Unfolding Operation: Revealing Hidden Structures

• Unfolding operation is a transformation that removes distortions from representations to expose underlying structural information.
• It is applied in diverse domains such as particle physics, materials science, geometry, graph theory, CHR, and molecular docking.
• Techniques range from matrix inversion and iterative Bayesian methods to mesh relaxation and algebraic unfolding, providing controlled, interpretable outputs.

An unfolding operation is a transformation that removes folding, smearing, compression, or indirection from a representation so that latent structure becomes explicit in a target domain. In arXiv usage, the term is markedly polysemous rather than singular: in particle physics it denotes correction for detector effects; in computational materials science it maps supercell bands or phonons back to a primitive-cell Brillouin zone; in computational geometry it cuts and flattens surfaces into non-overlapping nets; in graph theory it constructs cospectral graph pairs; in Constraint Handling Rules (CHR) it replaces a constraint conjunction by the body of a matching rule; and in molecular docking it expands a ligand to an unfolded conformation by maximizing internal distances (Andreassen et al., 2019, Zheng et al., 2014, Zheng et al., 2016, Kannan et al., 2024, 0807.3979, Mato et al., 2021).

1. Terminological scope and common structure

Across the cited literature, “unfolding” always acts on an object that has become difficult to interpret because information is hidden by a representation change. In detector physics, the obstruction is smearing and inefficiency; in supercell band theory, it is Brillouin-zone reduction and band folding; in polyhedral geometry, it is the impossibility of reading a planar net directly from a closed surface; in CHR, it is indirection through a rule body; and in graph theory, it is a constructive operation on partitioned adjacency structure (Gagunashvili, 2020, Zheng et al., 2014, Zawallich et al., 2024, 0807.3979, Kannan et al., 2024).

Domain	Folded object	Unfolded output
Particle physics	Detector-level or measured distribution	Truth-level distribution, moments, or differential cross section
Materials science	Supercell electronic or phonon dispersion	Primitive-cell spectral weights in the primitive BZ
Computational geometry	Polyhedral or mesh surface	Single-patch planar net or overlap-free approximation
Graph theory	Partitioned graph construction	Cospectral graphs for the adjacency matrix
CHR	Constraint conjunction in a rule body	Inlined rule body preserving qualified answers
Molecular docking	Compact ligand conformation	Configuration maximizing molecular area or internal distances

This diversity suggests that “unfolding operation” is best treated as a family-resemblance term. What unifies the usages is not a single algebraic formalism, but a recurring objective: expose a structurally preferred representation while controlling ambiguity, overlap, or instability.

2. Statistical unfolding in particle and nuclear physics

In detector-correction problems, unfolding is posed as an inverse problem relating a true distribution $p(x)$ to a measured distribution $P(x')$ through efficiency $A(x)$ and response $R(x,x')$: $P(x') = \int_{\Omega} p(x) A(x) R(x, x') \, dx .$ A linearized histogram version writes the measured histogram as $$\mathbf{f} = \mathsf{P}\boldsymbol{\phi} + \boldsymbol{\epsilon}$$, with $\mathsf{P}$ the apparatus response matrix, $\boldsymbol{\phi}$ the true histogram, and $\boldsymbol{\epsilon}$ statistical error; least-squares recovery uses $$(\mathsf{P}'\Sigma^{-1}\mathsf{P})^{-1}\mathsf{P}'\Sigma^{-1}\mathbf{f}$$ when $P(x')$0 is identified (Gagunashvili, 2020, Gagunashvili, 2010).

The matrix itself need not be taken as given. A machine-learning-oriented system-identification approach constructs a training sample of true distributions from a priori information, propagates them through Monte Carlo simulation, estimates rows of $P(x')$1 by regression, applies stepwise regression to select relevant matrix elements, and uses a D-optimality criterion to minimize the determinant of the unfolded covariance matrix (Gagunashvili, 2010). Parametric unfolding instead assumes a model $P(x')$2, reweights simulated events with

$P(x')$3

and fits the parameters by minimizing a specialized chi-square-like statistic for weighted histograms; goodness-of-fit is assessed by an approximate $P(x')$4 law and residuals, while uncertainties are estimated by bootstrap resampling (Gagunashvili, 2020).

Recent work replaces binned inversion by unbinned, fully differential reweighting. OmniFold iteratively reweights simulation with classifier-estimated likelihood ratios, performing one update at detector level and one at particle level; it is unbinned, works for arbitrarily high-dimensional data, naturally incorporates information from the full phase space, and reduces to Iterative Bayesian Unfolding in the binned limit (Andreassen et al., 2019). IcINN uses a conditional invertible neural network to learn $P(x')$5 event-by-event and then iteratively reweights the Monte Carlo prior to mitigate data-simulation mismatch (Backes et al., 2022). Unbinned Profiled Unfolding retains detector-level binned likelihoods but uses a learned truth-level reweighting function and a detector-response reweighting function $P(x')$6 so that nuisance parameters can be profiled simultaneously during unfolding (Chan et al., 2023).

Alternative formulations explicitly avoid matrix inversion. “Unfolding by Folding” samples candidate truth-space distributions, folds each one with the response matrix $P(x')$7, and selects the generator-level distribution whose folded version is closest to the observed data distribution (Vischia, 2020). “Moment Unfolding” does not target the full spectrum at all: it directly unfolds moments as functions of another observable with a GAN-inspired reweighting function of Boltzmann form,

$P(x')$8

and is reported to be more precise than bin-based approaches and as or more precise than completely unbinned methods in the jet-substructure example studied (Desai et al., 2024). A Wiener-SVD variant has also been introduced and described as maximizing signal-to-noise ratios in an effective frequency domain while being free from a regularization parameter (Tang et al., 2017).

3. Unfolding in electronic and phononic structure theory

In materials calculations, folding arises because supercell calculations enlarge the real-space periodicity and therefore shrink the Brillouin zone. Electronic and phononic dispersions then appear crowded and folded, obscuring comparison with experiments that probe the primitive-cell Brillouin zone (Zheng et al., 2014, Zheng et al., 2016).

For electronic bands, Quantum Unfolding reconstructs Bloch wavefunctions from Wannier functions produced by Wannier90, computes their Fourier coefficients, and assigns a primitive-cell spectral weight through

$P(x')$9

Here the unfolding weight is the expectation value of a projection operator and measures how much a supercell eigenstate resembles a primitive-cell Bloch wave at $A(x)$0. In the examples reported, perfect graphene and diamond yield weights 0 or 1 and coincide with primitive-cell bands, whereas Si-doped diamond and antiferromagnetic FeSe exhibit intermediate weights, faint bands, and fragmentation associated with broken translational symmetry (Zheng et al., 2014).

For phonons, Phonon Unfolding uses a generalized projection algorithm. The central quantity is the unfolding weight

$A(x)$1

with $A(x)$2 obtained by projecting the supercell phonon polarization vector onto plane-wave basis functions (Zheng et al., 2016). The method is presented as applicable in principle to any kind of atomic system, including surface reconstructions, point defects, alloys, and glasses, and it provides an intensity map in the primitive-cell BZ suitable for comparison with experiment (Zheng et al., 2016).

The significance of these operations is representational rather than merely visual. They restore hidden translational structure when it exists and quantify its breakdown when it does not. In both electronic and phononic settings, the unfolded result is not simply a reindexed band structure; it is a weighted spectral object whose noninteger intensity records the degree to which primitive-cell periodicity survives in a supercell calculation (Zheng et al., 2014, Zheng et al., 2016).

4. Geometric unfolding of meshes, terrains, and polyhedra

In computational geometry and fabrication, unfolding usually means cutting a surface so that it can be flattened into the plane without overlap. For triangle meshes, edge unfolding cuts only along edges, and for genus-0 meshes the cut set forms a spanning tree of the edge graph while the dual unfold-tree is a spanning tree over the dual graph of faces (Zawallich et al., 2024). This classical setting is difficult because many meshes are ununfoldable as single patches or become computationally intractable at high resolution.

A recent response is to couple unfolding with approximation instead of treating approximation as fixed preprocessing. “Unfolding via Progressive Mesh Approximation” combines edge-collapse simplification with Tabu Unfolding. The coarse target face count is

$A(x)$3

where $A(x)$4 is the original face count and $A(x)$5 the genus. A simplified mesh is first unfolded, overlaps are resolved with the Tabu Unfolder, and then the mesh is progressively uncollapsed while preserving a valid single-patch layout when possible; if any stage fails, the last overlap-free approximation is returned (Zawallich et al., 2024). On 2800 manifold, single-component meshes from Thingi10k at resolutions from 100 to 2000 faces, the progressive approach—especially the Q/Q strategy using quadric error metrics for both edge selection and vertex placement—is reported as faster and more reliable than baseline Tabu Unfolding (Zawallich et al., 2024).

“Mesh Simplification For Unfolding” makes the relaxation even more explicit. It seeks a single-patch, isometric, overlap-free unfolding by minimally modifying the input mesh geometry until such an unfolding exists. The method alternates unfolding-aware vertex displacement with overlap-aware edge collapse, repairs the dual spanning tree rather than recomputing it after each collapse, and then postprocesses the result by optimizing

$A(x)$6

On processed Thingi10K-derived datasets with 500 and 1000 faces, the method is compared against baselines derived from Takahashi et al. and Zawallich, and the supplied summary reports, for the “fine” dataset, a change from 10.55% success for Takahashi’s baseline to 84.10% success when combined with the method (Bhargava et al., 2024).

For orthogonal polyhedra, the problem is often attacked by structured cutting. Every orthogonal terrain—an orthogonal polyhedron based on a rectangle that meets every vertical line in a segment—has a $A(x)$7 grid unfolding obtained by cutting along grid edges induced by coordinate planes through every vertex (0707.0610). For general orthogonal polyhedra of genus at most 2, a separate result shows that every such polyhedron can be unfolded without overlap while using only a linear number of orthogonal cuts; the construction uses an unfolding tree and in the worst case requires $A(x)$8 refinement, which remains linear (Damian et al., 2016).

“Apple peel unfolding” defines yet another geometric regime: a spiral-like sequential face selection around an axis through the centroid and the centroid of the initial face. The left-side test is

$A(x)$9

A polyhedron is classified as perfectly peelable, possibly peelable, or non-peelable depending on whether all, some, or no starting pairs produce a complete face sequence. For the 13 Archimedean and 13 Catalan solids, the reported classification gives three perfect and three possible Archimedean solids, and six perfect and three possible Catalan solids (Yoshino et al., 17 Apr 2026).

5. Algebraic, rewriting, and conformational meanings of unfolding

In graph theory, unfolding is a constructive operation used to generate cospectral graphs. Building on Butler’s bipartite construction, a generalization to non-bipartite graphs employs the partitioned tensor product

$R(x,x')$0

for block-partitioned matrices $R(x,x')$1 and $R(x,x')$2. The framework covers reflexive bipartite, semi-reflexive bipartite, and multipartite graphs, and it yields pairs of graphs that are cospectral for the adjacency matrix; nonisomorphism follows under conditions such as $R(x,x')$3 not being permutation-equivalent to $R(x,x')$4 or $R(x,x')$5 and $R(x,x')$6 being nonisomorphic (Kannan et al., 2024).

In CHR, unfolding is a program transformation rather than a geometric or statistical inversion. A conjunction of constraints in the body of one rule is replaced by the body of another rule whose head matches it, while guards, identifiers, and propagation-history tokens are updated according to a formal unfolding rule (0807.3979). The transformation is correct as a program extension: if $R(x,x')$7, then the set of qualified answers is preserved for every goal. Under U-sequences of such transformations, normal termination and normal confluence are preserved, and safe rule replacement gives conditions under which the original rule can be deleted without changing program meaning (0807.3979).

In molecular docking, Molecular Unfolding refers to expanding a ligand so as to remove initial conformational bias before docking. The objective is to maximize the sum of squared internal atomic distances,

$R(x,x')$8

over torsional configurations $R(x,x')$9. The problem is discretized by one-hot angle variables $P(x') = \int_{\Omega} p(x) A(x) R(x, x') \, dx .$0, written as a High-order Unconstrained Binary Optimization problem,

$P(x') = \int_{\Omega} p(x) A(x) R(x, x') \, dx .$1

and then reduced to QUBO for solution on D-Wave hardware (Mato et al., 2021). In the reported comparisons, all approaches reach the optimum quickly for small problem sizes, D-Wave Advantage is competitive on medium-sized instances, and classical methods dominate on larger ones (Mato et al., 2021).

These three literatures show that unfolding need not involve flattening or deconvolution. It can instead mean structured graph inflation, semantics-preserving inlining, or conformational expansion. The shared theme is again the removal of an obstructive intermediate representation.

6. Exactness, approximation, and recurring misconceptions

A persistent misconception is that unfolding always means response-matrix inversion. The literature is more heterogeneous. Some methods are genuinely inverse-problem procedures in detector physics (Gagunashvili, 2020, Andreassen et al., 2019, Chan et al., 2023); some are reciprocal-space projections (Zheng et al., 2014, Zheng et al., 2016); some are overlap-avoidance constructions on surfaces (Zawallich et al., 2024, Damian et al., 2016); some are algebraic graph constructions (Kannan et al., 2024); and some are source-to-source program transformations (0807.3979).

A second misconception is that successful unfolding always reproduces the original object exactly. Many modern methods are explicitly approximate or relaxed. Progressive mesh methods may return the last valid overlap-free approximation rather than the full-resolution input (Zawallich et al., 2024). Geometric-relaxation methods modify the shape until a simple unfolding exists (Bhargava et al., 2024). Moment Unfolding targets moments directly rather than the full distribution (Desai et al., 2024). In detector unfolding, priors, regularization, iterative reweighting, or likelihood constraints are often indispensable because the inverse problem is ill-posed or high-dimensional (Vischia, 2020, Backes et al., 2022, Chan et al., 2023).

A third misconception is that unfolding is purely representational and therefore neutral. In practice, every formulation encodes a structural preference. Quantum Unfolding prefers primitive-cell Bloch character via plane-wave Fourier components (Zheng et al., 2014); Phonon Unfolding prefers a chosen primitive reciprocal structure through generalized projections (Zheng et al., 2016); CHR unfolding preserves qualified answers but alters operational structure (0807.3979); and apple peel unfolding imposes a specific axis-driven ordering criterion that is stricter than the mere existence of a Hamiltonian path in the dual graph (Yoshino et al., 17 Apr 2026).

Taken together, these results indicate that an unfolding operation is best understood as a technically disciplined re-expression of a system under constraints specific to a field: statistical identifiability in high-energy physics, translational symmetry in materials science, non-overlap and manufacturability in geometry, spectral equivalence in graph theory, semantic preservation in CHR, and conformational expansion in molecular docking. The term is unified less by a single formula than by a recurring methodological ambition: recover an interpretable structure from a representation in which that structure has been obscured.