Causal Influence of Communication

• Causal Influence of Communication measures causal effects across varied systems, distinguishing causation from correlation.
• Extensive methodologies, like KL divergence and directed information, enable precise estimation of communication impact.
• Applications span social media, quantum channels, and agent interactions, enhancing system design and communication strategy.

The Causal Influence of Communication (CIC) refers to the rigorous quantification of how interventions in communication channels, strategies, or behaviors causally affect downstream information processing, agent behavior, or systemic outcomes. CIC frameworks span settings from information theory and multi-agent reinforcement learning to social media engagement, quantum channels, and dyadic human interaction, employing both statistical and physical operationalizations. These formulations unify the formal estimation of influence with causal inference—explicitly distinguishing causation from mere correlation and enabling prescriptive reasoning about the design and impact of communication systems.

1. Formal Definitions and Mathematical Foundations

CIC originates from multiple lines of research, each providing a domain-specific definitional core. In classical causal inference on graphical models, the influence of a communication channel is defined via interventions that sever or corrupt specific edges in a directed acyclic graph, and quantify the change in the joint distribution via the Kullback–Leibler divergence between the original and intervened systems. The prototypical measure is

$C_{S}(P) = D\bigl(P \mid\mid P_{S}\bigr),$

where $S$ is a set of communication edges, $P$ is the original joint law, and $P_{S}$ is the intervention distribution with $S$ cut and replaced by independent sampling (Janzing et al., 2012).

In information theory, the continuous-time directed information $I(X_0^T \to Y_0^T)$ between two stochastic processes $X_t$ and $Y_t$ characterizes the net flow of information when feedback or causally adaptive encoding is allowed: $$I(X_0^T\to Y_0^T) = \inf_{t\in T(0,T)}\, \sum_{i=1}^n I\bigl(X_0^{t_i};\,Y_{t_{i-1}}^{t_i}\mid Y_0^{t_{i-1}}\bigr)$$ over time partitions $t_0< \cdots < t_n$ (Weissman et al., 2011). This formulation is fundamental for channels with feedback and in the estimation of feedback capacity.

In quantum information, CIC is treated via functional measures of communication-enabled correlation change, notably the signalling power (how much one agent's input can alter another's output state) and the causal-influence power (how much one agent's operation can modify joint correlations), both formalized using diamond-norm distances over quantum channels (Barsse et al., 2023, Barsse et al., 20 May 2025). In resource theoretic terms, the signalling robustness $S$0 quantifies the minimal mixing needed to render a process causally disconnected (Milz et al., 2021).

In agent-based systems, especially multi-agent reinforcement learning, CIC is concretized as the counterfactual effect of one agent’s action on another’s behavior, operationalized via immediate KL divergences between action distributions with and without a given action, and in expectation, as mutual information (Jaques et al., 2018).

2. Identification Strategies and Estimands

Rigorous estimation of CIC requires causal identification conditions. In agent-based social systems and conversations, this involves defining potential-outcome models for treatments (e.g., communication tendencies) and outcomes (e.g., ratings, engagement) and formulating relevant average treatment effects (ATE): $S$1 for two communication policies or agent behaviors (Zhang et al., 2020, Tian et al., 25 May 2025).

In networked settings, such as social media, the network causal inference framework extends the Rubin potential-outcomes theory to interference structures represented by exposure matrices (e.g., narrative propagation on Twitter). The causal influence of making account $S$2 a source is defined as

$S$3

where $S$4 differ only in the sourcing status of $S$5 (Smith et al., 2018). Model-based identification requires (i) correct specification of exposure as a function of the communication graph, (ii) adjustment for confounding via covariates, and (iii) assumptions of partial interference.

3. Methodological Approaches and Practical Implementation

A broad toolkit applies to the estimation of CIC depending on the system:

• DAG-based systems: Channel-cut interventions generate a “corrupted” distribution; KL divergence quantifies edge influence (Janzing et al., 2012).
• Continuous-time channels: Directed information is computed via infimum/supremum over time/space partitions, with feedback effects handled through causal estimation identities (generalizations of Duncan’s theorem, Poisson analogues) (Weissman et al., 2011).
• Networked social media: Causal estimands are computed from network potential outcome models using Poisson GLM fits, with exposures and covariates explicitly modeled. Bayesian inference (e.g., Gibbs-MCMC) is used for posterior estimation, with model selection guided by Fisher information geometry (Cramér–Rao bounds) (Smith et al., 2018).
• Multi-agent systems: Influence rewards are computed via counterfactual reasoning, using deep models of other agents to empirically estimate mutual information and train policy networks in reinforcement learning settings (Jaques et al., 2018).
• Quantum channels: Quantification employs the diamond norm between original and decoupled (no-influence) channels, with resource-theoretic monotones (e.g., $S$6) computed using semidefinite programming (Milz et al., 2021).
• Dyadic human communication: Granger causality is applied on selected intervals of synchronized signals (typically facial expressions), with interval selection based on high pairwise correlations and causal direction assigned through regression-based hypothesis tests (Shadaydeh et al., 2020, Müller et al., 2018).

4. Properties, Theoretical Guarantees, and Limitations

Key properties of CIC measures are domain-dependent, but share critical invariances and monotonicities:

• Non-negativity: No intervention or purely noise-additive channels yield zero causal influence (Janzing et al., 2012, Weissman et al., 2011, Milz et al., 2021).
• Additivity/sub-additivity: Influence may be additive (quantum parallelizations), non-superadditive (XOR pathologies), or bounded from above by resource-theoretic maxima (Janzing et al., 2012, Milz et al., 2021).
• Monotonicity: Under non-signalling transformations or local operations, causal influence cannot increase (Milz et al., 2021).
• Operational characterization: In continuous-time, directed information equates exactly to estimation-theoretic error integrals; in quantum channels, the signalling and causal-influence powers may diverge at small scale (CNOT gate gap), but asymptotically converge under parallel uses (Weissman et al., 2011, Barsse et al., 20 May 2025, Barsse et al., 2023).
• Causal identification limitations: All approaches face challenges related to model misspecification, confounding, network interference, finite-sample error, and measurement of counterfactual scenarios in practice (Smith et al., 2018, Tian et al., 25 May 2025, Zhang et al., 2020).

5. Applications and Empirical Findings

CIC frameworks enable principled analysis across diverse systems:

• Social media influence estimation: Formal network-based causal inference on large-scale Twitter graphs exposes high-impact accounts in misinformation campaigns; notably, volume of content is not predictive of network-level causal influence, with Bayesian estimates validated by external investigative findings (Smith et al., 2018).
• Conversational recommendation policies: Causal inference on behavioral tendencies in crisis counseling conversations permits design of allocation policies that modestly but robustly improve outcomes, with controlled estimation strategies breaking assignment and interactional confounding (Zhang et al., 2020).
• Agent coordination via intrinsic motivation: Influence-based intrinsic reward functions enable decentralized, communication-efficient coordination among deep RL agents, often exceeding prosocial and mean-reward baselines (Jaques et al., 2018).
• Quantum communication capacity and resource theory: Quantum process matrices, resource monotones ($S$7), and channel-specific influence-power formalism extend the unification of communication and causality into indefinite-order and multi-agent quantum settings (Milz et al., 2021, Barsse et al., 2023, Barsse et al., 20 May 2025).
• Emotional influence in human interaction: Targeted interval Granger-causality analyses, leveraging computer vision AU extraction, attribute directionality in transient mutual adaptation of affective signals in dyadic dialogs (Shadaydeh et al., 2020, Müller et al., 2018).
• CIC in misinformation engagement: Joint treatment–outcome deep models quantify the causal effect of exogenous signals on engagement, outperforming baseline measures (e.g. exposure, structural centrality), and aligning more closely with expert evaluations of influence (Tian et al., 25 May 2025).

6. Comparative Table: CIC Measures in Representative Domains

Domain	CIC Measure	Operationalization/Estimand
Classical DAGs	KL-div. $S$8	Edge cut, distributional divergence (Janzing et al., 2012)
Continuous-time comm.	Directed information $S$9	Inf/sup partition info flow (Weissman et al., 2011)
Quantum processes	$P$0, $P$1, $P$2	SDP monotones, diamond-norm distances (Milz et al., 2021, Barsse et al., 2023, Barsse et al., 20 May 2025)
Social media networks	$P$3: network causal effect	Network potential outcomes (Smith et al., 2018)
Multi-agent systems	Influence reward, mutual info	KL divergence, MI between agent actions (Jaques et al., 2018)
Dyadic nonverbal	Granger causality on intervals	Interval VAR model and F-test (Shadaydeh et al., 2020)

7. Open Challenges and Theoretical Questions

Open problems in CIC research include:

• Nonlinearity and High-dimensionality: Existing measures are often formally optimal in low-dimensional, linear, or discretized settings; generalizing to complex and nonlinear dynamical systems remains an open area.
• Causal inference under interference: Social and communication networks frequently violate the partial interference assumption.
• Temporal granularity and feedback: Continuous-time and multi-scale feedback loops complicate identifiability and operational meaning of information flows (Weissman et al., 2011).
• Quantum indefinite order: No unique maximally influential process exists in quantum-indefinite causal structures; resource-interconvertibility is, in general, frustrated (Milz et al., 2021).
• Validity of counterfactuals: Robustness of causal-effect estimators under un observed confounders and the design of plausible interventions remain practical bottlenecks, especially on social media and agent systems (Tian et al., 25 May 2025).

In summary, CIC frameworks provide a mathematically rigorous, operationally grounded basis for measuring the causal effect of communication mechanisms across classical, social, agent-based, and quantum systems. Their continued development is central for understanding, optimizing, and regulating influence in increasingly complex communicative environments.