Systematic control of the $1/M$ loop expansion around the Bethe approximation: subleading corrections, degenerate critical spectra, and finite-$M$ spurious disorder in the M-layer construction

The Bethe M-layer construction (Altieri, Angelini, Lucibello, Parisi, Ricci-Tersenghi and Rizzo, 2017) builds, for any model on a lattice or factor graph, a one-parameter family of models indexed by $M$: $M$ copies of the graph are made and edges are rewired at random between layers, so that $M=1$ recovers the original model while $M\to\infty$ is exactly solved by the Bethe approximation. This generates a loop expansion in $1/M$ in terms of fat diagrams whose values are line-connected observables, providing the missing small parameter for finite-dimensional corrections to Bethe mean field. Between 2019 and 2026 the advertised applications were carried out at one-loop order (random-field Ising model at zero temperature; spin glass in a field at zero temperature, where the construction changed the accepted upper critical dimension from 6 to at most 8 and produced a new zero-temperature fixed point with its critical exponents; Anderson localization, with exponentially diverging one-loop corrections supporting an infinite upper critical dimension; percolation; bootstrap percolation; supercooled-liquid models), and the full one-loop computation with all numerical constants was completed for the pure Ising model. What remains open, and what this problem records, is the systematic control of the expansion itself: (1) the $1/M$ series beyond the leading divergences, through two-loop order, i.e. subleading $1/M$ corrections from multiple-occupancy projections and the fat-diagram combinatorics required at that order; (2) critical regimes in which an infinite or degenerate set of transfer-matrix eigenvalues becomes critical simultaneously, beyond the single-critical-eigenvalue case treated so far; (3) quantitative criteria for when the rewiring-induced (spurious) quenched disorder at finite $M$ changes the universality class, the associated crossover scales, and how large $M$ must be in practice; and (4) a rigorous derivation of the series, controlling the null modes that obstruct it.

Background

The Bethe approximation (equivalently, belief propagation or the cavity method) predicts the thermodynamics of a model defined on a graph by ignoring the loops of the graph and solving exactly the corresponding problem on a tree — the Bethe lattice, an infinite regular tree with fixed connectivity $c$. It is exact on trees and on locally tree-like random graphs, and uncontrolled but often accurate on finite-dimensional lattices, which contain many short loops. A traditional route to finite-dimensional corrections is perturbation around the fully connected (Curie–Weiss) mean-field model, in which every spin interacts with every other; this yields the standard Landau–Ginzburg field theory and its loop expansion. However, several important second-order phase transitions are absent or qualitatively different in the fully connected limit. The spin glass (a magnet with random, frustrated couplings) in an external field has no transition at zero temperature in the fully connected model, whereas the Bethe lattice has one at a finite critical field. Anderson localization — the suppression of diffusion of waves or quantum particles by sufficiently strong quenched (sample-fixed) disorder — is impossible by construction on a fully connected graph, but exists at finite connectivity. The random-field Ising model (RFIM), while possessing a fully connected transition, has a zero-temperature critical behavior on the Bethe lattice dominated by rare-region fluctuations and by avalanches (collective domain flips triggered by a single field step) in a way the fully connected solution misses.

Altieri, Angelini, Lucibello, Parisi, Ricci-Tersenghi and Rizzo (2017) introduced a construction designed to expand around the Bethe solution instead. Given a model on a lattice or factor graph (a bipartite-graph representation of the interactions), take $M$ replicas ("layers") of every vertex and, for each edge of the original graph, randomly rewire its $M$ copies so as to connect spins in different layers, using an independent random permutation per edge. At $M=1$ the original model is recovered; as $M\to\infty$ the rewired graph becomes locally tree-like at any fixed distance, so the Bethe approximation becomes exact. Observables then admit a perturbative series in $1/M$ — the expansion parameter that the fully connected construction lacks. The series organizes itself as a loop expansion in fat diagrams: each line of a Feynman-like diagram is a bundle of edges, and the diagram's weight is computed from non-backtracking walks (walks that never step back to the vertex they just left) on the original lattice, with infinite Bethe trees attached to internal points to restore the local connectivity. The contribution of each fat diagram is a line-connected observable of the original model, defined by Mobius inversion so that the value of any observable on a diagram equals the sum of the line-connected observables over all its sub-diagrams; this is what adds and subtracts the sub-diagram contributions with the appropriate prefactors. When the projection of a fat diagram onto the original lattice visits the same site or bond more than once (a "multiple occupancy"), the different visits must land in different layers, a constraint that is automatic as $M\to\infty$ but modifies the diagram weight by relative factors of order $1/M$; the 2017 computation of the weights keeps only the leading order and neglects precisely these subleading corrections. Near a second-order transition of the Bethe solution — where the second-largest eigenvalue of the transfer matrix $T(\sigma,\tau)$, the operator encoding correlations along a one-dimensional chain with the same local couplings, reaches the critical value $\lambda_c = 1/(c-1)$ — the $1/M$ corrections diverge at all orders. The 2017 paper showed that, keeping the leading divergences, the critical series for a generic observable has exactly the structure of a standard field-theory loop expansion, with the usual symmetry factors, so that renormalization-group (RG) tools can be transferred to expansions around the Bethe solution: the Ginzburg criterion, which locates the upper critical dimension $D_U$ above which mean-field theory is correct, and the $\epsilon$-expansion in $\epsilon = D_U - D$, which yields critical exponents in dimension $D$ below $D_U$. The 2017 paper also presents this expansion as a saddle-point (partition-function) evaluation of the series, and reports that developing the $1/M$ expansion around that saddle point is obstructed by technical difficulties, one of them the presence of null modes.

Earlier attempts at loop corrections to the Bethe approximation (Montanari and Rizzo 2005; Parisi and Slanina 2006; and the one-loop topological expansion of Angelini, Parisi and Ricci-Tersenghi 2018 in the different large-connectivity limit) either lacked a systematic small parameter tied to the Bethe solution or stopped at low order. The 2017 paper itself left the systematic control of its own $1/M$ series open in several specific ways, recorded in the problem statement below.

Between 2019 and 2026 the construction was applied at one-loop order to a series of targets: bootstrap percolation (a percolation variant in which a site becomes stable only once enough of its neighbors are already stable, so that the Bethe-lattice transition is of mixed first/second-order type; this mixed transition was shown not to survive in finite dimension); the RFIM at $T=0$; a kinetically constrained model of supercooled liquids (a lattice model whose dynamics forbids certain local moves unless neighboring sites are in the right configuration, capturing the slow structural relaxation of glass-forming liquids) on the three-dimensional M-layer lattice; the spin glass in a field at $T=0$, where the generalized Ginzburg criterion gives an upper critical dimension $D_U \le 8$, at variance with the classical value 6 obtained from finite-temperature replica field theory (the standard technique of introducing copies of a disordered system in order to average over the disorder), and where a subsequent first-order $\epsilon$-expansion around $D=8$ found a new stable zero-temperature fixed point with three independent critical exponents; Anderson localization in finite dimension, where the one-loop $1/M$ corrections diverge exponentially at criticality in any finite dimension, supporting the decades-old hypothesis of an infinite upper critical dimension; and percolation (the emergence of a system-spanning cluster of occupied sites at a threshold occupation probability, obtained as the $q\to 1$ limit of the Potts model, the standard $q$-state generalization of the Ising model), where the known $\epsilon$-expansion is recovered without any analytic continuation in the number of Potts states. For the pure ferromagnetic Ising model, the full one-loop computation with all numerical constants was completed in 2024, recovering the continuum $\phi^4$ theory and identifying its coupling in terms of the lattice spacing, the dimension, and the transfer-matrix eigenvalues. What all these works have in common — and this is the crux of the open problem — is that at each loop order they keep only the leading order in $1/M$ and expand around a single critical transfer-matrix eigenvalue.

Problem Statement

Within the Bethe M-layer construction, turn the $1/M$ loop expansion from a leading-order tool into a fully controlled systematic expansion. Concretely, for models defined on finite-dimensional lattices or factor graphs whose Bethe-approximation solution is known, the following linked objectives remain open:

1. (Subleading and higher-order terms.) Compute and control the $1/M$ series beyond the leading divergences, through two-loop order: the subleading $1/M$ corrections to the fat-diagram weights arising from projections with multiple occupancy of lines and vertices, including the boucle (French for "loop") corrections coming from diagrams in which a line does not close into a loop and which first contribute when even-degree vertices are present, the diagram combinatorics required at that order, and the two-loop corrections to the two- and four-point functions, so that the identification of the continuum field theory and of its coupling constants extends beyond one loop.

2. (Degenerate critical spectra.) Extend the critical analysis to regimes in which an infinite or degenerate set of transfer-matrix eigenvalues becomes critical simultaneously, rather than only the single-critical-eigenvalue case treated so far.

3. (Finite-$M$ validity.) Establish quantitative criteria — in the renormalization-group sense of relevant versus irrelevant perturbations — for when the rewiring-induced (spurious) quenched disorder present at finite $M$ changes the universality class relative to the $M=1$ model, determine the associated crossover scales, and thereby settle how large $M$ must be in practice for numerical studies to observe the asymptotic Bethe-dominated critical window without finite-$M$ artifacts.

4. (Rigorous derivation.) Provide a rigorous derivation of the $1/M$ series, controlling the null modes that obstruct the derivation.

Components 1–3 jointly constitute the systematic control of the expansion: component 1 supplies the higher-order series through two-loop order with its listed sub-results jointly required, component 2 its domain of validity in critical regimes, and component 3 its validity at finite $M$ with the relevance criterion, the crossover scales, and the practical $M$ threshold jointly required; component 4 is supplementary, providing an independent rigorous foundation for the series. The originally advertised model applications — the random-field Ising model at zero temperature, the spin glass in a field at zero temperature, and Anderson localization in finite dimensions — were carried out at one-loop order between 2020 and 2026 and are not part of what remains open, except insofar as their extension beyond one loop, to finite temperature, or to resummed non-mean-field critical behavior depends on components 1–3 above.

References

1. A. Altieri, M. C. Angelini, C. Lucibello, G. Parisi, F. Ricci-Tersenghi, T. Rizzo, "Loop expansion around the Bethe approximation through the M-layer construction", J. Stat. Mech. (2017) 113303. DOI: 10.1088/1742-5468/aa8c3c; arXiv:1707.08499; https://arxiv.org/abs/1707.08499
2. A. Montanari, T. Rizzo, "How to compute loop corrections to the Bethe approximation", J. Stat. Mech. (2005) P10011. DOI: 10.1088/1742-5468/2005/10/p10011; https://doi.org/10.1088/1742-5468/2005/10/p10011
3. G. Parisi, F. Slanina, "Loop expansion around the Bethe–Peierls approximation for lattice models", J. Stat. Mech. (2006) L02003. DOI: 10.1088/1742-5468/2006/02/l02003; https://doi.org/10.1088/1742-5468/2006/02/l02003
4. M. C. Angelini, G. Parisi, F. Ricci-Tersenghi, "One-loop topological expansion for spin glasses in the large connectivity limit", EPL 121, 27001 (2018). DOI: 10.1209/0295-5075/121/27001; arXiv:1708.09581; https://arxiv.org/abs/1708.09581
5. T. Rizzo, "Fate of the Hybrid Transition of Bootstrap Percolation in Physical Dimension", Phys. Rev. Lett. 122, 108301 (2019). DOI: 10.1103/PhysRevLett.122.108301; arXiv:1807.08066 (preprint titled "The fate of the bootstrap percolation hybrid critical point in finite dimension"); https://arxiv.org/abs/1807.08066
6. M. C. Angelini, C. Lucibello, G. Parisi, F. Ricci-Tersenghi, T. Rizzo, "Loop expansion around the Bethe solution for the random magnetic field Ising ferromagnets at zero temperature", Proc. Natl. Acad. Sci. USA 117, 2268-2274 (2020). DOI: 10.1073/pnas.1909872117; arXiv:1906.04437 (preprint titled "New loop expansion for the Random Magnetic Field Ising Ferromagnets at zero temperature"); https://arxiv.org/abs/1906.04437
7. T. Rizzo, T. Voigtmann, "Solvable Models of Supercooled Liquids in Three Dimensions", Phys. Rev. Lett. 124, 195501 (2020). DOI: 10.1103/PhysRevLett.124.195501; arXiv:1903.01773; https://arxiv.org/abs/1903.01773
8. G. Perrupato, M. C. Angelini, G. Parisi, F. Ricci-Tersenghi, T. Rizzo, "Ising spin glass on random graphs at zero temperature: Not all spins are glassy in the glassy phase", Phys. Rev. B 106, 174202 (2022). DOI: 10.1103/PhysRevB.106.174202; arXiv:2207.06142; https://arxiv.org/abs/2207.06142
9. M. C. Angelini, C. Lucibello, G. Parisi, G. Perrupato, F. Ricci-Tersenghi, T. Rizzo, "Unexpected Upper Critical Dimension for Spin Glass Models in a Field Predicted by the Loop Expansion around the Bethe Solution at Zero Temperature", Phys. Rev. Lett. 128, 075702 (2022). DOI: 10.1103/PhysRevLett.128.075702; arXiv:2103.17080; https://arxiv.org/abs/2103.17080
10. M. Baroni, G. Garcia Lorenzana, T. Rizzo, M. Tarzia, "Corrections to the Bethe lattice solution of Anderson localization", Phys. Rev. B 109, 174216 (2024). DOI: 10.1103/PhysRevB.109.174216; arXiv:2304.10365; https://arxiv.org/abs/2304.10365
11. M. C. Angelini, S. Palazzi, G. Parisi, T. Rizzo, "Bethe M-layer construction on the Ising model", J. Stat. Mech. (2024) 063301. DOI: 10.1088/1742-5468/ad526e; arXiv:2403.01171; https://arxiv.org/abs/2403.01171
12. M. C. Angelini, S. Palazzi, G. Parisi, T. Rizzo, "Erratum: Bethe M-layer construction on the Ising model (2024 J. Stat. Mech. 063301)", J. Stat. Mech. (2024) 129901. DOI: 10.1088/1742-5468/ad9a96; https://doi.org/10.1088/1742-5468/ad9a96
13. M. C. Angelini, S. Palazzi, T. Rizzo, M. Tarzia, "Bethe M-layer construction for the percolation problem", SciPost Phys. 18, 030 (2025). DOI: 10.21468/SciPostPhys.18.1.030; arXiv:2407.00338; https://arxiv.org/abs/2407.00338
14. M. C. Angelini, S. Palazzi, G. Parisi, T. Rizzo, "Critical exponents of the spin-glass transition in a field at zero temperature", Proc. Natl. Acad. Sci. USA 122, e2511882122 (2025). DOI: 10.1073/pnas.2511882122; arXiv:2502.21089; https://arxiv.org/abs/2502.21089
15. M. C. Angelini, S. Palazzi, G. Parisi, T. Rizzo, "Long-range spin glass in a field at zero temperature", Phys. Rev. B 114, 014410 (2026). DOI: 10.1103/kfpr-xh2l; arXiv:2602.03488; https://arxiv.org/abs/2602.03488
16. Y. Huang, P. O. Vontobel, "Graph-Cover-based Characterization of the Bethe Partition Function of Double-Edge Factor Graphs" (2025). arXiv:2506.16250; https://arxiv.org/abs/2506.16250
17. T. Leleu, S. Reifenstein, A. Yamamura, S. Ganguli, "Reshaping Global Loop Structure to Accelerate Local Optimization by Smoothing Rugged Landscapes" (2026). arXiv:2602.01490; https://arxiv.org/abs/2602.01490