Conservation-Based Internal State Updates

• Conservation-based internal state updates are computational methods that enforce invariant quantities, such as entropy and mass, during state transitions across diverse domains.
• They enhance model stability and performance by ensuring that key properties remain conserved in neural networks, quantum simulations, and streaming algorithms.
• These approaches employ both denotational and operational constraints to achieve precise updates, leading to improved accuracy, consistency, and efficiency in complex systems.

Conservation-based internal state updates are computational and modeling strategies in which state transitions are explicitly constrained by underlying invariants—often entropy, conserved physical quantities, or user-specified partial information—in order to guarantee that particular forms of information, structure, or resource are preserved through the update process. These approaches are deployed in diverse domains, including comparison-based algorithms (Schellekens, 2020), neural sequence models (Thornton et al., 2019, Deng et al., 2020), streaming algorithms (Jayaram et al., 2024), physical neural operators (Dong et al., 20 Nov 2025, Hansen et al., 2023), quantum simulation and measurement (Muth, 2011, Collins et al., 2024, Aharonov et al., 2024, Polo-Gómez et al., 23 Jun 2025), constraint programming (Lin et al., 2016), RL environment modeling (Scholz et al., 2021), and bidirectional programming (Matsuda et al., 8 Jan 2026). Conservation principles in internal state updates serve to ensure that no irrecoverable loss of information or violation of domain laws occurs, and provide rigorous invariants for modular analysis or compositional design.

1. Formal Modeling of State Conservation Constraints

The formulation of conservation-based updates depends on the semantics and structure of the system:

• Comparison-based algorithms: Internal states are modeled as multisets of topological sorts of finite partial orders, with operations corresponding to poset refinements. Entropy—a measure of the disorder or freedom in data arrangements—is quantified on these states, and update rules are dictated by a conservation law ensuring the sum of quantitative entropy and positional entropy remains invariant at $\log_2(n!)$, the maximal entropy of $n$ elements (Schellekens, 2020).
• Neural and stochastic systems: In architectures such as Gaussian-gated LSTMs (Thornton et al., 2019) and Alternating ConvLSTM (Deng et al., 2020), gates or alternate accumulators that conserve state (by copying or accumulating rather than overwriting) promote memory persistence and prevent vanishing gradients.
• Probabilistic PDE solvers: Frameworks such as ProbConserv (Hansen et al., 2023) update predictive distributions via Bayesian conditioning on exact integral conservation laws, enforcing volume or mass conservation in the post-update posterior.
• Quantum systems: Conservation of total charge, particle number, energy, or angular momentum is enforced at each update, either by block-structuring tensors (MPO/MPS) (Muth, 2011), reformulating measurement channels to include quantum reference frames (Collins et al., 2024, Aharonov et al., 2024), or partitioning multipartite states into contextual registers ("polyperspective" formalism) to guarantee local and global invariance (Polo-Gómez et al., 23 Jun 2025).
• Bidirectional programming: Partial-state lenses formally capture conservation of user intentions by partial ordering and compositional merging of internal state specifications, ensuring no intent is lost in multi-view reconciliation (Matsuda et al., 8 Jan 2026).

2. Entropy and Conservation Laws in State Update Mechanisms

Central to conservation-based updates is the rigorous quantification and transfer of entropy, information, or conserved quantity:

• Entropy conservation in comparison-based computation: For posets with dual structures (e.g., discrete and total orders), the sum $H_q + H_p = \log_2(n!)$ partitions between label freedom and positional freedom, with every primitive (swap, split, merge) maintaining this invariant. In sorting, the decrease in quantitative entropy is exactly matched by an increase in positional entropy, exemplifying strict conservation (Schellekens, 2020).
• Quantum relative entropy in quantum games: Matrix Multiplicative Weights Update (MMWU) and continuous quantum replicator dynamics conserve the total quantum relative entropy from Nash equilibrium, providing a constant of motion analogous to Hamiltonian flows. Poincaré recurrence and boundedness of orbits in the state space are consequences of this invariance (Jain et al., 2022).
• Physical conservation laws in ML models: In both ProbConserv and the Exterior-Embedded Conservation Framework (ECF), explicit integral constraints (e.g., $\int u(x)dx = \textrm{const}$ for mass) are imposed post-prediction to restore physical law compliance, yielding both exact conservation and provable reductions in error (Hansen et al., 2023, Dong et al., 20 Nov 2025).

3. Operational Recipes and Algorithmic Construction

State conservation is achieved through specific operational protocols tailored to the domain:

Domain	State Conservation Recipe	Paper
Comparison-based algs.	Identify poset $P, P^*$, compute $H_q, H_p$, ensure $H_q+H_p$ preserved per primitive	(Schellekens, 2020)
RNN/LSTM	Insert gating or alternation to copy/skip updates, accumulate state without overwriting previous contents	(Thornton et al., 2019, Deng et al., 2020)
Streaming algorithms	Use sample-and-hold schemes and approximate counters to match lower bounds on state changes with optimal space	(Jayaram et al., 2024)
SciML/PDE solvers	Update predictive mean/covariance by Bayesian conditioning on integral (conserved) constraints; maintain UQ	(Hansen et al., 2023, Dong et al., 20 Nov 2025)
Quantum simulation	Block structure tensors by conserved quantum numbers, project updates to invariant sectors	(Muth, 2011, Collins et al., 2024, Aharonov et al., 2024, Polo-Gómez et al., 23 Jun 2025)
Bidirectional lenses	Merge partial specifications using $\oplus$, enforce update and intention conservation via poset laws (C), (A), (S)	(Matsuda et al., 8 Jan 2026)

4. Denotational versus Operational Conservation

There is a distinction between conservation defined over entire states (denotational) and conservation guaranteed at each operational step:

• Denotational conservation: Provides invariance over the sequence as a whole; e.g., entropy before and after sorting in comparison-based algorithms, or global post-measurement statistics in quantum mechanics (Schellekens, 2020, Collins et al., 2024, Aharonov et al., 2024).
• Operational conservation: Seeks to ensure reversibility and invariance at each primitive operation; only fully realized in some frameworks such as block-structured quantum updates and partial-state lens well-behavedness laws (C),(A),(S) (Muth, 2011, Matsuda et al., 8 Jan 2026). Operational entropy conservation enables granular analysis and design of local update steps with global guarantees.

5. Impact on Performance, Complexity, and Consistency

Conservation-based updates confer tractable complexity bounds, improved stability, and quality:

• Algorithmic complexity: Streaming moment estimation incurs $\Theta(n^{1-1/p})$ state changes, matched by near-optimal algorithms that also minimize space (Jayaram et al., 2024).
• Model accuracy and stability: Enforcement of conservation losses in data-driven models can lead to RMSE reductions of 10–40% and bring global conservation error to numerical machine precision (Dong et al., 20 Nov 2025, Hansen et al., 2023).
• Consistency and compositionality: In bidirectional transformations, partial-state lenses allow arbitrary merging of user intentions without information loss, supporting compositional reasoning across multiple views and edits (Matsuda et al., 8 Jan 2026).
• Restoration efficiency: Recollection in constraint programming avoids domain recomputation by conserving changed variable domains, outperforming recomputation in settings with high propagation cost (Lin et al., 2016).

6. Domain-specific Interpretations and Extensions

The explanatory significance varies across fields:

• Physical modeling: Conservation-imposed updates restore physical or statistical law compliance in machine-learned models, even for complex shocks, discontinuities, or heteroscedastic uncertainty (Hansen et al., 2023, Dong et al., 20 Nov 2025).
• Quantum information: Conservation-enforced update rules address foundational quantum measurement paradoxes, requiring explicit inclusion of reference frames and apparatus to ensure per-run invariance, not just ensemble statistics (Collins et al., 2024, Aharonov et al., 2024, Polo-Gómez et al., 23 Jun 2025).
• Neural architectures: Sparse update mechanisms (Gaussian gating) preserve memory and gradient stability over long temporal spans while dramatically reducing compute (Thornton et al., 2019).
• Bidirectional data management: Compositional merge and refinement laws ensure that updates from distinct views are never lost, yielding robust multi-user and multi-view data correspondence (Matsuda et al., 8 Jan 2026).

7. Open Problems and Future Directions

Conservation-based update frameworks stimulate further research:

• Extension to broader streaming models (turnstile), non-randomized algorithms, and broader summary structures in streaming contexts (Jayaram et al., 2024).
• Operational conservation for arbitrary primitives in comparison-based algorithms, and compositional entropy-preserving semantics (Schellekens, 2020).
• Improved mechanisms for conservation enforcement in neural operators and uncertainty quantification, including adaptive or soft constraint integration (Dong et al., 20 Nov 2025, Hansen et al., 2023).
• Explicit canonical modeling of per-run conservation in quantum measurement theory, impacting quantum cryptography, thermodynamics, and clock synchronization (Collins et al., 2024, Aharonov et al., 2024).
• Formal generalization of partial-state lens laws to databases, provenance tracking, and collaborative data transformation (Matsuda et al., 8 Jan 2026).

A plausible implication is that conservation-based internal state update principles will underpin algorithm design, model regularization, and foundational semantics in any context where invariants—information-theoretic, physical, or user-centric—must be robustly maintained throughout computation.