Ultrametricity of the three-dimensional Edwards-Anderson Ising spin glass: thermodynamic-limit behavior of three-replica overlap observables

Whether the low-temperature phase of the three-dimensional Edwards-Anderson (EA) Ising spin glass obeys Parisi's replica-symmetry-breaking (RSB) theory down to its state-space geometry — in particular whether the three mutual overlaps of independently sampled equilibrium configurations are organized ultrametrically in the thermodynamic limit — is a central open question of spin-glass physics. The 1996 study that introduced three-replica simulations of the 3D EA model with Gaussian couplings measured two families of three-overlap Binder cumulants, including the ultrametricity-violation cumulant $B_{q-q}$, and the distribution of the violation variable $b=(|q_{12}|-|q_{13}|)^2/q_{23}^2$, finding that violations decrease only very slowly with lattice size (a fitted power-law exponent of 0.13(2) at $T=0.8$) and concluding that sizes $L\le10$ cannot distinguish genuine lack of ultrametricity from finite-size violations. This problem asks for a determination of the thermodynamic-limit behavior: do $B_{q-q}$, its signed variant, and the tail of the $b$ distribution vanish as $L\to\infty$ below $T_c$, collapsing onto the mean-field ultrametric form of the three-overlap distribution, or do violations persist, contradicting the ultrametric overlap structure that both the mean-field and the trivial two-state (droplet) pictures require? Later work — simulations up to $L=32$, a rigorous proof of ultrametricity for the mean-field Sherrington-Kirkpatrick model, and conditional finite-dimensional results — has not settled this: the numerical evidence conflicts, and a 2022 review by leading practitioners states that direct detection of ultrametricity in 3D spin glasses remains elusive.

Background

A spin glass is a magnet whose interactions are randomly ferromagnetic or antiferromagnetic, so that no arrangement of the spins can satisfy all of them simultaneously; the resulting low-temperature phase is 'glassy', with extremely slow dynamics and a debated equilibrium structure. The standard lattice model is the Edwards-Anderson (EA) Ising spin glass with Hamiltonian $H=-\sum_{\langle i,j\rangle}J_{ij}\sigma_i\sigma_j$ on a cubic lattice of linear size $L$, where the quenched (frozen) nearest-neighbor couplings $J_{ij}$ are independent Gaussian random variables of zero mean. In three dimensions this model undergoes a spin-glass phase transition at $T_c\approx1.0$ in lattice units (the originating study estimated $T_c\approx1.02(5)$; the value $T_c\approx1.1$ often quoted in the recent simulation literature refers to the bimodal $\pm J$ variant of the couplings).

The order parameter is the overlap between two independent equilibrium replicas (copies) of the system evolving with the same disorder, $q=(1/L^3)\sum_i\sigma_i\tau_i$, whose probability density $P(q)$ differs sharply between the two competing theories of the spin-glass phase. In the droplet picture the low-temperature phase contains only a single pair of pure states related by global spin reversal, so in the thermodynamic limit $P(q)$ collapses onto two delta functions at $\pm q_{\rm EA}$. In Parisi's replica-symmetry-breaking (RSB) theory, imported from the exactly solved infinite-range Sherrington-Kirkpatrick (SK) model, there are infinitely many pure states with a nontrivial $P(q)$ carrying weight at all overlaps between $0$ and $q_{\rm EA}$. In RSB theory these states are moreover organized hierarchically: using the overlap to define a distance between states, the resulting space is ultrametric, meaning every triplet of states satisfies $d(A,B)\le\max[d(A,C),d(B,C)]$; equivalently, the three mutual overlaps of three independently sampled equilibrium configurations always form an equilateral or isosceles triangle, with the two smaller overlaps equal. For the SK model this ultrametricity is now a theorem (Panchenko 2013), derived from the Ghirlanda-Guerra identities among replica moments, which follow from stochastic stability — the invariance of the free-energy statistics under weak random perturbations of the couplings. In finite dimensions no comparable theorem exists; whether the nontrivial, hierarchical ultrametric organization of many pure states holds in the three-dimensional EA model, rather than the trivial two-state overlap structure of droplet theory, is one of the central open questions of spin-glass physics.

A Binder cumulant is a dimensionless ratio of moments of an order-parameter distribution whose size dependence locates phase transitions, since curves for different $L$ cross near $T_c$. Iñiguez, Parisi and Ruiz-Lorenzo (1996) simulated the 3D EA model with Gaussian couplings — using simulated tempering, a Monte Carlo scheme that lets the temperature itself fluctuate in order to accelerate equilibration, with $L\le10$ and $T\in[0.7,1.3]$ (plus an $L=4$ run down to $T=0.3$) — with three replicas per disorder sample. Three replicas yield three overlaps $q_{12},q_{13},q_{23}$ and hence three-overlap observables that cannot be constructed from two replicas. They studied $B_{qqq}=\langle|q_{12}q_{13}q_{23}|\rangle/\langle q^2\rangle^{3/2}$, whose ultrametricity-predicted value can be computed from $P(q)$ alone, and, crucially, the violation observables

$$B_{q-q}=\frac{\langle(|q_{12}|-|q_{13}|)^2\rangle}{\langle q_{23}^2\rangle},\qquad B'_{q-q}=\frac{\langle(q_{12}-q_{13}\,{\rm sign}\,q_{23})^2\rangle}{\langle q_{23}^2\rangle},$$

where $q_{23}$ is the largest of the three overlaps in absolute value: these vanish identically if the three overlaps satisfy ultrametricity exactly (isosceles triangles with the two smaller overlaps equal). The per-configuration violation variable is $b=(|q_{12}|-|q_{13}|)^2/q_{23}^2$. The authors found that $B_{qqq}$ approaches its ultrametricity-predicted value at low temperature but that this trend is not discriminating — nearly independent overlaps subject only to the combinatorial constraint that one of them is the largest would produce the same trend toward 1 — while the crossing behavior of $B_{q-q}$ yields a critical-temperature estimate $T_c\approx1.02(5)$ and, at fixed $T$, the violation observables decrease only very slowly with $L$: at $T=0.8$ the tail probability $\Pr(b>b_0)$ at $b_0=0.05$ falls with $L$ approximately as a power law with a small exponent $0.13(2)$, the small-$b_0$ region follows $b_0^{-0.71}$, and an exponential plateau sits in between. They concluded that these results were not conclusive and that substantially larger lattices, equilibrated with new thermalization methods such as parallel tempering (simulating many temperatures in parallel with swap moves), would be needed to distinguish genuine lack of ultrametricity from finite-size violations.

Problem Statement

Determine whether the equilibrium three-replica overlap structure of the three-dimensional Edwards-Anderson Ising spin glass with nearest-neighbor Gaussian couplings is ultrametric in the infinite-volume low-temperature phase, as replica-symmetry-breaking theory predicts, or whether violations persist, which would contradict the ultrametric structure common to both the mean-field many-state picture and the trivial two-state (droplet) picture. Concretely, with three independently equilibrated replicas of the same disorder having mutual overlaps $q_{12},q_{13},q_{23}$, of which $q_{23}$ is the largest in absolute value:

1. Do the ultrametricity-violation Binder cumulants $B_{q-q}=\langle(|q_{12}|-|q_{13}|)^2\rangle/\langle q_{23}^2\rangle$ and $B'_{q-q}=\langle(q_{12}-q_{13}\,{\rm sign}\,q_{23})^2\rangle/\langle q_{23}^2\rangle$ tend to zero as $L\to\infty$ at fixed $T<T_c$, with what finite-size scaling form (including corrections to scaling), or do they saturate at a nonzero limit?

2. Does the distribution of the violation variable $b=(|q_{12}|-|q_{13}|)^2/q_{23}^2$ collapse onto $b=0$ in the thermodynamic limit — equivalently, does $\Pr(b>b_0)\to0$ for every fixed $b_0>0$ below $T_c$ (the 1996 data at $T=0.8$ decayed with $L$ only as a power law with exponent $0.13(2)$) — and what is the asymptotic form of this distribution in its small-$b_0$ and intermediate-$b_0$ regimes?

3. More generally, does the joint three-overlap probability density $P_3(q,q',q'')$ converge to the mean-field ultrametric form, a mixture of equilateral and isosceles contributions fixed by $P(q)$ alone (eq. (8) of the 1996 paper), including the vanishing of the negative-weight region of the normalized triple product $b'_{qqq}=q_{12}q_{13}q_{23}/[(q_{12}^2+q_{13}^2+q_{23}^2)/3]^{3/2}$, which must go to zero if mean field holds?

Genuine ultrametric organization must be distinguished from the combinatorial null hypothesis of nearly independent overlaps, which the originating study showed can mimic low-temperature ultrametric trends. The three items are facets of one objective — the thermodynamic-limit ultrametricity, or lack thereof, of the three-dimensional spin-glass phase.

References

1. D. Iñiguez, G. Parisi, J. J. Ruiz-Lorenzo, 'Simulation of three-dimensional Ising spin glass model using three replicas: study of Binder cumulants', Journal of Physics A: Mathematical and General 29, 4337-4345 (1996), DOI 10.1080/0305-4470/29/15/009, https://arxiv.org/abs/cond-mat/9603083
2. A. Cacciuto, E. Marinari, G. Parisi, 'A numerical study of ultrametricity in finite-dimensional spin glasses', Journal of Physics A: Mathematical and General 30, L263-L269 (1997), DOI 10.1088/0305-4470/30/9/002, https://arxiv.org/abs/cond-mat/9608161
3. Silvio Franz, Federico Ricci-Tersenghi, 'Ultrametricity in three-dimensional Edwards-Anderson spin glasses', Physical Review E 61, 1121-1124 (2000), DOI 10.1103/PhysRevE.61.1121, https://arxiv.org/abs/cond-mat/9903130
4. G. Parisi, F. Ricci-Tersenghi, 'On the origin of ultrametricity', Journal of Physics A: Mathematical and General 33, 113-129 (1999), DOI 10.1088/0305-4470/33/1/307, https://doi.org/10.1088/0305-4470/33/1/307
5. G. Hed, A. P. Young, E. Domany, 'Lack of Ultrametricity in the Low-Temperature Phase of Three-Dimensional Ising Spin Glasses', Physical Review Letters 92, 157201 (2004), DOI 10.1103/PhysRevLett.92.157201, https://arxiv.org/abs/cond-mat/0310609
6. Pierluigi Contucci, Cristian Giardinà, Claudio Giberti, Giorgio Parisi, Cecilia Vernia, 'Ultrametricity in the Edwards-Anderson Model', Physical Review Letters 99, 057206 (2007), DOI 10.1103/PhysRevLett.99.057206, https://arxiv.org/abs/cond-mat/0607376
7. Thomas Jörg, Florent Krząkała, 'Comment on "Ultrametricity in the Edwards-Anderson Model"', Physical Review Letters 100, 159701 (2008), DOI 10.1103/PhysRevLett.100.159701, https://arxiv.org/abs/0709.0894
8. Pierluigi Contucci, Cristian Giardinà, Claudio Giberti, Giorgio Parisi, Cecilia Vernia, 'Answer to Comment on "Ultrametricity in the Edwards-Anderson Model" arXiv:0709.0894' (registered arXiv title of the unpublished reply), arXiv:0712.1431 (2007), https://arxiv.org/abs/0712.1431
9. H. Katzgraber, A. Hartmann, 'Ultrametricity and Clustering of States in Spin Glasses: A One-Dimensional View', Physical Review Letters 102, 037207 (2009), DOI 10.1103/PhysRevLett.102.037207, https://arxiv.org/abs/0807.3513
10. Helmut G. Katzgraber, Thomas Jörg, Florent Krząkała, Alexander K. Hartmann, 'Ultrametric probe of the spin-glass state in a field', Physical Review B 86, 184405 (2012), DOI 10.1103/PhysRevB.86.184405, https://arxiv.org/abs/1205.4200
11. R. Alvarez Baños, A. Cruz, L. A. Fernandez, J. M. Gil-Narvion, A. Gordillo-Guerrero, M. Guidetti, A. Maiorano, F. Mantovani, E. Marinari, V. Martín-Mayor, J. Monforte-Garcia, A. Muñoz Sudupe, D. Navarro, G. Parisi, S. Perez-Gaviro, J. J. Ruiz-Lorenzo, S. F. Schifano, B. Seoane, A. Tarancón, R. Tripiccione, D. Yllanes (The Janus Collaboration), 'Nature of the spin-glass phase at experimental length scales', Journal of Statistical Mechanics: Theory and Experiment P06026 (2010), DOI 10.1088/1742-5468/2010/06/P06026, https://arxiv.org/abs/1003.2569
12. The Janus Collaboration: R. A. Baños, A. Cruz, L. A. Fernandez, J. M. Gil-Narvion, A. Gordillo-Guerrero, M. Guidetti, D. Iñiguez, A. Maiorano, F. Mantovani, E. Marinari, V. Martin-Mayor, J. Monforte-Garcia, A. Muñoz Sudupe, D. Navarro, G. Parisi, S. Perez-Gaviro, F. Ricci-Tersenghi, J. J. Ruiz-Lorenzo, S. F. Schifano, B. Seoane, A. Tarancón, R. Tripiccione, D. Yllanes, 'Sample-to-sample fluctuations of the overlap distributions in the three-dimensional Edwards-Anderson spin glass', Physical Review B 84, 174209 (2011), DOI 10.1103/PhysRevB.84.174209, https://arxiv.org/abs/1107.5772
13. A. Maiorano, G. Parisi, D. Yllanes, 'Ultrametricity and long-range correlations in the Edwards-Anderson spin glass', arXiv:1312.2790 (2013), https://arxiv.org/abs/1312.2790
14. Dmitry Panchenko, 'The Parisi ultrametricity conjecture', Annals of Mathematics (2) 177, 383-393 (2013), DOI 10.4007/annals.2013.177.1.8, https://arxiv.org/abs/1112.1003
15. Juan J. Ruiz-Lorenzo, 'Nature of the Spin Glass Phase in Finite Dimensional (Ising) Spin Glasses', in Order, Disorder and Criticality: Advanced Problems of Phase Transition Theory (ed. Yu. Holovatch; World Scientific, 2020), pp. 1-52, DOI 10.1142/9789811216220_0001, https://arxiv.org/abs/2006.12930
16. V. Martín-Mayor, J. J. Ruiz-Lorenzo, B. Seoane, A. P. Young, 'Numerical Simulations and Replica Symmetry Breaking', arXiv:2205.14089 (2022), https://arxiv.org/abs/2205.14089 (contribution to the edited volume 'Spin Glass Theory and Far Beyond', World Scientific)