Stability, ultrametricity, and TAP verification of the multilevel constrained-replica cluster counting in the spherical p-spin glass

Baviera and Virasoro (J. Stat. Mech. 2015) introduced a constrained-real-replica method that, for the spherical p-spin glass analyzed at two replica-symmetry-breaking levels, yields analytic multiplicity formulae for states and clusters organized in a multi-level ultrametric tree of the energy landscape. The physical validation of this construction was left open: the linear stability (fluctuation spectrum) of the constrained two-level saddle points is uncomputed, the conjecture that fixing diagonal overlaps at their saddle-point values forces ultrametricity of all triplets is unproved, the multiplicity formulae do not separate stable from unstable stationary points, no TAP or cavity computation cross-checks the predicted cluster multiplicities, and extensions to mixtures of different p-spin interactions and to algorithmic exploitation are unexplored. No resolution of these questions in the constrained-saddle formulation is known in the literature to date.

Background

Spin glasses are disordered systems whose energy hypersurface over configuration space is extremely rugged, with exponentially many local minima and saddle points. A solvable mean-field example is the spherical p-spin glass: $N$ continuous spins confined to the sphere (sum of squared components equal to $N$) interact through frozen random couplings in a Hamiltonian that is a random degree-$p$ polynomial on that sphere; for $p \ge 3$ its low-temperature phase exhibits one-step replica symmetry breaking, and its thermodynamics was solved exactly by Crisanti and Sommers. In the replica approach one computes disorder-averaged free energies by introducing $n$ virtual copies ("replicas") of the system and analytically continuing $n$ to zero; below the glass transition the assumption that all replicas are equivalent becomes unstable and is replaced by a hierarchical scheme (replica symmetry breaking), whose order parameter is the Parisi function $q(x)$, a function on $[0,1]$ encoding the distribution of the overlap $q$ between two independent configurations (their normalized scalar product). The free-energy cost of violating ultrametricity has itself been computed (Franz, Parisi and Virasoro, 1993). A foundational result of the theory is ultrametricity: overlaps between any three pure states form isosceles triangles, so states sit at the leaves of a tree; this was proved rigorously for mixed p-spin models in the Ising setting. Beyond equilibrium thermodynamics one wants the statistics of the metastable states and saddles themselves: how many states of a given free energy lie inside a cluster, and how clusters nest. Kurchan, Parisi and Virasoro showed that such multiplicities can be computed as saddle points of a partition function of several "real" replicas (independent copies sharing the same disorder) with constrained overlaps, and the TAP (Thouless-Anderson-Palmer) analysis of Crisanti and Sommers gives the matching zero-temperature count of TAP solutions; the TAP equations are mean-field self-consistency equations whose solutions are the candidate metastable states, and the cavity method, which reconstructs the measure site by site, is the standard alternative to replicas. Baviera and Virasoro (2015) pushed this program to a two-level hierarchy: they study the free energy $\log Z(\beta, R_1, R_2)$ of $R_1$ groups of $R_2$ real replicas each, with configurations inside a group constrained to mutual overlap $q_2$ and configurations in different groups to overlap $q_1 < q_2$, both fixed at the saddle point of $\log Z$; in the replica calculation the virtual replicas carry a two-level Parisi block structure with block sizes $m_1$ and $m_2$ ($m_2 > m_1$), and an ansatz of three Parisi-type functions $q(x)$, $p_1(x)$, $p_2(x)$ encodes the constrained overlap distributions. They derive analytic formulae for the multiplicity of states inside a cluster, $\log N_{S \in V}$, and of clusters, $\log N_V$, as functions of labels $y_1$ and $Y$ that act as cluster- and state-level free-energy variables. They find that the decreasing branch of the Parisi function $q(x)$ codes a hidden multi-level ultrametric tree, reminiscent of Derrida's generalized random energy model (GREM), a solvable model whose random energies are correlated along an explicitly hierarchical tree. The authors are explicit that the construction rests on unverified assumptions: the ultrametricity of the constrained ensembles is conjectured, having been checked only on examples; the stability of the constrained saddle points was not analyzed; and the counting does not distinguish physically stable clusters from unstable ones. The Plefka criterion, a standard admissibility condition requiring the Hessian of the TAP free energy to remain non-negative so that a TAP solution is a genuine local minimum, is the natural tool on the TAP side.

Problem Statement

Establish the physical validity, and refine, the multilevel constrained-replica description of the ultrametric organization of the spherical p-spin glass energy landscape proposed by Baviera and Virasoro. The components below are linked by this single objective - deciding whether the constrained-replica cluster counting describes real, stable, hierarchically organized states. Components (1)-(4) form the mandatory validation core and name four distinct, non-overlapping deliverables - the fluctuation spectrum together with the location of the stability boundaries (1), whether the overlap constraints alone force ultrametricity (2), the multiplicity formulae restricted to the stable domain (3), and whether the replica counts match the TAP/cavity analysis (4); the validation is complete only once all four are resolved. Components (5)-(6) are open extensions and applications that lie outside this core, may remain unresolved, and are not required for completion: (1) Stability analysis: compute the full linear stability of the constrained two-level saddle-point solutions of $\log Z(\beta, R_1, R_2)$ in the $(q_1, q_2, m_1, m_2, p_1(x), p_2(x))$ ansatz - that is, the fluctuation spectrum, the replicon-type eigenvalues of the Hessian of the constrained replica free energy, the sector associated with fluctuations of the overlap order parameter; determine which constrained solutions correspond to physically stable clusters and states, and locate the stability boundaries (including Plefka-type criteria on the TAP side) that delimit the stable domain of the cluster and state labels. (2) Ultrametricity of the constrained ensembles: prove or disprove the conjecture that fixing the diagonal overlaps at their saddle-point values forces all triplets of configurations in the constrained ensembles to obey ultrametricity, so far verified by the authors only on examples; a TAP or cavity analysis of the constrained clusters is the suggested route. (3) Stable versus unstable clusters: obtain the closed-form refined multiplicity formulae $\log N_V$ and $\log N_{S \in V}$ restricted to the stable domain located in (1), so that the counts refer only to physically stable states and clusters. (4) TAP/cavity cross-check: connect the constrained-replica multiplicities and predicted thresholds to the multiplicity and organization of solutions of the TAP equations via cavity or direct TAP analysis, testing in particular the match between the replica-based counts and the Crisanti-Sommers TAP counts and the Plefka bound $y_1 \le 1/(p-1)$ on the cluster label $y_1$ (announced in the source's footnote as an unproved consequence of the Plefka criterion, attributed there to work in progress by Parisi, Ricci-Tersenghi and Virasoro). (5) Generalizations: apply the multilevel constrained-replica method to mixtures of different p-spin interactions and explore alternative functional forms of the order-parameter functions $q(x)$, $p_1(x)$, $p_2(x)$ beyond the stepwise ansatz. (6) Algorithmic consequences: determine whether or how the revealed multilevel structure can be exploited to design efficient hierarchical search algorithms for the ground state, given that the number of clusters at the relevant label $y_1$ is exponentially large and no local signal is known that identifies the clusters containing the lowest states.

References

1. R. Baviera and M. A. Virasoro, "A method that reveals the multi-level ultrametric tree hidden in p-spin glass like systems", Journal of Statistical Mechanics: Theory and Experiment 2015, P12007 (2015). DOI: 10.1088/1742-5468/2015/12/P12007; arXiv:1508.01232; https://arxiv.org/abs/1508.01232
2. A. Crisanti and H.-J. Sommers, "The spherical p-spin interaction spin glass model: the statics", Zeitschrift fuer Physik B Condensed Matter 87, 341-354 (1992). DOI: 10.1007/BF01309287; https://doi.org/10.1007/BF01309287
3. A. Crisanti and H.-J. Sommers, "Thouless-Anderson-Palmer approach to the spherical p-spin spin glass model", Journal de Physique I France 5, 805-813 (1995). DOI: 10.1051/jp1:1995164; https://doi.org/10.1051/jp1:1995164
4. J. Kurchan, G. Parisi, and M. A. Virasoro, "Barriers and metastable states as saddle points in the replica approach", Journal de Physique I France 3, 1819-1838 (1993). DOI: 10.1051/jp1:1993217; https://doi.org/10.1051/jp1:1993217
5. S. Franz, G. Parisi, and M. A. Virasoro, "Free-energy cost for ultrametricity violations in spin glasses", Europhysics Letters 22, 405-411 (1993). DOI: 10.1209/0295-5075/22/6/002; https://doi.org/10.1209/0295-5075/22/6/002
6. M. E. Ferrero and M. A. Virasoro, "Saddle points stability in the replica approach off equilibrium", Journal de Physique I France 4, 1819-1827 (1994). DOI: 10.1051/jp1:1994109; https://doi.org/10.1051/jp1:1994109
7. B. Derrida, "A generalization of the random energy model which includes correlations between energies", Journal de Physique Lettres 46, 401-407 (1985). DOI: 10.1051/jphyslet:01985004609040100; https://doi.org/10.1051/jphyslet:01985004609040100
8. D. Panchenko and M. Talagrand, "On the overlap in the multiple spherical SK models", The Annals of Probability 35 (2007). DOI: 10.1214/009117907000000015; arXiv:math/0604082; https://arxiv.org/abs/math/0604082
9. A. Auffinger, G. Ben Arous, and J. Černý, "Random matrices and complexity of spin glasses", Communications on Pure and Applied Mathematics 66, 165-201 (2013). DOI: 10.1002/cpa.21422; arXiv:1003.1129; https://arxiv.org/abs/1003.1129
10. E. Subag, "The complexity of spherical p-spin models - a second moment approach", The Annals of Probability 45, 3385-3450 (2017). DOI: 10.1214/16-aop1139; arXiv:1504.02251; https://arxiv.org/abs/1504.02251
11. E. Subag, "The geometry of the Gibbs measure of pure spherical spin glasses", Inventiones mathematicae 210, 135-209 (2017). DOI: 10.1007/s00222-017-0726-4; arXiv:1604.00679; https://arxiv.org/abs/1604.00679
12. J. Ko, "Free energy of multiple systems of spherical spin glasses with constrained overlaps", Electronic Journal of Probability 25 (2020). DOI: 10.1214/20-ejp431; arXiv:1806.09772; https://arxiv.org/abs/1806.09772
13. D. Belius and N. Kistler, "The TAP-Plefka variational principle for the spherical SK model", Communications in Mathematical Physics 367, 991-1017 (2019). DOI: 10.1007/s00220-019-03304-y; https://doi.org/10.1007/s00220-019-03304-y
14. D. Belius, L. Fröber, and J. Ko, "TAP variational principle for the constrained overlap multiple spherical Sherrington-Kirkpatrick model" (2023). arXiv:2304.04031; https://arxiv.org/abs/2304.04031
15. B. Huang and M. Sellke, "A constructive proof of the spherical Parisi formula" (2023). arXiv:2311.15495; https://arxiv.org/abs/2311.15495
16. A. Jagannath and I. Tobasco, "Bounds on the complexity of replica symmetry breaking for spherical spin glasses", Proceedings of the American Mathematical Society 146, 3127-3142 (2018). DOI: 10.1090/proc/13875; arXiv:1607.02134; https://arxiv.org/abs/1607.02134
17. A. Auffinger and Q. Zeng, "Existence of two-step replica symmetry breaking for the spherical mixed p-spin glass at zero temperature", Communications in Mathematical Physics 360, 377-402 (2018). DOI: 10.1007/s00220-018-3252-3; arXiv:1802.02991; https://arxiv.org/abs/1802.02991
18. E. Subag, "Following the ground states of full-RSB spherical spin glasses", Communications on Pure and Applied Mathematics 74, 1021-1044 (2021). DOI: 10.1002/cpa.21922; arXiv:1812.04588; https://arxiv.org/abs/1812.04588
19. B. Huang and M. Sellke, "Tight Lipschitz hardness for optimizing mean field spin glasses", Communications on Pure and Applied Mathematics 78, 60-119 (2024). DOI: 10.1002/cpa.22222; arXiv:2110.07847; https://arxiv.org/abs/2110.07847
20. T. Kim, "Heavy-tailed mixed p-spin spherical model: breakdown of ultrametricity and failure of the Parisi formula" (2025). arXiv:2506.23987; https://arxiv.org/abs/2506.23987
21. J. Yeo and M. A. Moore, "Possible instability of one-step replica symmetry breaking in p-spin Ising models outside mean-field theory", Physical Review E 101, 032127 (2020). DOI: 10.1103/PhysRevE.101.032127; arXiv:1911.02719; https://arxiv.org/abs/1911.02719
22. S. Franz, G. Parisi, M. Sevelev, P. Urbani, and F. Zamponi, "Universality of the SAT-UNSAT (jamming) threshold in non-convex continuous constraint satisfaction problems", SciPost Physics 2, 019 (2017). DOI: 10.21468/SciPostPhys.2.3.019; arXiv:1702.06919; https://arxiv.org/abs/1702.06919