Justification and validation of the replica block-saddle moment method for finite-volume corrections in one-step replica-symmetry-broken mean-field models

The replica moment/block-saddle method of Campellone, Parisi and Virasoro (J. Stat. Phys. 138, 29, 2010) computes finite-volume corrections to the quenched free energy of mean-field disordered systems whose low-temperature phase has one-step replica symmetry breaking (1RSB) with zero inter-block overlap: the integer moments $\overline{Z_N^k}$ are evaluated by summing block-decomposition saddle points, the 1RSB block size $m$ is promoted to an integration variable, and subleading corrections — including a characteristic $\tfrac{1}{2m}\ln N$ term and a Hessian determinant term $\Delta F_Q$ — are extracted from a complex contour integral. Verified only against the exactly solvable Random Energy Model, the method rests on unproven conjectures: sufficiency of the selected saddle family, uniqueness of the analytic continuation of moments growing as $\exp(Ck^2)$, legitimacy of the contour rotation, factorization of subleading terms, and the model-dependent value of $\Delta F_Q$. Later work by Derrida and Mottishaw (2015–2024) made an equivalent construction exact and ansatz-free within the random-energy model family, showing that block sizes must fluctuate with negative (indeed complex) variances, but left genuinely interacting models untouched. The open problem is to justify the method — or refute it by exhibiting a non-negligible omitted saddle or correction — and to compute and verify its predicted finite-$N$ corrections in an interacting 1RSB model such as the p-spin spherical or p-spin Ising model.

Background

Spin glasses and other disordered systems are defined by Hamiltonians containing random couplings, and their physical free energy is the quenched average — the disorder average of the logarithm of the partition function, $\overline{\ln Z_N}$ (the overline denoting averaging over the random couplings), rather than the logarithm of the average. The replica method computes this by introducing $n$ identical copies (replicas) of the system: the moments $\overline{Z_N^n}$ are well defined for integer $n$, and an analytic continuation to $n \to 0$ is taken, the order parameter being the overlap matrix $Q_{ab}$, whose elements measure the normalized similarity between the configurations of replicas $a$ and $b$ and distinguish whether two copies of the system occupy the same thermodynamic state. In mean-field (fully connected) models the resulting integral over $Q$ is evaluated for large system size $N$ by saddle points. At low temperature the replica-symmetric saddle (all off-diagonal elements equal) becomes unstable and replica symmetry breaking (RSB) occurs. In the simplest, one-step case (1RSB) the matrix $Q$ splits into blocks of size $m$: the overlap takes one value inside blocks and another between them; the parameter $m$ (eventually restricted to $0 < m < 1$) is fixed by extremizing the free energy. Models with 1RSB low-temperature phases — notably p-spin models, whose energy is a sum of random couplings acting jointly on $p$ spins — are the mean-field prototype of the random-first-order-transition (RFOT) theory of the glass transition, in which a supercooled liquid is described as freezing into an exponential number of long-lived metastable states. The Random Energy Model (REM) of Derrida, in which the energies of the $2^N$ configurations are independent Gaussian variables, is exactly solvable and is recovered as the $p \to \infty$ limit of p-spin models (Gross–Mézard).

Although the thermodynamic-limit 1RSB solution is well understood and has rigorous counterparts, the replica method's finite-$N$ (finite-volume) corrections are not under control. Nieuwenhuizen (1996) and Ferrero, Parisi and Ranieri (1996) showed that the naive 1RSB replica bookkeeping gives the wrong finite-size corrections already for the REM, and identified the missing ingredient as the fluctuations of the block-size parameter $m$. Campellone, Parisi and Virasoro (2010) proposed a repair. One computes the integer moments $\overline{Z_N^k}$ (no analytic continuation in the replica number at this stage) by summing over all saddle points of the $k$-replica integral whose matrices decompose into blocks of sizes $m_1, \dots, m_l$ summing to $k$, with $Q_{ab} = q_i$ inside block $i$ and $Q_{ab} = 0$ between blocks — a family generalizing the 1RSB structure, natural when the thermodynamic inter-block overlap $q_0$ vanishes. Inserted into the integral representation $\ln Z_N = \int_0^\infty \frac{dt}{t}(e^{-t} - e^{-t Z_N})$ and deformed into a contour integral, this sum produces a saddle point in a complex variable $x$ obeying $f'(x_{\rm sp}) = 0$ — reproducing the usual 1RSB equation for $m = x_{\rm sp}$ — plus a computable layer of subleading corrections, $N\beta F_N = N\beta F(x_{\rm sp}) + \Delta F_M + \Delta F_Q$, where $\Delta F_M$ arises from the fluctuations of the (now integration-variable) $m$ and contains a distinctive $\tfrac{1}{2m}\ln N$ term, and $\Delta F_Q = f_1(m) = \tilde{\mathrm{Tr}}(\ln(\beta\mathcal{H}))$ is the Gaussian-fluctuation determinant contribution, where $\mathcal{H}$ is the Hessian matrix of the replicated free-energy functional at the saddle point, $\beta$ is the inverse temperature, and $\tilde{\mathrm{Tr}}(A) = \lim_{n\to 0}\mathrm{Tr}(A^{(n)})/n$. In the REM, where $\Delta F_Q = 0$, the result reproduces Derrida's exact asymptotics.

The paper is explicit that this construction rests on unproven steps: the sufficiency of the chosen saddle family; the uniqueness of the analytic continuation when the moments grow as $\exp(Ck^2)$ for Gaussian disorder (fast enough to make the classical moment problem indeterminate, so the integer moments alone may not select the correct continuation); the legitimacy of the contour rotation that replaces a generally divergent sum over integers by an integral along the imaginary axis; the assumed factorization of subleading corrections into block-local pieces; and the model-dependent evaluation of $\Delta F_Q$. The paper's own suggested decisive test is a genuinely interacting model — the p-spin spherical or p-spin Ising model — where the predicted corrections, in particular the $\ln N$ term, could be checked against exact or numerical finite-$N$ results.

Problem Statement

For the class of mean-field disordered systems whose low-temperature phase is described by one-step replica symmetry breaking with vanishing inter-block overlap ($q_0 = 0$) — the p-spin spherical and p-spin Ising models being the canonical interacting examples, with the REM as the exactly solvable $p \to \infty$ limit — establish whether the replica moment/block-saddle method of Campellone, Parisi and Virasoro gives the correct finite-volume corrections to the disorder-averaged free energy, and turn it into a controlled theory. Concretely:

1. (Saddle-family sufficiency.) Prove, or refute by an explicit counterexample, that for all integer $k$ the large-$N$ asymptotics of the disorder-averaged moments $\overline{Z_N^k}$ (the overline denoting the average over the random couplings) is exhausted, up to remainders negligible uniformly in $k$, by the saddle points whose overlap matrices decompose into blocks of sizes $m_1, \dots, m_l$ with $\sum_i m_i = k$, $Q_{ab} = q_i$ inside block $i$, and $Q_{ab} = 0$ between blocks. It suffices to bound, collectively, the contribution of all other stationary configurations of the $k$-replica free energy $F(k, Q)$; a full enumeration or classification of those configurations is not required.

2. (Analytic continuation.) Settle the moment problem raised by the growth $\overline{Z_N^k} \sim \exp(Ck^2)$ for Gaussian disorder: determine whether the integer moments determine the relevant analytic continuation, and justify (or invalidate) the rotation of the contour that turns the — generally divergent — sum over integer replica numbers into the integral along the imaginary axis from which the $m$-saddle and the subleading corrections are extracted.

3. (Subleading corrections.) Compute, for at least one genuinely interacting model of this class (p-spin spherical or p-spin Ising), the finite-$N$ free energy to subleading order: the determinant/Hessian term $\Delta F_Q = f_1(m) = \tilde{\mathrm{Tr}}(\ln(\beta\mathcal{H}))$, the validity of the assumed factorization of the $O(1)$ and $O(\ln N)$ contributions into block-local pieces, and the coefficient of the predicted $\tfrac{1}{2m}\ln N$ term; and verify the result against exact or numerical finite-$N$ computations, deciding whether the block-saddle method's predictions are correct beyond the REM.

Items 1–3 are jointly required: the target is to establish items 1–3, or to refute the method by exhibiting, in a model of this class, an omitted saddle family or correction term that changes the leading or subleading finite-$N$ free energy. Invalidating the analytic continuation or the contour rotation settles item 2 only; items 1 and 3 then remain required. The extension to short-range models, and the proposed relation between $m$-fluctuations and sample-to-sample fluctuations of the critical temperature or diverging correlation lengths near $T_c$, are further directions flagged in the source paper and lie outside the scope of this problem.

References

1. M. Campellone, G. Parisi, M. A. Virasoro, 'Replica Method and Finite Volume Corrections', Journal of Statistical Physics 138, 29–39 (2010), doi:10.1007/s10955-009-9891-1, https://doi.org/10.1007/s10955-009-9891-1 (preprint arXiv:0911.1313, https://arxiv.org/abs/0911.1313)
2. B. Derrida, 'Random-energy model: An exactly solvable model of disordered systems', Physical Review B 24, 2613–2626 (1981), doi:10.1103/PhysRevB.24.2613, https://doi.org/10.1103/PhysRevB.24.2613
3. D. J. Gross, M. Mézard, 'The simplest spin glass', Nuclear Physics B 240, 431–452 (1984), doi:10.1016/0550-3213(84)90237-2, https://doi.org/10.1016/0550-3213(84)90237-2
4. Th. M. Nieuwenhuizen, 'A Puzzle on Fluctuations of Weights in Spin Glasses', Journal de Physique I France 6, 109–117 (1996), doi:10.1051/jp1:1996132, https://doi.org/10.1051/jp1:1996132
5. M. E. Ferrero, G. Parisi, P. Ranieri, 'Fluctuations in a spin-glass model with one replica symmetry breaking', Journal of Physics A: Mathematical and General 29, L569–L574 (1996), doi:10.1088/0305-4470/29/22/003, https://doi.org/10.1088/0305-4470/29/22/003
6. B. Derrida, P. Mottishaw, 'Finite size corrections in the random energy model and the replica approach', Journal of Statistical Mechanics: Theory and Experiment 2015, P01021 (2015), doi:10.1088/1742-5468/2015/01/P01021, https://doi.org/10.1088/1742-5468/2015/01/P01021 (preprint arXiv:1410.1432, https://arxiv.org/abs/1410.1432)
7. B. Derrida, P. Mottishaw, 'Finite Size Corrections to the Parisi Overlap Function in the GREM', Journal of Statistical Physics 172, 592–610 (2018), doi:10.1007/s10955-018-1953-9, https://doi.org/10.1007/s10955-018-1953-9 (preprint arXiv:1710.04611, https://arxiv.org/abs/1710.04611)
8. B. Derrida, P. Mottishaw, 'One step replica symmetry breaking and overlaps between two temperatures', Journal of Physics A: Mathematical and Theoretical 54, 045002 (2021), doi:10.1088/1751-8121/abd4ad, https://doi.org/10.1088/1751-8121/abd4ad (preprint arXiv:2009.03545, https://arxiv.org/abs/2009.03545)
9. B. Derrida, P. Mottishaw, 'The discrete random energy model and one step replica symmetry breaking', Journal of Physics A: Mathematical and Theoretical 55, 265002 (2022), doi:10.1088/1751-8121/ac71e9, https://doi.org/10.1088/1751-8121/ac71e9 (preprint arXiv:2202.12584, https://arxiv.org/abs/2202.12584)
10. B. Derrida, P. Mottishaw, 'Generalizations of Parisi's replica symmetry breaking and overlaps in random energy models', Comptes Rendus Physique 25, 329–351 (2024), doi:10.5802/crphys.199, https://doi.org/10.5802/crphys.199 (preprint arXiv:2408.15125, https://arxiv.org/abs/2408.15125)
11. U. Ferrari, C. Lucibello, F. Morone, G. Parisi, F. Ricci-Tersenghi, T. Rizzo, 'Finite size corrections to disordered systems on Erdős–Rényi random graphs', Physical Review B 88, 184201 (2013), doi:10.1103/PhysRevB.88.184201, https://doi.org/10.1103/PhysRevB.88.184201 (preprint arXiv:1305.3187, https://arxiv.org/abs/1305.3187)
12. C. Lucibello, 'Finite Size Corrections to Disordered Systems: mean field results and applications to finite dimensional models', PhD thesis, Sapienza – University of Rome (2015), arXiv:1502.02471, https://arxiv.org/abs/1502.02471
13. G. Folena, G. Biroli, P. Charbonneau, Y. Hu, F. Zamponi, 'Equilibrium fluctuations in mean-field disordered models', Physical Review E 106, 024605 (2022), doi:10.1103/PhysRevE.106.024605, https://doi.org/10.1103/PhysRevE.106.024605 (preprint arXiv:2202.07560, https://arxiv.org/abs/2202.07560)
14. G. Parisi, L. Sarra, L. Talamanca, 'Study of longitudinal fluctuations of the Sherrington–Kirkpatrick model', Journal of Statistical Mechanics: Theory and Experiment 2019, 033302 (2019), doi:10.1088/1742-5468/ab0c1b, https://doi.org/10.1088/1742-5468/ab0c1b (preprint arXiv:1812.06334, https://arxiv.org/abs/1812.06334)
15. T. Aspelmeier, A. Billoire, E. Marinari, M. A. Moore, 'Finite size corrections in the Sherrington–Kirkpatrick model', Journal of Physics A: Mathematical and Theoretical 41, 324008 (2008), doi:10.1088/1751-8113/41/32/324008, https://doi.org/10.1088/1751-8113/41/32/324008 (preprint arXiv:0711.3445, https://arxiv.org/abs/0711.3445)
16. B. F. Svaiter, N. F. Svaiter, 'The Distributional Zeta-Function in Disordered Field Theory', International Journal of Modern Physics A 31, 1650144 (2016), doi:10.1142/S0217751X1650144X, https://doi.org/10.1142/S0217751X1650144X (preprint arXiv:1603.05919, https://arxiv.org/abs/1603.05919)
17. T. Obuchi, 'Role of the Finite Replica Analysis in the Mean-Field Theory of Spin Glasses', PhD thesis, Tokyo Institute of Technology (2010), arXiv:1508.05225, https://arxiv.org/abs/1508.05225 (posted to arXiv in 2015)