A cram\'er type moderate deviation theorem for the critical curie-weiss model

In this paper we study the moderate deviations for the magnetization of critical Curie-Weiss model. Chen, Fang and Shao considered a similar problem for non-critical model by using Stein method. By direct and simple arguments based on Laplace method, we provide an explicit formula of the error and deduce a Cram\'er-type result.


Introduction
Let (X i ) be a sequence of i.i.d. random variables satisfying EX 1 = 0, Var(X 1 ) = 1. Then the classic Central limit theorem says that the normalized sum W n = (X 1 + . . . + X n )/ √ n converges in law to a standard normal random variable W . A natural question is to understand the rate of the convergence of the tail probability P(W n > x) to P(W > x) for the largest possible range of x. There are two major approaches to measure the approximation error. The first approach is to study the absolute error by Berry-Esseen type bounds. The other one is to study the relative error of the tail probability. One of the first result in this approach is the following Cramér type moderate deviation theorem. If E(e α|X 1 | 1/2 ) < ∞, for some α > 0, then for 0 ≤ x ≤ n 1/6 , with Φ the standard normal distribution function. It has been also shown that the assumptions on the exponential moment of X 1 and the length of range n 1/6 are optimal. We refer the reader to the book [12] for a proof of this result and a more detailed discussion.
The Cramér type moderate deviation results have been proved to be useful in designing statistical tests since they give a relation between the size and the accuracy of tests, see e.g. [9,10]. Hence, a lot of attention has been drawn in investigating this problem not only for independent variables but also for dependent structures as stationary process [1,15], self-linear process [11], normalized sums [4,14], and L-statistics [8]. On the other hand, Cramér type moderate deviation theorems for nonnormal limit distribution are also provided, such as for chi-squared distribution [10], for sub-Gaussian or exponential distribution [2].
In this paper, we study the case of the critical Curie-Weiss model, where the spin variables are dependent and the limit distribution is nonnormal. Let us first recall some definitions and existing results for Curie-Weiss model. For n ∈ N, let Ω n = {±1} n be the space of spin configurations. The spin configuration probability is given by Boltzman-Gibbs distribution, i.e. for any σ ∈ Ω n , where Z n is the normalizing factor, β > 0 and h ∈ R are inverse temperature and external field respectively. The Curie-Weiss model has been shown to exhibit a phase transition at β c = 1. More precisely, the asymptotic behavior of the total spin (also called the magnetization) S n = σ 1 + . . . + σ n changes when β crosses the critical value 1. Let us consider the following fixed-point equation .
The equation (1) has a unique solution m 0 , such that m 0 h ≥ 0. In this case, S/n is concentrated around m 0 and has a Gaussian limit under proper standardization, see [6]. Moreover, in [3] the authors prove the following moderate deviation theorem for the magnetization by using Stein method.
Then we have µ n (σ : for 0 ≤ x ≤ n 1/6 . Case 2. β > 1, h = 0 (the low temperature regime without external field). The equation (1) has two nonzero solutions m 1 < 0 < m 2 , where m 1 = −m 2 . In this case, one has the conditional central limit theorems as follows: conditionally on S n < 0 (resp. S n > 0), S/n is concentrated around m 1 (resp. m 2 ) and has a Gaussian limit after proper scaling, see [6]. Similarly to case 1, a moderate deviation result has been also proved.
We remark that in [5], the authors generalize Theorem 1.3 to a near critical regime of inverse temperature β = 1 + O( 1 √ n ). They also consider a general class of Curie-Weiss model, where the distribution of a single spin is a generic probability measure instead of Bernoulli distribution as in the classical model.
In this paper, we will prove a Cramér type moderate deviation theorem for the total spin in the critical case. Our main result is as follows. Then there exists a positive constant C, such that for all n large enough and 0 ≤ x ≤ n 1/12 , where It is worth noting that Theorem 1.4 gives the exact formula of the error term of order n −1/2 , while moderate deviation results in Theorems 1.1 and 1.2 only show asymptotic estimates of the error terms. The range of estimate n 1/6 is replaced by n 1/12 due to the change of scaling and limit distribution. The proof of Theorem 1.4 is simple and direct, based on Laplace method-like arguments. We have a direct corollary.
Moreover, for any fixed real number x, The first part of this corollary is a Cramér moderate deviation result in classic form, whereas the second part is an improvement of Theorem 1.3.
The paper is organized as follows. In Section 2, we provide some preliminary results. In Section 3, we prove the main theorem 1.4. We fix here some notation. If f and g are two real functions, we write

A lemma on the integral approximations.
Lemma 2.1. Let m, q, p be positive real numbers.
and m<ℓ<q 2∤ℓ (ii) Assume that f (t) is a differentiable function on R and there exists a positive con- and m<ℓ<q 2∤ℓ Proof. The proof of (i) is simple, so we safely leave it to the reader. For (ii), by using the mean value theorem, we get that for any ℓ, Therefore, by summing over ℓ we get desired results.

2.2.
Estimates on the binomial coefficients. We first recall a version of Stirling approximation (see [13]) that for all n ≥ 1, .
Using this approximation, we can show that n k ≤ e nI(k/n) , for all k = 0, . . . , n, and where I(0) = I(1) = 0 and for t ∈ (0, 1), We will see in Section 3.1 that the function J(t) defined by (6) J(t) = I(t) + (2t − 1) 2 2 plays an important role in the expression of the distribution function of W n . We prove here a lemma to describe the behavior of J(t).

Proof of Theorem 1.4
3.1. An expression of the distribution function of W n . Let us denote by F n (x) the distribution function of W n , i.e. for x ∈ R F n (x) = µ n (σ : W n ≤ x) = µ n (σ : σ 1 + . . . + σ n ≤ n 3/4 x).

3.2.
Estimates of A n andÂ n .
By letting ℓ = 2k − n, we obtain where 12 . The proof of the following lemma is simple, so we omit it.  Moreover, Combining the last three estimates gives that We now can deduce from (20), (21), (22) and (23) an estimate of B n that 3.4. Estimates of B n,x . Using the same arguments for (20), we also have In the sequel, we consider two cases: x > 10 and x ≤ 10. For the case x > 10, we will use Part (i) of Lemma 2.1 to obtain a sharp estimate on B n,x , while for the case x ≤ 10, as for B n , we apply Part (ii) to get a suitable estimate. The choice of the number 10 is flexible. We just need the fact that the functions p 1 (·), p 2 (·) and r(·) are decreasing in the interval  Therefore, Notice that 1 − F (x) =P 1 (x)/P 1 (−∞). Therefore, √ nP 2 (x) P 1 (x) + O(1) n −1R (x) P 1 (x) + n −3/4 p 1 (x) P 1 (x) + n −5/4 p 2 (x) P 1 (x) + n −7/4 r(x) where for the last line we have used that R(x) P 1 (x) = Θ(x 12 ), p 1 (x) P 1 (x) = Θ(1), p 2 (x) P 1 (x) = Θ(x 6 ), r(x) P 1 (x) = Θ(x 12 ).