Recoverability in Complex Systems

• Recoverability is defined as the ability to reconstruct a target object from degraded, missing, or corrupted data using structured methods like optimization and decoding.
• It is evaluated through domain-specific criteria, from exact recovery in compressive sensing and coding theory to fidelity bounds in quantum information theory.
• The concept underpins reliable system design in applications such as error correction, crash recovery, and causal inference under missing-data scenarios.

Searching arXiv for the cited works to ground the article in current literature. arxiv_search query: "Recoverability Has a Law: The ERR Measure for Tool-Augmented Agents (Vuddanti et al., 29 Jan 2026)" arxiv_search query: "recoverability quantum information theory (Wilde, 2015)" Recoverability is a domain-dependent term for the possibility of reconstructing a lost, corrupted, partially observed, or semantically invalid object, state, trajectory, or relation from available information. In contemporary research, it appears in at least three closely related senses: exact reconstruction of a latent target, approximate reversibility with quantitative error control, and preservation of operation within a regime from which correction remains possible. These senses occur in tool-augmented language agents, quantum information theory, compressive sensing, coding theory, persistent-memory systems, causal inference under missingness, privacy-utility tradeoffs, and decision theory (Vuddanti et al., 29 Jan 2026, Wilde, 2015, Zhang et al., 2012, Brakensiek et al., 25 Feb 2026, Cai et al., 2020, Zhang et al., 2023).

1. Conceptual scope and formal meanings

Across the cited literatures, recoverability is attached to different mathematical objects but always asks whether a target survives degradation in a form that permits principled restoration. In modified compressive sensing, the target is a sparse vector $x^*$, and recoverability means exact equality between the modified-CS optimizer and the truth, $x^{(1)} = x^*$. In coding theory, an erasure pattern $E$ is recoverable if the erased coordinates are uniquely determined by the parity constraints, equivalently $\mathrm{rank}(H|_E)=|E|$. In persistent memory allocation, recoverability means that after crash recovery the allocator metadata indicates that all and only the “in use” blocks are allocated. In missing-data causality, a query is recoverable if it can be consistently estimated from observed data alone, or equivalently expressed in terms of $P(O)$, the distribution of observable variables. In quantum information, recoverability denotes reversibility of a channel action on a set of states, exactly or approximately, via an explicit recovery map (Zhang et al., 2012, Brakensiek et al., 25 Feb 2026, Cai et al., 2020, Zhang et al., 2023, Wilde, 2015).

Domain	Recoverable object	Formal criterion
Modified-CS	Sparse signal	$x^{(1)} = x^*$
Coding theory	Erasure pattern	$\mathrm{rank}(H|_E)=|E|$
Persistent memory	Heap state after crash	All and only in-use blocks allocated
Missing-data causality	Query or ACE	Expressible from observed-data distribution
Quantum information	State/channel action	Existence of recovery channel; fidelity or entropy bound

This multiplicity does not imply vagueness. Rather, the term marks a precise relation between degradation and restoration. The degradation may be stochastic execution noise, erasures, aggregation, measurement, missingness, compression, or irreversible commitments. The restoration may be optimization, decoding, garbage collection, rollback, recovery channels, or observed-data identification. A plausible synthesis is that recoverability formalizes the boundary between information loss that is merely inconvenient and information loss that is structurally irreversible.

2. Execution-level recoverability in agents, runtimes, and adaptive decoders

In tool-augmented language-model agents, recoverability is formalized as Expected Recovery Regret (ERR), the expected gap between a given recovery policy and the optimal recovery policy under stochastic execution noise: $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$ The same work introduces the Efficiency Score (ES) as an observable trajectory-level surrogate built from recovery rate and normalized cumulative cost, with the central first-order law

$$\text{ERR}(\pi)\le \frac{1}{1-\gamma}(1-\text{ES})+O(\lambda c_{\max}), \qquad \widehat{\text{ERR}}_{\text{law}}=\frac{1}{1-\gamma}(1-\text{ES}).$$

Empirically, across five tool-use benchmarks, model scales from 8B to 32B, multiple recovery policies, and recovery horizons, predicted regret closely matched observed post-failure regret from 200 Monte Carlo rollouts per combination, with normalized deviation typically at or below $\Delta_{\text{norm}} \le 0.05$. The law is explicitly first-order: the paper reports a low-variance linear regime, a curvature regime, and a breakdown regime under large execution variance (Vuddanti et al., 29 Jan 2026).

A distinct runtime notion appears in DART, where local rollback is not accepted merely because it is controller-legal. DART defines semantic recoverability for a boundary $x^{(1)} = x^*$0 of instance $x^{(1)} = x^*$1 by

$x^{(1)} = x^*$2

The admissibility condition for local recovery further requires identification of the failed instance, checkpoint stability, scope confinement, absence of committed downstream conflict, and effect-policy permission. The key claim is that controller legality does not imply semantic validity. On the reported evaluation, DART correctly recovered all evaluated commitment-sensitive cases where baseline local recovery fails, and a five-domain safety audit reported $x^{(1)} = x^*$3 unsafe admissions and $x^{(1)} = x^*$4 false-blocked events in the audited blocked set (Yang et al., 22 May 2026).

A more thermodynamic and regime-level formulation is given for self-referential information decoders. There, recoverability requires a nonnegative feasibility margin

$x^{(1)} = x^*$5

together with local invertibility of the measurement-action mapping. The coarse-grained diagnostic is the stability ratio

$x^{(1)} = x^*$6

which diverges at the feasibility boundary. With explicit feedback, each uncertified commitment spawning on average $x^{(1)} = x^*$7 new candidates turns the transition first-order, yielding a lucid branch, a collapsed branch, a cusp-organized bistable region, spinodals, and hysteresis. For $x^{(1)} = x^*$8, load reduction alone cannot restore operation; only gating or reset can (Rooyen, 23 Jun 2026).

In on-policy robot learning, HORIZON uses recoverability as a predictive finite-horizon training signal rather than a viability-kernel statement. A candidate frontier is committed only if it passes both a locomotion-feasibility gate and a checkpoint-comparison gate: $x^{(1)} = x^*$9 This yields a checkpointed frontier curriculum with rollback and boundary refinement. The reported regularities are that direct domain widening is uneven across physical axes, domain composition is non-monotonic, and offline distillation of isolated experts does not substitute for joint on-policy interaction (Bai et al., 3 Jun 2026).

3. Reconstruction from incomplete, compressed, or noisy observations

In modified compressive sensing with partially known support, recoverability is exact sparse-signal recovery under the program

$E$0

The central result is a sufficient and necessary condition: $E$1 iff, for every subset $E$2, an associated null-space optimization has positive optimal value whenever solvable. This condition depends only on the support pattern and the signs outside $E$3, not on nonzero magnitudes, and it remains valid when the known support contains errors. The paper further derives an exact recovery-probability formula by averaging over support and sign configurations, and a sampling approximation $E$4 justified by the law of large numbers (Zhang et al., 2012).

For randomly compressed low-CP-rank tensors, recoverability is guaranteed under a subgaussian sensing map when the number of measurements is on the same order as the number of CP model parameters, without requiring sparsity or continuous-distribution assumptions on the latent factors. The argument proceeds through a CP-specific restricted isometry property, a covering-number bound for normalized tensors with bounded condition measure

$E$5

and a subgaussian embedding theorem. The dependence on $E$6 makes factor conditioning part of the recoverability theory rather than a peripheral identifiability issue (Ibrahim et al., 2020).

Recoverability of tree-structured Markov random fields under unknown $E$7-ary symmetric noise is governed by a different obstruction: leaf-cluster ambiguity. The paper proves that ambiguity is confined to leaf clusters, that the $E$8 case is universally ambiguous up to leaf-cluster permutations, and that for $E$9 recoverability depends on the latent joint PMF. Exact recovery is possible iff every leaf cluster is identifiable, which is decided by a quadratic feasibility condition. The accompanying algorithm is polynomial-time and sample-efficient, with $\mathrm{rank}(H|_E)=|E|$0 runtime for constant $\mathrm{rank}(H|_E)=|E|$1, and it returns the exact tree when possible or the maximal identifiable equivalence class otherwise (Katiyar et al., 2021).

For discrete-time signals observed on sparse, irregular, and even non-periodic subsets $\mathrm{rank}(H|_E)=|E|$2, the relevant notion is finitely robust $\mathrm{rank}(H|_E)=|E|$3-recoverability: every finite trace on $\mathrm{rank}(H|_E)=|E|$4 can be linearly and stably reconstructed from sufficiently many observations on $\mathrm{rank}(H|_E)=|E|$5, with robustness to small $\mathrm{rank}(H|_E)=|E|$6-noise. The main theorem gives dense classes $\mathrm{rank}(H|_E)=|E|$7 that satisfy this property under very mild one-sided or two-sided infinitude conditions on $\mathrm{rank}(H|_E)=|E|$8. The result is existential but constructive through explicit Z-domain kernels (Dokuchaev, 2019).

4. Recoverability in coding, storage, and crash recovery

In coding theory, recoverability becomes an architectural property. For a linear $\mathrm{rank}(H|_E)=|E|$9-code $P(O)$0, the recoverable erasure patterns form $P(O)$1, and maximal recoverability (MR) for a family $P(O)$2 means

$P(O)$3

For MR locally recoverable codes with parameters $P(O)$4, a pattern $P(O)$5 is recoverable iff one can delete at most $P(O)$6 erased symbols so that each local group retains at most $P(O)$7 erasures. For MR tensor and grid codes, the global parities again reduce recoverability to a topology-constrained core problem. In special cases the characterization is exact and combinatorial: for $P(O)$8-MR tensor codes, recoverable patterns are exactly forests in the associated bipartite graph; for $P(O)$9, they are exactly the $x^{(1)} = x^*$0-regular patterns; and for $x^{(1)} = x^*$1, recoverability is characterized by a 2-coloring condition on cycles. The survey further connects MR constructions to skew polynomial codes, higher-order MDS codes, optimal list decoding, and graph rigidity (Brakensiek et al., 25 Feb 2026).

Persistent-memory allocation uses recoverability as a correctness criterion distinct from crash-atomicity of each individual operation. Ralloc persists only the information needed to reconstruct the heap after a full-system crash, and then performs tracing garbage collection from persistent roots to rebuild descriptors, partial lists, and the superblock free list. Under the paper’s assumptions, the recovered allocator metadata marks exactly the blocks that are logically live as allocated. The design uses filter functions to identify pointer locations inside persistent blocks and offset-based pointers so that persistent regions can be remapped at arbitrary virtual addresses (Cai et al., 2020).

Transaction logging yields another variant. Poplar defines recoverability through the minimal ordering constraints needed for safe crash recovery: $x^{(1)} = x^*$2 Thus commit order must preserve RAW dependencies, and log order must preserve WAW dependencies; WAR dependencies need not be tracked. Poplar’s scalable sequence number is assigned in a decentralized manner from tuple and log-buffer metadata, and recovery replays partially constrained logs in SSN order. The underlying claim is that conventional global total ordering is stronger than necessary for correctness (Zhou et al., 2019).

5. Quantum-information-theoretic recoverability

Quantum-information theory develops the most systematic approximate-reversibility notion of recoverability. A central refinement of relative-entropy monotonicity states that for states $x^{(1)} = x^*$3 and a channel $x^{(1)} = x^*$4,

$x^{(1)} = x^*$5

Here $x^{(1)} = x^*$6 is a rotated Petz map. The interpretation is that a small drop in relative entropy implies the existence of a concrete recovery channel that exactly recovers $x^{(1)} = x^*$7 and approximately recovers $x^{(1)} = x^*$8. The same framework sharpens strong subadditivity and other entropy inequalities by attaching physically meaningful recovery terms (Wilde, 2015).

A complementary characterization uses binary hypothesis testing. For a channel $x^{(1)} = x^*$9 and pair of states $\mathrm{rank}(H|_E)=|E|$0, exact sufficiency with respect to $\mathrm{rank}(H|_E)=|E|$1 is equivalent to preservation of all Bayes-optimal binary testing errors and equivalently to preservation of the trace norm

$\mathrm{rank}(H|_E)=|E|$2

The approximate version states that near-preservation of these testing quantities implies the existence of a recovery channel with trace-norm error bounded in terms of the loss and the Hilbert projective metric $\mathrm{rank}(H|_E)=|E|$3 (Jenčová, 2023).

These results extend beyond relative entropy. For optimized quantum $\mathrm{rank}(H|_E)=|E|$4-divergences, the loss under a channel upper-bounds how well a rotated Petz map can reverse the channel action, and equality in a regular optimized $\mathrm{rank}(H|_E)=|E|$5-divergence becomes equivalent to quantum sufficiency. The same paper extends the definition of the optimized $\mathrm{rank}(H|_E)=|E|$6-divergence, the data-processing inequality, and the recoverability theorems to the general von Neumann algebraic setting (Gao et al., 2020).

A fidelity-based line of work defines the tripartite fidelity of recovery

$\mathrm{rank}(H|_E)=|E|$7

and its surprisal $\mathrm{rank}(H|_E)=|E|$8. This quantity obeys non-negativity, monotonicity under local operations on $\mathrm{rank}(H|_E)=|E|$9 or $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$0, duality for pure four-party states, dimension bounds, and continuity. It further underlies geometric squashed entanglement and the surprisal of measurement recoverability, a discord-like quantity characterizing recovery after measurement by entanglement-breaking channels (Seshadreesan et al., 2014).

6. Causal, statistical, and decision-theoretic recoverability

In causal inference with missing data, recoverability means that a target causal estimand can be consistently estimated from observed-data distributions compatible with a missingness graph. For the average causal effect,

$$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$1

recoverability is determined by the joint causal and missingness assumptions represented in m-DAGs. One paper’s central classification is that, among the canonical graphs without unmeasured common causes of missingness indicators, the ACE is recoverable in graphs A, B, and C, and non-recoverable in D, E, F, I, and J, with G and H conjectured non-recoverable. The decisive barrier is self-missingness: no incomplete variable causing its own missingness implies recoverability, while self-missingness generally destroys it. Simulations further show that compatible multiple imputation can be approximately unbiased in some theoretically non-recoverable settings, but outcome-driven missingness produces substantial bias across methods (Zhang et al., 2023).

A longitudinal case study makes the sensitivity sharper. In CHAPAS-3, the causal concentration-response curves under static interventions on efavirenz concentration were recoverable under one clinically motivated m-DAG, not recoverable under another differing by a small added arrow, and recoverable again under a third alternative. The paper introduces closed missingness mechanisms, under which there is no path between any substantive variable and its own missingness indicator. Under such mechanisms, the joint distribution $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$2, and therefore any marginal or conditional distribution, is recoverable from available cases, even when the missingness mechanism is of MNAR type (Holovchak et al., 2024).

Recoverability under aggregation is far more restrictive. For bulk gene expression data, the paper formalizes two preservation requirements: functional-form consistency, meaning that a micro-level relation $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$3 induces an aggregate relation $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$4, and conditional-independence consistency, meaning that conditional independences survive aggregation. The derived necessary and sufficient conditions show that recoverability is preserved only under linear aggregation, such as sum or mean, coupled with affine structural equations. Analyses of four bulk and four single-cell datasets further report limited empirical support for those linearity assumptions (Luo et al., 30 May 2026).

Decision theory uses a different distinction: recoverability of utility is stronger than identification of preferences. The paper defines recovery as convergence of finite-experiment utility estimates $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$5 to the true utility $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$6, $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$7, and shows that preference recovery alone does not guarantee this. In a monetary environment with objective monotonicity, however, convergence of standard or aggregative preferences implies convergence of the economically meaningful utility representation, including subjective expected utility and variational preferences. The same work also proves consistency and finite-sample rates for utility recovery under noisy choice in Lipschitz and homothetic environments (Chambers et al., 2023).

In Gaussian privacy under linear function recoverability, the recoverability constraint itself is an $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$8-distortion bound: $$\text{ERR}(\pi_\theta)=\mathbb{E}_{\mathcal F}\!\left[L(\pi_\theta,\mathcal F)-L(\pi^*,\mathcal F)\right].$$9 Privacy is measured by $$\text{ERR}(\pi)\le \frac{1}{1-\gamma}(1-\text{ES})+O(\lambda c_{\max}), \qquad \widehat{\text{ERR}}_{\text{law}}=\frac{1}{1-\gamma}(1-\text{ES}).$$0, and the paper exactly characterizes the maximal privacy $$\text{ERR}(\pi)\le \frac{1}{1-\gamma}(1-\text{ES})+O(\lambda c_{\max}), \qquad \widehat{\text{ERR}}_{\text{law}}=\frac{1}{1-\gamma}(1-\text{ES}).$$1 as a nondecreasing piecewise affine function of $$\text{ERR}(\pi)\le \frac{1}{1-\gamma}(1-\text{ES})+O(\lambda c_{\max}), \qquad \widehat{\text{ERR}}_{\text{law}}=\frac{1}{1-\gamma}(1-\text{ES}).$$2 determined by the singular values of $$\text{ERR}(\pi)\le \frac{1}{1-\gamma}(1-\text{ES})+O(\lambda c_{\max}), \qquad \widehat{\text{ERR}}_{\text{law}}=\frac{1}{1-\gamma}(1-\text{ES}).$$3. The optimal mechanism attenuates the singular modes of $$\text{ERR}(\pi)\le \frac{1}{1-\gamma}(1-\text{ES})+O(\lambda c_{\max}), \qquad \widehat{\text{ERR}}_{\text{law}}=\frac{1}{1-\gamma}(1-\text{ES}).$$4 and adds independent Gaussian noise (Nageswaran, 2023).

7. Recurring structures and limits

Taken together, these literatures suggest that recoverability is rarely a generic consequence of scale or redundancy alone. Tool-agent recoverability is reported not to be an artifact of architecture or model size; storage and coding recoverability depend on topology and parity structure; quantum recoverability depends on channel-state geometry; and causal recoverability can be destroyed by a single missingness edge or by nonlinear aggregation (Vuddanti et al., 29 Jan 2026, Brakensiek et al., 25 Feb 2026, Wilde, 2015, Holovchak et al., 2024, Luo et al., 30 May 2026).

A second recurring feature is the use of observable surrogates for latent restoration quality. ES predicts ERR in tool agents; stability ratios diagnose overload in self-referential decoders; recovery probability summarizes exact sparse recovery under random support/sign patterns; fidelity and testing error quantify quantum reversibility; and piecewise privacy curves encode the distortion budget needed to conceal singular directions of a queried function (Vuddanti et al., 29 Jan 2026, Rooyen, 23 Jun 2026, Zhang et al., 2012, Jenčová, 2023, Nageswaran, 2023).

A third recurring feature is the existence of sharply defined failure modes. Recoverability breaks under support errors that invalidate combinatorial conditions, under supercritical feedback $$\text{ERR}(\pi)\le \frac{1}{1-\gamma}(1-\text{ES})+O(\lambda c_{\max}), \qquad \widehat{\text{ERR}}_{\text{law}}=\frac{1}{1-\gamma}(1-\text{ES}).$$5, under rollback across committed downstream dependencies, under self-missingness, and under nonlinear aggregation of nonlinear causal mechanisms. This suggests that recoverability is best understood not as a generic synonym for robustness, but as a structurally constrained property of systems, representations, and interventions.