A new information theory swaps bit fidelity for action effectiveness: the authors define 'pragmatic' entropy, mutual information and a behavioural capacity that bounds how much net utility an agent can extract under resource constraints, unifying communication, control and decision-making in a single mathematical framework.

Main Finding

The paper proposes a formal mathematical theory of pragmatic information that extends and unifies classical (syntactic) information theory, semantic information theory, control theory, and decision theory. Its core is the isoteleia mapping—formalizing equifinality (different semantic representations that lead to the same optimal action)—which yields a three-tier hierarchy (syntactic → semantic → pragmatic). The authors define pragmatic counterparts to entropy, mutual information, capacity, and rate-distortion, prove pragmatic versions of Shannon’s coding theorems, and introduce decision-theoretic duals: pragmatic value of information (VoI) and pragmatic cost of information (CoI). A Lagrangian dual framework yields a pragmatic efficiency bound Ep(λ) = sup_R [Φp(R) − λ CoIp(R)], interpreted as a behavioral capacity: a fundamental limit on the net utility that a resource-constrained intelligent agent can extract from its environment.

Key Points

• Conceptual core
  • Isoteleia mapping: groups semantic classes that produce the same optimal terminal action (equifinality).
  • Reification mapping: composition that connects pragmatic intents to syntactic signals.
  • Three-tier hierarchy: syntactic (symbols/fidelity) ≤ semantic (meaning) ≤ pragmatic (action effectiveness); pragmatic abstractions discard distinctions irrelevant to decision outcomes.
• Pragmatic information measures
  • Pragmatic entropy Hp(W).
  • Up/down pragmatic mutual information Ip(X;Y).
  • Pragmatic channel capacity Cp and pragmatic rate-distortion Rp(D).
  • Hierarchical inequalities showing potential performance gains: Hp ≤ Hs ≤ H, Cp ≥ Cs ≥ C, Rp(D) ≤ Rs(D) ≤ R(D).
• Coding theorems
  • Pragmatic versions of three fundamental Shannon theorems: lossless source coding, channel coding (pragmatic capacity), and rate-distortion (pragmatic rate-distortion).
• Decision-theoretic duals and Lagrangian framework
  • Pragmatic Value of Information (VoI): utility gain from task-relevant information; dual to rate-distortion.
  • Pragmatic Cost of Information (CoI): minimal resource expenditure to achieve a given pragmatic rate; dual to channel capacity.
  • Global pragmatic Lagrangian Lglobal(R; λ) = Φp(R) − λ CoIp(R); optimality where marginal value equals marginal cost. λ serves as a shadow price converting physical resources (power, bandwidth, latency) into utility units.
  • Pragmatic efficiency functional Ep(λ) = sup_R [Φp(R) − λ CoIp(R)] — a behavioral capacity for goal-directed systems.
• Extensions and dynamics
  • Continuous-domain formulation via isoteleic volumes; closed-form expressions for Gaussian sources/channels.
  • Dynamic (sequential) settings addressed via a Bellman-equation–style formulation linking to reinforcement learning / dynamic programming.
• Applications emphasized
  • Task-oriented communication, semantic/pragmatic communication systems, networked control, embodied AI, LLMs/AI agents, autonomous driving, robotics, human learning.

Data & Methods

• Nature of work: theoretical and mathematical; no empirical datasets.
• Methods and formal tools used:
  • Formal mappings (isoteleia, synonymous, reification) to define equivalence classes at semantic and pragmatic layers.
  • Information-theoretic derivations to define pragmatic entropy, mutual information, capacity, rate-distortion.
  • Proofs extending Shannon coding theorems to pragmatic settings (lossless source coding, channel coding, rate-distortion).
  • Lagrangian duality to connect value (VoI) and cost (CoI) and derive the pragmatic efficiency bound and shadow price λ.
  • Measure-theoretic extension to continuous domains via isoteleic volumes; closed-form Gaussian results obtained.
  • Dynamic programming / Bellman-equation approach to model sequential decision-making and time-varying resource constraints.
• Validation: formal proofs and derived closed-form special cases (Gaussian). The paper frames implications and examples (LLMs, robotics) but does not present empirical implementation results.

Implications for AI Economics

• Pricing and valuation of information
  • VoI gives a principled, task-centered metric to value outputs of AI systems (e.g., LLM responses) in terms of expected utility improvement rather than accuracy or likelihood alone.
  • CoI provides a framework to quantify the resource cost (bandwidth, latency, energy, compute) required to deliver a given level of pragmatic information; this supports cost-based pricing of API calls, model access tiers, or communication services.
  • The shadow price λ is a natural economic price signal mapping physical resource consumption to utility units; it can guide dynamic pricing, congestion fees, and market-clearing mechanisms.
• Platform design and monetization
  • Platforms can use Ep(λ) to evaluate the maximum net utility extractable per resource budget—informing subscription tiers, per-call pricing, or bundled service design that optimizes user utility per dollar.
  • Auction or allocation mechanisms for scarce communication/computation resources (e.g., edge compute, V2X bandwidth) should be designed to allocate capacity where marginal pragmatic value exceeds the marginal pragmatic cost (VoI ≥ λ·CoI).
• Investment and cost-benefit analysis
  • Behavioral capacity Ep(λ) provides a normative upper bound to assess ROI on investments in infrastructure (bandwidth, specialized hardware) or model improvements: does increasing resources increase net pragmatic utility?
  • The framework helps quantify trade-offs between investing in better sensors/models (increasing VoI) versus investing in communication/computation infrastructure (reducing CoI).
• Mechanism and policy design
  • Regulators and standards bodies can adopt pragmatic-efficiency metrics (Ep(λ), VoI/CoI) to evaluate claims of “intelligent” or “task-oriented” services, ensuring incentives align toward real-world utility rather than signal fidelity.
  • Policies for transparency could require providers to report pragmatic-cost and value estimates for services that affect safety-critical decisions (e.g., autonomous vehicle stacks, medical decision support).
• Market modeling and welfare analysis
  • The VoI/CoI duality enables microeconomic models of information markets where buyers value pragmatic outcomes; welfare-maximizing allocations and price schedules can be derived by maximizing aggregate Ep(λ).
  • Platform competition: providers that deliver higher pragmatic efficiency (higher Φp(R) per CoIp(R)) can dominate markets even if their syntactic fidelity or throughput is lower.
• Operational research and empirical work to enable economics applications
  • Estimation: economists and data scientists can empirically estimate VoI by randomized experiments measuring outcome improvements from model outputs; estimate CoI by measuring marginal resource costs (energy, latency, bandwidth, compute).
  • Dynamic pricing: implement λ as a real-time scarcity signal (e.g., surge pricing for compute or bandwidth) to align supply with marginal pragmatic value.
  • Mechanism evaluation: test auction designs that allocate limited communication/computation to tasks maximizing net pragmatic value.
• Research directions
  • Empirical validation: apply the framework to real AI services (LLMs, perception-to-action pipelines) to measure pragmatic entropy, VoI, CoI, and Ep(λ).
  • Market design: design and test markets and contracts that buy/sell pragmatic information (e.g., APIs sold by VoI or priced by CoI-adjusted rates).
  • Welfare and distributional effects: analyze how pragmatic pricing affects different user groups (who may have different utilities and resource budgets), and design regulation to avoid adverse outcomes.

Overall, the paper supplies a rigorous theoretical foundation that can be operationalized in AI economics to price, allocate, and regulate information-intensive services using task-oriented, utility-based metrics instead of throughput or fidelity alone.