# DRCT v5 — Statistical Closure From One Noisy Continuous Stream ## 0. What changes in v5 DRCT v1–v4 used exact equality of future behavior in finite deterministic systems. Physical data are typically: - continuous; - noisy; - stochastic or effectively stochastic; - observed as one uninterrupted trajectory; - not resettable; - not accompanied by a known hidden-state bound. Exact future-word equality is therefore the wrong primitive. DRCT v5 replaces it with **equality of predictive probability laws**. This is close to computational mechanics, predictive state representations, conditional distribution embeddings, nonlinear state reconstruction and information-theoretic time-series analysis. The information-theoretic identities below are standard. DRCT's research candidate is their integration with the closure-defect / measurement-synthesis stack. --- # 1. Stochastic predictive closure Let \[ \ldots,Y_{t-2},Y_{t-1},Y_t,Y_{t+1},\ldots \] be a stationary stochastic process. Let \[ \mathcal P_t=\sigma(Y_s:s\le t) \] denote the complete observable past and \[ \mathcal F_t^+=\sigma(Y_s:s>t) \] the observable future. A descriptor \[ Z_t=\phi(Y_{\le t}) \] is **predictively closed** when \[ \boxed{ \mathcal F_t^+ \perp\!\!\!\perp \mathcal P_t \mid Z_t. } \] Equivalently, the conditional future law satisfies \[ \boxed{ \Pr(\mathcal F_t^+\in A\mid\mathcal P_t) = \Pr(\mathcal F_t^+\in A\mid Z_t) } \] for every measurable future event \(A\), almost surely. This is stochastic predictive sufficiency. --- # 2. Statistical predictive deficit Define \[ \boxed{ D(Z) = I(\mathcal F_t^+;\mathcal P_t\mid Z_t). } \] ## Theorem 29 — Statistical Closure Criterion \[ \boxed{ D(Z)=0 } \] iff the future and complete past are conditionally independent given \(Z\) (up to the standard almost-sure qualifications). Thus \(D(Z)=0\) is the stochastic analogue of DRCT's deterministic \(D_0\). This does **not** imply zero future entropy. A perfectly closed stochastic state may still satisfy \[ H(\mathcal F_t^+\mid Z_t)>0. \] Closure means no *additional past information* improves the future law; it does not mean the future itself must be deterministic. This distinction is essential. --- # 3. Finite-horizon deficit Infinite futures are not directly estimable, so define \[ F_t^{(L)} = (Y_{t+1},\ldots,Y_{t+L}) \] and \[ \boxed{ D_L(Z) = I(F_t^{(L)};\mathcal P_t\mid Z_t). } \] ## Theorem 30 — Horizon Monotonicity \[ \boxed{ D_1(Z)\le D_2(Z)\le\cdots\le D(Z) } \] whenever the information quantities are well-defined. ### Proof By the chain rule, \[ I(F^{(L+1)};\mathcal P\mid Z) = I(F^{(L)};\mathcal P\mid Z) + I(Y_{t+L+1};\mathcal P\mid Z,F^{(L)}), \] and the second term is nonnegative. QED. Therefore a descriptor can appear closed at one horizon and fail at a deeper future horizon. --- # 4. New dimension gain Let \(Q_t\) be any additional measurable coordinate from history, another sensor, context or intervention response. Define \[ Z_t'=(Z_t,Q_t). \] ## Theorem 31 — Predictive Gain Identity For a finite future block, \[ \boxed{ H(F^{(L)}\mid Z) - H(F^{(L)}\mid Z,Q) = I(F^{(L)};Q\mid Z)\ge0. } \] This is the stochastic analogue of DRCT's relational-dimension gain. A candidate dimension contributes no predictive information exactly when \[ \boxed{ I(F^{(L)};Q\mid Z)=0. } \] --- # 5. Optimal log-loss theorem Let a probabilistic predictor output a conditional density or probability law \(q(f\mid z)\). Its expected logarithmic score is \[ \mathcal R(q;Z) = \mathbb E[-\log q(F\mid Z)]. \] The Bayes-optimal predictor is the true conditional law. ## Theorem 32 — Log-Loss Closure Identity \[ \boxed{ \mathcal R^\*(Z) - \mathcal R^\*(Z,Q) = I(F;Q\mid Z). } \] ### Proof The Bayes log-loss risk is conditional entropy (or the corresponding expected negative log density in the continuous case): \[ \mathcal R^\*(Z)=H(F\mid Z). \] Apply Theorem 31. QED. This gives DRCT an operational score for noisy continuous systems: > how many predictive bits per observation does the proposed new dimension > remove? --- # 6. No information from deterministic re-expression survives noise If \[ Q=g(Z) \] almost surely, then \[ \boxed{ I(F;Q\mid Z)=0. } \] Therefore deterministic feature engineering of the same descriptor cannot create new Bayes-predictive information. It may make a restricted model class easier to train, but it does not change the information available to an optimal predictor. This distinction separates **representation convenience** from **genuine new state information**. --- # 7. History-depth curve Let \[ H_t^{(m)} = (Y_t,Y_{t-1},\ldots,Y_{t-m+1}). \] Define captured predictive information \[ \boxed{ G_{m,L} = I(F_t^{(L)};H_t^{(m)}). } \] Since \(H^{(m)}\) is a function of \(H^{(m+1)}\), ## Theorem 33 — History Information Monotonicity \[ \boxed{ G_{1,L}\le G_{2,L}\le\cdots. } \] The incremental gain is \[ \boxed{ \Delta_{m,L} = I(F_t^{(L)};Y_{t-m}\mid H_t^{(m)}). } \] For a finite-order Markov process of order \(k\), \[ \boxed{ \Delta_{m,L}=0 \quad\forall m\ge k } \] for all future horizons \(L\). Thus a history-gain curve is the noisy/stochastic analogue of DRCT's closure depth. --- # 8. Predictive causal states Define two complete pasts \(p,p'\) to be predictively equivalent when \[ \boxed{ \Pr(\mathcal F^+\in A\mid\mathcal P=p) = \Pr(\mathcal F^+\in A\mid\mathcal P=p') } \] for every measurable future event \(A\). Their equivalence classes form the process's predictive causal states. This is standard computational mechanics: causal states are minimal sufficient statistics of the past for prediction. DRCT does not claim this construction as novel. DRCT uses the causal-state object as the stochastic target \(C_\*\) for its deficit, sensor and intervention machinery. --- # 9. Continuous observations Exact equality of continuous conditional laws is impractical. One may represent a predictive law by: - parametric conditional densities; - characteristic functions; - quantiles; - kernel mean embeddings; - conditional mean embeddings in an RKHS; - a sufficiently rich collection of future tests. If a characteristic kernel is used, equality of kernel mean embeddings implies equality of the represented distributions. Thus the finite DRCT signature \[ \text{future word} \] generalizes to \[ \boxed{ \text{predictive distribution representation}. } \] Hilbert-space embeddings of PSRs are established prior art for continuous actions/observations. --- # 10. One-stream asymptotic theorem Reset access is not necessary for statistical testing. Let the process be stationary and ergodic. Let two predictors \(q_0,q_1\) be fixed before evaluation, and define their per-time log-loss difference \[ X_t = -\log q_0(F_t\mid Z_t) + \log q_1(F_t\mid Z_t,Q_t). \] Assume \(X_t\) is integrable. ## Theorem 34 — Forward Stream Risk Convergence By Birkhoff's ergodic theorem, \[ \boxed{ \frac1n\sum_{t=1}^n X_t \longrightarrow \mathbb E[X_t] \quad\text{almost surely}. } \] Therefore, after freezing the predictors, a genuine expected predictive gain can be estimated consistently from one sufficiently long ergodic trajectory. This is asymptotic, not a finite-sample universal certificate. --- # 11. Finite-sample defect witness Suppose two fixed probabilistic predictors are compared on a future block that was not used to choose or fit them. If the augmented predictor repeatedly achieves lower held-out log-loss, that is empirical evidence that the tested representation/model extracts predictive information unavailable to the baseline model. However: \[ \boxed{ \text{restricted-model improvement} \not\Rightarrow I(F;Q\mid Z)>0 } \] without assumptions about model consistency. A derived feature \(Q=g(Z)\) can improve a weak model's finite-sample score even though its true conditional mutual information is zero. Therefore DRCT distinguishes: 1. **information gain** — property of the underlying distribution; 2. **model-class gain** — property of a chosen decoder. --- # 12. Statistical defect rejection versus closure certification From noisy finite data, one can often reject a descriptor more easily than certify it. A significant and replicated improvement from genuinely additional information \(Q\) can reject the hypothesis that \(Z\) is sufficient for the tested future family under the statistical assumptions used. But failure to find an improvement does not prove \[ D(Z)=0. \] An untested future event, rare regime, inaccessible sensor or insufficient estimator may still reveal a defect. Thus: \[ \boxed{ \text{finite statistical non-rejection} \neq \text{universal closure}. } \] --- # 13. Statistical closure spectrum Choose a nested family of future tests \[ \mathcal T_1\subseteq\mathcal T_2\subseteq\cdots. \] Let \(F^{(n)}\) denote the vector of outcomes of tests through level \(n\). Define \[ \boxed{ D_n(Z)=I(F^{(n)};\mathcal P\mid Z). } \] Then \[ D_n(Z) \] is nondecreasing with test richness. This is the statistical analogue of DRCT's deterministic closure spectrum: deeper/richer future questions can expose distinctions hidden by shallow tests. --- # 14. Statistical operational synthesis For candidate observable \(Q_j\), define validated predictive gain \[ \widehat G_j = \widehat{\mathcal R}(Z) - \widehat{\mathcal R}(Z,Q_j). \] A practical DRCT search may rank dimensions by \[ \boxed{ \widehat G_j-\lambda K(Q_j) } \] and require: - chronological training; - frozen forward evaluation; - block-aware uncertainty estimates; - replication across regimes; - no promotion of dimensions selected only after viewing the final block. This is an empirical algorithm, not a theorem that the largest finite-sample gain is the physically correct hidden variable. --- # 15. What v5 proves The mathematical layer proves: 1. stochastic predictive closure is conditional independence; 2. the predictive deficit is conditional mutual information; 3. finite-horizon deficit grows with future richness; 4. added-information gain equals conditional mutual information; 5. optimal log-loss reduction equals that same information gain; 6. deterministic re-expression cannot add Bayes-predictive information; 7. nested history captures nondecreasing predictive information; 8. fixed forward risks converge on stationary ergodic streams. These are information-theoretic/probability results, many of them standard. --- # 16. What v5 does not prove It does not prove: - that physical reality is deterministic; - that every apparent stochasticity is reducible; - that one finite noisy stream identifies all causal states; - that a chosen ML model estimates conditional mutual information consistently; - that bootstrap intervals are distribution-free certificates; - that any particular real-world signal has reached closure. The point of v5 is to preserve DRCT's logic after exact equality becomes physically unrealistic. --- # 17. The reality-theory statement after v5 The strongest defensible DRCT hypothesis is now: > **Observed state descriptions are predictively closed only when no additional > accessible past, relational, contextual, sensor or intervention information > changes the conditional law of future observations. Apparent noise should be > retained as unresolved predictive structure until this conditional-independence > criterion survives progressively richer tests.** This formulation permits two outcomes: ### Reducible uncertainty A new coordinate \(Q\) yields \[ I(F;Q\mid Z)>0. \] The previous representation was incomplete. ### Irreducible relative uncertainty After a sufficient state has been identified, \[ I(F;\mathcal P\mid Z)=0 \] while \[ H(F\mid Z)>0. \] Then the state is closed even though the future remains probabilistic. That is where DRCT cleanly separates **missing state information** from **stochasticity conditional on a sufficient state**.