Contribution
It gives a formal, intrinsic-looking definition of “meaningful” information for any physical system: the part of its correlation with the environment that causally matters for the system’s continued existence. It defines two kinds (stored, from initial mutual information; observed, from transfer entropy), an information/viability curve, a semantic efficiency, semantic content of a state, and for stored information a thermodynamic multiplier comparing the viability benefit with the Landauer cost of acquiring the information.
Key insight
Scramble the system–environment correlation and watch viability. Scrambling by coarse-graining the environment as the system sees it lets one strip correlations away gradually; the smallest amount of correlation that still yields the actual viability is the meaningful part, and what it correlates with is the content.
Assumptions
- A world split into system X and environment Y, fixed over [0, τ], discrete finite state spaces, an initial joint distribution p(x₀, y₀).
- Viability V = −S(p_{X_τ}), plus internal entropies when states are macrostates (App. B). The authors call it exogenous: “the viability function is exogenously determined by the scientist analyzing the system”.
- Interventions by pre-garbling. p̂^φ(x₀ | y₀) = p(x₀ | φ(y₀)) for a coarse-graining φ of Y; for observed information the system’s update rule is coarse-grained the same way at each step, the environment’s left alone.
- First-order Markov, discrete-time dynamics for observed information; time-inhomogeneous allowed.
- Energetics ignored in interventions and in the multiplier (§V A 3), which the authors flag.
Key results
- Stored semantic information (§V A). S_stored = I_{p̂^opt}(X₀; Y₀), where p̂^opt minimizes I over φ subject to V(p̂^φ_{X_τ}) = V(p_{X_τ}). 0 ≤ S_stored ≤ I_p(X₀; Y₀); η_stored = S/I ∈ [0, 1]. The optimal intervention need not be unique (redundant sources).
- Thermodynamic multiplier. κ = ΔV_tot/I_p(X₀; Y₀), with W_min = k_B T ln 2 · I as the cost of measurement; κ = η · ΔV/S.
- Observed semantic information (§V B). The same construction with Σ_t T(Y_t → X_{t+1}) in place of I(X₀; Y₀).
- Example (§V A 4, App. B). Five locations, agent at the centre with exact knowledge of the food’s location, death macrostate of 100 bits’ internal entropy, τ = 5: I = log₂ 5 = 2.32 bits, S_stored ≈ 1.37, η ≈ 0.6, ΔV ≈ 22.1 bits, κ ≈ 9.5; for an agent that flees the food, the same S and ΔV ≈ −13.7, κ ≈ −5.9.
- Agency (§VI). “A physical system is an autonomous agent to the extent that it has a large measure of semantic information”, with decompositions, timescales and initial distributions to be chosen by maximizing such measures.
- Appendix A. p(X ∈ A) ≲ (log|X| − S)/(log|X| − log|A|) for a small viability set A.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Semantic information is bounded by the syntactic information it is drawn from | strong | the actual distribution is in the feasible set |
| C2 | The definition applies to any physical system | moderate | it is well defined for any discrete X, Y and τ once V is fixed |
| C3 | The definition is intrinsic to the system | weak | V, the split, τ and p are chosen by the analyst (§I B, §IV) |
| C4 | Negative entropy is a reasonable viability function | weak | Appendix A’s approximate bound; the authors list its failures |
| C5 | Agency can be identified by maximizing semantic information | weak | proposed, not tried |
| C6 | Semantic information can misrepresent (negative value) and is asymmetric | moderate | the fleeing-agent example; the definition |
Method
For stored information: enumerate coarse-grainings φ of Y, form p̂^φ(x₀, y₀), run the dynamics to τ, record V and I, take the information/viability curve and its optimum. Values are rounded to five decimals to smooth the curve.
Concepts
- Syntactic information: mutual information or transfer entropy.
- Viability (value of information) ΔV_tot: V(actual) − V(fully scrambled).
- Viability-optimal intervention: the most scrambled distribution with unchanged viability.
- Semantic content of x₀: p̂^opt(y₀ | x₀).
- Semantic efficiency η, thermodynamic multiplier κ.
Connections
It formalizes the autonomous-agents tradition (Ruiz-Mirazo and Moreno, Bickhard, Barandiaran et al., autopoiesis) and contrasts itself with teleosemantics (Millikan, Godfrey-Smith) and utility-based value of information. Its interventions are Blackwell-style garblings (Rauh et al. 2017). It cites Krakauer et al.’s individuality (LIT-1137) as related fitness-relevant information, and Touchette and Lloyd for entropy as a control objective.
Bearing on the record
- CLAIM-152. The LIT records A10’s caution that “the system’s continued existence must be defined first, or the test is circular”. The paper agrees with the first half and does not escape the second. Viability is a function the analyst picks, and the boundary of the system is an input; §VI then proposes to choose the boundary, the timescale and the initial distribution by maximizing semantic information, which is measured against that viability. A boundary found that way is relative to the chosen viability function, an instance of CLAIM-152 with viability as the target. The paper also concedes that negative entropy does not track identity (“both a rhinoceros and a human have low entropy”), so “keeping existing” here means staying concentrated, not staying the same system. Proposed dated note below.
- LIT-1149’s own label. The LIT says it is the likeliest source of the owner’s half-remembered “useful information”. The reading neither confirms nor rules that out: the paper’s name for what it measures is “semantic information”, and it lists “meaningful”, “relevant”, “functional” and “pragmatic information” as names others have used (footnote 1); “useful” is not among them. The idea, information that matters for staying in existence, fits the description.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- One toy example, five locations, deterministic agent, two variants.
- Choices carry the result. Viability function, split, τ, initial distribution and the class of interventions are all the analyst’s.
- Errors and slips. Appendix B writes the viability as V = S(p_{X_τ}) + Σ p(x_τ) S_int(x_τ), the total entropy without the minus sign that the definition (2) requires. It twice calls the semantic efficiency κ_stored (≈ 0.6) though κ is the multiplier. The main text starts the agent “in location 3 (the middle of the world)”; Appendix B puts it at ⌊n/2⌋ = 2 for n = 5. Appendix A’s bound is approximate: the derivation drops a term to reach the displayed form, which the text marks only by ”≈”. Eq. (10)‘s text says S_stored “must be less than” the mutual information where it can equal it.
Open questions
- Is there a viability function derivable from the dynamics alone (the attractor landscape the authors mention), which would remove the circularity?
- Does maximizing semantic information over splits pick out an agent in any system where the answer is known, rather than the whole world?