Distributed Side-Information Games

• Distributed side-information games are frameworks where multiple agents with asymmetric observations influence coding, estimation, and strategic decision-making.
• The models range from continuous-time Brownian games adjusting signal-to-noise ratios to resilient source coding that leverages graph-theoretic concepts like chromatic numbers.
• Side information fundamentally alters system design by affecting equilibrium strategies, rate-distortion tradeoffs, secrecy, and reliable communication in distributed scenarios.

As the surrounding literature suggests, a distributed side-information game is a class of strategic, coding-theoretic, and estimation problems in which multiple agents, encoders, or users hold asymmetric observations about a common source, state, action profile, or channel, while a receiver, decoder, server, or mediator combines those observations under explicit equilibrium, rate, distortion, secrecy, or reliability criteria. In "A game of information" (Brody, 14 Jun 2026), two players choose signal-to-noise ratios of noisy information streams about a binary state in order to push a Bayesian posterior across a threshold; in "Resilient Source Coding" (Treust et al., 2012), a single encoder communicates a multi-dimensional source to a decoder with side information under orthogonal deviation; in "Strategic Communication with Side Information at the Decoder" (Treust et al., 2019), the encoder controls posterior beliefs induced by a Wyner–Ziv source encoding scheme; and in "Wyner-Ziv Estimators for Distributed Mean Estimation with Side Information and Optimization" (Mayekar et al., 2020), clients send simultaneous messages to a server that alone possesses side information.

1. Defining features and information structure

Across the literature, the defining components are distributed observations, asymmetric information, and a downstream decision rule that is itself part of the strategic environment. The side information may be available only at the destination, as in resilient source coding with decoder-only side information $Y^n$ through a memoryless channel $\daleth: X \to \Delta(Y)$ (Treust et al., 2012); only at the decoder in strategic communication, where $(U,Z)$ are i.i.d. and the decoder acts on the basis of a channel output and $Z$ (Treust et al., 2019); at the server in non-interactive simultaneous-message-passing mean estimation, where each client observes $x_i\in\mathbb{R}^d$ while the server has a guess $y_i$ (Mayekar et al., 2020); or asymmetrically at multiple legitimate receivers and absent at an eavesdropper in the broadcast wiretap model (Treust et al., 2012).

The side information need not be passive. In the side-information “vending machine” framework, the decoder receives messages from distributed encoders and then chooses an action sequence $A^n$, with $Y_i$ drawn from a memoryless channel $p(y\mid a,x_1,x_2)$ and subject to an average cost constraint $\mathbb{E}[\Lambda(A)]\le \Gamma$ (Ahmadi et al., 2011). In the secure coordination model with a two-sided helper, Helen broadcasts a public message $\daleth: X \to \Delta(Y)$0, Alice sends a public message $\daleth: X \to \Delta(Y)$1, Alice and Bob share a secret key $\daleth: X \to \Delta(Y)$2, and Eve sees all public communication together with causal disclosure $\daleth: X \to \Delta(Y)$3 (Satpathy et al., 2014).

What counts as “distributed” therefore varies by model. In some formulations it means several senders or players, each with a local observation; in others it means a single encoder facing a decoder whose side information is private; and in yet others it means many clients whose local messages are fused by a server or many users whose costs depend on an aggregate quantity that must be estimated or exchanged. A plausible implication is that the term is best understood functionally: it names the interaction between decentralized information acquisition and centralized or strategic inference, rather than a single fixed protocol class.

2. Canonical mathematical formulations

A basic continuous-time instantiation appears in the two-player information game of Brody, Hughston, and Yang. The binary state is $\daleth: X \to \Delta(Y)$4 with prior $\daleth: X \to \Delta(Y)$5, $\daleth: X \to \Delta(Y)$6, and the players control the information processes

$\daleth: X \to \Delta(Y)$7

where $\daleth: X \to \Delta(Y)$8 are standard Brownian motions, independent of $\daleth: X \to \Delta(Y)$9, with correlation $(U,Z)$0. The receiver updates rationally via

$(U,Z)$1

and the event $(U,Z)$2 has probability

$(U,Z)$3

The game is zero-sum, with

$(U,Z)$4

and action space $(U,Z)$5 (Brody, 14 Jun 2026).

A discrete, adversarial coding formulation is given by resilient source coding under orthogonal deviation. The source is $(U,Z)$6, one unknown component may vary arbitrarily across the block, and the decoder has side information through $(U,Z)$7. The main single-letter rate is

$(U,Z)$8

where $(U,Z)$9 is the chromatic number of the confusion graph $Z$0. The graph-coloring term captures the cost of identifying the deviating component when the side-information channel can confound symbols, while the conditional-entropy term is the Slepian–Wolf cost of the remaining components (Treust et al., 2012).

A high-dimensional estimation version appears in distributed mean estimation with side information. Client $Z$1 observes $Z$2, the server has $Z$3, and the goal is to estimate

$Z$4

subject to a per-client budget of $Z$5 bits in a simultaneous-message-passing protocol. The mean-squared error decomposes as

$Z$6

which is the core lemma underlying the rotated modulo quantizer, the subsampled Wyner–Ziv quantizer, and the correlated-sampling constructions (Mayekar et al., 2020).

These formulations differ in alphabet, timing, and objective, but they share a common structure: the side information changes the feasible posterior, feasible rate, or feasible estimator, and that change is itself subject to strategic choice, adversarial variation, or communication constraints.

3. Equilibrium, posterior beliefs, and strategic information design

In the Brownian information game, equilibrium behavior is completely characterized by the prior-threshold polarity and the correlation $Z$7. If $Z$8, player $Z$9 wants to minimize the effective information flow $x_i\in\mathbb{R}^d$0 and player $x_i\in\mathbb{R}^d$1 wants to maximize it; if $x_i\in\mathbb{R}^d$2, the roles flip. For $x_i\in\mathbb{R}^d$3, the equilibrium is pure with $x_i\in\mathbb{R}^d$4. For $x_i\in\mathbb{R}^d$5, the equilibrium is again pure with $x_i\in\mathbb{R}^d$6. For $x_i\in\mathbb{R}^d$7, the equilibrium is mixed: $x_i\in\mathbb{R}^d$8 and $x_i\in\mathbb{R}^d$9 randomizes between $y_i$0 and $y_i$1 with probabilities $y_i$2 and $y_i$3, respectively (Brody, 14 Jun 2026).

A more general Stackelberg formulation appears in strategic communication with side information at the decoder. The admissible target laws are

$y_i$4

and the encoder’s single-letter value is

$y_i$5

The encoder chooses $y_i$6 to steer the decoder’s posteriors $y_i$7, the decoder best-responds by minimizing its own distortion, and the Wyner–Ziv feasibility constraint $y_i$8 limits which posterior splittings are implementable. When $y_i$9 is constant and the capacity constraint is inactive, the model reduces to Bayesian persuasion; when $A^n$0, it collapses to standard Wyner–Ziv or Merhav–Shamai separation (Treust et al., 2019).

Quadratic–Gaussian strategic communication sharpens the role of side information. With source $A^n$1, encoder bias $A^n$2, decoder side information $A^n$3, and quadratic costs $A^n$4 and $A^n$5, the strategic Wyner–Ziv problem exhibits no rate loss: the minimum equilibrium rate is the same whether or not the encoder is also informed of $A^n$6. In strategic joint source–channel coding, the uncoded linear mapping $A^n$7 with MMSE decoding is equilibrium-optimal if and only if

$A^n$8

This condition makes the side information orthogonal to the transmitted direction; otherwise digital schemes that exploit binning can outperform uncoded mappings (Akyol et al., 2016).

A recurring consequence is that side information alters not only estimation quality but also the geometry of best responses. The game is therefore not merely one of compression or persuasion alone; it is one of feasible posterior design under decentralized information constraints.

4. Coding, secrecy, reliability, and controllable side information

In multiterminal source coding with a side-information vending machine, encoder messages play a dual role: they describe source data and convey control information for the decoder’s action policy. With non-causal $A^n$9, the achievable distributed region combines a Berger–Tung control layer $Y_i$0 with a distributed Wyner–Ziv refinement layer $Y_i$1; with causal $Y_i$2, a single-stage Berger–Tung-type description is used because binning cannot exploit future side information. Tight characterizations are obtained for degraded source sets with causal $Y_i$3 and for a one-distortion, lossless-$Y_i$4, non-causal scenario where only one encoder controls actions (Ahmadi et al., 2011).

Secrecy variants replace distortion alone by rate-payoff regions under adversarial observation. In the broadcast wiretap channel with asymmetric side information, the achievable region is described by auxiliary variables $Y_i$5 and inequalities such as

$Y_i$6

together with the analogous constraint for $Y_i$7 and a coupled sum-rate bound. The scheme blends Marton coding, Gel’fand–Pinsker precoding, and wiretap binning, and the paper establishes an explicit connection between multiuser channels and observation structures in dynamic games (Treust et al., 2012).

A closely related coordination problem is secure source coding with a two-sided helper. The single-letter rate-payoff region is given by

$Y_i$8

Here Helen’s public broadcast $Y_i$9 is known to Alice, Bob, and Eve; Alice’s public layer $p(y\mid a,x_1,x_2)$0 is also visible to Eve; and the secret-key-protected layer $p(y\mid a,x_1,x_2)$1 carries the portion that must be concealed against causal disclosure. Setting $p(y\mid a,x_1,x_2)$2 recovers the no-helper case (Satpathy et al., 2014).

Reliability theory yields another game-theoretic layer. In source coding with side information, the error-exponent analysis is formulated as a two-player game between nature and the code designer, with nature choosing empirical distributions or covariance perturbations to minimize the exponent and the designer choosing auxiliary test channels to maximize it. For Wyner–Ziv coding, the analysis exposes a tension in the choice of test channel: the optimal test channel must balance a covering error event against a binning or decoding error event (Kelly et al., 2011).

5. Distributed algorithms, optimization, and representative applications

Distributed side-information games are not restricted to abstract coding problems; they also generate concrete iterative algorithms. In distributed mean estimation, the rotated modulo quantizer and its subsampled version yield a low-precision guarantee

$p(y\mid a,x_1,x_2)$3

while the universal, $p(y\mid a,x_1,x_2)$4-unknown construction satisfies

$p(y\mid a,x_1,x_2)$5

The same estimators feed communication-constrained distributed optimization: WZ-SGD attains the optimal convergence rate by exploiting the fact that the gradient estimates are close to each other, and the overall complexity is $p(y\mid a,x_1,x_2)$6 real operations per vector (Mayekar et al., 2020).

In autonomous demand side management, the side information is the aggregate load profile $p(y\mid a,x_1,x_2)$7. Consumer $p(y\mid a,x_1,x_2)$8 minimizes

$p(y\mid a,x_1,x_2)$9

the resulting Nash equilibrium solves a variational inequality, and uniqueness holds if

$\mathbb{E}[\Lambda(A)]\le \Gamma$0

The paper gives a one-timescale distributed iterative proximal-point algorithm when a central unit broadcasts $\mathbb{E}[\Lambda(A)]\le \Gamma$1, and an agreement-based algorithm when consumers only exchange local estimates. In the reported setup, the aggregated load profile becomes flatter after DSM, the peak-to-average ratio is reduced from $\mathbb{E}[\Lambda(A)]\le \Gamma$2 to $\mathbb{E}[\Lambda(A)]\le \Gamma$3, and the synchronous schemes stabilize within $\mathbb{E}[\Lambda(A)]\le \Gamma$4 iterations while the gossip method approaches the same Nash equilibrium within $\mathbb{E}[\Lambda(A)]\le \Gamma$5 iterations (Chen et al., 2013).

Wireless interference channels provide a further distributed-information regime. Users may know all channel gains, only the gains incident on their receiver, or only their own direct-link gain. In all three cases, Nash equilibria are formulated as variational inequalities; complete information yields an affine VI with water-filling best responses, while partial-information regimes yield nonaffine VIs defined by gradients of expected rate functions. The paper also provides lower-bound water-filling policies that require only minimal side information, a distributed augmented-Lagrangian method for Pareto and Nash-bargaining points, and a Bayesian-learning scheme that converges to an $\mathbb{E}[\Lambda(A)]\le \Gamma$6-Nash equilibrium for the direct-link-only game with finite power sets (A et al., 2016).

The mediator interpretation of resilient source coding makes the application to games explicit. When players in a dynamic game receive compressed messages plus exogenous side information, feasible utility vectors satisfy

$\mathbb{E}[\Lambda(A)]\le \Gamma$7

and the framework is applied in the paper to the prisoner’s dilemma and the battle of the sexes (Treust et al., 2012).

6. Conceptual relations, misconceptions, and limitations

One conceptual extension replaces communication protocols by coalition values. In cooperative-game-based information decomposition, the “players” are the nonempty subsets $\mathbb{E}[\Lambda(A)]\le \Gamma$8 of source variables, the coalition value is

$\mathbb{E}[\Lambda(A)]\le \Gamma$9

and the total mutual information is decomposed into non-negative generalized Shapley shares assigned to singleton and higher-order coalitions. The resulting decomposition has analogs of synergy and unique information, but lacks terms corresponding to redundancy; it obeys equivalents of local positivity and identity, and the contribution of a higher-order coalition is zero when the information it adds is already implied by lower-order constraints (Ay et al., 2019). This suggests a broader reading of distributed side-information games in which “distribution” refers not only to agents or terminals, but also to coalitional access to predictor subsets.

A common misconception is that side information is necessarily exogenous and fixed. The literature shows several contrary mechanisms: a decoder can choose an action policy to shape $\daleth: X \to \Delta(Y)$00 under a cost budget (Ahmadi et al., 2011); players can choose information flow rates or SNRs of noisy streams (Brody, 14 Jun 2026); and a helper can publicly broadcast $\daleth: X \to \Delta(Y)$01 while a secret key protects $\daleth: X \to \Delta(Y)$02 (Satpathy et al., 2014). Another misconception is that additional side information benefits all participants. In strategic quadratic–Gaussian communication, receiver side information reduces the decoder’s distortion but can erode the encoder’s influence, and transmitter-only side information is neutral or harmful to the encoder at equilibrium unless it is statistically redundant with $\daleth: X \to \Delta(Y)$03 (Akyol et al., 2016). In the disinformation generalization of the Brownian information game, if $\daleth: X \to \Delta(Y)$04 and $\daleth: X \to \Delta(Y)$05, the bias term becomes negative and disinformation can “backfire” (Brody, 14 Jun 2026).

The main limitations are structural. Several exact results rely on Gaussian or Brownian noise, memorylessness, and low-dimensional latent states; in the Brownian information game, tractability hinges on additive Gaussian noise and a binary state (Brody, 14 Jun 2026). In secure coordination, the converse is proved under $\daleth: X \to \Delta(Y)$06, and the pruning step constrains the payoff to be independent of $\daleth: X \to \Delta(Y)$07 (Satpathy et al., 2014). In error-exponent analysis for Wyner–Ziv coding, the converse is not proven tight in general, and a sharper converse would establish an exponent loss for one-sided Wyner–Ziv (Kelly et al., 2011). In cooperative information decomposition, the number of simplicial complexes and maximal chains grows super-exponentially with the number of sources, so exact computation is practical only for small $\daleth: X \to \Delta(Y)$08 (Ay et al., 2019).

Taken together, these models define a research program rather than a single theorem. The unifying claim is that once side information is distributed, strategic, coded, estimated, or controlled, the central object is no longer merely a rate-distortion function or a Nash equilibrium in isolation. It is the coupled geometry of information access, belief formation, and strategic response.