Learning Nonlinear Encoders for the Causal-State Masking Trade-off

A transmitter often has to deal with interference it can only see as it happens. Examples are a radio that measures a jammer's signal sample by sample, or a sensor node that observes the environment it is reporting on. Information theory calls this causal state information: the transmitter sees the current value of a random "state" that corrupts the channel, but never future values. Such a transmitter usually wants several things that pull against each other. It wants to send its own message quickly. It wants to let the intended receiver learn the state, which information theorists call state amplification. And it wants to keep an eavesdropper from learning that same state, called masking.

The only characterization of the amplification–masking trade-off assumes the transmitter knows the whole state sequence in advance. That is the non-causal case, and State Amplification Subject to Masking Constraints explicitly leaves the causal case open. This proposal computes, as numerical inner bounds, the three-way trade-off surface between message rate, amplification and leakage for the causal Gaussian channel. It then asks a sharp question: are the linear encoders that information theory already recommends for causal interference good enough on this surface, or do nonlinear encoders found by gradient-based multi-objective optimization reach points the linear ones cannot?

Core Claim

For the single-objective Gaussian "dirty tape" channel, Achievable Rate Regions for Dirty Tape Channels and "Joint Writing on Dirty Paper and Dirty Tape" shows three things:

  • Linear encoders of the form $X = U - \beta S$ are optimal among jointly Gaussian schemes. Here $X$ is the channel input, $S$ the current state, and $U$ an auxiliary codeword symbol carrying the message.
  • The dirty tape channel is the channel $Y = X + S + Z$, with Gaussian noise $Z$ and the state known causally to the transmitter.
  • The best known lower bound, $C_4(P)$, comes from time-sharing between linear compensation and a nonlinear inflated-lattice (modulo) scheme.

We conjecture that masking changes this picture. With a linear encoder, the eavesdropper sees $U + (1-\beta)S + Z_e$. Hiding $S$ therefore requires $\beta \to 1$, which spends power $\beta^2 Q$ just cancelling the state, where $Q$ is the state variance. A modulo-type mapping can instead make the eavesdropper's output nearly independent of $S$ at much lower power, the way a dither hides a signal. It can also leave the intended receiver, whose channel is stronger, able to recover state information.

The claim is therefore this. Over part of the causal message–amplification–leakage region, nonlinear single-letter strategies strictly dominate the convex hull of linear, compensation and lattice strategies. A neural encoder trained by primal-dual multi-objective optimization will find those strategies. If the claim is false, the same machinery yields a clean negative result: known strategy families suffice even under masking. That result is also new, because no causal masking region has been computed.

Research Program

The inner bound. The first step makes the objective computable exactly, so that no learned mutual-information estimator is involved. We use Shannon strategies. The transmitter picks a codeword $U^n$ from its message, independent of the state, and sends $X_i = f(U_i, S_i)$ for a deterministic map $f$ subject to the power constraint $\mathbb{E}[X^2] \le P$. Standard arguments give three quantities:

  • The message rate $R_m \le I(U;Y_b)$, where $Y_b$ is the intended receiver Bob's output.
  • An amplification rate $R_a \ge I(S;Y_b \mid U)$, which Bob obtains after decoding $U^n$.
  • An upper bound on the eavesdropper Eve's leakage.

The leakage bound follows because $U^n$ is independent of $S^n$:

$$ \tfrac{1}{n} I(S^n; Y_e^n) \;\le\; \tfrac{1}{n} I(S^n; Y_e^n, U^n) \;=\; I(S; Y_e \mid U), $$

where $Y_e$ is Eve's output. This bound is conservative, since Eve may not be able to decode $U^n$. The first research task is to verify this inner bound carefully and look for a tighter leakage term that exploits Eve's inability to decode.

The outer reference is the non-causal region of State Amplification Subject to Masking Constraints. It must contain the causal region, and its own inner and outer bounds are reported to lie within 0.5 bits of each other.

Exact, differentiable evaluation. We give $U$ a finite alphabet of size $K$ and quantize the Gaussian state finely. All three quantities then become one-dimensional integrals over Gaussian mixtures. These can be computed to high numerical precision and differentiated with respect to the encoder map $f_\theta$.

Trade-off geometry. A time-sharing variable averages all three quantities, so the achievable region is convex. The time-sharing variable is known to all parties, so leakage conditioned on it also averages. This convexity is the information-theoretic twin of the result in Constrained Reinforcement Learning Has Zero Duality Gap. That paper shows constrained policy optimization has zero duality gap because randomized policies make the feasible set convex. Here, time-sharing plays the role of randomization. Weighted sums with Lagrange multipliers on the leakage and power constraints are therefore enough to trace the whole region. The non-convex-front arguments for hypervolume scalarization in Navigating Conflicting Objectives with Pareto Optimization do not apply to the region itself.

The remaining difficulty. The same caveat raised in that constrained-RL paper applies: the zero-gap guarantee assumes a global maximization of the Lagrangian at each step. Our parameterization of $f_\theta$ is non-convex, so we will measure the empirical duality gap directly with random restarts. We reuse the gradient ascent–descent structure of Batch Policy Learning under Constraints, with exponentiated-gradient updates on the multipliers, but in an exact rather than a sampled setting.

Encoder parameterization. We train one encoder network conditioned on the preference weights and the leakage budget, in the spirit of the single universal network of PD-MORL and the preference-adaptive policies of Pareto Policy Adaptation. Its only purpose is to amortize compute across the surface. It is not a claimed contribution, and the baselines below test whether it costs anything.

We will also inspect the learned maps $f_\theta(u,\cdot)$ to see whether they are periodic (lattice-like), piecewise linear, or something else. Where they are simple, we will try to write them in closed form and verify them analytically.

Baselines

Structured-family baseline. This is the strongest competitor. It is the full family of known causal strategies, evaluated with exactly the same inner-bound expressions and numerical integration as the learned encoder:

  • linear strategies $X = U - \beta S$, swept over $\beta$ and the power split;
  • Willems-style compensation;
  • Erez-style inflated lattice strategies, swept over the inflation factor and lattice scale;
  • all time-sharing combinations, which means taking the convex hull.

This is the $C_4(P)$ construction of Khosravi-Farsani et al., extended to three objectives. It is the natural thing an information theorist would do first, and it already includes one nonlinear family. On masking, no reported result found on Lacuna for the causal region, so we compute this baseline ourselves. On pure rate, the file reports no numerical value for $C_4(P)$, so we reproduce that curve as a calibration target.

Naive reference. The do-nothing reference treats the state as noise, with $f(U,S) = U$, and applies no masking effort. It anchors how much any state-aware strategy buys.

Non-causal outer reference. The non-causal region from the masking paper is not implementable causally. It bounds how much room remains.

Fairness. Every method gets the same quantization grid, the same alphabet size $K$, the same power and state variances, and the same integration precision. The parametric families are swept on a grid whose number of evaluations matches the number of gradient steps given to the network, multiplied across restarts.

Isolating ablations. Two ablations test which ingredient matters:

  • A linear-restricted network has the same optimizer and multipliers but $f_\theta$ is forced to be affine in $S$. It isolates nonlinearity from the optimizer.
  • Independent per-weight optimization fits one encoder separately for each preference vector. It isolates preference conditioning.

Win criterion. Beating the structured baseline means producing an operating point outside its convex hull. Concretely, that is a strictly larger message rate at the same leakage and amplification, with the margin exceeding the numerical integration error estimated by doubling the grid resolution. The point must also persist across at least half of random restarts.

Why not just…?

Why not just sweep linear and lattice strategies and time-share? This is the real threat, and it is exactly our strongest baseline. Khosravi-Farsani et al. already show that time-sharing linear and lattice schemes beats pure linear on rate. The masking objective may simply be served well by the lattice family too. The decisive comparison below settles this. If the learned encoder never exits the hull, we report that known families suffice, which is itself the first computed causal masking region.

Why not just use a single fixed weighted sum, or hypervolume scalarization? Because the region is convex under time-sharing, weighted sums with multipliers reach every boundary point. The concern in Navigating Conflicting Objectives with Pareto Optimization about concave fronts does not arise here. What can fail is optimization of a non-convex encoder, and we measure that directly as a duality gap. A single fixed weight gives one point, not the surface engineers need to choose an operating point.

Why not just train an independent encoder per preference? For a three-objective convex surface this may work as well as a conditioned network. The parameter-count argument from PD-MORL (0.34M versus 2.8M parameters on Walker) matters less at this scale. We run it as an ablation. If it matches the conditioned network, we say so and treat conditioning purely as a convenience.

Why not just assume the state is known ahead of time and use dirty paper coding? A causal transmitter cannot implement it. State-Dependent Relay Channel with Private Messages finds the causal region strictly inside the non-causal one in its setting, so the non-causal scheme only serves as our outer reference.

Proposed Experiments

Calibration. On the single-objective Gaussian dirty tape channel, we check that our numerical pipeline reproduces the known ordering: linear below the $C_4(P)$ time-sharing curve, and both below the interference-free bound. The unconstrained learned encoder must match or exceed $C_4(P)$ across signal-to-noise ratios and state powers. Failure here would indicate optimizer problems, not an information-theoretic finding.

Main experiment. We trace the causal message–amplification–leakage surface on degraded Gaussian broadcast settings in which Bob's noise is smaller than Eve's, following the regimes of the masking paper. We sweep power, state variance and the noise ratio. For every method we report three metrics:

  • the hypervolume of the achieved region, relative to a common reference point;
  • the maximal message rate at fixed leakage budgets;
  • the gap to the non-causal outer region.

Decisive checks. The first check asks whether any learned operating point lies outside the convex hull of the structured baseline by more than the integration error. It also asks whether the gain concentrates, as predicted, at low leakage budgets and moderate power, where linear masking is expensive.

The second check reports the empirical duality gap across restarts. This connects the local-versus-global concern from Constrained Reinforcement Learning Has Zero Duality Gap to a setting where the true convex region is well defined.

The third check examines any dominating strategies. We extract them, fit them with closed-form maps, and re-evaluate them analytically, so that the headline claim rests on a checkable strategy rather than on a network.

Discrete check. A binary state and channel version repeats the comparison with exact finite sums. This rules out quantization artifacts.