On the Replica Symmetry of a Variant of the Sherrington-Kirkpatrick Spin Glass

Christian Brennecke, Adrien Schertzer · 2024

Read On the Replica Symmetry of a Variant of the Sherrington-Kirkpatrick Spin Glass by Christian Brennecke, Adrien Schertzer (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-04336.txt.

Introduction

Christian Brennecke and Adrien Schertzer study a spin glass variant in a 2024 math paper. The classic Sherrington Kirkpatrick model assigns random couplings between spins. Their version adds a quartic correction and lets the spins carry a nonzero average. Bolthausen proposed the toy to test whether a simple moment method can prove the replica symmetric formula at all high temperatures.

A condition in the style of de Almeida and Thouless marks the conjectured boundary for that formula. Small coupling is already known to sit safely inside. The pair push the analysis toward the full range the conjecture claims. Toy models like this serve as proving grounds: a method that works here may crack the harder original model.

Catalog record

CREATORChristian Brennecke, Adrien Schertzer
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-04336

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

On the Replica Symmetry of a Variant of the Sherrington-Kirkpatrick Spin Glass Christian Brennecke, Adrien Schertzer Abstract We consider N N i.i.d. Ising spins with mean m ∈ ( − 1 , 1 ) m\in(-1,1) whose interactions are described by a Sherrington-Kirkpatrick Hamiltonian with a quartic correction. This model was recently introduced by Bolthausen in [ 3 ] as a toy model to understand whether a second moment argument can be used to derive the replica symmetric formula in the full high temperature regime if m ≠ 0 m\neq 0 . In [ 3 ] , Bolthausen suggested that a natural analogue of the de Almeida-Thouless condition for the toy model is β 2 ​ ( 1 − m 2 ) 2 ≤ 1 . \beta^{2}(1-m^{2})^{2}\leq 1. (1) Here, β ≥ 0 \beta\geq 0 corresponds to the inverse temperature. While the second moment method implies replica symmetry for β \beta sufficiently small, Bolthausen showed that the method fails to prove replica symmetry in the full region described by ( 1 ). A natural question that was left open in [ 3 ] is whether ( 1 ) correctly characterizes the high temperature phase of the toy model. In this note, we show that this is indeed not the case. We prove that if | m | ≥ m ∗ |m|\geq m_{*} , for some m ∗ ∈ ( 0 , 1 ) m_{*}\in(0,1) , the limiting free energy of the toy model is negative for suitable β \beta that satisfy ( 1 ). 1 Introduction and Main Result For \beta\geq 0 and m\in(-1,1) , consider the mean field spin glass with partition function Z_{N,\beta,m}:=\sum_{\sigma\in\{-1,1\}^{N}}p_{m}(\sigma)\exp\bigg(\sum_{1\leq i,j\leq N}\frac{\beta}{\sqrt{2}}g_{ij}\widehat{\sigma}_{i}\widehat{\sigma}_{j}-\frac{\beta^{2}}{4N}\sum_{1\leq i,j\leq N}\widehat{\sigma}_{i}^{2}\widehat{\sigma}_{j}^{2}\bigg). (2) Here, \sigma=(\sigma_{1},\ldots,\sigma_{N})\in\{-1,1\}^{N},\widehat{\sigma}_{i}:=\sigma_{i}-m , p_{m}(\sigma):=\prod_{i=1}^{N}\frac{1}{2}(1+m\sigma_{i}) and the couplings (g_{ij})_{1\leq i,j\leq N} are i.i.d. Gaussian random variables with distribution \mathcal{N}(0,N^{-1}) . The model in ( 2 ) was recently introduced by Bolthausen in [ 3 ] . The main result of [ 3 ] is a new proof of the replica symmetric (RS) formula for the classical Sherrington-Kirkpatrick (SK) model [ 13 ] with external field strength h\in\mathbb{R} and for sufficiently small \beta\geq 0 . The proof is based on a second moment argument, conditionally on an approximate solution of the TAP equations [ 18 ] that was constructed in [ 2 ] . The RS formula is expected to be true under the de Almeida-Thouless [ 8 ] condition \beta^{2}E(1-m_{\beta,h}^{2})^{2}\leq 1,\;\text{ for }\;m_{\beta,h}=\text{tanh}(h+\beta\sqrt{q}Z),\;q=E\,\text{tanh}^{2}(h+\beta\sqrt{q}Z), (3) where Z is a standard Gaussian random variable. To date, this was verified in a very large subregion of the phase diagram in [ 19 , 16 , 17 , 10 ] , based on the analysis of the Parisi formula [ 11 , 12 ] proved in [ 9 , 15 ] . A complete proof that the RS phase is fully characterized by ( 3 ) still remains open, however. Note for a centered Gaussian external field, the RS phase indeed turns out to be characterized by an analogue of ( 3 ), see [ 6 ] . It is a natural question whether a second moment argument as in [ 3 ] can be improved to verify the replica symmetry in the expected region ( 3 ). Although the temperature range from [ 3 ] can be improved through basic restriction arguments [ 5 ] , Bolthausen already suggested in [ 3 ] that understanding the full RS phase likely requires additional techniques. His argument is based on the analysis of the toy model defined in ( 2 ). For this model, the second moment method can be evaluated explicitly and it is straightforward to show that for \beta\geq 0 small enough, one has that \lim_{N\to\infty}\frac{1}{N}\log\mathbb{E}Z_{N,\beta,m}=\lim_{N\to\infty}\frac{1}{N}\log\mathbb{E}Z_{N,\beta,m}^{2}=0. (4) By Gaussian concentration, see e.g. [ 16 , Th. 1.3.4.] , this implies for small \beta that \lim_{N\to\infty}\mathbb{E}\frac{1}{N}\log Z_{N,\beta,m}=0. (5) By analogy to ( 3 ), Bolthausen then suggested that the high temperature region \text{HT}_{m}:=\Big\{\beta\geq 0:\lim_{N\to\infty}\mathbb{E}\frac{1}{N}\log Z_{N,\beta,m}=0\Big\} (6) the set in which the quenched equals the annealed free energy, should correspond to \text{AT}_{m}:=\big\{\beta\geq 0:\beta^{2}(1-m^{2})^{2}\leq 1\big\} (7) and it was left open whether ( 5 ) indeed holds true for all \beta\in\text{AT}_{m} (cf. [ 3 , Eq. (6.1)] ). Note that one might expect that \text{AT}_{m}\subset\text{HT}_{m} based on different arguments. For example, an application of the TAP heuristics suggests that \textbf{m}=(m,m,\ldots,m)\in(-1,1)^{N} solves an analogue of the TAP equations associated with ( 2 ). Alternatively, it is not difficult to see that, for \beta sufficiently small, the free energy fluctuations converge to a normal variable with a variance that contains the term \log\big(1-\beta^{2}(1-m^{2})^{2}\big) . This indicates a phase transition when \beta^{2}(1-m^{2})^{2}\to 1 , similarly as in the SK model without external field analyzed in detail in [ 1 ] (in which case the corresponding divergence at \beta\to 1 detects the RS-RSB transition temperature correctly). It turns out 1 1 1 We thank W.-K. Chen for this remark and for pointing out [ 7 ] to us. We thank W.-K. Chen and S. Tang for sharing some related numerical simulations on ( 6 ) and ( 7 ) with us. that a precise characterization of \text{HT}_{m} may be obtained by adapting the methods from [ 7 ] to the present setting. Related numerical simulations suggest that for large |m| , one can find \beta\in\text{AT}_{m} with \beta\not\in\text{HT}_{m} . Perhaps surprisingly, this fact admits an elementary proof which does neither require precise knowledge of \text{HT}_{m} nor of a Parisi formula for the limiting free energy of the model. Theorem 1 . Let I(m):=-\frac{1+m}{2}\log\frac{1+m}{2}-\frac{1-m}{2}\log\frac{1+m}{2} and define m_{*}>0 to be such that I(m_{*})=\frac{1}{4} . Then, for every |m|>m_{*} , there exists \beta\in\emph{AT}_{m} with \beta\not\in\emph{HT}_{m} . 2 Proof of Theorem 1 Our proof is elementary and adapts an argument from the analysis of the random energy model, see [ 4 , p. 167] , to the present setting. By symmetry, we assume in the rest of this section without loss of generality that m>0 . Proof of Theorem 1 . In the following, denote by E_{m} the expectation with regards to the base measure p_{m} and, for \epsilon>0 , define the set S_{N,m,\epsilon}:=\Big\{\sigma\in\{-1,1\}^{N}:\frac{1}{N}\sum_{i=1}^{N}\sigma_{i}\in[m-\epsilon,m+\epsilon]\Big\}. Moreover, denote by H_{N,\beta,m} the Hamiltonian H_{N,\beta,m}(\sigma):=\sum_{1\leq i,j\leq N}\frac{\beta}{\sqrt{2}}g_{ij}\widehat{\sigma}_{i}\widehat{\sigma}_{j}-\frac{\beta^{2}}{4N}\sum_{1\leq i,j\leq N}\widehat{\sigma}_{i}^{2}\widehat{\sigma}_{j}^{2} so that Z_{N,\beta,m}=E_{m}\exp(H_{N,\beta,m}) . We first claim that for every \beta\in\text{HT}_{m} and every \epsilon>0 , we have that \lim_{N\to\infty}\Big|\frac{1}{N}\log(Z_{N,\beta,m})-\frac{1}{N}\log E_{m}\left(\textbf{1}_{S_{N,m,\epsilon}}\exp(H_{N,\beta,m})\right)\Big|=0 (8) in the sense of probability. Here, \textbf{1}_{S}:\{-1,1\}^{N}\to\{0,1\} denotes the characteristic function of the subset S\subset\{-1,1\}^{N} . To prove ( 8 ), set Z_{N,\beta,m}^{\leq\epsilon}:=E_{m}(\textbf{1}_{S_{N,m,\epsilon}}\exp H_{N,\beta,m}),\;Z_{N,\beta,m}^{>\epsilon}:=E_{m}(\textbf{1}_{S_{N,m,\epsilon}^{\complement}}\exp H_{N,\beta,m}) so that Z_{N,\beta,m}=Z_{N,\beta,m}^{\leq\epsilon}+Z_{N,\beta,m}^{>\epsilon} . Then, for every \delta>0 , we have that \begin{split}&\mathbb{P}\Big(\Big|\frac{1}{N}\log(Z_{N,\beta,m})-\frac{1}{N}\log\big(Z_{N,\beta,m}^{\leq\epsilon}\big)\Big|\geq\delta\Big)\\ &=\mathbb{P}\Big(Z_{N,\beta,m}^{>\epsilon}/Z_{N,\beta,m}^{\leq\epsilon}\geq e^{\delta N}-1,\;Z_{N,\beta,m}^{\leq\epsilon}\leq e^{-\delta N}\Big)\\ &\hskip 11.38092pt+\mathbb{P}\Big(Z_{N,\beta,m}^{>\epsilon}/Z_{N,\beta,m}^{\leq\epsilon}\geq e^{\delta N}-1,\;Z_{N,\beta,m}^{\leq\epsilon}>e^{-\delta N}\Big)\\ &\leq\mathbb{P}\Big(Z_{N,\beta,m}^{\leq\epsilon}\leq e^{-\delta N}\Big)+\mathbb{P}\Big(Z_{N,\beta,m}^{>\epsilon}\geq 1-e^{-\delta N}\Big)\\ &\leq\mathbb{P}\Big(Z_{N,\beta,m}^{\leq\epsilon}\leq e^{-\delta N}\Big)+\big(1-e^{-\delta N}\big)^{-1}p_{m}\big(S_{N,m,\epsilon}^{\complement}\big),\end{split} (9) where in the last step, we applied Markov’s inequality. Similarly, we have that \begin{split}&\mathbb{P}\Big(Z_{N,\beta,m}^{\leq\epsilon}\leq e^{-\delta N}\Big)\\ &\leq\mathbb{P}\Big(Z_{N,\beta,m}^{\leq\epsilon}\leq e^{-\delta N},\;Z_{N,\beta,m}^{>\epsilon}\leq e^{-\delta N}\Big)+\mathbb{P}\Big(Z_{N,\beta,m}^{>\epsilon}>e^{-\delta N}\Big)\\ &\leq\mathbb{P}\Big(\frac{1}{N}\log Z_{N,\beta,m}\leq-\frac{\delta}{2}\Big)+e^{\delta N}p_{m}\big(S_{N,m,\epsilon}^{\complement}\big)\end{split} for N large enough so that \begin{split}&\mathbb{P}\Big(\Big|\frac{1}{N}\log(Z_{N,\beta,m})-\frac{1}{N}\log\big(Z_{N,\beta,m}^{\leq\epsilon}\big)\Big|\geq\delta\Big)\\ &\leq\mathbb{P}\Big(\Big|\frac{1}{N}\log Z_{N,\beta,m}\Big|\geq\frac{\delta}{2}\Big)+\frac{e^{\delta N}}{1-e^{-\delta N}}\,p_{m}\big(S_{N,m,\epsilon}^{\complement}\big).\end{split} Markov’s inequality and Gaussian concentration (see [ 16 , Th. 1.3.4.] ) imply that \begin{split}\mathbb{P}\Big(\Big|\frac{1}{N}\log Z_{N,\beta,m}\Big|\geq\frac{\delta}{2}\Big)&\leq\mathbb{P}\Big(\Big|\frac{1}{N}\log Z_{N,\beta,m}-\mathbb{E}\frac{1}{N}\log Z_{N,\beta,m}\Big|\geq\frac{\delta}{4}\Big)\!+\frac{4}{\delta}\Big|\mathbb{E}\frac{1}{N}\log Z_{N,\beta,m}\Big|\\ &\leq Ce^{-cN\delta^{2}/\beta^{2}}+\frac{4}{\delta}\Big|\mathbb{E}\frac{1}{N}\log Z_{N,\beta,m}\Big|\end{split} for suitable constants C,c>0 (independent of all parameters). On the other hand, setting h:=\tanh^{-1}(m) , a standard exponential moment bound shows that \begin{split}p_{m}\big(S_{N,m,\epsilon}^{\complement}\big)&\leq\inf_{\lambda\geq 0}\exp\big(-N\epsilon\lambda-Nm\lambda+N(\log\cosh(h+\lambda)-\log\cosh(h)\big)\\ &\hskip 14.22636pt+\inf_{\lambda\geq 0}\exp\big(-N\epsilon\lambda+Nm\lambda+N(\log\cosh(h-\lambda)-\log\cosh(h)\big)\\ &\leq 2\inf_{\lambda\geq 0}\exp\big(-N\epsilon\lambda+N\lambda^{2}/2\big)=2\exp(-N\epsilon^{2}/2).\end{split} Now, for every \delta^{\prime}>0 , we can find 0<\delta<\min(\delta^{\prime},\epsilon^{2}/4) so that we arrive at \begin{split}&\limsup_{N\to\infty}\mathbb{P}\Big(\Big|\frac{1}{N}\log(Z_{N,\beta,m})-\frac{1}{N}\log\big(Z_{N,\beta,m}^{\leq\epsilon}\big)\Big|\geq\delta^{\prime}\Big)\\

Read the complete text