Law of Data Reconstruction

• Law of Data Reconstruction is a set of principles that define when incomplete observations can uniquely recover an underlying structure or dataset.
• It encompasses discrete continuity laws, such as the discrete Lipschitz condition in digital topology, ensuring gradual variation in reconstructions.
• The law also underpins dynamic system recovery and privacy guarantees by establishing capacity thresholds and optimal operators for secure data reconstruction.

The expression “Law of Data Reconstruction” is used in the literature to denote a family of law-like principles governing when incomplete observations suffice to recover an underlying object, process, or dataset, and when such recovery is impossible, non-unique, or unsafe. Across arXiv work, the phrase appears in digital topology, nonequilibrium statistical mechanics, data-driven dynamical systems, information propagation on trees, functional data analysis, and privacy theory; taken together, these uses suggest that a reconstruction law is typically formulated as an existence condition, an identifiability theorem, a threshold phenomenon, or a universal reconstruction principle relative to a specified observation model (Chen, 2010, Rau, 2015, Kneip et al., 2017, Cohen et al., 2024).

1. Scope of the concept

A central feature of the topic is that it does not admit a single context-independent definition. One line of work explicitly argues that “a single all-encompassing definition may not exist,” and therefore approaches the subject by separating two questions: what conditions guarantee protection against reconstruction attacks, and under what circumstances an attack clearly indicates that a system is not protected (Cohen et al., 2024). Another line formulates reconstruction categorically and treats uniqueness only up to equivalence: dynamical systems, observed dynamical systems, and timeseries data are organized as categories, the data-generation map is a functor, and reconstruction algorithms are proper only when they are functors from the category of timeseries-data into the category of dynamical systems (Das et al., 2024).

A Bayesian formulation further shifts attention from absolute recovery to recovery relative to a prior, a meta-distribution $\Pi$, and attacker side information $K$. In that setting, reconstruction is defined through a relation $R$, and security is assessed by comparing success on the actual dataset $S$ with success on a fresh dataset $T\leftarrow \mathcal D$ drawn from the same underlying distribution, possibly conditioned on the same side information (Kaplan et al., 29 May 2025). This suggests that the “law” is not a single formula but a structured family of criteria specifying when data determine hidden structure and when they do not.

2. Discrete continuity laws on graphs and manifolds

In digital-discrete surface reconstruction, the basic problem is to reconstruct a function $F$ on a discrete domain $D$ from values prescribed on a finite guiding set $J\subset D$. The domain may be a graph, a digital manifold, a triangulated mesh, or a general CW-complex, and the reconstructed function is required to satisfy a discrete continuity condition called gradual variation. If $F(a)=A_i$ and $a,b$ are adjacent, then $K$0 must equal $K$1, $K$2, or $K$3; equivalently, for adjacent points, $K$4 in level index (Chen, 2010).

The central existence theorem gives a necessary and sufficient condition for a gradually varied extension. If $K$5 with $K$6 and $K$7, then there exists an extension $K$8 with $K$9 and $R$0 gradually varied on $R$1 if and only if

$R$2

for all $R$3, where $R$4 is graph distance. This is explicitly a discrete Lipschitz condition and functions as a law of existence: if the sample data vary faster than the graph metric allows, no continuous discrete extension exists (Chen, 2010).

The same work characterizes the method as “truly universal and nonlinear.” Its universality consists in working on any graph or manifold, requiring adjacency but not a particular triangulation, Voronoi diagram, Delaunay triangulation, polynomial basis, or spline family. Its nonlinearity lies in the constraint $R$5, which defines a feasible set by inequalities rather than by a linear basis expansion $R$6. Higher-order smoothness is then introduced by combining gradually varied reconstruction with finite-difference operators and smooth approximations, yielding discrete analogues of $R$7, $R$8, and $R$9 fitting (Chen, 2010).

3. Reconstruction of dynamical laws from incomplete observations

In nonequilibrium Markovian dynamics, the law of reconstruction concerns recovery of time and equations of motion from untimed macrostate snapshots. The setting is a manifold of Gibbs macrostates $S$0 equipped with the Bogoliubov–Kubo–Mori metric $S$1. For genuinely irreversible processes, entropy is strictly increasing, $S$2, which permits entropy to serve as a proxy time parameter. One defines a vector field $S$3 by $S$4, and the true physical vector field is

$S$5

Near equilibrium, the macrodynamics satisfies

$S$6

so reconstruction reduces to determining the entropy production rate $S$7, then recovering $S$8, and finally inferring the generator tensor $S$9. Under the paper’s assumptions—Markovianity, genuine irreversibility, canonical equilibrium, and an effective Hamiltonian—this reconstruction is unique up to multiplicative time scaling (Rau, 2015).

A different dynamical formulation reconstructs normal forms directly from data using informed observation geometries. Observations are arranged in a tensor

$T\leftarrow \mathcal D$0

whose axes correspond to parameters, variables, and time. The method builds informed distances by coupling these axes through partition trees, multiscale filter banks, and diffusion maps, producing intrinsic embeddings for parameter space, state space, and time. In the Bogdanov–Takens example, the parameter embedding is described as homeomorphic to the true bifurcation diagram and the variable embedding recovers the minimal two-dimensional state space; in the coupled-pendula example, the time embedding recovers the intrinsic normal-mode frequencies from movies and even from random projections of the frames (Yair et al., 2016).

A categorical formalization generalizes these ideas by treating a dynamical system as a functor $T\leftarrow \mathcal D$1, an observed dynamical system as an object in a comma category, and the passage from observed dynamics to timeseries as a composite Data functor. In this framework, a proper reconstruction algorithm is itself a functor

$T\leftarrow \mathcal D$2

The inversion of data into dynamics is then expressed through Kan extensions: the best outer approximation is the left Kan extension of the projection functor along Data, and the best inner approximation is the corresponding right Kan extension. Under the paper’s assumptions, both exist, and for observable discrete-time systems the best outer approximation is consistent, so exact reconstruction holds on the observable subcategory (Das et al., 2024).

4. Threshold laws in hierarchical stochastic systems

In broadcasting processes on infinite $T\leftarrow \mathcal D$3-ary trees, the reconstruction problem asks whether boundary data at level $T\leftarrow \mathcal D$4 retains non-vanishing information about the root as $T\leftarrow \mathcal D$5. For a finite-state Markov channel with transition matrix $T\leftarrow \mathcal D$6, a standard criterion is reconstructibility when, for some root states $T\leftarrow \mathcal D$7,

$T\leftarrow \mathcal D$8

The classical baseline is the Kesten–Stigum bound: if $T\leftarrow \mathcal D$9, where $F$0 is the second eigenvalue in absolute value, then reconstruction is possible (Liu et al., 2018).

The $F$1 asymmetric model with community effects shows that this baseline need not be tight. In that model, states correspond to $F$2, the stationary distribution satisfies

$F$3

and the channel has two communities, $F$4 and $F$5. The main theorem proves that when

$F$6

the Kesten–Stigum bound is not sharp: there exist channels with $F$7 for which reconstruction still holds (Liu et al., 2018).

The proof is based on refined moment recursion, concentration estimates, and analysis of an asymptotic four-dimensional nonlinear second-order dynamical system. The resulting interpretation is that reconstruction on trees is governed not only by the spectral quantity $F$8, but also by asymmetry, community structure, and nonlinear amplification of correlations. This suggests a broader threshold law: exponential growth of observations, spectral decay, and higher-order interaction terms jointly determine whether information survives to large depth (Liu et al., 2018).

5. Optimal reconstruction of partially observed functional data

For partially observed functional data, the law of reconstruction is formulated as an optimal linear operator problem. One observes random functions $F$9 only on a subinterval $D$0, with missing domain $D$1. The goal is to reconstruct $D$2 from $D$3. The relevant operator class is much larger than classical functional regression: a linear operator $D$4 is admissible provided

$D$5

for each $D$6. Such operators admit an RKHS representation

$D$7

where $D$8 is the reproducing kernel Hilbert space induced by the covariance kernel on $D$9 (Kneip et al., 2017).

The optimal reconstruction operator is

$J\subset D$0

where $J\subset D$1 are the Karhunen–Loève scores on the observed domain and

$J\subset D$2

Its error is orthogonal to the observed part,

$J\subset D$3

and it minimizes pointwise mean squared error among all linear reconstruction operators with finite variance (Kneip et al., 2017).

A major consequence is that the usually considered regression operators

$J\subset D$4

generally cannot be optimal reconstruction operators. The argument is especially sharp at boundaries: optimal reconstruction must connect continuously to the observed fragment, whereas a Hilbert–Schmidt regression kernel cannot realize point evaluation in $J\subset D$5 (Kneip et al., 2017).

Estimation proceeds through nonparametric mean and covariance estimation, empirical eigenanalysis, and a truncated FPCA-based estimator. The theory allows autocorrelated functional data and the practically relevant design in which each of the $J\subset D$6 functions is observed at $J\subset D$7 discretization points. In the regime where $J\subset D$8 is considerably smaller than $J\subset D$9, the functional principal components based estimator can provide better rates of convergence than conventional nonparametric smoothing methods (Kneip et al., 2017).

6. Memorization, privacy, and security formulations

In modern machine learning, the law of data reconstruction appears as a capacity threshold. For random features regression, the key distinction is between label interpolation and input reconstruction. Fitting arbitrary labels typically requires $F(a)=A_i$0, but reconstructing the actual inputs $F(a)=A_i$1 requires a stronger scaling: under the paper’s assumptions,

$F(a)=A_i$2

In this regime, the span of the training feature vectors uniquely encodes the training inputs up to small perturbations and permutation; the paper therefore states a law of data reconstruction according to which the entire training dataset can be recovered as $F(a)=A_i$3 exceeds the threshold $F(a)=A_i$4. An optimization-based reconstruction procedure minimizes

$F(a)=A_i$5

and experiments show the same qualitative threshold in random features, two-layer fully connected networks, and deep residual networks, with strong reconstruction emerging when the last-layer parameter count is of order $F(a)=A_i$6 (Iurada et al., 26 Sep 2025).

A different line of work argues that reconstruction must be defined relative to a baseline. Narcissus Resiliency states that a mechanism is protected against a reconstruction relation $F(a)=A_i$7 if no attacker $F(a)=A_i$8 can produce outputs $F(a)=A_i$9 whose success probability on the true dataset $a,b$0 is much larger than the success probability of the same output on an independent fresh dataset $a,b$1: $a,b$2 This self-referential definition is presented as a general framework that captures differential privacy, one-way functions, encryption, membership inference, and predicate singling out as special cases. The same work links convincing reconstruction attacks to Kolmogorov complexity: an attack is compelling when the model output enables a short program to produce a valid reconstruction that could not be generated by a comparably short program without the model (Cohen et al., 2024).

The Bayesian extension makes the prior and attacker side information explicit. It introduces Bayesian Narcissus-resiliency and a stronger Bayesian Extraction-Safe notion based on a meta-distribution $a,b$3, side information $a,b$4, and surprisal terms such as $a,b$5. Within this framework, the paper argues that fingerprinting code attacks are really forms of membership inference rather than reconstruction attacks, and that if the goal is solely to prevent reconstruction, some impossibility results derived from fingerprinting codes no longer apply. Under Tardos-Prior, the paper gives positive results for exact averages and for Laplace-noised averages, showing that reconstruction can be ruled out even in settings where statistical memorization and membership inference remain unavoidable (Kaplan et al., 29 May 2025).

Across these privacy-theoretic formulations, the law of reconstruction becomes a statement about specificity: a release or trained model enables reconstruction only when it makes a target significantly more recoverable from the actual training dataset than from an independent sample drawn from the same underlying distribution. In that sense, the memorization literature turns the phrase from a geometric or analytical principle into a threshold-and-security doctrine: beyond $a,b$6, datasets may be encoded in parameters, whereas Narcissus-style conditions specify when a release is no more useful for reconstruction than the attacker’s own baseline (Iurada et al., 26 Sep 2025, Cohen et al., 2024, Kaplan et al., 29 May 2025).

The literature therefore presents the Law of Data Reconstruction not as a single theorem but as a recurring structural pattern. In graph settings it is a discrete Lipschitz existence law; in nonequilibrium dynamics it is an identifiability law for time and generators; in data-driven dynamics it is an intrinsic-geometry or Kan-extension law; in tree models it is a threshold law modified by asymmetry and nonlinear effects; in functional data it is an RKHS-optimality law; and in privacy theory it is a capacity or baseline-comparison law. What unifies these formulations is the claim that reconstruction becomes mathematically precise only after specifying the observation model, admissible operators, governing geometry, and the criterion by which recovery is judged.