A quadratic ancestor inequality for spanning trees

By (?), for every x,y\in U_B,

Furthermore, \partial_x^2 Q_B(z)=0 as z_x occurs in at most one factor in each product in (?). Hence, by Proposition 2.2, we obtain that -\nabla^2\log Q_B(0)=H_BH_B^{\mathsf T}-J_B is positive semidefinite.

For x,y\in U_B, the event \{\operatorname{LCA}_T(x,y)\in B\} implies that x,y,r lie in three different components of T\setminus B. Thus a_B(x,y)\le J_B(x,y), and (1.1) follows for f supported on U_B. Taking two-by-two principal minors of H_BH_B^{\mathsf T}-J_B gives

and hence J_B(x,y)\le2H_B(x)H_B(y).

It remains to remove the assumption that f is supported on U_B. Terms involving r vanish, while H_B(x)=1 for x\in B and a_B(x,y)\le H_B(y) whenever x\in B. Put a=\sum_{x\in B}f(x) and b=\sum_{x\in U_B}f(x)H_B(x). Decomposing the left side of (1.1) according to whether neither, exactly one, or both of x,y lie in B, the three contributions are bounded by b^2, 2ab, and a^2, respectively. This proves (1.1) for arbitrary nonnegative f, and also (1.1).

Fix B\subseteq V, and let U,W be two independent uniform vertices of V, independent of \mathfrak F_m. Applying (1.1) with root \partial and f=\mathbf 1_V/n yields

A spanning tree of the augmented graph containing j cemetery edges has weight proportional to m^j. As m\downarrow0, its law therefore converges to a weighted spanning tree T of G with an independent uniform vertex R joined to \partial. The joint law of (T,U,W,R) is invariant under the transitive automorphism group, so \operatorname{LCA}_R(U,W) is uniform on V. Taking m\downarrow0 in (Proof) gives

which proves (1.2). For a partition \mathcal P=\{B_1,\ldots,B_k\} of V, we then have \mathbb E N_{\mathcal P} =\sum_{i=1}^k\mathbb P(\gamma\cap B_i\ne\varnothing) \ge\sum_{i=1}^k\sqrt{|B_i|/n}; this implies (1.2). Taking \mathcal P to be the partition into singletons gives \mathbb E|\gamma|\ge\sqrt n. Since |\gamma|=d_T(X,Y)+1, we obtain (1.2).

We work with continuous-time simple random walk with unit jump rate. We first obtain a local connectivity estimate at killing rate R^{-2}, and then extend it to all 0<m\le R^{-2} by monotonicity in the killing rate.

Local connectivity estimate. We first prove a second-moment estimate on any transitive weighted undirected graph with \sum_y c_{xy}=1 for every x\in V, which will also be used in Section ?. Let p_t(x,y) be its transition kernel, fix an origin o, and put

Note that

For independent walks X,Y started at x,y, with independent exponential lifetimes \zeta_X,\zeta_Y of rate m, define

Writing g=g_m, we have

The last inequality follows from Cauchy–Schwarz inequality and that ab\le(a^2+b^2)/2. Wilson’s algorithm and the inequality of [LPS, Lemma 1.2 and Remark 1.3] imply

Now on \mathbb Z^3, set \mu=R^{-2}, and take x,y\in B_R. The heat-kernel bounds give

Here, the first estimate follows by integrating over t\in[R^2,2R^2], and the second follows from p_t(0,0)\le C(1+t)^{-3/2} and (?); see [LL, Theorem 2.1.3]. Combining (?) and (?) gives \mathbb E I_{x,y}\ge cR and \mathbb E I_{x,y}^2\le CR^2. Hence the second-moment inequality gives \mathbb P(I_{x,y}>0)\ge c, and (?) gives \inf_{x,y\in B_R}\mathbb P_\mu(y\in C_\mu(x)) \ge a_0>0, where a_0 is independent of R.

Write x\longleftrightarrow_D y if x and y are connected in \mathfrak F_m and their unique path in \mathfrak F_m is contained in D. Doob’s inequality applied up to the exponential lifetime \zeta_X gives

If both X[0,\zeta_X) and Y[0,\zeta_Y) stay in B_{KR}, and Y hits \operatorname{LE}(X[0,\zeta_X)) before \zeta_Y, then Wilson’s algorithm gives x\longleftrightarrow_{B_{KR}}y. Choose K>1 sufficiently large such that 2C/(K-1)^2\le a_0/2. Then with a=a_0/2, we obtain

Removing the killing. For 0<m\le\mu, the massive spanning forests can be coupled so that \mathfrak F_\mu\subseteq\mathfrak F_m. To see this, orient the edges arbitrarily and define \nabla:\ell^2(E)\to\ell^2(V) by \nabla\delta_e=\sqrt{c_e}(\delta_{e^+}-\delta_{e^-}). By the transfer-current theorem [LP, Section 4.2], the edge set of \mathfrak F_m is determinantal with kernel \mathcal K_m=\nabla^*(\nabla\nabla^*+mI)^{-1}\nabla. The same formula holds in infinite volume by exhaustion; see e.g. [LP, Section 10.1]. Now monotonicity of the resolvent gives 0\preceq \mathcal{K}_\mu\preceq \mathcal{K}_m\preceq I, and the determinantal domination theorem (see [Lyons, Theorem 2.9]) therefore couples the spatial edge sets with \mathfrak F_\mu\subseteq\mathfrak F_m. Consequently, (?) holds for every 0<m\le\mu as the event \{x\longleftrightarrow_{B_{KR}}y\} there is increasing.

On this event, \operatorname{LCA}(x,y) lies on their connecting path, hence in B_{KR}. Put

We apply (1.1) with f=\mathbf 1_{B_R} and B=\{v\} to each v\in B_{KR}, and then sum over v. Translation invariance then gives

Thus A_m((K+1)R)\ge cR^{3/2}, uniformly for 0<m\le R^{-2}. By transience, for every fixed u, the indicator \mathbf1_{\{u\in\operatorname{LE}(S[0,t])\}} converges to \mathbf1_{\{u\in\eta\}} as t\to\infty. Consequently, h_m(u)\to\mathbb P(u\in\eta) as m\downarrow0. Taking m\downarrow0 in the finite sum A_m((K+1)R) and rescaling R by the fixed factor K+1 proves (1.3) after adjusting the constant c.

The growth exponent. For independent simple random walks S^1,S^2 started at 0, define

where \tau_r^i is the first exit time of S^i from B_r. By [LS, Corollary 1.3], we have

The finite-volume one-point estimate in [ShiG, Proposition 8.2], combined with transience and the Green-function bound \sum_{j\ge0}\mathbb P_0(S_j=z)\le C/(1+|z|) [LL, Theorem 4.3.1], implies

Summing over lattice thus yields \mathbb E|\eta\cap B_R|\le CR^\beta. Combined with (1.3) gives \beta\ge3/2.

The complete log-concavity of the first polynomial is proved in [ALOV, Theorem 4.1]; the same property for both polynomials follows from [BH, Proposition 4.25 and Theorem 2.30]. The closure properties follow from [BH, Theorem 2.10 and Corollary 2.32].

Put U=V\setminus B, k=|V|-1, and E_U=E(G[U]). Add a new vertex * and the star edges M=\{s_u=\{*,u\}:u\in U\}. On E\sqcup M, take the direct sum \mathsf N of the graphic matroid of G[U]+M and the free matroid on E\setminus E_U. Set N=|E|+|U|.

Introduce variables t_e for e\in E, variables z_u for u\in U, and a homogenizing variable s. The polynomial

is homogeneous completely log-concave by Theorem 2.1 after the change of variables t_e\mapsto c_et_e. So is the basis generating polynomial of the dual graphic matroid

Extracting the coefficient of \prod_{e\in E}t_e gives

The closure properties above show that F is also homogeneous completely log-concave, of degree D=N-k. For each spanning tree T, a subset S\subseteq M can be added to T\cap E_U without creating a cycle precisely when S contains at most one edge s_u with u in each component of T\setminus B. Consequently,

All exponents in (?) are nonnegative, since D\ge |U|\ge|\mathcal C_B(T)|.

Differentiate F(s,z) with respect to z_r and then set s=1. Since only the factor belonging to C_r(T) is differentiated, we have

Note that partial differentiation preserves homogeneous complete log-concavity, and specialization at s=1 preserves complete log-concavity in the remaining variables. We conclude.

Put q=3-\beta and \kappa=3-2q=2\beta-3\ge0. By (?) and (?), \mathbb P(z\in\eta_x)\le C(1+|z-x|)^{-q}. For n\ge1, let

and let \sigma_n^x be the first exit time of \eta_x from B(x,n):=x+B_n.

Choose n_0\ge10^5(1+|a|+|b|) and set n_j=2^jn_0. The sets Q_j:=D_{n_j}(a)\cap D_{n_j}(b) are pairwise disjoint, satisfy |Q_j|\asymp n_j^3, and lie at distance comparable to n_j from both a and b. Define

By [ACHS, Proposition 3.16], applied with K=\varnothing, for z\ne w in D_n(x),

Independence therefore gives \mathbb E Z_j\asymp n_j^\kappa and \mathbb E Z_j^2 \le Cn_j^\kappa \left(1+\sum_{r=1}^{n_j}r^{2-2q}\right) \le Cn_j^{2\kappa}(1+\log n_j), where the last inequality uses 2-2q=\kappa-1 and \kappa\ge0.

For i<j, z\in Q_i, w\in Q_j, and x\in\{a,b\}, we have

Applying [ShiG, Proposition 8.2, case (I), k=2] in B(x,N) and letting N\to\infty yield

Consequently, \mathbb E(Z_iZ_j)\le Cn_i^\kappa n_j^\kappa.

Set Y_j=n_j^{-\kappa}Z_j and W_K=\sum_{j=1}^K Y_j. The preceding bounds imply

Since n_j=2^jn_0, we have

Thus, for K\ge1+\log n_0, the second-moment inequality gives \mathbb P(W_K\ge cK)\ge c_0>0, with constants independent of a,b. Since \kappa\ge0 and the sets Q_j are disjoint, we have W_K\le\sum_{j=1}^K Z_j \le|\eta_a\cap\eta_b|. Letting K\to\infty yields

Finally, note that three-dimensional simple random walk has infinitely many cut times a.s. [LawlerCut, Theorem 1.2]. Consequently, removing any finite initial segment of an infinite random walk changes its loop-erasure by only finitely many vertices. Let A=\{|\eta_a\cap\eta_b|=\infty\}, and let \mathcal F_n=\sigma(S_j^a,S_j^b:0\le j\le n), where S^a,S^b are the underlying independent random walks. The Markov property and (?) imply \mathbb P(A\mid\mathcal F_n) =\mathbb P\bigl( |\eta_{S_n^a}\cap\eta_{S_n^b}|=\infty \bigr) \ge c_0. By martingale convergence, \mathbb P(A\mid\mathcal F_n)\to\mathbf 1_A a.s. Therefore, \mathbb P(A)=1.

We first prove \chi(m)\le L(m)^2. For each v\in V and each finite set F\subseteq V, apply (1.1) with B=\{v\} and f=\mathbf 1_F. Letting F\uparrow V and using monotone convergence gives

Applying (?) to F(x,v)=\mathbb P_m(v\in\gamma_x^m) and to F(x,v)=\sum_y\mathbb P_m(\operatorname{LCA}(x,y)=v) gives

By Tonelli’s theorem, both identities hold in [0,\infty]. It follows that \chi(m)\le L(m)^2. The expected number of jumps before killing is m^{-1}, so L(m)\le1+m^{-1}.

For the lower bound, write I_y=I_{o,y} in (?). Summing (?)–(?) over y, using (?) and transitivity implies \sum_y\mathbb E I_y=m^{-2} and \sum_y\mathbb E I_y^2=4m^{-2}\mathcal B(m). Cauchy–Schwarz inequality, first on finite sets and then by exhaustion, gives

Summing (?) yields the lower bound on \chi(m). The lower bound on L(m) follows from \chi(m)\le L(m)^2.

For r>0, let X^{(r)} be the continuous-time random walk started at 0 with jump rates q_r(z)=q(z)\mathbf1_{\{|z|\le r\}}, killed at an independent exponential time \zeta of rate m_r=\sum_{|z|>r}q(z)\asymp r^{-\alpha}. Write L_r=|\operatorname{LE}(X^{(r)}[0,\zeta))|. Then the corresponding \mathcal B(m_r) is bounded by a constant multiple of that for S. By Proposition 5.2 and the heat-kernel estimates [BL, Theorem 1.1], \mathbb E L_r is thus at least c r^{3/2} for some c>0.

We claim that restricting to random walk trajectories contained in B_{Ar\log r} preserves this lower bound for sufficiently large A. Indeed, an exponential supermartingale argument (cf. [BL, Proposition 2.5]) gives

Since the number N_\zeta of jumps before killing is geometric with parameter m_r, we have \mathbb E L_r^2\le\mathbb E(N_\zeta+1)^2 \le Cr^{2\alpha}. Cauchy–Schwarz inequality then implies that trajectories leaving B_{Ar\log r} contribute o(r^{3/2}) to \mathbb E L_r for sufficiently large A.

We view S as a continuous-time random walk with unit jump rate. Its jumps longer than r occur at rate m_r, independently of the shorter jumps, so its trajectory before the first such jump has the law of X^{(r)}[0,\zeta). This jump exceeds 2R with probability m_{2R}/m_r. If the preceding trajectory stays in B_{R/2}, the jump exits B_R without erasing its loop-erasure. Hence, by taking r=\lfloor R/(4A\log R)\rfloor, we obtain

where the last inequality follows from the previous paragraph. Finally, since \mathbb E|\operatorname{LE}(S[0,\tau_R])|\le1+\mathbb E\tau_R\le CR^\alpha, we conclude.