O0-MATH-002

Markov Blanket Formation: Measure-Theoretic Framework

**Version:** 1.0

**Research status:** ACTIVE

CLAIM STATUS: FORMAL DEFINITION
EVIDENCE TYPE: MATHEMATICAL FORMALIZATION
PHYSICAL VALIDATION: NONE
INDEPENDENT REPLICATION: N/A
PHILOSOPHICAL PROVENANCE: O/0 ARCHIVE
ARCHIVE ENDORSEMENT: LIMITED TO REPORTED RESULT

Abstract

This document provides a rigorous measure-theoretic formulation of Markov blanket formation within the O/0 substrate framework. We define Markov blankets using conditional independence on probability spaces, establish Shannon entropy measures over weight vectors, and derive formal conditions under which blanket structures emerge from initially homogeneous substrates. The treatment connects to the Free Energy Principle (FEP) literature while maintaining strict separation between mathematical formalism and physical or metaphysical claims.

Source proposition

The O/0 archive claims that boundaries between "self" and "environment" are not fundamental but emerge dynamically from a continuous substrate. The Markov blanket formalism from the FEP literature provides a natural mathematical framework for defining such functional boundaries in terms of conditional statistical independence.

Scientific audit

The Markov blanket concept originates in Bayesian network theory (Pearl, 1988) and has been extensively applied in theoretical neuroscience (Friston, 2013). Its use here is mathematically standard. The novel claim is that blanket structures emerge *spontaneously* in a continuous substrate simulation without pre-encoded topology — this is an empirical claim requiring computational verification, not a mathematical theorem. This document provides the formal scaffolding for that empirical investigation.

Research question

Under what measure-theoretic conditions does a probability distribution over substrate states exhibit conditional independence structures interpretable as Markov blankets, and what is the minimum entropy configuration required for stable blanket maintenance?

Operational definitions

1. **Probability space**: A triple (Omega, F, P) where Omega is the sample space, F is a sigma-algebra of events, and P is a probability measure.

2. **Substrate state**: A random variable S: Omega -> R^{N x d} representing the full configuration of N grid cells, each with d-dimensional state.

3. **Local state**: For cell i, the random variable S_i: Omega -> R^d giving its individual state.

4. **Weight vector**: For cell i, the vector w_i = (w_{i1}, ..., w_{iK}) in R^K representing connection strengths to K neighbors.

5. **Markov blanket**: A subset MB(i) of {1, ..., N} such that S_i is conditionally independent of S_{ext(i)} given S_{MB(i)}, where ext(i) = {1,...,N} \ ({i} union MB(i)).

Hypothesis

In a substrate initialized with uniform random weights, local prediction-error minimization dynamics will drive the weight distribution toward configurations exhibiting conditional independence structures (Markov blankets) with Shannon entropy of weight vectors significantly below the maximum-entropy baseline.

Null hypothesis

Weight evolution under prediction-error minimization produces distributions indistinguishable from maximum-entropy (uniform) distributions, and no statistically significant conditional independence structures emerge.

Competing explanations

1. Apparent blanket structures may be artifacts of finite sample size in the conditional independence tests.

2. Blanket formation may require specific initial conditions rather than emerging generically.

3. The conditional independence may be approximate but never exact, making the "blanket" designation a matter of threshold choice.

Formal model

**Definition 1 (Substrate Probability Space).** Let (Omega, F, P) be a probability space. Define the substrate state as a random vector S = (S_1, ..., S_N) where each S_i takes values in R^d. Let mu_S denote the joint distribution of S on R^{Nd}.

**Definition 2 (Conditional Independence).** Random variables X and Y are conditionally independent given Z, written X _||_ Y | Z, if for all measurable sets A, B:

P(X in A, Y in B | Z) = P(X in A | Z) * P(Y in B | Z) P-a.s.

Equivalently, the conditional mutual information vanishes:

I(X; Y | Z) = 0

where I(X; Y | Z) = H(X|Z) + H(Y|Z) - H(X,Y|Z).

**Definition 3 (Markov Blanket).** For cell i, a set MB(i) subset {1,...,N}\{i} is a *Markov blanket* of i if:

S_i _||_ S_{ext(i)} | S_{MB(i)}

where ext(i) = {1,...,N} \ ({i} union MB(i)). The *minimal Markov blanket* (Markov boundary) is the smallest such set.

**Definition 4 (Weight Entropy).** For cell i with weight vector w_i = (w_{i1}, ..., w_{iK}), define the normalized weight distribution:

p_{ij} = |w_{ij}| / sum_{k=1}^{K} |w_{ik}|

The *weight entropy* of cell i is:

H_w(i) = -sum_{j=1}^{K} p_{ij} * log(p_{ij})

with the convention 0 * log(0) = 0. The maximum value is H_w^{max} = log(K), attained when all weights are equal.

**Definition 5 (Blanket Sharpness).** The *blanket sharpness* of cell i is:

beta(i) = 1 - H_w(i) / H_w^{max}

This quantity lies in [0, 1]. beta(i) = 0 corresponds to uniform weights (no blanket structure), and beta(i) = 1 corresponds to a single nonzero weight (maximally sharp blanket).

**Theorem 1 (Blanket Entropy Bound).** If cell i has a well-defined Markov blanket MB(i) with |MB(i)| = m < K, then the weight entropy satisfies:

H_w(i) <= log(m)

That is, the effective entropy is bounded by the logarithm of the blanket size, not the total number of neighbors.

*Proof.* If S_i _||_ S_{ext(i)} | S_{MB(i)}, then the optimal prediction of S_i depends only on {S_j : j in MB(i)}. Under prediction-error minimization, weights w_{ij} for j not in MB(i) are driven toward zero (they provide no additional predictive information). At convergence, at most m weights are nonzero, so H_w(i) <= log(m). QED

**Proposition 1 (Conditional Mutual Information Criterion).** Cell i possesses an approximate Markov blanket at threshold epsilon > 0 if there exists MB_epsilon(i) subset {1,...,N}\{i} such that:

I(S_i; S_{ext(i)} | S_{MB_epsilon(i)}) < epsilon

The *blanket quality* is Q(i) = -log(epsilon_min), where epsilon_min is the smallest epsilon for which such a set exists with |MB_epsilon(i)| <= K.

**Theorem 2 (Entropy Production at Blanket).** Let the substrate evolve according to the dynamical rule:

S_i(t+1) = f(sum_{j} w_{ij}(t) * S_j(t)) + eta_i(t)

where f is a nonlinear activation, w_{ij}(t) are time-dependent weights, and eta_i ~ N(0, sigma^2) is noise. Define the local free energy:

F_i(t) = E_Q[log Q(S_i(t)) - log P(S_i(t) | S_{MB(i)}(t))]

where Q is the variational (internal model) distribution and P is the true conditional. Then F_i(t) >= 0 (by Gibbs' inequality), and under gradient descent on F_i with respect to w_i:

dw_{ij}/dt = -alpha * partial F_i / partial w_{ij}

the weight entropy H_w(i)(t) is non-increasing whenever dF_i/dt < 0.

*Proof sketch.* The free energy F_i decomposes as:

F_i = D_KL(Q || P) + constant

where D_KL is the Kullback-Leibler divergence. Gradient descent on F_i with respect to w_{ij} adjusts weights to minimize the discrepancy between the internal model and the true conditional distribution. Weights corresponding to cells outside the true Markov blanket contribute no information to reducing D_KL and are thus driven toward zero by the L2 implicit regularization of gradient descent. As weights concentrate on fewer connections, H_w decreases. QED

**Definition 6 (Blanket Emergence Time).** The *blanket emergence time* T_b(epsilon) is defined as:

T_b(epsilon) = inf{t >= 0 : I(S_i; S_{ext(i)} | S_{MB(i)}(t)) < epsilon for all i in A}

where A is a designated subset of cells (typically those that form persistent agents in simulations).

**Proposition 2 (Spatial Blanket Structure).** Define the *blanket indicator field* on the substrate:

B(x, t) = 1 if x is in some cell's Markov blanket MB(i) for some i in A

B(x, t) = 0 otherwise

If blankets are spatially coherent (connected subsets of the grid), then B(x, t) partitions the substrate into interior (agent), blanket, and exterior (environment) regions, recovering the tripartite structure of the FEP.

Methods

1. Initialize substrate with N = 10000 cells, K = 8 neighbors (Moore neighborhood), random weights w_{ij} ~ Uniform(0, 1).

2. Run prediction-error minimization dynamics for T = 50000 timesteps.

3. At each checkpoint (every 1000 steps), estimate conditional mutual information I(S_i; S_{ext(i)} | S_{MB(i)}) using k-nearest-neighbor estimators (Kraskov et al., 2004).

4. Compute weight entropy H_w(i) for all cells. Track beta(i) over time.

5. Test Theorem 1 by comparing H_w(i) to log(|MB(i)|) at convergence.

Controls

1. Shuffled-weight control: randomly permute weights at each checkpoint. Verify that no conditional independence structure appears.

2. Fixed-weight control: run dynamics with frozen weights. Verify that blanket sharpness does not increase.

3. Independent-noise control: replace S_j inputs with independent noise. Verify that all weights converge to equal values (maximum entropy).

Predictions

1. Weight entropy H_w(i) will decrease monotonically during learning for cells that form blankets.

2. Blanket sharpness beta(i) will converge to values > 0.5 for agent-interior cells and remain < 0.1 for diffuse substrate cells.

3. Blanket emergence time T_b will scale as O(N * log(K)) under standard learning rates.

4. The spatial blanket indicator B(x, t) will form connected, closed curves (in 2D) around agent regions.

Falsification criteria

1. If weight entropy does not decrease under prediction-error minimization, the entropy bound (Theorem 1) is violated, and the formal model is incorrect.

2. If conditional mutual information does not decrease below threshold for any cell, no Markov blankets form and the hypothesis is refuted.

3. If blanket structures are not spatially coherent, the FEP tripartite structure does not apply to this substrate.

Results / Expected Outcomes

No simulation results are reported in this document. The formal framework is complete and ready for computational testing under O0-SIM protocols. Preliminary analysis suggests that the entropy bounds in Theorem 1 are tight for simple test cases (2-cell, 4-cell systems).

Uncertainty

1. The k-nearest-neighbor mutual information estimator has known bias in high dimensions; d > 10 may produce unreliable estimates.

2. The proof of Theorem 2 relies on implicit regularization of gradient descent, which is empirically observed but not universally guaranteed.

3. Approximate Markov blankets (Proposition 1) depend on the threshold epsilon, and there is no canonical choice of epsilon.

Limitations

1. The measure-theoretic framework assumes the substrate state has a well-defined joint probability density, which requires ergodicity assumptions about the dynamics.

2. Finite-sample estimation of conditional mutual information is computationally expensive (O(N^2 * K)) and may be infeasible for large substrates.

3. The connection between Markov blankets in probability spaces and physical boundaries is metaphorical — conditional independence is a statistical property, not a spatial one, though spatial coherence is a testable prediction.

Replication status

Mathematical framework is self-contained and independently verifiable. No computational results to replicate at this stage.

Data and code

No data or code generated. Mutual information estimation to use the JIDT library (Lizier, 2014) or custom k-NN estimators.

Relationship to philosophical archive

The archive's claim that boundaries are "functional appearances" maps naturally to the Markov blanket formalism, where boundaries are defined by conditional independence rather than physical substance. However, this mapping is an interpretive choice, not a deduction. The mathematical framework is valid regardless of the archive's philosophical claims.

References

1. Pearl, J. (1988). *Probabilistic Reasoning in Intelligent Systems*. Morgan Kaufmann.

2. Friston, K. J. (2013). "Life as we know it." *Journal of the Royal Society Interface*, 10(86), 20130475.

3. Kraskov, A., Stögbauer, H., & Grassberger, P. (2004). "Estimating mutual information." *Physical Review E*, 69(6), 066138.

4. Cover, T. M., & Thomas, J. A. (2006). *Elements of Information Theory* (2nd ed.). Wiley.

5. Lizier, J. T. (2014). "JIDT: An information-theoretic toolkit for studying the dynamics of complex systems." *Frontiers in Robotics and AI*, 1, 11.

6. Parr, T., Da Costa, L., & Friston, K. J. (2020). "Markov blankets, information geometry, and stochastic thermodynamics." *Philosophical Transactions of the Royal Society A*, 378(2164), 20190159.

Revision history

• v1.0: Initial formalization. Definitions 1–6 established. Theorems 1–2 stated with proof sketches. Propositions 1–2 formulated.