betaFeedback

WorksMathematics

A proof of the Lyons-Pemantle-Peres monotonicity conjecture for high offspring laws

Abstract

We prove the Lyons-Pemantle-Peres conjecture for high enough offspring laws, that is we show that the speed satisfies ℓ′(λ)<0\ell'(λ) < 0 for the entire transient regime of the bias λ∈(0,m)λ\in (0, m) in the special case that the number of children ZZ in the Galton-Watson tree satisfies Z≥4Z \geq 4 almost surely.

Introduction

Let ZZ be a N={1,2,…}\mathbb{N} = \{1,2,\ldots\} valued random variable and denote by T\mathcal{T} the corresponding (leafless) Galton-Watson tree having ZZ as the offspring law at each vertex vv and call the resulting law of T\mathcal{T} as P\mathbb{P}. On this random environment, we then introduce a random walk (ξkλ)k∈N0(\xi_k^\lambda)_{k\in\mathbb{N}_0} starting at the root ϱ\varrho of T\mathcal{T}, and which follows the following transition probabilities for each vertex vv of T\mathcal{T} which is not the root ϱ\varrho:

Pλ,T(ξk+1λ=parent⁡(v)∣ξkλ=v)=λλ+Z(v)P_{\lambda,\mathcal{T}}(\xi_{k+1}^\lambda=\operatorname{parent}(v)\mid\xi_k^\lambda=v)=\frac{\lambda}{\lambda+Z(v)}
Pλ,T(ξk+1λ=child⁡i(v)∣ξkλ=v)=1λ+Z(v),i=1,…,Z(v)P_{\lambda,\mathcal{T}}(\xi_{k+1}^\lambda=\operatorname{child}_i(v)\mid\xi_k^\lambda=v)=\frac{1}{\lambda+Z(v)},\qquad i=1,\ldots,Z(v)

as well as for v=ϱv=\varrho

Pλ,T(ξk+1λ=child⁡i(ϱ)∣ξkλ=ϱ)=1Z(ϱ).P_{\lambda,\mathcal{T}}(\xi_{k+1}^\lambda=\operatorname{child}_i(\varrho)\mid\xi_k^\lambda=\varrho)=\frac{1}{Z(\varrho)}.

Here λ∈R≥0\lambda\in\mathbb{R}_{\ge0} is called the bias parameter, indicating the bias to return to the parent nodes during the random walk. Going from the quenched law Pλ,TP_{\lambda,\mathcal{T}}, we write for the annealed probability measure

(P∘(T,ξλ)−1)(dτ,dx)=P(dτ)Pλ,τ(dx).(\mathbb{P}\circ(\mathcal{T},\xi^\lambda)^{-1})(d\tau,dx)=\mathbb{P}(d\tau)P_{\lambda,\tau}(dx).

Here we model everything on one probability space (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}) furnishing T,ξλ\mathcal{T},\xi^\lambda and any additional random variables as needed on the fly. One can prove that for λ∈(0,m)\lambda\in(0,m) with m:=EZm:=\mathbb{E}Z, the asymptotic speed

ℓ(λ):=lim⁡k→∞∣ξkλ∣k\ell(\lambda):=\lim_{k\to\infty}\frac{|\xi_k^\lambda|}{k}

exists and is deterministic P\mathbb{P}-a.s. [5 Theorem 3.1 + Section 4]. Given this context, we are going to prove the following.

Theorem 1.1 Let ZZ be such that Z≥4Z \ge4 almost surely and m=EZ<∞m = \mathbb{E}Z < \infty. Then ℓ:(0,m)→R≥0\ell: (0,m) \to\mathbb{R}_{\ge0} is differentiable and

∀λ∈(0,m):ℓ′(λ)<0.\forall\lambda\in(0,m) : \ell'(\lambda) < 0 .

In recent years variants of Theorem 1.1 have been treated, providing partial resolutions to the Lyons-Pemantle-Peres monotonicity conjecture [6]. We mention for example [AFS11] and [SWX25], where ℓ′(λ)<0\ell'(\lambda) < 0 was shown for λ≤m11160\lambda\le\frac{m_1}{1160} and λ≤m11+1−1m1\lambda\le\frac{m_1}{1+\sqrt{1-\frac{1}{m_1}}} respectively, with m1=min⁡{k≥1:P(Z=k)>0}m_1 = \min\{k \ge1 : \mathbb{P}(Z = k) > 0\}. Our advance is the extension to the entire transient regime λ∈(0,m)\lambda\in(0,m), albeit at the cost of requiring high offspring laws Z≥4Z \ge4. With similar techniques as presented in this note, it is possible to prove Theorem 1.1 for the less restrictive case Z≥2Z \ge2. In the interest of clarity we refrain from doing so here, as this requires just lengthier arguments that are merely technical in nature which nonetheless don’t appear to be sufficient yet to resolve the Z≥1Z \ge1 case.

To comment on the main technical concept appearing in the proof, recall the tree-dependent escape probability

Xλ(T)=P(ξλ never returns to ϱ∗∣T)=Pλ,T(ξ never returns to ϱ∗).X_\lambda(\mathcal{T}) = \mathbb{P}(\xi^\lambda\text{ never returns to } \varrho^* \mid\mathcal{T}) = \mathbb{P}_{\lambda,\mathcal{T}}(\xi\text{ never returns to } \varrho^*) .

Here, ϱ∗\varrho^* is an artificial parent of the root ϱ\varrho, and the random walk (ξkλ)(\xi_k^\lambda) starting at ϱ\varrho on the extended tree T∗\mathcal{T}_* is similarly defined as the ordinary random walk on T\mathcal{T}, resulting though in the same speed ℓ(λ)\ell(\lambda) as in the original random walk [SWX25, Lemma 1] for λ∈(0,m)\lambda\in(0,m). Calling the law of XλX_\lambda under P\mathbb{P} as μλ\mu_\lambda, we recall the nonlinear fixed point equation satisfied by the law μλ\mu_\lambda expressed via XλX_\lambda and ZZ as

Xλ=d∑i=1ZXλ,iλ+∑i=1ZXλ,i,(1)X_\lambda\overset{d}{=} \frac{\sum_{i=1}^{Z} X_{\lambda,i}}{\lambda+\sum_{i=1}^{Z} X_{\lambda,i}} , \tag*{(1)}

where Xλ,iX_{\lambda,i} are i.i.d. with law μλ\mu_\lambda, and equality is in distribution under P\mathbb{P}. This equation (1) is explicitly mentioned for example in [2 Section 13.3.3], but can essentially be traced back to [4, 5]. Whereas [SWX25] now differentiates in a pathwise sense the random variable XλX_\lambda in (1)1, we differentiate the law μλ\mu_\lambda, and thus derive the linearized equation around a given λ\lambda for a measure Γλ\Gamma_\lambda, resulting in an instance of the Dyson equation

Γλ=σλ+TλΓλ.(2)\Gamma_\lambda= \sigma_\lambda+ T_\lambda\Gamma_\lambda. \tag*{(2)}

Let us elaborate more in an informal manner how to arrive at (2). For every test function ff, we can write (1) as an explicit fixed point equation for μλ\mu_\lambda:

∫f(x) dμλ(x)=∑npn∫f(∥x∥1λ+∥x∥1) dμλ⊗n(x)(3)\int f(x)\,\mathrm{d}\mu_\lambda(x) = \sum_n p_n \int f\left(\frac{\lVert x\rVert_1}{\lambda+\lVert x\rVert_1}\right)\,\mathrm{d}\mu_\lambda^{\otimes n}(x) \tag*{(3)}

where pn=P(Z=n)p_n = \mathbb{P}(Z = n) and ∥x∥1=x1+⋯+xn\lVert x\rVert_1 = x_1+\cdots+x_n. We want to study now the first order change to μλ\mu_\lambda when λ↦λ+ε\lambda\mapsto\lambda+\varepsilon, that is, we linearize μλ+ε≈μλ+εμ˙λ\mu_{\lambda+\varepsilon} \approx\mu_\lambda+\varepsilon\dot{\mu}_\lambda. Plugging into (3), to order ε\varepsilon, we get easily the equation satisfied by μ˙λ\dot{\mu}_\lambda:

∫f(x) dμ˙λ=−1λEf′(Xλ)Xλ(1−Xλ)+∑nnpn∫f(∥x∥1λ+∥x∥1) d(μλ⊗(n−1)⊗μ˙λ)(x).(4)\int f(x)\,\mathrm{d}\dot{\mu}_\lambda= -\frac{1}{\lambda}\mathbb{E}f'(X_\lambda)X_\lambda(1-X_\lambda)+\sum_n np_n\int f\left(\frac{\lVert x\rVert_1}{\lambda+\lVert x\rVert_1}\right)\,\mathrm{d}(\mu_\lambda^{\otimes(n-1)}\otimes\dot{\mu}_\lambda)(x). \tag*{(4)}

On the other hand, for the family (μλ)λ∈(0,m)(\mu_\lambda)_{\lambda\in(0,m)}, pretending λ\lambda to be a time parameter, we have a fairly general ‘continuity equation’ of the form

μ˙λ(dx)=∂λμλ(dx)=∂λ[Fλ(x+dx)−Fλ(x)dxdx]=∂x∂λFλ(x)dx=:∂xΓλ(dx).\dot{\mu}_\lambda(\mathrm{d}x)=\partial_\lambda\mu_\lambda(\mathrm{d}x)=\partial_\lambda\left[\frac{F_\lambda(x+\mathrm{d}x)-F_\lambda(x)}{\mathrm{d}x}\mathrm{d}x\right]=\partial_x\partial_\lambda F_\lambda(x)\mathrm{d}x=:\partial_x\Gamma_\lambda(\mathrm{d}x).

where Fλ(x)=μλ((−∞,x])F_\lambda(x)=\mu_\lambda((-\infty,x]). This terminology is chosen, as it resembles the continuity equation from fluid dynamics ∂tϱ+∇vϱ=0\partial_t\varrho+\nabla v\varrho=0, with Γλ\Gamma_\lambda representing the ‘velocity field’ in this context. This is to say, that it is more natural to formulate (4) in terms of Γλ\Gamma_\lambda than with μ˙\dot{\mu}. Hence, with formal integration by parts

∫f(x)μ˙(dx)=∫f(x)∂xΓλ(dx)=−∫f′(x)Γλ(dx),\int f(x)\dot{\mu}(\mathrm{d}x)=\int f(x)\partial_x\Gamma_\lambda(\mathrm{d}x)=-\int f'(x)\Gamma_\lambda(\mathrm{d}x),

we arrive at, by calling the measure arising in the first term of (4) as σλ\sigma_\lambda,

∫f′(x)Γλ(dx)=∫f′(x)σλ(dx)+∑nnpn∫f′(∥x∥1λ+∥x∥1)λ(λ+∥x∥1)2d(μλn−1⊗Γλ)(x).\int f'(x)\Gamma_\lambda(\mathrm{d}x)=\int f'(x)\sigma_\lambda(\mathrm{d}x)+\sum_n np_n\int f'\left(\frac{\lVert x\rVert_1}{\lambda+\lVert x\rVert_1}\right)\frac{\lambda}{(\lambda+\lVert x\rVert_1)^2}\mathrm{d}(\mu_\lambda^{n-1}\otimes\Gamma_\lambda)(x).

As this holds for every ff nice enough, we can read off the operator TλT_\lambda and arrive at (2).

This entire preprocessing step lets us eventually express the derivative ℓ′(λ)\ell'(\lambda) in a way where its sign can be easily read off from evaluations of some bilinear form on explicit measures related to μλ\mu_\lambda and Γλ\Gamma_\lambda.

Notation: We will frequently use Xλ(T)X_\lambda(\mathcal{T}) as the escape probability of the random walk on a given tree T\mathcal{T} starting at the root. We will drop T\mathcal{T} though from the notation. Furthermore, many of the appearing objects Xλ,μλ,Qλ,TλX_\lambda,\mu_\lambda,Q_\lambda,T_\lambda will be dependent on λ\lambda, however for legibility’s sake, inside the proofs we will often drop the subscript λ\lambda.

A formula for the speed ℓ′(λ)\ell'(\lambda)

We first recall from Aidekon [Aïd14] and Rousselin [8] the formula for the speed

ℓ(λ)=AλBλ,(5)\ell(\lambda)=\frac{A_\lambda}{B_\lambda}, \tag*{(5)}

with

Aλ:=EXYDλ(X,Y),Bλ:=EX+Y−XYDλ(X,Y),A_\lambda:=\mathbb{E}\frac{XY}{D_\lambda(X,Y)},\quad B_\lambda:=\mathbb{E}\frac{X+Y-XY}{D_\lambda(X,Y)},

where Dλ(x,y)=λ−(1−x)(1−y)D_\lambda(x,y)=\lambda-(1-x)(1-y). Here X,YX,Y are i.i.d and X∼μλ=Law⁡P(Xλ)X\sim\mu_\lambda=\operatorname{Law}_{\mathbb{P}}(X_\lambda). It turns out that for the technical details appearing later on, it is very convenient to have knowledge of where μλ\mu_\lambda is supported, since this gives control on the kernel DλD_\lambda.

Lemma 2.1 For any offspring law such that Z≥4Z\ge4, it holds that

Xλ≥(1−λ4)+,P-a.s.X_\lambda\ge\left(1-\frac{\lambda}{4}\right)_+,\quad\mathbb{P}\text{-a.s.}

and hence supp⁡μλ⊂[(1−λ4)+,1]\operatorname{supp}\mu_\lambda\subset\left[\left(1-\frac{\lambda}{4}\right)_+,1\right]. As a consequence, we get two-sided bound on the kernel DλD_\lambda when sampling from the distribution μλ\mu_\lambda:

∀x,y∈supp⁡μλ:34λ≤Dλ(x,y)≤λ(6)\forall x,y\in\operatorname{supp}\mu_\lambda:\quad\frac{3}{4}\lambda\le D_\lambda(x,y)\le\lambda \tag*{(6)}

Proof. This is a special case of [SWX25, Lemma 3], applied to m1:=inf⁡{n≥1:P(Z=n)>0}≥4m_1:=\inf\{n\ge1:\mathbb{P}(Z=n)>0\}\ge4, with (6) an immediate computation.

It will be useful to introduce short-hand notations for the following sets of measures

M:={ν∣ν is signed measure on ([0,1],B([0,1]))}\mathcal{M}:=\{\nu\mid\nu\text{ is signed measure on }([0,1],\mathcal{B}([0,1]))\}
Mλ:={ν∈M∣supp⁡ν⊂[(1−λ/4)+,1]}\mathcal{M}_\lambda:=\{\nu\in\mathcal{M}\mid\operatorname{supp}\nu\subset[(1-\lambda/4)_+,1]\}
M+:={ν∈M∣ν≥0}\mathcal{M}^+:=\{\nu\in\mathcal{M}\mid\nu\ge0\}
Mλ+:=Mλ∩M+.\mathcal{M}^+_\lambda:=\mathcal{M}_\lambda\cap\mathcal{M}^+.

The introduction of signed measures in this context is natural, as we recall from the introduction that we study a family of probability measures {μλ}λ\{\mu_\lambda\}_\lambda and their linearization μλ+ε≈μλ+εμ˙\mu_{\lambda+\varepsilon}\approx\mu_\lambda+\varepsilon\dot{\mu}, from which it follows that μ˙([0,1])=0\dot{\mu}([0,1])=0, hence μ˙\dot{\mu} is in general signed.

Definition 2.2 Fix λ∈(0,m)\lambda\in(0,m). We let T=TλT=T_\lambda be a linear operator mapping

Tλ:M→MT_\lambda:\mathcal{M}\to\mathcal{M}
η↦∑n=4∞npn∫1(⋅)(∥y∥1λ+∥y∥1)λ(λ+∥y∥1)2 d(η⊗μλ⊗(n−1))(y),\eta\mapsto\sum_{n=4}^{\infty}np_n\int\mathbf{1}_{(\cdot)}\left(\frac{\lVert y\rVert_1}{\lambda+\lVert y\rVert_1}\right)\frac{\lambda}{(\lambda+\lVert y\rVert_1)^2}\,d\left(\eta\otimes\mu_\lambda^{\otimes(n-1)}\right)(y),

where pn=P(Z=n)p_n=\mathbb{P}(Z=n) and ∥y∥1=y1+⋯+yn\lVert y\rVert_1=y_1+\cdots+y_n.

We note that we can define TλT_\lambda equivalently by testing TληT_\lambda\eta against functions f∈C([0,1])f\in C([0,1]), resulting in the representation

∫f d(Tλη)=∑n=4∞npnE[∫λ(λ+x+Sn−1)2f(x+Sn−1λ+x+Sn−1)η(dx)],\int f\,d(T_\lambda\eta)=\sum_{n=4}^{\infty}np_n\mathbb{E}\left[\int\frac{\lambda}{(\lambda+x+S_{n-1})^2}f\left(\frac{x+S_{n-1}}{\lambda+x+S_{n-1}}\right)\eta(dx)\right],

where Sk=∑i=1kXλ,iS_k=\sum_{i=1}^{k}X_{\lambda,i} with Xλ,iX_{\lambda,i} are i.i.d. with law μλ\mu_\lambda. We record two immediate properties about TT.

Lemma 2.3

  1. TλT_\lambda is a positive linear operator, that is η≥0⇒Tλη≥0\eta\geq0\Rightarrow T_\lambda\eta\geq0. Furthermore, η∈Mλ⇒Tλη∈Mλ\eta\in\mathcal{M}_\lambda\Rightarrow T_\lambda\eta\in\mathcal{M}_\lambda.

  1. TλT_\lambda is bounded in the total variation norm, that is ∥Tλη∥TV≤mλ∥η∥TV\lVert T_\lambda\eta\rVert_{\mathrm{TV}}\leq\frac{m}{\lambda}\lVert\eta\rVert_{\mathrm{TV}} for any signed measure η∈M\eta\in\mathcal{M}.

Proof. The first part follow from straightforward inspection of the definition of TT; boundedness of the total variation norm follows by the characterization ∥ν∥TV=sup⁡f:∣f∣≤1∫f dν\lVert\nu\rVert_{\mathrm{TV}}=\sup_{f:\lvert f\rvert\leq1}\int f\,d\nu.

Eventually we want to differentiate AλA_\lambda and BλB_\lambda with respect to λ\lambda. As XλX_\lambda is dependent on λ\lambda, the prototypical quantity which we need to differentiate is Ef(Xλ)=∫f(x)dμλ(x)\mathbb{E}f(X_\lambda)=\int f(x)d\mu_\lambda(x). We establish first, that in a suitable way the law μλ\mu_\lambda is differentiable with ‘derivative’ Γλ\Gamma_\lambda. More precisely:

Proposition 2.4 Let λ∈(0,m)\lambda\in(0,m).

  1. There exists a unique positive measure Γλ∈Mλ+\Gamma_\lambda\in\mathcal{M}_\lambda^+ satisfying the fixed point equation

Γλ=σλ+TλΓλ,(7)\Gamma_\lambda=\sigma_\lambda+T_\lambda\Gamma_\lambda, \tag*{(7)}

where σλ(dx)=x(1−x)μλ(dx)\sigma_\lambda(dx)=x(1-x)\mu_\lambda(dx).

  1. The measure Γλ\Gamma_\lambda is the derivative of μλ\mu_\lambda in the sense that for all f∈C1[0,1]f\in C^1[0,1], the map λ↦∫f dμλ\lambda\mapsto\int f\,d\mu_\lambda is differentiable and

ddλ∫f(x)dμλ(x)=1λ∫f′(x)dΓλ(x).(8)\frac{d}{d\lambda}\int f(x)d\mu_\lambda(x)=\frac{1}{\lambda}\int f'(x)d\Gamma_\lambda(x). \tag*{(8)}

In order to prove Proposition 2.4 the following bilinear operator will prove useful.

Definition 2.5 For fixed λ∈(0,m)\lambda\in(0,m) we define the mapping

Qλ:Mλ×Mλ→RQ_\lambda:\mathcal{M}_\lambda\times\mathcal{M}_\lambda\to\mathbb{R}
(η,ζ)↦∫λDλ(x,y)2η(dx)ζ(dy).(\eta,\zeta)\mapsto\int\frac{\lambda}{D_\lambda(x,y)^2}\eta(dx)\zeta(dy).

Because of the kernel bounds (6) on DλD_\lambda, QλQ_\lambda is well-defined and it immediately follows that for positive measures η,ζ∈Mλ+\eta,\zeta\in\mathcal{M}_\lambda^+:

1λ∥η∥TV∥ζ∥TV≤Qλ(η,ζ)≤169λ∥η∥TV∥ζ∥TV.(9)\frac{1}{\lambda}\lVert\eta\rVert_{\mathrm{TV}}\lVert\zeta\rVert_{\mathrm{TV}}\leq Q_\lambda(\eta,\zeta)\leq\frac{16}{9\lambda}\lVert\eta\rVert_{\mathrm{TV}}\lVert\zeta\rVert_{\mathrm{TV}}. \tag*{(9)}

We will also need a simple technical lemma about the interplay of QQ and TT and some explicit evaluations of the latter.

Lemma 2.6 It holds

  1. The linear operator TλT_\lambda is symmetric with respect to QλQ_\lambda, that is for any signed measures η,ζ∈Mλ\eta,\zeta\in\mathcal{M}_\lambda

Qλ(Tλη,ζ)=Qλ(η,Tλζ).(10)Q_\lambda(T_\lambda\eta,\zeta)=Q_\lambda(\eta,T_\lambda\zeta). \tag*{(10)}

Introducing the measures αλ,σλ≪μλ\alpha_\lambda, \sigma_\lambda\ll\mu_\lambda as

αλ(dx)=xμλ(dx), σλ(dx)=x(1−x)μλ(dx),\alpha_\lambda(dx)=x\mu_\lambda(dx),\ \sigma_\lambda(dx)=x(1-x)\mu_\lambda(dx),

they satisfy

Tα=σ,(11)T\alpha=\sigma, \tag*{(11)}

Proof. Symmetry of TT with respect to QQ, that is (10), follows directly by plugging in the definitions in each other:

Q(Tη,ζ)=∑nnpnE[∫[∫λ(λ+z+Sn−1)2Dλ(z+Sn−1λ+z+Sn−1,y)η(dz)]ζ(dy)]=∑nnpn∫(E(λ+z+y+Sn−1−1)−2)η(dz)ζ(dy)=Q(η,Tζ).\begin{aligned} Q(T\eta,\zeta)&=\sum_n np_n\mathbb{E}\left[\int\left[\int\frac{\lambda}{(\lambda+z+S_{n-1})^2}D_\lambda\left(\frac{z+S_{n-1}}{\lambda+z+S_{n-1}},y\right)\eta(dz)\right]\zeta(dy)\right]\\ &=\sum_n np_n\int\left(\mathbb{E}(\lambda+z+y+S_{n-1}-1)^{-2}\right)\eta(dz)\zeta(dy)=Q(\eta,T\zeta). \end{aligned}

For (11) this follows as straightforwardly, that is, take any f∈C([0,1])f\in C([0,1]) and compute:

∫f d(Tα)=∑nnpnE[∫λy(λ+y+Sn−1)2f(y+Sn−1λ+y+Sn−1)μ(dy)]=∑npnEλ(nXn)(λ+Sn)2f(Snλ+Sn)=∑npnE∑i=1nλXi(λ+Sn)2f(Snλ+Sn)=∑npnEλλ+SnSnλ+Snf(Snλ+Sn)=∑npnE(1−Snλ+Sn)Snλ+Snf(Snλ+Sn)=E(1−X)Xf(X)=∫f dσ,\begin{aligned} \int f\,d(T\alpha)&=\sum_n np_n\mathbb{E}\left[\int\frac{\lambda y}{(\lambda+y+S_{n-1})^2}f\left(\frac{y+S_{n-1}}{\lambda+y+S_{n-1}}\right)\mu(dy)\right]\\ &=\sum_n p_n\mathbb{E}\frac{\lambda(nX_n)}{(\lambda+S_n)^2}f\left(\frac{S_n}{\lambda+S_n}\right)=\sum_n p_n\mathbb{E}\sum_{i=1}^n\frac{\lambda X_i}{(\lambda+S_n)^2}f\left(\frac{S_n}{\lambda+S_n}\right)\\ &=\sum_n p_n\mathbb{E}\frac{\lambda}{\lambda+S_n}\frac{S_n}{\lambda+S_n}f\left(\frac{S_n}{\lambda+S_n}\right)=\sum_n p_n\mathbb{E}\left(1-\frac{S_n}{\lambda+S_n}\right)\frac{S_n}{\lambda+S_n}f\left(\frac{S_n}{\lambda+S_n}\right)\\ &=\mathbb{E}(1-X)Xf(X)=\int f\,d\sigma, \end{aligned}

where in the penultimate equality we used the fixed point of the law of XX from (1).

This suffices to proof the first part of Proposition 2.4.

Proof of 1. of Proposition 2.4. As usual we split the existence and uniqueness part of the proof.

  1. Existence: We let ΓN:=∑i=0NTiσ\Gamma^N:=\sum_{i=0}^N T^i\sigma and aim to show convergence for N→∞N\to\infty. To this end, first note that as σ≥0\sigma\ge0, by positivity of TT and induction ΓN≥0\Gamma^N\ge0. The trick now is to observer that α−σ\alpha-\sigma is a positive measure due to α−σ=x2μ(dx)≥0\alpha-\sigma=x^2\mu(dx)\ge0 and compute

Q(ΓN,x2μ(dx))=Q(ΓN,α−σ)=∑j=0NQ(Tjσ,α)−Q(Tjσ,σ)=Q(σ,α)−Q(TNσ,σ)≤Q(σ,α),\begin{aligned} Q(\Gamma^N,x^2\mu(dx))&=Q(\Gamma^N,\alpha-\sigma)=\sum_{j=0}^N Q(T^j\sigma,\alpha)-Q(T^j\sigma,\sigma)\\ &=Q(\sigma,\alpha)-Q(T^N\sigma,\sigma)\le Q(\sigma,\alpha), \end{aligned}

where the penultimate equality comes from Q(Tjσ,α)=Q(Tj−1σ,Tα)=Q(Tj−1σ,σ)Q(T^j\sigma,\alpha)=Q(T^{j-1}\sigma,T\alpha)=Q(T^{j-1}\sigma,\sigma) by (10) and (11). Now utilizing the equivalence (9) twice, we get:

∥ΓN∥TV∥α−σ∥TV≤λQ(ΓN,α−σ)≤Q(σ,α)≤169∥σ∥TV∥α∥TV.\|\Gamma^N\|_{\mathrm{TV}}\|\alpha-\sigma\|_{\mathrm{TV}}\le\lambda Q(\Gamma^N,\alpha-\sigma)\le Q(\sigma,\alpha)\le\frac{16}{9}\|\sigma\|_{\mathrm{TV}}\|\alpha\|_{\mathrm{TV}}.

Hence we obtain the uniform in NN bound

∥ΓN∥TV≤169(EXλ)(EXλ−Xλ2)EXλ2<∞(12)\|\Gamma^N\|_{\mathrm{TV}}\le\frac{16}{9}\frac{(\mathbb{E}X_\lambda)(\mathbb{E}X_\lambda-X_\lambda^2)}{\mathbb{E}X_\lambda^2}<\infty \tag*{(12)}

It follows that the sequence of positive measures (ΓN)N∈N(\Gamma^N)_{N\in\mathbb{N}} on [0,1][0,1] is Cauchy in (M,∥⋅∥TV)(\mathcal{M},\|\cdot\|_{\mathrm{TV}}) and we denote the limit by Γ:=lim⁡N→∞ΓN\Gamma:=\lim_{N\to\infty}\Gamma^N. It is immediately seen that Γ∈Mλ+\Gamma\in\mathcal{M}_\lambda^+, and by continuity of TT by Lemma 2.3 we get

TΓ=lim⁡N→∞TΓN=lim⁡N→∞(ΓN+1−σ)=Γ−σ.T\Gamma=\lim_{N\to\infty}T\Gamma^N=\lim_{N\to\infty}(\Gamma^{N+1}-\sigma)=\Gamma-\sigma.
  1. Uniqueness: Let Γ~∈Mλ+\widetilde{\Gamma} \in\mathcal{M}^{+}_{\lambda} another positive measure satisfying the fixed point equation. Then for any N≥1N \ge1 we get

Γ~=(σ+…TNσ)⏟=ΓN+TN+1Γ~⏟≥0≥ΓN.\widetilde{\Gamma} = \underbrace{(\sigma+ \ldots T^{N}\sigma)}_{=\Gamma^{N}} + \underbrace{T^{N+1}\widetilde{\Gamma}}_{\ge0} \ge\Gamma^{N}.

Hence taking limits we get Γ~≥Γ\widetilde{\Gamma} \ge\Gamma. So we apply (9) to Γ~−Γ\widetilde{\Gamma} - \Gamma, α−σ≥0\alpha- \sigma\ge0 and use again symmetry of TT with respect to QQ and Tα=σT\alpha= \sigma:

0≤1λ∥Γ~−Γ∥TV∥α−σ∥TV≤Q(Γ~−Γ,α−σ)=Q(T(Γ~−Γ),α)−Q(Γ~−Γ,σ)=0,0 \le\frac{1}{\lambda}\lVert\widetilde{\Gamma} - \Gamma\rVert_{\mathrm{TV}}\lVert\alpha- \sigma\rVert_{\mathrm{TV}} \le Q(\widetilde{\Gamma} - \Gamma,\alpha- \sigma) = Q(T(\widetilde{\Gamma} - \Gamma),\alpha) - Q(\widetilde{\Gamma} - \Gamma,\sigma) = 0,

where the penultimate equality is by the fact that T(Γ~−Γ)=Γ~−ΓT(\widetilde{\Gamma} - \Gamma) = \widetilde{\Gamma} - \Gamma. Thus we get ∥Γ~−Γ∥TV=0\lVert\widetilde{\Gamma} - \Gamma\rVert_{\mathrm{TV}} = 0.

Remark 2.7

  1. Intuitively any Γ\Gamma satisfying (7) should be of the form Γ=σ+Tσ+T2σ+…\Gamma= \sigma+ T\sigma+ T^{2}\sigma+ \ldots, so one would be tempted to construct Γ\Gamma via some Neumann series ∑kTk\sum_k T^k. Unfortunately, the operator norm of TT is only guaranteed to be bounded by mλ≥1\frac{m}{\lambda} \ge1, which is the reason for the above proof.

  1. One can attach a probabilistic meaning to Γλ\Gamma_{\lambda} in the following way. We formally differentiate (1) to get

∂λXλ=−Sλ(λ+Sλ)2+λ(λ+Sλ)2∑iXλ,i=−1λXλ(1−Xλ)+Xλ(1−Xλ)∑i=1ZXλ,iSλ.\begin{aligned} \partial_{\lambda}X_{\lambda} = -\frac{S_{\lambda}}{(\lambda+ S_{\lambda})^{2}} + \frac{\lambda}{(\lambda+ S_{\lambda})^{2}}\sum_i X_{\lambda,i} \\ &= -\frac{1}{\lambda}X_{\lambda}(1-X_{\lambda}) + X_{\lambda}(1-X_{\lambda})\frac{\sum_{i=1}^{Z}X_{\lambda,i}}{S_{\lambda}}. \end{aligned}

Rewriting this further, while emphasizing the dependence on the vertex vv yields

−λ∂λlog⁡Xλ(v)=(1−Xλ(v))[1+∑i=1Z(−λ∂λlog⁡Xλ(vi))Xλ(vi)Sλ].(13)-\lambda\partial_{\lambda}\log X_{\lambda}(v) = (1-X_{\lambda}(v))\left[1+\frac{\sum_{i=1}^{Z}(-\lambda\partial_{\lambda}\log X_{\lambda}(v_i))X_{\lambda}(v_i)}{S_{\lambda}}\right]. \tag*{(13)}

where viv_i is the ii-th child of vv. It turns out that (13) can be given a meaning in terms of a stopping time τ(v)\tau(v) of an auxiliary random walk on the given Galton Watson tree TT. This auxiliary random walk is defined via transition probabilities at vertex vv as

(a) with probability Xλ(v)X_{\lambda}(v) it becomes extinct.

(b) with probability 1−Xλ(v)1-X_{\lambda}(v) it moves to one of the children of viv_i, each with probability Xλ(vi)Sλ\frac{X_{\lambda}(v_i)}{S_{\lambda}}.

With respect to this construction, let τ(v)\tau(v) be the stopping time until extinction of a random walk started at the vertex vv. We can immediately see by a first step analysis that

E[τ(v)∣T]=(1−Xλ)(1+∑i=1ZXλ(vi)Sλ(v)E[τ(vi)∣T]).\mathbb{E}[\tau(v)\mid T] = (1-X_{\lambda})\left(1+\sum_{i=1}^{Z}\frac{X_{\lambda}(v_i)}{S_{\lambda}(v)}\mathbb{E}[\tau(v_i)\mid T]\right).

Hence this suggests, assuming uniqueness of solutions, that

E[τ(v)∣T]=−λ∂λlog⁡Xλ(v).(14)\mathbb{E}[\tau(v)\mid T] = -\lambda\partial_{\lambda}\log X_{\lambda}(v). \tag*{(14)}

To connect this now to a representation of Γλ\Gamma_{\lambda}, we furthermore make a formal calculation

−1λ∫f′(x)Γλ(dx)=∂λEf(Xλ)=E[f′(Xλ)∂λXλ]=∫f′(x)E[∂λXλδXλ(dx)].-\frac{1}{\lambda}\int f'(x)\Gamma_{\lambda}(dx) = \partial_{\lambda}\mathbb{E}f(X_{\lambda}) = \mathbb{E}[f'(X_{\lambda})\partial_{\lambda}X_{\lambda}] = \int f'(x)\mathbb{E}[\partial_{\lambda}X_{\lambda}\delta_{X_{\lambda}}(dx)].

from which we are able to identify

Γλ(dx)=−λE[∂λXλδXλ(dx)].\Gamma_{\lambda}(dx) = -\lambda\mathbb{E}[\partial_{\lambda}X_{\lambda}\delta_{X_{\lambda}}(dx)].

Substituting (14) we arrive at:

Γλ(dx)=E[Xλτ(v)δXλ(dx)].(15)\Gamma_{\lambda}(dx) = \mathbb{E}[X_{\lambda}\tau(v)\delta_{X_{\lambda}}(dx)]. \tag*{(15)}
  1. Starting from (15), we can also make sense of the Dyson series in a probabilistic manner. To this end note that by the layer-cake representation

Γλ(dx)=EXλτ(ϱ)δXλ(dx)=∑n=1∞EXλ1τ≥nδXλ(dx)=∑n=1∞E[XλE[1τ≥n∣T]δXλ(dx)].\Gamma_{\lambda}(dx) = \mathbb{E}X_{\lambda}\tau(\varrho)\delta_{X_{\lambda}}(dx) = \sum_{n=1}^{\infty}\mathbb{E}X_{\lambda}\mathbf{1}_{\tau\ge n}\delta_{X_{\lambda}}(dx) = \sum_{n=1}^{\infty}\mathbb{E}[X_{\lambda}\mathbb{E}[\mathbf{1}_{\tau\ge n}\mid\mathcal{T}]\delta_{X_{\lambda}}(dx)].

Let us compute now explicitly

(a) n=1n=1: It is immediate that by definition of the auxiliary process P(τ(ϱ)≥1∣T)=(1−Xλ(ϱ))\mathbb{P}(\tau(\varrho)\ge1\mid\mathcal{T})=(1-X_{\lambda}(\varrho)). Hence

E[XλE[1τ≥1∣T]δXλ(dx)]=E[Xλ(1−Xλ)δXλ(dx)]=x(1−x)μλ(dx)=σλ(dx).\mathbb{E}[X_{\lambda}\mathbb{E}[\mathbf{1}_{\tau\ge1}\mid\mathcal{T}]\delta_{X_{\lambda}}(dx)] = \mathbb{E}[X_{\lambda}(1-X_{\lambda})\delta_{X_{\lambda}}(dx)] = x(1-x)\mu_{\lambda}(dx) = \sigma_{\lambda}(dx).

(b) n=2n=2: Again, by inspection of the auxiliary random walk, we see the probability that the walk doesn’t get extinct for at least two steps is

P(τ≥2∣T)=(1−Xλ(ϱ))[∑iXλ(ϱi)Sλ(ϱ)(1−Xλ(ϱi))].\mathbb{P}(\tau\ge2\mid\mathcal{T})=(1-X_{\lambda}(\varrho))\left[\sum_i\frac{X_{\lambda}(\varrho_i)}{S_{\lambda}(\varrho)}(1-X_{\lambda}(\varrho_i))\right].

And so

E[XλE[1τ≥2∣T]δXλ(dx)]=E[λ(λ+Sλ(ϱ))2(∑iXλ(ϱi)(1−Xλ(ϱi)))δXλ(dx)].\mathbb{E}[X_{\lambda}\mathbb{E}[\mathbf{1}_{\tau\ge2}\mid\mathcal{T}]\delta_{X_{\lambda}}(dx)] = \mathbb{E}\left[\frac{\lambda}{(\lambda+S_{\lambda}(\varrho))^2}\left(\sum_i X_{\lambda}(\varrho_i)(1-X_{\lambda}(\varrho_i))\right)\delta_{X_{\lambda}}(dx)\right].

Testing against some function ff, we see

∫f(x)E[XλE[1τ≥2∣T]δXλ(dx)]=∑nnpn∫λ(λ+∥x∥1)2f(∥x∥1λ+∥x∥1)d(μλn−1⊗σλ)(x)=∫fd(Tλσλ).\begin{aligned} \int f(x)\mathbb{E}[X_{\lambda}\mathbb{E}[\mathbf{1}_{\tau\ge2}\mid\mathcal{T}]\delta_{X_{\lambda}}(dx)] = \sum_n np_n\int\frac{\lambda}{(\lambda+\lVert x\rVert_1)^2}f\left(\frac{\lVert x\rVert_1}{\lambda+\lVert x\rVert_1}\right)d(\mu_{\lambda}^{n-1}\otimes\sigma_{\lambda})(x) \\ &= \int f d(\mathcal{T}_{\lambda}\sigma_{\lambda}). \end{aligned}

And generally for n∈Nn\in\mathbb{N}

E[XλE[1τ≥n∣T]δXλ(dx)]=Tλn−1σλ(dx).\mathbb{E}[X_{\lambda}\mathbb{E}[\mathbf{1}_{\tau\ge n}\mid\mathcal{T}]\delta_{X_{\lambda}}(dx)] = \mathcal{T}_{\lambda}^{n-1}\sigma_{\lambda}(dx).

Thus we have rederived the Dyson series Γ=σ+Tσ+⋯\Gamma=\sigma+\mathcal{T}\sigma+\cdots in a purely probabilistic manner. It is also worthwhile to note that the analytic convergence question of ΓN\Gamma^{N} can be answered (with a bit more work to make the previous steps fully rigorous) by the probabilistic statement of (the weighted) integrability of the existence time τ\tau of the auxiliary process:

∥Γλ∥TV=E[Xλ(ϱ)τ(ϱ)]<∞.\lVert\Gamma_{\lambda}\rVert_{\mathrm{TV}}=\mathbb{E}[X_{\lambda}(\varrho)\tau(\varrho)]<\infty.

To prove the second part, we will need a technical lemma showing a certain Lipschitz continuity in law:

Lemma 2.8 For any u,v≥0u,v\ge0 such that [u,v]⊂(0,m)[u,v]\subset(0,m) there exists a constant Cu,v>0C_{u,v}>0 such that

E[Xa−Xb]≤Cu,v(b−a)\mathbb{E}[X_a-X_b]\le C_{u,v}(b-a)

uniformly in a,ba,b with u≤a<b≤vu\le a<b\le v.

Proof. See Lemma A.1.

Proof of 2. of Proposition 2.4. Let f∈C1([0,1])f\in C^1([0,1]). For a,b∈(0,m)a,b\in(0,m) with a<ba<b one gets using Xa≥XbX_a\ge X_b (pathwise for every fixed tree by an application of Rayleigh’s monotonicity principle [7 Section 2.4]), that

f(Xb)−f(Xa)=−∫[0,1]1[Xb,Xa](y)f′(y) dy,f(X_b)-f(X_a)=-\int_{[0,1]}\mathbf{1}_{[X_b,X_a]}(y)f'(y)\,dy,

and hence by taking the expectation we get the finite difference equation:

∫f dμb−∫f dμab−a=−1a(∫f′(y)ab−aE[1[Xb,Xa](y)] dy)=−1a∫f′(y)Ra,b(dy).\frac{\int f\,d\mu_b-\int f\,d\mu_a}{b-a}=-\frac{1}{a}\left(\int f'(y)\frac{a}{b-a}\mathbb{E}[\mathbf{1}_{[X_b,X_a]}(y)]\,dy\right)=-\frac{1}{a}\int f'(y)R_{a,b}(dy).

The rest is now rigorously establishing that we can take the limit as b→ab\to a in the following steps.

  1. The map λ↦Ef(Xλ),(0,m)→R\lambda\mapsto\mathbb{E}f(X_\lambda),(0,m) \to\mathbb{R} is locally Lipschitz: Take some u,vu,v with [u,v]⊂(0,m)[u,v] \subset(0,m), then for any a,ba,b with u≤a<b≤vu \le a < b \le v we have

∣Ef(Xb)−Ef(Xa)∣≤Lip⁡(f)E∣Xb−Xa∣≤Lip⁡(f)Cu,v(b−a),\left|\mathbb{E}f(X_b)-\mathbb{E}f(X_a)\right| \le\operatorname{Lip}(f)\mathbb{E}|X_b-X_a| \le\operatorname{Lip}(f)C_{u,v}(b-a),

by Lemma 2.8. Hence, by Rademacher’s theorem, for Lebesgue almost every λ∈(0,m)\lambda\in(0,m) the map λ↦Ef(Xλ)\lambda\mapsto\mathbb{E}f(X_\lambda) is differentiable. By the usual arguments, for Lebesgue almost every a∈(0,m)a \in(0,m), for every rational coefficient polynomial ff the following exists

ddλEf(Xλ)∣λ=a∈R.\left.\frac{\mathrm{d}}{\mathrm{d}\lambda}\mathbb{E}f(X_\lambda)\right|_{\lambda=a} \in\mathbb{R}.

For further readability, call PQ⊂C([0,1])\mathcal{P}_{\mathbb{Q}} \subset C([0,1]) the countable set of rational coefficient polynomials and N⊂(0,m)\mathcal{N} \subset(0,m) the exceptional set of Lebesgue measure zero.

  1. Take any a∈(0,m)\Na \in(0,m)\backslash\mathcal{N} and choose a sequence bn↓ab_n \downarrow a. Looking at the positive measure Ra,bR_{a,b}, we get immediately by Lemma 2.8

∥Ra,bn∥TV=Ra,bn([0,1])=aEXa−EXbnbn−a≤aCa,b0\|R_{a,b_n}\|_{\mathrm{TV}}=R_{a,b_n}([0,1])=a\frac{\mathbb{E}X_a-\mathbb{E}X_{b_n}}{b_n-a}\le aC_{a,b_0}

Hence the family (Ra,bn)n⊂M+(R_{a,b_n})_n \subset\mathcal{M}^{+} is tight, and so we can extract a subsequential weak limit which we call

Γ~a:=w−limk→∞Ra,bnk.\widetilde{\Gamma}_a := \mathop{\mathrm{w-lim}}_{k\to\infty}R_{a,b_{n_k}}.
  1. We can now identify

−1a∫f′(y)Γ~a(dy)=−1alim⁡k→∞∫f′(y)Ra,bnk(dy)=lim⁡k→∞EXbnk−EXab−a=ddλEf(Xλ)∣λ=a.-\frac{1}{a}\int f'(y)\widetilde{\Gamma}_a(\mathrm{d}y)=-\frac{1}{a}\lim_{k\to\infty}\int f'(y)R_{a,b_{n_k}}(\mathrm{d}y)=\lim_{k\to\infty}\frac{\mathbb{E}X_{b_{n_k}}-\mathbb{E}X_a}{b-a}=\left.\frac{\mathrm{d}}{\mathrm{d}\lambda}\mathbb{E}f(X_\lambda)\right|_{\lambda=a}.
  1. Moving on to show the fixed point equation, we again note that Xλ=∑i=1ZXλ,iλ+∑i=1ZXλ,iX_\lambda=\dfrac{\sum_{i=1}^{Z}X_{\lambda,i}}{\lambda+\sum_{i=1}^{Z}X_{\lambda,i}} in distribution, that is, for any λ∈(0,m)\lambda\in(0,m) and f∈PQf \in\mathcal{P}_{\mathbb{Q}} we have

∫f(x) μλ(x)=Ef(Xλ)=Ef(Sλλ+Sλ)=∑n=4∞pn∫φλ,n dμλ⊗n.\int f(x)\,\mu_\lambda(x)=\mathbb{E}f(X_\lambda)=\mathbb{E}f\left(\frac{S_\lambda}{\lambda+S_\lambda}\right)=\sum_{n=4}^{\infty}p_n\int\varphi_{\lambda,n}\,\mathrm{d}\mu_\lambda^{\otimes n}.

where pn=P(Z=n)p_n=\mathbb{P}(Z=n) and φλ,n:y↦f(∥y∥1λ+∥y∥1)\varphi_{\lambda,n}:y\mapsto f\left(\dfrac{\|y\|_1}{\lambda+\|y\|_1}\right). Here we can write again ∥y∥1=y1+⋯+yn\|y\|_1=y_1+\cdots+y_n since y∈[0,1]ny\in[0,1]^n. We can differentiate this fixed point equation at λ∈(0,m)\N\lambda\in(0,m)\backslash\mathcal{N} to get

−1λ∫f′(x) dΓλ(x)=∑npn[∫∂λφλ,n dμλ⊗n−nλ∫∂1φλ,n d(Γ~λ⊗μλ⊗(n−1))].-\frac{1}{\lambda}\int f'(x)\,\mathrm{d}\Gamma_\lambda(x)=\sum_n p_n\left[\int\partial_\lambda\varphi_{\lambda,n}\,\mathrm{d}\mu_\lambda^{\otimes n}-\frac{n}{\lambda}\int\partial_1\varphi_{\lambda,n}\,\mathrm{d}\left(\widetilde{\Gamma}_\lambda\otimes\mu_\lambda^{\otimes(n-1)}\right)\right].

Here, to arrive rigorously at the right-hand-side we employed Lemma A.2. Now, since

∂λφλ,n(y)=f′(∥y∥1λ+∥y∥1)(−∥y∥1(λ+∥y∥1)2)\partial_\lambda\varphi_{\lambda,n}(y)=f'\left(\frac{\|y\|_1}{\lambda+\|y\|_1}\right)\left(-\frac{\|y\|_1}{(\lambda+\|y\|_1)^2}\right)
∂1φλ,n(y)=f′(∥y∥1λ+∥y∥1)λ(λ+∥y∥1)2,\partial_1\varphi_{\lambda,n}(y)=f'\left(\frac{\|y\|_1}{\lambda+\|y\|_1}\right)\frac{\lambda}{(\lambda+\|y\|_1)^2},

the first term can be seen to be

∫∂λφλ,n dμλ⊗n=−1λ∫f′(∥y∥1λ+∥y∥1)∥y∥1λ+∥y∥1(1−∥y∥1λ+∥y∥1) dμλ⊗n(y)=−1λE[f′(X)X(1−X)∣Z=n],\int\partial_\lambda\varphi_{\lambda,n}\,\mathrm{d}\mu_\lambda^{\otimes n}=-\frac{1}{\lambda}\int f'\left(\frac{\|y\|_1}{\lambda+\|y\|_1}\right)\frac{\|y\|_1}{\lambda+\|y\|_1}\left(1-\frac{\|y\|_1}{\lambda+\|y\|_1}\right)\,\mathrm{d}\mu_\lambda^{\otimes n}(y)=-\frac{1}{\lambda}\mathbb{E}[f'(X)X(1-X)\mid Z=n],

and hence summing over nn yields directly ∫f′(x) dσλ(x)\int f'(x)\,\mathrm{d}\sigma_\lambda(x). The second term is already in the form of the operator TT and so we are done. Hence we arrive at

∫f′ dΓ~λ=∫f′ dσλ+∫f′ d(TΓ~λ),\int f'\,\mathrm{d}\widetilde{\Gamma}_\lambda=\int f'\,\mathrm{d}\sigma_\lambda+\int f'\,\mathrm{d}(T\widetilde{\Gamma}_\lambda),

and as f∈PQf\in\mathcal{P}_{\mathbb{Q}} and the polynomials are dense in C([0,1])C([0,1]), we get Γ~λ=σλ+TΓ~λ\widetilde{\Gamma}_\lambda=\sigma_\lambda+T\widetilde{\Gamma}_\lambda. By uniqueness of the fixed point equation (7), we get Γ~a=Γa\widetilde{\Gamma}_a=\Gamma_a for any a∈(0,m)\Na\in(0,m)\backslash\mathcal{N}.

  1. Finally we want to upgrade the convergence result. So fix f′∈PQf' \in\mathcal{P}_{Q}. From the previous steps we know that λ↦Ef(Xλ)\lambda\mapsto\mathbb{E}f(X_{\lambda}) is locally Lipschitz continuous and so on every compact interval [a,b]⊂(0,m)[a,b] \subset(0,m) it is absolutely continuous. As such, by Lebesgue’s fundamental theorem of calculus, for any a,b∈(0,m)a,b \in(0,m):

Ef(Xb)−Ef(Xa)=∫[a,b]ddλEf(Xλ)dλ=∫[a,b](−1λ∫f′(y)Γλ(dy))dλ,\mathbb{E}f(X_b)-\mathbb{E}f(X_a)=\int_{[a,b]}\frac{\mathrm{d}}{\mathrm{d}\lambda}\mathbb{E}f(X_{\lambda})\mathrm{d}\lambda=\int_{[a,b]}\left(-\frac{1}{\lambda}\int f'(y)\Gamma_{\lambda}(\mathrm{d}y)\right)\mathrm{d}\lambda,

where the last equality comes from the almost everywhere equivalence. Using the fact that λ↦Γλ\lambda\mapsto\Gamma_{\lambda} is weakly continuous by Lemma A.3, we know that λ↦−1λ∫f′(y)Γλ(dy)\lambda\mapsto-\frac{1}{\lambda}\int f'(y)\Gamma_{\lambda}(\mathrm{d}y) is continuous, and hence the right-hand-side is equivalent to the Riemann integral. Hence, by the fundamental theorem of calculus

∀λ∈(0,m):ddλEf(Xλ)=−1λ∫f′(y)Γλ(dy).\forall\lambda\in(0,m):\qquad\frac{\mathrm{d}}{\mathrm{d}\lambda}\mathbb{E}f(X_{\lambda})=-\frac{1}{\lambda}\int f'(y)\Gamma_{\lambda}(\mathrm{d}y).

Now it is straightforward to extend the argument to f∈C1([0,1])⊂PQ‾∥⋅∥∞f \in C^{1}([0,1]) \subset\overline{\mathcal{P}_{Q}}^{\|\cdot\|_{\infty}}.

Having established the derivative formula (8), we record in passing that it is not difficult to show (using weak continuity of λ↦μλ\lambda\mapsto\mu_{\lambda} from Lemma A.3) that for f:(0,m)×[0,1]n→Rf:(0,m)\times[0,1]^{n}\to\mathbb{R} continuously differentiable, we have the formula2

ddλ∫fλdμλ⊗n=∫∂λfλdμλ⊗n−1λ∑i=1n∫∂ifλd(μλ⊗(i−1)⊗Γλ⊗μλ⊗(n−i)).(16)\frac{\mathrm{d}}{\mathrm{d}\lambda}\int f_{\lambda}\mathrm{d}\mu_{\lambda}^{\otimes n}=\int\partial_{\lambda}f_{\lambda}\mathrm{d}\mu_{\lambda}^{\otimes n}-\frac{1}{\lambda}\sum_{i=1}^{n}\int\partial_{i}f_{\lambda}\mathrm{d}\left(\mu_{\lambda}^{\otimes(i-1)}\otimes\Gamma_{\lambda}\otimes\mu_{\lambda}^{\otimes(n-i)}\right). \tag*{(16)}

We can now explicitly compute the derivative of (5).

Proposition 2.9 The speed ℓ\ell is differentiable on (0,m)(0,m) and it holds that

−λ2Bλℓ′(λ)=2Qλ(Ωλ,wλ)+(1+ℓ(λ))Qλ(σλ,σλ).(17)-\lambda^{2}B_{\lambda}\ell'(\lambda)=2Q_{\lambda}(\Omega_{\lambda},w_{\lambda})+(1+\ell(\lambda))Q_{\lambda}(\sigma_{\lambda},\sigma_{\lambda}). \tag*{(17)}

where Ωλ=σλ+(λ−1)Γλ\Omega_{\lambda}=\sigma_{\lambda}+(\lambda-1)\Gamma_{\lambda} and wλ(dx)=(x−ℓ(λ)(1−x))μλ(dx)w_{\lambda}(\mathrm{d}x)=(x-\ell(\lambda)(1-x))\mu_{\lambda}(\mathrm{d}x).

Proof. We start with the rewriting of Rousselins formula (5)

0=Aλ−Bλℓ(λ)=EXY−ℓ(λ)(X+Y−XY)Dλ(X,Y)=:EHλ(X,Y)Dλ(X,Y)=∫Hλ(x,y)Dλ(x,y)dμλ(x)dμλ(y),0=A_{\lambda}-B_{\lambda}\ell(\lambda)=\mathbb{E}\frac{XY-\ell(\lambda)(X+Y-XY)}{D_{\lambda}(X,Y)}=:\mathbb{E}\frac{H_{\lambda}(X,Y)}{D_{\lambda}(X,Y)}=\int\frac{H_{\lambda}(x,y)}{D_{\lambda}(x,y)}\mathrm{d}\mu_{\lambda}(x)\mathrm{d}\mu_{\lambda}(y),

and differentiate with respect to λ\lambda. We obtain with (16)2

0=∫∂λ(Hλ(x,y)Dλ(x,y))dμλ(x)dμλ(y)−1λ∂x(Hλ(x,y)Dλ(x,y))dΓλ(x)dμλ(y)−1λ∂y(Hλ(x,y)Dλ(x,y))dΓλ(y)dμλ(x).0=\int\partial_{\lambda}\left(\frac{H_{\lambda}(x,y)}{D_{\lambda}(x,y)}\right)\mathrm{d}\mu_{\lambda}(x)\mathrm{d}\mu_{\lambda}(y)-\frac{1}{\lambda}\partial_{x}\left(\frac{H_{\lambda}(x,y)}{D_{\lambda}(x,y)}\right)\mathrm{d}\Gamma_{\lambda}(x)\mathrm{d}\mu_{\lambda}(y)-\frac{1}{\lambda}\partial_{y}\left(\frac{H_{\lambda}(x,y)}{D_{\lambda}(x,y)}\right)\mathrm{d}\Gamma_{\lambda}(y)\mathrm{d}\mu_{\lambda}(x).

Performing the algebraic manipulations and correctly computing the derivatives leads to

−λ2Bλℓ′(λ)=λ2E[Hλ(X,Y)Dλ2(X,Y)]+2Q(Γ,x2μλ(dx))+2(λ−1)Q(Γ,w).-\lambda^{2}B_{\lambda}\ell'(\lambda)=\lambda^{2}\mathbb{E}\left[\frac{H_{\lambda}(X,Y)}{D_{\lambda}^{2}(X,Y)}\right]+2Q(\Gamma,x^{2}\mu_{\lambda}(\mathrm{d}x))+2(\lambda-1)Q(\Gamma,w).

Here w(dx)=(x−ℓ(λ)(1−x))μλ(dx)w(\mathrm{d}x)=(x-\ell(\lambda)(1-x))\mu_{\lambda}(\mathrm{d}x). We continue with the rewriting by noting that λ=Dλ+(1−x)(1−y)\lambda=D_{\lambda}+(1-x)(1-y) and hence

λ2E[Hλ(X,Y)Dλ2(X,Y)]=λE[(Dλ(X,Y)+(1−X)(1−Y))Hλ(X,Y)Dλ2(X,Y)]=λE[(1−X)(1−Y)Hλ(X,Y)Dλ2(X,Y)],\lambda^{2}\mathbb{E}\left[\frac{H_{\lambda}(X,Y)}{D_{\lambda}^{2}(X,Y)}\right]=\lambda\mathbb{E}\left[\frac{(D_{\lambda}(X,Y)+(1-X)(1-Y))H_{\lambda}(X,Y)}{D_{\lambda}^{2}(X,Y)}\right]=\lambda\mathbb{E}\left[\frac{(1-X)(1-Y)H_{\lambda}(X,Y)}{D_{\lambda}^{2}(X,Y)}\right],

where the last equality follows as we have EHD=0\mathbb{E}\frac{H}{D}=0. By noticing that

(1−x)(1−y)Hλ(x,y)=(1+ℓ)(x(1−x))(y(1−y))−ℓ(x(1−x)(1−y)+y(1−y)(1−x)),(1-x)(1-y)H_{\lambda}(x,y)=(1+\ell)(x(1-x))(y(1-y))-\ell(x(1-x)(1-y)+y(1-y)(1-x)),

even though Hλ/DλH_{\lambda}/D_{\lambda} is not C1C^{1} on (0,m)×[0,1]2(0,m)\times[0,1]^{2}, we can easily remedy this fact by noting that μλ\mu_{\lambda} has restricted support on which 1/Dλ1/D_{\lambda} is C1C^{1}, so the usual argument with smooth bump functions does the job. we arrive at

−λ2Bλℓ′(λ)=(1+ℓ)Q(σ,σ)−2ℓ(λ)Q(σ,(1−x)μλ(dx))+2Q(Γ,x2μλ(dx))+2(λ−1)Q(Γ,w).-\lambda^{2}B_{\lambda}\ell'(\lambda)=(1+\ell)Q(\sigma,\sigma)-2\ell(\lambda)Q(\sigma,(1-x)\mu_{\lambda}(\mathrm{d}x))+2Q(\Gamma,x^{2}\mu_{\lambda}(\mathrm{d}x))+2(\lambda-1)Q(\Gamma,w).

Now we are almost there, we use that Γ=σ+TΓ\Gamma=\sigma+T\Gamma and that σ=Tα\sigma=T\alpha to get at

Q(Γ,x2μλ(dx))=Q(Γ,xμλ(dx))−Q(Γ,σ)=Q(Γ,xμλ(dx))−Q(Γ,T(xμλ(dx)))=Q(Γ−TΓ,xμλ(dx))=Q(σ,xμλ(dx)).\begin{aligned} Q(\Gamma,x^{2}\mu_{\lambda}(\mathrm{d}x))&=Q(\Gamma,x\mu_{\lambda}(\mathrm{d}x))-Q(\Gamma,\sigma)=Q(\Gamma,x\mu_{\lambda}(\mathrm{d}x))-Q(\Gamma,T(x\mu_{\lambda}(\mathrm{d}x)))\\ &=Q(\Gamma-T\Gamma,x\mu_{\lambda}(\mathrm{d}x))=Q(\sigma,x\mu_{\lambda}(\mathrm{d}x)). \end{aligned}

And hence

−λ2Bλℓ′(λ)=(1+ℓ)Q(σ,σ)+2Q(σ+(λ−1)Γ,w).-\lambda^{2}B_{\lambda}\ell'(\lambda)=(1+\ell)Q(\sigma,\sigma)+2Q(\sigma+(\lambda-1)\Gamma,w).

The proof

Having established (17), we can now quite straightforwardly prove Theorem 1.1. We start with a technical bound on ℓ(λ)\ell(\lambda) which is independent of the assumption on the offspring law of Z≥4Z\ge4.

Lemma 3.1 For λ∈(0,m)\lambda\in(0,m) it holds that

ℓ(λ)≤EXλ2−EXλ.(18)\ell(\lambda)\le\frac{\mathbb{E}X_{\lambda}}{2-\mathbb{E}X_{\lambda}}. \tag*{(18)}

Proof. The observation we use is that

Aλ=E[XYDλ(X,Y)]≤E[Y]E[XDλ(X,Y)].(19)A_{\lambda}=\mathbb{E}\left[\frac{XY}{D_{\lambda}(X,Y)}\right]\le\mathbb{E}[Y]\mathbb{E}\left[\frac{X}{D_{\lambda}(X,Y)}\right]. \tag*{(19)}

Indeed, the function r:y↦EXDλ(X,y)=EXλ−(1−X)(1−y)r:y\mapsto\mathbb{E}\frac{X}{D_{\lambda}(X,y)}=\mathbb{E}\frac{X}{\lambda-(1-X)(1-y)} is monotonically decreasing, hence we have for any random variables Y,ZY,Z that

(Y−Z)(r(Y)−r(Z))≤0.(Y-Z)(r(Y)-r(Z))\le0.

Hence taking in particular Y,ZY,Z i.i.d. and distributed according to μλ\mu_{\lambda}, we get by taking expectation values the desired inequality (19). Finally, by definition we have

Bλ=E[X+Y−XYDλ(X,Y)]=2E[XDλ(X,Y)]−Aλ,B_{\lambda}=\mathbb{E}\left[\frac{X+Y-XY}{D_{\lambda}(X,Y)}\right]=2\mathbb{E}\left[\frac{X}{D_{\lambda}(X,Y)}\right]-A_{\lambda},

and so

ℓ(λ)=AλBλ=Aλ2E[XDλ(X,Y)]−Aλ≤E[Y]E[XDλ(X,Y)]2E[XDλ(X,Y)]−E[Y]E[XDλ(X,Y)]=EY2−EY.\ell(\lambda)=\frac{A_{\lambda}}{B_{\lambda}}=\frac{A_{\lambda}}{2\mathbb{E}\left[\frac{X}{D_{\lambda}(X,Y)}\right]-A_{\lambda}}\le\frac{\mathbb{E}[Y]\mathbb{E}\left[\frac{X}{D_{\lambda}(X,Y)}\right]}{2\mathbb{E}\left[\frac{X}{D_{\lambda}(X,Y)}\right]-\mathbb{E}[Y]\mathbb{E}\left[\frac{X}{D_{\lambda}(X,Y)}\right]}=\frac{\mathbb{E}Y}{2-\mathbb{E}Y}.

Having established now this bound on the speed ℓ\ell we come now to the proof of Theorem 1.1. The point is that we can lower-bound very crudely each of the appearing terms inside of (17):

Proof of Theorem 1.1. We start by inspecting the formula (17) and note that as Qλ(σλ,σλ)≥0Q_{\lambda}(\sigma_{\lambda},\sigma_{\lambda})\ge0 we are done the moment we can show that the sign of Qλ(Ωλ,wλ)Q_{\lambda}(\Omega_{\lambda},w_{\lambda}) is (strictly) positive as well. Explicitly we can write

Qλ(Ωλ,wλ)=∫∫y−ℓ(λ)(1−y)Dλ2(x,y) μλ(dy)Ωλ(dx)Q_{\lambda}(\Omega_{\lambda},w_{\lambda})=\int\int\frac{y-\ell(\lambda)(1-y)}{D_{\lambda}^{2}(x,y)}\,\mu_{\lambda}(\mathrm{d}y)\Omega_{\lambda}(\mathrm{d}x)

and the inner integrand we can estimate directly for every fixed xx

∫y−ℓ(λ)(1−y)Dλ2(x,y) μλ(dy)≥∫yλ2 μλ(dy)−ℓ(λ)∫1−y(34λ)2 μλ(dy)≥EXλλ2−EXλ2−EXλ169λ2(1−EXλ)=EXλλ2(1−1691−EXλ2−EXλ)≥EXλ9λ2>0.\begin{aligned} \int\frac{y-\ell(\lambda)(1-y)}{D_{\lambda}^{2}(x,y)}\,\mu_{\lambda}(\mathrm{d}y)\ge\int\frac{y}{\lambda^{2}}\,\mu_{\lambda}(\mathrm{d}y)-\ell(\lambda)\int\frac{1-y}{\left(\frac{3}{4}\lambda\right)^{2}}\,\mu_{\lambda}(\mathrm{d}y) \\ &\ge\frac{\mathbb{E}X_\lambda}{\lambda^2}-\frac{\mathbb{E}X_\lambda}{2-\mathbb{E}X_\lambda}\frac{16}{9\lambda^2}(1-\mathbb{E}X_\lambda) \\ &= \frac{\mathbb{E}X_\lambda}{\lambda^2}\left(1-\frac{16}{9}\frac{1-\mathbb{E}X_\lambda}{2-\mathbb{E}X_\lambda}\right)\ge\frac{\mathbb{E}X_\lambda}{9\lambda^2}>0. \end{aligned}

Here in the second inequality we used (18) and in the last inequality the elementary inequality 1−x2−x≥12\frac{1-x}{2-x}\ge\frac{1}{2}. Since Ωλ\Omega_\lambda is a positive measure by Lemma A.4, we obtain that Qλ(Ωλ,wλ)>0Q_\lambda(\Omega_\lambda,w_\lambda)>0 and hence that −λ2Bλℓ′(λ)>0-\lambda^2B_\lambda\ell'(\lambda)>0 and so the theorem follows, as λ2Bλ>0\lambda^2B_\lambda>0.

Some technical lemmas

We collect some lemmas that were used, which are technical in nature and use often just a slight (but tedious) variant of arguments already presented in the main text.

Lemma A.1 For any u,v≥0u,v\ge0 such that [u,v]⊂(0,m)[u,v]\subset(0,m) there exists a constant Cu,v>0C_{u,v}>0 such that

E[Xa−Xb]≤Cu,v(b−a)\mathbb{E}[X_a-X_b]\le C_{u,v}(b-a)

uniformly in a,ba,b with u≤a<b≤vu\le a<b\le v.

Proof. We fix such u,vu,v as in the assumption and choose some a,ba,b with u≤a<b≤vu\le a<b\le v. We will now define depending on a,ba,b analogous objects to QQ, TT and α\alpha, σ\sigma as follows. First we set

Da(x,x′)=a−(1−x)(1−x′),Db(y,y′)=b−(1−y)(1−y′).D_a(x,x')=a-(1-x)(1-x'),\qquad D_b(y,y')=b-(1-y)(1-y').

The symmetric bilinear form Qa,bQ_{a,b} we set as

Qa,b:Ma,2×Mb,2→RQ_{a,b}:\mathcal{M}_{a,2}\times\mathcal{M}_{b,2}\to\mathbb{R}
(η,ζ)↦∫abDa(x,x′)Db(y,y′)η(dx,dx′)ζ(dy,dy′).(\eta,\zeta)\mapsto\int\frac{\sqrt{ab}}{D_a(x,x')D_b(y,y')}\eta(dx,dx')\zeta(dy,dy').

Here, Mλ,2={ν∣supp⁡ν⊂[(1−λ/4)+,1]2}\mathcal{M}_{\lambda,2}=\{\nu\mid\operatorname{supp}\nu\subset[(1-\lambda/4)_+,1]^2\}. We get similar to (9), that for positive measures η∈Ma,2+\eta\in\mathcal{M}_{a,2}^+, ζ∈Mb,2+\zeta\in\mathcal{M}_{b,2}^+:

1v∥η∥TV∥ζ∥TV≤Qa,b(η,ζ)≤169u∥η∥TV∥ζ∥TV.(20)\frac{1}{v}\|\eta\|_{\mathrm{TV}}\|\zeta\|_{\mathrm{TV}}\le Q_{a,b}(\eta,\zeta)\le\frac{16}{9u}\|\eta\|_{\mathrm{TV}}\|\zeta\|_{\mathrm{TV}}. \tag*{(20)}

as well as an operator Ta,b:M→MT_{a,b}:\mathcal{M}\to\mathcal{M} defined via actions on test functions f∈C([0,1]2)f\in C([0,1]^2) as

∫fd(Ta,bη)=∑nnpn∫E[ab(a+x+Sa,n−1)(b+y+Sb,n−1)f(x+Sa,n−1λ+x+Sa,n−1,y+Sb,n−1λ+y+Sb,n−1)]η(dx,dy).\int f\mathrm{d}(T_{a,b}\eta)=\sum_n np_n\int\mathbb{E}\left[\frac{\sqrt{ab}}{(a+x+S_{a,n-1})(b+y+S_{b,n-1})}f\left(\frac{x+S_{a,n-1}}{\lambda+x+S_{a,n-1}},\frac{y+S_{b,n-1}}{\lambda+y+S_{b,n-1}}\right)\right]\eta(dx,dy).

This map Ta,bT_{a,b} has analogous properties as in Lemmas 2.3 and 2.6, in particular it is a positive linear operator symmetric with respect to Qa,bQ_{a,b}, that is for η∈Ma,2\eta\in\mathcal{M}_{a,2}, ζ∈Mb,2\zeta\in\mathcal{M}_{b,2}

Qa,b(Ta,bη,ζ)=Qa,b(η,Ta,bζ).Q_{a,b}(T_{a,b}\eta,\zeta)=Q_{a,b}(\eta,T_{a,b}\zeta).

We introduce furthermore

μa,b=Law⁡P(Xa,Xb),\mu_{a,b}=\operatorname{Law}_{\mathbb{P}}(X_a,X_b),

as well as

αa,b(dx,dy)=(xa+yb)μa,b(dx,dy),\alpha_{a,b}(dx,dy)=\left(\frac{x}{\sqrt{a}}+\frac{y}{\sqrt{b}}\right)\mu_{a,b}(dx,dy),
σa,b(dx,dy)=(x(1−y)b+y(1−x)a)μa,b(dx,dy).\sigma_{a,b}(dx,dy)=\left(\frac{x(1-y)}{\sqrt{b}}+\frac{y(1-x)}{\sqrt{a}}\right)\mu_{a,b}(dx,dy).

We can check by similar arguments as before that

Ta,bαa,b=σa,b.T_{a,b}\alpha_{a,b}=\sigma_{a,b}.

Finally we define

Γa,b(dx,dy)=x−yb−aμa,b(dx,dy).\Gamma_{a,b}(\mathrm{d}x,\mathrm{d}y)=\frac{x-y}{\sqrt{b}-\sqrt{a}}\mu_{a,b}(\mathrm{d}x,\mathrm{d}y).

We can now by straightforward algebraic manipulations check that in fact

Γa,b=σa,b+Ta,bΓa,b.\Gamma_{a,b}=\sigma_{a,b}+T_{a,b}\Gamma_{a,b}.

Now we have done enough preprocessing to obtain a bound on ∥Γa,b∥TV\|\Gamma_{a,b}\|_{\mathrm{TV}}. In fact, the procedure will be mimicking again the main text. That is, we note that αa,b−σa,b\alpha_{a,b}-\sigma_{a,b} is positive:

d(αa,b−σa,b)dμa,b(x,y)=((x−y)(1a−1b)+xy(1a+1b))≥0.\frac{\mathrm{d}(\alpha_{a,b}-\sigma_{a,b})}{\mathrm{d}\mu_{a,b}}(x,y)=\left((x-y)\left(\frac{1}{\sqrt{a}}-\frac{1}{\sqrt{b}}\right)+xy\left(\frac{1}{\sqrt{a}}+\frac{1}{\sqrt{b}}\right)\right)\geq0.

Then we compute using symmetry of Ta,bT_{a,b} with respect to Qa,bQ_{a,b} and Ta,bαa,b=σa,bT_{a,b}\alpha_{a,b}=\sigma_{a,b}:

Qa,b(Γa,b,αa,b−σa,b)=Qa,b(σa,b,αa,b).Q_{a,b}(\Gamma_{a,b},\alpha_{a,b}-\sigma_{a,b})=Q_{a,b}(\sigma_{a,b},\alpha_{a,b}).

Hence, with (20) we obtain

1v∥Γa,b∥TV∥αa,b−σa,b∥TV≤Qa,b(Γa,b,αa,b−σa,b)=Qa,b(σa,b,αa,b)≤169u∥αa,b∥TV∥σa,b∥TV.\frac{1}{v}\|\Gamma_{a,b}\|_{\mathrm{TV}}\|\alpha_{a,b}-\sigma_{a,b}\|_{\mathrm{TV}}\leq Q_{a,b}(\Gamma_{a,b},\alpha_{a,b}-\sigma_{a,b})=Q_{a,b}(\sigma_{a,b},\alpha_{a,b})\leq\frac{16}{9u}\|\alpha_{a,b}\|_{\mathrm{TV}}\|\sigma_{a,b}\|_{\mathrm{TV}}.

Now all that is left is to explicitly compute bounds on the total mass of αa,b\alpha_{a,b}, σa,b\sigma_{a,b} and their difference:

∥αa,b∥TV=1aEXa+1bEXb≤2u\|\alpha_{a,b}\|_{\mathrm{TV}}=\frac{1}{\sqrt{a}}\mathbb{E}X_a+\frac{1}{\sqrt{b}}\mathbb{E}X_b\leq\frac{2}{\sqrt{u}}
∥σa,b∥TV=1bEXa(1−Xb)+1bEXb(1−Xa)≤2u\|\sigma_{a,b}\|_{\mathrm{TV}}=\frac{1}{\sqrt{b}}\mathbb{E}X_a(1-X_b)+\frac{1}{\sqrt{b}}\mathbb{E}X_b(1-X_a)\leq\frac{2}{\sqrt{u}}

For the difference, we want to compute a lower bound

∥αa,b−σa,b∥TV≥(1a+1b)EXaXb≥2vEXv2>0.\|\alpha_{a,b}-\sigma_{a,b}\|_{\mathrm{TV}}\geq\left(\frac{1}{\sqrt{a}}+\frac{1}{\sqrt{b}}\right)\mathbb{E}X_aX_b\geq\frac{2}{\sqrt{v}}\mathbb{E}X_v^2>0.

Consequently we obtain

∥Γa,b∥TV≤32v3/29u2EXv2.\|\Gamma_{a,b}\|_{\mathrm{TV}}\leq\frac{32v^{3/2}}{9u^2\mathbb{E}X_v^2}.

Hence, we get the desired Lipschitz in law estimate:

E[Xa−Xb]=(b−a)∥Γa,b∥TV≤(16v3/29u5/2EXv2)(b−a).\mathbb{E}[X_a-X_b]=(\sqrt{b}-\sqrt{a})\|\Gamma_{a,b}\|_{\mathrm{TV}}\leq\left(\frac{16v^{3/2}}{9u^{5/2}\mathbb{E}X_v^2}\right)(b-a).

Lemma A.2 In the context of the proof for Proposition 2.9, it holds that for λ∈(0,m)∖N\lambda\in(0,m)\setminus\mathbb{N} once we set φλ,n(y)=y1+⋯+ynλ+y1+⋯+yn\varphi_{\lambda,n}(y)=\frac{y_1+\cdots+y_n}{\lambda+y_1+\cdots+y_n} we have

ddλ∑npn∫φλ,n dμλ⊗n=∑npn[∫∂λφλ,n dμλ⊗n−nλ∫∂1φλ,n d(μ~λ⊗μλ⊗(n−1))].\frac{\mathrm{d}}{\mathrm{d}\lambda}\sum_n p_n\int\varphi_{\lambda,n}\,\mathrm{d}\mu_\lambda^{\otimes n} =\sum_n p_n\left[\int\partial_\lambda\varphi_{\lambda,n}\,\mathrm{d}\mu_\lambda^{\otimes n}-\frac{n}{\lambda}\int\partial_1\varphi_{\lambda,n}\,\mathrm{d}\left(\tilde{\mu}_\lambda\otimes\mu_\lambda^{\otimes(n-1)}\right)\right].

Proof. We evaluate the derivative by taking the limit of the finite difference quotient along a specific sequence λk↓λ\lambda_k \downarrow\lambda for which we know that Rλ,λk⇀Γ~λR_{\lambda,\lambda_k} \rightharpoonup\widetilde{\Gamma}_\lambda. Without loss of generality this sequence will be contained in some compact interval (λk)k⊂[u,v]⊂(0,m)(\lambda_k)_k \subset[u,v] \subset(0,m). That is we form

Ef(Xλk)−Ef(Xλ)λk−λ=∑npn∫φλk,ndμλk⊗n−∫φλ,ndμλ⊗nλk−λ=:∑npnIn(λ,λk).\frac{\mathbb{E}f(X_{\lambda_k})-\mathbb{E}f(X_\lambda)}{\lambda_k-\lambda} =\sum_n p_n\frac{\int\varphi_{\lambda_k,n}\mathrm{d}\mu_{\lambda_k}^{\otimes n}-\int\varphi_{\lambda,n}\mathrm{d}\mu_\lambda^{\otimes n}}{\lambda_k-\lambda} =:\sum_n p_n I_n(\lambda,\lambda_k).

We split now In=In(1)+In(2)I_n=I_n^{(1)}+I_n^{(2)} with

In(1)(λ,λk)=∫(φλk,n−φλ,n)dμλk⊗nλk−λ,In(2)(λ,λk)=∫φλ,nd(μλk⊗n−μλ⊗n)λk−λ.I_n^{(1)}(\lambda,\lambda_k)=\frac{\int(\varphi_{\lambda_k,n}-\varphi_{\lambda,n})\mathrm{d}\mu_{\lambda_k}^{\otimes n}}{\lambda_k-\lambda},\qquad I_n^{(2)}(\lambda,\lambda_k)=\frac{\int\varphi_{\lambda,n}\mathrm{d}(\mu_{\lambda_k}^{\otimes n}-\mu_\lambda^{\otimes n})}{\lambda_k-\lambda}.

We note first a convenient representation for In(2)I_n^{(2)} in the form of a telescoping sum

In(2)(λ,λk)=1λk−λE[φλ,n(Xλk,1,…,Xλk,n)−φλ,n(Xλ,1,…,Xλ,n)]=1λk−λ∑j=1nE[φλ,n(Xλk,1,…,Xλk,j,Xλ,j+1,…,Xλ,n)−φλ,n(Xλk,1,…,Xλk,j−1,Xλ,j,…,Xλ,n)].\begin{aligned} I_n^{(2)}(\lambda,\lambda_k)=\frac{1}{\lambda_k-\lambda}\mathbb{E}\left[\varphi_{\lambda,n}(X_{\lambda_k,1},\ldots,X_{\lambda_k,n})-\varphi_{\lambda,n}(X_{\lambda,1},\ldots,X_{\lambda,n})\right] \\ &=\frac{1}{\lambda_k-\lambda}\sum_{j=1}^{n}\mathbb{E}\left[\varphi_{\lambda,n}(X_{\lambda_k,1},\ldots,X_{\lambda_k,j},X_{\lambda,j+1},\ldots,X_{\lambda,n})-\varphi_{\lambda,n}(X_{\lambda_k,1},\ldots,X_{\lambda_k,j-1},X_{\lambda,j},\ldots,X_{\lambda,n})\right]. \end{aligned}

We will first argue that we can interchange limits and the sum over nn. Indeed, we can immediately bound

∣In(1)∣≤∥f′∥∞u\left|I_n^{(1)}\right|\leq\frac{\lVert f'\rVert_\infty}{u}
∣In(2)∣≤1λk−λ∑j=1nE[∥∂jφλ,n∥∞(Xλk,j−Xλ,j)]≤nCu,vu∥f′∥∞.\left|I_n^{(2)}\right|\leq\frac{1}{\lambda_k-\lambda}\sum_{j=1}^{n}\mathbb{E}\left[\lVert\partial_j\varphi_{\lambda,n}\rVert_\infty(X_{\lambda_k,j}-X_{\lambda,j})\right]\leq\frac{nC_{u,v}}{u}\lVert f'\rVert_\infty.

Here we used Lemma 2.8 to obtain the very last inequality. Hence we see that the map n↦Inn\mapsto I_n is pointwise bounded by a function which is integrable:

∑npn(1+nCu,v)∥f′∥∞u=(1+Cu,vEZ)∥f′∥∞u<∞\sum_n p_n(1+nC_{u,v})\frac{\lVert f'\rVert_\infty}{u} =(1+C_{u,v}\mathbb{E}Z)\frac{\lVert f'\rVert_\infty}{u}<\infty

Thus by dominated convergence theorem we have

lim⁡k→∞∑npnIn(λ,λk)=∑npnlim⁡k→∞In(λ,λk),\lim_{k\to\infty}\sum_n p_n I_n(\lambda,\lambda_k)=\sum_n p_n\lim_{k\to\infty}I_n(\lambda,\lambda_k),

granted that the pointwise limit lim⁡kIn(λ,λk)\lim_k I_n(\lambda,\lambda_k) exists. This we will show now. For In(1)I_n^{(1)} this is quite straightforward, since μλk⇀μλ\mu_{\lambda_k}\rightharpoonup\mu_\lambda and φλk,n−φλ,nλk−λ→∂λφλ,n\frac{\varphi_{\lambda_k,n}-\varphi_{\lambda,n}}{\lambda_k-\lambda}\to\partial_\lambda\varphi_{\lambda,n} uniformly as f∈C1([0,1])f\in C^1([0,1]). Hence

lim⁡k→∞In(1)(λ,λk)=∫∂λφλ,ndμλ⊗n.\lim_{k\to\infty}I_n^{(1)}(\lambda,\lambda_k)=\int\partial_\lambda\varphi_{\lambda,n}\mathrm{d}\mu_\lambda^{\otimes n}.

For In(2)I_n^{(2)} we extract from the telescoping sum with the fundamental theorem of calculus and the independence of Xλk,1,…,Xλ,nX_{\lambda_k,1},\ldots,X_{\lambda,n}:

In(2)(λ,λk)=1λk−λ∑j=1nE[∫Xλ,jXλk,j∂jφλ,n(Xλk,1,…,z,…,Xλ,n)dz]=−1λ∑j=1n∫∂jφλ,nd(μλk⊗(j−1)⊗Rλ,λk⊗μλ⊗(n−j)).\begin{aligned} I_n^{(2)}(\lambda,\lambda_k)=\frac{1}{\lambda_k-\lambda}\sum_{j=1}^{n}\mathbb{E}\left[\int_{X_{\lambda,j}}^{X_{\lambda_k,j}}\partial_j\varphi_{\lambda,n}(X_{\lambda_k,1},\ldots,z,\ldots,X_{\lambda,n})\mathrm{d}z\right] \\ &=-\frac{1}{\lambda}\sum_{j=1}^{n}\int\partial_j\varphi_{\lambda,n}\mathrm{d}\left(\mu_{\lambda_k}^{\otimes(j-1)}\otimes R_{\lambda,\lambda_k}\otimes\mu_\lambda^{\otimes(n-j)}\right). \end{aligned}

Now as ∂jφλ,n\partial_j\varphi_{\lambda,n} is itself bounded and continuous due to f∈C1([0,1])f\in C^1([0,1]), we can pass to the limit, using μλk⇀μλ\mu_{\lambda_k}\rightharpoonup\mu_\lambda and Rλ,λk⇀Γ~λR_{\lambda,\lambda_k}\rightharpoonup\widetilde{\Gamma}_\lambda, to obtain

lim⁡k→∞In(2)(λ,λk)=−1λ∑j=1n∫∂jφλ,nd(μλ⊗(j−1)⊗Γ~λ⊗μλ⊗(n−j))=−nλ∫∂1φλ,nd(Γ~λ⊗μλ⊗(n−1)).\lim_{k\to\infty}I_n^{(2)}(\lambda,\lambda_k) =-\frac{1}{\lambda}\sum_{j=1}^{n}\int\partial_j\varphi_{\lambda,n}\mathrm{d}\left(\mu_\lambda^{\otimes(j-1)}\otimes\widetilde{\Gamma}_\lambda\otimes\mu_\lambda^{\otimes(n-j)}\right) =-\frac{n}{\lambda}\int\partial_1\varphi_{\lambda,n}\mathrm{d}\left(\widetilde{\Gamma}_\lambda\otimes\mu_\lambda^{\otimes(n-1)}\right).

The last equality follows from the fact that φλ,n\varphi_{\lambda,n} factors through y↦∥y∥1y \mapsto\lVert y\rVert_1 which is permutation invariant. This concludes the proof.

Lemma A.3. We have

  1. λ↦μλ,(0,m)→M+\lambda\mapsto\mu_{\lambda,(0,m)} \to\mathcal{M}^{+} is weakly continuous.

  1. λ↦Γλ,(0,m)→M+\lambda\mapsto\Gamma_{\lambda,(0,m)} \to\mathcal{M}^{+} is weakly continuous.

Proof. 1. It easy to see that by the Lipschitz continuity of Lemma 2.8, and the fact that Lipschitz continuous functions are dense in C([0,1])C([0,1]), that we have λ↦μλ\lambda\mapsto\mu_{\lambda} is weakly continuous.

  1. Let λn→λ\lambda_n \to\lambda. To prove Γλn⇀Γλ\Gamma_{\lambda_n} \rightharpoonup\Gamma_{\lambda} we use the usual subsequence argument. So let (λn)n(\lambda_n)_n be any subsequence of the original sequence3. First, note that by (12) and Jensen’s inequality, we get a uniform in λ\lambda bound

∥Γλ∥TV≤lim⁡N→∞∥ΓλN∥TV≤169(EXλ)(EXλ−Xλ2)EXλ2≤169.\lVert\Gamma_{\lambda}\rVert_{\mathrm{TV}} \le\lim_{N\to\infty}\lVert\Gamma_{\lambda}^{N}\rVert_{\mathrm{TV}} \le\frac{16}{9}\frac{(\mathbb{E}X_{\lambda})(\mathbb{E}X_{\lambda}-X_{\lambda}^{2})}{\mathbb{E}X_{\lambda}^{2}} \le\frac{16}{9}.

Hence, by tightness, we can extract a subsequence such that (Γλnk)k(\Gamma_{\lambda_{n_k}})_k converges weakly in M+\mathcal{M}^{+}, and denote its limit by Γ~λ=w−limk→∞Γλnk\widetilde{\Gamma}_{\lambda} = \mathop{\mathrm{w-lim}}_{k\to\infty}\Gamma_{\lambda_{n_k}}. Now for every kk it holds

Γλnk=σλnk+TλnkΓλnk.\Gamma_{\lambda_{n_k}}=\sigma_{\lambda_{n_k}}+T_{\lambda_{n_k}}\Gamma_{\lambda_{n_k}}.

As clearly λ↦σλ\lambda\mapsto\sigma_{\lambda} is weakly continuous, all that remains is to compute the weak limit of TλnkΓλnkT_{\lambda_{n_k}}\Gamma_{\lambda_{n_k}}. To this end, let f∈C([0,1])f\in C([0,1]) and compute ∫f dTλnkΓλnk\int f\,\mathrm{d}T_{\lambda_{n_k}}\Gamma_{\lambda_{n_k}}. We denote by Φλ=Φλ(m,f)∈C([0,1]m;R)\Phi_{\lambda}=\Phi_{\lambda}^{(m,f)}\in C([0,1]^m;\mathbb{R}) the function such that

∫f dTλΓλ=∑m=4∞mpm∫Φλ d(Γλ⊗μλ⊗(m−1)).\int f\,\mathrm{d}T_{\lambda}\Gamma_{\lambda}=\sum_{m=4}^{\infty}mp_m\int\Phi_{\lambda}\,\mathrm{d}\left(\Gamma_{\lambda}\otimes\mu_{\lambda}^{\otimes(m-1)}\right).

Explicitly, Φλ(x,y1,…,ym−1)=λλ+x+∑iyif(x+∑iyiλ+x+∑iyi)\Phi_{\lambda}(x,y_1,\ldots,y_{m-1})=\frac{\lambda}{\lambda+x+\sum_i y_i}f\left(\frac{x+\sum_i y_i}{\lambda+x+\sum_i y_i}\right). Since ∥Φλnk−Φλ∥∞→0\lVert\Phi_{\lambda_{n_k}}-\Phi_{\lambda}\rVert_{\infty}\to0 and Γλnk⊗μλnk⊗(m−1)⇀Γ~λ⊗μλ⊗(m−1)\Gamma_{\lambda_{n_k}}\otimes\mu_{\lambda_{n_k}}^{\otimes(m-1)}\rightharpoonup\widetilde{\Gamma}_{\lambda}\otimes\mu_{\lambda}^{\otimes(m-1)} it is easy to see that for every fixed m∈Nm\in\mathbb{N}:

lim⁡k→∞∫Φλnk d(Γλnk⊗μλnk⊗(m−1))=∫Φλ d(Γ~λ⊗μλ⊗(m−1)).\lim_{k\to\infty}\int\Phi_{\lambda_{n_k}}\,\mathrm{d}\left(\Gamma_{\lambda_{n_k}}\otimes\mu_{\lambda_{n_k}}^{\otimes(m-1)}\right)=\int\Phi_{\lambda}\,\mathrm{d}\left(\widetilde{\Gamma}_{\lambda}\otimes\mu_{\lambda}^{\otimes(m-1)}\right).

Finally, by inspection

∥Φλ∥∞≤∥f∥∞u,\lVert\Phi_{\lambda}\rVert_{\infty}\le\frac{\lVert f\rVert_{\infty}}{u},

where u>0u>0 is some lower bound such that for all n∈Nn\in\mathbb{N}: λn>u\lambda_n>u which we can assume without loss of generality. This, combined with the bound on ∥Γλ∥TV\lVert\Gamma_{\lambda}\rVert_{\mathrm{TV}}, yields the following uniform in λnk\lambda_{n_k} bound

∣∑m≥Nmpm∫Φλnk d(Γλnk⊗μλnk⊗(m−1))∣≤169∥f∥∞u∑m≥Nmpm.\left\lvert\sum_{m\ge N}mp_m\int\Phi_{\lambda_{n_k}}\,\mathrm{d}\left(\Gamma_{\lambda_{n_k}}\otimes\mu_{\lambda_{n_k}}^{\otimes(m-1)}\right)\right\rvert\le\frac{16}{9}\frac{\lVert f\rVert_{\infty}}{u}\sum_{m\ge N}mp_m.

The right-hand-side tends to 0 for N→∞N\to\infty as EZ=m<∞\mathbb{E}Z=m<\infty, and hence we can interchange limit taking over λnk→λ\lambda_{n_k}\to\lambda and summing over m∈Nm\in\mathbb{N}. Therefore TλnkΓλnk⇀TλΓ~λT_{\lambda_{n_k}}\Gamma_{\lambda_{n_k}}\rightharpoonup T_{\lambda}\widetilde{\Gamma}_{\lambda}, and thus

Γ~λ=σλ+TλΓ~λ,\widetilde{\Gamma}_{\lambda}=\sigma_{\lambda}+T_{\lambda}\widetilde{\Gamma}_{\lambda},

from which, by Γ~λ∈M+\widetilde{\Gamma}_{\lambda}\in\mathcal{M}^{+} and uniqueness follows Γ~λ=Γλ\widetilde{\Gamma}_{\lambda}=\Gamma_{\lambda}. So by the subsequence argument, we have proven that Γλn⇀Γλ\Gamma_{\lambda_n}\rightharpoonup\Gamma_{\lambda}.

Lemma A.4. The measure Ωλ\Omega_{\lambda} is positive. Proof. Since Ωλ=σλ−(1−λ)Γλ\Omega_{\lambda}=\sigma_{\lambda}-(1-\lambda)\Gamma_{\lambda}, and σλ,Γλ≥0\sigma_{\lambda},\Gamma_{\lambda}\geq0, positivity is clear when λ∈[1,m)\lambda\in[1,m). So let λ∈(0,1)\lambda\in(0,1). Then observe that

Ωλ=σ−(1−λ)(σ+TΓ)=λσ+λTΓ−TΓ=λΓ−TΓ=∑j=0∞Tj+1(λα−σ).\Omega_{\lambda}=\sigma-(1-\lambda)(\sigma+T\Gamma)=\lambda\sigma+\lambda T\Gamma-T\Gamma=\lambda\Gamma-T\Gamma=\sum_{j=0}^{\infty}T^{j+1}(\lambda\alpha-\sigma).

Now for λ<1\lambda<1 one sees that with (λα−σ)(dx)=x(λ−(1−x))μ(dx)(\lambda\alpha-\sigma)(dx)=x(\lambda-(1-x))\mu(dx), that as supp⁡μλ⊂[(1−λ/4)+,1]\operatorname{supp}\mu_{\lambda}\subset[(1-\lambda/4)+,1], the elementary inequality

x+(λ−1)≥1−λ/4+(λ−1)=34λ>0x+(\lambda-1)\geq1-\lambda/4+(\lambda-1)=\frac{3}{4}\lambda>0

yields λα−σ≥0\lambda\alpha-\sigma\geq0. As TT is a positive operator, positivity of Ωλ\Omega_{\lambda} follows.

Acknowledgements

Acknowledgements. FS is funded by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant agreement No. 101045082), funded also by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy EXC 2044/2 –390685587, Mathematics Münster: Dynamics–Geometry–Structure, and has been supported by the Research Training Group “Rigorous Analysis of Complex Random Systems” (RTG 3027) funded by the DFG (Project Number 524444762). The author is grateful to Jan Nagel for helpful discussions and for first bringing this theorem to their attention.

Tool usage statement. OpenAI’s ChatGPT 6 Astra was used during the initial stages of this work. The proof in its final form was developed by the author, who takes full responsibility for all claims and arguments made in the paper.

References

  1. [2]D. Aldous and J. Fill. Reversible markov chains and random walks on graphs. 2002.
  2. [4]R. Lyons, R. Pemantle, and Y. Peres. Ergodic theory on galton–watson trees: speed of random walk and dimension of harmonic measure. Ergodic Theory and Dynamical Systems, 15(3):593–619, 1995.DOI
  3. [5]R. Lyons, R. Pemantle, and Y. Peres. Biased random walks on galton–watson trees. Probability theory and related fields, 106(2):249–264, 1996.DOI
  4. [6]R. Lyons, R. Pemantle, and Y. Peres. Unsolved problems concerning random walks on trees. In Classical and Modern Branching Processes, pages 223–237. Springer New York, 1997.DOI
  5. [7]R. Lyons, Y. Peres, and Y. Peres. Probability on trees and networks, volume 42. Cambridge University Press, Cambridge, 2016.DOI
  6. [8]P. Rousselin. Invariant measures, hausdorff dimension and dimension drop of some harmonic measures on galton–watson trees. Electronic Journal of Probability, 23(46):1–31, 2018.arxiv.org/abs/1708.06965

Paper details

vlnc:W_a8xvu29y
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

GPT-6 Astra (bare)

View statement
Looking for related papers…
Contents