Abstract
We analyze the statics for pure p-spin spherical spin glass models with \(p\ge 3\), at low enough temperature. With \(F_{N,\beta }\) denoting the free energy, we compute the (logarithmic) second order term of \( NF _{N,\beta }\) and prove that, for an appropriate centering \(c_{N,\beta },\, NF _{N,\beta }-c_{N,\beta }\) is a tight sequence. We further establish the absence of temperature chaos. Those results follow from the following geometric picture we prove for the Gibbs measure, of interest by itself: asymptotically, the measure splits into infinitesimal spherical ‘bands’ centered at deep minima, playing the role of ‘pure states’. For the pure models, the latter makes precise the picture of ‘many valleys separated by high mountains’ and significant parts of the TAP analysis from the physics literature.
Similar content being viewed by others
Notes
An even mixed model \(H_{N}(\varvec{\sigma })=\sum _{p\ge 1}\gamma _{p}H_{N,2p}(\varvec{\sigma })\), either with spherical or Ising spins, is generic if and only if \(\sum p^{-1}\mathbf {1}\left( \gamma _{p}\ne 0\right) =\infty \).
We note that though we work with critical points, by Corollary 3 we could have replaced everywhere in the results \(\varvec{\sigma }_{0}^{i}\) by the corresponding enumeration of local minima instead of general critical points.
Below we will apply [7, Theorem 1.2] for the 2-spin model defined by (1) with \(p=2\) and the inverse-temperature \(\beta _{\mathrm{eff}}=\beta _{\mathrm{eff}}\left( N,q_{*}\right) \) defined in (91). In the notation of [7, Theorem 1.2], this corresponds to centered Gaussian \(J_{ij}=J_{ji}\) with variance 2 if \(i=j\) and 1 if \(i\ne j\) and inverse-temperature \(\beta _{\mathrm{eff}}/\sqrt{2}\). Since \(\beta _{\mathrm{eff}}\in \left( 0,1/\sqrt{2}\right) \), in our application we will only use (12).
I.e., the overlap between the projections of the points to the orthogonal space to the center point \(\varvec{\sigma }_{0}\in \mathbb {R}^{N}\).
The fact that such a frame field exists can be seen from the following. If we let \(\left\{ \frac{\partial }{\partial x_{i}}\right\} _{i=1}^{N-1}\) be the pullback of \(\left\{ \frac{d}{dx_{i}}\right\} _{i=1}^{N-1}\) by \(P_{\hat{\mathbf {n}}}\), then \(\left\{ \frac{\partial }{\partial x_{i}}(\hat{\mathbf {n}})\right\} _{i=1}^{N-1}\) is an orthonormal frame at the north pole. For any point in a small neighborhood of \(\hat{\mathbf {n}}\) we can define an orthonormal frame as the parallel transport of \(\left\{ \frac{\partial }{\partial x_{i}}(\hat{\mathbf {n}})\right\} _{i=1}^{N-1}\) along a geodesic from \(\hat{\mathbf {n}}\) to that point. This yields an orthonormal frame field on that neighborhood, say \(E_{i}(\varvec{\sigma })=\sum _{j=1}^{N-1}a_{ij}(\varvec{\sigma })\frac{\partial }{\partial x_{j}}(\varvec{\sigma }),\,i=1,\ldots ,N-1\). Working with the coordinate system \(P_{\hat{\mathbf {n}}}\) one can verify that at \(x=0\) the Christoffel symbols \(\varGamma _{ij}^{k}\) are equal to 0, and therefore (see e.g. [22, Eq. (2), P. 53]) the derivatives \(\frac{d}{dx_{k}}a_{ij}(P_{\hat{\mathbf {n}}}^{-1}(x))\) at \(x=0\) are also equal to 0.
Substituting (169) and (52) into (170), canceling like-terms and dividing by \(\beta q_{*}^{p}/2\), one finds that (170) holds if and only if
$$\begin{aligned} c_{p}+\frac{1}{c_{p}}\log \left( \frac{1-q_{*}^{2}\left( 1-c_{p}/\left( \beta q_{*}^{p}\right) \right) }{1-q_{*}^{2}}\right) -2E_{0}+\beta pq_{*}^{p-2}\left( 1-q_{*}^{2}\right) =0. \end{aligned}$$Using (18), (54) and (56), by straightforward algebra, one can see that the left-hand side of the above is equal to \(2\varTheta _{p}\left( -E_{0}\right) \) [see (177)], which by the definition of \(E_{0}\) (16), verifies (170).
If we apply the Kac–Rice formula to compute the expectation as is (174) with the additional restriction that \(\varvec{\sigma }_{0}\) belongs to some measurable subset \(B\subset \mathbb {S}^{N-1}\), then what changes in the integral formula on the right-hand side of (174) is that the domain of integration \(\mathbb {S}^{N-1}\) is replaced by B. Thus, the corresponding integrand is a continuous Radon–Nikodym derivative w.r.t the Lebesgue measure which, since \(\left( H_{N}\left( \varvec{\sigma }\right) ,g_{N}\left( \varvec{\sigma }\right) \right) \) is stationary, is invariant to rotations of the sphere. It is therefore a constant function.
References
Adler, R.J., Taylor, J.E.: Random Fields and Geometry. Springer Monographs in Mathematics. Springer, New York (2007)
Aizenman, M., Lebowitz, J.L., Ruelle, D.: Some rigorous results on the Sherrington–Kirkpatrick spin glass model. Commun. Math. Phys. 112(1), 3–20 (1987)
Anderson, G.W., Guionnet, A., Zeitouni, O.: An Introduction to Random Matrices, Cambridge Studies in Advanced Mathematics, vol. 118. Cambridge University Press, Cambridge (2010)
Auffinger, A., Ben Arous, G.: Complexity of random smooth functions on the high-dimensional sphere. Ann. Probab. 41(6), 4214–4247 (2013)
Auffinger, A., Ben Arous, G., Černý, J.: Random matrices and complexity of spin glasses. Commun. Pure Appl. Math. 66(2), 165–201 (2013)
Bai, Z., Silverstein, J.W.: Spectral Analysis of Large Dimensional Random Matrices, Springer Series in Statistics, 2nd edn. Springer, New York (2010)
Baik, J., Lee, J.O.: Fluctuations of the free energy of the spherical Sherrington–Kirkpatrick model. J. Stat. Phys. 165(2), 185–224 (2016)
Ben Arous, G., Dembo, A., Guionnet, A.: Aging of spherical spin glasses. Probab. Theory Relat. Fields 120(1), 1–67 (2001)
Ben Arous, G., Guionnet, A.: Large deviations for Wigner’s law and Voiculescu’s non-commutative entropy. Probab. Theory Relat. Fields 108(4), 517–542 (1997)
Bovier, A., Kurkova, I., Löwe, M.: Fluctuations of the free energy in the REM and the \(p\)-spin SK models. Ann. Probab. 30(2), 605–651 (2002)
Bray, A.J., Moore, M.A.: Chaotic nature of the spin-glass phase. Phys. Rev. Lett. 58, 57–60 (1987)
Chatterjee, S.: Disorder Chaos and Multiple Valleys in Spin Glasses. arXiv:0907.3381 (2009)
Chen, W.-K.: The Aizenman–Sims–Starr scheme and Parisi formula for mixed \(p\)-spin spherical models. Electron. J. Probab. 18(94), 14 (2013)
Chen, W.-K.: Chaos in the mixed even-spin models. Commun. Math. Phys. 328(3), 867–901 (2014)
Chen, W.-K., Dey, P., Panchenko, D.: Fluctuations of the Free Energy in the Mixed \(p\)-Spin Models with External Field. arXiv:1509.07071 (2015)
Chen, W.-K., Panchenko, D.: An approach to chaos in some mixed \(p\)-spin models. Probab. Theory Relat. Fields 157(1–2), 389–404 (2013)
Chen, W.-K., Panchenko, D.: Temperature chaos in some spherical mixed p-spin models. J. Stat. Phys. 166(5), 1151–1162 (2017)
Comets, F., Neveu, J.: The Sherrington–Kirkpatrick model of spin glasses and stochastic calculus: the high temperature case. Commun. Math. Phys. 166(3), 549–564 (1995)
Crisanti, A., Sommers, H.-J.: The spherical p-spin interaction spin glass model: the statics. Z. Phys. B Condens. Matter 87(3), 341–354 (1992)
Crisanti, A., Sommers, H.-J.: Thouless–Anderson–Palmer approach to the spherical p-spin spin glass model. J. Phys. I Fr. 5, 805–813 (1995)
Dembo, A., Zeitouni, O.: Large Deviations Techniques and Applications, Volume 38 of Applications of Mathematics (New York), 2nd edn. Springer, New York (1998)
do Carmo, M.P.: Riemannian Geometry. Mathematics: Theory & Applications. Birkhäuser Boston, Inc., Boston, MA, (1992). Translated from the second Portuguese edition by Francis Flaherty
Faraut, J.: Logarithmic Potential Theory, Orthogonal Polynomials, and Random Matrices. In: Modern Methods in Multivariate Statistics, Lecture Notes of CIMPA-FECYT-UNESCO-ANR. Hermann (2014)
Fisher, D.S., Huse, D.A.: Ordered phase of short-range ising spin-glasses. Phys. Rev. Lett. 56, 1601–1604 (1986)
Jagannath, A.: Approximate Ultrametricity for Random Measures and Applications to Spin Glasses. arXiv:1412.7076 (2014)
Kurchan, J., Parisi, G., Virasoro, M.A.: Barriers and metastable states as saddle points in the replica approach. J. Phys. I Fr. 3, 1819–1838 (1993)
Liggett, T.M.: Random invariant measures for Markov chains, and independent particle systems. Z. Wahrsch. Verw. Geb. 45(4), 297–313 (1978)
Panchenko, D.: The Sherrington–Kirkpatrick Model. Springer Monographs in Mathematics. Springer, New York (2013)
Panchenko, D.: Chaos in temperature in generic \(2p\)-spin models. Commun. Math. Phys. 346(2), 703–739 (2016)
Panchenko, D., Talagrand, M.: On the overlap in the multiple spherical SK models. Ann. Probab. 35(6), 2321–2355 (2007)
Parisi, G.: Order parameter for spin-glasses. Phys. Rev. Lett. 50, 1946–1948 (1983)
Rizzo, T.: Spin Glasses: Statics and Dynamics: Summer School, Paris 2007, chapter Chaos in Mean-field Spin-Glass Models, pp. 143–157. Birkhäuser Basel, Basel (2009)
Rizzo, T., Yoshino, H.: Chaos in glassy systems from a Thouless–Anderson–Palmer perspective. Phys. Rev. B 73, 064416 (2006)
Subag, E.: The complexity of spherical \(p\)-spin models—a second moment approach. to appear in Ann. Probab. (2017)
Subag, E., Zeitouni, O.: The extremal process of critical points of the pure \(p\)-spin spherical spin glass model. Probab. Theory Relat. Fields (2016). doi:10.1007/s00440-016-0724-2
Talagrand, M.: Free energy of the spherical mean field model. Probab. Theory Relat. Fields 134(3), 339–382 (2006)
Talagrand, M.: Construction of pure states in mean field models for spin glasses. Probab. Theory Relat. Fields 148(3–4), 601–643 (2010)
Thouless, D.J., Anderson, P.W., Palmer, R.G.: Solution of ‘solvable model of a spin glass’. Philos. Mag. 35(3), 593–601 (1977)
Acknowledgements
I am grateful to my adviser Ofer Zeitouni for many very fruitful conversations and for carefully reading the manuscript. I would also like to thank Gérard Ben Arous and Aukosh Jagannath for helpful discussions.
Author information
Authors and Affiliations
Corresponding author
Additional information
This work is supported by the Adams Fellowship Program of the Israel Academy of Sciences and Humanities and by a US-Israel BSF Grant.
Appendix: The Kac–Rice formula and related auxiliary results
Appendix: The Kac–Rice formula and related auxiliary results
The Kac–Rice formula [1, Theorem 12.1.1] is a basic tool in our analysis, allowing us to relate several quantities of interest to the conditional probability \(\mathbb {P}_{u,0}\). Below we prove a number of simple derivatives of the formula; some using the results of Sect. 2.2 on critical points and values.
The Kac–Rice formula can be used to express the mean number of critical points \(\varvec{\sigma }_{0}\) on the sphere at which the values of some other fields belong to some target set. For us, those other fields will be usually related to masses of bands around \(\varvec{\sigma }_{0}\). The variant of the Kac–Rice formula that we use is [1, Theorem 12.1.1]. In the notation of [1, Theorem 12.1.1] we will consider situations where, with some function \(g_{N}\left( \varvec{\sigma }\right) \),
where we recall that, with \(E=(E_{i})_{i=1}^{N-1}\) being an orthonormal frame field on the sphere, we denote
The application of [1, Theorem 12.1.1] requires \(h\left( \varvec{\sigma }\right) \) to satisfy certain non-degeneracy conditions—namely, conditions (a)–(g) in [1, Theorem 12.1.1]. We will say that \(g_{N}\left( \varvec{\sigma }\right) \) is tame if the conditions are satisfied and if \(\{h\left( \varvec{\sigma }\right) \}_{\varvec{\sigma }}\) is a stationary random field. Using (1), the conditions are easy to check in any case we will apply the formula and this will be left to the reader. Let
denote the surface area of the \(N-1\)-dimensional unit sphere. The following is obtained from a direct application of the formula.
Lemma 13
Suppose that \(g_{N}\left( \varvec{\sigma }\right) \) is tame and \(D_{N}\) and \(J_{N}\) are some intervals, then
Proof
Applying [1, Theorem 12.1.1] with (171) yields the integral formula
where \(\phi _{\nabla H_{N}\left( \varvec{\sigma }\right) }\left( x\right) \) is the Gaussian density of \(\nabla H_{N}\left( \varvec{\sigma }\right) \) and \(\nu \) is the standard measure on the sphere (not normalized). Since the integrand above is independent of \(\varvec{\sigma }\),Footnote 10 we can replace the integral with the value of the integrand evaluated at \(\varvec{\sigma }=\hat{\mathbf {n}}\) and multiply by \(\omega _{N}N^{\frac{N-1}{2}}\), the volume of \(\mathbb {S}^{N-1}\). By Lemma 2, \(\nabla H_{N}\left( \hat{\mathbf {n}}\right) \sim \mathcal {N}\left( 0,p\mathbf {I}_{N-1}\right) \), so that \(\phi _{\nabla H_{N}\left( \varvec{\sigma }\right) }\left( 0\right) =(2\pi p)^{-\frac{N-1}{2}}\), and \(H_{N}\left( \hat{\mathbf {n}}\right) \) and \(\nabla H_{N}\left( \hat{\mathbf {n}}\right) \) are independent. The proof is completed by conditioning on \(H_{N}\left( \hat{\mathbf {n}}\right) \) in addition to \(\nabla H_{N}\left( \varvec{\sigma }\right) =0\) and some calculus. \(\square \)
Several derivatives of (173) are of use to us. Their proofs will involve the rate function \(\varTheta _{p}\left( E\right) \) of Theorem 5 which we now define explicitly. Let \(\nu ^{*}\) denote the semicircle measure, the density of which with respect to Lebesgue measure is
and define the function (see e.g., [23, Proposition II.1.2])
The exponential growth rate function of (15) is given [5] by
where \(\gamma _{p}=\sqrt{p/\left( p-1\right) }\).
Lemma 14
Suppose that \(g_{N}\left( \varvec{\sigma }\right) \) is tame, \(D_{N}\) are some intervals and \(J_{N}= NJ \) where \(J\subset \left( -\infty ,0\right) \) is a fixed finite interval. Let \(\varphi {:}\,\mathbb {R}\rightarrow \mathbb {R}\) be a continuous function and assume that
Then
Proof
From (173) and Hölder’s inequality,
where
\(a>1\) is an arbitrary number, and \(b:=b(a)=a/\left( a-1\right) \).
First,
Second, by a change of variables \(u\mapsto Nv\),
where \(\gamma _{p}=\sqrt{p/\left( p-1\right) }\) and \(\varOmega \) is defined in (176). Thus, by (177), the lemma follows if we show that for any \(a>1\),
and that
The inequality (185) obviously follows from (178), since \(\lim _{a\rightarrow \infty }b(a)=1\).
From (38) and (42), the conditional law of \(\mathbf {V}\) under \(\mathbb {P}_{u,0}\left\{ \cdot \right\} \) is identical to that of
where \(\mathbf {M}\) is a GOE matrix and \(\mathbf {I}\) is the identity matrix, both of dimension \(N-1\). For any \(0<\epsilon<1<\kappa \), with \(\lambda _{i}\) denoting the eigenvalues of \(\mathbf {M}\) and \(\lambda _{*}\left( u\right) =\max _{i}\left| \lambda _{i}-\gamma _{p}\frac{u}{\sqrt{N\left( N-1\right) }}\right| \),
where
From the upper bound on the maximal eigenvalue of [8, Lemma 6.3],
uniformly for \(u\in J_{N}\). The empirical measure of eigenvalues of GOE matrices \(L_{N}=\frac{1}{N-1}\sum _{i=1}^{N-1}\lambda _{i}\) satisfies a large deviation principle with speed \(N^{2}\) and a good rate function \(J_{0}\left( \nu \right) \), in the space of Borel probability measures on \(\mathbb {R}\) equipped with the weak topology, which is compatible with the Lipschitz bounded metric; see [9, Theorem 2.1.1]. The good rate function \(J_{0}\left( \nu \right) \) satisfies \(J_{0}\left( \nu \right) =0\) if and only if \(\nu =\nu ^{*}\) is the semicircle law (175). Combined with the fact that the functions \(\log _{\epsilon }^{\kappa }\left( \left| \cdot -u'\right| \right) \) have the same Lipschitz constant and bound for all \(u'\in \mathbb {R}\), this implies that for the event
we have for any \(\delta >0\) some positive number \(d\left( \delta \right) \) such that
By taking \(\kappa \) large enough and \(\epsilon \) small enough, combining this with (187) and (189) proves (184) and completes the proof. \(\square \)
For intervals \(J_{N}\) of length \(o\left( N\right) \) around \(-NE_{0}\) the following is also useful for us.
Lemma 15
Suppose that \(g_{N}\left( \varvec{\sigma }\right) \) is tame, \(D_{N}\) are some intervals and \(J_{N}=\left( m_{N}-a_{N},m_{N}+a_{N}^{\prime }\right) \) for some sequences \(a_{N},\,a_{N}^{\prime }=o\left( N\right) \). Let \(c_{p}\) be as defined in (18). Then for large enough N,
where \(C>0\) is an appropriate constant.
Proof
We shall use the notation (181) introduced in the proof of Lemma 14. Since the conditional law of \(\mathbf {V}\) under \(\mathbb {P}_{u,0}\left\{ \cdot \right\} \) is identical to that of the shifted GOE matrix (186), as a particular case of [34, Corollary 23],
uniformly in \(u\in J_{N}\), for some constant \(C>0\). As in the proof of Lemma 14, (180) holds and therefore, taking \(a=2\) and using 191, we obtain that
By Lemma 13,
By definition, up to a factor of 2 in the even p case [related to the normalization factor in (19)], the integrand above is equal to the density of the intensity measure of the extremal point process of critical points \(\xi _{N}\), defined in (19), shifted by \(m_{N}\). Thus, by [35, Proposition 3] it converges uniformly to \(e^{c_{p}\left( u-m_{N}\right) }\) (see also Theorem 7 above). Combined with (192) this yields (190). \(\square \)
The non-negative random variable
can be approximated by
where \(\left[ t_{i},t_{i+1}\right) ,\,i\le k\), form a finite partition of \(\left[ 0,\infty \right) \). Combining this with the monotone convergence theorem and (173) one obtains the following.
Corollary 8
We have that
We have the following exponential bound on the expectation above.
Lemma 16
Let \(J_{N}= NJ \) where \(J\subset \mathbb {R}\) is a fixed interval. Let \(\varphi {:}\,\mathbb {R}\rightarrow \mathbb {R}\) be a continuous function and suppose that
Then
Proof
Abbreviate \(\mathrm{Band}\left( \varvec{\sigma }_{0}\right) =\mathrm{Band}\left( \varvec{\sigma }_{0},q_{N},q_{N}^{\prime }\right) \). By Hölder’s inequality,
is a convex function. Hence, for \(b\in \left( 1,2\right) \), for any \(u\in J_{N}\),
where we define
and use the fact that
For any \(b>1\), with \(a:=a(b)=b/\left( b-1\right) \), we have, using the notation (181),
Using (196) and (184) and letting \(b\searrow 1\) we obtain
From (193), (182) and (183), we conclude that (195) follows. \(\square \)
We finish with the following lemma
Lemma 17
Let \(J_{N}=\left( m_{N}-a_{N},m_{N}+a_{N}^{\prime }\right) \) with some sequences \(a_{N},\,a_{N}^{\prime }=o\left( N\right) \), and define \(c_{p}\) as in (18). Then, for large enough N,
where \(C>0\) is an appropriate constant.
Proof
By (193), (197) and (191) it follows that, for an appropriate \(C>0\), for large N,
where \(\mathbf {V}\) is defined in (181). The proof now follows from the argument used in the proof of Lemma 15 right after (192). \(\square \)
Rights and permissions
About this article
Cite this article
Subag, E. The geometry of the Gibbs measure of pure spherical spin glasses. Invent. math. 210, 135–209 (2017). https://doi.org/10.1007/s00222-017-0726-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00222-017-0726-4