Simultaneous Info & Control Signaling (SICS)

• Simultaneous Information and Control Signaling (SICS) is a framework where one control signal performs dual roles of system regulation and information transmission within stochastic and wireless environments.
• It employs advanced methods from stochastic control and information theory, optimizing trade-offs between directed information and control cost through belief-state dynamics and meta dynamic programming.
• RIS implementations of SICS leverage techniques like SWIPT and NOMA to balance energy harvesting and data transmission, effectively integrating control and communication in real time.

Searching arXiv for the cited SICS-related papers and exact identifiers. Simultaneous Information and Control Signaling (SICS) denotes a class of coupled communication–control problems in which a single action sequence or RF waveform serves a dual operational role: it must regulate, actuate, or reconfigure a physical system while also conveying information required by another agent or subsystem. In partially observable stochastic control, actions $A_t$ simultaneously shape the hidden state evolution and embed information into observations $Y_t$; in reconfigurable intelligent surface (RIS) systems, the same wireless signal can carry control commands together with either harvested power for the control circuitry or user-plane data for a terminal while the RIS microcontroller decodes its command (Charalambous et al., 16 Jun 2026, Charalambous et al., 2024, Kisseleff et al., 2021, Koutsonas et al., 6 Aug 2025).

1. Scope and canonical model classes

In the control-theoretic formulation, SICS is posed on a finite-horizon partially observable Markov decision process (POMDP) or, more generally, a nonlinear partially observable stochastic system. The state $X_t$ is hidden, the controller applies action $A_t$, and the system produces observation $Y_t$. The controller does not observe $X_t$ directly; it only has access to past actions and outputs. The dynamics are described by a controlled state-transition kernel $S_{t+1}(dx_{t+1}\mid x_t,a_t)$ and an observation kernel $Q_t(dy_t\mid x_t,a_t)$. The defining SICS step is to reinterpret this controlled stochastic system as a channel with feedback: a message $M^{(n)}$ must be conveyed to a decoder observing $Y^n$, while the controller-encoder chooses causal actions according to

$Y_t$0

so that the same control input affects both the plant evolution and the observation law (Charalambous et al., 16 Jun 2026).

In the nonlinear partially observable stochastic-system formulation, the same idea appears in state-space form: $Y_t$1 with a cost constraint

$Y_t$2

Here SICS means that the control sequence $Y_t$3 is used simultaneously for plant regulation and for transmission of an independent message to a decoder that observes $Y_t$4. The capacity notion is therefore operational, in the Shannon sense, but the channel is induced by the controlled stochastic dynamics rather than being exogenous (Charalambous et al., 2024).

The RIS literature instantiates the same principle in wireless control planes. One line considers a base station (BS) sending a dedicated RIS control waveform that simultaneously conveys the configuration bits for each reflecting element and provides harvested energy for the RIS control circuitry through simultaneous wireless information and power transfer (SWIPT). Another line considers in-band SICS, where the BS superimposes user data and RIS-control information by non-orthogonal multiple access (NOMA), while a simultaneous transmission and reflection (STAR) RIS directs one component toward the user equipment (UE) and the other toward the RIS microcontroller (MC) (Kisseleff et al., 2021, Koutsonas et al., 6 Aug 2025).

2. Directed information and control-coding capacity

The central information-theoretic object in SICS is directed information from actions to observations,

$Y_t$5

which is the natural causal metric for channels with feedback. In the finite-horizon POMDP formulation, randomized control laws are allowed: $Y_t$6 subject to the causal Markov condition

$Y_t$7

and an average-cost constraint

$Y_t$8

The finite-horizon optimization is

$Y_t$9

and the control-coding capacity is

$X_t$0

obtained through converse and achievability arguments using Fano’s inequality, random coding, and a stability condition (Charalambous et al., 16 Jun 2026).

The nonlinear stochastic-systems treatment gives the same capacity structure under Dobrushin information stability: $X_t$1 This formulation makes the SICS trade-off explicit. Without the cost constraint, the problem degenerates into pure communication over a controlled channel; with the constraint, higher directed information generally requires more aggressive or more variable actions, while lower control cost pushes the policy toward conservative regulation. The same work defines the dual cost–rate problem

$X_t$2

and shows that

$X_t$3

Hence, if $X_t$4, it is necessary that $X_t$5. In this sense, nonzero signalling rate requires control effort beyond the minimum needed for regulation alone (Charalambous et al., 2024).

3. Meta dynamic programming and sufficient statistics in POMDPs

A distinctive feature of the POMDP formulation is that standard belief-state dynamic programming is not sufficient when the objective includes directed information. The first information state is the usual posterior distribution of the hidden state,

$X_t$6

which evolves by a policy-independent nonlinear filtering recursion

$X_t$7

Given $X_t$8, the conditional law of $X_t$9 is

$A_t$0

Thus $A_t$1 is sufficient for the control side and for the instantaneous channel law from action to observation (Charalambous et al., 16 Jun 2026).

Directed information also depends on the one-step marginal prediction $A_t$2, which is not determined by $A_t$3 alone. This motivates the second information state,

$A_t$4

a probability measure on the belief space itself. It satisfies a Markov recursion under a lifted kernel $A_t$5, and the marginal prediction of $A_t$6 can be written purely in terms of $A_t$7 and the randomized policy. The pair $A_t$8 is therefore the minimal structural augmentation required by SICS: $A_t$9 tracks plant uncertainty, while $Y_t$0 tracks the distribution of beliefs that governs the denominator in the directed-information density (Charalambous et al., 16 Jun 2026).

This leads to a “meta” dynamic program on the space of measures over beliefs. After Lagrangian relaxation with multiplier $Y_t$1, the value process is defined by

$Y_t$2

and the Bellman equations optimize a per-stage term comprising immediate directed-information gain, expected control cost, and the future meta-value. Necessary and sufficient optimality conditions follow from these equations, and the resulting optimal policy can be chosen in separated randomized form depending only on the information states $Y_t$3 rather than the full raw history (Charalambous et al., 16 Jun 2026).

4. Information-state realizations in nonlinear and LQG systems

In the nonlinear partially observable stochastic setting, the posterior distribution itself is the fundamental information state. The controller-encoder can be realized through exogenous randomization variables $Y_t$4 or arbitrary independent $Y_t$5, so that

$Y_t$6

and optimal strategies can be taken in separated form,

$Y_t$7

with $Y_t$8 the nonlinear-filter posterior. For information-lossless mappings, the directed-information objective can be written equivalently in terms of the randomization variables: $Y_t$9 and analogously for $X_t$0. This identifies the signalling content of SICS with the fresh randomness causally injected into the control signal through the information state (Charalambous et al., 2024).

The linear-quadratic Gaussian partially observable stochastic system (LQG-POSS) yields a finite-dimensional specialization. The posterior of $X_t$1 is Gaussian with mean $X_t$2 and covariance $X_t$3, where $X_t$4 obeys a Kalman filter and $X_t$5 evolves through a covariance Riccati equation. Optimal signalling strategies can be taken Gaussian: $X_t$6 and then further reduced to affine policies of the form

$X_t$7

A second Kalman filter appears through

$X_t$8

the decoder’s estimate of the encoder’s estimate. The sufficient statistics are therefore $X_t$9, $S_{t+1}(dx_{t+1}\mid x_t,a_t)$0, the signalling covariance $S_{t+1}(dx_{t+1}\mid x_t,a_t)$1, and the Riccati sequences $S_{t+1}(dx_{t+1}\mid x_t,a_t)$2 and $S_{t+1}(dx_{t+1}\mid x_t,a_t)$3. Directed information admits the innovations-covariance form

$S_{t+1}(dx_{t+1}\mid x_t,a_t)$4

and a decentralized separation principle holds: for fixed signalling parameters $S_{t+1}(dx_{t+1}\mid x_t,a_t)$5, the optimal control component is linear,

$S_{t+1}(dx_{t+1}\mid x_t,a_t)$6

with $S_{t+1}(dx_{t+1}\mid x_t,a_t)$7 generated by a control matrix difference Riccati equation. Estimation, control, and signalling are therefore structurally separated, even though they remain coupled through shared dynamics and the common cost budget (Charalambous et al., 2024).

5. Zero-signalling limit and recovery of classical stochastic control

One of the sharpest structural results is that the SICS formulation reduces to ordinary POMDP control when signalling is absent. If the policy is restricted to deterministic separated controls

$S_{t+1}(dx_{t+1}\mid x_t,a_t)$8

then the directed information vanishes: $S_{t+1}(dx_{t+1}\mid x_t,a_t)$9 In this case $Q_t(dy_t\mid x_t,a_t)$0 almost surely, because the belief is itself measurable with respect to the observation history, and the high-level meta-state collapses to the standard belief state. The classical POMDP Bellman recursion is recovered: $Q_t(dy_t\mid x_t,a_t)$1

$Q_t(dy_t\mid x_t,a_t)$2

The minimum cost at rate $Q_t(dy_t\mid x_t,a_t)$3 is exactly the classical deterministic POMDP optimal cost (Charalambous et al., 16 Jun 2026).

The nonlinear stochastic-systems treatment states the same relationship in cost–rate form. If the target information rate is zero, then

$Q_t(dy_t\mid x_t,a_t)$4

and randomized and deterministic stochastic-control optima coincide: $Q_t(dy_t\mid x_t,a_t)$5 This establishes classical stochastic control as the zero-signalling face of the SICS problem. A plausible implication is that SICS should be regarded not as a separate discipline from stochastic control, but as a strict extension in which the controller is permitted, and sometimes required, to spend control effort on communication (Charalambous et al., 2024).

6. RIS implementations: control with power transfer and in-band control with user data

A concrete wireless implementation of SICS appears in RIS control via SWIPT. The BS transmits a dedicated control waveform $Q_t(dy_t\mid x_t,a_t)$6, where $Q_t(dy_t\mid x_t,a_t)$7 is a beamforming vector and $Q_t(dy_t\mid x_t,a_t)$8 is a control packet composed of $Q_t(dy_t\mid x_t,a_t)$9 sub-packets, each updating one RIS reflecting element. After equal-gain combining at the RIS, the received useful-signal power is

$M^{(n)}$0

with $M^{(n)}$1 chosen as maximum-ratio transmission. The SWIPT receiver then partitions this power between information detection (ID) and energy harvesting (EH). The harvested energy follows a sigmoid-type nonlinear model, each reflecting-element update requires at least $M^{(n)}$2, and each control sub-packet must satisfy $M^{(n)}$3. Four architectures are studied: time sharing (TS), power splitting (PS), dynamic power splitting (DPS), and antenna selection (AS). The corresponding optimization problems maximize the number $M^{(n)}$4 of RIS elements that can be updated after one control sequence, subject to the SNR and energy constraints, and closed-form solutions are derived, including

$M^{(n)}$5

$M^{(n)}$6

and

$M^{(n)}$7

With perfect received-power knowledge, PS, DPS, and AS reach $M^{(n)}$8 updated reflecting elements at around $M^{(n)}$9 dBm transmit power, while TS performs better at very low powers because it avoids splitter noise. With received-power fluctuations and no bias, PS, DPS, and AS can saturate around $Y^n$0 average updated reflecting elements, whereas introducing a negative bias in the design of the splitting parameters can restore nearly $Y^n$1 updates at sufficiently high transmit power. With $Y^n$2 dBm and a 2-std bias, increasing $Y^n$3 degrades all architectures, and PS and AS outperform TS and DPS (Kisseleff et al., 2021).

A second RIS realization uses in-band SICS rather than dedicated control-plane SWIPT. The BS superimposes user data $Y^n$4 and MC control signal $Y^n$5 using NOMA,

$Y^n$6

and a STAR-RIS sends the reflected component to the UE and the transmitted component to the MC. The UE effective SNR is

$Y^n$7

while the MC first decodes the user stream,

$Y^n$8

then decodes the control stream after SIC,

$Y^n$9

The design objective is to maximize the UE rate $Y_t$00 subject to the STAR energy constraint $Y_t$01, the non-degenerate constraints $Y_t$02, $Y_t$03, and the MC thresholds $Y_t$04, $Y_t$05. For fixed RIS coefficients, the optimal power split is

$Y_t$06

so the control signal receives exactly the minimum power required by the MC SNR constraint. For fixed $Y_t$07, the RIS-coefficient problem reduces to a convex optimization in the transmission magnitudes $Y_t$08, solved by interior-point methods; the total complexity is

$Y_t$09

and one alternating-optimization pass suffices to reach the fixed point reported in the paper. Simulations averaged over 100 channel realizations show that $Y_t$10, so less than $Y_t$11 of BS power is used for control in the shown scenarios. Increasing $Y_t$12 or $Y_t$13 improves the UE rate, while increasing $Y_t$14 slightly decreases it; for example, at $Y_t$15 W, raising $Y_t$16 from $Y_t$17 dB to $Y_t$18 dB decreases the rate from about $Y_t$19 to $Y_t$20 bit/s/Hz (Koutsonas et al., 6 Aug 2025).

7. Context, trade-offs, and limitations

Across these formulations, SICS unifies several previously distinct viewpoints. In classical information-constrained control and control over communication channels, the communication link is usually exogenous; in the POMDP and nonlinear-state-space treatments, the plant and observation mechanism themselves constitute the communication medium, so the channel is endogenous to the control policy. In feedback-capacity theory, directed information is the canonical objective for channels with memory and feedback; SICS imports that objective into control design under performance constraints. In dual control, actions are already known to both regulate and probe; SICS makes the second role explicitly information-theoretic by optimizing directed information or, in RIS settings, by jointly enforcing decoding and harvested-energy constraints (Charalambous et al., 16 Jun 2026, Charalambous et al., 2024).

The main trade-off is the allocation of actuation or RF resources between regulation and information transfer. In the POMDP meta-DP, varying the Lagrange multiplier $Y_t$21 moves along the rate–cost frontier: small $Y_t$22 prioritizes directed information, while large $Y_t$23 prioritizes control cost. In the nonlinear capacity formulation, $Y_t$24 is the extra control effort required to sustain rate $Y_t$25. In RIS SWIPT, the resource split is between information detection and harvested energy. In STAR-RIS NOMA, it is between reflected user service and transmitted MC control, as well as between $Y_t$26 and $Y_t$27 in the superposition code. This suggests that SICS is fundamentally a constrained co-design problem rather than a simple overlay of communication on top of control.

The available results also delineate clear limitations. The meta-DP theory is developed for finite horizon, and its exact solution is generally intractable because the state lives on a space of probability measures; the paper explicitly points to approximate dynamic programming or reinforcement learning on the lifted space rather than closed-form solutions. The nonlinear and LQG capacity results rely on information-stability or asymptotic time-invariance assumptions and on the tractability of posterior-state recursions. The RIS SWIPT study assumes a single RIS control link, quasi-static channels during one control sequence, a specific nonlinear EH model, and perfect channel knowledge for beamforming. The in-band STAR-RIS NOMA study assumes one UE and one MC, ideal STAR constraints, perfect CSI, and focuses on the SICS phase rather than a fully joint protocol across control and data phases. None of these limitations negate the formalism; they specify the operating regimes in which current SICS results are rigorous (Charalambous et al., 16 Jun 2026, Charalambous et al., 2024, Kisseleff et al., 2021, Koutsonas et al., 6 Aug 2025).