Entropic Inequalities and Limit Theorems in Boolean Probability

First assume that \Psi(X_i)<\infty for all i=1,\ldots,n. By the Stam inequality (?), we have \Psi(S_A)<\infty for all A\subset [n]. Set \rho_i=\mathcal L(X_i^2), \rho_A=\mathcal L(S_A^2). By (?),

Thus Lemma 2.1 gives

Since each r_{\rho_i} is nonnegative,

Set

If a_A\ge T for some A, then (Proof) immediately gives

We may therefore assume a_A<T for every A. Define

Then \sum_Aw_A=1 and w_Ac_A=\beta_A. Let \tilde\rho_A=\rho_A^{\uplus c_A}. By the definition of fractional partition and (Proof),

Since K_\rho(z)=z-F_\rho(z) and \sum_A w_A=1, we obtain

Evaluating at z=-s for s>0, we have F_{\rho_{[n]}}(-s)<0 and F_{\tilde\rho_A}(-s)<0 for every A. Since x\mapsto x^{-2} is convex on (-\infty,0), Jensen’s inequality therefore gives

Integrating with respect to s and applying Lemma 2.1 and Lemma 2.2,

Taking reciprocals yields the desired inequality.

If \Psi(X_i)=\infty for some i\in[n], we regularize the corresponding law by setting \rho_i(\varepsilon):=\rho_i\uplus\delta_\varepsilon. Then r_{\rho_i(\varepsilon)}(s)=r_{\rho_i}(s+\varepsilon). Applying the previous argument to the perturbed laws and then by Lemma 2.1, letting \varepsilon\downarrow0, the monotone convergence theorem completes the proof.

For any s>0, set

Fix s>0. Differentiating with respect to t and using the semigroup relation (?) gives

The second term on the right-hand side is nonnegative, and we deduce that t\partial_th_t(s)\le 0. Moreover,

Therefore, by Cauchy-Schwarz,

Substituting this inequality into (Proof) yields

The right-hand inequality implies that t\mapsto h_t(s) is decreasing. For the left-hand inequality, we observe that \partial_t\bigl(t h_t(s)\bigr) =h_t(s)+t\partial_t h_t(s)\ge0. Thus t\mapsto t h_t(s) is increasing, which implies that t h_t(s)\ge h_1(s), and consequently,

For \varepsilon>0, introduce the regularized logarithmic integral

Tonelli’s theorem gives,

Assume first that \Gamma(\mu)>-\infty. Taking t=1 in (?) shows that h_1 is integrable on every interval (0,\varepsilon). For fixed s>0, weak convergence \mu_t\Rightarrow \mathbf{b} gives

The upper bound in (Proof) and dominated convergence therefore yield

On the other hand, for sufficiently large M>0, decompose

Since the family \{\mu_t\}_{t\ge1} has uniformly bounded second moments, it is uniformly integrable. Thus, the second term can be made uniformly arbitrarily small. The first term tends to \Gamma_\varepsilon(\mathbf{b}) by the weak convergence \mu_t\Rightarrow \mathbf{b}. Therefore, letting M\to \infty yields

Combining this with (?) proves \Gamma(\mu_t)\longrightarrow0.

Now assume \Gamma(\mu)=-\infty. Then (?) at t=1 implies

The lower bound in (Proof) gives

Since \Gamma_\varepsilon(\mu_t) is finite, (?) forces \Gamma(\mu_t)=-\infty.

By (?), (?), let \nu be such that F_\mu(z)=z-G_\nu(z), and hence

Write b_3=m_1(\nu)=m_3(\mu),\ b_4= m_2(\nu)=m_4(\mu)-1. For any s>0, define

so that G_\nu(is)=-R_s-isI_s. By Lemma 2.3, after the change of variables s\mapsto s^2/n, (2.3) becomes

where

The compact region. Fix M>0. By Jensen’s inequality,

Hence, for all sufficiently large n, I_s\geq n^{-1} on (0,M], and therefore

On the other hand, F_\mu(is) = is-G_\nu(is) = R_s+is(1+I_s), so

By the lower bound I_s\geq (b_4+M^2)^{-1}, there exists a constant C_M>1 such that 1+I_s\leq C_M I_s for all 0<s\leq M, and consequently sf_\infty(s)\leq C^2_M\bigl[-\operatorname{Im}G_\mu(is)\bigr]. Moreover, note that

Here the finiteness follows from \Gamma(\mu)>-\infty. Thus

Finally,

Therefore

Expansion on the tail. We now consider s\geq M, and set A_n(s):=s^{-2}+n^{-1}. Since m_2(\nu)=b_4,

Hence

where the convergence follows from the dominated convergence theorem.

Similarly,

By |x|/(x^2+s^2)\leq 1/(2s), we obtain

Set

Then I_s+n^{-1}=A_n(s)+d_s, and (Proof) gives

Let

Since \frac{n}{n+s^2}=\frac{s^{-2}}{A_n(s)}, we have the exact identity

Write

By (?),

Substituting

into (?), we obtain

Therefore, if we define

then

Estimate of the remainder. From (?) and (?), there exists some constant C_\mu>0 such that

Since A_n(s)\geq s^{-2}, for M sufficiently large,

Hence

for some constants c,C>0 depending only on \mu. Moreover,

so that

Consequently,

For fixed M, let

By (?), \theta_M\to0 as M\to\infty. Using the previous estimates and |d_s|\leq C_\mu s^{-4}, we obtain

With the change of variable u=s^{-2}, we have

and it follows that

The first, second, fourth and fifth integrals are uniformly bounded in n, while

Thus

Integration of the leading terms. From (?), the two leading contributions are

and

where y=s^2. Using the elementary decompositions

and

we obtain

and

Combining (?), (?), (?), and (?), we deduce that

Finally, letting M\to\infty and using \theta_M\to0, we conclude that

as claimed.

Let \rho=\mu^{(2)}. Then \rho is supported on \mathbb{R}_+, \int x\,d\rho(x)=1, and

For any s>0, set

Jensen’s inequality gives

so 0<r(s)\le1 and 0\le q(s)<1. By (?),

Hence, for any s>0,

Applying the change of variables u = ns to the identity \int_0^\infty (s+1)^{-2} \, ds = 1, and setting b_n(u) \coloneqq 1 + u/n, we have

Substituting the identity in the integral representation of \Psi(\mu_n) in Lemma 2.1 yields

where H_n(u) \coloneqq (b_n(u) - q(u))^{-2} - b_n(u)^{-2}. A straightforward calculation gives

Recall that b_4=m_4(\mu)-1. We first establish the following asymptotic behavior of q:

Indeed, a rearrangement gives

and by \int(x-1)d\rho(x)=0, we have

The dominated convergence theorem yields (Proof). If b_4=0, then \rho=\delta_1 and symmetry forces \mu=\mathbf{b}, so the conclusion is immediate. Assume henceforth b_4>0, and it remains to prove

We split the integral into a bounded part, an intermediate bulk, and a tail.

Bounded region. From (Proof), 0\le H_n(u)\le r(u)^{-2}. Moreover, r(s)/s=(sh(s))^{-1}-1 is decreasing in s, since sh(s) is increasing in s. Thus for any 0<u\le M,

Since

we obtain

Bulk region. Fix \eta\in(0,b_4). By (Proof), there exists some sufficiently large M such that

For M\le u\le\delta n, write

Because \Phi(1,0)=2, one may first choose M large and then \delta>0 small enough that

uniformly on [M,\delta n]. Hence

Tail. Again by (Proof), there exists C>0 such that q(u)\le C/u for u\ge M. For fixed \delta>0 and sufficiently large n, u\ge\delta n implies q(u)\le1/2, and hence

Therefore

Dividing (?), (?), and (?) by n, then letting n\to\infty and \eta\downarrow0, yields (Proof). Substitution into (Proof) gives

as claimed.