)𝑓(f)( italic_f ) holds by the derivations in the sequel, and (g)𝑔(g)( italic_g ) holds by applying Bhatia–Davis inequality where the variance calculation is limited to non-zero entries in WY|X=r(x)subscript𝑊conditionalYX𝑟superscript𝑥W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT. Hence, the random variable logWY|X=r(x)(y)QY,η(y)subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{% \star}(y)}roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG is bounded as

log(κ(x))miny𝒴:WY|X=r(x)(y)0logWY|X=r(x)(y)QY,η(y)𝜅𝑥subscript:𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦0subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\log(\kappa(x))\leq\min_{y\in\mathcal{Y}:W_{\mathrm{Y}|\mathrm{X}% =r(x^{\star})}(y)\neq 0}\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q% _{\mathrm{Y},\eta}^{\star}(y)}roman_log ( italic_κ ( italic_x ) ) ≤ roman_min start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y : italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ≠ 0 end_POSTSUBSCRIPT roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG
logWY|X=r(x)(y)QY,η(y)absentsubscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\leq\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{% \mathrm{Y},\eta}^{\star}(y)}≤ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG
maxy𝒴:WY|X=r(x)(y)0logWY|X=r(x)(y)QY,η(y)log(1η).absentsubscript:𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦0subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦1𝜂\displaystyle\leq\max_{y\in\mathcal{Y}:W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(% y)\neq 0}\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},% \eta}^{\star}(y)}\leq\log\left(\frac{1}{\eta}\right).≤ roman_max start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y : italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ≠ 0 end_POSTSUBSCRIPT roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ≤ roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) .

(f)𝑓(f)( italic_f ) follows from (e)𝑒(e)( italic_e ) since

VarWY|X=r(x)[logWY|X=r(x)(y)QY,η(y)]subscriptVarsubscript𝑊conditionalYX𝑟superscript𝑥delimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\text{Var}_{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}\left[\log% \frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y% )}\right]Var start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ]
=𝔼WY|X=r(x)[logWY|X=r(x)(y)QY,η(y)]2+𝔼WY|X=r(x)[log2WY|X=r(x)(y)QY,η(y)]absentmissing-subexpressionsubscript𝔼subscript𝑊conditionalYX𝑟superscript𝑥superscriptdelimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦2missing-subexpressionsubscript𝔼subscript𝑊conditionalYX𝑟superscript𝑥delimited-[]superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle=\begin{aligned} &\mathbb{E}_{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star% })}}\left[\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},% \eta}^{\star}(y)}\right]^{2}\\ &+\mathbb{E}_{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}\left[\log^{2}\frac{W_{% \mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}\right]% \end{aligned}= start_ROW start_CELL end_CELL start_CELL blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] end_CELL end_ROW
=y𝒴WY|X=r(x)(y)[logWY|X=r(x)(y)QY,η(y)]2+y𝒴WY|X=r(x)(y)[log2WY|X=r(x)(y)QY,η(y)]absentmissing-subexpressionsubscript𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptdelimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦2missing-subexpressionsubscript𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦delimited-[]superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle=\begin{aligned} &\sum_{y\in\mathcal{Y}}W_{\mathrm{Y}|\mathrm{X}=% r(x^{\star})}(y)\left[\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{% \mathrm{Y},\eta}^{\star}(y)}\right]^{2}\\ &+\sum_{y\in\mathcal{Y}}W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\left[\log^{2% }\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(% y)}\right]\end{aligned}= start_ROW start_CELL end_CELL start_CELL ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) [ roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] end_CELL end_ROW
=(a)y𝒴:WY|X=r(x)(y)0WY|X=r(x)(y)[logWY|X=r(x)(y)QY,η(y)]2+y𝒴:WY|X=r(x)(y)0WY|X=r(x)(y)[log2WY|X=r(x)(y)QY,η(y)]𝑎missing-subexpressionsubscript:𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦0subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptdelimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦2missing-subexpressionsubscript:𝑦𝒴subscript𝑊conditionalYX𝑟superscript𝑥𝑦0subscript𝑊conditionalYX𝑟superscript𝑥𝑦delimited-[]superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\overset{(a)}{=}\begin{aligned} &\sum_{y\in\mathcal{Y}:W_{\mathrm% {Y}|\mathrm{X}=r(x^{\star})}(y)\neq 0}\!\!\!\!\!\!\!W_{\mathrm{Y}|\mathrm{X}=r% (x^{\star})}(y)\left[\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{% \mathrm{Y},\eta}^{\star}(y)}\right]^{2}\\ &+\sum_{y\in\mathcal{Y}:W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\neq 0}\!\!\!% \!\!\!\!W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)\left[\log^{2}\frac{W_{% \mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}\right]% \end{aligned}start_OVERACCENT ( italic_a ) end_OVERACCENT start_ARG = end_ARG start_ROW start_CELL end_CELL start_CELL ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y : italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ≠ 0 end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + ∑ start_POSTSUBSCRIPT italic_y ∈ caligraphic_Y : italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) ≠ 0 end_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) [ roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] end_CELL end_ROW
=(b)𝔼WY|X=r(x)[logWY|X=r(x)(y)QY,η(y)]2+𝔼WY|X=r(x)[log2WY|X=r(x)(y)QY,η(y)]𝑏missing-subexpressionsubscript𝔼subscriptsuperscript𝑊conditionalYX𝑟superscript𝑥superscriptdelimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦2missing-subexpressionsubscript𝔼subscriptsuperscript𝑊conditionalYX𝑟superscript𝑥delimited-[]superscript2subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle\overset{(b)}{=}\begin{aligned} &\mathbb{E}_{W^{\prime}_{\mathrm{% Y}|\mathrm{X}=r(x^{\star})}}\left[\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{% \star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y)}\right]^{2}\\ &+\mathbb{E}_{W^{\prime}_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}\left[\log^{2}% \frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}^{\star}(y% )}\right]\end{aligned}start_OVERACCENT ( italic_b ) end_OVERACCENT start_ARG = end_ARG start_ROW start_CELL end_CELL start_CELL blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + blackboard_E start_POSTSUBSCRIPT italic_W start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] end_CELL end_ROW
=VarWY|X=r(x)[logWY|X=r(x)(y)QY,η(y)],absentsubscriptVarsubscriptsuperscript𝑊conditionalYX𝑟superscript𝑥delimited-[]subscript𝑊conditionalYX𝑟superscript𝑥𝑦superscriptsubscript𝑄Y𝜂𝑦\displaystyle=\text{Var}_{W^{\prime}_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}}% \left[\log\frac{W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}(y)}{Q_{\mathrm{Y},\eta}% ^{\star}(y)}\right],= Var start_POSTSUBSCRIPT italic_W start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT end_POSTSUBSCRIPT [ roman_log divide start_ARG italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ( italic_y ) end_ARG start_ARG italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_y ) end_ARG ] ,

where (a)𝑎(a)( italic_a ) holds since xlogx=0𝑥𝑥0x\log x=0italic_x roman_log italic_x = 0 and (b)𝑏(b)( italic_b ) holds since WY|X=r(x)subscript𝑊conditionalYX𝑟superscript𝑥W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT is a valid probability distribution over a subset of 𝒴𝒴\mathcal{Y}caligraphic_Y. This concludes the proof. ∎

-D Proof of Theorem 1

To prove Theorem 1, we rely on the following lemma to turn the conditional distributions in the convex hull that are close to symbols outside the convex hull in terms of χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-divergence into their nearest neighbors.

Lemma 2.

Let \mathcal{R}caligraphic_R be a set of symbols whose conditional distributions span the convex hull Conv({WY|X=r}r)Convsubscriptsubscript𝑊conditionalYX𝑟𝑟\operatorname{Conv}\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\right)roman_Conv ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ) such that each symbol not in the convex hull n𝒩𝒳𝑛𝒩𝒳n\in\mathcal{N}\subset\mathcal{X}\setminus\mathcal{R}italic_n ∈ caligraphic_N ⊂ caligraphic_X ∖ caligraphic_R has a bounded difference of ε(n)subscript𝜀𝑛\varepsilon_{\mathcal{R}}(n)italic_ε start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ( italic_n ) in terms of χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT divergence. Let for each n𝒩𝑛𝒩n\in\mathcal{N}italic_n ∈ caligraphic_N the distribution WY,nsuperscriptsubscript𝑊𝑌𝑛W_{Y,n}^{\star}italic_W start_POSTSUBSCRIPT italic_Y , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT be the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-closest of WY|X=nsubscript𝑊conditionalYX𝑛W_{\mathrm{Y}|\mathrm{X}=n}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT to the convex hull Conv({WY|X=r}r)Convsubscriptsubscript𝑊conditionalYX𝑟𝑟\operatorname{Conv}\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\right)roman_Conv ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ), then we have

C𝐶\displaystyle Citalic_C ({WY|X=x}x𝒳)C({WY|X=r}r)subscriptsubscript𝑊conditionalYX𝑥𝑥𝒳𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟\displaystyle\left(\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{X}}\right)-C% \left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\right)( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT ) - italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT )
C({WY|X=r}r{WY|X=n}n𝒩)C({WY|X=r}r{WY,n}n𝒩).absentmissing-subexpression𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsubscript𝑊conditionalYX𝑛𝑛𝒩missing-subexpression𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsuperscriptsubscript𝑊𝑌𝑛𝑛𝒩\displaystyle\leq\begin{aligned} &C\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in% \mathcal{R}}\cup\{W_{\mathrm{Y}|\mathrm{X}=n}\}_{n\in\mathcal{N}}\right)\\ &-C(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\cup\{W_{Y,n}^{\star}\}_{% n\in\mathcal{N}}).\end{aligned}≤ start_ROW start_CELL end_CELL start_CELL italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT ) end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL - italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT italic_Y , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT ) . end_CELL end_ROW
Proof.

We prove this lemma in Section -E. ∎

By applying the above lemma, bounding the capacity loss based on the χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT distance of each symbol outside the convex hull to its nearest neighbor in \mathcal{R}caligraphic_R is equivalent to considering the distance to the convex hull, since we can artificially transform the channel to contain all the points that minimize the distance to the convex hull to nearest neighbors that are actually part of the channel.

Proof of Theorem 1.

Let QY,ηsuperscriptsubscript𝑄Y𝜂Q_{\mathrm{Y},\eta}^{\star}italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT be the unique minimizer of

minQY𝒫η(𝒴)maxrD(WY|X=rQY).subscriptsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟subscript𝑄Y\displaystyle\min_{Q_{\mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal{Y})}\max_{r\in% \mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{\mathrm{Y}}\right).roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) .

Then, using triangle inequality, Lemma 1, and results from [1, Lemma 1], we can write for any x𝒳superscript𝑥𝒳x^{\star}\in\mathcal{X}italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ∈ caligraphic_X that maximizes D(WY|X=xQY,η(W))Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂subscript𝑊\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}(W_{% \mathcal{R}})\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) ) and the distance to its nearest neighbor, which at the same time is the distribution WY,xsuperscriptsubscript𝑊𝑌superscript𝑥W_{Y,x^{\star}}^{\star}italic_W start_POSTSUBSCRIPT italic_Y , italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT on the convex hull closest to WY|X=xsubscript𝑊conditionalYXsuperscript𝑥W_{\mathrm{Y}|\mathrm{X}=x^{\star}}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT in terms of χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT-distance (χ2(WY|X=x,WY|X=r(x))superscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}},W_{\mathrm{Y}|\mathrm{X}=r(x% ^{\star})}\right)italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ), that

minQY𝒫(𝒴)maxx𝒳D(WY|X=xQY)maxrD(WY|X=rQY)subscriptsubscript𝑄Y𝒫𝒴subscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥subscript𝑄Ysubscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟superscriptsubscript𝑄𝑌\displaystyle\min_{Q_{\mathrm{Y}}\in\mathcal{P}(\mathcal{Y})}\max_{x\in% \mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}\right)% -\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q_{Y}^{% \star}\right)roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) - roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT italic_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )
|minQYmaxx𝒳D(WY|X=xQY)minQY𝒫η(𝒴)maxx𝒳D(WY|X=xQY)|+|minQY𝒫η(𝒴)maxx𝒳D(WY|X=xQY)maxrD(WY|X=rQY,η)|+|maxrD(WY|X=rQY,η)maxrD(WY|X=rQY)|\displaystyle\leq\!\!\!\begin{aligned} &\bigg{|}\!\min_{Q_{\mathrm{Y}}}\!\max_% {x\in\mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}% \right)\!-\!\!\!\!\!\!\!\!\min_{Q_{\mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal{Y% })}\!\max_{x\in\mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{% \mathrm{Y}}\right)\!\!\bigg{|}\\ &\!\!\!\!\!\!\!\!+\bigg{|}\!\min_{Q_{\mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal% {Y})}\!\max_{x\in\mathcal{X}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{% \mathrm{Y}}\right)\!-\!\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|% \mathrm{X}=r}\|Q_{\mathrm{Y},\eta}^{\star}\right)\!\!\bigg{|}\\ &+\bigg{|}\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=r}\|Q% _{\mathrm{Y},\eta}^{\star}\right)-\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{% \mathrm{Y}|\mathrm{X}=r}\|Q_{\mathrm{Y}}^{\star}\right)\bigg{|}\end{aligned}≤ start_ROW start_CELL end_CELL start_CELL | roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) - roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) | end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + | roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ) - roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) | end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + | roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) - roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) | end_CELL end_ROW
(a)4|𝒴|η+minQY𝒫η(𝒴)maxx𝒳D(WY|X=xQY)𝑎4𝒴𝜂subscriptsubscript𝑄Ysubscript𝒫𝜂𝒴subscript𝑥𝒳Dconditionalsubscript𝑊conditionalYX𝑥subscript𝑄Y\displaystyle\overset{(a)}{\leq}\!\!4|\mathcal{Y}|\eta+\!\!\!\!\!\!\!\min_{Q_{% \mathrm{Y}}\in\mathcal{P}_{\eta}(\mathcal{Y})}\max_{x\in\mathcal{X}}\mathrm{D}% \left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}\right)\!start_OVERACCENT ( italic_a ) end_OVERACCENT start_ARG ≤ end_ARG 4 | caligraphic_Y | italic_η + roman_min start_POSTSUBSCRIPT italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT ∈ caligraphic_P start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( caligraphic_Y ) end_POSTSUBSCRIPT roman_max start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT )
maxrD(WY|X=rQY,η)subscript𝑟Dconditionalsubscript𝑊conditionalYX𝑟superscriptsubscript𝑄Y𝜂\displaystyle\qquad\qquad\qquad-\max_{r\in\mathcal{R}}\mathrm{D}\left(W_{% \mathrm{Y}|\mathrm{X}=r}\|Q_{\mathrm{Y},\eta}^{\star}\right)- roman_max start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )
(b)4|𝒴|η+χ2(WY|X=x,WY|X=r(x))+Cη(W)log(κ(x)η)+log(1η)log(1κ(x))χ2(WY|X=x,WY|X=r(x))𝑏missing-subexpression4𝒴𝜂superscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥missing-subexpressionsubscript𝐶𝜂subscript𝑊𝜅𝑥𝜂1𝜂1𝜅𝑥missing-subexpressionabsentsuperscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥\displaystyle\overset{(b)}{\leq}\begin{aligned} &4|\mathcal{Y}|\eta+\chi^{2}% \left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}},W_{\mathrm{Y}|\mathrm{X}=r(x^{\star}% )}\right)\\ &+\sqrt{C_{\eta}(W_{\mathcal{R}})\log\left(\frac{\kappa(x)}{\eta}\right)+\log% \left(\frac{1}{\eta}\right)\log\left(\frac{1}{\kappa(x)}\right)}\\ &\cdot\sqrt{\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}},W_{\mathrm{Y}|% \mathrm{X}=r(x^{\star})}\right)}\end{aligned}start_OVERACCENT ( italic_b ) end_OVERACCENT start_ARG ≤ end_ARG start_ROW start_CELL end_CELL start_CELL 4 | caligraphic_Y | italic_η + italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + square-root start_ARG italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) roman_log ( divide start_ARG italic_κ ( italic_x ) end_ARG start_ARG italic_η end_ARG ) + roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) roman_log ( divide start_ARG 1 end_ARG start_ARG italic_κ ( italic_x ) end_ARG ) end_ARG end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL ⋅ square-root start_ARG italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_ARG end_CELL end_ROW
(c)4|𝒴|η+χ2(WY|X=x,WY|X=r(x))+log(1η)(C(W)+log(1κ(x)))χ2(WY|X=x,WY|X=r(x)),𝑐missing-subexpression4𝒴𝜂superscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥missing-subexpression1𝜂𝐶subscript𝑊1𝜅𝑥missing-subexpressionabsentsuperscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥\displaystyle\overset{(c)}{\leq}\begin{aligned} &4|\mathcal{Y}|\eta+\chi^{2}% \left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}},W_{\mathrm{Y}|\mathrm{X}=r(x^{\star}% )}\right)\\ &+\sqrt{\log\left(\frac{1}{\eta}\right)\left(C(W_{\mathcal{R}})+\log\left(% \frac{1}{\kappa(x)}\right)\right)}\\ &\cdot\sqrt{\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x^{\star}},W_{\mathrm{Y}|% \mathrm{X}=r(x^{\star})}\right)},\end{aligned}start_OVERACCENT ( italic_c ) end_OVERACCENT start_ARG ≤ end_ARG start_ROW start_CELL end_CELL start_CELL 4 | caligraphic_Y | italic_η + italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL + square-root start_ARG roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) ( italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) + roman_log ( divide start_ARG 1 end_ARG start_ARG italic_κ ( italic_x ) end_ARG ) ) end_ARG end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL ⋅ square-root start_ARG italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_ARG , end_CELL end_ROW

where (a)𝑎(a)( italic_a ) follows from applying [19, Lemma 1] twice; (b)𝑏(b)( italic_b ) follows from Lemma 1; and (c)𝑐(c)( italic_c ) is by bounding Cη(W)C(W)subscript𝐶𝜂subscript𝑊𝐶subscript𝑊C_{\eta}(W_{\mathcal{R}})\leq C(W_{\mathcal{R}})italic_C start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) ≤ italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) and from log(κη)log(1η)𝜅𝜂1𝜂\log\left(\frac{\kappa}{\eta}\right)\leq\log\left(\frac{1}{\eta}\right)roman_log ( divide start_ARG italic_κ end_ARG start_ARG italic_η end_ARG ) ≤ roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ).

This emits a bias-variance-type trade-off based on the parameter η𝜂\etaitalic_η, and it remains to appropriately choose η𝜂\etaitalic_η. The function

log(1η)(C(W)+log(1κ(x)))1𝜂𝐶subscript𝑊1𝜅𝑥\displaystyle\sqrt{\log\left(\frac{1}{\eta}\right)\left(C(W_{\mathcal{R}})+% \log\left(\frac{1}{\kappa(x)}\right)\right)}square-root start_ARG roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) ( italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) + roman_log ( divide start_ARG 1 end_ARG start_ARG italic_κ ( italic_x ) end_ARG ) ) end_ARG

is strictly decreasing in η𝜂\etaitalic_η, with a sharp decrease around 00. With constants k1:=4|𝒴|assignsubscript𝑘14𝒴k_{1}:=4|\mathcal{Y}|italic_k start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT := 4 | caligraphic_Y | and k2:=C(W)log(κ(x))assignsubscript𝑘2𝐶subscript𝑊𝜅𝑥k_{2}:=\sqrt{C(W_{\mathcal{R}})-\log\left(\kappa(x)\right)}italic_k start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT := square-root start_ARG italic_C ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) - roman_log ( italic_κ ( italic_x ) ) end_ARG, we find a good η𝜂\etaitalic_η by solving

ηsuperscript𝜂\displaystyle\eta^{\star}italic_η start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT =argminηk1η+k2log(1η).absentsubscriptargmin𝜂subscript𝑘1𝜂subscript𝑘21𝜂\displaystyle=\operatorname*{arg\,min}_{\eta}\,k_{1}\eta+k_{2}\sqrt{\log\left(% \frac{1}{\eta}\right)}.= start_OPERATOR roman_arg roman_min end_OPERATOR start_POSTSUBSCRIPT italic_η end_POSTSUBSCRIPT italic_k start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT italic_η + italic_k start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT square-root start_ARG roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) end_ARG .

Therefore, we analyze the derivatives w.r.t. η𝜂\etaitalic_η:

log(1η)η=12ηlog(1η)1η0.07,1𝜂𝜂12𝜂1𝜂1𝜂0.07\displaystyle\frac{\partial\sqrt{\log\left(\frac{1}{\eta}\right)}}{\partial% \eta}=-\frac{1}{2\eta\sqrt{\log\left(\frac{1}{\eta}\right)}}\approx-\frac{1}{% \sqrt{\eta}-0.07},divide start_ARG ∂ square-root start_ARG roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) end_ARG end_ARG start_ARG ∂ italic_η end_ARG = - divide start_ARG 1 end_ARG start_ARG 2 italic_η square-root start_ARG roman_log ( divide start_ARG 1 end_ARG start_ARG italic_η end_ARG ) end_ARG end_ARG ≈ - divide start_ARG 1 end_ARG start_ARG square-root start_ARG italic_η end_ARG - 0.07 end_ARG ,

where the approximation is valid in the regime of interest (η1much-less-than𝜂1\eta\ll 1italic_η ≪ 1). Hence, we can choose η𝜂\etaitalic_η to equalize the derivatives of the above and the linear dependency 4η|𝒴|4𝜂𝒴4\eta|\mathcal{Y}|4 italic_η | caligraphic_Y | on η𝜂\etaitalic_η. We have

k1=k2χ2(WY|X=x,WY|X=r(x))1η0.07subscript𝑘1subscript𝑘2superscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥1𝜂0.07absent\displaystyle k_{1}=k_{2}\cdot\sqrt{\chi^{2}\left(W_{\mathrm{Y}|\mathrm{X}=x^{% \star}},W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}\right)}\cdot\frac{1}{\sqrt{\eta% }-0.07}\Leftrightarrowitalic_k start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = italic_k start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ⋅ square-root start_ARG italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_ARG ⋅ divide start_ARG 1 end_ARG start_ARG square-root start_ARG italic_η end_ARG - 0.07 end_ARG ⇔
η=(k2χ2(WY|X=x,WY|X=r(x))k1+0.07)2𝜂superscriptsubscript𝑘2superscript𝜒2subscript𝑊conditionalYXsuperscript𝑥subscript𝑊conditionalYX𝑟superscript𝑥subscript𝑘10.072\displaystyle\eta=\left(\frac{k_{2}\cdot\sqrt{\chi^{2}\left(W_{\mathrm{Y}|% \mathrm{X}=x^{\star}},W_{\mathrm{Y}|\mathrm{X}=r(x^{\star})}\right)}}{k_{1}}+0% .07\right)^{2}italic_η = ( divide start_ARG italic_k start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ⋅ square-root start_ARG italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r ( italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ) end_POSTSUBSCRIPT ) end_ARG end_ARG start_ARG italic_k start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT end_ARG + 0.07 ) start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT

However, the value for η𝜂\etaitalic_η is conditioned on the choice of the symbol xsuperscript𝑥x^{\star}italic_x start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT, which is supposed to be the symbol that maximizes D(WY|X=xQY,η(W))Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂subscript𝑊\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}(W_{% \mathcal{R}})\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) ). To determine this symbol, we must fix η𝜂\etaitalic_η on the other hand. Therefore, we use the fact that the bound holds uniformly for all η𝜂\etaitalic_η. Hence, we choose for the calculation of η𝜂\etaitalic_η the symbol that maximizes the difference to the capacity achieving output distribution D(WY|X=xQY(W))Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Ysubscript𝑊\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y}}^{\star}(W_{% \mathcal{R}})\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) ), which is expected to be close to D(WY|X=xQY,η(W))Dconditionalsubscript𝑊conditionalYX𝑥superscriptsubscript𝑄Y𝜂subscript𝑊\mathrm{D}\left(W_{\mathrm{Y}|\mathrm{X}=x}\|Q_{\mathrm{Y},\eta}^{\star}(W_{% \mathcal{R}})\right)roman_D ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT ∥ italic_Q start_POSTSUBSCRIPT roman_Y , italic_η end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT caligraphic_R end_POSTSUBSCRIPT ) ) and consequently a good choice for computing η𝜂\etaitalic_η. This concludes the proof of the theorem. ∎

-E Proof of Lemma 2

Proof.

Let 𝒳𝒳\mathcal{I}\subset\mathcal{X}\setminus\mathcal{R}caligraphic_I ⊂ caligraphic_X ∖ caligraphic_R be the symbols whose conditionals WY|X=i,isubscript𝑊conditionalYX𝑖𝑖W_{\mathrm{Y}|\mathrm{X}=i},i\in\mathcal{I}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_i end_POSTSUBSCRIPT , italic_i ∈ caligraphic_I are contained in the convex hull of the conditionals of symbols in \mathcal{R}caligraphic_R. Let for each symbol n𝒩𝒳𝑛𝒩𝒳n\in\mathcal{N}\subset\mathcal{X}\setminus\mathcal{R}italic_n ∈ caligraphic_N ⊂ caligraphic_X ∖ caligraphic_R whose conditional WY|X=isubscript𝑊conditionalYX𝑖W_{\mathrm{Y}|\mathrm{X}=i}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_i end_POSTSUBSCRIPT is not contained in the convex hull of symbols in \mathcal{R}caligraphic_R the distribution WY,nsuperscriptsubscript𝑊𝑌𝑛W_{Y,n}^{\star}italic_W start_POSTSUBSCRIPT italic_Y , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT in the convex hull be closest to WY|X=nsubscript𝑊conditionalYX𝑛W_{\mathrm{Y}|\mathrm{X}=n}italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT in terms of χ2superscript𝜒2\chi^{2}italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT distance, i.e.,

WY,n=argminWYConv({WY|X=r}r)χ2(WY|X=n,WY).superscriptsubscript𝑊𝑌𝑛subscriptargminsubscript𝑊𝑌Convsubscriptsubscript𝑊conditionalYX𝑟𝑟superscript𝜒2subscript𝑊conditionalYX𝑛subscript𝑊𝑌\displaystyle W_{Y,n}^{\star}=\operatorname*{arg\,min}_{W_{Y}\in\operatorname{% Conv}\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\right)}\chi^{2}% \left(W_{\mathrm{Y}|\mathrm{X}=n},W_{Y}\right).italic_W start_POSTSUBSCRIPT italic_Y , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT = start_OPERATOR roman_arg roman_min end_OPERATOR start_POSTSUBSCRIPT italic_W start_POSTSUBSCRIPT italic_Y end_POSTSUBSCRIPT ∈ roman_Conv ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ) end_POSTSUBSCRIPT italic_χ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ( italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT , italic_W start_POSTSUBSCRIPT italic_Y end_POSTSUBSCRIPT ) .

Since symbols are either in the convex hull or not, all sets are disjoint and we have that 𝒳=𝒩𝒳𝒩\mathcal{X}=\mathcal{R}\cup\mathcal{I}\cup\mathcal{N}caligraphic_X = caligraphic_R ∪ caligraphic_I ∪ caligraphic_N. Consider a set of conditional probabilities containing those given by symbols in \mathcal{R}caligraphic_R and \mathcal{I}caligraphic_I and those given by distributions in the convex hull of \mathcal{R}caligraphic_R closest to all symbols in 𝒩𝒩\mathcal{N}caligraphic_N. Then we can bound the capacity difference as follows:

C({WY|X=x}x𝒳)C({WY|X=r}r)𝐶subscriptsubscript𝑊conditionalYX𝑥𝑥𝒳𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟\displaystyle C\left(\{W_{\mathrm{Y}|\mathrm{X}=x}\}_{x\in\mathcal{X}}\right)-% C\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\right)italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_x end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_x ∈ caligraphic_X end_POSTSUBSCRIPT ) - italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT )
=C({WY|X=r}r{WY|X=n}n𝒩{WY|X=i}i)absent𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsubscript𝑊conditionalYX𝑛𝑛𝒩subscriptsubscript𝑊conditionalYX𝑖𝑖\displaystyle=C\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\!\cup\!% \{W_{\mathrm{Y}|\mathrm{X}=n}\}_{n\in\mathcal{N}}\!\cup\!\{W_{\mathrm{Y}|% \mathrm{X}=i}\}_{i\in\mathcal{I}}\right)= italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_i end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_i ∈ caligraphic_I end_POSTSUBSCRIPT ) (3)
C({WY|X=r}r{WY|X=n}n𝒩)𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsubscript𝑊conditionalYX𝑛𝑛𝒩\displaystyle-C\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\cup\{W_% {\mathrm{Y}|\mathrm{X}=n}\}_{n\in\mathcal{N}}\right)- italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT ) (4)
+C({WY|X=r}r{WY|X=n}n𝒩)𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsubscript𝑊conditionalYX𝑛𝑛𝒩\displaystyle+C\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\cup\{W_% {\mathrm{Y}|\mathrm{X}=n}\}_{n\in\mathcal{N}}\right)+ italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT )
C({WY|X=r}r{WY,n}n𝒩)𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsuperscriptsubscript𝑊𝑌𝑛𝑛𝒩\displaystyle-C(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\cup\{W_{Y,n}% ^{\star}\}_{n\in\mathcal{N}})- italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT italic_Y , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT )
+C({WY|X=r}r{WY,n}n𝒩)C({WY|X=r}r)𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsuperscriptsubscript𝑊𝑌𝑛𝑛𝒩𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟\displaystyle+\!C(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\!\cup\!\{W% _{Y,n}^{\star}\}_{n\in\mathcal{N}})\!-\!C\left(\!\{W_{\mathrm{Y}|\mathrm{X}=r}% \}_{r\in\mathcal{R}}\!\right)+ italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT italic_Y , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT ) - italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ) (5)
C({WY|X=r}r{WY|X=n}n𝒩)absent𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsubscript𝑊conditionalYX𝑛𝑛𝒩\displaystyle\leq C\left(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\cup% \{W_{\mathrm{Y}|\mathrm{X}=n}\}_{n\in\mathcal{N}}\right)≤ italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_n end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT )
C({WY|X=r}r{WY,n}n𝒩),𝐶subscriptsubscript𝑊conditionalYX𝑟𝑟subscriptsuperscriptsubscript𝑊𝑌𝑛𝑛𝒩\displaystyle-C(\{W_{\mathrm{Y}|\mathrm{X}=r}\}_{r\in\mathcal{R}}\cup\{W_{Y,n}% ^{\star}\}_{n\in\mathcal{N}}),- italic_C ( { italic_W start_POSTSUBSCRIPT roman_Y | roman_X = italic_r end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_r ∈ caligraphic_R end_POSTSUBSCRIPT ∪ { italic_W start_POSTSUBSCRIPT italic_Y , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT } start_POSTSUBSCRIPT italic_n ∈ caligraphic_N end_POSTSUBSCRIPT ) ,

where (3)+(4) and (5) are 00 by the application of Proposition 1. This proves the equivalence of comparing the distance of removed symbols to either the nearest neighbor, or the closest distribution in the convex hull. ∎