Abstract
In this paper we study a multi-species disordered model on the Nishimori line. The typical properties of this line, a set of identities and inequalities among correlation functions, allow us to prove the replica symmetry i.e. the concentration of the order parameter. When the interaction structure is elliptic we rigorously compute the exact solution of the model in terms of a finite-dimensional variational principle and we study its properties.
Similar content being viewed by others
References
Agliari, E., Albanese, L., Barra, A., Ottaviani, G.: Replica symmetry breaking in neural networks: a few steps toward rigorous results. J. Phys. A Math. Theor. 53(41), 415005 (2020). https://doi.org/10.1088/1751-8121/abaf2c
Agliari, E., Barra, A., Sollich, P., Zdeborova, L.: Machine learning and statistical physics: theory, inspiration, application. J. Phys. A Math. Theor. (2020). https://doi.org/10.1088/1751-8121/abaf2c
Alberici, D., Barra, A., Contucci, P., Mingione, E.: Annealing and replica-symmetry in deep Boltzmann machines. J. Stat. Phys. 180, 1–13 (2020)
Alberici, D., Camilli, F., Contucci, P., Mingione, E.: The deep Boltzmann machine on the Nishimori line. (In preparation)
Alberici, D., Contucci, P., Mingione, E.: Deep Boltzmann machines: rigorous results at arbitrary depth. arXiv:2004.04495v1 (2020)
Barbier, J., Dia, M., Macris, N., Krzakala, F., Lesieur, T., Zdeborová, L.: Mutual information for symmetric rank-one matrix estimation: a proof of the replica formula. In: Lee, D.D., Sugiyama, M., Luxburg, U.V., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems, vol. 29, pp. 424–432. Curran Associates Inc., New York (2016)
Barbier, J., Dia, M., Macris, N., Krzakala, F., Zdeborová, L.: Rank-one matrix estimation: analysis of algorithmic and information theoretic limits by the spatial coupling method. arXiv:1812.02537 (2018)
Barbier, J., Macris, N.: The adaptive interpolation method for proving replica formulas applications to the curie-weiss and wigner spike models. J. Phys. A 52(29), 294002 (2019). https://doi.org/10.1088/1751-8121/ab2735
Barbier, J., Macris, N., Miolane, L.: The layered structure of tensor estimation and its mutual information. In: 2017 55th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 1056–1063 (2017). https://doi.org/10.1109/ALLERTON.2017.8262854
Barra, A., Contucci, P., Mingione, E., Tantari, D.: Multi-species mean field spin glasses rigorous results. Ann. Henri Poinc. (2013). https://doi.org/10.1007/s00023-014-0341-5
Barra, A., Genovese, G., Guerra, F.: Equilibrium statistical mechanics of bipartite spin systems. J. Phys. A Math. Theor. 44, 245002 (2011)
Chen W (2014) On the mixed even-spin Sherrington-Kirkpatrick model with ferromagnetic interaction. J. Ann. l’inst. Henri Poinc. B Probab. Stat. 50(1), 63–83
Contucci, P., Giardina, C.: The Ghirlanda-Guerra identities. J. Stat. Phys. (2005). https://doi.org/10.1007/978-3-642-22253-5_5
Contucci, P., Giardinà, C.: Perspectives on Spin Glasses. Cambridge University Press, Cambridge (2012)
Contucci, P., Morita, S., Nishimori, H.: Surface terms on the Nishimori line of the Gaussian Edwards-Anderson model. J. Stat. Phys. (2005). https://doi.org/10.1007/s10955-005-8020-z
Genovese, G.: Minimax formula for the replica symmetric free energy of deep restricted Boltzmann machines. arXiv:2005.09424 (2020)
Ghirlanda, S., Guerra, F.: General properties of overlap probability distributions in disordered spin systems. Towards Parisi ultrametricity. J. Phys. A Math. General 31(46), 9149–9155 (1998). https://doi.org/10.1088/0305-4470/31/46/006
Guerra, F.: Broken replica symmetry bounds in the mean field spin glass model. Commun. Math. Phys. 233(1), 1–12 (2003). https://doi.org/10.1007/s00220-002-0773-5
Guerra, F., Toninelli, F.L.: The thermodynamic limit in mean field spin glass models. Commun. Math. Phys. 230(1), 71–79 (2002). https://doi.org/10.1007/s00220-002-0699-y
Morita, S., Nishimori, H., Contucci, P.: Griffiths inequalities for the Gaussian spin glass. J. Phys. A Math. General 37(18), L203–L209 (2004). https://doi.org/10.1088/0305-4470/37/18/L03
Morita, S., Nishimori, H., Contucci, P.: Griffiths inequalities in the Nishimori line. Progress Theor. Phys. Suppl. 157, 73–76 (2005). https://doi.org/10.1143/PTPS.157.73
Nishimori, H.: Internal energy, specific heat and correlation function of the bond-random Ising model. Progress Theoret. Phys. 66(4), 1169–1181 (1981). https://doi.org/10.1143/PTP.66.1169
Nishimori, H.: Statistical Physics of Spin Glasses and Information Processing: An Introduction. Oxford University Press, Oxford (2001)
Panchenko, D.: The free energy in a multi-speciessherrington-kirkpatrick model. Ann. Probab. 43(6), 3494–3513 (2015). https://doi.org/10.1214/14-AOP967
Panchenko, D.: The Sherrington-Kirkpatrick Model. Springer Monographs in Mathematics. Springer, New York (2015)
Talagrand, M.: Mean Field Models for Spin Glasses: Volume I: Basic Examples. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics. Springer, Berlin (2010)
Acknowledgements
The authors thank Jean Barbier, Adriano Barra, Francesco Guerra and Nicolas Macris for interesting discussions. In particular we thank Jean Barbier for pointing out the inferential interpretation of our model. D.A. and E.M. acknowledge support from Progetto Alma Idea 2018, Università di Bologna.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Simone Warzel.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
A Appendix: Proof of the Concentration Lemma
Proof of Lemma 3
Let us split the proof into three steps for the sake of clarity. As anticipated, it is convenient to split the total fluctuation of \({\mathcal {L}}_r\) into two parts, thus proving that:
From this moment on, we neglect sub and superscripts in the Gibbs brackets as well as t-dependencies. We start by proving the inequality (45).
Proof of inequality (45): To begin with, we compute:
where integration by parts has been used.
Then, we proceed with:
We treat the three terms \(R_1\), \(R_2\) and \(R_3\) separately with repeated integrations by parts.
where we have used the Nishimori identity (16).
Hence:
Summing up all the contributions:
Proof of (A.1): Notice that:
From the last one, after an integration by parts and using the regularity of the map \(\epsilon \longmapsto {\mathbf {Q}}_\epsilon (\cdot )\) and Lemma 2 we get:
where:
Proof of (A.2): Let \(p_{N,\epsilon }\) be the random interpolating pressure, such that \({\mathbb {E}}p_{N,\epsilon }={\bar{p}}_{N,\epsilon }\). Define:
Let us evaluate:
where:
Thanks to the independence of the \(J_i^r\) it is immediate to verify that \(\exists \, a\ge 0\) s.t.:
Using Lemma 3.2 in [25], with notations used in [8]:
with:
where for simplicity we have kept the dependence on \(Q_{\epsilon ,r}\) only, \(\hat{{\bar{p}}}'_{N,\epsilon }\) is the derivative w.r.t. it and \(\delta >0\). Notice that \(\delta \) will be chosen strictly smaller than \(s_N\), so that \(Q_{\epsilon ,r}-\delta \ge \epsilon -\delta \ge s_N-\delta >0\).
Then, using (A.15), (A.16) and (A.18), and thanks to the fact that \((\sum _{i=1}^p\nu _i)^2\le p\sum _{i=1}^p\nu _i^2\), we get:
We first evaluate the two terms containing \(C_\delta ^\pm \):
Taking the expectation \({\mathbb {E}}_{\varvec{\epsilon }}{\mathbb {E}}\) in (A.22), and defining \(W_r\) s.t. \(Q_{\epsilon ,r}\le W_r\), we get:
We can make the r.h.s. vanish by choosing for example: \(\delta =s_N^{2K/3}N^{-1/3}\). The choice \(s_N\propto N^{-1/16K}\) makes the r.h.s. (A.24) behave like \({\mathcal {O}}(N^{-1/4})\). \(\square \)
B Appendix: the SK Case
In the case \(K=1\) the equation (46) reduces to:
while (48) simply becomes:
We collect the main results on this model in the following proposition.
Proposition 4
Define:
The following hold:
-
1.
if \(\mu <1\) then \({\bar{p}}_{var}\) is concave in x. Equivalently if \(\mu <1\) then \(T(x;\mu ,h)\) is a contraction, and if further \(h=0\) then \(x=0\) is its fixed point;
-
2.
the stable solution of the consistency equation (B.2) is continuous at \((\mu ,h)=(1,0)\):
$$\begin{aligned} \lim _{(\mu ,h)\rightarrow (1,0)}{\bar{x}}(\mu ,h)=0={\bar{x}}(1,0)\;; \end{aligned}$$(B.4) -
3.
for fixed \(h=0\), the magnetization goes to 0 linearly with \(\mu -1\) as \(\mu \rightarrow 1_+\), more precisely:
$$\begin{aligned} {\bar{x}}=(1+o(1))\frac{\mu -1}{\mu ^2} \end{aligned}$$(B.5)where o(1) goes to 0 when \(\mu \rightarrow 1_+\). Therefore the critical exponent \(\beta \) (in the Landau classification) is 1, which means that the derivative of the magnetization w.r.t. \(\mu \) does not diverge at the critical point, it only jumps from 0 to 1 and then decreases;
-
4.
Along the line \((\mu ,\lambda (\mu -1)),\;\lambda >0\) in the plane \((\mu ,h)\) the magnetization goes to 0 as follows:
$$\begin{aligned} {\bar{x}}=\sqrt{\frac{\lambda (\mu -1)}{\mu ^2}}(1+o(1))\; \end{aligned}$$(B.6)when \(\mu \rightarrow 1_+\), therefore with a critical exponent 1/2;
-
5.
For fixed \(\mu =1\) and \(h\rightarrow 0_+\) the magnetization behaves as:
$$\begin{aligned} {\bar{x}}^2=h(1+o(1)) \end{aligned}$$(B.7)where \(o(1)\rightarrow 0\) when \(h\rightarrow 0_+\). Therefore we have a critical exponent \(\delta =2\) (according to Landau’s classification).
Proof
1. The fist assertion follows immediately from (56), since \({\hat{\alpha }}\equiv 1\) and \(\varDelta \equiv \mu \). Then, by (51):
that implies T is a contraction. It is easy to see that if \(h=0\) then \(x=0\) is a solution of the fixed point equation which must be unique by Banach’s fixed point theorem.
2. Using continuity and monotonicity of T (see (51)):
hence both \(\limsup _{(\mu ,h)\rightarrow (1,0)}{\bar{x}}(\mu ,h)\) and \(\liminf _{(\mu ,h)\rightarrow (1,0)}{\bar{x}}(\mu ,h)\) satisfy the consistency equation:
whose solution \(m=0\) is unique, since the derivative of T(m; 1, 0) is \(\le 1\) and equality holds only at \(m=0\). We conclude that there exists
3. We first notice that
which simply follows from the third relation in (51) and the identity (14). Indeed, the quantity in (B.9) is nothing but the quenched average magnetization of a free system. Now, by computing the first and second derivatives of the map \(T(x;\mu ,0)\) and using (B.9) we get:
which implies that, in proximity of \(\mu =1\), the magnetization goes to 0 with a critical exponent \(\beta =1\) (not to be confused with inverse absolute temperature) and with slope 1.
4. An analogous expansion of T yields:
which in turn entails:
5. Here by o(1) we mean a quantity that approaches 0 as \(h\rightarrow 0_+\). As in the previous steps:
then we get:
\(\square \)
Rights and permissions
About this article
Cite this article
Alberici, D., Camilli, F., Contucci, P. et al. The Multi-species Mean-Field Spin-Glass on the Nishimori Line. J Stat Phys 182, 2 (2021). https://doi.org/10.1007/s10955-020-02684-z
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s10955-020-02684-z