Introduction
Overview
Many one-dimensional random growth and interacting particle models share a characteristic large-scale behavior: at large time t t t , height fluctuations are of order t 1 / 3 t^{1/3} t 1/3 and spatial correlations extend over distances of order t 2 / 3 t^{2/3} t 2/3 . This 1:2:3 scaling is characteristic of the Kardar–Parisi–Zhang (KPZ) universality class [22 ] . Three related objects describe the expected universal limits. The KPZ fixed point describes height evolution from general initial profiles. The directed landscape is a random directed metric that records passage values between space–time points and couples these evolutions. The stationary horizon is the jointly stationary coupling of profiles with different slopes for the KPZ fixed point, after recentering at the origin.
Early progress in KPZ universality concerned fluctuation laws for special initial conditions. Baik, Deift, and Johansson proved that rescaled lengths of longest increasing subsequences of uniform random permutations converge to the Tracy–Widom law of the Gaussian unitary ensemble (GUE) [3 ] ; Johansson established corresponding one-point laws for last-passage percolation and step-initial TASEP [20 ] . To describe spatial correlations, Prähofer and Spohn introduced multilayer PNG and identified the Airy process in the PNG droplet [26 ] . Its one-point marginals recover the GUE Tracy–Widom law. Johansson proved functional convergence for discrete PNG [21 ] . The multilayer construction also led to the Airy line ensemble, whose top curve is the Airy process. Corwin and Hammond established its continuous, nonintersecting realization and, after a parabolic shift, its Brownian Gibbs property [11 ] . For general initial profiles, growth at an endpoint depends on competition among paths from different starting locations. The Airy sheet couples passage values as both starting and ending positions vary. Corwin, Quastel, and Remenik proposed a variational formula for universal height evolution, maximizing initial height plus an Airy-sheet passage value [12 ] . Building on earlier determinantal representations [28 , 5 ] , Matetski, Quastel, and Remenik constructed the KPZ fixed point as a Feller Markov process by taking the scaling limit of TASEP transition probabilities from general initial profiles [23 ] . Dauvergne, Ortmann, and Virág characterized the Airy-sheet law using the Airy line ensemble and constructed the directed landscape as the locally uniform limit of Brownian last-passage percolation, together with convergence of maximizing paths to geodesics in the directed landscape [15 ] . The directed landscape records both endpoint positions and times; passage values compose by maximization over intermediate positions, and disjoint time intervals carry independent passage fields. Using one-sided reflected Brownian motions, Nica, Quastel, and Remenik showed that maximizing initial height plus a passage value in the directed landscape gives the KPZ fixed point [25 ] . The directed landscape thus couples evolutions from different initial profiles. Subsequent results established fixed-point convergence for the O’Connell–Yor polymer [30 ] ; convergence to the directed landscape and the KPZ fixed point for classical last-passage models and PNG [14 ] and the KPZ equation [31 ] ; fixed-point convergence for ASEP and the stochastic six-vertex model, with directed landscape convergence for their colored versions [2 , 1 ] ; and directed landscape convergence for the log-gamma polymer [33 ] . Dauvergne and Zhang characterized the directed landscape coupling by fixed-point laws, independent time increments, monotonicity, and commutation with vertical shifts [16 ] , providing a framework for deriving directed landscape convergence from fixed-point convergence and control of coupled evolutions.
The stationary horizon describes the joint stationary behavior of profiles with different slopes. Its origins lie in Busemann functions, obtained from limits of passage-time differences in a prescribed direction. Fan and Seppäläinen gave a queueing description of the joint Busemann profiles of exponential last-passage percolation [18 ] . Building on this description, Busani introduced the stationary horizon as their diffusive scaling limit, with convergence of the entire process indexed by drift [6 ] . Each profile is a two-sided Brownian motion with its prescribed drift. The same coupling independently appeared in the Brownian last-passage Busemann process studied by Seppäläinen and Sorensen [29 ] . Busani, Seppäläinen, and Sorensen then proved that the stationary horizon is jointly invariant under KPZ fixed-point evolutions driven by one directed landscape, with each evolved profile recentered at the origin. Under prescribed asymptotic slope conditions, it is the unique jointly invariant law and an attractor. They also identified it with the directed landscape’s Busemann profiles, connecting stationary growth with semi-infinite geodesics [8 ] . Their subsequent convergence criterion uses directed landscape convergence, joint invariance, and control of maximizers to identify the stationary horizon as the limit. It yields joint convergence at finitely many slopes for Brownian, exponential, geometric, and Poisson last-passage percolation, the Poisson lines model, and the Seppäläinen–Johansson model [7 ] . Convergence at finitely many drift parameters also holds for the KPZ horizon, which describes jointly invariant profiles for the KPZ equation, in the low-temperature limit [19 ] , and for stationary ASEP and stochastic six-vertex profiles [1 ] . The stationary profiles encoded by the TASEP speed process also converge to the stationary horizon, with convergence of the entire process indexed by drift [9 ] .
In half-space, boundary interactions produce parameter-dependent analogues of these objects. Zhang obtained half-space TASEP transition Pfaffians and their fixed-time KPZ scaling limits [32 ] . Dauvergne and Zhang constructed the half-space stationary horizon from jointly stationary last-passage profiles and obtained the half-space directed landscape as the scaling limit of exponential last-passage percolation and a family of Poisson-avoiding metrics generalizing colored TASEP [17 ] . Their variational construction realizes Zhang’s transition laws as a continuous Markov process [17 ] , the half-space KPZ fixed point. Their directed landscape characterization is the universality criterion used here.
We study half-space PNG, a growth process on the half-line driven by bulk and boundary Poisson nucleations [27 ] . We construct its stationary measures, jointly stationary couplings, and PNG horizon. We prove convergence of the PNG horizon, height evolution, and common passage field to the half-space stationary horizon, KPZ fixed point, and directed landscape, respectively, in the critical regime and at fixed subcritical boundary rates. We prove the half-space KPZ fixed-point limit from the transition Pfaffian of Chen and Zhang [10 ] and Zhang’s limiting formula [32 ] , together with tightness and a Markov argument. We obtain the directed landscape limit through the characterization of Dauvergne and Zhang [17 ] .
Model and setting
We work on the half-line R + : = [ 0 , ∞ ) \mathbb{R}_{+}:=[0,\infty) R + := [ 0 , ∞ ) . We use the normalization of full-space PNG from [24 ] , in which bulk nucleations have intensity 2 2 2 , and a source at the origin as in [27 , 10 ] .
Definition 1.1 (Half-line height profiles). Let U C + \mathrm{UC}_{+} UC + consist of locally bounded-above, upper-semicontinuous functions h : [ 0 , ∞ ) → Z ∪ { − ∞ } h:[0,\infty)\to\mathbb{Z}\cup\{-\infty\} h : [ 0 , ∞ ) → Z ∪ { − ∞ } , with local Hausdorff convergence of their closed hypographs hypo ( h ) = { ( x , r ) : x ≥ 0 , r ≤ h ( x ) } \operatorname{hypo}(h)=\{(x,r):x\ge0,\ r\le h(x)\} hypo ( h ) = {( x , r ) : x ≥ 0 , r ≤ h ( x )} .
Upper semicontinuity selects the upper value at a step and includes height-one nucleation spikes. The value − ∞ -\infty − ∞ allows localized initial data. Write u + = max ( u , 0 ) u_{+}=\max(u,0) u + = max ( u , 0 ) .
Definition 1.2 (Half-space PNG). Fix a boundary parameter c ≥ 0 c\ge0 c ≥ 0 . The half-space PNG process with boundary rate c c c is the evolution on U C + \mathrm{UC}_{+} UC + determined by the following rules.
In the absence of new nucleations, a profile h h h evolves for time s > 0 s>0 s > 0 according to
( D s h ) ( x ) : = sup y ∈ [ ( x − s ) + , x + s ] h ( y ) , x ∈ R + . (1) (D_{s}h)(x):=\sup_{y\in[(x-s)_{+},x+s]}h(y),\qquad x\in\mathbb{R}_{+}. \tag*{(1)} ( D s h ) ( x ) := y ∈ [( x − s ) + , x + s ] sup h ( y ) , x ∈ R + . (1)
For locally finite unit-step profiles, an up-step, where the height rises by one from left to right, moves to the left with speed one, whereas a down-step moves to the right with speed one. A down-step followed by an up-step annihilates when the two steps meet. An up-step that reaches the origin is absorbed, so that the upper terrace has expanded to the boundary.
(ii) Bulk nucleations are the atoms of a Poisson point process Π b u l k \Pi_{\mathrm{bulk}} Π bulk on R > 0 × R > 0 \mathbb{R}_{>0} \times\mathbb{R}_{>0} R > 0 × R > 0 with intensity 2 d t d z 2\,\mathrm{d}t\,\mathrm{d}z 2 d t d z . At an atom ( t , z ) (t,z) ( t , z ) , the creation map C z C_z C z sets ( C z h ) ( z ) = h ( z ) + 1 (C_z h)(z)=h(z)+1 ( C z h ) ( z ) = h ( z ) + 1 and leaves the profile unchanged elsewhere. Where h ( z ) h(z) h ( z ) is finite, this adds a height-one upper spike. On a flat terrace it creates a coincident up-step/down-step pair; immediately afterwards the two steps separate and form an expanding island.
(iii) Independently of Π b u l k \Pi_{\mathrm{bulk}} Π bulk , boundary nucleations are the atoms of a Poisson process Π ∂ \Pi_{\partial} Π ∂ of rate c c c on R > 0 \mathbb{R}_{>0} R > 0 . At each boundary atom, the creation map C 0 C_0 C 0 sets ( C 0 h ) ( 0 ) = h ( 0 ) + 1 (C_0 h)(0)=h(0)+1 ( C 0 h ) ( 0 ) = h ( 0 ) + 1 and leaves the profile unchanged for x > 0 x>0 x > 0 . Where h ( 0 ) h(0) h ( 0 ) is finite, the raised boundary height propagates into the half-line under the deterministic evolution; on a flat terrace it emits a right-moving down-step. When an incoming up-step reaches the origin, it is absorbed, raising the height at 0 0 0 by one.
We use the convention − ∞ + 1 = − ∞ -\infty+1=-\infty − ∞ + 1 = − ∞ : an atom at a location whose current height is − ∞ -\infty − ∞ leaves the profile unchanged. Figure 1 illustrates these rules on finite terraces.
Figure 1. Local rules of half-space PNG. (a) A bulk atom creates a coincident up-step/down-step pair, which separates into an expanding island. (b) A down-step followed by an up-step annihilates on contact. (c) A boundary atom raises the height at the origin and emits a down-step, whereas an incoming up-step is absorbed at the origin.
The maps D s D_s D s , C z C_z C z , and C 0 C_0 C 0 preserve U C + \mathrm{UC}_+ UC + , and D s D r = D s + r D_sD_r=D_{s+r} D s D r = D s + r . Indeed, for each integer a a a , the superlevel set of D s h D_s h D s h is the closed s s s -neighbourhood of { h ≥ a } \{h\ge a\} { h ≥ a } in R + \mathbb{R}_+ R + , while a creation map either leaves a superlevel set unchanged or adds its creation location. To construct the evolution on [ 0 , R ] [0,R] [ 0 , R ] up to time T T T , only h 0 ∣ [ 0 , R + T ] h_0|_{[0,R+T]} h 0 ∣ [ 0 , R + T ] , bulk atoms in ( 0 , T ] × ( 0 , R + T ] (0,T]\times(0,R+T] ( 0 , T ] × ( 0 , R + T ] , and boundary atoms in ( 0 , T ] (0,T] ( 0 , T ] can contribute. Their total number N R , T N_{R,T} N R , T is almost surely finite, and the heights in this window are bounded above by sup [ 0 , R + T ] h 0 + N R , T \sup_{[0,R+T]}h_0+N_{R,T} sup [ 0 , R + T ] h 0 + N R , T . Ordering these atoms in time and alternating deterministic evolution with the creation maps therefore defines a pathwise evolution in U C + \mathrm{UC}_+ UC + . Finite propagation makes these constructions agree on overlapping windows. We denote the resulting height by h ( c ) ( t , x ; h 0 ) h^{(c)}(t,x;h_0) h ( c ) ( t , x ; h 0 ) , t , x ≥ 0 t,x\ge0 t , x ≥ 0 , for h 0 ∈ U C + h_0\in\mathrm{UC}_+ h 0 ∈ UC + . Evolving several profiles with one environment Π = ( Π b u l k , Π ∂ ) \Pi=(\Pi_{\mathrm{bulk}},\Pi_{\partial}) Π = ( Π bulk , Π ∂ ) defines the graphical, basic, coupling. We write h ( c ) ( t , x ; h 0 ; Π ) h^{(c)}(t,x;h_0;\Pi) h ( c ) ( t , x ; h 0 ; Π ) when specifying the environment.
The value at ( t , x ) (t,x) ( t , x ) depends on the initial profile only through the interval I t , x I_{t,x} I t , x and on the bulk atoms in the backward light cone C t , x C_{t,x} C t , x , where
I t , x : = [ ( x − t ) + , x + t ] , C t , x : = { ( s , z ) : 0 < s ≤ t , z ≥ 0 , ∣ z − x ∣ ≤ t − s } . I_{t,x}:=[(x-t)_+,x+t], \qquad C_{t,x}:=\{(s,z):0<s\le t,\ z\ge0,\ |z-x|\le t-s\}. I t , x := [( x − t ) + , x + t ] , C t , x := {( s , z ) : 0 < s ≤ t , z ≥ 0 , ∣ z − x ∣ ≤ t − s } .
Boundary atoms can contribute only for 0 < s ≤ t − x 0<s\le t-x 0 < s ≤ t − x , an interval of length ( t − x ) + (t-x)_+ ( t − x ) + .
Definition 1.3 (Narrow wedge and point-to-point field). For y ∈ R + y\in\mathbb{R}_+ y ∈ R + , the narrow wedge based at y y y is the profile ∂ y ∈ U C + \partial_y\in\mathrm{UC}_+ ∂ y ∈ UC + with ∂ y ( y ) = 0 \partial_y(y)=0 ∂ y ( y ) = 0 and ∂ y ( z ) = − ∞ \partial_y(z)=-\infty ∂ y ( z ) = − ∞ for z ≠ y z\ne y z = y . The point-to-point half-space PNG field is A t ( c ) ( x , y ) : = h ( c ) ( t , x ; ∂ y ) A_t^{(c)}(x,y):=h^{(c)}(t,x;\partial_y) A t ( c ) ( x , y ) := h ( c ) ( t , x ; ∂ y ) . We write A t ( c ) ( x , y ; Π ) A_t^{(c)}(x,y;\Pi) A t ( c ) ( x , y ; Π ) when specifying the graphical environment.
Equivalently, A t ( c ) ( x , y ) A_t^{(c)}(x,y) A t ( c ) ( x , y ) is the largest number of bulk and boundary atoms that can be collected by a path from ( y , 0 ) (y,0) ( y , 0 ) to ( x , t ) (x,t) ( x , t ) which remains in the half-space and whose spatial coordinate is 1 1 1 -Lipschitz in time. Collected atoms have time coordinates in ( 0 , t ] (0,t] ( 0 , t ] . If ∣ x − y ∣ > t |x-y|>t ∣ x − y ∣ > t , there is no such path and we set A t ( c ) ( x , y ) = − ∞ A_t^{(c)}(x,y)=-\infty A t ( c ) ( x , y ) = − ∞ . A feasible path that collects no atoms has score zero.
Each of the maps D s D_s D s , C z C_z C z , and C 0 C_0 C 0 preserves pointwise suprema whenever the supremum belongs to U C + \mathrm{UC}_+ UC + , and commutes with integer vertical shifts. The profile identically equal to − ∞ -\infty − ∞ is absorbing. Applying these identities to the pointwise decomposition h 0 = sup y ≥ 0 { h 0 ( y ) + ∂ y } h_0=\sup_{y\ge0}\{h_0(y)+\partial_y\} h 0 = sup y ≥ 0 { h 0 ( y ) + ∂ y } through the finitely many updates in a dependence cone gives the pathwise variational formula
h ( c ) ( t , x ; h 0 ) = sup y ∈ I t , x { h 0 ( y ) + A t ( c ) ( x , y ) } . (2) h^{(c)}(t,x;h_0)=\sup_{y\in I_{t,x}}\{h_0(y)+A_t^{(c)}(x,y)\}. \tag*{(2)} h ( c ) ( t , x ; h 0 ) = y ∈ I t , x sup { h 0 ( y ) + A t ( c ) ( x , y )} . (2)
Thus a general initial profile is propagated by adding to each possible starting height h 0 ( y ) h_0(y) h 0 ( y ) the point-to-point passage value from y y y , and then selecting the largest contribution. Figure 2 illustrates (1.2) for flat initial data.
Stationary measures
We formulate stationarity after pinning at the origin. Write U C + , 0 : = { h ∈ U C + : h ( 0 ) = 0 } \mathrm{UC}_{+,0}:=\{h\in\mathrm{UC}_+:h(0)=0\} UC + , 0 := { h ∈ UC + : h ( 0 ) = 0 } , Rec ( h ) : = h − h ( 0 ) \operatorname{Rec}(h):=h-h(0) Rec ( h ) := h − h ( 0 ) , where recentering is defined when h ( 0 ) h(0) h ( 0 ) is finite. Equip U C + , 0 \mathrm{UC}_{+,0} UC + , 0 with its induced Borel σ \sigma σ -field.
Figure 2. Last passage from the flat profile h 0 ≡ 0 h_0 \equiv0 h 0 ≡ 0 . All atoms in the backward cone are shown in coordinates ( z , s ) (z,s) ( z , s ) . Each displayed path maximizes from its starting point; the blue path also maximizes over the starting location.
Definition 1.4 (Stationary measure). Fix c ≥ 0 c \ge0 c ≥ 0 . A law ν \nu ν on U C + , 0 \mathrm{UC}_{+,0} UC + , 0 is stationary for half-space PNG if, for H ∼ ν H \sim\nu H ∼ ν independent of a future graphical environment Π \Pi Π ,
Rec ( h ( c ) ( t , ⋅ ; H ; Π ) ) = d H , t ≥ 0. \operatorname{Rec}(h^{(c)}(t,\cdot;H;\Pi)) \overset{\mathrm{d}}{=} H,\qquad t \ge0. Rec ( h ( c ) ( t , ⋅ ; H ; Π )) = d H , t ≥ 0. Set r c : = c ∨ 1 r_c := c \vee1 r c := c ∨ 1 , G c : = [ r c , ∞ ) \mathcal{G}_c := [r_c,\infty) G c := [ r c , ∞ ) , and N 0 : = { 0 , 1 , … } \mathbb{N}_0 := \{0,1,\ldots\} N 0 := { 0 , 1 , … } . Our geometric convention is P ( Geo ( q ) = n ) = ( 1 − q ) q n \mathbb{P}(\operatorname{Geo}(q)=n)=(1-q)q^n P ( Geo ( q ) = n ) = ( 1 − q ) q n for q ∈ [ 0 , 1 ) q\in[0,1) q ∈ [ 0 , 1 ) and n ∈ N 0 n\in\mathbb{N}_0 n ∈ N 0 . A signed Poisson walk with up/down rates ( a , b ) (a,b) ( a , b ) is N + − N − N^+ - N^- N + − N − , where N + N^+ N + , N − N^- N − are independent Poisson processes with rates a , b a,b a , b , respectively, started at zero.
For a spectral parameter g ∈ G c g \in\mathcal{G}_c g ∈ G c , let S 1 ( g ) , S 2 ( g ) S_1^{(g)},S_2^{(g)} S 1 ( g ) , S 2 ( g ) be independent signed Poisson walks with respective up/down rates ( g , g − 1 ) (g,g^{-1}) ( g , g − 1 ) and ( g − 1 , g ) (g^{-1},g) ( g − 1 , g ) . If g > c g>c g > c , independently take Y ( c , g ) ∼ Geo ( c / g ) Y^{(c,g)}\sim\operatorname{Geo}(c/g) Y ( c , g ) ∼ Geo ( c / g ) and set
H ~ ( c , g ) ( x ) : = S 2 ( g ) ( x ) + ( sup 0 ≤ y ≤ x { S 1 ( g ) ( y ) − S 2 ( g ) ( y ) } − Y ( c , g ) ) . (3) \widetilde{H}^{(c,g)}(x):=S_2^{(g)}(x)+\left(\sup_{0\le y\le x}\{S_1^{(g)}(y)-S_2^{(g)}(y)\}-Y^{(c,g)}\right). \tag*{(3)} H ( c , g ) ( x ) := S 2 ( g ) ( x ) + ( 0 ≤ y ≤ x sup { S 1 ( g ) ( y ) − S 2 ( g ) ( y )} − Y ( c , g ) ) . (3)
For locally bounded f f f , its upper-semicontinuous regularization is
( usc f ) ( x ) : = lim ε ↓ 0 sup { f ( y ) : y ≥ 0 , ∣ y − x ∣ < ε } . (\operatorname{usc}f)(x):=\lim_{\varepsilon\downarrow0}\sup\{f(y):y\ge0,\ |y-x|<\varepsilon\}. ( usc f ) ( x ) := ε ↓ 0 lim sup { f ( y ) : y ≥ 0 , ∣ y − x ∣ < ε } .
Define the profiles and their laws by
H ( c , g ) : = usc H ~ ( c , g ) , g > c , (4) H^{(c,g)}:=\operatorname{usc}\widetilde{H}^{(c,g)},\qquad g>c, \tag*{(4)} H ( c , g ) := usc H ( c , g ) , g > c , (4)
H ( c , c ) : = usc S 2 ( c ) , c ≥ 1 , (5) H^{(c,c)}:=\operatorname{usc}S_2^{(c)},\qquad c\ge1, \tag*{(5)} H ( c , c ) := usc S 2 ( c ) , c ≥ 1 , (5)
and ν c , g : = Law ( H ( c , g ) ) \nu_{c,g}:=\operatorname{Law}(H^{(c,g)}) ν c , g := Law ( H ( c , g ) ) .
Theorem 1.5 (A stationary family for half-space PNG). For every c ≥ 0 c\ge0 c ≥ 0 and g ∈ G c g\in\mathcal{G}_c g ∈ G c , the law ν c , g \nu_{c,g} ν c , g is stationary for half-space PNG with boundary rate c c c .
Section 2.1 proves this by generator balance.
Definition 1.6 (Jointly stationary measure). Fix c ≥ 0 c\ge0 c ≥ 0 and k ∈ N k\in\mathbb{N} k ∈ N . A law ν \nu ν on ( U C + , 0 ) k (\mathrm{UC}_{+,0})^k ( UC + , 0 ) k is jointly stationary if, for ( H 1 , … , H k ) ∼ ν (H_1,\ldots,H_k)\sim\nu ( H 1 , … , H k ) ∼ ν independent of one common future environment Π \Pi Π ,
( Rec ( h ( c ) ( t , ⋅ ; H i ; Π ) ) ) i = 1 k = d ( H i ) i = 1 k , t ≥ 0. \left(\operatorname{Rec}(h^{(c)}(t,\cdot;H_i;\Pi))\right)_{i=1}^k \overset{\mathrm{d}}{=} (H_i)_{i=1}^k,\qquad t\ge0. ( Rec ( h ( c ) ( t , ⋅ ; H i ; Π )) ) i = 1 k = d ( H i ) i = 1 k , t ≥ 0.
For g 1 > ⋯ > g k ≥ r c g_1>\cdots>g_k\ge r_c g 1 > ⋯ > g k ≥ r c , write g = ( g 1 , … , g k ) \mathbf{g}=(g_1,\ldots,g_k) g = ( g 1 , … , g k ) . When g k > c g_k>c g k > c , let ( R i ( c ; g ) ) i = 1 k \left(R_i^{(c;\mathbf{g})}\right)_{i=1}^k ( R i ( c ; g ) ) i = 1 k be the profiles of Definition 2.6; Figure 3 illustrates their shared line and bridge environment for k = 2 k=2 k = 2 . At g k = c ≥ 1 g_k=c\ge1 g k = c ≥ 1 , omit the singular terminal bridge and use the shifted source in (2.8). For an arbitrary list g 1 , … , g k ∈ G c g_1,\ldots,g_k\in\mathcal{G}_c g 1 , … , g k ∈ G c , let a 1 > ⋯ > a ℓ a_1>\cdots>a_\ell a 1 > ⋯ > a ℓ be its distinct values, set a = ( a 1 , … , a ℓ ) \mathbf{a}=(a_1,\ldots,a_\ell) a = ( a 1 , … , a ℓ ) , and define κ ( i ) \kappa(i) κ ( i ) by g i = a κ ( i ) g_i=a_{\kappa(i)} g i = a κ ( i ) . Set
Figure 3. The shared k = 2 k=2 k = 2 environment. Squares are bridges; the highlighted paths run from the two sources to ( x , 4 ) (x,4) ( x , 4 ) . The signed Poisson walks are schematic.
ν c ; g 1 , … , g k : = Law ( ( R κ ( i ) ( c ; a ) ) i = 1 k ) . \nu_{c;g_1,\ldots,g_k}:=\operatorname{Law}\left(\left(R_{\kappa(i)}^{(c;\mathbf{a})}\right)_{i=1}^k\right). ν c ; g 1 , … , g k := Law ( ( R κ ( i ) ( c ; a ) ) i = 1 k ) .
Theorem 1.7 (A jointly stationary family for half-space PNG). For every c ≥ 0 c\ge0 c ≥ 0 , k ∈ N k\in\mathbb{N} k ∈ N , and g 1 , … , g k ∈ G c g_1,\ldots,g_k\in\mathcal{G}_c g 1 , … , g k ∈ G c , the law ν c ; g 1 , … , g k \nu_{c;g_1,\ldots,g_k} ν c ; g 1 , … , g k is jointly stationary for half-space PNG with boundary rate c c c . Its i i i -th marginal is ν c , g i \nu_{c,g_i} ν c , g i , and the family is consistent under deleting, permuting, and repeating parameters.
Remark 1.8 (Conjecture). We expect that the stationary laws in Theorem 1.5 are all the extremal stationary measures of half-space PNG, and that the jointly stationary laws in Theorem 1.7 are all the extremal jointly stationary measures under the basic coupling. Proving these classifications requires further work, which we do not pursue here. For the corresponding classifications in half-space geometric last-passage percolation, see [13 ] .
Section 2.3 constructs a jointly measurable field whose finite-parameter distributions are the jointly stationary laws ν c ; g 1 , … , g k \nu_{c;g_1,\ldots,g_k} ν c ; g 1 , … , g k . Uniqueness below concerns regular representations of this family in the sense of Definition 2.20 .
Theorem 1.9 (Half-space PNG horizon). For every c ≥ 0 c \ge0 c ≥ 0 , there exists a regular representation ( b c P N G , μ c P N G ) (b_c^{\mathrm{PNG}},\mu_c^{\mathrm{PNG}}) ( b c PNG , μ c PNG ) , unique in law, with the following properties.
(i) Construction. The base profile b c P N G b_c^{\mathrm{PNG}} b c PNG is a pinned, locally finite, integer-valued càdlàg step function. The measure μ c P N G \mu_c^{\mathrm{PNG}} μ c PNG is a positive locally finite Borel measure on G c × R + \mathcal{G}_c \times\mathbb{R}_+ G c × R + , zero on { r c } × R + \{r_c\} \times\mathbb{R}_+ { r c } × R + and G c × { 0 } \mathcal{G}_c \times\{0\} G c × { 0 } . Define
H c P N G , → ( g , x ) : = b c P N G ( x ) + μ c P N G ( ( r c , g ] × ( 0 , x ] ) , H c P N G ( g , ⋅ ) : = usc H c P N G , → ( g , ⋅ ) . H_c^{\mathrm{PNG},\to}(g,x) := b_c^{\mathrm{PNG}}(x) + \mu_c^{\mathrm{PNG}}\bigl((r_c,g] \times(0,x]\bigr), \qquad H_c^{\mathrm{PNG}}(g,\cdot) := \operatorname{usc} H_c^{\mathrm{PNG},\to}(g,\cdot). H c PNG , → ( g , x ) := b c PNG ( x ) + μ c PNG ( ( r c , g ] × ( 0 , x ] ) , H c PNG ( g , ⋅ ) := usc H c PNG , → ( g , ⋅ ) . We call H c P N G H_c^{\mathrm{PNG}} H c PNG the half-space PNG horizon.
(ii) Finite-parameter laws and regularity. The field H c P N G H_c^{\mathrm{PNG}} H c PNG is jointly measurable. For every deterministic finite list g 1 , … , g k ∈ G c g_1,\ldots,g_k \in\mathcal{G}_c g 1 , … , g k ∈ G c ,
Law ( H c P N G ( g 1 , ⋅ ) , … , H c P N G ( g k , ⋅ ) ) = ν c ; g 1 , … , g k . \operatorname{Law}\bigl(H_c^{\mathrm{PNG}}(g_1,\cdot),\ldots,H_c^{\mathrm{PNG}}(g_k,\cdot)\bigr) = \nu_{c;g_1,\ldots,g_k}. Law ( H c PNG ( g 1 , ⋅ ) , … , H c PNG ( g k , ⋅ ) ) = ν c ; g 1 , … , g k . Almost surely, every coordinate is a locally finite unit-step profile in U C + , 0 \mathrm{UC}_{+,0} UC + , 0 , and H c P N G , → ( g n , ⋅ ) H_c^{\mathrm{PNG},\to}(g_n,\cdot) H c PNG , → ( g n , ⋅ ) converges locally uniformly to H c P N G , → ( g , ⋅ ) H_c^{\mathrm{PNG},\to}(g,\cdot) H c PNG , → ( g , ⋅ ) whenever g n ↓ g g_n \downarrow g g n ↓ g .
(iii) Stationarity. Let Π \Pi Π be a future graphical environment independent of H c P N G H_c^{\mathrm{PNG}} H c PNG . For each fixed t ≥ 0 t \ge0 t ≥ 0 , define
H c , t e v ( g , ⋅ ) : = Rec ( h c ( t , ⋅ ; H c P N G ( g , ⋅ ) ; Π ) ) . H_{c,t}^{\mathrm{ev}}(g,\cdot) := \operatorname{Rec}\bigl(h^{\mathrm{c}}(t,\cdot;H_c^{\mathrm{PNG}}(g,\cdot);\Pi)\bigr). H c , t ev ( g , ⋅ ) := Rec ( h c ( t , ⋅ ; H c PNG ( g , ⋅ ) ; Π ) ) . The coordinates indexed by Q c : = { r c } ∪ ( Q ∩ ( r c , ∞ ) ) \mathbb{Q}_c := \{r_c\} \cup(\mathbb{Q} \cap(r_c,\infty)) Q c := { r c } ∪ ( Q ∩ ( r c , ∞ )) determine a canonical regular reconstruction H ^ c , t P N G \widehat{H}_{c,t}^{\mathrm{PNG}} H c , t PNG . For each fixed deterministic g ∈ G c g \in\mathcal{G}_c g ∈ G c ,
H ^ c , t P N G ( g , ⋅ ) = H c , t e v ( g , ⋅ ) almost surely , H ^ c , t P N G = d H c P N G in the ( b , μ ) representation . (6) \widehat{H}_{c,t}^{\mathrm{PNG}}(g,\cdot) = H_{c,t}^{\mathrm{ev}}(g,\cdot) \quad\text{almost surely}, \qquad\widehat{H}_{c,t}^{\mathrm{PNG}} \overset{\mathrm{d}}{=} H_c^{\mathrm{PNG}} \quad\text{in the }(b,\mu)\text{ representation}. \tag*{(6)} H c , t PNG ( g , ⋅ ) = H c , t ev ( g , ⋅ ) almost surely , H c , t PNG = d H c PNG in the ( b , μ ) representation . (6)
KPZ universality
We prove 1 : 2 : 3 1:2:3 1 : 2 : 3 limits for the stationary horizon, the height evolution from general initial data, and the common passage field. Use the boundary convention of [17 ] and set
c ε ( ρ ) = e ρ ε 1 / 2 , λ ρ = ( 2 ρ ) + ( ρ ∈ R ) , c ε ( − ∞ ) = c ⋆ ∈ [ 0 , 1 ) , λ − ∞ = 0. (7) c_\varepsilon(\rho) = e^{\rho\varepsilon^{1/2}}, \qquad\lambda_\rho= (2\rho)_+ \quad(\rho\in\mathbb{R}), \qquad c_\varepsilon(-\infty) = c_\star\in[0,1), \qquad\lambda_{-\infty} = 0. \tag*{(7)} c ε ( ρ ) = e ρ ε 1/2 , λ ρ = ( 2 ρ ) + ( ρ ∈ R ) , c ε ( − ∞ ) = c ⋆ ∈ [ 0 , 1 ) , λ − ∞ = 0. (7)
Every fixed c ⋆ ∈ [ 0 , 1 ) c_\star\in[0,1) c ⋆ ∈ [ 0 , 1 ) has the same subcritical limit. The critical window includes rates above one when ρ > 0 \rho> 0 ρ > 0 ; a fixed rate c > 1 c > 1 c > 1 is outside these scaling limits. Zhang’s parameter in [32 ] is − ρ -\rho − ρ ; the affine boundary parameter in [10 ] , § 1.3 is − ρ + O ( ε 1 / 2 ) -\rho+ O(\varepsilon^{1/2}) − ρ + O ( ε 1/2 ) . For the stationary family, scale the spectral parameter g g g by:
g ε ( λ ) = e λ ε 1 / 2 / 2 , λ ≥ λ ρ , (8) g_\varepsilon(\lambda) = e^{\lambda\varepsilon^{1/2}/2}, \qquad\lambda\ge\lambda_\rho, \tag*{(8)} g ε ( λ ) = e λ ε 1/2 /2 , λ ≥ λ ρ , (8)
equivalently g ε ( λ ) ≥ c ε ( ρ ) ∨ 1 g_\varepsilon(\lambda) \ge c_\varepsilon(\rho) \vee1 g ε ( λ ) ≥ c ε ( ρ ) ∨ 1 . Write H ρ \mathcal{H}^\rho H ρ for the half-space stationary horizon of [17 ] , Theorem 2.19, restricted here to x ≥ 0 x \ge0 x ≥ 0 , and L h \mathcal{L}^{\mathrm{h}} L h for the half-space directed landscape and KPZ fixed point.
For the passage and height fields below, local uniform convergence is understood through continuous mesh interpolation, with uninterpolated and interpolated fields asymptotically uniformly close on each compact set. We write C l o c ( D ) C_{\mathrm{loc}}(D) C loc ( D ) for continuous real-valued functions on D D D with uniform convergence on compact subsets, and C c ( D ) C_c(D) C c ( D ) for continuous functions with compact support in D D D . For a scalar function f f f and a compact set K K K , write ∥ f ∥ K : = sup x ∈ K ∣ f ( x ) ∣ \lVert f\rVert_K := \sup_{x \in K} |f(x)| ∥ f ∥ K := sup x ∈ K ∣ f ( x ) ∣ . We write ⇒ \Rightarrow ⇒ for convergence in distribution.
Convergence to the half-space stationary horizon
For x = ε ( n + θ ) x = \varepsilon(n+\theta) x = ε ( n + θ ) , n ∈ N 0 n \in\mathbb{N}_0 n ∈ N 0 , 0 ≤ θ < 1 0 \le\theta< 1 0 ≤ θ < 1 , define the linear interpolation by ( Lin ε h ) ( x ) = ( 1 − θ ) h ( n ) + θ h ( n + 1 ) (\operatorname{Lin}_\varepsilon h)(x) = (1-\theta)h(n) + \theta h(n+1) ( Lin ε h ) ( x ) = ( 1 − θ ) h ( n ) + θ h ( n + 1 ) , and set
H ε P N G , ρ ( λ , x ) = ε 1 / 2 ( Lin ε H c ε ( ρ ) P N G , → ( g ε ( λ ) , ⋅ ) ) ( x ) . \mathcal{H}_\varepsilon^{\mathrm{PNG},\rho}(\lambda,x) = \varepsilon^{1/2}\bigl(\operatorname{Lin}_\varepsilon H_{c_\varepsilon(\rho)}^{\mathrm{PNG},\to}(g_\varepsilon(\lambda),\cdot)\bigr)(x). H ε PNG , ρ ( λ , x ) = ε 1/2 ( Lin ε H c ε ( ρ ) PNG , → ( g ε ( λ ) , ⋅ ) ) ( x ) .
Here H c P N G , → H_c^{\mathrm{PNG},\to} H c PNG , → is the canonical right-continuous representative from Theorem 1.9 . The space G ( λ ) \mathcal{G}(\lambda) G ( λ ) consists of pairs ( b , μ ) (b,\mu) ( b , μ ) , where b b b is continuous with b ( 0 ) = 0 b(0)=0 b ( 0 ) = 0 and μ \mu μ is a positive locally finite measure on [ λ ρ , ∞ ) × [ 0 , ∞ ) [\lambda_\rho,\infty) \times[0,\infty) [ λ ρ , ∞ ) × [ 0 , ∞ ) with zero mass on the two lower faces. Identify the pair with G ( λ , x ) = b ( x ) + μ ( [ λ ρ , λ ] × ( 0 , x ] ) G(\lambda,x) = b(x) + \mu([\lambda_\rho,\lambda] \times(0,x]) G ( λ , x ) = b ( x ) + μ ([ λ ρ , λ ] × ( 0 , x ]) . Its topology is local uniform convergence of b b b and vague convergence of the bivariate Stieltjes measure μ \mu μ . This is Definition 2.13 restricted to x ≥ 0 x \ge0 x ≥ 0 , with zero mass assigned to the two lower faces.
Theorem 1.10 (Stationary horizon limit). For ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R} \cup\{-\infty\} ρ ∈ R ∪ { − ∞ } , H ε P N G , ρ \mathcal{H}_\varepsilon^{\mathrm{PNG},\rho} H ε PNG , ρ is a Borel G ( λ ρ ) \mathcal{G}(\lambda_\rho) G ( λ ρ ) -valued random element and
H ε P N G , ρ ⇒ H ρ ∣ [ λ ρ , ∞ ) × R + in G ( λ ρ ) . \mathcal{H}_\varepsilon^{\mathrm{PNG},\rho} \Rightarrow\mathcal{H}^{\rho}\big|_{[\lambda_\rho,\infty)\times\mathbb{R}_+} \quad\text{in }\mathcal{G}(\lambda_\rho). H ε PNG , ρ ⇒ H ρ [ λ ρ , ∞ ) × R + in G ( λ ρ ) . For every λ 1 > ⋯ > λ k ≥ λ ρ \lambda_1 > \cdots> \lambda_k \ge\lambda_\rho λ 1 > ⋯ > λ k ≥ λ ρ , also
( H ε P N G , ρ ( λ i , ⋅ ) ) i = 1 k ⇒ ( H ρ ( λ i , ⋅ ) ) i = 1 k in C l o c ( [ 0 , ∞ ) ) k . \bigl(\mathcal{H}_\varepsilon^{\mathrm{PNG},\rho}(\lambda_i,\cdot)\bigr)_{i=1}^k \Rightarrow\bigl(\mathcal{H}^{\rho}(\lambda_i,\cdot)\bigr)_{i=1}^k \quad\text{in } C_{\mathrm{loc}}([0,\infty))^k. ( H ε PNG , ρ ( λ i , ⋅ ) ) i = 1 k ⇒ ( H ρ ( λ i , ⋅ ) ) i = 1 k in C loc ([ 0 , ∞ ) ) k .
The proof is in § 3.1 .
Convergence to the half-space KPZ fixed point
Let U C + R \mathrm{UC}^{\mathbb{R}}_+ UC + R consist of locally bounded above, upper-semicontinuous functions [ 0 , ∞ ) → R ∪ { − ∞ } [0,\infty) \to\mathbb{R} \cup\{-\infty\} [ 0 , ∞ ) → R ∪ { − ∞ } not identically − ∞ -\infty − ∞ . Local hypograph convergence means that x ε → x x_\varepsilon\to x x ε → x implies lim sup f ε ( x ε ) ≤ f ( x ) \limsup f_\varepsilon(x_\varepsilon) \le f(x) lim sup f ε ( x ε ) ≤ f ( x ) , and each finite f ( x ) f(x) f ( x ) has a recovery sequence attaining its value in the limit. For f f f with a linear upper bound and ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R} \cup\{-\infty\} ρ ∈ R ∪ { − ∞ } , let h ρ ( t , y ; f ) h^\rho(t,y;f) h ρ ( t , y ; f ) denote the half-space KPZ fixed point started from f f f , realized as a continuous Markov process in [17 ] , Theorem 4.11. Its transition laws are the Fredholm Pfaffians of [32 ] , recalled in § 3.4.1 for finite ρ \rho ρ , and their subcritical limit for ρ = − ∞ \rho=-\infty ρ = − ∞ . Let X a d m ⊂ U C + R \mathcal{X}_{\mathrm{adm}} \subset\mathrm{UC}^{\mathbb{R}}_+ X adm ⊂ UC + R be the state space of [10 ] , § 1.2 : on every [ 0 , R o b s ] [0,R_{\mathrm{obs}}] [ 0 , R obs ] , a profile is constant on the open intervals of some finite partition. It contains the unit-step class X s t e p \mathcal{X}_{\mathrm{step}} X step . of Definition 2.1 and hence supports every stationary law ν c , g \nu_{c,g} ν c , g . For deterministic h 0 , ε ∈ X a d m h_{0,\varepsilon} \in\mathcal{X}_{\mathrm{adm}} h 0 , ε ∈ X adm , set f ε ( x ) = ε 1 / 2 h 0 , ε ( ε − 1 x ) f_\varepsilon(x) = \varepsilon^{1/2}h_{0,\varepsilon}(\varepsilon^{-1}x) f ε ( x ) = ε 1/2 h 0 , ε ( ε − 1 x ) and
h ε ρ ( t , y ) = ε 1 / 2 [ h ( c ( ε ) ) ( ε − 3 / 2 t , ε − 1 y ; h 0 , ε ) − 2 ε − 3 / 2 t ] . h^\rho_\varepsilon(t,y)=\varepsilon^{1/2}\left[h^{(c(\varepsilon))}\left(\varepsilon^{-3/2}t,\varepsilon^{-1}y;h_{0,\varepsilon}\right)-2\varepsilon^{-3/2}t\right]. h ε ρ ( t , y ) = ε 1/2 [ h ( c ( ε )) ( ε − 3/2 t , ε − 1 y ; h 0 , ε ) − 2 ε − 3/2 t ] .
Theorem 1.11 (Half-space KPZ fixed-point limit). Fix ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R} \cup\{-\infty\} ρ ∈ R ∪ { − ∞ } . Suppose f ε → f ∈ U C + R f_\varepsilon\to f \in\mathrm{UC}_{+}^{\mathbb{R}} f ε → f ∈ UC + R locally in hypograph and f ε ( x ) ≤ C ( 1 + x ) f_\varepsilon(x) \le C(1+x) f ε ( x ) ≤ C ( 1 + x ) for one finite C ≥ 0 C \ge0 C ≥ 0 . Then
h ε ρ ⟹ h ρ ( ⋅ , ⋅ ; f ) locally uniformly on ( 0 , ∞ ) × [ 0 , ∞ ) . h^\rho_\varepsilon\Longrightarrow h^\rho(\cdot,\cdot;f)\quad\text{locally uniformly on }(0,\infty)\times[0,\infty). h ε ρ ⟹ h ρ ( ⋅ , ⋅ ; f ) locally uniformly on ( 0 , ∞ ) × [ 0 , ∞ ) .
The proof is in § 3.4 .
Convergence to the half-space directed landscape
Take bulk atoms of intensity 2 d s d z 2\,ds\,dz 2 d s d z on R × ( 0 , ∞ ) \mathbb{R}\times(0,\infty) R × ( 0 , ∞ ) and independent rate-c c c boundary atoms on R × { 0 } \mathbb{R}\times\{0\} R × { 0 } , with union Π ( c ) \Pi^{(c)} Π ( c ) . Put
H = [ 0 , ∞ ) × R , H ↑ 2 = { ( x , s ; y , t ) ∈ H 2 : s < t } . \mathbb{H}=[0,\infty)\times\mathbb{R},\qquad\mathbb{H}^{2}_{\uparrow}=\{(x,s;y,t)\in\mathbb{H}^{2}:s<t\}. H = [ 0 , ∞ ) × R , H ↑ 2 = {( x , s ; y , t ) ∈ H 2 : s < t } .
Definition 1.12 (Half-space PNG directed metric). With ϑ s Π = { ( r − s , z ) : ( r , z ) ∈ Π } \vartheta_s\Pi=\{(r-s,z):(r,z)\in\Pi\} ϑ s Π = {( r − s , z ) : ( r , z ) ∈ Π } and the passage field of Definition 1.3, set
L P N G ( c ) ( x , s ; y , t ) = A t − s ( c ) ( y , x ; ϑ s Π ( c ) ) . \mathcal{L}^{(c)}_{\mathrm{PNG}}(x,s;y,t)=\mathcal{A}^{(c)}_{t-s}(y,x;\vartheta_s\Pi^{(c)}). L PNG ( c ) ( x , s ; y , t ) = A t − s ( c ) ( y , x ; ϑ s Π ( c ) ) .
It counts the greatest number of distinct atoms in ( s , t ] (s,t] ( s , t ] collected by a 1-Lipschitz path γ : [ s , t ] → [ 0 , ∞ ) \gamma:[s,t]\to[0,\infty) γ : [ s , t ] → [ 0 , ∞ ) with γ ( s ) = x \gamma(s)=x γ ( s ) = x , γ ( t ) = y \gamma(t)=y γ ( t ) = y , and is − ∞ -\infty − ∞ if ∣ x − y ∣ > t − s |x-y|>t-s ∣ x − y ∣ > t − s .
For passage from time s s s to time t t t , split an admissible path at its position z z z at time r r r . Any atom at time r r r is counted in the first segment and excluded from the second. Maximizing over the segments and their meeting position gives
L P N G ( c ) ( x , s ; y , t ) = max z ≥ 0 { L P N G ( c ) ( x , s ; z , r ) + L P N G ( c ) ( z , r ; y , t ) } , s < r < t . (9) \mathcal{L}^{(c)}_{\mathrm{PNG}}(x,s;y,t)=\max_{z\ge0}\{\mathcal{L}^{(c)}_{\mathrm{PNG}}(x,s;z,r)+\mathcal{L}^{(c)}_{\mathrm{PNG}}(z,r;y,t)\},\qquad s<r<t. \tag*{(9)} L PNG ( c ) ( x , s ; y , t ) = z ≥ 0 max { L PNG ( c ) ( x , s ; z , r ) + L PNG ( c ) ( z , r ; y , t )} , s < r < t . (9)
Define
L ε ρ ( x , s ; y , t ) = ε 1 / 2 L P N G ( c ( ε ) ) ( ε − 1 x , ε − 3 / 2 s ; ε − 1 y , ε − 3 / 2 t ) − 2 ε − 1 ( t − s ) . L^\rho_\varepsilon(x,s;y,t)=\varepsilon^{1/2}\mathcal{L}^{(c(\varepsilon))}_{\mathrm{PNG}}(\varepsilon^{-1}x,\varepsilon^{-3/2}s;\varepsilon^{-1}y,\varepsilon^{-3/2}t)-2\varepsilon^{-1}(t-s). L ε ρ ( x , s ; y , t ) = ε 1/2 L PNG ( c ( ε )) ( ε − 1 x , ε − 3/2 s ; ε − 1 y , ε − 3/2 t ) − 2 ε − 1 ( t − s ) .
The max-plus representation gives the exact identity
h ε ρ ( t , y ) = sup x ≥ 0 { f ε ( x ) + L ε ρ ( x , 0 ; y , t ) } . h^\rho_\varepsilon(t,y)=\sup_{x\ge0}\{f_\varepsilon(x)+L^\rho_\varepsilon(x,0;y,t)\}. h ε ρ ( t , y ) = x ≥ 0 sup { f ε ( x ) + L ε ρ ( x , 0 ; y , t )} .
Theorem 1.13 (Half-space directed landscape limit). For every ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R}\cup\{-\infty\} ρ ∈ R ∪ { − ∞ } , there are continuous mesh interpolations L ~ ε ρ \widetilde L^\rho_\varepsilon L ε ρ such that
L ~ ε ρ ⟹ L ρ in C l o c ( H ↑ 2 ) , ∥ L ~ ε ρ − L ε ρ ∥ K ⟶ 0 in probability \widetilde L^\rho_\varepsilon\Longrightarrow L^\rho\quad\text{in }C_{\mathrm{loc}}(\mathbb{H}^{2}_{\uparrow}),\qquad\|\widetilde L^\rho_\varepsilon-L^\rho_\varepsilon\|_K\longrightarrow0\quad\text{in probability} L ε ρ ⟹ L ρ in C loc ( H ↑ 2 ) , ∥ L ε ρ − L ε ρ ∥ K ⟶ 0 in probability
for every compact K ⊂ H ↑ 2 K\subset\mathbb{H}^{2}_{\uparrow} K ⊂ H ↑ 2 . Thus the raw metrics converge locally uniformly in the convention above. For ρ = − ∞ \rho=-\infty ρ = − ∞ , this holds for each fixed c ⋆ ∈ [ 0 , 1 ) c_\star\in[0,1) c ⋆ ∈ [ 0 , 1 ) .
The centering retains the parabola − ( x − y ) 2 / ( t − s ) -(x-y)^2/(t-s) − ( x − y ) 2 / ( t − s ) in the limit. § 3.6 proves this theorem using the finite-wedge specialization of the transition limits in § 3.4 . The half-space KPZ fixed point has the variational representation [17 Theorem 4.11] :
h ρ ( t , y ; f ) = sup x ≥ 0 { f ( x ) + L ρ ( x , 0 ; y , t ) } . (10) h^\rho(t,y;f)=\sup_{x\ge0}\{f(x)+L^\rho(x,0;y,t)\}. \tag*{(10)} h ρ ( t , y ; f ) = x ≥ 0 sup { f ( x ) + L ρ ( x , 0 ; y , t )} . (10)
Corollary 3.27 gives joint convergence of the metric and finitely many height evolutions driven by the same Poisson environment. Every limiting height field is given by this variational formula with the same L ρ L^\rho L ρ .
Proof strategy and organization
§ 2 encodes the one-profile law by a reflected spatial walk and proves generator balance at the wall. The joint construction combines geometric last-passage stationarity and symmetries from [17 ] with a finite line rearrangement and a limit retaining one common future. The singular endpoint uses a shifted source, and planar crossing gives the regular PNG horizon. § 3.1 proves its limit by an invariance principle for the lines and bridges and convergence of the representing Stieltjes measures.
§ 3.4 proves uniform boundary and hitting estimates for the exact transition Pfaffian [10 Theorem 1.1] . Trace and Hilbert–Schmidt bounds justify its limit to Zhang’s transition law [32 ] . Fixed subcritical rates require a skew derivative form and stationary comparison at the wall, where an individual kernel block has only a distributional limit. Localization, profile tightness, Markov iteration, and short-time bounds give the height-field limit.
§ 3.6 uses finite-wedge tests and [17 Theorem 5.1 and Corollary 5.2] to identify the rational metric law, then proves four-variable tightness. The exact variational formula and localization identify the joint height limits as functions of that same metric.
AI assistance
The main ideas and overall research direction of this work originated with the author. The author acknowledges substantial assistance from OpenAI’s ChatGPT in research discussions, proof development, and manuscript preparation, and takes full responsibility for the mathematical arguments and the final manuscript.
Acknowledgements
The author thanks the organizers of the Virginia Integrable Probability Summer School 2024 and the workshop “All roads to the KPZ universality class” at the American Institute of Mathematics for providing stimulating research environments. The author is grateful to Xincheng Zhang and Lingfu Zhang for helpful discussions.
Stationary measures
We first prove stationarity of the one-parameter laws ν c , g \nu_{c,g} ν c , g by a generator calculation in § 2.1 . § 2.3 constructs the laws ν c ; g 1 , … , g k \nu_{c;g_1,\ldots,g_k} ν c ; g 1 , … , g k and proves their joint stationarity under a common graphical environment, together with parameter-deletion consistency. Finally, § 2.6 assembles this family into a canonical measurable field, the half-space PNG horizon.
Stationarity at one parameter
We prove Theorem 1.5 by checking generator balance for a marked reflected random-walk representation of ν c , g \nu_{c,g} ν c , g .
Step coordinates and a local stationarity criterion
For bounded measurable F F F on U C + , 0 \mathrm{UC}_{+,0} UC + , 0 , set
( S t ( c ) F ) ( h ) : = E h [ F ( Rec ( h ( c ) ( t , ⋅ ; h ) ) ) ] . (S_t^{(c)}F)(h):=\mathbb{E}_h\left[F\left(\operatorname{Rec}\left(h^{(c)}(t,\cdot;h)\right)\right)\right]. ( S t ( c ) F ) ( h ) := E h [ F ( Rec ( h ( c ) ( t , ⋅ ; h ) ) ) ] .
The graphical cocycle, independent time slabs, and vertical-shift covariance give S 0 ( c ) = I S_0^{(c)}=I S 0 ( c ) = I and S s + t ( c ) = S s ( c ) S t ( c ) S_{s+t}^{(c)}=S_s^{(c)}S_t^{(c)} S s + t ( c ) = S s ( c ) S t ( c ) .
For the index set of a finite or infinite step configuration, write
I 0 = ∅ , I N = { 1 , … , N } ( 1 ≤ N < ∞ ) , I ∞ = N . I_0=\varnothing,\qquad I_N=\{1,\ldots,N\}\quad(1\leq N<\infty),\qquad I_\infty=\mathbb{N}. I 0 = ∅ , I N = { 1 , … , N } ( 1 ≤ N < ∞ ) , I ∞ = N .
Definition 2.1 (Step-coordinate state space). A profile belongs to X s t e p ⊂ U C + \mathcal{X}_{\mathrm{step}}\subset\mathrm{UC}_+ X step ⊂ UC + if it has a representation h = ( m ; ξ ) h=(m;\xi) h = ( m ; ξ ) , where m ∈ Z m\in\mathbb{Z} m ∈ Z and ξ = ( ( y i , σ i ) ) i ∈ I N \xi=((y_i,\sigma_i))_{i\in I_N} ξ = (( y i , σ i ) ) i ∈ I N , satisfying
y i ≥ 0 , σ i ∈ { + 1 , − 1 } , y i ≤ y i + 1 , # { i : y i ≤ R } < ∞ ( R < ∞ ) , y_i\geq0,\quad\sigma_i\in\{+1,-1\},\quad y_i\leq y_{i+1},\quad\#\{i:y_i\leq R\}<\infty\quad(R<\infty), y i ≥ 0 , σ i ∈ { + 1 , − 1 } , y i ≤ y i + 1 , # { i : y i ≤ R } < ∞ ( R < ∞ ) ,
y i = y i + 1 > 0 ⟹ ( σ i , σ i + 1 ) = ( + 1 , − 1 ) , y i = 0 ⟹ i = 1 and σ i = − 1 , y_i=y_{i+1}>0\quad\Longrightarrow\quad(\sigma_i,\sigma_{i+1})=(+1,-1),\qquad y_i=0\quad\Longrightarrow\quad i=1\text{ and }\sigma_i=-1, y i = y i + 1 > 0 ⟹ ( σ i , σ i + 1 ) = ( + 1 , − 1 ) , y i = 0 ⟹ i = 1 and σ i = − 1 ,
and
h ( x ) = m + ∑ y i < x σ i + # { i : y i = x , σ i = + 1 } , x ≥ 0. h(x)=m+\sum_{y_i<x}\sigma_i+\#\{i:y_i=x,\ \sigma_i=+1\},\qquad x\geq0. h ( x ) = m + y i < x ∑ σ i + # { i : y i = x , σ i = + 1 } , x ≥ 0.
Thus m = h ( 0 ) m=h(0) m = h ( 0 ) , and reconstruction takes the upper value at each step, including a coincident ( + 1 , − 1 ) (+1,-1) ( + 1 , − 1 ) -spike. The representation is unique. Put X s t e p , 0 : = X s t e p ∩ U C + , 0 \mathcal{X}_{\mathrm{step},0}:=\mathcal{X}_{\mathrm{step}}\cap\mathrm{UC}_{+,0} X step , 0 := X step ∩ UC + , 0 . Reconstruction is a Borel injection from the standard Borel space of admissible marked configurations; its image and inverse are therefore Borel. Thus X s t e p , 0 \mathcal{X}_{\mathrm{step},0} X step , 0 is standard Borel.
Definition 2.2 (Regular configuration). A configuration in X s t e p \mathcal{X}_{\mathrm{step}} X step is regular if its step locations are distinct and strictly positive.
Write ξ < z \xi_{<z} ξ < z and ξ > z \xi_{>z} ξ > z for the subconfigurations with locations below and above z z z ; the symbols ξ − \xi^- ξ − and ξ + \xi^+ ξ + denote the ordered configurations preceding and following a displayed block of marks. Between events, y ˙ i = − σ i \dot y_i=-\sigma_i y ˙ i = − σ i . At locations without an existing step, nucleations act by
C z ( m ; ξ ) = ( m ; ξ < z , ( z , + 1 ) , ( z , − 1 ) , ξ > z ) , C 0 ( m ; ξ ) = ( m + 1 ; ( 0 , − 1 ) , ξ ) . C_z(m;\xi)=(m;\xi_{<z},(z,+1),(z,-1),\xi_{>z}),\qquad C_0(m;\xi)=(m+1;(0,-1),\xi). C z ( m ; ξ ) = ( m ; ξ < z , ( z , + 1 ) , ( z , − 1 ) , ξ > z ) , C 0 ( m ; ξ ) = ( m + 1 ; ( 0 , − 1 ) , ξ ) .
An adjacent ( − 1 , + 1 ) (-1,+1) ( − 1 , + 1 ) -pair is deleted on contact, and an incoming up-step is absorbed at the wall:
( m ; ξ − , ( q , − 1 ) , ( q , + 1 ) , ξ + ) ⟼ ( m ; ξ − , ξ + ) , ( m ; ( y , + 1 ) , ξ ) ⟶ y ↓ 0 ( m + 1 ; ξ ) . (m;\xi^-,(q,-1),(q,+1),\xi^+)\longmapsto(m;\xi^-,\xi^+),\qquad(m;(y,+1),\xi)\mathrel{\mathop{\longrightarrow}^{y\downarrow0}}(m+1;\xi). ( m ; ξ − , ( q , − 1 ) , ( q , + 1 ) , ξ + ) ⟼ ( m ; ξ − , ξ + ) , ( m ; ( y , + 1 ) , ξ ) ⟶ y ↓ 0 ( m + 1 ; ξ ) .
The incoming collision and absorption faces are identified with the configurations after deletion. For each fixed initial state, these coordinate rules almost surely reconstruct graphical PNG and preserve X s t e p \mathcal{X}_{\mathrm{step}} X step , and, after recentering, X s t e p , 0 \mathcal{X}_{\mathrm{step},0} X step , 0 . Indeed, the level sets A a ( h ) : = { x ≥ 0 : h ( x ) ≥ a } A_a(h):=\{x\geq0:h(x)\geq a\} A a ( h ) := { x ≥ 0 : h ( x ) ≥ a } satisfy
A a ( D s h ) = { x ≥ 0 : dist R + ( x , A a ( h ) ) ≤ s } , a ∈ Z . A_a(D_sh)=\{x\geq0:\operatorname{dist}_{\mathbb{R}_+}(x,A_a(h))\leq s\},\qquad a\in\mathbb{Z}. A a ( D s h ) = { x ≥ 0 : dist R + ( x , A a ( h )) ≤ s } , a ∈ Z .
Their endpoints have the stated velocities and deletion rules; isolated components produce outgoing ( + 1 , − 1 ) (+1,-1) ( + 1 , − 1 ) pairs. Poisson atoms almost surely avoid existing steps, so the creation rules agree with one-point raisings. Induction over the finitely many events in each compact dependence cone, followed by countable exhaustion, proves agreement.
Lift observables on X s t e p , 0 \mathcal{X}_{\mathrm{step},0} X step , 0 to vertically shift-invariant functions on X s t e p \mathcal{X}_{\mathrm{step}} X step . Let D p o l \mathcal{D}_{\mathrm{pol}} D pol consist of bounded observables F F F admitting finite radii 0 < R 0 < R 0<R_0<R 0 < R 0 < R such that F ( h ) F(h) F ( h ) depends only on h ∣ [ 0 , R 0 ] h|_{[0,R_0]} h ∣ [ 0 , R 0 ] . For y = ( y 1 , … , y n ) y=(y_1,\ldots,y_n) y = ( y 1 , … , y n ) and σ = ( σ 1 , … , σ n ) \boldsymbol{\sigma}=(\sigma_1,\ldots,\sigma_n) σ = ( σ 1 , … , σ n ) , let F n ( m ; y , σ ) F_n(m;y,\boldsymbol{\sigma}) F n ( m ; y , σ ) be its coordinate form when there are n n n marks in [ 0 , R ) [0,R) [ 0 , R ) ; locality makes the continuation beyond R R R irrelevant. On each ordered simplex with fixed signs, these forms are locally Lipschitz and piecewise C 1 C^1 C 1 , independent of m m m , and satisfy the boundary matching conditions
lim ε ↓ 0 F n ( m ; ξ − , ( q , − 1 ) , ( q + ε , + 1 ) , ξ + ) = F n − 2 ( m ; ξ − , ξ + ) , q > 0 , \lim_{\varepsilon\downarrow0}F_n(m;\xi^-,(q,-1),(q+\varepsilon,+1),\xi^+)=F_{n-2}(m;\xi^-,\xi^+),\qquad q>0, ε ↓ 0 lim F n ( m ; ξ − , ( q , − 1 ) , ( q + ε , + 1 ) , ξ + ) = F n − 2 ( m ; ξ − , ξ + ) , q > 0 ,
lim y ↓ 0 F n ( m ; ( y , + 1 ) , ξ ) = F n − 1 ( m + 1 ; ξ ) . \lim_{y\downarrow0}F_n(m;(y,+1),\xi)=F_{n-1}(m+1;\xi). y ↓ 0 lim F n ( m ; ( y , + 1 ) , ξ ) = F n − 1 ( m + 1 ; ξ ) .
At bulk and boundary creation faces, F ( C z h ) F(C_zh) F ( C z h ) and F ( C 0 h ) F(C_0h) F ( C 0 h ) are the traces from the outgoing configurations. Require, almost everywhere,
∑ i : y i < R ∣ ∂ y i F ( h ) ∣ ≤ C ( 1 + N R ( h ) ) p , N R ( h ) : = # { i : y i ≤ R } , (11) \sum_{i:y_i<R}|\partial_{y_i}F(h)|\leq C(1+N_R(h))^p,\qquad N_R(h):=\#\{i:y_i\leq R\}, \tag*{(11)} i : y i < R ∑ ∣ ∂ y i F ( h ) ∣ ≤ C ( 1 + N R ( h ) ) p , N R ( h ) := # { i : y i ≤ R } , (11)
for some C < ∞ C<\infty C < ∞ and integer p ≥ 0 p\geq0 p ≥ 0 . On cells of regular configurations, define almost everywhere, T F ( h ) : = − ∑ i σ i ∂ y i F n \mathcal{T}F(h):=-\sum_i\sigma_i\partial_{y_i}F_n T F ( h ) := − ∑ i σ i ∂ y i F n , and the local generator
( G c F ) ( h ) = T F ( h ) + 2 ∫ 0 ∞ ( F ( C z h ) − F ( h ) ) d z + c ( F ( C 0 h ) − F ( h ) ) . (\mathcal{G}_cF)(h)=\mathcal{T}F(h)+2\int_0^\infty\bigl(F(C_zh)-F(h)\bigr)\,\mathrm{d}z+c\bigl(F(C_0h)-F(h)\bigr). ( G c F ) ( h ) = T F ( h ) + 2 ∫ 0 ∞ ( F ( C z h ) − F ( h ) ) d z + c ( F ( C 0 h ) − F ( h ) ) .
Set the bulk integrand to zero at existing step locations, a Lebesgue-null set. The integral vanishes beyond R 0 R_0 R 0 ; the matching conditions account for deterministic deletions.
Lemma 2.3 (A local generator criterion). Fix c ≥ 0 c \ge0 c ≥ 0 . Let μ \mu μ be a probability measure supported on regular configurations in X s t e p , 0 \mathcal{X}_{\mathrm{step},0} X step , 0 . Assume that its finite-window coordinate distributions are absolutely continuous on each fixed-count, fixed-sign ordered simplex, and that
∫ N R ( h ) p μ ( d h ) < ∞ , 0 < R < ∞ , p ∈ N . \int N_R(h)^p\,\mu(dh)<\infty,\qquad0<R<\infty,\qquad p\in\mathbb{N}. ∫ N R ( h ) p μ ( d h ) < ∞ , 0 < R < ∞ , p ∈ N . If
∫ G c F d μ = 0 , F ∈ D p o l , \int\mathcal{G}_cF\,d\mu=0,\qquad F\in\mathcal{D}_{\mathrm{pol}}, ∫ G c F d μ = 0 , F ∈ D pol , then μ S t ( c ) = μ \mu S_t^{(c)}=\mu μ S t ( c ) = μ for every t ≥ 0 t\ge0 t ≥ 0 .
Proof. Step 1: A determining family. For h = ( 0 ; ξ ) h=(0;\xi) h = ( 0 ; ξ ) , write η h = ∑ i σ i δ y i \eta_h=\sum_i \sigma_i\delta_{y_i} η h = ∑ i σ i δ y i and consider
F ( h ) = Φ ( ⟨ η h , φ 1 ⟩ , … , ⟨ η h , φ j ⟩ ) , φ a ∈ C c ∞ ( ( 0 , ∞ ) ) , Φ ∈ C b ∞ ( R j ) . F(h)=\Phi\bigl(\langle\eta_h,\varphi_1\rangle,\ldots,\langle\eta_h,\varphi_j\rangle\bigr),\qquad\varphi_a\in C_c^\infty((0,\infty)),\qquad\Phi\in C_b^\infty(\mathbb{R}^j). F ( h ) = Φ ( ⟨ η h , φ 1 ⟩ , … , ⟨ η h , φ j ⟩ ) , φ a ∈ C c ∞ (( 0 , ∞ )) , Φ ∈ C b ∞ ( R j ) .
These functions determine laws on regular pinned profiles: the smooth spatial tests determine η h \eta_h η h , and pinning fixes the additive height constant. They belong to D p o l \mathcal{D}_{\mathrm{pol}} D pol , since opposite marks cancel at collisions and the spatial tests vanish near the wall. Their total coordinate gradient is bounded by a F N R a_FN_R a F N R , where
a F : = ∑ a = 1 j ∥ ∂ a Φ ∥ ∞ ∥ φ a ′ ∥ ∞ a_F:=\sum_{a=1}^{j}\|\partial_a\Phi\|_\infty\|\varphi'_a\|_\infty a F := a = 1 ∑ j ∥ ∂ a Φ ∥ ∞ ∥ φ a ′ ∥ ∞
and R R R contains all spatial supports.
Step 2: Truncated backward observables. Fix T < ∞ T<\infty T < ∞ and R s r c > R + T + 1 R_{\mathrm{src}}>R+T+1 R src > R + T + 1 . Let N t s r c N_t^{\mathrm{src}} N t src count the bulk and wall sources in ( 0 , t ] × ( 0 , R s r c ) (0,t]\times(0,R_{\mathrm{src}}) ( 0 , t ] × ( 0 , R src ) and ( 0 , t ] (0,t] ( 0 , t ] , respectively, and put Λ = 2 R s r c + c \Lambda=2R_{\mathrm{src}}+c Λ = 2 R src + c . With h t = h ( c ) ( t , ⋅ ; h ) h_t=h^{(c)}(t,\cdot;h) h t = h ( c ) ( t , ⋅ ; h ) , define
U k ( t , h ) : = E h [ F ( Rec h t ) 1 { N t s r c = k } ] , V M ( t , h ) : = ∑ k = 0 M U k ( t , h ) . U_k(t,h):=\mathbb{E}_h\left[F(\operatorname{Rec}h_t)\mathbf{1}_{\{N_t^{\mathrm{src}}=k\}}\right],\qquad V_M(t,h):=\sum_{k=0}^{M}U_k(t,h). U k ( t , h ) := E h [ F ( Rec h t ) 1 { N t src = k } ] , V M ( t , h ) := k = 0 ∑ M U k ( t , h ) .
We verify directly that V M ( t , ⋅ ) ∈ D p o l V_M(t,\cdot)\in\mathcal{D}_{\mathrm{pol}} V M ( t , ⋅ ) ∈ D pol for $0\le t\le T.
For fixed source types and k ≤ M k\le M k ≤ M , the ordered source coordinates range over a bounded box of dimension at most 2 M 2M 2 M . With n = N R s r c ( h ) n=N_{R_{\mathrm{src}}}(h) n = N R src ( h ) initial marks, all candidate steps follow affine rays. Spatial insertion order and comparisons of births, meetings, wall arrivals, and the observation time partition this box into history cells using at most C M ( n + 1 ) 4 C_M(n+1)^4 C M ( n + 1 ) 4 affine hyperplanes. Inside each cell the terminal observable is smooth, with total initial-coordinate gradient bounded by C F ( n + 2 M ) C_F(n+2M) C F ( n + 2 M ) . Interfaces involving source coordinates have nonzero source normals from a finite set depending only on M M M and uniformly bounded displacement speeds in the initial coordinates. Under an initial-coordinate displacement, their swept volumes are bounded by a polynomial in n n n times that displacement, and the integrand jumps by at most 2 ∥ F ∥ ∞ 2\|F\|_\infty 2∥ F ∥ ∞ . Across interfaces independent of the source coordinates, the observable is continuous by collision cancellation and vanishing of the spatial tests near the wall. Integrating over the source coordinates gives local Lipschitz continuity and
∑ i : y i < R s r c ∣ ∂ y i V M ( t , h ) ∣ ≤ C F , R s r c , T , M ( 1 + N R s r c ( h ) ) p M . \sum_{i:y_i<R_{\mathrm{src}}}|\partial_{y_i}V_M(t,h)|\le C_{F,R_{\mathrm{src}},T,M}(1+N_{R_{\mathrm{src}}}(h))^{p_M}. i : y i < R src ∑ ∣ ∂ y i V M ( t , h ) ∣ ≤ C F , R src , T , M ( 1 + N R src ( h ) ) p M .
On regions of fixed combinatorial type, the history polytopes have affine vertices; triangulation and differentiation under the integral give piecewise C 1 C^1 C 1 regularity.
The boundary matching conditions follow by bounded convergence. Indeed, for almost every fixed source list, a collapsing incoming pair, or an up-step approaching zero, disappears before the first source and the observation time. The subsequent evolutions then agree with those started after deletion. Outgoing creation traces similarly agree with the graphical creation states. At t = 0 t=0 t = 0 the matching holds for F F F . Finally, finite propagation gives locality in [ 0 , R + T ] [0,R+T] [ 0 , R + T ] , proving the claimed domain membership.
Step 3: The truncated backward equation. Write
( N f ) ( h ) : = 2 ∫ 0 R s r c f ( C z h ) d z + c f ( C 0 h ) . (\mathcal{N}f)(h):=2\int_0^{R_{\mathrm{src}}}f(C_z h)\,dz+cf(C_0h). ( N f ) ( h ) := 2 ∫ 0 R src f ( C z h ) d z + c f ( C 0 h ) .
Differentiating the ordered source integrals gives
∂ t U k = ( T − Λ ) U k + Λ U k − 1 , U − 1 = 0. \partial_tU_k=(\mathcal{T}-\Lambda)U_k+\Lambda U_{k-1},\qquad U_{-1}=0. ∂ t U k = ( T − Λ ) U k + Λ U k − 1 , U − 1 = 0.
By locality, G c = T − Λ + N \mathcal{G}_c=\mathcal{T}-\Lambda+\mathcal{N} G c = T − Λ + N on these observables. Summing over k < M k<M k < M yields, coordinate-almost everywhere,
∂ t V M = G c V M − N U M , V M ( 0 ) = F . (12) \partial_tV_M=\mathcal{G}_cV_M-\mathcal{N}U_M,\qquad V_M(0)=F. \tag*{(12)} ∂ t V M = G c V M − N U M , V M ( 0 ) = F . (12)
The absolute-continuity assumption makes this identity valid μ \mu μ -almost everywhere. Moreover, F F F is continuous through graphical events, and between events its speed is at most a F ( N R s r c ( h ) + 2 N T s r c ) a_F(N_{R_{\mathrm{src}}}(h)+2N_T^{\mathrm{src}}) a F ( N R src ( h ) + 2 N T src ) . The truncation indicator changes only at a source, so
∣ V M ( t , h ) − V M ( s , h ) ∣ ≤ [ a F ( N R s r c ( h ) + 2 Λ T ) + Λ ∥ F ∥ ∞ ] ∣ t − s ∣ . |V_M(t,h)-V_M(s,h)|\le\left[a_F\bigl(N_{R_{\mathrm{src}}}(h)+2\Lambda T\bigr)+\Lambda\|F\|_\infty\right]|t-s|. ∣ V M ( t , h ) − V M ( s , h ) ∣ ≤ [ a F ( N R src ( h ) + 2Λ T ) + Λ∥ F ∥ ∞ ] ∣ t − s ∣.
This integrable bound and the polynomial gradient bound justify time integration and Fubini against μ \mu μ .
Step 4: Generator balance and removal of the truncation. Since V M ( t , ⋅ ) ∈ D p o l V_M(t,\cdot)\in\mathcal{D}_{\mathrm{pol}} V M ( t , ⋅ ) ∈ D pol , the balance hypothesis and (12) give
∫ V M ( t , ⋅ ) d μ − ∫ F d μ = − ∫ 0 t ∫ L U M ( s , ⋅ ) d μ d s . \int V_M(t,\cdot)\,\mathrm{d}\mu-\int F\,\mathrm{d}\mu=-\int_0^t\int\mathcal{L}U_M(s,\cdot)\,\mathrm{d}\mu\,\mathrm{d}s. ∫ V M ( t , ⋅ ) d μ − ∫ F d μ = − ∫ 0 t ∫ L U M ( s , ⋅ ) d μ d s .
The bound ∣ U M ( s , h ) ∣ ≤ ∥ F ∥ ∞ e − Λ s ( Λ s ) M / M ! |U_M(s,h)|\leq\|F\|_\infty e^{-\Lambda s}(\Lambda s)^M/M! ∣ U M ( s , h ) ∣ ≤ ∥ F ∥ ∞ e − Λ s ( Λ s ) M / M ! implies
∣ ∫ V M ( t , ⋅ ) d μ − ∫ F d μ ∣ ≤ ∥ F ∥ ∞ P ( Pois ( Λ t ) > M ) . \left|\int V_M(t,\cdot)\,\mathrm{d}\mu-\int F\,\mathrm{d}\mu\right|\leq\|F\|_\infty\mathbb{P}(\operatorname{Pois}(\Lambda t)>M). ∫ V M ( t , ⋅ ) d μ − ∫ F d μ ≤ ∥ F ∥ ∞ P ( Pois ( Λ t ) > M ) .
Also, uniformly in h h h ,
∣ S t ( c ) F ( h ) − V M ( t , h ) ∣ ≤ ∥ F ∥ ∞ P ( Pois ( Λ t ) > M ) . \left|S_t^{(c)}F(h)-V_M(t,h)\right|\leq\|F\|_\infty\mathbb{P}(\operatorname{Pois}(\Lambda t)>M). S t ( c ) F ( h ) − V M ( t , h ) ≤ ∥ F ∥ ∞ P ( Pois ( Λ t ) > M ) .
Letting M → ∞ M\to\infty M → ∞ proves ∫ S t ( c ) F d μ = ∫ F d μ \int S_t^{(c)}F\,\mathrm{d}\mu=\int F\,\mathrm{d}\mu ∫ S t ( c ) F d μ = ∫ F d μ . At each fixed deterministic time the evolved configuration is regular almost surely: incoming collision and wall steps are deleted immediately, and a Poisson source occurs at that time with probability zero. The determining family therefore identifies the evolved law with μ \mu μ . Since T T T was arbitrary, the conclusion holds for every t ≥ 0 t\geq0 t ≥ 0 .
A reflected random-walk representation
We encode ν c , g \nu_{c,g} ν c , g by a reflected walk whose clock marks record the visible steps of the height profile. This gives the finite-window densities used in the generator calculation below.
The reflected random-walk. Fix g ∈ G c g\in\mathcal{G}_c g ∈ G c with c < g c<g c < g , and abbreviate S j = S j ( g ) S_j=S_j^{(g)} S j = S j ( g ) , Y = Y ( c , g ) Y=Y^{(c,g)} Y = Y ( c , g ) , H ~ = H ~ ( c , g ) \widetilde{H}=\widetilde{H}^{(c,g)} H = H ( c , g ) , and H = H ( c , g ) H=H^{(c,g)} H = H ( c , g ) . Set
W ( x ) : = S 1 ( x ) − S 2 ( x ) , M ( x ) : = sup 0 ≤ y ≤ x W ( y ) , Q ( x ) : = max { Y , M ( x ) } − W ( x ) . W(x):=S_1(x)-S_2(x),\qquad M(x):=\sup_{0\leq y\leq x}W(y),\qquad Q(x):=\max\{Y,M(x)\}-W(x). W ( x ) := S 1 ( x ) − S 2 ( x ) , M ( x ) := 0 ≤ y ≤ x sup W ( y ) , Q ( x ) := max { Y , M ( x )} − W ( x ) .
Thus Q Q Q is the reflection of Y − W Y-W Y − W at zero, and
Q ( 0 ) = Y , H ~ ( x ) = S 1 ( x ) + Q ( x ) − Y , π q : = P ( Q ( 0 ) = q ) = ( 1 − c / g ) ( c / g ) q . Q(0)=Y,\qquad\widetilde{H}(x)=S_1(x)+Q(x)-Y,\qquad\pi_q:=\mathbb{P}(Q(0)=q)=(1-c/g)(c/g)^q. Q ( 0 ) = Y , H ( x ) = S 1 ( x ) + Q ( x ) − Y , π q := P ( Q ( 0 ) = q ) = ( 1 − c / g ) ( c / g ) q .
At c = 0 c=0 c = 0 , the last expression means π = δ 0 \pi=\delta_0 π = δ 0 . We regard π = ( π q ) q ≥ 0 \pi=(\pi_q)_{q\geq0} π = ( π q ) q ≥ 0 as a row vector. It specifies the initial law of this auxiliary spatial process. Upper-semicontinuous regularization preserves the locations and signs of the jumps, so the same step marks describe H H H .
Call a clock event visible if it changes H ~ \widetilde{H} H . For the up- and down-jump clocks N j , + , N j , − N_{j,+},N_{j,-} N j , + , N j , − of S j S_j S j , the identity above gives
In particular, the down-step intensity at the spatial origin is g P ( Q ( 0 ) > 0 ) = c g\mathbb{P}(Q(0)>0)=c g P ( Q ( 0 ) > 0 ) = c . The clock marks must be retained: at q = 0 q=0 q = 0 , both N 1 , + N_{1,+} N 1 , + and N 2 , − N_{2,-} N 2 , − leave Q Q Q unchanged, but only N 1 , + N_{1,+} N 1 , + is visible.
Visible-step matrices. Let I I I be the identity matrix on N 0 \mathbb{N}_0 N 0 , and let 1 \mathbf{1} 1 be the column vector of ones. The rate matrices for visible up-steps and down-steps are, respectively,
( Q + ) q p = g 1 { q = p = 0 } + g − 1 1 { p = q + 1 } , ( Q − ) q p = g 1 { q ≥ 1 , p = q − 1 } . (Q_+)_{qp}=g\mathbf{1}_{\{q=p=0\}}+g^{-1}\mathbf{1}_{\{p=q+1\}},\qquad(Q_-)_{qp}=g\mathbf{1}_{\{q\geq1,\ p=q-1\}}. ( Q + ) q p = g 1 { q = p = 0 } + g − 1 1 { p = q + 1 } , ( Q − ) q p = g 1 { q ≥ 1 , p = q − 1 } .
Put r = g + g − 1 r=g+g^{-1} r = g + g − 1 and A = Q + + Q − A=Q_++Q_- A = Q + + Q − . The table shows that the invisible events have the same total hidden-state rate matrix A A A . Since the four clocks have total rate 2 r 2r 2 r , evolution with no visible step is described by
Q 0 : = A − 2 r I , E ( u ) : = e u Q 0 . Q_0:=A-2rI,\qquad E(u):=e^{uQ_0}. Q 0 := A − 2 r I , E ( u ) := e u Q 0 .
The entry E ( u ) q p E(u)_{qp} E ( u ) q p is the probability of reaching state p p p from q q q over spatial length u u u without a visible step. Allowing visible steps as well gives the reflected-walk generator
Q = Q 0 + Q + + Q − = 2 ( A − r I ) . Q=Q_0+Q_++Q_-=2(A-rI). Q = Q 0 + Q + + Q − = 2 ( A − r I ) .
These matrices act boundedly on ℓ ∞ ( N 0 ) \ell^\infty(\mathbb{N}_0) ℓ ∞ ( N 0 ) , paired with row vectors in ℓ 1 ( N 0 ) \ell^1(\mathbb{N}_0) ℓ 1 ( N 0 ) . Since A 1 = r 1 A\mathbf{1}=r\mathbf{1} A 1 = r 1 ,
E ( u ) = e − 2 r u ∑ m ≥ 0 u m A m m ! , E ( u ) 1 = e − r u 1. E(u)=e^{-2ru}\sum_{m\geq0}\frac{u^mA^m}{m!},\qquad E(u)\mathbf{1}=e^{-ru}\mathbf{1}. E ( u ) = e − 2 r u m ≥ 0 ∑ m ! u m A m , E ( u ) 1 = e − r u 1 .
Thus E ( u ) E(u) E ( u ) is positive with operator norm e − r u e^{-ru} e − r u , and Q Q Q is a bounded conservative Markov generator. The following identities will be used in the generator balance:
Q + + Q − = I , π Q − = c π , C : = Q − Q + − I . Q_++Q_-=I,\qquad\pi Q_-=c\pi,\qquad\mathcal{C}:=Q_-Q_+-I. Q + + Q − = I , π Q − = c π , C := Q − Q + − I .
Finite-window jump densities. For L > 0 L>0 L > 0 , let Δ n ( L ) = { 0 < y 1 < ⋯ < y n < L } \Delta_n(L)=\{0<y_1<\cdots<y_n<L\} Δ n ( L ) = { 0 < y 1 < ⋯ < y n < L } , with the empty configuration when n = 0 n=0 n = 0 , and write Q ± 1 = Q ± Q_{\pm1}=Q_\pm Q ± 1 = Q ± . For prescribed ordered jump locations y y y and signs σ \sigma σ , the density is
J L ( y , σ ) : = π E ( y 1 ) Q σ 1 E ( y 2 − y 1 ) ⋯ Q σ n E ( L − y n ) 1. \mathcal{J}_L(y,\sigma):=\pi E(y_1)Q_{\sigma_1}E(y_2-y_1)\cdots Q_{\sigma_n}E(L-y_n)\mathbf{1}. J L ( y , σ ) := π E ( y 1 ) Q σ 1 E ( y 2 − y 1 ) ⋯ Q σ n E ( L − y n ) 1 .
Each E E E factor accounts for a gap with no visible step; each Q σ i Q_{\sigma_i} Q σ i inserts a step of the specified sign; the final multiplication by 1 \mathbf{1} 1 sums over the terminal hidden state. For n = 0 n=0 n = 0 , the product is π E ( L ) 1 \pi E(L)\mathbf{1} π E ( L ) 1 . This is the ordered marked Janossy density, meaning the joint density of exactly these steps and no others in ( 0 , L ) (0,L) ( 0 , L ) .
Summing over signs replaces each Q σ i Q_{\sigma_i} Q σ i by A A A . The identities above give
∑ σ ∈ { ± 1 } n J L ( y , σ ) = r n e − r L . \sum_{\sigma\in\{\pm1\}^n}\mathcal{J}_L(y,\sigma)=r^n e^{-rL}. σ ∈ { ± 1 } n ∑ J L ( y , σ ) = r n e − r L .
Since Δ n ( L ) \Delta_n(L) Δ n ( L ) has volume L n / n ! L^n/n! L n / n ! , these densities have total mass one. They also show that the unmarked step locations form a homogeneous Poisson process of rate r r r ; in particular, N L ∼ Pois ( r L ) N_L \sim\operatorname{Pois}(rL) N L ∼ Pois ( r L ) .
There is almost surely no step at a fixed endpoint L L L . Hence, for bounded measurable F F F depending only on h ∣ [ 0 , L ] h|_{[0,L]} h ∣ [ 0 , L ] ,
∫ F ( h ) ν c , g ( d h ) = ∑ n ≥ 0 ∑ σ ∈ { ± 1 } n ∫ Δ n ( L ) J L ( y , σ ) F n ( y , σ ) d y . \int F(h)\,\nu_{c,g}(dh)=\sum_{n\geq0}\sum_{\sigma\in\{\pm1\}^n}\int_{\Delta_n(L)}\mathcal{J}_L(\mathbf{y},\sigma)F_n(\mathbf{y},\sigma)\,d\mathbf{y}. ∫ F ( h ) ν c , g ( d h ) = n ≥ 0 ∑ σ ∈ { ± 1 } n ∑ ∫ Δ n ( L ) J L ( y , σ ) F n ( y , σ ) d y .
Here F n F_n F n is evaluated on the pinned upper-semicontinuous profile with the indicated step marks. Thus ν c , g \nu_{c,g} ν c , g has regular support, absolutely continuous finite-window coordinate laws, and finite moments of all orders for every local step count, as required by Lemma 2.3 .
Remark 2.4 (Relation to full-space PNG stationary measures). For a > 0 a>0 a > 0 , let μ a + \mu^+_a μ a + be the restriction to [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , pinned at zero, of the full-space PNG stationary law with independent Poisson up-steps and down-steps of rates a a a and a − 1 a^{-1} a − 1 [24 Section 4.3, Proposition 4.2] . For every c > 0 c>0 c > 0 ,
μ 1 / c + = { ν c , 1 / c , 0 < c < 1 , ν c , c , c ≥ 1. \mu^+_{1/c}=
\begin{cases}
\nu_{c,1/c}, & 0<c<1,\\
\nu_{c,c}, & c\geq1.
\end{cases} μ 1/ c + = { ν c , 1/ c , ν c , c , 0 < c < 1 , c ≥ 1.
For c ≥ 1 c\geq1 c ≥ 1 , this follows directly from (5) , whose up/down rates are ( c − 1 , c ) (c^{-1},c) ( c − 1 , c ) . For 0 < c < 1 0<c<1 0 < c < 1 , take g = 1 / c g=1/c g = 1/ c . Then π q = ( 1 − g − 2 ) g − 2 q \pi_q=(1-g^{-2})g^{-2q} π q = ( 1 − g − 2 ) g − 2 q , and direct multiplication gives π Q + = g π \pi Q_+=g\pi π Q + = g π , π Q − = g − 1 π \pi Q_-=g^{-1}\pi π Q − = g − 1 π , π E ( u ) = e − r u π \pi E(u)=e^{-ru}\pi π E ( u ) = e − r u π . Consequently, J L ( y , σ ) = e − r L ∏ i = 1 n g σ i \mathcal{J}_L(\mathbf{y},\sigma)=e^{-rL}\prod_{i=1}^n g^{\sigma_i} J L ( y , σ ) = e − r L ∏ i = 1 n g σ i , which is the density of independent Poisson up-steps and down-steps with rates ( g , g − 1 ) = ( c − 1 , c ) (g,g^{-1})=(c^{-1},c) ( g , g − 1 ) = ( c − 1 , c ) . For general g g g , the signs need not be independent, although the unmarked step locations remain Poisson.
Generator balance
Fix F ∈ D p o l F\in\mathcal{D}_{\mathrm{pol}} F ∈ D pol , with locality radii 0 < R 0 < R 0<R_0<R 0 < R 0 < R , and take the integration cutoff R 0 < L < R R_0<L<R R 0 < L < R . Locality gives
2 ∫ 0 ∞ ( F ( C z h ) − F ( h ) ) d z = 2 ∫ 0 L F ( C z h ) d z − 2 L F ( h ) . 2\int_0^\infty\bigl(F(C_zh)-F(h)\bigr)\,dz=2\int_0^L F(C_zh)\,dz-2LF(h). 2 ∫ 0 ∞ ( F ( C z h ) − F ( h ) ) d z = 2 ∫ 0 L F ( C z h ) d z − 2 L F ( h ) .
Abbreviate
∑ ∗ G : = ∑ n ≥ 0 ∑ σ ∈ { ± 1 } n ∫ Δ n ( L ) G n ( y , σ ) d y . \sum\nolimits^*G:=\sum_{n\geq0}\sum_{\sigma\in\{\pm1\}^n}\int_{\Delta_n(L)}G_n(\mathbf{y},\sigma)\,d\mathbf{y}. ∑ ∗ G := n ≥ 0 ∑ σ ∈ { ± 1 } n ∑ ∫ Δ n ( L ) G n ( y , σ ) d y .
Write J L ( i ) [ M ] \mathcal{J}_L^{(i)}[M] J L ( i ) [ M ] for the matrix product with Q σ i Q_{\sigma_i} Q σ i replaced by M M M . For z ∈ ( y k , y k + 1 ) z\in(y_k,y_{k+1}) z ∈ ( y k , y k + 1 ) , with y 0 = 0 , y n + 1 = L y_0=0,y_{n+1}=L y 0 = 0 , y n + 1 = L , write J L , z [ M ] \mathcal{J}_{L,z}[M] J L , z [ M ] for the matrix product with its gap factor replaced by E ( z − y k ) M E ( y k + 1 − z ) E(z-y_k)ME(y_{k+1}-z) E ( z − y k ) M E ( y k + 1 − z ) , using one-sided endpoint conventions. The mean generator comprises
T F = − ∑ ∗ J L ∑ i = 1 n σ i ∂ y i F n , T b u l k + = 2 ∑ ∗ J L ∫ 0 L F ( C z h ) d z , \mathcal{T}_F=-\sum\nolimits^*\mathcal{J}_L\sum_{i=1}^n\sigma_i\partial_{y_i}F_n,\qquad
\mathcal{T}_{\mathrm{bulk}}^+=2\sum\nolimits^*\mathcal{J}_L\int_0^L F(C_zh)\,dz, T F = − ∑ ∗ J L i = 1 ∑ n σ i ∂ y i F n , T bulk + = 2 ∑ ∗ J L ∫ 0 L F ( C z h ) d z ,
T b u l k − = − 2 L ∑ ∗ J L F n , C + = c ∑ ∗ J L F ( C 0 h ) , C − = − c ∑ ∗ J L F n . \mathcal{T}_{\mathrm{bulk}}^-=-2L\sum\nolimits^*\mathcal{J}_LF_n,\qquad
\mathcal{C}_+=c\sum\nolimits^*\mathcal{J}_LF(C_0h),\qquad
\mathcal{C}_-=-c\sum\nolimits^*\mathcal{J}_LF_n. T bulk − = − 2 L ∑ ∗ J L F n , C + = c ∑ ∗ J L F ( C 0 h ) , C − = − c ∑ ∗ J L F n .
Set Z = Q + − Q − \mathcal{Z}=Q_+-Q_- Z = Q + − Q − , and write [ U , V ] : = U V − V U [U,V]:=UV-VU [ U , V ] := U V − V U for the commutator. Direct multiplication also gives
[ Q 0 , Z ] = 2 C , [ Z , Q σ ] = − C , σ ∂ y ( E ( y − u ) Q σ E ( v − y ) ) = E ( y − u ) C E ( v − y ) . [Q_0,\mathcal{Z}]=2\mathcal{C},\qquad[\mathcal{Z},Q_\sigma]=-\mathcal{C},\qquad
\sigma\partial_y\bigl(E(y-u)Q_\sigma E(v-y)\bigr)=E(y-u)\mathcal{C}E(v-y). [ Q 0 , Z ] = 2 C , [ Z , Q σ ] = − C , σ ∂ y ( E ( y − u ) Q σ E ( v − y ) ) = E ( y − u ) C E ( v − y ) .
Integration by parts yields
− ∫ Δ n ( L ) J L ∑ i σ i ∂ y i F n = ∫ Δ n ( L ) J L ( i ) [ C ] F n + ∫ ∂ i − σ i J L F n − ∫ ∂ i + σ i J L F n . (13) -\int_{\Delta_n(L)}\mathcal{J}_L\sum_i\sigma_i\partial_{y_i}F_n
=\int_{\Delta_n(L)}\mathcal{J}_L^{(i)}[\mathcal{C}]F_n+\int_{\partial_i^-}\sigma_i\mathcal{J}_LF_n-\int_{\partial_i^+}\sigma_i\mathcal{J}_LF_n. \tag*{(13)} − ∫ Δ n ( L ) J L i ∑ σ i ∂ y i F n = ∫ Δ n ( L ) J L ( i ) [ C ] F n + ∫ ∂ i − σ i J L F n − ∫ ∂ i + σ i J L F n . (13)
where ∂ i − \partial_i^- ∂ i − is the face y i = y i − 1 y_i=y_{i-1} y i = y i − 1 and ∂ i + \partial_i^+ ∂ i + is the face y i = y i + 1 y_i=y_{i+1} y i = y i + 1 , with y 0 = 0 y_0=0 y 0 = 0 and y n + 1 = L y_{n+1}=L y n + 1 = L . Face integrals use the remaining n − 1 n-1 n − 1 ordered coordinates. Internal-face coefficients σ i + 1 − σ i \sigma_{i+1}-\sigma_i σ i + 1 − σ i are zero for equal signs, − 2 -2 − 2 for ( + 1 , − 1 ) (+1,-1) ( + 1 , − 1 ) , and + 2 +2 + 2 for ( − 1 , + 1 ) (-1,+1) ( − 1 , + 1 ) .
The creation face has observable F ( C z h ) F(C_zh) F ( C z h ) and factor Q + Q − = I Q_+Q_-=I Q + Q − = I , so cancels T b u l k + \mathcal{T}_{\mathrm{bulk}}^+ T bulk + . At the annihilation face, the matching condition deletes the pair. Its Q − Q + Q_-Q_+ Q − Q + factor and the − 2 I -2I − 2 I insertion representing T b u l k − \mathcal{T}_{\mathrm{bulk}}^- T bulk − leave 2 ∑ ∗ ∫ 0 L J L , z [ C ] F n d z 2\sum\nolimits^*\int_0^L\mathcal{J}_{L,z}[\mathcal{C}]F_n\,dz 2 ∑ ∗ ∫ 0 L J L , z [ C ] F n d z .
Within gaps, ∂ z J L , z [ Z ] = 2 J L , z [ C ] \partial_z\mathcal{J}_{L,z}[\mathcal{Z}]=2\mathcal{J}_{L,z}[\mathcal{C}] ∂ z J L , z [ Z ] = 2 J L , z [ C ] . At y i y_i y i , its left trace minus right trace is J L ( i ) [ [ Z , Q σ i ] ] = − J L ( i ) [ C ] \mathcal{J}_L^{(i)}[[\mathcal{Z},Q_{\sigma_i}]]=-\mathcal{J}_L^{(i)}[\mathcal{C}] J L ( i ) [[ Z , Q σ i ]] = − J L ( i ) [ C ] . Telescoping therefore gives
2 ∫ 0 L J L , z [ C ] d z = J L , L [ Z ] − J L , 0 [ Z ] − ∑ i = 1 n J L ( i ) [ C ] . 2\int_0^L\mathcal{J}_{L,z}[\mathcal{C}]\,dz=\mathcal{J}_{L,L}[\mathcal{Z}]-\mathcal{J}_{L,0}[\mathcal{Z}]-\sum_{i=1}^n\mathcal{J}_L^{(i)}[\mathcal{C}]. 2 ∫ 0 L J L , z [ C ] d z = J L , L [ Z ] − J L , 0 [ Z ] − i = 1 ∑ n J L ( i ) [ C ] .
After multiplication by F n F_n F n and summation, the last term cancels the density derivative in (2.3). Since F F F does not depend on steps near L L L , the boundary contributions there equal − J L , L [ Q + ] F ′ + J L , L [ Q − ] F ′ = − J L , L [ Z ] F ′ -\mathcal{J}_{L,L}[Q_+]F'+\mathcal{J}_{L,L}[Q_-]F'=-\mathcal{J}_{L,L}[\mathcal{Z}]F' − J L , L [ Q + ] F ′ + J L , L [ Q − ] F ′ = − J L , L [ Z ] F ′ , canceling the right endpoint.
At zero, shift invariance and absorption matching give
F ( 0 ; ( 0 , − 1 ) , ξ ) = F ( C 0 ( 0 ; ξ ) ) , lim y ↓ 0 F ( 0 ; ( y , + 1 ) , ξ ) = F ( 0 ; ξ ) . F(0;(0,-1),\xi)=F(C_0(0;\xi)),\qquad
\lim_{y\downarrow0}F(0;(y,+1),\xi)=F(0;\xi). F ( 0 ; ( 0 , − 1 ) , ξ ) = F ( C 0 ( 0 ; ξ )) , y ↓ 0 lim F ( 0 ; ( y , + 1 ) , ξ ) = F ( 0 ; ξ ) .
For the remaining column vector
v r e m : = E ( y 1 ) Q σ 1 E ( y 2 − y 1 ) ⋯ Q σ n E ( L − y n ) 1 , \mathbf{v}_{\mathrm{rem}}:=E(y_1)Q_{\sigma_1}E(y_2-y_1)\cdots Q_{\sigma_n}E(L-y_n)\mathbf{1}, v rem := E ( y 1 ) Q σ 1 E ( y 2 − y 1 ) ⋯ Q σ n E ( L − y n ) 1 ,
with v r e m = E ( L ) 1 v_{\mathrm{rem}} = \mathrm{E}(L)\mathbf{1} v rem = E ( L ) 1 when n = 0 n=0 n = 0 , the boundary contributions at zero combine as
− π Q − v r e m F ( C 0 h ) + π Q + v r e m F ( h ) − π ( Q + − Q − ) v r e m F ( h ) = π Q − v r e m ( F ( h ) − F ( C 0 h ) ) = c π v r e m ( F ( h ) − F ( C 0 h ) ) . -\pi Q_{-}v_{\mathrm{rem}}F(C_{0}h)+\pi Q_{+}v_{\mathrm{rem}}F(h)-\pi(Q_{+}-Q_{-})v_{\mathrm{rem}}F(h)
=\pi Q_{-}v_{\mathrm{rem}}\left(F(h)-F(C_{0}h)\right)=c\pi v_{\mathrm{rem}}\left(F(h)-F(C_{0}h)\right). − π Q − v rem F ( C 0 h ) + π Q + v rem F ( h ) − π ( Q + − Q − ) v rem F ( h ) = π Q − v rem ( F ( h ) − F ( C 0 h ) ) = c π v rem ( F ( h ) − F ( C 0 h ) ) .
Summing over the remaining products cancels C + + C − C_{+}+C_{-} C + + C − .
By locality, deleting the tail beyond L L L bounds the observable derivatives by C ( 1 + N L ) p C(1+N_{L})^{p} C ( 1 + N L ) p , integrable under the Poisson step counts. If e q \mathbf{e}_{q} e q denotes the q q q th coordinate vector, then E = ( g 2 e 1 − e 0 ) e 0 T \mathcal{E}=(g^{2}\mathbf{e}_{1}-\mathbf{e}_{0})\mathbf{e}_{0}^{\mathsf{T}} E = ( g 2 e 1 − e 0 ) e 0 T . Positivity of E E E , its row sums, and A 1 = r 1 A\mathbf{1}=r\mathbf{1} A 1 = r 1 give, for every nonnegative bounded matrix D D D ,
∑ ∗ ∫ 0 L J L , z [ D ] d z ≤ L ∥ D ∥ ∞ → ∞ , \sum^{*}\int_{0}^{L}\mathcal{J}_{L,z}[D]\,\mathrm{d}z\le L\lVert D\rVert_{\infty\to\infty}, ∑ ∗ ∫ 0 L J L , z [ D ] d z ≤ L ∥ D ∥ ∞ → ∞ ,
∑ n ≥ 1 ∑ σ ∫ Δ n ( L ) ∑ i = 1 n ∣ J L ( i ) [ E ] ∣ d y ≤ 2 L ∥ ∣ E ∣ ∥ ∞ → ∞ . \sum_{n\ge1}\sum_{\sigma}\int_{\Delta_{n}(L)}\sum_{i=1}^{n}\left|\mathcal{J}_{L}^{(i)}[\mathcal{E}]\right|\,\mathrm{d}y\le2L\lVert|\mathcal{E}|\rVert_{\infty\to\infty}. n ≥ 1 ∑ σ ∑ ∫ Δ n ( L ) i = 1 ∑ n J L ( i ) [ E ] d y ≤ 2 L ∥ ∣ E ∣ ∥ ∞ → ∞ .
Here ∣ E ∣ |\mathcal{E}| ∣ E ∣ is taken entrywise; the factor two accounts for the two signs of the replaced mark. Boundedness of Q σ Q_{\sigma} Q σ and Q σ Q τ Q_{\sigma}Q_{\tau} Q σ Q τ similarly controls the reindexed face terms. Thus Fubini, integration by parts, and face reindexing are justified absolutely. Continuity cancels boundary terms between the C 1 C^{1} C 1 regions. Hence
∫ ( G c F ) ( h ) ν c , g ( d h ) = 0 , F ∈ D p o l . (14) \int(\mathcal{G}_{c}F)(h)\,\nu_{c,g}(\mathrm{d}h)=0,\qquad F\in\mathcal{D}_{\mathrm{pol}}. \tag*{(14)} ∫ ( G c F ) ( h ) ν c , g ( d h ) = 0 , F ∈ D pol . (14)
For the endpoint g = c ≥ 1 g=c\ge1 g = c ≥ 1 , the right-continuous profile is S 2 ( c ) S_{2}^{(c)} S 2 ( c ) , whose regularization has the same signed steps and scalar Janossy data Q + = c − 1 Q_{+}=c^{-1} Q + = c − 1 , Q − = c Q_{-}=c Q − = c , Q 0 = − ( c + c − 1 ) Q_{0}=-(c+c^{-1}) Q 0 = − ( c + c − 1 ) , and π = 1 = 1 \pi=1=1 π = 1 = 1 . Here Q + Q − = 1 Q_{+}Q_{-}=1 Q + Q − = 1 , π Q − = c π \pi Q_{-}=c\pi π Q − = c π , and E = 0 \mathcal{E}=0 E = 0 ; the same face calculation proves (2.4).
Proof of Theorem 1.5. For every g ∈ G c g\in\mathcal{G}_{c} g ∈ G c , ν c , g \nu_{c,g} ν c , g has regular support (Definition 2.2), absolutely continuous finite-window Janossy densities, and Poisson local step counts. The balance and Lemma 2.3 give ν c , g S t ( c ) = ν c , g \nu_{c,g}S_{t}^{(c)}=\nu_{c,g} ν c , g S t ( c ) = ν c , g .
Jointly stationary measures
We prove Theorem 1.7. For h = ( h 1 , … , h k ) ∈ ( U C + , 0 ) k h=(h_{1},\ldots,h_{k})\in(\mathrm{UC}_{+,0})^{k} h = ( h 1 , … , h k ) ∈ ( UC + , 0 ) k and bounded measurable F F F , write
Θ t ( c ) , [ k ] ( h ; Π ) : = ( Rec h ( c ) ( t , ⋅ ; h i ; Π ) ) i = 1 k , ( P ‾ t ( c ) , [ k ] F ) ( h ) : = E Π F ( Θ t ( c ) , [ k ] ( h ; Π ) ) . \Theta_{t}^{(c),[k]}(h;\Pi):=\left(\operatorname{Rec}_{h}^{(c)}(t,\cdot;h_{i};\Pi)\right)_{i=1}^{k},\qquad
\left(\overline{P}_{t}^{(c),[k]}F\right)(h):=\mathbb{E}_{\Pi}F\left(\Theta_{t}^{(c),[k]}(h;\Pi)\right). Θ t ( c ) , [ k ] ( h ; Π ) := ( Rec h ( c ) ( t , ⋅ ; h i ; Π ) ) i = 1 k , ( P t ( c ) , [ k ] F ) ( h ) := E Π F ( Θ t ( c ) , [ k ] ( h ; Π ) ) .
The future Π \Pi Π is independent of the input and shared by all components. Joint stationarity means ν P ‾ t ( c ) , [ k ] = ν \nu\overline{P}_{t}^{(c),[k]}=\nu ν P t ( c ) , [ k ] = ν . A common geometric update gives joint invariance, and parameter deletion identifies the marginals. Its finite line representation then allows the initial profiles and this shared update environment to converge together to PNG.
The Poisson-line construction
We give the precise Poisson-line construction of the jointly stationary measures ν c ; g 1 , … , g k \nu_{c;g_{1},\ldots,g_{k}} ν c ; g 1 , … , g k from § 1.3 .
Bridge-inclusive càdlàg last passage. Let E = ( E p , … , E q ) E=(E_{p},\ldots,E_{q}) E = ( E p , … , E q ) be càdlàg functions. Write
d E j ( ( r , s ] ) = E j ( s ) − E j ( r ) , d E j ( { r } ) = E j ( r ) − E j ( r − ) \mathrm{d}E_{j}((r,s])=E_{j}(s)-E_{j}(r),\qquad\mathrm{d}E_{j}(\{r\})=E_{j}(r)-E_{j}(r-) d E j (( r , s ]) = E j ( s ) − E j ( r ) , d E j ({ r }) = E j ( r ) − E j ( r − )
for increments and jumps; these agree with signed Stieltjes measures when the lines have locally finite variation.
Definition 2.5 (Last passage with bridge atoms). Fix a finite set M b r ⊂ ( − ∞ , 0 ) \mathcal{M}_{\mathrm{br}}\subset(-\infty,0) M br ⊂ ( − ∞ , 0 ) of marked coordinates. For u = ( a , p ) u=(a,p) u = ( a , p ) , v = ( d , q ) v=(d,q) v = ( d , q ) , a ≤ d a\le d a ≤ d , define paths, weights, and passage values by
a = x p − 1 ≤ x p ≤ ⋯ ≤ x q = d . a=x_{p-1}\le x_{p}\le\cdots\le x_{q}=d. a = x p − 1 ≤ x p ≤ ⋯ ≤ x q = d . wt E ; M b r ( x p − 1 , … , x q ) : = ∑ j = p q { d E j ( ( x j − 1 , x j ] ) + 1 { x j − 1 ∈ M b r } d E j ( { x j − 1 } ) } . \operatorname{wt}_{E;\mathcal{M}_{\mathrm{br}}}(x_{p-1},\ldots,x_{q})
:=\sum_{j=p}^{q}\left\{\mathrm{d}E_{j}((x_{j-1},x_{j}])+\mathbf{1}_{\{x_{j-1}\in\mathcal{M}_{\mathrm{br}}\}}\mathrm{d}E_{j}(\{x_{j-1}\})\right\}. wt E ; M br ( x p − 1 , … , x q ) := j = p ∑ q { d E j (( x j − 1 , x j ]) + 1 { x j − 1 ∈ M br } d E j ({ x j − 1 }) } . L E ; M b r ( u ; v ) : = sup a = x p − 1 ≤ ⋯ ≤ x q = d wt E ; M b r ( x p − 1 , … , x q ) . \mathcal{L}_{E;\mathcal{M}_{\mathrm{br}}}(u;v):=
\sup_{a=x_{p-1}\le\cdots\le x_{q}=d}
\operatorname{wt}_{E;\mathcal{M}_{\mathrm{br}}}(x_{p-1},\ldots,x_{q}). L E ; M br ( u ; v ) := a = x p − 1 ≤ ⋯ ≤ x q = d sup wt E ; M br ( x p − 1 , … , x q ) . The value is − ∞ -\infty − ∞ if a > d a>d a > d or p > q p>q p > q . When the marked set is fixed by the environment, we abbreviate L E ; M b r \mathcal{L}_{E;\mathcal{M}_{\mathrm{br}}} L E ; M br to L E \mathcal{L}_{E} L E .
An ordinary switch atom belongs to the departing line. At marked coordinates, the singleton also collects the bridge on each entered line, including the source line, exactly once.
Every path from ( a , p ) (a,p) ( a , p ) to ( d , q ) (d,q) ( d , q ) has absolute weight at most
2 ∑ j = p q max { sup [ a , d ] ∣ E j ∣ , ∣ E j ( a − ) ∣ } < ∞ , 2\sum_{j=p}^{q}\max\left\{\sup_{[a,d]}|E_{j}|,\left|E_{j}(a-)\right|\right\}<\infty, 2 j = p ∑ q max { [ a , d ] sup ∣ E j ∣ , ∣ E j ( a − ) ∣ } < ∞ , so the passage value is finite for a ≤ d a\le d a ≤ d , p ≤ q p\le q p ≤ q . For finite step environments, switches at atoms or in complementary gaps give finitely many weight patterns, so the supremum is attained.
Candidate jointly stationary measures. We specialize Definition 2.5 to a shared environment of signed Poisson lines and geometric bridge atoms. Figure 3 illustrates the construction for k = 2 k=2 k = 2 .
Definition 2.6 (Poisson-line candidate law). Fix c ≥ 0 c \ge0 c ≥ 0 and g = ( g 1 , … , g k ) \mathbf{g}=(g_1,\ldots,g_k) g = ( g 1 , … , g k ) with g 1 > ⋯ > g k ≥ r c g_1>\cdots>g_k\ge r_c g 1 > ⋯ > g k ≥ r c and g k > c g_k>c g k > c . Write j ˉ : = 2 k + 1 − j \bar{j}:=2k+1-j j ˉ := 2 k + 1 − j and
( γ 1 , … , γ 2 k ) : = ( g 1 , … , g k , g k − 1 , … , g 1 − 1 ) . (\gamma_1,\ldots,\gamma_{2k}):=(g_1,\ldots,g_k,g_k^{-1},\ldots,g_1^{-1}). ( γ 1 , … , γ 2 k ) := ( g 1 , … , g k , g k − 1 , … , g 1 − 1 ) . The environment and its passage profiles are defined as follows.
(i) Poisson lines. Let S 1 g , … , S 2 k g S_1^{\mathbf{g}},\ldots,S_{2k}^{\mathbf{g}} S 1 g , … , S 2 k g be independent signed Poisson walks, started at zero, with line-j j j up/down rates ( γ j , γ j − 1 ) (\gamma_j,\gamma_j^{-1}) ( γ j , γ j − 1 ) .
(ii) Geometric bridges. Independently of the walks, take independent variables Z ( − i , j ) Z(-i,j) Z ( − i , j ) , 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k and i < j ≤ i ˉ i<j\le\bar{i} i < j ≤ i ˉ , with
Z ( − i , j ) ∼ { Geo ( γ j / g i ) , i < j < i ˉ , Geo ( c / g i ) , j = i ˉ . (15) Z(-i,j)\sim
\begin{cases}
\operatorname{Geo}(\gamma_j/g_i), & i<j<\bar{i},\\
\operatorname{Geo}(c/g_i), & j=\bar{i}.
\end{cases} \tag*{(15)} Z ( − i , j ) ∼ { Geo ( γ j / g i ) , Geo ( c / g i ) , i < j < i ˉ , j = i ˉ . (15) For each line j j j , let A j ( c ; g ) A_j^{(c;\mathbf{g})} A j ( c ; g ) be the right-continuous step function on ( − ∞ , 0 ] (-\infty,0] ( − ∞ , 0 ] , normalized by A j ( c ; g ) ( 0 ) = 0 A_j^{(c;\mathbf{g})}(0)=0 A j ( c ; g ) ( 0 ) = 0 , whose Stieltjes measure is
d A j ( c ; g ) = ∑ 1 ≤ i ≤ k i ≤ j ≤ i ˉ Z ( − i , j ) δ − i . \mathrm{d}A_j^{(c;\mathbf{g})}=\sum_{\substack{1\le i\le k\\i\le j\le\bar{i}}} Z(-i,j)\,\delta_{-i}. d A j ( c ; g ) = 1 ≤ i ≤ k i ≤ j ≤ i ˉ ∑ Z ( − i , j ) δ − i . where δ − i \delta_{-i} δ − i is the unit point mass at − i -i − i . Concatenate the bridges and Poisson lines at zero:
E j ( c ; g ) ( x ) : = { A j ( c ; g ) ( x ) , x ≤ 0 , S j g ( x ) , x ≥ 0. E_j^{(c;\mathbf{g})}(x):=
\begin{cases}
A_j^{(c;\mathbf{g})}(x), & x\le0,\\
S_j^{\mathbf{g}}(x), & x\ge0.
\end{cases} E j ( c ; g ) ( x ) := { A j ( c ; g ) ( x ) , S j g ( x ) , x ≤ 0 , x ≥ 0. (iii) Passage profiles and their joint law. Use the passage functional of Definition 2.5 with marked bridge columns M b r = { − k , … , − 1 } \mathcal{M}_{\mathrm{br}}=\{-k,\ldots,-1\} M br = { − k , … , − 1 } and sources s i = ( − i , i ) s_i=(-i,i) s i = ( − i , i ) . In this shared environment, define
G i ( c ; g ) ( x ) : = LE E ( c ; g ) ( s i ; ( x , 2 k ) ) , (16) G_i^{(c;\mathbf{g})}(x):=\operatorname{LE}_{E^{(c;\mathbf{g})}}(s_i;(x,2k)), \tag*{(16)} G i ( c ; g ) ( x ) := LE E ( c ; g ) ( s i ; ( x , 2 k )) , (16) R ~ i ( c ; g ) ( x ) : = G i ( c ; g ) ( x ) − G i ( c ; g ) ( 0 ) , R i ( c ; g ) : = usc R ~ i ( c ; g ) , (17) \widetilde{R}_i^{(c;\mathbf{g})}(x):=G_i^{(c;\mathbf{g})}(x)-G_i^{(c;\mathbf{g})}(0),\qquad R_i^{(c;\mathbf{g})}:=\operatorname{usc}\widetilde{R}_i^{(c;\mathbf{g})}, \tag*{(17)} R i ( c ; g ) ( x ) := G i ( c ; g ) ( x ) − G i ( c ; g ) ( 0 ) , R i ( c ; g ) := usc R i ( c ; g ) , (17) for 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k and x ≥ 0 x\ge0 x ≥ 0 . The candidate jointly stationary law is
ν c ; g 1 , … , g k : = Law ( R 1 ( c ; g ) , … , R k ( c ; g ) ) . \nu_{c;g_1,\ldots,g_k}:=\operatorname{Law}(R_1^{(c;\mathbf{g})},\ldots,R_k^{(c;\mathbf{g})}). ν c ; g 1 , … , g k := Law ( R 1 ( c ; g ) , … , R k ( c ; g ) ) . All coordinates use the same Poisson lines and bridge variables.
All geometric parameters in the definition lie in [ 0 , 1 ) [0,1) [ 0 , 1 ) . In particular, when c < 1 c<1 c < 1 and g k = 1 g_k=1 g k = 1 , the central terminal bridge has law Geo ( c ) \operatorname{Geo}(c) Geo ( c ) and remains well defined. Each G i G_i G i is a measurable, locally finite càdlàg step function. Almost surely, the Poisson lines have no atom in some interval immediately to the right of zero, so upper-semicontinuous regularization preserves pinning. Hence R i ( c ; g ) ∈ U C + , 0 R_i^{(c;\mathbf{g})}\in\mathrm{UC}_{+,0} R i ( c ; g ) ∈ UC + , 0 .
For illustration, take k = 2 k=2 k = 2 with g 2 > c g_2>c g 2 > c ; Figure 3 shows the shared environment. The bridge functions and weight distributions in this example are
d A 1 = 0 , Z ( − 1 , 2 ) ∼ Geo ( g 2 / g 1 ) , d A 2 = Z ( − 1 , 2 ) δ − 1 , Z ( − 2 , 3 ) ∼ Geo ( c / g 2 ) , d A 3 = Z ( − 2 , 3 ) δ − 2 + Z ( − 1 , 3 ) δ − 1 , Z ( − 1 , 3 ) ∼ Geo ( ( g 1 g 2 ) − 1 ) , d A 4 = Z ( − 1 , 4 ) δ − 1 , Z ( − 1 , 4 ) ∼ Geo ( c / g 1 ) . \begin{aligned}
\mathrm{d}A_1&=0, & Z(-1,2)&\sim\operatorname{Geo}(g_2/g_1),\\
\mathrm{d}A_2&=Z(-1,2)\,\delta_{-1}, & Z(-2,3)&\sim\operatorname{Geo}(c/g_2),\\
\mathrm{d}A_3&=Z(-2,3)\,\delta_{-2}+Z(-1,3)\,\delta_{-1}, & Z(-1,3)&\sim\operatorname{Geo}((g_1g_2)^{-1}),\\
\mathrm{d}A_4&=Z(-1,4)\,\delta_{-1}, & Z(-1,4)&\sim\operatorname{Geo}(c/g_1).
\end{aligned} d A 1 d A 2 d A 3 d A 4 = 0 , = Z ( − 1 , 2 ) δ − 1 , = Z ( − 2 , 3 ) δ − 2 + Z ( − 1 , 3 ) δ − 1 , = Z ( − 1 , 4 ) δ − 1 , Z ( − 1 , 2 ) Z ( − 2 , 3 ) Z ( − 1 , 3 ) Z ( − 1 , 4 ) ∼ Geo ( g 2 / g 1 ) , ∼ Geo ( c / g 2 ) , ∼ Geo (( g 1 g 2 ) − 1 ) , ∼ Geo ( c / g 1 ) .
The two profiles use the same four lines and bridges, with entrances ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) and ( − 2 , 2 ) (-2,2) ( − 2 , 2 ) , respectively.
The one-parameter consistency check.
Lemma 2.7 (The one-parameter marginal). For k = 1 k = 1 k = 1 and g > c g > c g > c , the profile R 1 ( c ; g ) R_1^{(c;g)} R 1 ( c ; g ) equals H ( c , g ) H^{(c,g)} H ( c , g ) almost surely under the shared-walk coupling, where H ( c , g ) H^{(c,g)} H ( c , g ) is defined in (1.4).
Proof. The bridge is Y : = Z ( − 1 , 2 ) ∼ Geo ( c / g ) Y := Z(-1, 2) \sim\operatorname{Geo}(c/g) Y := Z ( − 1 , 2 ) ∼ Geo ( c / g ) . Switching at − 1 -1 − 1 , in ( − 1 , 0 ] (-1, 0] ( − 1 , 0 ] , or at y ∈ ( 0 , x ] y \in(0, x] y ∈ ( 0 , x ] gives Y + S 2 ( g ) ( x ) Y + S_2^{(g)}(x) Y + S 2 ( g ) ( x ) , S 2 ( g ) ( x ) S_2^{(g)}(x) S 2 ( g ) ( x ) , or S 1 ( g ) ( y ) + S 2 ( g ) ( x ) − S 2 ( g ) ( y ) S_1^{(g)}(y) + S_2^{(g)}(x) - S_2^{(g)}(y) S 1 ( g ) ( y ) + S 2 ( g ) ( x ) − S 2 ( g ) ( y ) , respectively. Including y = 0 y = 0 y = 0 in the last supremum gives
G 1 ( c ; g ) ( x ) = S 2 ( g ) ( x ) + max { Y , sup 0 ≤ y ≤ x ( S 1 ( g ) ( y ) − S 2 ( g ) ( y ) ) } , G_1^{(c;g)}(x) = S_2^{(g)}(x) + \max\left\{Y,\sup_{0 \le y \le x}\left(S_1^{(g)}(y)-S_2^{(g)}(y)\right)\right\}, G 1 ( c ; g ) ( x ) = S 2 ( g ) ( x ) + max { Y , 0 ≤ y ≤ x sup ( S 1 ( g ) ( y ) − S 2 ( g ) ( y ) ) } ,
R ~ 1 ( c ; g ) ( x ) = S 2 ( g ) ( x ) + ( sup 0 ≤ y ≤ x { S 1 ( g ) ( y ) − S 2 ( g ) ( y ) } − Y ) + . \widetilde{R}_1^{(c;g)}(x) = S_2^{(g)}(x) + \left(\sup_{0 \le y \le x}\left\{S_1^{(g)}(y)-S_2^{(g)}(y)\right\}-Y\right)_+. R 1 ( c ; g ) ( x ) = S 2 ( g ) ( x ) + ( 0 ≤ y ≤ x sup { S 1 ( g ) ( y ) − S 2 ( g ) ( y ) } − Y ) + .
Here G 1 ( 0 ) = Y G_1(0)=Y G 1 ( 0 ) = Y ; the second identity is therefore exactly (1.3), and applying usc proves the claim. □
This includes c = 0 c = 0 c = 0 , when Y = 0 Y = 0 Y = 0 , and c < 1 c < 1 c < 1 , g = 1 g = 1 g = 1 , when the walks have rates ( 1 , 1 ) (1, 1) ( 1 , 1 ) and Y ∼ Geo ( c ) Y \sim\operatorname{Geo}(c) Y ∼ Geo ( c ) .
The singular endpoint. To complete the construction at g k = c ≥ 1 g_k = c \ge1 g k = c ≥ 1 , observe that the terminal bridge Z ( − k , k + 1 ) Z(-k, k + 1) Z ( − k , k + 1 ) would have parameter one. For g ∂ = ( g 1 , … , g k − 1 , c ) \boldsymbol{g}^{\partial} = (g_1, \ldots, g_{k-1}, c) g ∂ = ( g 1 , … , g k − 1 , c ) , set g k = c g_k = c g k = c in the remaining laws, omit this bridge, and replace the last source by
s k ∂ : = ( − k + 1 , k + 1 ) . (18) s_k^{\partial} := (-k + 1, k + 1). \tag*{(18)} s k ∂ := ( − k + 1 , k + 1 ) . (18)
All retained geometric parameters are below one. Put s i ∂ = ( − i , i ) s_i^{\partial} = (-i, i) s i ∂ = ( − i , i ) for i < k i < k i < k , and define the profiles by (2.6)–(2.7) using the resulting environment E ∂ E^{\partial} E ∂ and sources s i ∂ s_i^{\partial} s i ∂ .
Figure 4 illustrates the source shift for k = 2 k = 2 k = 2 . The following lemma justifies the construction.
Figure 4. The singular endpoint for k = 2 k = 2 k = 2 . Removing the diverging bridge Y n = Z n ( − 2 , 3 ) Y_n = Z_n(-2, 3) Y n = Z n ( − 2 , 3 ) shifts s 2 = ( − 2 , 2 ) s_2 = (-2, 2) s 2 = ( − 2 , 2 ) to s 2 ∂ = ( − 1 , 3 ) s_2^{\partial} = (-1, 3) s 2 ∂ = ( − 1 , 3 ) . The bridge Z ( − 1 , 3 ) Z(-1, 3) Z ( − 1 , 3 ) remains included at the new source. Highlighted paths illustrate the shift; the first source is unchanged.
Lemma 2.8 (A diverging terminal bridge). Let c ≥ 1 c \ge1 c ≥ 1 , fix the first k − 1 k - 1 k − 1 spectral parameters, and let g k ( n ) ↓ c g_k^{(n)} \downarrow c g k ( n ) ↓ c through strictly ordered admissible vectors g ( n ) = ( g 1 , … , g k − 1 , g k ( n ) ) \boldsymbol{g}^{(n)} = (g_1, \ldots, g_{k-1}, g_k^{(n)}) g ( n ) = ( g 1 , … , g k − 1 , g k ( n ) ) . Writing g ∂ = ( g 1 , … , g k − 1 , c ) \boldsymbol{g}^{\partial} = (g_1, \ldots, g_{k-1}, c) g ∂ = ( g 1 , … , g k − 1 , c ) , one can couple the profiles so that, for every L < ∞ L < \infty L < ∞ ,
P ( R i ( c ; g ( n ) ) = R i ( c ; g ∂ ) on [ 0 , L ] for every i ≤ k ) ⟶ 1. \mathbb{P}\left(R_i^{(c;g^{(n)})} = R_i^{(c;g^{\partial})}\text{ on }[0,L]\text{ for every }i \le k\right) \longrightarrow1. P ( R i ( c ; g ( n ) ) = R i ( c ; g ∂ ) on [ 0 , L ] for every i ≤ k ) ⟶ 1.
Proof. Remove Y n = Z n ( − k , k + 1 ) ∼ Geo ( c / g k ( n ) ) Y_n = Z_n(-k, k + 1) \sim\operatorname{Geo}(c/g_k^{(n)}) Y n = Z n ( − k , k + 1 ) ∼ Geo ( c / g k ( n ) ) to obtain E ( n ) , ∘ E^{(n),\circ} E ( n ) , ∘ . Its strip [ − k , − k + 1 ) [-k, -k + 1) [ − k , − k + 1 ) is empty, so
L E ( n ) , ∘ ( ( − k , k + 1 ) ; ( x , 2 k ) ) = L E ( n ) , ∘ ( s k ∂ ; ( x , 2 k ) ) . \mathcal{L}_{E^{(n),\circ}}\left((-k, k + 1); (x, 2k)\right) = \mathcal{L}_{E^{(n),\circ}}\left(s_k^{\partial}; (x, 2k)\right). L E ( n ) , ∘ ( ( − k , k + 1 ) ; ( x , 2 k ) ) = L E ( n ) , ∘ ( s k ∂ ; ( x , 2 k ) ) .
Indeed, moving switches to − k + 1 -k + 1 − k + 1 can only gain nonnegative bridges; conversely extend a shifted path backwards on line k + 1 k + 1 k + 1 . The interval atom at − k + 1 -k + 1 − k + 1 equals the shifted source singleton.
The total variation of E ( n ) , ∘ E^{(n),\circ} E ( n ) , ∘ on lines k , … , 2 k k, \ldots, 2k k , … , 2 k and [ − k , L + 1 ] [-k, L + 1] [ − k , L + 1 ] gives a tight bound w n w_n w n for all nonterminal weights. Since Y n → ∞ Y_n \to\infty Y n → ∞ in probability, P ( Y n > 2 w n ) → 1 \mathbb{P}(Y_n > 2w_n) \to1 P ( Y n > 2 w n ) → 1 . On this event every maximizer for the last source and an exit in [ 0 , L + 1 ] [0, L + 1] [ 0 , L + 1 ] uses Y n Y_n Y n , and
G k ( c ; g ( n ) ) ( x ) = Y n + L E ( n ) ( s k ∂ ; ( x , 2 k ) ) . G_k^{(c;g^{(n)})}(x) = Y_n + \mathcal{L}_{E^{(n)}}\left(s_k^{\partial}; (x, 2k)\right). G k ( c ; g ( n ) ) ( x ) = Y n + L E ( n ) ( s k ∂ ; ( x , 2 k ) ) .
Pinning cancels Y n Y_n Y n . Couple the remaining bridges maximally and the Poisson clocks by common rate cutoffs; their configurations agree with the endpoint environment on this compact window with probability tending to one. The other sources cannot access Y n Y_n Y n , and working on [ 0 , L + 1 ] [0, L + 1] [ 0 , L + 1 ] preserves equality on [ 0 , L ] [0, L] [ 0 , L ] after upper-semicontinuous regularization. □
For k = 1 k = 1 k = 1 , the shifted source is s 1 ∂ = ( 0 , 2 ) s_1^{\partial} = (0, 2) s 1 ∂ = ( 0 , 2 ) . There is then no line choice and no bridge at the source, so
R 1 ( c ; c ) = usc S 2 ( c ) , (19) R_1^{(c;c)} = \operatorname{usc} S_2^{(c)}, \tag*{(19)} R 1 ( c ; c ) = usc S 2 ( c ) , (19)
in agreement with (1.5).
For arbitrary parameter lists, use the repetition and permutation convention of Section 1.3.
Geometric stationarity and consistency
We construct a geometric approximation with three properties: consistent spatial-window laws, joint invariance under a shared discrete PNG update, and consistency under deletion of spectral parameters. The next subsections identify its Poisson-line limit and transfer these properties to PNG. Assume
g 1 > ⋯ > g k > c ∨ 1 , c > 0. (20) g_1>\cdots>g_k>c\vee1,\qquad c>0. \tag*{(20)} g 1 > ⋯ > g k > c ∨ 1 , c > 0. (20)
The physical mesh δ \delta δ remains fixed throughout this subsection. The auxiliary parameter η \eta η regularizes reciprocal pairs; we take η ↓ 0 \eta\downarrow0 η ↓ 0 after cancelling source weights. The limit δ ↓ 0 \delta\downarrow0 δ ↓ 0 is taken later. Write j ˉ = 2 k + 1 − j \bar{j}=2k+1-j j ˉ = 2 k + 1 − j and set
γ i ( η ) = g i , γ i ˉ ( η ) = ( 1 − η ) g i − 1 , 1 ≤ i ≤ k . \gamma_i^{(\eta)}=g_i,\qquad\gamma_{\bar{i}}^{(\eta)}=(1-\eta)g_i^{-1},\qquad1\le i\le k. γ i ( η ) = g i , γ i ˉ ( η ) = ( 1 − η ) g i − 1 , 1 ≤ i ≤ k .
Choose 0 < δ < 1 0<\delta<1 0 < δ < 1 , c δ < 1 c\delta<1 cδ < 1 , δ g 1 < 1 \delta g_1<1 δ g 1 < 1 , and 0 < η < 1 − δ g 1 0<\eta<1-\delta g_1 0 < η < 1 − δ g 1 . For a cut base a ∈ Z a\in\mathbb{Z} a ∈ Z , define
α p δ , η , a = { γ p − a ( η ) , a + 1 ≤ p ≤ a + 2 k , δ , otherwise . \alpha_p^{\delta,\eta,a}=
\begin{cases}
\gamma_{p-a}^{(\eta)}, & a+1\le p\le a+2k,\\
\delta, & \text{otherwise}.
\end{cases} α p δ , η , a = { γ p − a ( η ) , δ , a + 1 ≤ p ≤ a + 2 k , otherwise .
Let X δ , η , a X^{\delta,\eta,a} X δ , η , a be the extended half-space geometric environment of [17 Definition 2.2] , with boundary parameter c c c and coordinate parameters α δ , η , a \alpha^{\delta,\eta,a} α δ , η , a . Thus the variables are independent over reflection orbits, X ( p , q ) = X ( q , p ) X(p,q)=X(q,p) X ( p , q ) = X ( q , p ) , and
X ( p , q ) ∼ { Geo ( α p δ , η , a α q δ , η , a ) , p ≠ q , Geo ( c α p δ , η , a ) , p = q , X(p,q)\sim
\begin{cases}
\operatorname{Geo}\left(\alpha_p^{\delta,\eta,a}\alpha_q^{\delta,\eta,a}\right), & p\ne q,\\
\operatorname{Geo}\left(c\alpha_p^{\delta,\eta,a}\right), & p=q,
\end{cases} X ( p , q ) ∼ { Geo ( α p δ , η , a α q δ , η , a ) , Geo ( c α p δ , η , a ) , p = q , p = q ,
for parameters in ( 0 , 1 ) (0,1) ( 0 , 1 ) , and + ∞ +\infty + ∞ otherwise. For u , v ∈ Z 2 u,v\in\mathbb{Z}^2 u , v ∈ Z 2 , let P ( u , v ) \mathcal{P}(u,v) P ( u , v ) consist of the vertex paths from u u u to v v v with steps ( 1 , 0 ) (1,0) ( 1 , 0 ) and ( 0 , 1 ) (0,1) ( 0 , 1 ) , subject to ( r , r ) ∈ π ⇒ ( r − 1 , r ) ∉ π (r,r)\in\pi\Rightarrow(r-1,r)\notin\pi ( r , r ) ∈ π ⇒ ( r − 1 , r ) ∈ / π for every r ∈ Z r\in\mathbb{Z} r ∈ Z . Define the extended half-space last-passage value by
L X δ , η , a ( u ; v ) : = max π ∈ P ( u , v ) ∑ w ∈ π X ( w ) , max ∅ : = − ∞ . L_{X^{\delta,\eta,a}}(u;v):=\max_{\pi\in\mathcal{P}(u,v)}\sum_{w\in\pi}X(w),\qquad\max\varnothing:=-\infty. L X δ , η , a ( u ; v ) := π ∈ P ( u , v ) max w ∈ π ∑ X ( w ) , max ∅ := − ∞.
Both endpoint weights are included.
For 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k and M ∈ N 0 M\in\mathbb{N}_0 M ∈ N 0 , put
s i a = ( a + i ˉ , a + i ) , N M = a + 2 k + M , v m M = ( N M + m , N M − m ) , 0 ≤ m ≤ M . s_i^a=(a+\bar{i},a+i),\qquad N_M=a+2k+M,\qquad v_m^M=(N_M+m,N_M-m),\qquad0\le m\le M. s i a = ( a + i ˉ , a + i ) , N M = a + 2 k + M , v m M = ( N M + m , N M − m ) , 0 ≤ m ≤ M .
Define the pinned antidiagonal field
H i , M δ , η , a ( m ) = L X δ , η , a ( s i a ; v m M ) − L X δ , η , a ( s i a ; v 0 M ) . (21) H_{i,M}^{\delta,\eta,a}(m)=L_{X^{\delta,\eta,a}}(s_i^a;v_m^M)-L_{X^{\delta,\eta,a}}(s_i^a;v_0^M). \tag*{(21)} H i , M δ , η , a ( m ) = L X δ , η , a ( s i a ; v m M ) − L X δ , η , a ( s i a ; v 0 M ) . (21)
For the sources and terminals in (2.11), the path restriction keeps every admissible path in { ( p , q ) : p ≥ q } \{(p,q):p\ge q\} {( p , q ) : p ≥ q } . Simultaneous translation of the parameters, sources, and terminals by ( − a , − a ) (-a,-a) ( − a , − a ) shows that these window laws do not depend on a a a .
Define the light-cone map by
Φ δ , N ( p , q ) = ( δ ( p + q 2 − N ) , δ 2 ( p − q ) ) . (22) \Phi_{\delta,N}(p,q)=\left(\delta\left(\frac{p+q}{2}-N\right),\frac{\delta}{2}(p-q)\right). \tag*{(22)} Φ δ , N ( p , q ) = ( δ ( 2 p + q − N ) , 2 δ ( p − q ) ) . (22)
With N = N M N=N_M N = N M , this map sends v m M v_m^M v m M to ( 0 , δ m ) (0,\delta m) ( 0 , δ m ) and v m M + ( n , n ) v_m^M+(n,n) v m M + ( n , n ) to ( n δ , δ m ) (n\delta,\delta m) ( n δ , δ m ) , so the antidiagonal field is read at one physical time. Figure 5 shows the cuts and the two fresh lattice layers in both coordinate systems.
Figure 5. Antidiagonal cuts and the two-layer update, with k = 2 k=2 k = 2 , a = 0 a=0 a = 0 , M = 2 M=2 M = 2 , and N = 6 N=6 N = 6 . The vertices v m = ( 6 + m , 6 − m ) v_m=(6+m,6-m) v m = ( 6 + m , 6 − m ) and v m ′ = v m + ( 1 , 1 ) v'_m=v_m+(1,1) v m ′ = v m + ( 1 , 1 ) , m = 0 , 1 , 2 m=0,1,2 m = 0 , 1 , 2 , map to ( t , z ) = ( 0 , m δ ) (t,z)=(0,m\delta) ( t , z ) = ( 0 , m δ ) and ( δ , m δ ) (\delta,m\delta) ( δ , m δ ) under Φ δ , N \Phi_{\delta,N} Φ δ , N ; panel (b) plots space horizontally and time vertically. Orange and blue vertices belong to the two fresh layers; the red square is the new diagonal vertex.
Lemma 2.9 (Accessible cells and cancellation of source weights). For a path starting at s i a s_i^a s i a , the source is the only accessible cell whose geometric parameter tends to one as η ↓ 0 \eta\downarrow0 η ↓ 0 . Its weight occurs in every path and cancels in (2.11). The remaining weights can be coupled so that, for every fixed M M M , jointly in i ≤ k i\le k i ≤ k and 0 ≤ m ≤ M 0\le m\le M 0 ≤ m ≤ M ,
H i , M δ , η , a ( m ) → η ↓ 0 a . s . H i , M δ , a ( m ) . H_{i,M}^{\delta,\eta,a}(m)\xrightarrow[\eta\downarrow0]{\mathrm{a.s.}}H_{i,M}^{\delta,a}(m). H i , M δ , η , a ( m ) a.s. η ↓ 0 H i , M δ , a ( m ) .
Proof. The source parameter is γ i ( η ) γ i ( η ) = 1 − η \gamma_i^{(\eta)}\gamma_i^{(\eta)}=1-\eta γ i ( η ) γ i ( η ) = 1 − η . An accessible mixed special cell with labels r ˉ , ℓ \bar r,\ell r ˉ , ℓ satisfies r ≤ i ≤ ℓ r\le i\le\ell r ≤ i ≤ ℓ , so its parameter is ( 1 − η ) g ℓ g r ≤ 1 − η (1-\eta)\frac{g_\ell}{g_r}\le1-\eta ( 1 − η ) g r g ℓ ≤ 1 − η , with equality only for r = ℓ = i r=\ell=i r = ℓ = i . Every other mixed parameter has a limit strictly below one. Two reflected labels give a parameter at most ( 1 − η ) 2 / g k 2 (1-\eta)^2/g_k^2 ( 1 − η ) 2 / g k 2 ; a background coordinate gives at most δ g 1 \delta g_1 δ g 1 or δ 2 \delta^2 δ 2 ; an accessible special diagonal parameter is at most c ( 1 − η ) / g i c(1-\eta)/g_i c ( 1 − η ) / g i . All these limits are below one. Cells with two unreflected special labels are inaccessible. Thus every accessible weight is finite at positive η \eta η , and all but the source weight have proper limits. Common-uniform coupling of the finitely many remaining weights and finite maximization prove the assertion after the source contribution is cancelled. □
This cancellation parallels [17 Proposition 2.8] . All parameter swaps below are performed at positive η \eta η .
Parameter swaps and moving sources. Let X ⟨ ℓ ⟩ X^{\langle\ell\rangle} X ⟨ ℓ ⟩ have the coordinate parameters at ℓ , ℓ + 1 \ell,\ell+1 ℓ , ℓ + 1 interchanged. The shifted version of [17 Lemma 2.6] gives
{ L X ( ( p , q ) ; ( r , s ) ) : p ≤ r , q ≤ s , p , q ≠ ℓ + 1 , r , s ≠ ℓ } = d { L X ⟨ ℓ ⟩ ( ( p , q ) ; ( r , s ) ) : p ≤ r , q ≤ s , p , q ≠ ℓ + 1 , r , s ≠ ℓ } . (23) \{L_X((p,q);(r,s)):p\le r,\ q\le s,\ p,q\ne\ell+1,\ r,s\ne\ell\}\mathrel{\overset{\mathrm d}{=}}\{L_{X^{\langle\ell\rangle}}((p,q);(r,s)):p\le r,\ q\le s,\ p,q\ne\ell+1,\ r,s\ne\ell\}. \tag*{(23)} { L X (( p , q ) ; ( r , s )) : p ≤ r , q ≤ s , p , q = ℓ + 1 , r , s = ℓ } = d { L X ⟨ ℓ ⟩ (( p , q ) ; ( r , s )) : p ≤ r , q ≤ s , p , q = ℓ + 1 , r , s = ℓ } . (23)
This identity applies jointly to all retained sources and terminals. We first apply it to unpinned passage values, then take their pinned differences.
The swaps also require transporting paired sources with their labels. The following observation explains these source shifts. In a post-swap environment, suppose that the new source w w w is one admissible north or east step from its old source u u u and can reach every retained terminal, that Y η : = X ( w ) ∼ Geo ( 1 − η ) Y^\eta:=X(w)\sim\operatorname{Geo}(1-\eta) Y η := X ( w ) ∼ Geo ( 1 − η ) , and that every other weight accessible from u u u to the retained terminals has a proper limit. Their finite sum w η w^\eta w η is tight, whereas Y η → ∞ Y^\eta\to\infty Y η → ∞ in probability. On { Y η > w η } \{Y^\eta>w^\eta\} { Y η > w η } , every maximizer visits w w w : a path through w w w has weight at least Y η Y^\eta Y η , while a path avoiding it has weight at most w η w^\eta w η . Endpoint inclusion and adjacency then give
L X ( u ; v ) = X ( u ) + L X ( w ; v ) L_X(u;v)=X(u)+L_X(w;v) L X ( u ; v ) = X ( u ) + L X ( w ; v )
for every retained terminal v v v . The extra term cancels under pinning. Thus exact equality in law under a swap can be combined with a source shift whose pinned fields agree with probability tending to one as η ↓ 0 \eta\downarrow0 η ↓ 0 .
Cut consistency and the common update. One discrete time step adds two independent lattice layers. Let ( ξ m ∘ ) m ≥ 0 (\xi_m^\circ)_{m\ge0} ( ξ m ∘ ) m ≥ 0 , ( ξ m ε ) m ≥ 1 (\xi_m^\varepsilon)_{m\ge1} ( ξ m ε ) m ≥ 1 , and d d d be mutually independent and independent of the input, with ξ m ∘ , ξ m ε ∼ Geo ( δ 2 ) \xi_m^\circ,\xi_m^\varepsilon\sim\operatorname{Geo}(\delta^2) ξ m ∘ , ξ m ε ∼ Geo ( δ 2 ) and d ∼ Geo ( c δ ) d\sim\operatorname{Geo}(c\delta) d ∼ Geo ( cδ ) . For pinned h h h , set
K h ( m + 1 2 ) : = ξ m ∘ + max { h ( m ) , h ( m + 1 ) } , m ≥ 0. (24) K_h\left(m+\frac12\right):=\xi_m^\circ+\max\{h(m),h(m+1)\},\qquad m\ge0. \tag*{(24)} K h ( m + 2 1 ) := ξ m ∘ + max { h ( m ) , h ( m + 1 )} , m ≥ 0. (24)
r h : = d + K h ( 1 2 ) , ( Ψ δ ( c ) h ) ( 0 ) : = 0 , r_h:=d+K_h\left(\frac12\right),\qquad(\Psi_\delta^{(c)}h)(0):=0, r h := d + K h ( 2 1 ) , ( Ψ δ ( c ) h ) ( 0 ) := 0 ,
( Ψ δ ( c ) h ) ( m ) : = ξ m ε + max { K h ( m − 1 2 ) , K h ( m + 1 2 ) } − r h , m ≥ 1. (25) (\Psi_\delta^{(c)}h)(m):=\xi_m^\varepsilon+\max\left\{K_h\left(m-\frac12\right),K_h\left(m+\frac12\right)\right\}-r_h,\qquad m\ge1. \tag*{(25)} ( Ψ δ ( c ) h ) ( m ) := ξ m ε + max { K h ( m − 2 1 ) , K h ( m + 2 1 ) } − r h , m ≥ 1. (25)
Use the same fresh array ξ = ( ξ ∘ , ξ ε , d ) \xi=(\xi^\circ,\xi^\varepsilon,d) ξ = ( ξ ∘ , ξ ε , d ) for every component, each with its own recentering constant. Write Ψ δ , ξ ( c ) , [ k ] \Psi_{\delta,\xi}^{(c),[k]} Ψ δ , ξ ( c ) , [ k ] for this shared update, or Ψ δ ( c ) , [ k ] \Psi_\delta^{(c),[k]} Ψ δ ( c ) , [ k ] when the noise is implicit. Different time steps use independent fresh arrays.
Lemma 2.10 (Cut consistency and stationarity). For every M ∈ N 0 M\in\mathbb{N}_0 M ∈ N 0 ,
( H i , M + 1 δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) = d ( H i , M δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) . (26) \left(H_{i,M+1}^{\delta,a}(m):i\le k,\ 0\le m\le M\right)\mathrel{\overset{\mathrm d}{=}}\left(H_{i,M}^{\delta,a}(m):i\le k,\ 0\le m\le M\right). \tag*{(26)} ( H i , M + 1 δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) = d ( H i , M δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) . (26) These consistent window laws define a law μ δ , g a \mu_{\delta,g}^{a} μ δ , g a on ( Z ≥ 0 N 0 ) k (\mathbb{Z}_{\ge0}^{\mathbb{N}_0})^k ( Z ≥ 0 N 0 ) k , independent of a a a . This half-line law is invariant under the common two-layer update Ψ δ ( c ) , [ k ] \Psi_{\delta}^{(c),[k]} Ψ δ ( c ) , [ k ] .
Proof. Step 1: consistency of the cut laws. At positive η \eta η , move the background parameter at a + 2 k + 1 a+2k+1 a + 2 k + 1 to a + 1 a+1 a + 1 through the swaps τ a + 2 k , τ a + 2 k − 1 , … , τ a + 1 \tau_{a+2k},\tau_{a+2k-1},\ldots,\tau_{a+1} τ a + 2 k , τ a + 2 k − 1 , … , τ a + 1 , where τ ℓ \tau_\ell τ ℓ interchanges coordinates ℓ , ℓ + 1 \ell,\ell+1 ℓ , ℓ + 1 . At stage j ∈ { 0 , … , 2 k } j\in\{0,\ldots,2k\} j ∈ { 0 , … , 2 k } , the special labels occupy
p h [ j ] : = a + h + 1 { h > j } , s i [ j ] : = ( p i [ j ] , p i [ j ] ) , 1 ≤ h ≤ 2 k . p_h^{[j]}:=a+h+\mathbf{1}_{\{h>j\}},\qquad s_i^{[j]}:=\left(p_i^{[j]},p_i^{[j]}\right),\qquad1\le h\le2k. p h [ j ] := a + h + 1 { h > j } , s i [ j ] := ( p i [ j ] , p i [ j ] ) , 1 ≤ h ≤ 2 k .
The background parameter is at a + j + 1 a+j+1 a + j + 1 . Let X η , [ j ] X^{\eta,[j]} X η , [ j ] be this environment, and let H i η , [ j ] ( m ) H_i^{\eta,[j]}(m) H i η , [ j ] ( m ) denote passage from s i [ j ] s_i^{[j]} s i [ j ] to v m M + 1 v_m^{M+1} v m M + 1 , pinned at v 0 M + 1 v_0^{M+1} v 0 M + 1 , for 0 ≤ m ≤ M 0\le m\le M 0 ≤ m ≤ M .
In the step j → j − 1 j\to j-1 j → j − 1 , put ℓ = a + j \ell=a+j ℓ = a + j . No source uses the background coordinate ℓ + 1 \ell+1 ℓ + 1 , and
min { ( v m M + 1 ) 1 , ( v m M + 1 ) 2 } = a + 2 k + M + 1 − m ≥ a + 2 k + 1 > ℓ . \min\{(v_m^{M+1})_1,(v_m^{M+1})_2\}=a+2k+M+1-m\ge a+2k+1>\ell. min {( v m M + 1 ) 1 , ( v m M + 1 ) 2 } = a + 2 k + M + 1 − m ≥ a + 2 k + 1 > ℓ .
Thus (23) applies. If H ~ i η , [ j ] \widetilde H_i^{\eta,[j]} H i η , [ j ] is the post-swap field from the old source s i [ j ] s_i^{[j]} s i [ j ] , it gives exact joint equality in law between H η , [ j ] H^{\eta,[j]} H η , [ j ] and H ~ η , [ j ] \widetilde H^{\eta,[j]} H η , [ j ] .
Only i j = min { j , 2 k + 1 − j } i_j=\min\{j,2k+1-j\} i j = min { j , 2 k + 1 − j } changes its source, since
s i j [ j − 1 ] − s i j [ j ] = ( 1 { j = i j } , 1 { j = i j } ) . s_{i_j}^{[j-1]}-s_{i_j}^{[j]}=(\mathbf{1}_{\{j=i_j\}},\mathbf{1}_{\{j=i_j\}}). s i j [ j − 1 ] − s i j [ j ] = ( 1 { j = i j } , 1 { j = i j } ) .
Its new source has parameter 1 − η 1-\eta 1 − η . The accessibility argument of Lemma 2.9 extends to these intermediate orders, which insert one background coordinate among the ordered special labels. From the affected old source, an accessible mixed special cell still has labels r ˉ , ℓ ′ \bar r,\ell' r ˉ , ℓ ′ with r ≤ i j ≤ ℓ ′ r\le i_j\le\ell' r ≤ i j ≤ ℓ ′ . Its limiting parameter g ℓ ′ / g r g_{\ell'}/g_r g ℓ ′ / g r is strictly below one except at the new paired source. Any additional cell involving the inserted background coordinate has parameter bounded by δ g 1 \delta g_1 δ g 1 or δ 2 \delta^2 δ 2 ; reflected and diagonal cells have the same proper bounds as in that lemma. Hence all weights other than the new source have parameters uniformly below one as η ↓ 0 \eta\downarrow0 η ↓ 0 . The source-shift observation above therefore yields
P ( H ~ η , [ j ] = H η , [ j − 1 ] ) ⟶ 1 , \mathbb{P}\left(\widetilde{H}^{\eta,[j]}=H^{\eta,[j-1]}\right)\longrightarrow1, P ( H η , [ j ] = H η , [ j − 1 ] ) ⟶ 1 ,
where the arrays include all i ≤ k i\leq k i ≤ k and 0 ≤ m ≤ M 0\leq m\leq M 0 ≤ m ≤ M . The same accessibility bounds give the pinned limits at each stage. Combining these limits with the exact swap identities and telescoping bounded-Lipschitz expectations over the 2 k 2k 2 k swaps identifies the limiting laws at stages 2 k 2k 2 k and 0 0 0 .
Finally, s i [ 0 ] = s i [ 2 k ] + ( 1 , 1 ) s_i^{[0]}=s_i^{[2k]}+(1,1) s i [ 0 ] = s i [ 2 k ] + ( 1 , 1 ) . Translation by ( − 1 , − 1 ) (-1,-1) ( − 1 , − 1 ) restores the original parameter sequence and sends v m M + 1 v_m^{M+1} v m M + 1 to v m M v_m^M v m M , proving (26) . Kolmogorov consistency now defines the half-line law μ c ; g δ \mu_{c;g}^{\delta} μ c ; g δ ; its independence of a a a follows from the diagonal translation noted above.
Step 2: invariance under the common update. Output coordinates 0 , … , M 0,\ldots,M 0 , … , M use only input coordinates 0 , … , M + 1 0,\ldots,M+1 0 , … , M + 1 . Realize this input prefix on the cut N = N M + 1 = a + 2 k + M + 1 N=N_{M+1}=a+2k+M+1 N = N M + 1 = a + 2 k + M + 1 . The fresh vertices are
o m = ( N + m + 1 , N − m ) , e m = ( N + m + 1 , N − m + 1 ) , 0 ≤ m ≤ M . o_m=(N+m+1,N-m),\qquad e_m=(N+m+1,N-m+1),\qquad0\leq m\leq M. o m = ( N + m + 1 , N − m ) , e m = ( N + m + 1 , N − m + 1 ) , 0 ≤ m ≤ M .
Both coordinates exceed a + 2 k a+2k a + 2 k , so these are background sites: every o m o_m o m and each e m e_m e m with m ≥ 1 m\geq1 m ≥ 1 have parameter δ 2 \delta^2 δ 2 , while e 0 = ( N + 1 , N + 1 ) e_0=(N+1,N+1) e 0 = ( N + 1 , N + 1 ) has parameter c δ c\delta cδ . The input passage values use vertices with p + q ≤ 2 N p+q\leq2N p + q ≤ 2 N , whereas the fresh layers have p + q = 2 N + 1 , 2 N + 2 p+q=2N+1,2N+2 p + q = 2 N + 1 , 2 N + 2 . Reflection preserves p + q p+q p + q , so their reflection orbits are disjoint and the fresh weights are independent of the entire input prefix.
The Bellman recurrence gives (2.14)–(2.15) simultaneously for all sources. The new height at the origin before recentering is d + K h ( 1 / 2 ) d+K_h(1/2) d + K h ( 1/2 ) , whose subtraction pins each output. The new cut is N + 1 = N M + 2 N+1=N_{M+2} N + 1 = N M + 2 , so the output prefix has law ( H i , M + 2 δ , c ( m ) : i ≤ k , 0 ≤ m ≤ M ) (H_{i,M+2}^{\delta,c}(m):i\leq k,\ 0\leq m\leq M) ( H i , M + 2 δ , c ( m ) : i ≤ k , 0 ≤ m ≤ M ) . Step 1 identifies this with the original M M M -window law.
Each output coordinate depends on finitely many input coordinates and fresh weights. Equality on all finite cylinders therefore proves invariance of the full law; equivalently, for every bounded measurable F F F ,
∫ E ξ F ( T δ , ξ ( c ) , [ k ] ( h ) ) μ c ; g δ ( d h ) = ∫ F ( h ) μ c ; g δ ( d h ) . \int\mathbb{E}_{\xi}F\left(\mathcal{T}_{\delta,\xi}^{(c),[k]}(h)\right)\,\mu_{c;g}^{\delta}(dh)
=
\int F(h)\,\mu_{c;g}^{\delta}(dh). ∫ E ξ F ( T δ , ξ ( c ) , [ k ] ( h ) ) μ c ; g δ ( d h ) = ∫ F ( h ) μ c ; g δ ( d h ) .
□
Deleting a spectral parameter. We next move the two labels representing one spectral parameter outside the region accessible to the retained sources. Write
g ^ m : = ( g 1 , … , g m − 1 , g m + 1 , … , g k ) , D m ( h 1 , … , h k ) : = ( h 1 , … , h m − 1 , h m + 1 , … , h k ) . \widehat{g}^{\,m}:=(g_1,\ldots,g_{m-1},g_{m+1},\ldots,g_k),\qquad
D_m(h_1,\ldots,h_k):=(h_1,\ldots,h_{m-1},h_{m+1},\ldots,h_k). g m := ( g 1 , … , g m − 1 , g m + 1 , … , g k ) , D m ( h 1 , … , h k ) := ( h 1 , … , h m − 1 , h m + 1 , … , h k ) .
Lemma 2.11 (Paired-parameter deletion). Assume k ≥ 2 k\geq2 k ≥ 2 and (2.10). For every admissible δ > 0 \delta>0 δ > 0 and 1 ≤ m ≤ k 1\leq m\leq k 1 ≤ m ≤ k ,
( D m ) ∗ μ c ; g δ = μ c ; g ^ m δ (27) (D_m)_*\mu_{c;g}^{\delta}=\mu_{c;\widehat{g}^{\,m}}^{\delta} \tag*{(27)} ( D m ) ∗ μ c ; g δ = μ c ; g m δ (27) on ( Z N 0 ) k − 1 (\mathbb{Z}^{\mathbb{N}_0})^{k-1} ( Z N 0 ) k − 1 .
Proof. By translation, take a = 0 a=0 a = 0 . Put N = m ‾ N=\overline{m} N = m and retain the regularization γ i ( η ) = g i \gamma_i^{(\eta)}=g_i γ i ( η ) = g i , γ i ( η ) = ( 1 − η ) / g i \gamma_i^{(\eta)}=(1-\eta)/g_i γ i ( η ) = ( 1 − η ) / g i . Move label N N N to the first position and then label m m m to the second, through adjacent swaps:
σ m i d : = ( N , 1 , … , N − 1 , N + 1 , … , 2 k ) , \sigma_{\mathrm{mid}}:=(N,1,\ldots,N-1,N+1,\ldots,2k), σ mid := ( N , 1 , … , N − 1 , N + 1 , … , 2 k ) ,
( 1 , … , 2 k ) ⟶ σ m i d ⟶ ( N , m , 1 , … , m − 1 , m + 1 , … , N − 1 , N + 1 , … , 2 k ) . (28) (1,\ldots,2k)\longrightarrow\sigma_{\mathrm{mid}}\longrightarrow(N,m,1,\ldots,m-1,m+1,\ldots,N-1,N+1,\ldots,2k). \tag*{(28)} ( 1 , … , 2 k ) ⟶ σ mid ⟶ ( N , m , 1 , … , m − 1 , m + 1 , … , N − 1 , N + 1 , … , 2 k ) . (28)
For an intermediate order σ \sigma σ , assign parameter γ σ ( p ) ( η ) \gamma_{\sigma(p)}^{(\eta)} γ σ ( p ) ( η ) to position 1 ≤ p ≤ 2 k 1\leq p\leq2k 1 ≤ p ≤ 2 k and δ \delta δ elsewhere, and denote the environment by X η , σ X^{\eta,\sigma} X η , σ . The retained source for i ≠ m i\ne m i = m is
s i σ : = ( σ − 1 ( i ) , σ − 1 ( i ) ) . s_i^\sigma:=(\sigma^{-1}(i),\sigma^{-1}(i)). s i σ := ( σ − 1 ( i ) , σ − 1 ( i )) .
Use finite antidiagonal terminal sets with coordinates at least 2 k 2k 2 k . Before a swap at d , d + 1 d,d+1 d , d + 1 , the moving deleted label is at d + 1 d+1 d + 1 , which no retained source uses. The terminals avoid d ≤ 2 k − 1 d\leq2k-1 d ≤ 2 k − 1 . Thus (2.13) applies jointly to every retained source and terminal.
When N N N crosses m m m , no retained source moves. In every other swap, exactly one source moves to an adjacent paired source of parameter 1 − η 1-\eta 1 − η . The possibilities are as follows; the affected component is i i i in each row, and the last column gives the post-swap parameter at its old source:
All three limiting parameters are strictly below one. The relative orders of the retained unreflected labels and of the retained reflected labels stay unchanged. Hence an accessible mixed cell with retained labels r ˉ , ℓ \bar r,\ell r ˉ , ℓ satisfies r ≤ i ≤ ℓ r\leq i\leq\ell r ≤ i ≤ ℓ , and its parameter ( 1 − η ) g ℓ / g r (1-\eta)g_\ell/g_r ( 1 − η ) g ℓ / g r has a proper limit unless r = ℓ = i r=\ell=i r = ℓ = i , the new source. Products involving a moved deleted label are controlled by the same inequalities i > m i>m i > m or i < m i<m i < m displayed in the table; the deleted pair’s own cell is inaccessible. Reflected, diagonal, and background cells satisfy the proper bounds used above. Thus the sum of all other accessible weights is tight on each finite terminal cone.
The source-shift observation now applies at every step: with probability tending to one, all retained-terminal maximizers from the affected old source visit the new source, and
L X p o s t ( η ) ( s − ; v ) = X p o s t ( η ) ( s − ) + L X p o s t ( η ) ( s + ; v ) . L_{X_{\mathrm{post}}^{(\eta)}}(s^-;v)
=
X_{\mathrm{post}}^{(\eta)}(s^-)
+
L_{X_{\mathrm{post}}^{(\eta)}}(s^+;v). L X post ( η ) ( s − ; v ) = X post ( η ) ( s − ) + L X post ( η ) ( s + ; v ) .
Pinning removes the extra term. Combining these source shifts with the exact joint swap identities and taking η ↓ 0 \eta\downarrow0 η ↓ 0 identifies the retained pinned fields before and after (28) .
In the final order, both coordinates of every retained source are at least 3, so the deleted positions 1, 2 are inaccessible. Remove them and translate by ( − 2 , − 2 ) (-2,-2) ( − 2 , − 2 ) . If pos ( h ) \operatorname{pos}(h) pos ( h ) denotes the final position of label h h h , then
( pos ( i ˉ ) , pos ( i ) ) = { ( i ˉ , i + 2 ) , i < m , ( i ˉ + 1 , i + 1 ) , i > m . \left(\operatorname{pos}(\bar{i}),\operatorname{pos}(i)\right)=
\begin{cases}
(\bar{i},i+2), & i<m,\\
(\bar{i}+1,i+1), & i>m.
\end{cases} ( pos ( i ˉ ) , pos ( i ) ) = { ( i ˉ , i + 2 ) , ( i ˉ + 1 , i + 1 ) , i < m , i > m .
After translation, these are precisely the paired sources of the reduced vector, indexed by i i i for i < m i<m i < m and i − 1 i-1 i − 1 for $i>m. The accessible parameter environment also becomes that of the reduced vector. Finally, for
V M ( r ) : = { ( r + M + j , r + M − j ) : 0 ≤ j ≤ M } , V M ( 2 k ) − ( 2 , 2 ) = V M ( 2 k − 2 ) , V_M^{(r)}:=\{(r+M+j,r+M-j):0\le j\le M\},\qquad V_M^{(2k)}-(2,2)=V_M^{(2k-2)}, V M ( r ) := {( r + M + j , r + M − j ) : 0 ≤ j ≤ M } , V M ( 2 k ) − ( 2 , 2 ) = V M ( 2 k − 2 ) ,
the corresponding M M M -window laws agree. Since this holds for every M M M , it proves (27) on the half-line sequence space. □ \square □
A line representation of the geometric field
We identify the geometric stationary law μ c δ ; g \mu_c^{\delta;g} μ c δ ; g with a discrete version of the Poisson-line construction. The representation uses independent geometric bridges and signed geometric increments. It is exact when the atoms in each mesh interval are placed in a specified order. Moving them to the mesh points changes their mesh readouts on a fixed window with probability at most O ( δ ) O(\delta) O ( δ ) .
Throughout, c > 0 c>0 c > 0 , g 1 > ⋯ > g k > c ∨ 1 g_1>\cdots>g_k>c\vee1 g 1 > ⋯ > g k > c ∨ 1 , and δ \delta δ satisfies the restrictions of the preceding subsection. Put n = 2 k n=2k n = 2 k , j ˉ = n + 1 − j \bar{j}=n+1-j j ˉ = n + 1 − j , γ i = g i \gamma_i=g_i γ i = g i , and γ ˉ i = g i − 1 \bar{\gamma}_i=g_i^{-1} γ ˉ i = g i − 1 . We first take η ↓ 0 \eta\downarrow0 η ↓ 0 at fixed δ \delta δ and set the cancelled source weights to zero, as in Lemma 2.9. The limit δ ↓ 0 \delta\downarrow0 δ ↓ 0 is taken only in the next subsection.
The representation. For a nonnegative bridge array Z = ( Z i , j : 1 ≤ i ≤ k , i < j ≤ i ˉ ) Z=(Z_{i,j}:1\le i\le k,\ i<j\le\bar{i}) Z = ( Z i , j : 1 ≤ i ≤ k , i < j ≤ i ˉ ) , define
d E Z , j : = ∑ 1 ≤ i ≤ k i < j ≤ i ˉ Z i , j δ − i , E Z , j ( 0 ) = 0 , \mathrm{d}E_{Z,j}:=\sum_{\substack{1\le i\le k\\i<j\le\bar{i}}}Z_{i,j}\,\delta^{-i},\qquad E_{Z,j}(0)=0, d E Z , j := 1 ≤ i ≤ k i < j ≤ i ˉ ∑ Z i , j δ − i , E Z , j ( 0 ) = 0 ,
C Z ( i , q ) : = L E Z ( ( − i , i ) ; ( 0 , q ) ) , 1 ≤ i ≤ k , 1 ≤ q ≤ n . C^Z(i,q):=\mathcal{L}_{E_Z}\bigl((-i,i);(0,q)\bigr),\qquad1\le i\le k,\quad1\le q\le n. C Z ( i , q ) := L E Z ( ( − i , i ) ; ( 0 , q ) ) , 1 ≤ i ≤ k , 1 ≤ q ≤ n .
Here L \mathcal{L} L is the passage functional of Definition 2.5 , with marked bridge coordinates { − k , … , − 1 } \{-k,\ldots,-1\} { − k , … , − 1 } . An impossible connection has value − ∞ -\infty − ∞ .
For compatible matrices, write
( U ⊙ V ) ( p , q ) : = max r { U ( p , r ) + V ( r , q ) } . (U\odot V)(p,q):=\max_r\{U(p,r)+V(r,q)\}. ( U ⊙ V ) ( p , q ) := r max { U ( p , r ) + V ( r , q )} .
We also use juxtaposition for this max-plus product and its row-vector version. For 1 ≤ p , q , j ≤ n 1\le p,q,j\le n 1 ≤ p , q , j ≤ n , the transfer matrices for a positive atom u u u and a negative atom − v -v − v on line j j j are
P j ( u ) ( p , q ) : = { u 1 { p ≤ j ≤ q } , p ≤ q , − ∞ , p > q , P_j(u)(p,q):=
\begin{cases}
u\mathbf{1}_{\{p\le j\le q\}}, & p\le q,\\
-\infty, & p>q,
\end{cases} P j ( u ) ( p , q ) := { u 1 { p ≤ j ≤ q } , − ∞ , p ≤ q , p > q ,
N j ( v ) ( p , q ) : = { − v , p = q = j , 0 , p ≤ q , ( p , q ) ≠ ( j , j ) , − ∞ , p > q . N_j(v)(p,q):=
\begin{cases}
-v, & p=q=j,\\
0, & p\le q,\ (p,q)\ne(j,j),\\
-\infty, & p>q.
\end{cases} N j ( v ) ( p , q ) := ⎩ ⎨ ⎧ − v , 0 , − ∞ , p = q = j , p ≤ q , ( p , q ) = ( j , j ) , p > q .
A positive atom can be collected whenever its line is crossed. A negative atom can be avoided unless the entrance and exit lines both equal j j j ; this includes atoms at the right endpoint under the one-sided convention in Definition 2.5 . For arrays U , V U,V U , V , set
T m : = N n ( V n , m ) ⊙ ⋯ ⊙ N 1 ( V 1 , m ) ⊙ P 1 ( U 1 , m ) ⊙ ⋯ ⊙ P n ( U n , m ) , T 1 : m : = T 1 ⊙ ⋯ ⊙ T m , T 1 : 0 : = I , T_m:=N_n(V_{n,m})\odot\cdots\odot N_1(V_{1,m})\odot P_1(U_{1,m})\odot\cdots\odot P_n(U_{n,m}),\qquad T_{1:m}:=T_1\odot\cdots\odot T_m,\qquad T_{1:0}:=I, T m := N n ( V n , m ) ⊙ ⋯ ⊙ N 1 ( V 1 , m ) ⊙ P 1 ( U 1 , m ) ⊙ ⋯ ⊙ P n ( U n , m ) , T 1 : m := T 1 ⊙ ⋯ ⊙ T m , T 1 : 0 := I ,
where I I I is the max-plus identity, with diagonal entries zero and all other entries − ∞ -\infty − ∞ .
Theorem 2.12 (Finite line representation). Fix M ∈ N M\in\mathbb{N} M ∈ N and a ∈ Z a\in\mathbb{Z} a ∈ Z . For every deterministic array X : K M → N 0 X:\mathcal{K}_M\to\mathbb{N}_0 X : K M → N 0 , where
K M : = ⋃ 1 ≤ i ≤ k 0 ≤ m ≤ M { z ∈ Z 2 : s i a ≤ z ≤ v m M , z 1 ≥ z 2 } , \mathcal{K}_M:=\bigcup_{\substack{1\le i\le k\\0\le m\le M}}\{z\in\mathbb{Z}^2:s_i^a\le z\le v_m^M,\ z_1\ge z_2\}, K M := 1 ≤ i ≤ k 0 ≤ m ≤ M ⋃ { z ∈ Z 2 : s i a ≤ z ≤ v m M , z 1 ≥ z 2 } , and with X ( s i a ) = 0 X(s_i^a)=0 X ( s i a ) = 0 for every i i i , the finite rearrangement below produces arrays Z o u t , U , V Z^{\mathrm{out}},U,V Z out , U , V and constants κ i \kappa_i κ i , independent of m m m , satisfying, for all 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k and 0 ≤ m ≤ M 0\le m\le M 0 ≤ m ≤ M ,
κ i : = L X ( s i a ; v 0 M ) − C Z o u t ( i , 2 k ) (29) \kappa_i:=L_X(s_i^a;v_0^M)-C^{Z^{\mathrm{out}}}(i,2k) \tag*{(29)} κ i := L X ( s i a ; v 0 M ) − C Z out ( i , 2 k ) (29) L X ( s i a ; v m M ) = κ i + [ C Z o u t ⊙ T 1 : m ] ( i , 2 k ) . (30) L_X(s_i^a;v_m^M)=\kappa_i+[C^{Z^{\mathrm{out}}}\odot T_{1:m}](i,2k). \tag*{(30)} L X ( s i a ; v m M ) = κ i + [ C Z out ⊙ T 1 : m ] ( i , 2 k ) . (30) For the product-geometric array with source weights set to zero, all entries of Z o u t , U Z^{\mathrm{out}},U Z out , U , and V V V are independent, with
Z i , j o u t ∼ { Geo ( γ i / q j ) , i < j < i ˉ , Geo ( c / g i ) , j = i ˉ , 1 ≤ i ≤ k , i < j ≤ i ˉ , (31) Z_{i,j}^{\mathrm{out}}\sim
\begin{cases}
\operatorname{Geo}(\gamma_i/q_j), & i<j<\bar{i},\\
\operatorname{Geo}(c/g_i), & j=\bar{i},
\end{cases}
\qquad1\le i\le k,\quad i<j\le\bar{i}, \tag*{(31)} Z i , j out ∼ { Geo ( γ i / q j ) , Geo ( c / g i ) , i < j < i ˉ , j = i ˉ , 1 ≤ i ≤ k , i < j ≤ i ˉ , (31) V j , m ∼ Geo ( δ / γ j ) , U j , m ∼ Geo ( δ γ j ) , 1 ≤ j ≤ 2 k , 1 ≤ m ≤ M . (32) V_{j,m}\sim\operatorname{Geo}(\delta/\gamma_j),\qquad U_{j,m}\sim\operatorname{Geo}(\delta\gamma_j),\qquad1\le j\le2k,\quad1\le m\le M. \tag*{(32)} V j , m ∼ Geo ( δ / γ j ) , U j , m ∼ Geo ( δ γ j ) , 1 ≤ j ≤ 2 k , 1 ≤ m ≤ M . (32) Consequently, the geometric cut field has the joint representation
( H i , M δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) = d ( [ C Z o u t ⊙ T 1 : m ] ( i , 2 k ) − C Z o u t ( i , 2 k ) : i ≤ k , 0 ≤ m ≤ M ) . (33) \left(H_{i,M}^{\delta,a}(m):i\le k,\ 0\le m\le M\right)\overset{\mathrm{d}}{=}\left([C^{Z^{\mathrm{out}}}\odot T_{1:m}](i,2k)-C^{Z^{\mathrm{out}}}(i,2k):i\le k,\ 0\le m\le M\right). \tag*{(33)} ( H i , M δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) = d ( [ C Z out ⊙ T 1 : m ] ( i , 2 k ) − C Z out ( i , 2 k ) : i ≤ k , 0 ≤ m ≤ M ) . (33)
The finite rearrangement. We now give the deterministic construction used in the theorem. The bridge array is
Z i , j : = X ( a + i ˉ , a + j ) , 1 ≤ i ≤ k , i < j ≤ i ˉ . Z_{i,j}:=X(a+\bar{i},a+j),\qquad1\le i\le k,\quad i<j\le\bar{i}. Z i , j := X ( a + i ˉ , a + j ) , 1 ≤ i ≤ k , i < j ≤ i ˉ .
The indices ( i , j ) (i,j) ( i , j ) correspond to the marked line atom Z ( − i , j ) Z(-i,j) Z ( − i , j ) in the Poisson-line construction.
For the fixed window size M M M , put
p h : = a + 2 k + M − h + 1 , r m : = a + 2 k + M + m , 1 ≤ h , m ≤ M . p_h:=a+2k+M-h+1,\qquad r_m:=a+2k+M+m,\qquad1\le h,m\le M. p h := a + 2 k + M − h + 1 , r m := a + 2 k + M + m , 1 ≤ h , m ≤ M .
Thus v 0 M = ( p 1 , p 1 ) v_0^M=(p_1,p_1) v 0 M = ( p 1 , p 1 ) and v m M = ( r m , p m − 1 ) v_m^M=(r_m,p_m-1) v m M = ( r m , p m − 1 ) for 1 ≤ m ≤ M 1\le m\le M 1 ≤ m ≤ M . For 1 ≤ j ≤ 2 k 1\le j\le2k 1 ≤ j ≤ 2 k and 1 ≤ h , m ≤ M 1\le h,m\le M 1 ≤ h , m ≤ M , initialize
A j ( h ) : = X ( p h , a + j ) , D h : = X ( p h , p h ) , U j , m : = X ( r m , a + j ) . A_j^{(h)}:=X(p_h,a+j),\qquad D_h:=X(p_h,p_h),\qquad U_{j,m}:=X(r_m,a+j). A j ( h ) := X ( p h , a + j ) , D h := X ( p h , p h ) , U j , m := X ( r m , a + j ) .
Under the geometric input law, these variables, the entries of Z Z Z , and the remaining weights on K M \mathcal{K}_M K M are independent, with
Z i , j ∼ Geo ( γ j / g i ) ( i < j < i ˉ ) , Z i , i ˉ ∼ Geo ( c / g i ) , Z_{i,j}\sim\operatorname{Geo}(\gamma_j/g_i)\quad(i<j<\bar{i}),\qquad Z_{i,\bar{i}}\sim\operatorname{Geo}(c/g_i), Z i , j ∼ Geo ( γ j / g i ) ( i < j < i ˉ ) , Z i , i ˉ ∼ Geo ( c / g i ) ,
A j ( h ) , U j , m ∼ Geo ( δ γ j ) , D h ∼ Geo ( c δ ) , A_j^{(h)},U_{j,m}\sim\operatorname{Geo}(\delta\gamma_j),\qquad D_h\sim\operatorname{Geo}(c\delta), A j ( h ) , U j , m ∼ Geo ( δ γ j ) , D h ∼ Geo ( cδ ) ,
X ( p ℓ , p h ) , X ( r m , p s ) ∼ Geo ( δ 2 ) ( ℓ < h , m < s ) . X(p_\ell,p_h),X(r_m,p_s)\sim\operatorname{Geo}(\delta^2)\qquad(\ell<h,\ m<s). X ( p ℓ , p h ) , X ( r m , p s ) ∼ Geo ( δ 2 ) ( ℓ < h , m < s ) .
For u , v , w ∈ N 0 u,v,w\in\mathbb{N}_0 u , v , w ∈ N 0 , set
R ( u , v , w ) : = ( w + ( u − v ) + , w + ( v − u ) + , u ∧ v ) . (34) \mathcal{R}(u,v,w):=(w+(u-v)_+,w+(v-u)_+,u\wedge v). \tag*{(34)} R ( u , v , w ) := ( w + ( u − v ) + , w + ( v − u ) + , u ∧ v ) . (34)
This involution preserves u + w u+w u + w and v + w v+w v + w . Each assignment below replaces its three displayed variables by the components of R \mathcal{R} R . The working variables C C C and w i n t w_{\mathrm{int}} w int carry the active coordinates through the updates; they are not additional independent copies.
The construction has three stages: fold each background coordinate p h p_h p h through the bridge block, move its resulting negative factor past the remaining p p p -coordinates, and interleave the positive r r r -factors with these negative factors.
For h = M , M − 1 , … , 1 h=M,M-1,\ldots,1 h = M , M − 1 , … , 1 , apply the following updates. For i = 1 , … , k i=1,\ldots,k i = 1 , … , k , set C = A i ( h ) C=A_i^{(h)} C = A i ( h ) , update in increasing j j j , and then apply the last update:
( Z i , j , C , A j ( h ) ) ⟵ R ( Z i , j , C , A j ( h ) ) , j = i + 1 , … , i ˉ − 1 , (35) (Z_{i,j},C,A_j^{(h)})\longleftarrow\mathcal{R}(Z_{i,j},C,A_j^{(h)}),\qquad j=i+1,\ldots,\bar{i}-1, \tag*{(35)} ( Z i , j , C , A j ( h ) ) ⟵ R ( Z i , j , C , A j ( h ) ) , j = i + 1 , … , i ˉ − 1 , (35)
( Z i , i ˉ , C , D h ) ⟵ R ( Z i , i ˉ , C , D h ) , A i ( h ) ⟵ C . (36) (Z_{i,\bar{i}},C,D_h)\longleftarrow\mathcal{R}(Z_{i,\bar{i}},C,D_h),\qquad A_i^{(h)}\longleftarrow C. \tag*{(36)} ( Z i , i ˉ , C , D h ) ⟵ R ( Z i , i ˉ , C , D h ) , A i ( h ) ⟵ C . (36)
For i = k , k − 1 , … , 1 i=k,k-1,\ldots,1 i = k , k − 1 , … , 1 , start with C = A i ˉ ( h ) C=A_{\bar{i}}^{(h)} C = A i ˉ ( h ) , apply
( Z i , j , A j ( h ) , C ) ⟵ R ( Z i , j , A j ( h ) , C ) , j = i ˉ − 1 , … , i + 1 , (Z_{i,j},A_j^{(h)},C)\longleftarrow\mathcal{R}(Z_{i,j},A_j^{(h)},C),\qquad j=\bar{i}-1,\ldots,i+1, ( Z i , j , A j ( h ) , C ) ⟵ R ( Z i , j , A j ( h ) , C ) , j = i ˉ − 1 , … , i + 1 ,
in decreasing j j j , and put A i ˉ ( h ) ⟵ C A_{\bar{i}}^{(h)}\longleftarrow C A i ˉ ( h ) ⟵ C . Set
V j , h : = A j ( h ) , 1 ≤ j ≤ 2 k . (37) V_{j,h}:=A_j^{(h)},\qquad1\le j\le2k. \tag*{(37)} V j , h := A j ( h ) , 1 ≤ j ≤ 2 k . (37)
Next, for ℓ = h − 1 , h − 2 , … , 1 \ell=h-1,h-2,\ldots,1 ℓ = h − 1 , h − 2 , … , 1 , start with w i n t = X ( p ℓ , p h ) w_{\mathrm{int}}=X(p_\ell,p_h) w int = X ( p ℓ , p h ) and apply
( V j , h , A j ( ℓ ) , w i n t ) ⟵ R ( V j , h , A j ( ℓ ) , w i n t ) , j = 2 k , … , 1. (38) (V_{j,h},A_j^{(\ell)},w_{\mathrm{int}})\longleftarrow\mathcal{R}(V_{j,h},A_j^{(\ell)},w_{\mathrm{int}}),\qquad j=2k,\ldots,1. \tag*{(38)} ( V j , h , A j ( ℓ ) , w int ) ⟵ R ( V j , h , A j ( ℓ ) , w int ) , j = 2 k , … , 1. (38)
After processing every p h p_h p h , for m = 1 , … , M − 1 m=1,\ldots,M-1 m = 1 , … , M − 1 and s = M , M − 1 , … , m + 1 s=M,M-1,\ldots,m+1 s = M , M − 1 , … , m + 1 , start with w i n t = X ( r m , p s ) w_{\mathrm{int}}=X(r_m,p_s) w int = X ( r m , p s ) and apply
( U j , m , V j , s , w i n t ) ⟵ R ( U j , m , V j , s , w i n t ) , j = 2 k , … , 1. (39) (U_{j,m},V_{j,s},w_{\mathrm{int}})\longleftarrow\mathcal{R}(U_{j,m},V_{j,s},w_{\mathrm{int}}),\qquad j=2k,\ldots,1. \tag*{(39)} ( U j , m , V j , s , w int ) ⟵ R ( U j , m , V j , s , w int ) , j = 2 k , … , 1. (39)
Deterministic passage identities. We first record two matrix identities used to verify the rearrangement. Put n = 2 k n=2k n = 2 k and write
P ( a ) = P 1 ( a 1 ) ⊙ ⋯ ⊙ P n ( a n ) , N ( v ) = N n ( v n ) ⊙ ⋯ ⊙ N 1 ( v 1 ) . P(a)=P_1(a_1)\odot\cdots\odot P_n(a_n),\qquad N(v)=N_n(v_n)\odot\cdots\odot N_1(v_1). P ( a ) = P 1 ( a 1 ) ⊙ ⋯ ⊙ P n ( a n ) , N ( v ) = N n ( v n ) ⊙ ⋯ ⊙ N 1 ( v 1 ) .
For nonnegative vectors, their entries are
P ( a ) ( p , q ) = ∑ j = p q a j , N ( v ) ( p , q ) = { − v p , p = q , 0 , p < q . P(a)(p,q)=\sum_{j=p}^{q}a_j,\qquad
N(v)(p,q)=
\begin{cases}
-v_p,&p=q,\\
0,&p<q.
\end{cases} P ( a ) ( p , q ) = j = p ∑ q a j , N ( v ) ( p , q ) = { − v p , 0 , p = q , p < q .
As usual entries below the diagonal are − ∞ -\infty − ∞ .
First start with any c 0 ≥ 0 c_0\ge0 c 0 ≥ 0 and, in increasing order of j j j , set ( b j , c j , a j ′ ) = R ( z j , c j − 1 , a j ) (b_j,c_j,a'_j)=\mathcal{R}(z_j,c_{j-1},a_j) ( b j , c j , a j ′ ) = R ( z j , c j − 1 , a j ) . Then the following matrix identity holds:
P ( z ) ⊙ P ( a ) = P ( a ′ ) ⊙ P ( b ) . (40) P(z)\odot P(a)=P(a')\odot P(b). \tag*{(40)} P ( z ) ⊙ P ( a ) = P ( a ′ ) ⊙ P ( b ) . (40)
Fix p ≤ q p\le q p ≤ q , put c = c p − 1 c=c_{p-1} c = c p − 1 , and set
Z j = ∑ ℓ = p j z ℓ , A j = ∑ ℓ = p j a ℓ , h j = Z j − A j − 1 , H j = max p ≤ r ≤ j h r , D j = ( H j − c ) + , Z_j=\sum_{\ell=p}^{j}z_\ell,\qquad A_j=\sum_{\ell=p}^{j}a_\ell,\qquad h_j=Z_j-A_{j-1},\qquad H_j=\max_{p\le r\le j}h_r,\qquad D_j=(H_j-c)_+, Z j = ℓ = p ∑ j z ℓ , A j = ℓ = p ∑ j a ℓ , h j = Z j − A j − 1 , H j = p ≤ r ≤ j max h r , D j = ( H j − c ) + ,
with A p − 1 = D p − 1 = 0 A_{p-1}=D_{p-1}=0 A p − 1 = D p − 1 = 0 . The recurrence for c j c_j c j gives
∑ ℓ = p j a ℓ ′ = Z j − D j , ∑ ℓ = p j b ℓ = A j + D j . \sum_{\ell=p}^{j}a'_\ell=Z_j-D_j,\qquad\sum_{\ell=p}^{j}b_\ell=A_j+D_j. ℓ = p ∑ j a ℓ ′ = Z j − D j , ℓ = p ∑ j b ℓ = A j + D j .
The left entry in (2.30) is A q + H q A_q+H_q A q + H q and the right entry is
A q + D q + max p ≤ r ≤ q { h r − D r − D r − 1 } . A_q+D_q+\max_{p\le r\le q}\{h_r-D_r-D_{r-1}\}. A q + D q + p ≤ r ≤ q max { h r − D r − D r − 1 } .
The last maximum is min { H q , c } \min\{H_q,c\} min { H q , c } : before the first crossing of c c c all D D D ’s vanish, and at the first crossing the displayed expression is exactly c c c ; all subsequent expressions are at most c c c . This also covers the case with no crossing.
Second, start with any t n + 1 ≥ 0 t_{n+1}\ge0 t n + 1 ≥ 0 and, in decreasing order of j j j , set ( v j ′ , a j ′ , t j ) = R ( v j , a j , t j + 1 ) (v'_j,a'_j,t_j)=\mathcal{R}(v_j,a_j,t_{j+1}) ( v j ′ , a j ′ , t j ) = R ( v j , a j , t j + 1 ) . Then
N ( v ) ⊙ P ( a ) = P ( a ′ ) ⊙ N ( v ′ ) . (41) N(v)\mathbin{\odot}P(a)=P(a')\mathbin{\odot}N(v'). \tag*{(41)} N ( v ) ⊙ P ( a ) = P ( a ′ ) ⊙ N ( v ′ ) . (41)
The diagonal entries agree because a j ′ − v j ′ = a j − v j a'_j-v'_j=a_j-v_j a j ′ − v j ′ = a j − v j . For p < q p<q p < q the two entries are respectively
∑ j = p + 1 q a j + ( a p − v p ) + , ∑ j = p q − 1 a j ′ + ( a q ′ − v q ′ ) + . \sum_{j=p+1}^{q}a_j+(a_p-v_p)_+,\qquad
\sum_{j=p}^{q-1}a'_j+(a'_q-v'_q)_+. j = p + 1 ∑ q a j + ( a p − v p ) + , j = p ∑ q − 1 a j ′ + ( a q ′ − v q ′ ) + .
They agree by a j ′ = min { v j + 1 , a j + 1 } + ( a j − v j ) + a'_j=\min\{v_{j+1},a_{j+1}\}+(a_j-v_j)_+ a j ′ = min { v j + 1 , a j + 1 } + ( a j − v j ) + for j < n j<n j < n , and min { v j , a j } + ( a j − v j ) + = a j \min\{v_j,a_j\}+(a_j-v_j)_+=a_j min { v j , a j } + ( a j − v j ) + = a j . The initial value t n + 1 ≥ 0 t_{n+1}\ge0 t n + 1 ≥ 0 is arbitrary. To track a terminal at the intersection being eliminated, we also need an identity for the intermediate endpoint: writing w i n t = t n + 1 w_{\mathrm{int}}=t_{n+1} w int = t n + 1 , for any incoming row ℓ \ell ℓ ,
[ ℓ P ( a ′ ) ] n = w i n t + max { ℓ n , [ ℓ N ( v ) P ( a ) ] n } . (42) [\ell P(a')]_n=w_{\mathrm{int}}+\max\{\ell_n,[\ell N(v)P(a)]_n\}. \tag*{(42)} [ ℓ P ( a ′ ) ] n = w int + max { ℓ n , [ ℓ N ( v ) P ( a ) ] n } . (42)
Indeed, for p < n p<n p < n , P ( a ′ ) ( p , n ) = w i n t + [ N ( v ) P ( a ) ] ( p , n ) P(a')(p,n)=w_{\mathrm{int}}+[N(v)P(a)](p,n) P ( a ′ ) ( p , n ) = w int + [ N ( v ) P ( a )] ( p , n ) , while its diagonal entry is w i n t + ( a n − v n ) + w_{\mathrm{int}}+(a_n-v_n)_+ w int + ( a n − v n ) + .
Lemma 2.13 (Passage identities for the updates). Let n = 2 k n=2k n = 2 k and apply (2.25)–(2.27) to one input ( Z , A , D ) (Z,A,D) ( Z , A , D ) , with outputs ( Z ′ , V ) (Z',V) ( Z ′ , V ) ; A A A in the identities below denotes its initial value. Record α i \alpha_i α i as the value of A i A_i A i at the start of forward row i i i , and β i \beta_i β i as the value of C C C at the end of backward row $i (before the subsequent updates). Put χ i = α i + β i \chi_i=\alpha_i+\beta_i χ i = α i + β i . Then, simultaneously for every source i i i and every q ≥ i q\ge i q ≥ i ,
[ C Z ⊙ P ( A ) ] ( i , q ) = χ i + [ C Z ′ ⊙ N ( V ) ] ( i , q ) , (43) [C^Z\mathbin{\odot}P(A)](i,q)=\chi_i+[C^{Z'}\mathbin{\odot}N(V)](i,q), \tag*{(43)} [ C Z ⊙ P ( A )] ( i , q ) = χ i + [ C Z ′ ⊙ N ( V )] ( i , q ) , (43) D + [ C Z ⊙ P ( A ) ] ( i , n ) = χ i + C Z ′ ( i , n ) . (44) D+[C^Z\mathbin{\odot}P(A)](i,n)=\chi_i+C^{Z'}(i,n). \tag*{(44)} D + [ C Z ⊙ P ( A )] ( i , n ) = χ i + C Z ′ ( i , n ) . (44) In particular, χ i \chi_i χ i does not depend on q q q .
Proof. For k = 1 k=1 k = 1 , write Z 1 , 2 = z Z_{1,2}=z Z 1 , 2 = z , A = ( a 1 , a 2 ) A=(a_1,a_2) A = ( a 1 , a 2 ) , and D = d D=d D = d . The updates give
z ′ = d + ( z − a 1 ) + , V 1 = a 2 , V 2 = d + ( a 1 − z ) + , χ 1 = a 1 + a 2 . z'=d+(z-a_1)_+,\qquad V_1=a_2,\qquad V_2=d+(a_1-z)_+,\qquad\chi_1=a_1+a_2. z ′ = d + ( z − a 1 ) + , V 1 = a 2 , V 2 = d + ( a 1 − z ) + , χ 1 = a 1 + a 2 .
Since C Z = ( 0 , z ) C^Z=(0,z) C Z = ( 0 , z ) and C Z ′ = ( 0 , z ′ ) C^{Z'}=(0,z') C Z ′ = ( 0 , z ′ ) , the all-port identity at q = 1 q=1 q = 1 is a 1 = χ 1 − V 1 a_1=\chi_1-V_1 a 1 = χ 1 − V 1 , and at q = 2 q=2 q = 2 it is a 2 + max { a 1 , z } = χ 1 + ( z − a 1 ) + a_2+\max\{a_1,z\}=\chi_1+(z-a_1)_+ a 2 + max { a 1 , z } = χ 1 + ( z − a 1 ) + . Adding d d d gives the diagonal identity.
Assume k ≥ 2 k\ge2 k ≥ 2 and the result for k − 1 k-1 k − 1 . Write the first row of the bridge array as z = ( 0 , Z 1 , 2 , … , Z 1 , n ) z=(0,Z_{1,2},\ldots,Z_{1,n}) z = ( 0 , Z 1 , 2 , … , Z 1 , n ) . For a source i > 1 i>1 i > 1 , let ℓ i Z \ell_i^Z ℓ i Z be its passage vector in the bridge array restricted to lines 2 , … , n − 1 2,\ldots,n-1 2 , … , n − 1 , extended by ℓ i Z ( 1 ) = − ∞ \ell_i^Z(1)=-\infty ℓ i Z ( 1 ) = − ∞ and ℓ i Z ( n ) = ℓ i Z ( n − 1 ) \ell_i^Z(n)=\ell_i^Z(n-1) ℓ i Z ( n ) = ℓ i Z ( n − 1 ) . Then C Z ( i , ⋅ ) = ℓ i Z ⊙ P ( z ) C^Z(i,\cdot)=\ell_i^Z\mathbin{\odot}P(z) C Z ( i , ⋅ ) = ℓ i Z ⊙ P ( z ) . The first source has C Z ( 1 , ⋅ ) = e ⊙ P ( z ) C^Z(1,\cdot)=e\mathbin{\odot}P(z) C Z ( 1 , ⋅ ) = e ⊙ P ( z ) , where e = ( 0 , … , 0 ) e=(0,\ldots,0) e = ( 0 , … , 0 ) .
Apply the first forward updates through j = 2 , … , n − 1 j=2,\ldots,n-1 j = 2 , … , n − 1 , with initial value C = A 1 C=A_1 C = A 1 . Write its first and third output components as b j b_j b j and a j a_j a j , respectively, and let x x x be the final value of C C C . Complete these vectors by
a 1 = 0 , b 1 = A 1 , d = a n = min { Z 1 , n , x } , b n = A n + ( Z 1 , n − x ) + . a_1=0,\qquad b_1=A_1,\qquad d=a_n=\min\{Z_{1,n},x\},\qquad b_n=A_n+(Z_{1,n}-x)_+. a 1 = 0 , b 1 = A 1 , d = a n = min { Z 1 , n , x } , b n = A n + ( Z 1 , n − x ) + .
For the positive-row identity, initialize its running variable at zero. Its first update is R ( 0 , 0 , A 1 ) = ( A 1 , A 1 , 0 ) \mathcal{R}(0,0,A_1)=(A_1,A_1,0) R ( 0 , 0 , A 1 ) = ( A 1 , A 1 , 0 ) , so the remaining updates start at C = A 1 C=A_1 C = A 1 and give a 1 = 0 a_1=0 a 1 = 0 , b 1 = A 1 b_1=A_1 b 1 = A 1 . Thus (2.30) gives P ( z ) P ( A ) = P ( a ) P ( b ) P(z)P(A)=P(a)P(b) P ( z ) P ( A ) = P ( a ) P ( b ) . The actual last update gives
Z 1 , n ′ = D + ( Z 1 , n − x ) + , V n = D + ( x − Z 1 , n ) + , Z'_{1,n}=D+(Z_{1,n}-x)_+,\qquad V_n=D+(x-Z_{1,n})_+, Z 1 , n ′ = D + ( Z 1 , n − x ) + , V n = D + ( x − Z 1 , n ) + ,
and leaves d d d as the diagonal input for the inner array.
Apply the induction hypothesis to the inner bridge array, the vector ( a 2 , … , a n − 1 ) (a_2,\ldots,a_{n-1}) ( a 2 , … , a n − 1 ) , and diagonal d d d . Denote its updated passage vector by C ‾ i \overline{C}_i C i and the weights of its negative atoms by w 2 , … , w n − 1 w_2,\ldots,w_{n-1} w 2 , … , w n − 1 . Extend w 1 = A 1 w_1=A_1 w 1 = A 1 , w n = A n + ( x − Z 1 , n ) + w_n=A_n+(x-Z_{1,n})_+ w n = A n + ( x − Z 1 , n ) + , and, for each i > 1 i>1 i > 1 , set C ‾ i ( 1 ) = − ∞ \overline{C}_i(1)=-\infty C i ( 1 ) = − ∞ , C ‾ i ( n ) = C ‾ i ( n − 1 ) + A n \overline{C}_i(n)=\overline{C}_i(n-1)+A_n C i ( n ) = C i ( n − 1 ) + A n . The two induction identities give the vector equality
ℓ i Z ⊙ P ( a ) = χ i + C ‾ i ⊙ N ( w ) . \ell_i^Z\mathbin{\odot}P(a)=\chi_i+\overline{C}_i\mathbin{\odot}N(w). ℓ i Z ⊙ P ( a ) = χ i + C i ⊙ N ( w ) .
Indeed, C ‾ i \overline{C}_i C i is nondecreasing on its feasible coordinates up to n − 1 n-1 n − 1 , and w n ≥ A n w_n\ge A_n w n ≥ A n , so
[ C ‾ i N ( w ) ] ( n ) = max { C ‾ i ( n − 1 ) , C ‾ i ( n ) − w n } = C ‾ i ( n − 1 ) . [\overline{C}_iN(w)](n)=\max\{\overline{C}_i(n-1),\overline{C}_i(n)-w_n\}=\overline{C}_i(n-1). [ C i N ( w )] ( n ) = max { C i ( n − 1 ) , C i ( n ) − w n } = C i ( n − 1 ) .
The last left coordinate is d + [ ℓ i Z P ( z ) ] ( n − 1 ) d+[\ell_i^ZP(z)](n-1) d + [ ℓ i Z P ( z )] ( n − 1 ) , which agrees by the inner diagonal identity. The other coordinates are the interface induction hypothesis for q < n q<n q < n .
Now commute N ( w ) P ( b ) N(w)P(b) N ( w ) P ( b ) using (2.31), with initial value t n + 1 = D t_{n+1}=D t n + 1 = D . The n n n -th update produces exactly ( V n , Z 1 , n ′ ) (V_n,Z'_{1,n}) ( V n , Z 1 , n ′ ) and leaves A n A_n A n as the next initial value. The updates for j = n − 1 , … , 2 j=n-1,\ldots,2 j = n − 1 , … , 2 are exactly the first backward row. At j = 1 j=1 j = 1 , since w 1 = b 1 = A 1 w_1=b_1=A_1 w 1 = b 1 = A 1 , both output components equal V V V . Thus the resulting vector of positive atom weights is b ′ = ( V 1 , Z 1 , 2 ′ , … , Z 1 , n ′ ) b'=(V_1,Z'_{1,2},\ldots,Z'_{1,n}) b ′ = ( V 1 , Z 1 , 2 ′ , … , Z 1 , n ′ ) , and the negative vector is V V V . For i > 1 i>1 i > 1 , C ‾ i P ( b ′ ) = C Z ′ ( i , ⋅ ) \overline{C}_iP(b')=C^{Z'}(i,\cdot) C i P ( b ′ ) = C Z ′ ( i , ⋅ ) . Before the last coordinate this is immediate; at the last, the entrance value C ‾ i ( n − 1 ) + A n \overline{C}_i(n-1)+A_n C i ( n − 1 ) + A n is dominated by the route through line n − 1 n-1 n − 1 , because the first backward update has Z 1 , n − 1 ′ ≥ A n Z'_{1,n-1}\ge A_n Z 1 , n − 1 ′ ≥ A n . This proves (2.33) for all inner sources.
For the first source, put s j = ∑ ℓ = 2 j Z 1 , ℓ s_j=\sum_{\ell=2}^j Z_{1,\ell} s j = ∑ ℓ = 2 j Z 1 , ℓ and H j = [ C Z P ( A ) ] ( 1 , j ) H_j=[C^ZP(A)](1,j) H j = [ C Z P ( A )] ( 1 , j ) . The scalar path recurrence is H j = A j + max { H j − 1 , s j } H_j=A_j+\max\{H_{j-1},s_j\} H j = A j + max { H j − 1 , s j } , and its auxiliary variable is c j = H j − s j c_j=H_j-s_j c j = H j − s j . Thus H 1 = A 1 H_1=A_1 H 1 = A 1 and H j − H j − 1 = A j + ( Z 1 , j − c j − 1 ) + = b j H_j-H_{j-1}=A_j+(Z_{1,j}-c_{j-1})_+=b_j H j − H j − 1 = A j + ( Z 1 , j − c j − 1 ) + = b j , giving
[ C Z P ( A ) ] ( 1 , q ) = A 1 + ∑ j = 2 q b j = A 1 + [ e N ( w ) P ( b ) ] ( q ) . [C^ZP(A)](1,q)=A_1+\sum_{j=2}^{q}b_j=A_1+[eN(w)P(b)](q). [ C Z P ( A )] ( 1 , q ) = A 1 + j = 2 ∑ q b j = A 1 + [ e N ( w ) P ( b )] ( q ) .
Here e N ( w ) = ( − A 1 , 0 , … , 0 ) \mathrm{e}_{N}(w)=(-A_{1},0,\ldots,0) e N ( w ) = ( − A 1 , 0 , … , 0 ) . After the same interchange, e P ( b ′ ) = V 1 + C Z ′ ( 1 , ⋅ ) \mathrm{e}_{P}(b')=V_{1}+C^{Z'}(1,\cdot) e P ( b ′ ) = V 1 + C Z ′ ( 1 , ⋅ ) , proving the first-row identity with χ 1 = A 1 + V 1 \chi_{1}=A_{1}+V_{1} χ 1 = A 1 + V 1 . The constants for the inner rows are unchanged on adding lines 1 1 1 and n n n , so they are precisely the recorded α i + β i \alpha_{i}+\beta_{i} α i + β i .
Finally every row of the bridge array satisfies C Z ′ ( i , n ) = C Z ′ ( i , n − 1 ) + Z 1 , n ′ C^{Z'}(i,n)=C^{Z'}(i,n-1)+Z'_{1,n} C Z ′ ( i , n ) = C Z ′ ( i , n − 1 ) + Z 1 , n ′ , since the only atom on the last line of the bridge environment is Z 1 , n ′ Z'_{1,n} Z 1 , n ′ . Therefore
[ C Z ′ N ( V ) ] ( i , n ) = C Z ′ ( i , n ) − min { V n , Z 1 , n ′ } = C Z ′ ( i , n ) − D , [C^{Z'}N(V)](i,n)=C^{Z'}(i,n)-\min\{V_{n},Z'_{1,n}\}=C^{Z'}(i,n)-D, [ C Z ′ N ( V )] ( i , n ) = C Z ′ ( i , n ) − min { V n , Z 1 , n ′ } = C Z ′ ( i , n ) − D ,
which proves (2.34). □
Proof of Theorem 2.12. Step 1: passage preservation. We maintain the following invariant: at every stage, all source-to-interface passage values and all retained terminal values are preserved up to one additive constant for each source. The folding identities introduce source constants; the matrix interchanges preserve the passage vectors, and the intermediate-endpoint identity tracks a terminal at an eliminated intersection. Max-plus multiplication carries an additive source constant unchanged. In the original grid the passage matrix for the special block is C Z C^{Z} C Z : its row at label ı ˉ \bar{\imath} ˉ has exactly the atoms Z i , j Z_{i,j} Z i , j , and its source cell is zero. An ordinary row with special-block weights a a a contributes the matrix P ( a ) P(a) P ( a ) . The scalar recurrence is H j = a j + max { H j − 1 , ℓ j i n } H_{j}=a_{j}+\max\{H_{j-1},\ell_{j}^{\mathrm{in}}\} H j = a j + max { H j − 1 , ℓ j in } , where ℓ j i n \ell_{j}^{\mathrm{in}} ℓ j in is the incoming passage value at line j j j ; hence these identities hold simultaneously for every source.
The initial background order is p M , … , p 1 , r 1 , … , r M p_{M},\ldots,p_{1},r_{1},\ldots,r_{M} p M , … , p 1 , r 1 , … , r M . Eliminate the rows indexed by p h p_{h} p h in the order h = M , … , 1 h=M,\ldots,1 h = M , … , 1 . Lemma 2.13 replaces the block consisting of the special coordinates and p h p_{h} p h by the updated bridge array followed by N ( V ⋅ , h ) N(V_{\cdot,h}) N ( V ⋅ , h ) , with one constant added to each source row. Its second identity gives the same constants at the diagonal endpoint; thus the constants are the same for every retained terminal. Move this negative factor past the remaining positive rows p h − 1 , … , p 1 p_{h-1},\ldots,p_{1} p h − 1 , … , p 1 using (2.31), initialized at their intersection w i n t = X ( p ℓ , p h ) w_{\mathrm{int}}=X(p_{\ell},p_{h}) w int = X ( p ℓ , p h ) . The updates in decreasing j j j are exactly (2.28). In particular each interchange preserves every exit coordinate of every source row, including the coordinates needed at later interchanges. Descending induction in h h h gives the negative-factor order p 1 , … , p M p_{1},\ldots,p_{M} p 1 , … , p M . The constants from successive folds accumulate into one constant κ i \kappa_{i} κ i for each source.
Next move the positive factor for r m r_{m} r m left past p M , … , p m + 1 p_{M},\ldots,p_{m+1} p M , … , p m + 1 , for m = 1 , … , M − 1 m=1,\ldots,M-1 m = 1 , … , M − 1 . The same identity applies with a = U ⋅ , m a=U_{\cdot,m} a = U ⋅ , m , v = V ⋅ , s v=V_{\cdot,s} v = V ⋅ , s , and initial value X ( r m , p s ) X(r_{m},p_{s}) X ( r m , p s ) . Swapping the first two inputs of R \mathcal{R} R swaps its first two outputs, so the order of variables is precisely (2.29). The final factor order is
p 1 , r 1 , p 2 , r 2 , … , p M , r M . (45) p_{1},r_{1},p_{2},r_{2},\ldots,p_{M},r_{M}. \tag*{(45)} p 1 , r 1 , p 2 , r 2 , … , p M , r M . (45)
All original grid weights are included in this composition. The bridge array contains the weights from the special block. The updates for p h p_{h} p h use its intersections with that block and its diagonal weight; the first interchanges use the ( p ℓ , p h ) (p_{\ell},p_{h}) ( p ℓ , p h ) cells, and U j , m U_{j,m} U j , m gives the intersections of r m r_{m} r m with the special block. A mixed cell is in the cone of terminal ℓ \ell ℓ exactly when ( r m , p s ) ⪯ v ℓ V (r_{m},p_{s})\preceq v_{\ell}^{\mathrm{V}} ( r m , p s ) ⪯ v ℓ V , m ≤ ℓ < s m\leq\ell<s m ≤ ℓ < s . These are exactly the second interchanges. No cell ( r m , r s ) (r_{m},r_{s}) ( r m , r s ) is accessible. Thus this composition is also a direct elimination of the grid Bellman recurrences: the forward and backward identities eliminate one block consisting of the special coordinates and p h p_{h} p h , and the row-interchange identity eliminates one remaining intersection with its displayed initial value. Each background intersection is eliminated once and every retained source remains. At an intersection of weight w i n t w_{\mathrm{int}} w int , (2.32) is the Bellman recurrence for the endpoint between the exchanged factors. It therefore gives the passage values at retained terminals there, as well as those after both factors.
Step 2: terminal identification and pinning. The terminal v 0 M v_{0}^{M} v 0 M is the end of the updated bridge array, by the diagonal identity (2.34) at each step. Terminal v m M v_{m}^{M} v m M is the end of the first m m m pairs in (2.35), since its background labels are ( r m , p m − 1 ) (r_{m},p_{m}-1) ( r m , p m − 1 ) . The factors indexed by p m p_{m} p m and r m r_{m} r m are N ( V ⋅ , m ) N(V_{\cdot,m}) N ( V ⋅ , m ) and P ( U ⋅ , m ) P(U_{\cdot,m}) P ( U ⋅ , m ) , respectively; their product is T m T_{m} T m . All preceding replacements preserve every passage-matrix entry, so they give, with the same row constants for every m m m ,
L X ( s i a ; v m M ) = κ i + [ C Z o u t ⊙ T 1 : m ] ( i , 2 k ) . L_{X}(s_{i}^{a};v_{m}^{M})=\kappa_{i}+[C^{Z^{\mathrm{out}}}\odot T_{1:m}](i,2k). L X ( s i a ; v m M ) = κ i + [ C Z out ⊙ T 1 : m ] ( i , 2 k ) .
The case m = 0 m=0 m = 0 determines (2.19).
Step 3: the geometric product law. Write ( u ′ , v ′ , w ′ ) = R ( u , v , w ) (u',v',w')=\mathcal{R}(u,v,w) ( u ′ , v ′ , w ′ ) = R ( u , v , w ) . For independent inputs with respective laws Geo ( r ) \operatorname{Geo}(r) Geo ( r ) , Geo ( s ) \operatorname{Geo}(s) Geo ( s ) , and Geo ( r s ) \operatorname{Geo}(rs) Geo ( r s ) , where 0 ≤ r , s < 1 0\leq r,s<1 0 ≤ r , s < 1 , the outputs have the same product law. For r , s > 0 r,s>0 r , s > 0 , use bijectivity and
r u s v ( r s ) w = r u + w s v + w = r u ′ + w ′ s v ′ + w ′ = r u ′ s v ′ ( r s ) w ′ ; r^{u}s^{v}(rs)^{w}=r^{u+w}s^{v+w}=r^{u'+w'}s^{v'+w'}=r^{u'}s^{v'}(rs)^{w'}; r u s v ( r s ) w = r u + w s v + w = r u ′ + w ′ s v ′ + w ′ = r u ′ s v ′ ( r s ) w ′ ;
the cases r = 0 r=0 r = 0 or s = 0 s=0 s = 0 follow directly from (2.24).
The forward, final, and backward updates use parameter triples ( r , s , r s ) (r,s,rs) ( r , s , r s ) :
( γ j q j , δ j , δ γ j ) , ( c γ j , δ j , c δ ) , ( γ j q j , δ γ j , δ q j ) , \left(\frac{\gamma_{j}}{q_{j}},\delta_{j},\delta\gamma_{j}\right),\qquad
\left(\frac{c}{\gamma_{j}},\delta_{j},c\delta\right),\qquad
\left(\frac{\gamma_{j}}{q_{j}},\frac{\delta}{\gamma_{j}},\frac{\delta}{q_{j}}\right), ( q j γ j , δ j , δ γ j ) , ( γ j c , δ j , cδ ) , ( q j γ j , γ j δ , q j δ ) ,
while the two sequences of background updates use ( δ / γ j , δ γ j , δ 2 ) (\delta/\gamma_{j},\delta\gamma_{j},\delta^{2}) ( δ / γ j , δ γ j , δ 2 ) or the same first two parameters in reverse order. At each step, the three active variables have the displayed product law and are independent of all untouched active variables. The involution therefore preserves the product law of the entire active array. Induction through the finite sequence of updates proves mutual independence of all retained outputs and (2.21)–(2.22). Combining the deterministic pinned identity with the source-cancellation limit in Lemma 2.9 gives (2.23). □
Staggered atoms and mesh approximation. We first realize the ordered product T m T_{m} T m exactly by signed line atoms. Take the bridge array and families ( U j , m , V j , m ) j = 1 2 k (U_{j,m},V_{j,m})_{j=1}^{2k} ( U j , m , V j , m ) j = 1 2 k , m ≥ 1 m\geq1 m ≥ 1 , independently with the laws (2.21)–(2.22). For m ≥ 1 m\geq1 m ≥ 1 , place the atoms with index m m m at
x m , j − : = ( m − 1 ) δ + 2 k + 1 − j 4 k + 1 δ , x m , j + : = ( m − 1 ) δ + 2 k + j 4 k + 1 δ . x^{-}_{m,j}:=(m-1)\delta+\frac{2k+1-j}{4k+1}\delta,\qquad x^{+}_{m,j}:=(m-1)\delta+\frac{2k+j}{4k+1}\delta. x m , j − := ( m − 1 ) δ + 4 k + 1 2 k + 1 − j δ , x m , j + := ( m − 1 ) δ + 4 k + 1 2 k + j δ .
Their left-to-right order is − V 2 k , m , … , − V 1 , m , U 1 , m , … , U 2 k , m -V_{2k,m},\ldots,-V_{1,m},U_{1,m},\ldots,U_{2k,m} − V 2 k , m , … , − V 1 , m , U 1 , m , … , U 2 k , m . Set
d S j δ : = − ∑ m ≥ 1 V j , m δ x m , j − + ∑ m ≥ 1 U j , m δ x m , j + , \mathrm{d}S^{\delta}_{j}:=-\sum_{m\geq1}V_{j,m}\delta_{x^{-}_{m,j}}+\sum_{m\geq1}U_{j,m}\delta_{x^{+}_{m,j}}, d S j δ := − m ≥ 1 ∑ V j , m δ x m , j − + m ≥ 1 ∑ U j , m δ x m , j + ,
normalize S ^ j δ ( 0 ) = 0 \widehat{S}^{\delta}_{j}(0)=0 S j δ ( 0 ) = 0 , retain the bridge atoms at − k , … , − 1 -k,\ldots,-1 − k , … , − 1 , and call the resulting line environment E ^ δ \widehat{E}^{\delta} E δ . Define
G ^ i δ ( x ) : = L E ^ δ ( s i ; ( x , 2 k ) ) , h ^ i δ ( x ) : = G ^ i δ ( x ) − G ^ i δ ( 0 ) . \widehat{G}^{\delta}_{i}(x):=\mathcal{L}_{\widehat{E}^{\delta}}(s_i;(x,2k)),\qquad\widehat{h}^{\delta}_{i}(x):=\widehat{G}^{\delta}_{i}(x)-\widehat{G}^{\delta}_{i}(0). G i δ ( x ) := L E δ ( s i ; ( x , 2 k )) , h i δ ( x ) := G i δ ( x ) − G i δ ( 0 ) .
Theorem 2.12 gives
( H i , M δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) = d ( h ^ i δ ( δ m ) : i ≤ k , 0 ≤ m ≤ M ) . (46) \left(H^{\delta,a}_{i,M}(m):i\leq k,\ 0\leq m\leq M\right)\mathrel{\overset{\mathrm{d}}{=}}\left(\widehat{h}^{\delta}_{i}(\delta m):i\leq k,\ 0\leq m\leq M\right). \tag*{(46)} ( H i , M δ , a ( m ) : i ≤ k , 0 ≤ m ≤ M ) = d ( h i δ ( δ m ) : i ≤ k , 0 ≤ m ≤ M ) . (46)
Next move all atoms in ( ( m − 1 ) δ , m δ ) ((m-1)\delta,m\delta) (( m − 1 ) δ , m δ ) to m δ m\delta m δ . Keep U U U and V V V as separate marked inputs; the resulting càdlàg line has their net increment U j , m − V j , m U_{j,m}-V_{j,m} U j , m − V j , m at each mesh point. With the same bridge array and the resulting environment E ~ δ \widetilde{E}^{\delta} E δ , set
S ~ j δ ( x ) : = ∑ 1 ≤ m ≤ ⌊ x / δ ⌋ ( U j , m − V j , m ) , G ~ i δ ( x ) : = L E ~ δ ( s i ; ( x , 2 k ) ) , h ~ i δ ( x ) : = G ~ i δ ( x ) − G ~ i δ ( 0 ) . \widetilde{S}^{\delta}_{j}(x):=\sum_{1\leq m\leq\lfloor x/\delta\rfloor}(U_{j,m}-V_{j,m}),\qquad\widetilde{G}^{\delta}_{i}(x):=\mathcal{L}_{\widetilde{E}^{\delta}}(s_i;(x,2k)),\qquad\widetilde{h}^{\delta}_{i}(x):=\widetilde{G}^{\delta}_{i}(x)-\widetilde{G}^{\delta}_{i}(0). S j δ ( x ) := 1 ≤ m ≤ ⌊ x / δ ⌋ ∑ ( U j , m − V j , m ) , G i δ ( x ) := L E δ ( s i ; ( x , 2 k )) , h i δ ( x ) := G i δ ( x ) − G i δ ( 0 ) .
This collapse need not preserve passage values when several atoms occur in one interval. For example, with two lines, U 2 , m = V 2 , m = 1 U_{2,m}=V_{2,m}=1 U 2 , m = V 2 , m = 1 and all other increments zero, a path from line 1 to line 2 can avoid the staggered negative atom and collect the positive atom, obtaining weight 1; after collapse the net jump is zero. The lemma below controls all such exceptional intervals.
Figure 6 shows the mesh collapse and the subsequent Poisson limit.
Figure 6. From geometric packets to Poisson lines, illustrated for k = 2 k=2 k = 2 . The exact packet order is − V 4 , m , … , − V 1 , m , U 1 , m , … , U 4 , m -V_{4,m},\ldots,-V_{1,m},U_{1,m},\ldots,U_{4,m} − V 4 , m , … , − V 1 , m , U 1 , m , … , U 4 , m . Moving atoms to m δ m\delta m δ retains their line and sign; equality of mesh passage values holds outside the exceptional cells of Lemma 2.14. The limiting positive and negative atoms are independent Poisson processes. The geometric bridge array is retained throughout.
Lemma 2.14 (Approximation on the mesh). For every 0 < L < ∞ 0<L<\infty 0 < L < ∞ , with M δ ( L ) : = ⌈ L / δ ⌉ + 1 M_{\delta}(L):=\lceil L/\delta\rceil+1 M δ ( L ) := ⌈ L / δ ⌉ + 1 ,
P ( h ^ i δ ( δ m ) ≠ h ~ i δ ( δ m ) for some i ≤ k , 0 ≤ m ≤ M δ ( L ) ) ≤ C g , L δ . \mathbb{P}\left(\widehat{h}^{\delta}_{i}(\delta m)\ne\widetilde{h}^{\delta}_{i}(\delta m)\text{ for some }i\leq k,\ 0\leq m\leq M_{\delta}(L)\right)\leq C_{g,L}\delta. P ( h i δ ( δ m ) = h i δ ( δ m ) for some i ≤ k , 0 ≤ m ≤ M δ ( L ) ) ≤ C g , L δ . Proof. Call the m m m -th interval exceptional if B m : = { ∑ j = 1 2 k ( U j , m + V j , m ) ≥ 2 } B_m:=\left\{\sum_{j=1}^{2k}(U_{j,m}+V_{j,m})\geq2\right\} B m := { ∑ j = 1 2 k ( U j , m + V j , m ) ≥ 2 } occurs. Otherwise its passage value from line p p p to line q q q , p ≤ q p\leq q p ≤ q , is
Moving these atoms to the right endpoint preserves this value: for a negative atom and p < q p<q p < q , a switch avoids it, including at the endpoint under the one-sided convention. Composition preserves all source-to-terminal passage values through successive nonexceptional intervals, before and after pinning.
Let Λ g : = ∑ j = 1 2 k ( γ j + γ j − 1 ) \Lambda_g:=\sum_{j=1}^{2k}(\gamma_j+\gamma_j^{-1}) Λ g := ∑ j = 1 2 k ( γ j + γ j − 1 ) . Since P ( Geo ( q ) ≥ 1 ) = q \mathbb{P}(\operatorname{Geo}(q)\geq1)=q P ( Geo ( q ) ≥ 1 ) = q and P ( Geo ( q ) ≥ 2 ) = q 2 \mathbb{P}(\operatorname{Geo}(q)\geq2)=q^2 P ( Geo ( q ) ≥ 2 ) = q 2 , independence and a union bound give
P ( a fixed interval is exceptional ) ≤ δ 2 [ ∑ j = 1 2 k ( γ j 2 + γ j − 2 ) + 1 2 Λ g 2 ] = : C g ′ δ 2 . \mathbb{P}(\text{a fixed interval is exceptional})\leq\delta^2\left[\sum_{j=1}^{2k}(\gamma_j^2+\gamma_j^{-2})+\frac{1}{2}\Lambda_g^2\right]=:C'_g\delta^2. P ( a fixed interval is exceptional ) ≤ δ 2 [ j = 1 ∑ 2 k ( γ j 2 + γ j − 2 ) + 2 1 Λ g 2 ] =: C g ′ δ 2 . A final union bound over the first M δ ( L ) M_{\delta}(L) M δ ( L ) intervals yields C g ( L + 2 ) δ C_g(L+2)\delta C g ( L + 2 ) δ for 0 < δ ≤ 1 0<\delta\leq1 0 < δ ≤ 1 , proving the claim. □ \square □
For a sequence f : N 0 → R f:\mathbb{N}_0\to\mathbb{R} f : N 0 → R , let I δ f \mathcal{I}_{\delta}f I δ f be the upper-semicontinuous regularization of the function equal to f ( m ) f(m) f ( m ) on [ δ m , δ ( m + 1 ) ) [\delta m,\delta(m+1)) [ δ m , δ ( m + 1 )) . For a function on δ N 0 \delta\mathbb{N}_0 δ N 0 , use the same notation after identifying its value at δ m \delta m δ m with the sequence entry at m m m ; apply I δ \mathcal{I}_{\delta} I δ componentwise to vectors of profiles. The collapsed fields are constant on these intervals, and therefore
I δ ( h ~ i ∣ δ N 0 δ ) = usc h ~ i δ \mathcal{I}_{\delta}(\widetilde{h}^{\delta}_{i|\delta N_0})=\operatorname{usc}\widetilde{h}^{\delta}_{i} I δ ( h i ∣ δ N 0 δ ) = usc h i δ . Using the same infinite family ( Z , U , V ) (Z,U,V) ( Z , U , V ) , construct the coupling by setting H i δ ( m ) : = h ~ i δ ( δ m ) H^{\delta}_{i}(m):=\widetilde{h}^{\delta}_{i}(\delta m) H i δ ( m ) := h i δ ( δ m ) for m ∈ N 0 m\in\mathbb{N}_0 m ∈ N 0 . The exact readout identity for every M M M identifies the entire sequence law as H δ ∼ μ c ; g δ H^{\delta}\sim\mu^{\delta}_{c;g} H δ ∼ μ c ; g δ . On the event that the first M δ ( L ) M_{\delta}(L) M δ ( L ) intervals are nonexceptional, the mesh values agree with those of the collapsed profiles. The extra mesh interval also covers the adjacent plateau values used by upper-semicontinuous regularization at L L L . Hence
P ( I δ H i δ = usc h ~ i δ on [ 0 , L ] for every i ≤ k ) ≥ 1 − C g , L δ . (47) \mathbb{P}\left(\mathcal{I}_{\delta}H^{\delta}_{i}=\operatorname{usc}\widetilde{h}^{\delta}_{i}\text{ on }[0,L]\text{ for every }i\le k\right)\ge1-C_{g,L}\delta. \tag*{(47)} P ( I δ H i δ = usc h i δ on [ 0 , L ] for every i ≤ k ) ≥ 1 − C g , L δ . (47)
Passage to the common-noise PNG evolution
We transfer joint stationarity from the geometric model to PNG by passing the initial profiles and their common-noise evolution to the limit together. Throughout, c > 0 c>0 c > 0 and g 1 > ⋯ > g k > c ∨ 1 g_1>\cdots>g_k>c\vee1 g 1 > ⋯ > g k > c ∨ 1 . Fix a deterministic t ≥ 0 t\ge0 t ≥ 0 and choose n δ ∈ N 0 n_{\delta}\in\mathbb{N}_0 n δ ∈ N 0 with t δ : = n δ δ → t t_{\delta}:=n_{\delta}\delta\to t t δ := n δ δ → t . Let H δ ∼ μ c ; g δ H^{\delta}\sim\mu^{\delta}_{c;g} H δ ∼ μ c ; g δ and write R = ( R i ( c ; g ) ) i = 1 k R=(R_i^{(c;g)})_{i=1}^{k} R = ( R i ( c ; g ) ) i = 1 k for the Poisson-line candidate.
For each time step, sample a fresh noise array ξ r \xi_r ξ r , independently of the input and the other time steps, and use that same array for all k k k coordinates. We use the shorthand
( Ψ δ ( c ) , [ k ] ) n : = Ψ δ , ξ n ( c ) , [ k ] ∘ ⋯ ∘ Ψ δ , ξ 1 ( c ) , [ k ] , \left(\Psi_{\delta}^{(c),[k]}\right)_n:=\Psi_{\delta,\xi_n}^{(c),[k]}\circ\cdots\circ\Psi_{\delta,\xi_1}^{(c),[k]}, ( Ψ δ ( c ) , [ k ] ) n := Ψ δ , ξ n ( c ) , [ k ] ∘ ⋯ ∘ Ψ δ , ξ 1 ( c ) , [ k ] ,
with the analogous convention for one coordinate and the identity map when n = 0 n=0 n = 0 . Our main claim is
( I δ H δ , I δ ( Ψ δ , n δ ( c ) , [ k ] H δ ) ) ⟹ ( R , Θ t ( c ) , [ k ] ( R ; Π ) ) , (48) \left(\mathcal{I}_{\delta}H^{\delta},\mathcal{I}_{\delta}\left(\Psi_{\delta,n_{\delta}}^{(c),[k]}H^{\delta}\right)\right)\Longrightarrow\left(R,\Theta_t^{(c),[k]}(R;\Pi)\right), \tag*{(48)} ( I δ H δ , I δ ( Ψ δ , n δ ( c ) , [ k ] H δ ) ) ⟹ ( R , Θ t ( c ) , [ k ] ( R ; Π ) ) , (48)
in the product local-hypograph topology. Here Π \Pi Π is independent of R R R and is shared by every coordinate. We first prove convergence of the initial profiles, then of the independent future environment, and finally of the initial and evolved profiles jointly, obtaining (48) . We then transfer stationarity and parameter-deletion consistency to PNG.
Convergence of the initial profiles. Recall the independent line increments from (32) , and set
N j , + δ : = ∑ m ≥ 1 U j , m δ m δ , N j , − δ : = ∑ m ≥ 1 V j , m δ m δ , 1 ≤ j ≤ 2 k . \mathcal{N}^{\delta}_{j,+}:=\sum_{m\ge1}U_{j,m}\delta_{m\delta},\qquad
\mathcal{N}^{\delta}_{j,-}:=\sum_{m\ge1}V_{j,m}\delta_{m\delta},\qquad
1\le j\le2k. N j , + δ := m ≥ 1 ∑ U j , m δ m δ , N j , − δ := m ≥ 1 ∑ V j , m δ m δ , 1 ≤ j ≤ 2 k .
For Y ∼ Geo ( q ) Y\sim\operatorname{Geo}(q) Y ∼ Geo ( q ) , as q ↓ 0 q\downarrow0 q ↓ 0 ,
log E e − u Y = log 1 − q 1 − q e − u = − q ( 1 − e − u ) + O ( q 2 ) , (49) \log\mathbb{E}e^{-uY}=\log\frac{1-q}{1-qe^{-u}}=-q(1-e^{-u})+O(q^2), \tag*{(49)} log E e − u Y = log 1 − q e − u 1 − q = − q ( 1 − e − u ) + O ( q 2 ) , (49)
uniformly for bounded u ≥ 0 u\ge0 u ≥ 0 . Apply this expansion with q = δ γ j q=\delta\gamma_j q = δ γ j and q = δ / γ j q=\delta/\gamma_j q = δ / γ j , insert a nonnegative continuous compactly supported test function in place of u u u , and sum over the mesh sites. The resulting Riemann sums give joint vague convergence
( N j , + δ , N j , − δ ) j = 1 2 k ⟹ ( N j , + , N j , − ) j = 1 2 k , rate ( N j , + ) = γ j , rate ( N j , − ) = γ j − 1 . \left(\mathcal{N}^{\delta}_{j,+},\mathcal{N}^{\delta}_{j,-}\right)_{j=1}^{2k}
\Longrightarrow
\left(\mathcal{N}_{j,+},\mathcal{N}_{j,-}\right)_{j=1}^{2k},
\qquad
\operatorname{rate}(\mathcal{N}_{j,+})=\gamma_j,\qquad
\operatorname{rate}(\mathcal{N}_{j,-})=\gamma_j^{-1}. ( N j , + δ , N j , − δ ) j = 1 2 k ⟹ ( N j , + , N j , − ) j = 1 2 k , rate ( N j , + ) = γ j , rate ( N j , − ) = γ j − 1 .
The limiting point measures are independent. The finite bridge array is independent of these measures and retains the laws (15) . We keep the two signs separate: convergence of their signed difference alone could conceal cancellations at coincident prelimit sites.
Lemma 2.15 (Continuity of the line profiles). Fix 0 < L < K < ∞ 0<L<K<\infty 0 < L < K < ∞ and finitely many sources. Suppose that the finite bridge arrays converge coordinatewise and that the positive and negative counting measures on every line converge separately and vaguely on [ 0 , K ] [0,K] [ 0 , K ] . If the limiting measures consist of finitely many unit atoms, with distinct locations across all lines and signs and no atom at 0 , L , K 0,L,K 0 , L , K , then the associated pinned last-passage profiles converge jointly on [ 0 , L ] [0,L] [ 0 , L ] in compact hypograph topology.
Proof. Integer-valued marked convergence gives an eventual matching of all ordinary atoms, preserving their line, sign, and spatial order. Choose increasing piecewise affine homeomorphisms ϕ ℓ \phi_{\ell} ϕ ℓ , equal to the identity outside ( 0 , K ) (0,K) ( 0 , K ) and fixing L L L , which map each limiting atom to its matched prelimit atom. Then ∥ ϕ ℓ − id ∥ [ 0 , K ] → 0 \lVert\phi_{\ell}-\operatorname{id}\rVert_{[0,K]}\to0 ∥ ϕ ℓ − id ∥ [ 0 , K ] → 0 . These maps preserve switching patterns and atom ownership. If b ℓ b_{\ell} b ℓ is the sum of the absolute bridge-weight errors, corresponding path weights differ by at most b ℓ b_{\ell} b ℓ . Maximization and subtraction of the value at zero therefore give
max 0 ≤ i ≤ k sup 0 ≤ x ≤ L ∣ R ~ i , ℓ ( ϕ ℓ ( x ) ) − R ~ i ( x ) ∣ ≤ 2 b ℓ → 0. \max_{0\le i\le k}\sup_{0\le x\le L}\left|\widetilde{R}_{i,\ell}(\phi_{\ell}(x))-\widetilde{R}_{i}(x)\right|\le2b_{\ell}\to0. 0 ≤ i ≤ k max 0 ≤ x ≤ L sup R i , ℓ ( ϕ ℓ ( x )) − R i ( x ) ≤ 2 b ℓ → 0. where the tildes denote the pinned càdlàg profiles. Upper-semicontinuous regularization commutes with ϕ ℓ \phi_{\ell} ϕ ℓ , which proves the claim.
Poisson configurations satisfy these hypotheses almost surely. The continuous mapping theorem and the coupling estimate (47) give
( usc h ~ i δ ) i = 1 k ⟹ R , (50) \left(\operatorname{usc}\widetilde{h}^{\delta}_{i}\right)_{i=1}^{k}\Longrightarrow R, \tag*{(50)} ( usc h i δ ) i = 1 k ⟹ R , (50)
I δ H δ ⟹ R , H δ ∼ μ c ; g δ (51) \mathcal{I}_{\delta}H^{\delta}\Longrightarrow R,\qquad H^{\delta}\sim\mu^{\delta}_{c;g} \tag*{(51)} I δ H δ ⟹ R , H δ ∼ μ c ; g δ (51)
locally in product hypograph topology.
Convergence of the fresh future environment. Sample all fresh update arrays independently of the entire initial line coupling. The time-zero cut belongs to the past; the next 2 n δ 2n_\delta 2 n δ lattice layers form the half-open future slab ( 0 , t δ ] (0,t_\delta] ( 0 , t δ ] . Under the embedding (2.12), push one representative of each off-diagonal reflection orbit, with its multiplicity, to its physical point. Do the same for the diagonal weights, obtaining Π b u l k δ \Pi^\delta_{\mathrm{bulk}} Π bulk δ and Π ∂ δ \Pi^\delta_{\partial} Π ∂ δ .
The bulk lattice has covolume δ 2 / 2 \delta^2/2 δ 2 /2 and geometric parameter δ 2 \delta^2 δ 2 ; the boundary lattice has time spacing δ \delta δ and geometric parameter c δ c\delta cδ . Thus (2.39), with q = δ 2 q=\delta^2 q = δ 2 and q = c δ q=c\delta q = cδ , gives
Π b u l k δ ⟹ Π b u l k , Π b u l k ∼ PPP ( 2 d s d z ) ∣ ( 0 , t ] × ( 0 , ∞ ) , \Pi^\delta_{\mathrm{bulk}} \Longrightarrow\Pi_{\mathrm{bulk}}, \qquad\Pi_{\mathrm{bulk}} \sim\operatorname{PPP}(2\,\mathrm{d}s\,\mathrm{d}z)\big|_{(0,t]\times(0,\infty)}, Π bulk δ ⟹ Π bulk , Π bulk ∼ PPP ( 2 d s d z ) ( 0 , t ] × ( 0 , ∞ ) ,
Π ∂ δ ⟹ Π ∂ , Π ∂ ∼ PPP ( c d s ) ∣ ( 0 , t ] . \Pi^\delta_{\partial} \Longrightarrow\Pi_{\partial}, \qquad\Pi_{\partial} \sim\operatorname{PPP}(c\,\mathrm{d}s)\big|_{(0,t]}. Π ∂ δ ⟹ Π ∂ , Π ∂ ∼ PPP ( c d s ) ( 0 , t ] .
Here PPP ( μ ) \operatorname{PPP}(\mu) PPP ( μ ) denotes a Poisson point measure of intensity μ \mu μ . The limits are independent of one another and of the initial line network. On a compact slab, the probability of any weight exceeding one is O ( δ 2 ) O(\delta^2) O ( δ 2 ) in the bulk and O ( δ ) O(\delta) O ( δ ) at the boundary. Embed the moving slabs in a fixed compact envelope, including the time faces, and extend their measures by zero. The limiting measures have no atoms at times 0 , t 0,t 0 , t , so the changing terminal time causes no restriction discontinuity.
Convergence of the evolution. We first express the discrete dynamics through last passage in the fresh slab. Let L δ , n ( c ) ( ℓ , m ) L^{(c)}_{\delta,n}(\ell,m) L δ , n ( c ) ( ℓ , m ) be its passage weight from
( N + ℓ , N − ℓ ) to ( N + n + m , N + n − m ) (N+\ell,N-\ell) \quad\text{to} \quad(N+n+m,N+n-m) ( N + ℓ , N − ℓ ) to ( N + n + m , N + n − m )
in p ≥ q p\ge q p ≥ q , where the initial cut is at level N N N . Count only weights strictly after the initial cut, including the terminal weight. Set the value to − ∞ -\infty − ∞ when ∣ ℓ − m ∣ > n |\ell-m|>n ∣ ℓ − m ∣ > n ; for n = 0 n=0 n = 0 , set it to zero if ℓ = m \ell=m ℓ = m and − ∞ -\infty − ∞ otherwise. For pinned f : N 0 → Z f:\mathbb{N}_0\to\mathbb{Z} f : N 0 → Z , define
( T δ , n ( c ) f ) ( m ) : = max ∣ ℓ − m ∣ ≤ n { f ( ℓ ) + L δ , n ( c ) ( ℓ , m ) } − max 0 ≤ ℓ ≤ n { f ( ℓ ) + L δ , n ( c ) ( ℓ , 0 ) } . (T^{(c)}_{\delta,n}f)(m):=\max_{\lvert\ell-m\rvert\le n}\{f(\ell)+L^{(c)}_{\delta,n}(\ell,m)\}-\max_{0\le\ell\le n}\{f(\ell)+L^{(c)}_{\delta,n}(\ell,0)\}. ( T δ , n ( c ) f ) ( m ) := ∣ ℓ − m ∣ ≤ n max { f ( ℓ ) + L δ , n ( c ) ( ℓ , m )} − 0 ≤ ℓ ≤ n max { f ( ℓ ) + L δ , n ( c ) ( ℓ , 0 )} .
The Bellman recurrences (2.14)–(2.15) and passage composition give the pathwise identity
T δ , n ( c ) f = ( Ψ δ ( c ) ) n f . (52) T^{(c)}_{\delta,n}f=(\Psi^{(c)}_\delta)^n f. \tag*{(52)} T δ , n ( c ) f = ( Ψ δ ( c ) ) n f . (52)
The initial heights are supplied by f f f , so initial-cut weights are not counted again. Intermediate recentering constants cancel because max-plus updates commute with addition of constants.
The coupling from the preceding subsection also identifies the raw mesh heights. For K < ∞ K<\infty K < ∞ , let
E δ , K : = { H i δ ( m ) = h ~ i δ ( m δ ) for all i ≤ k , 0 ≤ m ≤ ⌈ K / δ ⌉ + 1 } . \mathcal{E}_{\delta,K}:=\left\{H_i^\delta(m)=\widetilde h_i^\delta(m\delta)\text{ for all }i\le k,\ 0\le m\le\lceil K/\delta\rceil+1\right\}. E δ , K := { H i δ ( m ) = h i δ ( m δ ) for all i ≤ k , 0 ≤ m ≤ ⌈ K / δ ⌉ + 1 } .
The mesh-approximation lemma gives
P ( E δ , K ) ≥ 1 − C g , K δ . (53) \mathbb{P}(\mathcal{E}_{\delta,K})\ge1-C_{g,K}\delta. \tag*{(53)} P ( E δ , K ) ≥ 1 − C g , K δ . (53)
This raw equality is needed because the lattice update acts on mesh heights, whereas upper-semicontinuous regularization can change their values at downward jumps. If K > L + t + 1 K>L+t+1 K > L + t + 1 , finite propagation and (2.42) give identical interpolated outputs on [ 0 , L ] [0,L] [ 0 , L ] for these two inputs, on E δ , K \mathcal{E}_{\delta,K} E δ , K and for all sufficiently small δ \delta δ , when they use the same future slab.
We next state the continuity property used for these line inputs. For a future point p = ( s , z ) p=(s,z) p = ( s , z ) , put u ( p ) = s + z u(p)=s+z u ( p ) = s + z and v ( p ) = s − z v(p)=s-z v ( p ) = s − z . Let A ⊂ ( 0 , K ) A\subset(0,K) A ⊂ ( 0 , K ) be the set of ordinary incoming atom locations on every line P P P in P \mathcal{P} P , and let P P P be the finite future configuration in the compact envelope. Call ( A , P ) (A,P) ( A , P ) generic if the configurations are simple, P P P has no atom at times 0 0 0 or t t t , distinct future atoms have distinct u u u - and distinct v v v -coordinates, and the random entries in the critical sets
C 0 : = A ∪ { u ( p ) , − v ( p ) : p ∈ P } ∪ { 0 , K } , \mathcal{C}_0:=A\cup\{u(p),-v(p):p\in P\}\cup\{0,K\}, C 0 := A ∪ { u ( p ) , − v ( p ) : p ∈ P } ∪ { 0 , K } ,
C t : = { a − t , a + t : a ∈ A } ∪ { u ( p ) − t , t − v ( p ) : p ∈ P } ∪ { 0 , L } \mathcal{C}_t:=\{a-t,a+t:a\in A\}\cup\{u(p)-t,t-v(p):p\in P\}\cup\{0,L\} C t := { a − t , a + t : a ∈ A } ∪ { u ( p ) − t , t − v ( p ) : p ∈ P } ∪ { 0 , L }
do not coincide with one another or with a deterministic endpoint, apart from forced repetitions arising from a single atom or t = 0 t=0 t = 0 . Only entries in the corresponding spatial window are relevant. These conditions stabilize causal order, separate the feasible initial interval endpoints from incoming jumps, and prevent an output breakpoint at zero or L L L . Independent Poisson inputs are almost surely generic for each fixed deterministic t , L , K t,L,K t , L , K .
Lemma 2.16 (Continuity of the evolution on compact intervals). Fix L > 0 L>0 L > 0 , t ≥ 0 t\ge0 t ≥ 0 , and K > L + t K>L+t K > L + t . Suppose that the bridge arrays and separately marked incoming line measures converge on [ 0 , K ] [0,K] [ 0 , K ] as in Lemma 2.15. Suppose also that the future point measures in ( 0 , t ℓ ) (0,t_\ell) ( 0 , t ℓ ) , t ℓ → t t_\ell\to t t ℓ → t , including their multiplicities and bulk or boundary marks, converge vaguely on a fixed compact envelope containing the relevant backward cones and its time faces. If the limiting configuration is generic, the pinned initial and evolved profiles converge jointly on [ 0 , L ] [0,L] [ 0 , L ] in compact hypograph topology.
In the lattice case, use raw mesh inputs f i , ℓ ( m ) = R ~ i , ℓ ( m δ ℓ ) f_{i,\ell}(m)=\widetilde R_{i,\ell}(m\delta_\ell) f i , ℓ ( m ) = R i , ℓ ( m δ ℓ ) and outputs I δ ℓ T δ ℓ , n ℓ ( c ) f i , ℓ \mathcal{I}_{\delta_\ell}T^{(c)}_{\delta_\ell,n_\ell}f_{i,\ell} I δ ℓ T δ ℓ , n ℓ ( c ) f i , ℓ , where δ ℓ ↓ 0 \delta_\ell\downarrow0 δ ℓ ↓ 0 and n ℓ δ ℓ = t ℓ n_\ell\delta_\ell=t_\ell n ℓ δ ℓ = t ℓ . In the continuum case, use the half-space PNG max-plus evolution.
Proof. Initial-profile convergence follows from the preceding continuity lemma. First assume t > 0 t>0 t > 0 . There are finitely many causal chains in the limiting future. For each such chain C = ( p 1 ≺ ⋯ ≺ p r ) C=(p_1\prec\cdots\prec p_r) C = ( p 1 ≺ ⋯ ≺ p r ) , with p j = ( s j , z j ) p_j=(s_j,z_j) p j = ( s j , z j ) , define
b i , C : = r + sup ( z 1 − s 1 ) + ≤ y ≤ z 1 + s 1 R i ( y ) , I C : = [ ( z r − ( t − s r ) ) + , z r + t − s r ] . b_{i,C}:=r+\sup_{(z_1-s_1)+\le y\le z_1+s_1}R_i(y), \qquad I_C:=\left[(z_r-(t-s_r))_+,z_r+t-s_r\right]. b i , C := r + ( z 1 − s 1 ) + ≤ y ≤ z 1 + s 1 sup R i ( y ) , I C := [ ( z r − ( t − s r ) ) + , z r + t − s r ] .
The unpinned output is the maximum of these constant-height contributions b i , C b_{i,C} b i , C on their intervals I C I_C I C , together with the no-future-atom contribution sup y ≥ 0 , ∣ y − x ∣ ≤ t R i ( y ) \sup_{y\ge0,\lvert y-x\rvert\le t}R_i(y) sup y ≥ 0 , ∣ y − x ∣ ≤ t R i ( y ) . Marked convergence matches all future atoms. Their strict null-coordinate comparisons are stable, so the collection of feasible chains eventually agrees with that of the limit. Genericity of C 0 \mathcal{C}_0 C 0 separates each feasible starting-interval endpoint from the incoming jumps. Thus the corresponding initial suprema converge; in the lattice case, the mesh captures every relevant plateau because its limiting length is positive. Consequently, the chain heights converge to b i , C b_{i,C} b i , C and their support endpoints converge to those of I C I_C I C . The no-atom contribution is likewise a finite maximum over incoming plateaus, whose heights and expanded interval endpoints converge. Taking the finite maximum gives hypograph convergence of the unpinned outputs, jointly for all sources. At interval endpoints, upper-semicontinuous regularization selects the larger adjacent value.
Genericity of C t \mathcal{C}_t C t makes the limiting output constant in a neighborhood of zero. The matched finite description then also gives convergence of its value at zero. Subtracting this value therefore preserves convergence and proves the pinned assertion.
If t = 0 t=0 t = 0 , the limiting future measure is empty. Since convergence is on a compact envelope including time zero, integer-valued atom matching eliminates all prelimit future atoms for large ℓ \ell ℓ . The remaining evolution is a dilation of radius t ℓ → 0 t_\ell\to0 t ℓ → 0 , with a mesh restriction in the lattice case. The matched plateau description proves hypograph convergence, and the initial atom-free neighborhood of zero gives convergence after pinning. □
Joint marked convergence of the incoming lines, bridges, and the independent future, followed by this lemma and (2.43), proves (2.38) on [ 0 , L ] [0,L] [ 0 , L ] . For local convergence, take a deterministic increasing exhaustion with K > L + t + 1 K>L+t+1 K > L + t + 1 at each stage. Its endpoints are simultaneously noncritical almost surely, so the claim holds in product local-hypograph topology.
Stationarity and parameter consistency. Lemma 2.10, applied successively with fresh independent slabs, gives
( Ψ δ ( c ) , [ k ] n H δ ) = d H δ , n ∈ N 0 . \left(\Psi_{\delta}^{(c),[k]}{}_n H^\delta\right)\mathrel{\stackrel{\mathrm d}{=}}H^\delta,\qquad n\in\mathbb{N}_0. ( Ψ δ ( c ) , [ k ] n H δ ) = d H δ , n ∈ N 0 .
Hence the two marginals on the left of (2.38) have the same law. Their limits must agree, proving
Θ t ( c ) , [ k ] ( R ; Π ) = d R , t ≥ 0. (54) \Theta_t^{(c),[k]}(R;\Pi)\mathrel{\stackrel{\mathrm d}{=}}R,\qquad t\ge0. \tag*{(54)} Θ t ( c ) , [ k ] ( R ; Π ) = d R , t ≥ 0. (54)
under (2.10). The argument applies to each fixed deterministic t t t ; convergence of complete time trajectories is not needed for this invariance.
Finally, for k ≥ 2 k\ge2 k ≥ 2 and 1 ≤ m ≤ k 1\le m\le k 1 ≤ m ≤ k , coordinate deletion is continuous in the product topology and commutes with I δ \mathcal{I}_\delta I δ . Applying (2.41) to the two sides of (2.17) gives
( R i ( c ; g ) : i ≠ m ) = d ( R j ( c ; g m ^ ) : 1 ≤ j ≤ k − 1 ) , (55) \left(R_i^{(c;g)}:i\ne m\right)\mathrel{\stackrel{\mathrm d}{=}}\left(R_j^{(c;\widehat{g_m})}:1\le j\le k-1\right), \tag*{(55)} ( R i ( c ; g ) : i = m ) = d ( R j ( c ; g m ) : 1 ≤ j ≤ k − 1 ) , (55)
with the retained coordinates listed in their original order. Moreover, for every input and every realization of the common future,
D m Θ t ( c ) , [ k ] ( h ; Π ) = Θ t ( c ) , [ k − 1 ] [ D m h ; Π ] . D_m\Theta_t^{(c),[k]}(h;\Pi)=\Theta_t^{(c),[k-1]}[D_mh;\Pi]. D m Θ t ( c ) , [ k ] ( h ; Π ) = Θ t ( c ) , [ k − 1 ] [ D m h ; Π ] .
Thus deletion also preserves the retained common-noise evolutions.
2.2.5. Endpoints and completion of the proof. We extend joint stationarity and deletion consistency from c > 0 c>0 c > 0 , g 1 > ⋯ > g k > c ∨ 1 g_1>\cdots>g_k>c\vee1 g 1 > ⋯ > g k > c ∨ 1 to all admissible parameters. The argument couples the initial profiles and their common future so that they agree locally with their limits with probability tending to one. Proper bridge limits are treated by direct coupling; at g k = c ≥ 1 g_k=c\ge1 g k = c ≥ 1 , the diverging terminal bridge is removed by pinning and the source shift of Lemma 2.8.
Lemma 2.17 (Parameter limits with a common future). Let c n > 0 c_n>0 c n > 0 , c n → c c_n\to c c n → c , and g ( n ) → g g^{(n)}\to g g ( n ) → g , where g 1 ( n ) > ⋯ > g k ( n ) > c n ∨ 1 g_1^{(n)}>\cdots>g_k^{(n)}>c_n\vee1 g 1 ( n ) > ⋯ > g k ( n ) > c n ∨ 1 , g k ≥ c g_k\ge c g k ≥ c . Assume either
(i) g k > c g_k>c g k > c ; or
(ii) g k = c ≥ 1 g_k=c\ge1 g k = c ≥ 1 , c n = c c_n=c c n = c , g i ( n ) = g i g_i^{(n)}=g_i g i ( n ) = g i for i < k i<k i < k , and g k ( n ) ↓ c g_k^{(n)}\downarrow c g k ( n ) ↓ c .
Set R ( n ) = ( R i ( c n ; g ( n ) ) ) i = 1 k R^{(n)}=\left(R_i^{(c_n;g^{(n)})}\right)_{i=1}^k R ( n ) = ( R i ( c n ; g ( n ) ) ) i = 1 k and R = ( R i ( c ; g ) ) i = 1 k R=\left(R_i^{(c;g)}\right)_{i=1}^k R = ( R i ( c ; g ) ) i = 1 k . In case (ii), define the last limiting profile using the shifted source in (2.8). There is a coupling of these initial profiles and future environments Π n , Π \Pi_n,\Pi Π n , Π , with boundary rates c n , c c_n,c c n , c , such that each future is independent of its initial vector and shared by all its coordinates, and, for each fixed deterministic t ≥ 0 t\ge0 t ≥ 0 ,
( R ( n ) , Θ t ( c n ) , [ k ] ( R ( n ) ; Π n ) ) → P ( R , Θ t ( c ) , [ k ] ( R ; Π ) ) (56) \left(R^{(n)},\Theta_t^{(c_n),[k]}(R^{(n)};\Pi_n)\right)\xrightarrow{\mathbb{P}} \left(R,\Theta_t^{(c),[k]}(R;\Pi)\right) \tag*{(56)} ( R ( n ) , Θ t ( c n ) , [ k ] ( R ( n ) ; Π n ) ) P ( R , Θ t ( c ) , [ k ] ( R ; Π ) ) (56) in the product local-hypograph topology.
Proof. Regular limits. In case (i), every limiting bridge parameter is below one. For q n → q < 1 q_n\to q<1 q n → q < 1 , the masses ( 1 − q n ) q n m (1-q_n)q_n^m ( 1 − q n ) q n m converge to ( 1 − q ) q m (1-q)q^m ( 1 − q ) q m , giving
d T V ( Geo ( q n ) , Geo ( q ) ) ⟶ 0. d_{\mathrm{TV}}\bigl(\operatorname{Geo}(q_n),\operatorname{Geo}(q)\bigr)\longrightarrow0. d TV ( Geo ( q n ) , Geo ( q ) ) ⟶ 0.
Independent maximal couplings therefore make the finite bridge arrays agree with probability tending to one. Couple each incoming Poisson clock by a common rate-cutoff measure, using independent coupling randomness for distinct clocks and bridges: for rates a n → a a_n\to a a n → a , its mismatch probability on [ 0 , K ] [0,K] [ 0 , K ] is 1 − e − K ∣ a n − a ∣ ≤ K ∣ a n − a ∣ 1-e^{-K|a_n-a|}\le K|a_n-a| 1 − e − K ∣ a n − a ∣ ≤ K ∣ a n − a ∣ . Since all line rates converge, their configurations on [ 0 , K ] [0,K] [ 0 , K ] also agree with probability tending to one.
The singular limit. In case (ii), apply the same coupling to the retained bridges and line clocks. The remaining bridge Y n = Z n ( − k , k + 1 ) Y_n=Z_n(-k,k+1) Y n = Z n ( − k , k + 1 ) diverges in probability. Lemma 2.8 shows that, with probability tending to one, all maximizing paths from the last source collect Y n Y_n Y n throughout a fixed compact window. Pinning cancels this common contribution, leaving the shifted-source profile. The other sources cannot access Y n Y_n Y n . Thus in both cases the entire pinned input vector agrees with its limit on [ 0 , K ] [0,K] [ 0 , K ] with probability tending to one.
The common future. Use identical bulk futures and couple the boundary clocks by common rate cutoffs. Their mismatch probability up to time t t t is at most t ∣ c n − c ∣ t|c_n-c| t ∣ c n − c ∣ . Choose the randomness for this future coupling independently of the entire incoming coupling. For K > L + t + 1 K>L+t+1 K > L + t + 1 , finite propagation then gives identical pinned outputs on [ 0 , L ] [0,L] [ 0 , L ] whenever the inputs agree on [ 0 , K ] [0,K] [ 0 , K ] and the future configurations agree in the relevant backward cone. Both events have probability tending to one. This proves the joint convergence on every compact interval and hence in the product local-hypograph topology. □
Approximating the endpoints. For a distinct ordered admissible vector, all cases outside the strict interior are covered by the following choices, with ε n ↓ 0 \varepsilon_n\downarrow0 ε n ↓ 0 and 0 < ε n < 1 0<\varepsilon_n<1 0 < ε n < 1 :
If c = 0 c=0 c = 0 , take c n = ε n c_n=\varepsilon_n c n = ε n and g i ( n ) = g i + ε n g_i^{(n)}=g_i+\varepsilon_n g i ( n ) = g i + ε n for every i i i . This includes g k = 1 g_k=1 g k = 1 and gives proper limiting bridges.
If 0 < c < 1 0<c<1 0 < c < 1 and g k = 1 g_k=1 g k = 1 , fix c n = c c_n=c c n = c and the first k − 1 k-1 k − 1 parameters, and take g k ( n ) = 1 + ε n g_k^{(n)}=1+\varepsilon_n g k ( n ) = 1 + ε n . The last terminal bridge converges to Geo ( c ) \operatorname{Geo}(c) Geo ( c ) .
If c ≥ 1 c\ge1 c ≥ 1 and g k = c g_k=c g k = c , fix c n = c c_n=c c n = c and the first k − 1 k-1 k − 1 parameters, and take g k ( n ) = c + ε n g_k^{(n)}=c+\varepsilon_n g k ( n ) = c + ε n . The limit is the shifted-source construction; in particular, c = 1 c=1 c = 1 , g k = 1 g_k=1 g k = 1 belongs to this singular case.
For all sufficiently large n n n , these vectors remain strictly ordered and satisfy (2.10).
Proof of Theorem 1.7. Distinct ordered parameters. In the strict interior, joint stationarity and deletion consistency are (2.44) and (2.45). For an endpoint vector, choose an approximation above. For each fixed deterministic t t t , the two components of the approximating input–output pair have the same law. Passing to their limits in (2.46) gives Θ t ( c ) , [ k ] ( R ; Π ) = d R \Theta_t^{(c),[k]}(R;\Pi)\mathrel{\stackrel{\mathrm d}{=}}R Θ t ( c ) , [ k ] ( R ; Π ) = d R . For k ≥ 2 k\ge2 k ≥ 2 and 1 ≤ m ≤ k 1\le m\le k 1 ≤ m ≤ k , obtain the reduced approximating vector by deleting the same coordinate from g ( n ) g^{(n)} g ( n ) . The parameter-limit lemma applies to both the full and reduced vectors. At a singular endpoint, deleting m < k m<k m < k retains the same singular approximation in the smaller dimension, while deleting m = k m=k m = k leaves a fixed strictly interior vector. Coordinate deletion is continuous, so the prelimit deletion identity passes to
( R i ( c ; g ) : i ≠ m ) = d ( R j ( c ; g m ^ ) : 1 ≤ j ≤ k − 1 ) , (57) \left(R_i^{(c;g)}:i\ne m\right)\mathrel{\stackrel{\mathrm d}{=}}\left(R_j^{(c;\widehat{g^m})}:1\le j\le k-1\right), \tag*{(57)} ( R i ( c ; g ) : i = m ) = d ( R j ( c ; g m ) : 1 ≤ j ≤ k − 1 ) , (57)
with the retained coordinates in their original order.
Arbitrary parameter lists. Construct the joint law at the distinct values in decreasing order, then permute and duplicate coordinates according to the requested list. These operations commute with the common-noise evolution; in particular, identical copies remain identical under the same future. Joint stationarity follows. For deletion from a list of length at least two, if the parameter still occurs elsewhere, deletion removes an identical copy. If its occurrence is unique, apply (2.47) to the distinct-parameter family. This proves deletion consistency for all lists. Permutation and repetition consistency hold by construction.
One-profile marginals. Iterated deletion reduces each coordinate to its one-parameter law. For g > c g>c g > c , Lemma 2.7 identifies this law as ν c , g \nu_{c,g} ν c , g . At g = c ≥ 1 g=c\ge1 g = c ≥ 1 , (2.9) gives the same identification through the shifted source (0, 2). This proves all the assertions of the theorem. □
Half-space PNG horizon
We now realize the jointly stationary laws ν c ; g 1 , … , g k \nu_{c;g_1,\ldots,g_k} ν c ; g 1 , … , g k as one measurable field indexed by g ∈ G c g\in\mathcal{G}_c g ∈ G c . The field is determined by its profile at r c = c ∨ 1 r_c=c\vee1 r c = c ∨ 1 and a positive measure recording its parameter increments. Planar crossing gives positivity of the measure, while right continuity in the parameter selects a canonical realization. We then prove uniqueness in law of this regular representation and its stationarity, as asserted in Theorem 1.9.
Right-continuous representatives and planar crossing
For h ∈ X s t e p , 0 h\in\mathcal{X}_{\mathrm{step},0} h ∈ X step , 0 without coincident bulk pairs or a boundary mark at zero, define its right-continuous representative by
h → ( x ) : = lim y ↓ x y > x h ( y ) , h = usc h → , ( usc f ) ( x ) = { f ( 0 ) , x = 0 , max { f ( x − ) , f ( x ) } , x > 0. (58) h^{\to}(x):=\lim_{\substack{y\downarrow x\\y>x}}h(y),\qquad h=\operatorname{usc}h^{\to},\qquad(\operatorname{usc}f)(x)=
\begin{cases}
f(0), & x=0,\\
\max\{f(x-),f(x)\}, & x>0.
\end{cases} \tag*{(58)} h → ( x ) := y ↓ x y > x lim h ( y ) , h = usc h → , ( usc f ) ( x ) = { f ( 0 ) , max { f ( x − ) , f ( x )} , x = 0 , x > 0. (58)
Every finite-parameter law is supported on such profiles. Step coordinates (Definition 2.1) make c ( h ) : = h → \mathbf{c}(h):=h^{\to} c ( h ) := h → a Borel map, extended by the zero profile on the excluded set. For a field F F F , set
Δ □ F ( g 1 , g 2 ; x 1 , x 2 ) : = F ( g 1 , x 1 ) + F ( g 2 , x 2 ) − F ( g 1 , x 2 ) − F ( g 2 , x 1 ) . \Delta_{\Box}F(g_1,g_2;x_1,x_2):=F(g_1,x_1)+F(g_2,x_2)-F(g_1,x_2)-F(g_2,x_1). Δ □ F ( g 1 , g 2 ; x 1 , x 2 ) := F ( g 1 , x 1 ) + F ( g 2 , x 2 ) − F ( g 1 , x 2 ) − F ( g 2 , x 1 ) .
Lemma 2.18 (Planar crossing). For r c ≤ a < b r_c\le a<b r c ≤ a < b and ( R a , R b ) ∼ ν c ; a , b (R_a,R_b)\sim\nu_{c;a,b} ( R a , R b ) ∼ ν c ; a , b , almost surely
R a → ( x 1 ) + R b → ( x 2 ) − R a → ( x 2 ) − R b → ( x 1 ) ≥ 0 , 0 ≤ x 1 < x 2 . R_a^{\to}(x_1)+R_b^{\to}(x_2)-R_a^{\to}(x_2)-R_b^{\to}(x_1)\ge0,\qquad0\le x_1<x_2. R a → ( x 1 ) + R b → ( x 2 ) − R a → ( x 2 ) − R b → ( x 1 ) ≥ 0 , 0 ≤ x 1 < x 2 .
Thus R b → − R a → R_b^{\to}-R_a^{\to} R b → − R a → is nonnegative and nondecreasing, and R a ≤ R b R_a\le R_b R a ≤ R b .
Proof. First take a > c a>c a > c . The descending construction ( b , a ) (b,a) ( b , a ) has sources s 1 = ( − 1 , 1 ) s_1=(-1,1) s 1 = ( − 1 , 1 ) , s 2 = ( − 2 , 2 ) s_2=(-2,2) s 2 = ( − 2 , 2 ) , terminal line 4, and uncentered last-passage profiles G b , G a G_b,G_a G b , G a . Maximizers P b : s 1 → ( x 1 , 4 ) P_b:s_1\to(x_1,4) P b : s 1 → ( x 1 , 4 ) and P a : s 2 → ( x 2 , 4 ) P_a:s_2\to(x_2,4) P a : s 2 → ( x 2 , 4 ) must have intersecting closed horizontal intervals on a common line, including switching endpoints: otherwise P a P_a P a lies strictly left of P b P_b P b on line 2, and this order propagates through their common lines, forcing x 2 < x 1 x_2 < x_1 x 2 < x 1 . Switch their tails on the first line where they intersect. The overlap makes both crossed intervals admissible. On the meeting line, with entrances α i \alpha_i α i and exits β i \beta_i β i ,
d E ( ( α 1 , β 1 ] ) + d E ( ( α 2 , β 2 ] ) = d E ( ( α 1 , β 2 ] ) + d E ( ( α 2 , β 1 ] ) . \mathrm{d}E((\alpha_1,\beta_1])+\mathrm{d}E((\alpha_2,\beta_2])=\mathrm{d}E((\alpha_1,\beta_2])+\mathrm{d}E((\alpha_2,\beta_1]). d E (( α 1 , β 1 ]) + d E (( α 2 , β 2 ]) = d E (( α 1 , β 2 ]) + d E (( α 2 , β 1 ]) .
Entrance singleton contributions are unchanged and subsequent segments are exchanged, preserving signed weight. Thus G b ( x 1 ) + G a ( x 2 ) ≤ G b ( x 2 ) + G a ( x 1 ) G_b(x_1)+G_a(x_2)\leq G_b(x_2)+G_a(x_1) G b ( x 1 ) + G a ( x 2 ) ≤ G b ( x 2 ) + G a ( x 1 ) , and pinning leaves this cross increment unchanged.
For a = c ≥ 1 a=c\geq1 a = c ≥ 1 , take a n ↓ c a_n\downarrow c a n ↓ c with c < a n < b c<a_n<b c < a n < b . Lemma 2.8, with the common clock-and-bridge coupling on a larger compact interval, gives agreement of both pinned right-continuous representatives for ( b , a n ) (b,a_n) ( b , a n ) and ( b , c ) (b,c) ( b , c ) with probability tending to one. Pass the inequality at rational spatial points and then everywhere by right-continuity. The difference is nondecreasing and vanishes at zero; applying usc preserves the order. □
Lemma 2.19 (One-parameter right continuity). If g n ↓ g g_n\downarrow g g n ↓ g in G c \mathcal{G}_c G c , there is a coupling such that, for every L < ∞ L<\infty L < ∞ ,
P ( H ( c , g n ) , → ∣ [ 0 , L ] = H ( c , g ) , → ∣ [ 0 , L ] ) ⟶ 1. \mathbb{P}\left(H^{(c,g_n),\rightarrow}\big|_{[0,L]}=H^{(c,g),\rightarrow}\big|_{[0,L]}\right)\longrightarrow1. P ( H ( c , g n ) , → [ 0 , L ] = H ( c , g ) , → [ 0 , L ] ) ⟶ 1. Consequently ν c , g n ⇒ ν c , g \nu_{c,g_n}\Rightarrow\nu_{c,g} ν c , g n ⇒ ν c , g locally in hypograph topology.
Proof. An eventually constant tail is coupled identically; otherwise g n > g g_n>g g n > g . Let M ( g ) M^{(g)} M ( g ) be the running maximum in (1.3). For g > c g>c g > c , common Poisson random measures with rate cutoffs couple the four counters with mismatch probability at most 2 L ( ∣ g n − g ∣ + ∣ g n − 1 − g − 1 ∣ ) 2L(|g_n-g|+|g_n^{-1}-g^{-1}|) 2 L ( ∣ g n − g ∣ + ∣ g n − 1 − g − 1 ∣ ) . Maximal coupling makes the geometric variables agree with probability tending to one, since c / g n → c / g < 1 c/g_n\to c/g<1 c / g n → c / g < 1 . At g = c ≥ 1 g=c\geq1 g = c ≥ 1 , Y ( c , g n ) → ∞ Y^{(c,g_n)}\to\infty Y ( c , g n ) → ∞ in probability, whereas M ( g n ) ( L ) M^{(g_n)}(L) M ( g n ) ( L ) is tight, bounded by four counters with bounded rates. The regulator therefore vanishes on [ 0 , L ] [0,L] [ 0 , L ] with probability tending to one, leaving the coupled S 2 ( c ) = H ( c , c ) , → S_2^{(c)}=H^{(c,c),\rightarrow} S 2 ( c ) = H ( c , c ) , → . Applying usc gives the hypograph convergence. □
Canonical regularization and stationarity
Put r = r c r=r_c r = r c . Let D + , 0 \mathcal{D}_{+,0} D + , 0 be the standard Borel space of pinned, locally finite, integer-valued càdlàg step functions, with its jump-coordinate Borel structure. Let E c \mathcal{E}_c E c consist of ( b , μ ) (b,\mu) ( b , μ ) , where the base profile b ∈ D + , 0 b\in\mathcal{D}_{+,0} b ∈ D + , 0 and μ \mu μ is a positive locally finite Borel measure on G c × R + \mathcal{G}_c\times\mathbb{R}_+ G c × R + , zero on the lower faces { r } × R + \{r\}\times\mathbb{R}_+ { r } × R + and G c × { 0 } \mathcal{G}_c\times\{0\} G c × { 0 } . The vague measure structure and Borel lower-face conditions make E c \mathcal{E}_c E c standard Borel. Set
E b , μ ( g , x ) : = b ( x ) + μ ( ( r , g ] × ( 0 , x ] ) , R c ( b , μ ) ( g , ⋅ ) : = usc E b , μ ( g , ⋅ ) . E_{b,\mu}(g,x):=b(x)+\mu((r,g]\times(0,x]),\qquad\mathcal{R}_c(b,\mu)(g,\cdot):=\operatorname{usc}E_{b,\mu}(g,\cdot). E b , μ ( g , x ) := b ( x ) + μ (( r , g ] × ( 0 , x ]) , R c ( b , μ ) ( g , ⋅ ) := usc E b , μ ( g , ⋅ ) .
Local finiteness makes each E b , μ ( g , ⋅ ) E_{b,\mu}(g,\cdot) E b , μ ( g , ⋅ ) càdlàg. Rectangle evaluations, spatial left limits, and (2.48) make the reconstruction jointly Borel in ( b , μ , g , x ) (b,\mu,g,x) ( b , μ , g , x ) .
Definition 2.20 (Regular half-space PNG horizon). A regular half-space PNG horizon of boundary rate c c c is a random ( b c P N G , μ c P N G ) ∈ E c (b_c^{\mathrm{PNG}},\mu_c^{\mathrm{PNG}})\in\mathcal{E}_c ( b c PNG , μ c PNG ) ∈ E c whose reconstructed field H c P N G : = R c ( b c P N G , μ c P N G ) H_c^{\mathrm{PNG}}:=\mathcal{R}_c(b_c^{\mathrm{PNG}},\mu_c^{\mathrm{PNG}}) H c PNG := R c ( b c PNG , μ c PNG ) belongs to X s t e p , 0 \mathcal{X}_{\mathrm{step},0} X step , 0 at every g ∈ G c g\in\mathcal{G}_c g ∈ G c and satisfies
Law ( H c P N G ( g 1 , ⋅ ) , … , H c P N G ( g k , ⋅ ) ) = ν c ; g 1 , … , g k (59) \operatorname{Law}(H_c^{\mathrm{PNG}}(g_1,\cdot),\ldots,H_c^{\mathrm{PNG}}(g_k,\cdot))=\nu_{c;g_1,\ldots,g_k} \tag*{(59)} Law ( H c PNG ( g 1 , ⋅ ) , … , H c PNG ( g k , ⋅ )) = ν c ; g 1 , … , g k (59) for every deterministic finite list. Write H c P N G , → : = E b c P N G , μ c P N G H_c^{\mathrm{PNG},\rightarrow}:=E_{b_c^{\mathrm{PNG}},\mu_c^{\mathrm{PNG}}} H c PNG , → := E b c PNG , μ c PNG .
The representing measure satisfies
μ c P N G ( ( g 1 , g 2 ] × ( x 1 , x 2 ] ) = Δ □ H c P N G , → ( g 1 , g 2 ; x 1 , x 2 ) . (60) \mu_c^{\mathrm{PNG}}((g_1,g_2]\times(x_1,x_2])=\Delta_{\square}H_c^{\mathrm{PNG},\rightarrow}(g_1,g_2;x_1,x_2). \tag*{(60)} μ c PNG (( g 1 , g 2 ] × ( x 1 , x 2 ]) = Δ □ H c PNG , → ( g 1 , g 2 ; x 1 , x 2 ) . (60) For g n ↓ g g_n\downarrow g g n ↓ g , continuity from above gives
sup 0 ≤ x ≤ L ∣ H c P N G , → ( g n , x ) − H c P N G , → ( g , x ) ∣ = μ c P N G ( ( g , g n ] × ( 0 , L ] ) ⟶ 0. \sup_{0\leq x\leq L}\left|H_c^{\mathrm{PNG},\rightarrow}(g_n,x)-H_c^{\mathrm{PNG},\rightarrow}(g,x)\right|=\mu_c^{\mathrm{PNG}}((g,g_n]\times(0,L])\longrightarrow0. 0 ≤ x ≤ L sup H c PNG , → ( g n , x ) − H c PNG , → ( g , x ) = μ c PNG (( g , g n ] × ( 0 , L ]) ⟶ 0. Proof of Theorem 1.9. Construction. Let Q c = { r } ∪ ( Q ∩ ( r , ∞ ) ) \mathbb{Q}_c=\{r\}\cup(\mathbb{Q}\cap(r,\infty)) Q c = { r } ∪ ( Q ∩ ( r , ∞ )) . By projective consistency, the standard Borel step space, and Kolmogorov’s extension theorem, there is a family ( H q 0 ( g ) ) q ∈ Q c (H_q^0(g))_{q\in\mathbb{Q}_c} ( H q 0 ( g ) ) q ∈ Q c with the prescribed finite-dimensional laws. Put f q = c ( H q 0 ( q ) ) f_q=c(H_q^0(q)) f q = c ( H q 0 ( q )) , b = f r b=f_r b = f r , and F q i n c = f q − b F_q^{\mathrm{inc}}=f_q-b F q inc = f q − b . On one full-probability event these are the right-continuous representatives and F q 2 i n c − F q 1 i n c F_{q_2}^{\mathrm{inc}}-F_{q_1}^{\mathrm{inc}} F q 2 inc − F q 1 inc is nonnegative and nondecreasing. For q ∈ Q c q\in\mathbb{Q}_c q ∈ Q c and rational x ≥ 0 x\geq0 x ≥ 0 , choose p n ∈ Q c p_n\in\mathbb{Q}_c p n ∈ Q c decreasing to q q q from above. Then f q ( x ) ≤ f p n ( x ) ↓ Z f_q(x)\leq f_{p_n}(x)\downarrow Z f q ( x ) ≤ f p n ( x ) ↓ Z . Lemma 2.19 gives Z = d f q ( x ) Z\overset{\mathrm{d}}{=}f_q(x) Z = d f q ( x ) , so order forces equality. A countable intersection yields F q i n c ( x ) = inf p ∈ Q c , p > q F p i n c ( x ) F_q^{\mathrm{inc}}(x)=\inf_{p\in\mathbb{Q}_c,\ p>q}F_p^{\mathrm{inc}}(x) F q inc ( x ) = inf p ∈ Q c , p > q F p inc ( x ) for q ∈ Q c q\in\mathbb{Q}_c q ∈ Q c and x ∈ Q ≥ 0 x\in\mathbb{Q}_{\geq0} x ∈ Q ≥ 0 . Define
F ( g , x ) : = inf q ∈ Q c , q > g inf y ∈ Q ≥ 0 , y > x F q i n c ( y ) . F(g,x):=\inf_{q\in\mathbb{Q}_c,\ q>g}\inf_{y\in\mathbb{Q}_{\geq0},\ y>x}F_q^{\mathrm{inc}}(y). F ( g , x ) := q ∈ Q c , q > g inf y ∈ Q ≥ 0 , y > x inf F q inc ( y ) . Commuting the infima and using spatial right-continuity gives F ( g , x ) = inf q ∈ Q c , q > g F q i n c ( x ) F(g,x)=\inf_{q\in\mathbb{Q}_c,\ q>g}F_q^{\mathrm{inc}}(x) F ( g , x ) = inf q ∈ Q c , q > g F q inc ( x ) and F ( q , x ) = F q i n c ( x ) F(q,x)=F_q^{\mathrm{inc}}(x) F ( q , x ) = F q inc ( x ) for q ∈ Q c q\in\mathbb{Q}_c q ∈ Q c . For the second identity, also use the preceding right-continuity identity at rational y ↓ x y\downarrow x y ↓ x . The function F F F is finite-valued, integer-valued, nondecreasing and right-continuous in each variable; agreement on Q c \mathbb{Q}_c Q c gives F ( r , x ) = F ( g , 0 ) = 0 F(r,x)=F(g,0)=0 F ( r , x ) = F ( g , 0 ) = 0 . Taking synchronized q i , n ↓ g i q_{i,n}\downarrow g_i q i , n ↓ g i with q 1 , n < q 2 , n q_{1,n}<q_{2,n} q 1 , n < q 2 , n in the crossing inequality gives Δ □ F ≥ 0 \Delta_{\square}F\geq0 Δ □ F ≥ 0 .
For fixed G > g G>g G > g and sufficiently small u , v > 0 u,v>0 u , v > 0 , the rectangle inequality gives
0 ≤ F ( g + u , x + v ) − F ( g , x ) ≤ F ( G , x + v ) − F ( G , x ) + F ( g + u , x ) − F ( g , x ) . 0\leq F(g+u,x+v)-F(g,x)\leq F(G,x+v)-F(G,x)+F(g+u,x)-F(g,x). 0 ≤ F ( g + u , x + v ) − F ( g , x ) ≤ F ( G , x + v ) − F ( G , x ) + F ( g + u , x ) − F ( g , x ) . Separate right-continuity therefore gives joint continuity from above. The nonnegative rectangle increments define an additive content with its Carathéodory extension, a unique positive locally finite measure μ \mu μ , zero on the lower faces, with
μ ( ( g 1 , g 2 ] × ( x 1 , x 2 ] ) = Δ □ F ( g 1 , g 2 ; x 1 , x 2 ) , F ( g , x ) = μ ( ( r , g ] × ( 0 , x ] ) . \mu((g_1,g_2]\times(x_1,x_2])=\Delta_{\square}F(g_1,g_2;x_1,x_2),\qquad F(g,x)=\mu((r,g]\times(0,x]). μ (( g 1 , g 2 ] × ( x 1 , x 2 ]) = Δ □ F ( g 1 , g 2 ; x 1 , x 2 ) , F ( g , x ) = μ (( r , g ] × ( 0 , x ]) .
For finite G > r G>r G > r and L > 0 L>0 L > 0 , local finiteness follows from F ( G , L ) < ∞ F(G,L)<\infty F ( G , L ) < ∞ . Masses of rectangles with endpoints in the chosen countable dense sets are measurable and generate the vague Borel structure. The preceding construction defines the regularization map
K c : ( X s t e p , 0 ) Q c ⟶ E c , ( H ( q ) ) q ∈ Q c ⟼ ( b , μ ) . \mathcal{K}_{c}:(\mathcal{X}_{\mathrm{step},0})^{\mathbb{Q}_{c}}\longrightarrow\mathcal{E}_{c},\qquad(H(q))_{q\in\mathbb{Q}_{c}}\longmapsto(b,\mu). K c : ( X step , 0 ) Q c ⟶ E c , ( H ( q ) ) q ∈ Q c ⟼ ( b , μ ) .
The good event is specified by regularity of each skeleton profile, the crossing inequalities and parameter right-continuity identities, both tested at rational spatial points. These are countably many Borel conditions. Thus K c \mathcal{K}_{c} K c is Borel, taking the value ( 0 , 0 ) (0,0) ( 0 , 0 ) off this event.
Set ( b ^ P N G , μ ^ P N G ) = ( b , μ ) (\widehat{b}^{\mathrm{PNG}},\widehat{\mu}^{\mathrm{PNG}})=(b,\mu) ( b PNG , μ PNG ) = ( b , μ ) and reconstruct by R c \mathcal{R}_{c} R c . For every g g g , L L L and q n ↓ g q_{n}\downarrow g q n ↓ g in Q c \mathbb{Q}_{c} Q c , one has μ ( ( g , q n ] × ( 0 , L ] ) = F q n i n c ( L ) − F ( g , L ) ↓ 0 \mu((g,q_{n}]\times(0,L])=F^{\mathrm{inc}}_{q_{n}}(L)-F(g,L)\downarrow0 μ (( g , q n ] × ( 0 , L ]) = F q n inc ( L ) − F ( g , L ) ↓ 0 . These masses are nonnegative integers and eventually vanish. On this single skeleton event the argument applies to every g ∈ G c g\in\mathcal{G}_{c} g ∈ G c and finite L ≥ 0 L\geq0 L ≥ 0 . Thus every slice agrees on each compact interval with a skeleton profile, proving admissibility at all parameters and eventual compact equality from the right.
Identification and uniqueness. For a deterministic list g 1 , … , g k g_{1},\ldots,g_{k} g 1 , … , g k , extend the skeleton to Q c ∪ { g 1 , … , g k } \mathbb{Q}_{c}\cup\{g_{1},\ldots,g_{k}\} Q c ∪ { g 1 , … , g k } . For q n ↓ g i q_{n}\downarrow g_{i} q n ↓ g i , planar crossing gives c ( H c 0 ( g i ) ) ( x ) ≤ b ( x ) + F ( g i , x ) = lim n f q n ( x ) c(H^{0}_{c}(g_{i}))(x)\leq b(x)+F(g_{i},x)=\lim_{n}f_{q_{n}}(x) c ( H c 0 ( g i )) ( x ) ≤ b ( x ) + F ( g i , x ) = lim n f q n ( x ) . Marginal continuity identifies their laws, so order forces equality at rational x x x , then everywhere by right-continuity:
c ( H c 0 ( g i ) ) ( x ) = b ( x ) + F ( g i , x ) = H P N G , → ( g i , x ) , x ≥ 0. (61) c(H^{0}_{c}(g_{i}))(x)=b(x)+F(g_{i},x)=H^{\mathrm{PNG},\rightarrow}(g_{i},x),\qquad x\geq0. \tag*{(61)} c ( H c 0 ( g i )) ( x ) = b ( x ) + F ( g i , x ) = H PNG , → ( g i , x ) , x ≥ 0. (61)
For fixed g i g_{i} g i , this identification event is Borel: equality of the right-continuous profiles is determined by their rational spatial values. Intersect over the finite list and apply usc to obtain (2.49); rectangle subtraction gives (2.50). Each E b , μ ( g , ⋅ ) E_{b,\mu}(g,\cdot) E b , μ ( g , ⋅ ) is càdlàg, so ( usc E b , μ ( g , ⋅ ) ) → = E b , μ ( g , ⋅ ) (\operatorname{usc}E_{b,\mu}(g,\cdot))^{\rightarrow}=E_{b,\mu}(g,\cdot) ( usc E b , μ ( g , ⋅ ) ) → = E b , μ ( g , ⋅ ) . Thus the skeleton profiles recover b = f r b=f_{r} b = f r and all rectangle masses of μ \mu μ on the chosen dense sets, proving uniqueness of the law of ( b , μ ) (b,\mu) ( b , μ ) .
Stationarity. Fix t ≥ 0 t\geq0 t ≥ 0 and let H c , t e v H^{\mathrm{ev}}_{c,t} H c , t ev be the evolved field in Theorem 1.9. Graphical evolution followed by Rec \operatorname{Rec} Rec is Borel on the step space, and joint stationarity gives ( H c , t e v ( q ) ) q (H^{\mathrm{ev}}_{c,t}(q))_{q} ( H c , t ev ( q ) ) q the original skeleton’s countable-product law. Put ( b ^ t , μ ^ t ) = K c ( ( H c , t e v ( q ) ) q ) (\widehat{b}_{t},\widehat{\mu}_{t})=\mathcal{K}_{c}((H^{\mathrm{ev}}_{c,t}(q))_{q}) ( b t , μ t ) = K c (( H c , t ev ( q ) ) q ) and H ^ c , t P N G = R c ( b ^ t , μ ^ t ) \widehat{H}^{\mathrm{PNG}}_{c,t}=\mathcal{R}_{c}(\widehat{b}_{t},\widehat{\mu}_{t}) H c , t PNG = R c ( b t , μ t ) . Borel reconstruction preserves the law of ( b , μ ) (b,\mu) ( b , μ ) . For each fixed deterministic g g g , joint stationarity gives equality in law of the augmented countable families ( ( H c , t e v ( q ) ) q , H c , t e v ( g ) ) ((H^{\mathrm{ev}}_{c,t}(q))_{q},H^{\mathrm{ev}}_{c,t}(g)) (( H c , t ev ( q ) ) q , H c , t ev ( g )) and ( ( H P N G ( q ) ) q , H P N G ( g ) ) ((H^{\mathrm{PNG}}(q))_{q},H^{\mathrm{PNG}}(g)) (( H PNG ( q ) ) q , H PNG ( g )) . Thus the measurable identification event (2.51) transfers to the evolved field. Hence H ^ c , t P N G ( g ) = H c , t e v ( g ) \widehat{H}^{\mathrm{PNG}}_{c,t}(g)=H^{\mathrm{ev}}_{c,t}(g) H c , t PNG ( g ) = H c , t ev ( g ) almost surely, proving (1.6) for each fixed time. ○
KPZ universality
Using the scalings of Section 1.4, we first prove convergence of the PNG horizon to the half-space stationary horizon in Section 3.1. Section 3.2 establishes convergence of the height evolution to the half-space KPZ fixed point through transition-probability asymptotics, stationary comparison, and tightness. Finally, Section 3.3 proves convergence of the passage field to the half-space directed landscape, together with joint convergence of evolutions from several initial profiles driven by the same environment.
Convergence to the half-space stationary horizon
We prove convergence of the stationary PNG horizon in two senses: joint convergence at every deterministic finite parameter list, and convergence of its base profile and representing measure in D ( λ ρ ) \mathcal{D}(\lambda_{\rho}) D ( λ ρ ) . For the first, we pass from the shared Poisson-line construction to the Brownian-line construction of the half-space stationary horizon. For the second, positivity of rectangle increments gives tightness of the measures, and thin-strip estimates exclude mass on the two lower faces. Both conclusions are retained, since evaluation at a fixed parameter need not be continuous in the measure topology.
Fix ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R}\cup\{-\infty\} ρ ∈ R ∪ { − ∞ } and use c ε = c ε ( ρ ) c_{\varepsilon}=c_{\varepsilon}(\rho) c ε = c ε ( ρ ) , λ ρ \lambda_{\rho} λ ρ , and g ε g_{\varepsilon} g ε from (1.7)–(1.8). Write
r ε : = c ε ∨ 1 = g ε ( λ ρ ) , ℓ ε ( g ) : = 2 ε − 1 / 2 log g , ℓ ε − 1 = g ε . r_{\varepsilon}:=c_{\varepsilon}\vee1=g_{\varepsilon}(\lambda_{\rho}),\qquad\ell_{\varepsilon}(g):=2\varepsilon^{-1/2}\log g,\qquad\ell_{\varepsilon}^{-1}=g_{\varepsilon}. r ε := c ε ∨ 1 = g ε ( λ ρ ) , ℓ ε ( g ) := 2 ε − 1/2 log g , ℓ ε − 1 = g ε .
The rescaled PNG horizon
We first identify the representing pair of H ε P N G , ρ H_{\varepsilon}^{\mathrm{PNG},\rho} H ε PNG , ρ and verify its measurability. Spatial interpolation spreads each microscopic spatial atom uniformly across its mesh interval. The atom’s rescaled mass is ε 1 / 2 \varepsilon^{1/2} ε 1/2 times its original mass, so the corresponding density on an interval of length ε \varepsilon ε has factor ε − 1 / 2 \varepsilon^{-1/2} ε − 1/2 .
For y > 0 y>0 y > 0 , let j ( y ) = n j(y)=n j ( y ) = n on ( n , n + 1 ] (n,n+1] ( n , n + 1 ] , n ∈ N 0 n\in\mathbb{N}_{0} n ∈ N 0 , and put j ( 0 ) = 0 j(0)=0 j ( 0 ) = 0 . For a Borel set A ⊆ [ λ ρ , ∞ ) × R + A\subseteq[\lambda_{\rho},\infty)\times\mathbb{R}_{+} A ⊆ [ λ ρ , ∞ ) × R + , define
μ ε ( A ) : = ε − 1 / 2 ∫ ∫ ε j ( y ) ε ( j ( y ) + 1 ) 1 A ( ℓ ε ( g ) , x ) d x μ c P N G ( d g , d y ) . \mu_{\varepsilon}(A):=\varepsilon^{-1/2}\int\int_{\varepsilon j(y)}^{\varepsilon(j(y)+1)}\mathbf{1}_{A}(\ell_{\varepsilon}(g),x)\,\mathrm{d}x\,\mu_{c}^{\mathrm{PNG}}(\mathrm{d}g,\mathrm{d}y). μ ε ( A ) := ε − 1/2 ∫ ∫ ε j ( y ) ε ( j ( y ) + 1 ) 1 A ( ℓ ε ( g ) , x ) d x μ c PNG ( d g , d y ) .
For λ ρ ≤ λ 1 < λ 2 \lambda_{\rho}\leq\lambda_{1}<\lambda_{2} λ ρ ≤ λ 1 < λ 2 , put g a = g ε ( λ a ) g_{a}=g_{\varepsilon}(\lambda_{a}) g a = g ε ( λ a ) and
F n i n c : = H ε P N G , → ( g 2 , n ) − H ε P N G , → ( g 1 , n ) . F_{n}^{\mathrm{inc}}:=H_{\varepsilon}^{\mathrm{PNG},\rightarrow}(g_{2},n)-H_{\varepsilon}^{\mathrm{PNG},\rightarrow}(g_{1},n). F n inc := H ε PNG , → ( g 2 , n ) − H ε PNG , → ( g 1 , n ) .
At x = ε ( n + θ ) x=\varepsilon(n+\theta) x = ε ( n + θ ) , 0 ≤ θ < 1 0\leq\theta<1 0 ≤ θ < 1 , the microscopic rectangle identity gives
μ ε ( ( λ 1 , λ 2 ] × ( 0 , x ] ) = ε 1 / 2 [ ( 1 − θ ) F n i n c + θ F n + 1 i n c ] = H ε P N G , ρ ( λ 2 , x ) − H ε P N G , ρ ( λ 1 , x ) . \mu_{\varepsilon}((\lambda_{1},\lambda_{2}]\times(0,x])=\varepsilon^{1/2}\left[(1-\theta)F_{n}^{\mathrm{inc}}+\theta F_{n+1}^{\mathrm{inc}}\right]=H_{\varepsilon}^{\mathrm{PNG},\rho}(\lambda_{2},x)-H_{\varepsilon}^{\mathrm{PNG},\rho}(\lambda_{1},x). μ ε (( λ 1 , λ 2 ] × ( 0 , x ]) = ε 1/2 [ ( 1 − θ ) F n inc + θ F n + 1 inc ] = H ε PNG , ρ ( λ 2 , x ) − H ε PNG , ρ ( λ 1 , x ) .
Thus, with b ε ( x ) : = H ε P N G , ρ ( λ ρ , x ) b_{\varepsilon}(x):=H_{\varepsilon}^{\mathrm{PNG},\rho}(\lambda_{\rho},x) b ε ( x ) := H ε PNG , ρ ( λ ρ , x ) ,
H ε P N G , ρ ( λ , x ) = b ε ( x ) + μ ε ( ( λ ρ , λ ] × ( 0 , x ] ) . H_{\varepsilon}^{\mathrm{PNG},\rho}(\lambda,x)=b_{\varepsilon}(x)+\mu_{\varepsilon}((\lambda_{\rho},\lambda]\times(0,x]). H ε PNG , ρ ( λ , x ) = b ε ( x ) + μ ε (( λ ρ , λ ] × ( 0 , x ]) .
The measure is positive and locally finite. Its parameter lower face has zero mass by the microscopic normalization, and integration with respect to d x \mathrm{d}x d x gives zero mass to the spatial lower face. Interpolation makes every spatial profile continuous and pinned. Mesh evaluations make b ε b_{\varepsilon} b ε measurable, and Borel integration makes ∫ f d μ ε \int f\,\mathrm{d}\mu_{\varepsilon} ∫ f d μ ε measurable for every f ∈ C c ( [ λ ρ , ∞ ) × R + ) f\in C_{c}([\lambda_{\rho},\infty)\times\mathbb{R}_{+}) f ∈ C c ([ λ ρ , ∞ ) × R + ) . A countable determining family therefore makes ( b ε , μ ε ) (b_{\varepsilon},\mu_{\varepsilon}) ( b ε , μ ε ) a Borel D ( λ ρ ) \mathcal{D}(\lambda_{\rho}) D ( λ ρ ) -valued random element.
Finite-parameter convergence
Fix λ 1 > ⋯ > λ k ≥ λ ρ \lambda_1 > \cdots> \lambda_k \ge\lambda_\rho λ 1 > ⋯ > λ k ≥ λ ρ and put j ˉ = 2 k + 1 − j \bar{j} = 2k + 1 - j j ˉ = 2 k + 1 − j . The limiting shared environment consists of independent Brownian lines and exponential bridges. Let B 1 , … , B 2 k B_1,\ldots,B_{2k} B 1 , … , B 2 k be independent Brownian motions started at zero, with drift d j d_j d j and variance specified by
d j : = { λ j , j ≤ k , − λ j ˉ , j > k , Var ( B j ( y ) − B j ( x ) ) = 2 ( y − x ) , 0 ≤ x < y . d_j :=
\begin{cases}
\lambda_j, & j \le k,\\
-\lambda_{\bar{j}}, & j > k,
\end{cases}
\qquad\operatorname{Var}(B_j(y)-B_j(x)) = 2(y-x), \qquad0 \le x < y. d j := { λ j , − λ j ˉ , j ≤ k , j > k , Var ( B j ( y ) − B j ( x )) = 2 ( y − x ) , 0 ≤ x < y .
For 1 ≤ i ≤ k 1 \le i \le k 1 ≤ i ≤ k and i < j ≤ i ˉ i < j \le\bar{i} i < j ≤ i ˉ , define
a i j ρ : = { ( λ i − λ j ) / 2 , i < j ≤ k , ( λ i + λ j ˉ ) / 2 , k < j < i ˉ , λ i / 2 − ρ , j = i ˉ , ρ ∈ R , + ∞ , j = i ˉ , ρ = − ∞ . a_{ij}^{\rho} :=
\begin{cases}
(\lambda_i-\lambda_j)/2, & i < j \le k,\\
(\lambda_i+\lambda_{\bar{j}})/2, & k < j < \bar{i},\\
\lambda_i/2-\rho, & j=\bar{i},\ \rho\in\mathbb{R},\\
+\infty, & j=\bar{i},\ \rho=-\infty.
\end{cases} a ij ρ := ⎩ ⎨ ⎧ ( λ i − λ j ) /2 , ( λ i + λ j ˉ ) /2 , λ i /2 − ρ , + ∞ , i < j ≤ k , k < j < i ˉ , j = i ˉ , ρ ∈ R , j = i ˉ , ρ = − ∞.
For each positive rate take X ρ ( − i , j ) ∼ Exp ( a i j ρ ) X^\rho(-i,j) \sim\operatorname{Exp}(a_{ij}^{\rho}) X ρ ( − i , j ) ∼ Exp ( a ij ρ ) , independently of one another and of the Brownian lines; here Exp ( + ∞ ) \operatorname{Exp}(+\infty) Exp ( + ∞ ) is the point mass at zero. Set s i = ( − i , i ) s_i=(-i,i) s i = ( − i , i ) . The only zero rate occurs when ρ ≥ 0 \rho\ge0 ρ ≥ 0 , λ k = 2 ρ \lambda_k=2\rho λ k = 2 ρ , i = k i=k i = k , and j = k + 1 j=k+1 j = k + 1 . In this case omit that atom and use the shifted source s k = ( − k + 1 , k + 1 ) s_k=(-k+1,k+1) s k = ( − k + 1 , k + 1 ) .
As in the Poisson-line construction, concatenate the cumulative bridge functions on ( − ∞ , 0 ] (-\infty,0] ( − ∞ , 0 ] , normalized at zero, with the Brownian lines to obtain E ρ E^\rho E ρ . Using Definition 2.5 with marked columns { − k , … , − 1 } \{-k,\ldots,-1\} { − k , … , − 1 } , set
R i ρ ( x ) : = L E ρ ( s i ; ( x , 2 k ) ) − L E ρ ( s i ; ( 0 , 2 k ) ) . R_i^\rho(x):=\mathcal{L}_{E^\rho}(s_i;(x,2k))-\mathcal{L}_{E^\rho}(s_i;(0,2k)). R i ρ ( x ) := L E ρ ( s i ; ( x , 2 k )) − L E ρ ( s i ; ( 0 , 2 k )) .
The bridge-entry convention agrees with [17 Section 2.5] : the positive-axis lines are continuous and all atoms lie at the marked bridge columns. The terminal line is included, consistently with the construction and shifted-source formula in [17 Proposition 2.18] . By [17 Proposition 2.18 and Theorem 2.19] ,
( R i ρ ) i = 1 k = d ( H ρ ( λ i , ⋅ ) ) i = 1 k ∣ R + . (62) \left(R_i^\rho\right)_{i=1}^k \mathrel{\mathop{=}^{d}} \left(H^\rho(\lambda_i,\cdot)\right)_{i=1}^k\bigg|_{\mathbb{R}_+}. \tag*{(62)} ( R i ρ ) i = 1 k = d ( H ρ ( λ i , ⋅ ) ) i = 1 k R + . (62)
Proposition 3.1 (Finite-parameter invariance principle). For every deterministic list λ 1 > ⋯ > λ k ≥ λ ρ \lambda_1 > \cdots> \lambda_k \ge\lambda_\rho λ 1 > ⋯ > λ k ≥ λ ρ ,
( H ε P N G , ρ ( λ i , ⋅ ) ) i = 1 k ⟹ ( H ρ ( λ i , ⋅ ) ) i = 1 k in C l o c ( R + ) k . \left(H_{\varepsilon}^{\mathrm{PNG},\rho}(\lambda_i,\cdot)\right)_{i=1}^k \Longrightarrow\left(H^\rho(\lambda_i,\cdot)\right)_{i=1}^k \quad\text{in } C_{\mathrm{loc}}(\mathbb{R}_+)^k. ( H ε PNG , ρ ( λ i , ⋅ ) ) i = 1 k ⟹ ( H ρ ( λ i , ⋅ ) ) i = 1 k in C loc ( R + ) k .
Proof. Set g i , ε = g ε ( λ i ) g_{i,\varepsilon}=g_\varepsilon(\lambda_i) g i , ε = g ε ( λ i ) , g ε = ( g 1 , ε , … , g k , ε ) \mathbf{g}_\varepsilon=(g_{1,\varepsilon},\ldots,g_{k,\varepsilon}) g ε = ( g 1 , ε , … , g k , ε ) , and γ j , ε = e d j ε 1 / 2 / 2 \gamma_{j,\varepsilon}=e^{d_j\varepsilon^{1/2}/2} γ j , ε = e d j ε 1/2 /2 .
Step 1: convergence of the Poisson lines. Let S j , ε : = S j g ε S_{j,\varepsilon}:=S_{j}^{g_\varepsilon} S j , ε := S j g ε and put
S ^ j , ε ( x ) : = ε 1 / 2 S j , ε ( ε − 1 x ) , M ^ j , ε ( x ) : = S ^ j , ε ( x ) − ε − 1 / 2 ( γ j , ε − γ j , ε − 1 ) x . \widehat{S}_{j,\varepsilon}(x):=\varepsilon^{1/2}S_{j,\varepsilon}(\varepsilon^{-1}x),\qquad\widehat{M}_{j,\varepsilon}(x):=\widehat{S}_{j,\varepsilon}(x)-\varepsilon^{-1/2}(\gamma_{j,\varepsilon}-\gamma_{j,\varepsilon}^{-1})x. S j , ε ( x ) := ε 1/2 S j , ε ( ε − 1 x ) , M j , ε ( x ) := S j , ε ( x ) − ε − 1/2 ( γ j , ε − γ j , ε − 1 ) x .
The martingale jumps have size ε 1 / 2 \varepsilon^{1/2} ε 1/2 , and locally uniformly,
ε − 1 / 2 ( γ j , ε − γ j , ε − 1 ) x ⟶ d j x , ⟨ M ^ j , ε ⟩ ( x ) = ( γ j , ε + γ j , ε − 1 ) x ⟶ 2 x . \varepsilon^{-1/2}(\gamma_{j,\varepsilon}-\gamma_{j,\varepsilon}^{-1})x\longrightarrow d_jx,\qquad\langle\widehat{M}_{j,\varepsilon}\rangle(x)=(\gamma_{j,\varepsilon}+\gamma_{j,\varepsilon}^{-1})x\longrightarrow2x. ε − 1/2 ( γ j , ε − γ j , ε − 1 ) x ⟶ d j x , ⟨ M j , ε ⟩ ( x ) = ( γ j , ε + γ j , ε − 1 ) x ⟶ 2 x .
The vanishing jumps give the Lindeberg condition, and independence gives zero cross variations. The martingale functional central limit theorem therefore yields joint convergence of the lines to ( B j ) j = 1 2 k (B_j)_{j=1}^{2k} ( B j ) j = 1 2 k in D l o c ( R + ) 2 k D_{\mathrm{loc}}(\mathbb{R}_+)^{2k} D loc ( R + ) 2 k with the local Skorokhod J 1 J_1 J 1 topology.
Step 2: convergence of the bridges and treatment of endpoints. For each finite positive a i j ρ a_{ij}^{\rho} a ij ρ , the geometric parameter in (2.5) is e − a i j ρ ε 1 / 2 e^{-a_{ij}^{\rho}\varepsilon^{1/2}} e − a ij ρ ε 1/2 . If Z ε ∼ Geo ( e − a ε 1 / 2 ) Z_\varepsilon\sim\operatorname{Geo}(e^{-a\varepsilon^{1/2}}) Z ε ∼ Geo ( e − a ε 1/2 ) with a > 0 a>0 a > 0 , then, for x ≥ 0 x\ge0 x ≥ 0 ,
P ( ε 1 / 2 Z ε > x ) = e − a ε 1 / 2 ( ⌊ x ε − 1 / 2 ⌋ + 1 ) ⟶ e − a x . \mathbb{P}(\varepsilon^{1/2}Z_\varepsilon>x)=e^{-a\varepsilon^{1/2}(\lfloor x\varepsilon^{-1/2}\rfloor+1)}\longrightarrow e^{-ax}. P ( ε 1/2 Z ε > x ) = e − a ε 1/2 (⌊ x ε − 1/2 ⌋ + 1 ) ⟶ e − a x .
For ρ = − ∞ \rho=-\infty ρ = − ∞ , terminal bridge parameters are at most c ⋆ < 1 c_\star<1 c ⋆ < 1 , so E [ ε 1 / 2 Z ] ≤ ε 1 / 2 c ⋆ / ( 1 − c ⋆ ) → 0 \mathbb{E}[\varepsilon^{1/2}Z]\le\varepsilon^{1/2}c_\star/(1-c_\star)\to0 E [ ε 1/2 Z ] ≤ ε 1/2 c ⋆ / ( 1 − c ⋆ ) → 0 . If a k , k + 1 ρ = 0 a_{k,k+1}^{\rho}=0 a k , k + 1 ρ = 0 , then g k , ε = c ε ≥ 1 g_{k,\varepsilon}=c_\varepsilon\ge1 g k , ε = c ε ≥ 1 for every ε \varepsilon ε : both constructions omit the terminal bridge and use the same shifted source. This includes ρ = 0 \rho=0 ρ = 0 , λ k = 0 \lambda_k=0 λ k = 0 . Independence makes convergence of all the lines and retained bridges joint, with the specified independent limiting inputs.
Step 3: convergence of the passage profiles. Let E ε = ( E i , j ) j = 1 2 k E_\varepsilon=(E_{i,j})_{j=1}^{2k} E ε = ( E i , j ) j = 1 2 k be the corresponding shared PNG environment, using the shifted-source convention when needed. The negative coordinates label a finite bridge array, so we keep them fixed and rescale only the positive axis. Set
ϕ ε ( u ) : = { u , u ≤ 0 , ε u , u ≥ 0 , E ^ j , ε ( x ) : = ε 1 / 2 E j , ε ( ϕ ε − 1 ( x ) ) . \phi_\varepsilon(u):=
\begin{cases}
u, & u\le0,\\
\varepsilon u, & u\ge0,
\end{cases}
\qquad\widehat{E}_{j,\varepsilon}(x):=\varepsilon^{1/2}E_{j,\varepsilon}(\phi_\varepsilon^{-1}(x)). ϕ ε ( u ) := { u , ε u , u ≤ 0 , u ≥ 0 , E j , ε ( x ) := ε 1/2 E j , ε ( ϕ ε − 1 ( x )) .
This increasing coordinate change preserves switching paths and the bridge-entry convention. Hence
L E ^ ε ( s i ; ( x , 2 k ) ) − L E ^ ε ( s i ; ( 0 , 2 k ) ) = ε 1 / 2 R ^ i ( c ε ; g ε ) ( ε − 1 x ) . \mathcal{L}_{\widehat{E}_\varepsilon}(s_i;(x,2k))-\mathcal{L}_{\widehat{E}_\varepsilon}(s_i;(0,2k))=\varepsilon^{1/2}\widehat{R}_{i}^{(c_\varepsilon;\mathbf{g}_\varepsilon)}(\varepsilon^{-1}x). L E ε ( s i ; ( x , 2 k )) − L E ε ( s i ; ( 0 , 2 k )) = ε 1/2 R i ( c ε ; g ε ) ( ε − 1 x ) .
Along any sequence ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , use a joint Skorokhod representation of the lines and bridges. Continuity of the Brownian limits gives uniform convergence of the positive lines on compacts. Together with convergence of the bridge masses, this gives, almost surely,
η ε , L : = max 1 ≤ j ≤ 2 k sup − k − 1 ≤ x ≤ L ∣ E ^ j , ε ( x ) − E j ρ ( x ) ∣ ⟶ 0 , \eta_{\varepsilon,L}:=\max_{1\le j\le2k}\sup_{-k-1\le x\le L}\left|\widehat{E}_{j,\varepsilon}(x)-E_j^\rho(x)\right|\longrightarrow0, η ε , L := 1 ≤ j ≤ 2 k max − k − 1 ≤ x ≤ L sup E j , ε ( x ) − E j ρ ( x ) ⟶ 0 ,
also controlling left limits at the fixed bridge columns. Every path uses at most 2 k 2k 2 k lines, and each line contributes at most 4 η ε , L 4\eta_{\varepsilon,L} 4 η ε , L to the weight error. Taking suprema and then pinning therefore gives
sup 0 ≤ x ≤ L ∣ ε 1 / 2 R ^ i ( c ε ; g ε ) ( ε − 1 x ) − R i ρ ( x ) ∣ ≤ 16 k η ε , L . \sup_{0\le x\le L}\left|\varepsilon^{1/2}\widehat{R}_{i}^{(c_\varepsilon;\mathbf{g}_\varepsilon)}(\varepsilon^{-1}x)-R_i^\rho(x)\right|\le16k\eta_{\varepsilon,L}. 0 ≤ x ≤ L sup ε 1/2 R i ( c ε ; g ε ) ( ε − 1 x ) − R i ρ ( x ) ≤ 16 k η ε , L .
Step 4: interpolation and identification of the limiting law. By (59) and (61) , the canonical right-continuous horizon representatives have these pinned profiles’ finite joint laws. The Brownian passage profiles R i p R_i^p R i p are continuous, by uniform continuity of the finitely many Brownian lines on compacts. Write V i , ε V_{i,\varepsilon} V i , ε for the scaled càdlàg profile and V ‾ i , ε \overline{V}_{i,\varepsilon} V i , ε for its mesh interpolation. For a continuous f f f , write
ω f ( a ; L ) : = sup x , y ∈ [ 0 , L ] ∣ x − y ∣ ≤ a ∣ f ( x ) − f ( y ) ∣ . \omega_f(a;L) := \sup_{\substack{x,y\in[0,L]\\ |x-y|\le a}} |f(x)-f(y)|. ω f ( a ; L ) := x , y ∈ [ 0 , L ] ∣ x − y ∣ ≤ a sup ∣ f ( x ) − f ( y ) ∣.
For ε ≤ 1 \varepsilon\le1 ε ≤ 1 ,
∥ V ‾ i , ε − R i p ∥ [ 0 , L ] ≤ ∥ V i , ε − R i p ∥ [ 0 , L + ε ] + ω R i p ( ε ; L + 1 ) ⟶ 0. \left\|\overline{V}_{i,\varepsilon}-R_i^p\right\|_{[0,L]}
\le\left\|V_{i,\varepsilon}-R_i^p\right\|_{[0,L+\varepsilon]}+\omega_{R_i^p}(\varepsilon;L+1)\longrightarrow0. V i , ε − R i p [ 0 , L ] ≤ ∥ V i , ε − R i p ∥ [ 0 , L + ε ] + ω R i p ( ε ; L + 1 ) ⟶ 0.
The joint convergence and (62) prove the proposition. □
Convergence of the representing pair
The preceding proposition identifies every prescribed finite collection of profiles. We now use positivity to obtain tightness and identify the law of the complete representing pair.
Lemma 3.2 (Convergence in D ( λ 0 ) \mathcal{D}(\lambda_0) D ( λ 0 ) ). Let G n , G G_n,G G n , G be Borel D ( λ 0 ) \mathcal{D}(\lambda_0) D ( λ 0 ) -valued random elements, where D ( λ 0 ) \mathcal{D}(\lambda_0) D ( λ 0 ) is the space defined in Section 1.4.1 with lower parameter endpoint λ 0 \lambda_0 λ 0 . Assume that, for every deterministic λ ≥ λ 0 \lambda\ge\lambda_0 λ ≥ λ 0 , their spatial slices are C l o c ( R + ) C_{\mathrm{loc}}(\mathbb{R}_+) C loc ( R + ) -valued random elements. If, for every deterministic finite parameter list,
( G n ( λ i , ⋅ ) ) i = 1 m ⟹ ( G ( λ i , ⋅ ) ) i = 1 m in C l o c ( R + ) m , \left(G_n(\lambda_i,\cdot)\right)_{i=1}^m\Longrightarrow\left(G(\lambda_i,\cdot)\right)_{i=1}^m\quad\text{in }C_{\mathrm{loc}}(\mathbb{R}_+)^m, ( G n ( λ i , ⋅ ) ) i = 1 m ⟹ ( G ( λ i , ⋅ ) ) i = 1 m in C loc ( R + ) m , then G n ⟹ G G_n\Longrightarrow G G n ⟹ G in D ( λ 0 ) \mathcal{D}(\lambda_0) D ( λ 0 ) .
Proof. Write ( b n , μ n ) (b_n,\mu_n) ( b n , μ n ) and ( b , μ ) (b,\mu) ( b , μ ) for the representing pairs. Work first in the ambient Polish space
C l o c , 0 ( R + ) × M l o c , + ( [ λ 0 , ∞ ) × R + ) , C_{\mathrm{loc},0}(\mathbb{R}_+)\times\mathcal{M}_{\mathrm{loc},+}([\lambda_0,\infty)\times\mathbb{R}_+), C loc , 0 ( R + ) × M loc , + ([ λ 0 , ∞ ) × R + ) ,
where the first factor consists of continuous profiles pinned at zero with the local uniform topology, and the second of positive locally finite measures with the vague topology.
Step 1: tightness. The hypothesis at λ 0 \lambda_0 λ 0 makes ( b n ) (b_n) ( b n ) tight. For finite Λ > λ 0 \Lambda>\lambda_0 Λ > λ 0 and L > 0 L>0 L > 0 ,
μ n ( [ λ 0 , Λ ] × [ 0 , L ] ) = G n ( Λ , L ) − G n ( λ 0 , L ) \mu_n([\lambda_0,\Lambda]\times[0,L])=G_n(\Lambda,L)-G_n(\lambda_0,L) μ n ([ λ 0 , Λ ] × [ 0 , L ]) = G n ( Λ , L ) − G n ( λ 0 , L )
is tight by the two-parameter hypothesis. Positivity and a compact exhaustion give vague tightness of ( μ n ) (\mu_n) ( μ n ) , and hence tightness of the pairs in the ambient space.
Step 2: exclusion of lower-face mass. Zero lower-face masses need not be preserved by vague limits, so we control the mass of thin strips. Put
X n , δ L : = G n ( λ 0 + δ , L ) − G n ( λ 0 , L ) , Y n , δ Λ : = G n ( Λ , δ ) − G n ( λ 0 , δ ) . X_{n,\delta}^L:=G_n(\lambda_0+\delta,L)-G_n(\lambda_0,L),\qquad Y_{n,\delta}^{\Lambda}:=G_n(\Lambda,\delta)-G_n(\lambda_0,\delta). X n , δ L := G n ( λ 0 + δ , L ) − G n ( λ 0 , L ) , Y n , δ Λ := G n ( Λ , δ ) − G n ( λ 0 , δ ) .
These are the masses of ( λ 0 , λ 0 + δ ] × ( 0 , L ] (\lambda_0,\lambda_0+\delta]\times(0,L] ( λ 0 , λ 0 + δ ] × ( 0 , L ] and ( λ 0 , Λ ] × ( 0 , δ ] (\lambda_0,\Lambda]\times(0,\delta] ( λ 0 , Λ ] × ( 0 , δ ] , respectively. For fixed δ \delta δ they converge in law to their target counterparts X δ L , Y δ Λ X_\delta^L,Y_\delta^\Lambda X δ L , Y δ Λ , which decrease to zero as δ ↓ 0 \delta\downarrow0 δ ↓ 0 by continuity from above. Portmanteau therefore gives, for every η > 0 \eta>0 η > 0 ,
lim δ ↓ 0 lim sup n P ( X n , δ L > η ) = lim δ ↓ 0 lim sup n P ( Y n , δ Λ > η ) = 0. \lim_{\delta\downarrow0}\limsup_n\mathbb{P}(X_{n,\delta}^L>\eta)
=
\lim_{\delta\downarrow0}\limsup_n\mathbb{P}(Y_{n,\delta}^{\Lambda}>\eta)=0. δ ↓ 0 lim n lim sup P ( X n , δ L > η ) = δ ↓ 0 lim n lim sup P ( Y n , δ Λ > η ) = 0.
Take an ambient subsequential limit ( b ∗ , μ ∗ ) (b_*,\mu_*) ( b ∗ , μ ∗ ) and a Skorokhod coupling. A continuous compactly supported cutoff in [ 0 , 1 ] [0,1] [ 0 , 1 ] , equal to one on { λ 0 } × [ 0 , L ] \{\lambda_0\}\times[0,L] { λ 0 } × [ 0 , L ] and supported in [ λ 0 , λ 0 + δ ) × [ 0 , L + 1 ) [\lambda_0,\lambda_0+\delta)\times[0,L+1) [ λ 0 , λ 0 + δ ) × [ 0 , L + 1 ) , has μ n \mu_n μ n integral at most X n , δ L + 1 X_{n,\delta}^{L+1} X n , δ L + 1 . Similarly, a cutoff equal to one on [ λ 0 , Λ ] × { 0 } [\lambda_0,\Lambda]\times\{0\} [ λ 0 , Λ ] × { 0 } and supported in [ λ 0 , Λ + 1 ) × [ 0 , δ ) [\lambda_0,\Lambda+1)\times[0,\delta) [ λ 0 , Λ + 1 ) × [ 0 , δ ) has integral at most Y n , δ Λ + 1 Y_{n,\delta}^{\Lambda+1} Y n , δ Λ + 1 . Almost sure vague convergence yields
P ( μ ∗ ( { λ 0 } × [ 0 , L ] ) > η ) ≤ lim sup n P ( X n , δ L + 1 > η / 2 ) , \mathbb{P}\bigl(\mu_*(\{\lambda_0\}\times[0,L])>\eta\bigr)
\le\limsup_n\mathbb{P}(X_{n,\delta}^{L+1}>\eta/2), P ( μ ∗ ({ λ 0 } × [ 0 , L ]) > η ) ≤ n lim sup P ( X n , δ L + 1 > η /2 ) ,
P ( μ ∗ ( [ λ 0 , Λ ] × { 0 } ) > η ) ≤ lim sup n P ( Y n , δ Λ + 1 > η / 2 ) . \mathbb{P}\bigl(\mu_*([\lambda_0,\Lambda]\times\{0\})>\eta\bigr)
\le\limsup_n\mathbb{P}(Y_{n,\delta}^{\Lambda+1}>\eta/2). P ( μ ∗ ([ λ 0 , Λ ] × { 0 }) > η ) ≤ n lim sup P ( Y n , δ Λ + 1 > η /2 ) .
Let δ ↓ 0 \delta\downarrow0 δ ↓ 0 and exhaust the two faces. Both lower faces are μ ∗ \mu_* μ ∗ -null almost surely.
Step 3: identification of the subsequential limit. Choose deterministic countable dense grids Q ⊂ ( λ 0 , ∞ ) Q\subset(\lambda_0,\infty) Q ⊂ ( λ 0 , ∞ ) and T ⊂ ( 0 , ∞ ) T\subset(0,\infty) T ⊂ ( 0 , ∞ ) whose vertical and horizontal lines are μ ∗ \mu_* μ ∗ -null almost surely. Such grids exist: the countable projection sets of the finite measures
μ ‾ K ( A ) : = E [ μ ∗ ( A ∩ K ) 1 + μ ∗ ( K ) ] \overline{\mu}_K(A):=\mathbb{E}\left[\frac{\mu_*(A\cap K)}{1+\mu_*(K)}\right] μ K ( A ) := E [ 1 + μ ∗ ( K ) μ ∗ ( A ∩ K ) ]
have only countably many atoms, and one avoids these atoms over a compact exhaustion K K K . The resulting rectangles are continuity sets, so on the coupling,
G n ( q , t ) = b n ( t ) + μ n ( ( λ 0 , q ] × ( 0 , t ] ) ⟶ b ∗ ( t ) + μ ∗ ( ( λ 0 , q ] × ( 0 , t ] ) = : G ∗ ( q , t ) . G_n(q,t)=b_n(t)+\mu_n((\lambda_0,q]\times(0,t])\longrightarrow b_*(t)+\mu_*((\lambda_0,q]\times(0,t])=:G_*(q,t). G n ( q , t ) = b n ( t ) + μ n (( λ 0 , q ] × ( 0 , t ]) ⟶ b ∗ ( t ) + μ ∗ (( λ 0 , q ] × ( 0 , t ]) =: G ∗ ( q , t ) .
The finite-parameter hypothesis identifies every finite grid law, jointly with the base-profile values, with the corresponding law of G G G . Dense base values determine the continuous base profile, and rectangle increments on the grids determine the locally finite measure. Together with the zero lower-face masses, this proves ( b ∗ , μ ∗ ) = d ( b , μ ) (b_*,\mu_*)\overset{d}{=}(b,\mu) ( b ∗ , μ ∗ ) = d ( b , μ ) . Every ambient subsequential limit therefore lies in D ( λ 0 ) \mathcal{D}(\lambda_0) D ( λ 0 ) and has the target pair law. This proves the claimed convergence in its induced topology. □
Proof of Theorem 1.10. Restrict the horizon of [17 Theorem 2.19] to x ≥ 0 x \ge0 x ≥ 0 . Restrict its representing measure to ( λ ρ , ∞ ) × ( 0 , ∞ ) (\lambda_\rho,\infty) \times(0,\infty) ( λ ρ , ∞ ) × ( 0 , ∞ ) and extend by zero on the lower faces; this leaves all reconstruction rectangles unchanged. Positivity of cross increments from [17 Definition 2.13] gives
μ ρ ( [ λ ρ , Λ ] × [ 0 , L ] ) = H ρ ( Λ , L ) − H ρ ( λ ρ , L ) < ∞ , \mu^\rho([\lambda_\rho,\Lambda] \times[0,L]) = \mathcal{H}^\rho(\Lambda,L) - \mathcal{H}^\rho(\lambda_\rho,L) < \infty, μ ρ ([ λ ρ , Λ ] × [ 0 , L ]) = H ρ ( Λ , L ) − H ρ ( λ ρ , L ) < ∞ ,
so the measure is locally finite. [17 Proposition 2.18] gives continuous spatial profiles at each deterministic parameter. The continuous base profile and rectangle increments on countable dense grids give a Borel D ( λ ρ ) \mathcal{D}(\lambda_\rho) D ( λ ρ ) -valued representing pair. The rescaled PNG pair is Borel, Proposition 3.1 and Lemma 3.2 therefore give convergence in D ( λ ρ ) \mathcal{D}(\lambda_\rho) D ( λ ρ ) . The additional convergence at every prescribed finite parameter list is the separate conclusion of Proposition 3.1.
Convergence to the half-space KPZ fixed point
The transition limit requires separate critical and subcritical analyses of the boundary operators. In the critical window we prove uniform estimates for their Airy and first-hitting factors and pass to the Fredholm Pfaffian. At a fixed subcritical rate the boundary normalization changes, and stationary comparison supplies terminal wall observations. Localization and spatial tightness give profile limits; Markov iteration and short-time bounds then prove Theorem 1.11, following the transition-based approach of [23 Section 3] .
Use
δ = ε 1 / 2 , T = t δ − 3 , X i = x i δ − 2 , c δ = e ρ δ . \delta= \varepsilon^{1/2}, \qquad T = t\delta^{-3}, \qquad X_i = x_i\delta^{-2}, \qquad c_\delta= e^{\rho\delta}. δ = ε 1/2 , T = t δ − 3 , X i = x i δ − 2 , c δ = e ρ δ .
We index the operator estimates by δ \delta δ and the process estimates by ε = δ 2 \varepsilon=\delta^2 ε = δ 2 , with profiles reindexed by f ε = f ε f_\varepsilon=f_{\sqrt{\varepsilon}} f ε = f ε . We first treat finite ρ \rho ρ and then a fixed subcritical boundary rate. We use the Markov property and space–time continuity of the half-space KPZ fixed point from [17 Theorem 4.11] .
Critical-window transition probabilities
The boundary poles approach the cubic saddle under critical scaling. We control the corresponding residue terms and the first-hitting factors in weighted operator norms before passing to the Pfaffian.
Theorem 3.3 (Fixed-time transition limit for general initial data). Fix ρ ∈ R \rho\in\mathbb{R} ρ ∈ R and t > 0 t>0 t > 0 . Let h 0 , δ ∈ X a d m h_{0,\delta} \in\mathcal{X}_{\mathrm{adm}} h 0 , δ ∈ X adm be deterministic initial profiles, and put f δ ( s ) = δ h 0 , δ ( s δ − 2 ) f_\delta(s)=\delta h_{0,\delta}(s\delta^{-2}) f δ ( s ) = δ h 0 , δ ( s δ − 2 ) for s ≥ 0 s \ge0 s ≥ 0 . Suppose f δ → f ∈ U C R + f_\delta\to f \in\mathrm{UCR}_{+} f δ → f ∈ UCR + locally in hypograph and, for one finite constant C ≥ 0 C \ge0 C ≥ 0 ,
f δ ( s ) ≤ C ( 1 + s ) , s ≥ 0 , (63) f_\delta(s) \le C(1+s), \qquad s \ge0, \tag*{(63)} f δ ( s ) ≤ C ( 1 + s ) , s ≥ 0 , (63) for all sufficiently small δ \delta δ . Then, for every 0 ≤ x 1 < ⋯ < x m 0 \le x_1 < \cdots< x_m 0 ≤ x 1 < ⋯ < x m ,
( δ [ h ( c δ ) ( T , x i δ − 2 ; h 0 , δ ) − 2 T ] ) i = 1 m ⟹ ( h ρ ( t , x i ; f ) ) i = 1 m . (64) \left(\delta\left[h^{(c_\delta)}(T,x_i\delta^{-2};h_{0,\delta})-2T\right]\right)_{i=1}^{m}
\Longrightarrow
\left(h^\rho(t,x_i;f)\right)_{i=1}^{m}. \tag*{(64)} ( δ [ h ( c δ ) ( T , x i δ − 2 ; h 0 , δ ) − 2 T ] ) i = 1 m ⟹ ( h ρ ( t , x i ; f ) ) i = 1 m . (64) Here h ρ h^\rho h ρ is the half-space KPZ fixed point of [17 Theorem 4.11] ; its transition Pfaffian is Zhang’s formula [32 ] , displayed below in the convention of Section 1.4.
The transition formula and Airy factors. We first recall the symmetric kernel of [10 Theorem 1.1 and Section 1.3] . The free, hitting, and boundary factors encode propagation, the initial profile, and the wall, respectively. For this recall, fix c ≥ 0 c \ge0 c ≥ 0 , h ∈ X a d m h \in\mathcal{X}_{\mathrm{adm}} h ∈ X adm , T > 0 T>0 T > 0 , 0 ≤ X 1 < ⋯ < X m 0 \le X_1 < \cdots< X_m 0 ≤ X 1 < ⋯ < X m , and r j ∈ Z r_j \in\mathbb{Z} r j ∈ Z .
On functions on Z \mathbb{Z} Z , set
( τ ± f ) ( u ) = f ( u ± 1 ) , ∇ = 1 2 ( τ + − τ − ) , Δ = τ + + τ − − 2 I , (\tau_{\pm}f)(u)=f(u\pm1), \qquad\nabla=\tfrac{1}{2}(\tau_+-\tau_-), \qquad\Delta=\tau_+ + \tau_- - 2I, ( τ ± f ) ( u ) = f ( u ± 1 ) , ∇ = 2 1 ( τ + − τ − ) , Δ = τ + + τ − − 2 I ,
S T , X = e 2 T ∇ + X Δ , 1 r ( u , v ) = 1 { u > r } 1 { u = v } . S_{T,X}=e^{2T\nabla+X\Delta}, \qquad\mathbf{1}_r(u,v)=\mathbf{1}_{\{u>r\}}\mathbf{1}_{\{u=v\}}. S T , X = e 2 T ∇ + X Δ , 1 r ( u , v ) = 1 { u > r } 1 { u = v } .
Discrete kernel products use counting measure. The boundary kernel is
H c ( u , v ) = { 1 2 sgn ( v − u ) c ∣ v − u ∣ − 1 , u ≠ v , 0 , u = v , 0 0 = 1. H_c(u,v)=
\begin{cases}
\tfrac{1}{2}\operatorname{sgn}(v-u)c^{|v-u|-1}, & u\ne v,\\
0, & u=v,
\end{cases}
\qquad0^0=1. H c ( u , v ) = { 2 1 sgn ( v − u ) c ∣ v − u ∣ − 1 , 0 , u = v , u = v , 0 0 = 1.
For 0 ≤ c ≤ 1 0 \le c \le1 0 ≤ c ≤ 1 , it represents ( I − c τ + ) − 1 ∇ ( I − c τ − ) − 1 (I-c\tau_+)^{-1}\nabla(I-c\tau_-)^{-1} ( I − c τ + ) − 1 ∇ ( I − c τ − ) − 1 ; the displayed kernel continues it to every c ≥ 0 c \ge0 c ≥ 0 . The lower and upper one-sided inverse kernels are
H c , − − 1 = 2 ∑ k ≥ 0 τ − 2 k + 1 ( I − c τ + ) ( I − c τ − ) , H c , + − 1 = − 2 ∑ k ≥ 0 τ + 2 k + 1 ( I − c τ + ) ( I − c τ − ) . H_{c,-}^{-1}=2\sum_{k\ge0}\tau_-^{2k+1}(I-c\tau_+)(I-c\tau_-), \qquad
H_{c,+}^{-1}=-2\sum_{k\ge0}\tau_+^{2k+1}(I-c\tau_+)(I-c\tau_-). H c , − − 1 = 2 k ≥ 0 ∑ τ − 2 k + 1 ( I − c τ + ) ( I − c τ − ) , H c , + − 1 = − 2 k ≥ 0 ∑ τ + 2 k + 1 ( I − c τ + ) ( I − c τ − ) .
Their supports are u ≥ v u \ge v u ≥ v and u ≤ v u \le v u ≤ v , respectively, and H c ∗ = − H c H_c^*=-H_c H c ∗ = − H c , ( H c , − − 1 ) ∗ = − H c , + − 1 (H_{c,-}^{-1})^*=-H_{c,+}^{-1} ( H c , − − 1 ) ∗ = − H c , + − 1 . These one-sided kernels are used on the corresponding exponentially weighted spaces; grouped products for c ≥ 1 c \ge1 c ≥ 1 are interpreted by analytic continuation.
Let X X X be the continuous-time symmetric nearest-neighbor walk on Z \mathbb{Z} Z , with jumps + 1 +1 + 1 and − 1 -1 − 1 each occurring at rate one. Its generator is Δ \Delta Δ . We use its lower-semicontinuous version at jump times. Write P a , u \mathbb{P}_{a,u} P a , u for its law started from X ( a ) = u X(a)=u X ( a ) = u . For 0 ≤ a ≤ b 0 \le a \le b 0 ≤ a ≤ b , put
P a , b h ( u , v ) : = P a , u ( X ( y ) ≤ h ( y ) for some y ∈ [ a , b ] , X ( b ) = v ) , P^h_{a,b}(u,v):=\mathbb{P}_{a,u}\bigl(X(y)\le h(y)\text{ for some }y\in[a,b],\,X(b)=v\bigr), P a , b h ( u , v ) := P a , u ( X ( y ) ≤ h ( y ) for some y ∈ [ a , b ] , X ( b ) = v ) ,
Q a , b h : = e ( b − a ) Δ − P a , b h , Q^h_{a,b}:=e^{(b-a)\Delta}-P^h_{a,b}, Q a , b h := e ( b − a ) Δ − P a , b h ,
T a , b : = e a Δ P a , b h e − b Δ H c , − − 1 , T b : = T 0 , b . (65) T_{a,b}:=e^{a\Delta}P^h_{a,b}e^{-b\Delta}H_{c,-}^{-1}, \qquad T_b:=T_{0,b}. \tag*{(65)} T a , b := e a Δ P a , b h e − b Δ H c , − − 1 , T b := T 0 , b . (65)
Thus hitting is tested on the closed interval, and P a , a h \mathrm{P}^{h}_{a,a} P a , a h is multiplication by 1 { u ≤ h ( a ) } \mathbf{1}_{\{u\le h(a)\}} 1 { u ≤ h ( a )} . For an empty interval a > b ≥ 0 a>b\ge0 a > b ≥ 0 , set P a , b h = T a , b = 0 \mathrm{P}^{h}_{a,b}=\mathrm{T}_{a,b}=0 P a , b h = T a , b = 0 . For nonzero real p , q p,q p , q , define
W p , q : = { T q − T p ∗ + T p ∗ H c T q , p , q > 0 , T − p , q , p < 0 < q , − T − q , p ∗ , q < 0 < p , 0 , p , q < 0. \mathrm{W}_{p,q}:=
\begin{cases}
\mathrm{T}_{q}-\mathrm{T}_{p}^{*}+\mathrm{T}_{p}^{*}\mathrm{H}_{c}\mathrm{T}_{q}, & p,q>0,\\
\mathrm{T}_{-p,q}, & p<0<q,\\
-\mathrm{T}^{*}_{-q,p}, & q<0<p,\\
0, & p,q<0.
\end{cases} W p , q := ⎩ ⎨ ⎧ T q − T p ∗ + T p ∗ H c T q , T − p , q , − T − q , p ∗ , 0 , p , q > 0 , p < 0 < q , q < 0 < p , p , q < 0.
When T ≠ X j T\ne X_j T = X j for every j j j , the symmetric kernel uses
S j = ( S T , − X j 0 H c S T , X j S T , − X j ) , V i j = ( W T − X i , T − X j W T − X i , T + X j W T + X i , T − X j W T + X i , T + X j ) . (66) S_j=\begin{pmatrix}\mathrm{S}_{T,-X_j}&0\\ \mathrm{H}_{c}\mathrm{S}_{T,X_j}&\mathrm{S}_{T,-X_j}\end{pmatrix},
\qquad
V_{ij}=\begin{pmatrix}
\mathrm{W}_{T-X_i,T-X_j}&\mathrm{W}_{T-X_i,T+X_j}\\
\mathrm{W}_{T+X_i,T-X_j}&\mathrm{W}_{T+X_i,T+X_j}
\end{pmatrix}. \tag*{(66)} S j = ( S T , − X j H c S T , X j 0 S T , − X j ) , V ij = ( W T − X i , T − X j W T + X i , T − X j W T − X i , T + X j W T + X i , T + X j ) . (66)
Set J 0 = ( 0 1 − 1 0 ) J_0=\begin{pmatrix}0&1\\-1&0\end{pmatrix} J 0 = ( 0 − 1 1 0 ) . The projected kernel is K i j = R i j + K ~ i j K_{ij}=R_{ij}+\widetilde{K}_{ij} K ij = R ij + K ij , where
R i j = 1 r i ( 1 { j < i } e ( X i − X j ) Δ e X i Δ H c e X j Δ 0 1 { i < j } e ( X j − X i ) Δ ) 1 r j , R_{ij}=\mathbf{1}_{r_i}
\begin{pmatrix}
\mathbf{1}_{\{j<i\}}e^{(X_i-X_j)\Delta}&e^{X_i\Delta}\mathrm{H}_{c}e^{X_j\Delta}\\
0&\mathbf{1}_{\{i<j\}}e^{(X_j-X_i)\Delta}
\end{pmatrix}
\mathbf{1}_{r_j}, R ij = 1 r i ( 1 { j < i } e ( X i − X j ) Δ 0 e X i Δ H c e X j Δ 1 { i < j } e ( X j − X i ) Δ ) 1 r j ,
K ~ i j = 1 r i J 0 S i ∗ V i j S j J 0 1 r j . \widetilde{K}_{ij}=\mathbf{1}_{r_i}J_0S_i^{*}V_{ij}S_jJ_0\mathbf{1}_{r_j}. K ij = 1 r i J 0 S i ∗ V ij S j J 0 1 r j .
Here the adjoint includes height variables and internal components. Let J m J_m J m act by J 0 J_0 J 0 at each observation-height pair. The identity W p , q ∗ = − W q , p \mathrm{W}_{p,q}^{*}=-\mathrm{W}_{q,p} W p , q ∗ = − W q , p gives ( J m K ) ∗ = − J m K (J_mK)^{*}=-J_mK ( J m K ) ∗ = − J m K . Then
P h ( c ) ( h ( c ) ( T , X j ) ≤ r j , 1 ≤ j ≤ m ) = Pf ( J m + J m K ) \mathbb{P}^{(c)}_{h}\left(h^{(c)}(T,X_j)\le r_j,\ 1\le j\le m\right)
=\operatorname{Pf}(J_m+J_mK) P h ( c ) ( h ( c ) ( T , X j ) ≤ r j , 1 ≤ j ≤ m ) = Pf ( J m + J m K )
on ⨁ j = 1 m ℓ 2 ( Z ) ⊗ C 2 \bigoplus_{j=1}^{m}\ell^{2}(\mathbb{Z})\otimes\mathbb{C}^{2} ⨁ j = 1 m ℓ 2 ( Z ) ⊗ C 2 , with Pf ( J m ) = 1 \operatorname{Pf}(J_m)=1 Pf ( J m ) = 1 . The exponentially weighted formulation is justified by [10 Proposition 1.2] . For sufficiently small δ \delta δ , our scaling has T > X m T>X_m T > X m . The two hitting compositions that we will estimate directly are
T i − : = T T − X i S T , − X i , T i + : = T T + X i H c S T , X i . \mathrm{T}_{i}^{-}:=\mathrm{T}_{T-X_i}\mathrm{S}_{T,-X_i},
\qquad
\mathrm{T}_{i}^{+}:=\mathrm{T}_{T+X_i}\mathrm{H}_{c}\mathrm{S}_{T,X_i}. T i − := T T − X i S T , − X i , T i + := T T + X i H c S T , X i .
The limiting half-space KPZ kernel. We state Zhang’s limiting transition kernel [32 ] 1 in the conventions used here. Fix t > 0 t>0 t > 0 , ρ ∈ R \rho\in\mathbb{R} ρ ∈ R , and an upper-semicontinuous h 0 ≢ − ∞ h_0\not\equiv-\infty h 0 ≡ − ∞ with a linear upper bound. Here 0 ≤ y 1 < ⋯ < y m 0\le y_1<\cdots<y_m 0 ≤ y 1 < ⋯ < y m and s 1 , … , s m ∈ R s_1,\ldots,s_m\in\mathbb{R} s 1 , … , s m ∈ R . The parameter and variable correspondence with our PNG scaling is
ϱ = − ρ , h 0 = f , y j = x j , s j = r j . \varrho=-\rho,\qquad h_0=f,\qquad y_j=x_j,\qquad s_j=r_j. ϱ = − ρ , h 0 = f , y j = x j , s j = r j .
The corresponding microscopic height thresholds are ⌊ 2 T + δ − 1 s j ⌋ \lfloor2T+\delta^{-1}s_j\rfloor ⌊ 2 T + δ − 1 s j ⌋ . The affine boundary parameter of [10 Section 1.3] is δ − 1 ( 1 − c δ ) → ϱ \delta^{-1}(1-c\delta)\to\varrho δ − 1 ( 1 − cδ ) → ϱ . The sign change in the boundary parameter agrees with the convention of [17 Section 4.2] .
Let D = ∂ u D=\partial_u D = ∂ u , and let C a π / 3 C_a^{\pi/3} C a π /3 be the two Airy rays a + r e ± i π / 3 a+re^{\pm i\pi/3} a + r e ± iπ /3 , oriented from the lower ray to the upper ray. For t > 0 t>0 t > 0 , define
S t , x ( u , v ) = 1 2 π i ∫ C a π / 3 e t w 3 / 3 + x w 2 + ( u − v ) w d w , a > 0. \mathrm{S}^{t,x}(u,v)=\frac{1}{2\pi i}\int_{C_a^{\pi/3}}e^{tw^3/3+xw^2+(u-v)w}\,dw,\qquad a>0. S t , x ( u , v ) = 2 π i 1 ∫ C a π /3 e t w 3 /3 + x w 2 + ( u − v ) w d w , a > 0.
Write H = H ϱ \mathrm{H}=\mathrm{H}_{\varrho} H = H ϱ for the boundary kernel and use the lower-triangular inverse convention
H ( u , v ) = 1 2 sgn ( v − u ) e − ϱ ∣ u − v ∣ , \mathrm{H}(u,v)=\frac{1}{2}\operatorname{sgn}(v-u)e^{-\varrho|u-v|}, H ( u , v ) = 2 1 sgn ( v − u ) e − ϱ ∣ u − v ∣ ,
H − 1 = D − 1 ( ϱ − D ) ( ϱ + D ) , D − 1 ( u , v ) = 1 { u > v } . \mathrm{H}^{-1}=D^{-1}(\varrho-D)(\varrho+D),\qquad D^{-1}(u,v)=\mathbf{1}_{\{u>v\}}. H − 1 = D − 1 ( ϱ − D ) ( ϱ + D ) , D − 1 ( u , v ) = 1 { u > v } .
In particular, the inverse Airy factor has the kernel
( H − 1 S ) t , x ( u , v ) = 1 2 π i ∫ a + i R ϱ 2 − w 2 w e t w 3 / 3 + x w 2 + ( u − v ) w d w , a > max ( 0 , − x / t ) . (\mathrm{H}^{-1}\mathrm{S})^{t,x}(u,v)
=\frac{1}{2\pi i}\int_{a+i\mathbb{R}}\frac{\varrho^2-w^2}{w}e^{tw^3/3+xw^2+(u-v)w}\,dw,
\qquad a>\max(0,-x/t). ( H − 1 S ) t , x ( u , v ) = 2 π i 1 ∫ a + i R w ϱ 2 − w 2 e t w 3 /3 + x w 2 + ( u − v ) w d w , a > max ( 0 , − x / t ) .
For the boundary-modified Airy factor, let C ϱ C_{\varrho} C ϱ be a clockwise loop enclosing only w = ϱ w=\varrho w = ϱ , and set
S H t , x ( u , v ) = 1 2 π i ∫ C π / 3 a + ( 1 − 1 2 1 { ϱ = 0 } ) C ϱ w e t w 3 / 3 + x w 2 + ( u − v ) w ( ϱ + w ) ( ϱ − w ) d w , a > ∣ ϱ ∣ . \mathrm{S}_{\mathrm{H}}^{t,x}(u,v)
=\frac{1}{2\pi i}\int_{C_{\pi/3}^{a}+(1-\frac{1}{2}\mathbf{1}_{\{\varrho=0\}})C_{\varrho}}
\frac{w e^{tw^3/3+xw^2+(u-v)w}}{(\varrho+w)(\varrho-w)}\,dw,
\qquad a>|\varrho|. S H t , x ( u , v ) = 2 π i 1 ∫ C π /3 a + ( 1 − 2 1 1 { ϱ = 0 } ) C ϱ ( ϱ + w ) ( ϱ − w ) w e t w 3 /3 + x w 2 + ( u − v ) w d w , a > ∣ ϱ ∣.
The loop has half weight when ϱ = 0 \varrho=0 ϱ = 0 . Equivalently,
S H t , x ( u , v ) = 1 2 π i ∫ a + i R w ϱ 2 − w 2 e t w 3 / 3 + x w 2 + ( u − v ) w d w + 1 2 e t ϱ 3 / 3 + x ϱ 2 + ϱ ( u − v ) , a > ∣ ϱ ∣ + ∣ x ∣ / t . (67) \mathrm{S}_{\mathrm{H}}^{t,x}(u,v)
=\frac{1}{2\pi i}\int_{a+i\mathbb{R}}\frac{w}{\varrho^2-w^2}e^{tw^3/3+xw^2+(u-v)w}\,dw
+\frac{1}{2}e^{t\varrho^3/3+x\varrho^2+\varrho(u-v)},
\qquad a>|\varrho|+|x|/t. \tag*{(67)} S H t , x ( u , v ) = 2 π i 1 ∫ a + i R ϱ 2 − w 2 w e t w 3 /3 + x w 2 + ( u − v ) w d w + 2 1 e t ϱ 3 /3 + x ϱ 2 + ϱ ( u − v ) , a > ∣ ϱ ∣ + ∣ x ∣/ t . (67)
At ϱ = 0 \varrho=0 ϱ = 0 the last term is 1 / 2 1/2 1/2 . The notation S H \mathrm{S}_{\mathrm{H}} S H always uses this contour prescription; for nonpositive ϱ \varrho ϱ , it does not assert an absolutely convergent convolution on unweighted L 2 L^2 L 2 .
The initial profile enters through Brownian first hitting. Let B B B be Brownian motion of variance 2 2 2 , started at u u u , and put τ = inf { s ≥ 0 : B ( s ) ≤ h 0 ( s ) } \tau=\inf\{s\ge0:B(s)\le h_0(s)\} τ = inf { s ≥ 0 : B ( s ) ≤ h 0 ( s )} . Define
S h y p o ( h 0 ) , t , x ( u , v ) : = E u [ S t , x − τ ( B ( τ ) , v ) ; τ < ∞ ] , \mathrm{S}^{\mathrm{hypo}(h_0),t,x}(u,v):=\mathbb{E}_u\left[\mathrm{S}^{t,x-\tau}(B(\tau),v);\ \tau<\infty\right], S hypo ( h 0 ) , t , x ( u , v ) := E u [ S t , x − τ ( B ( τ ) , v ) ; τ < ∞ ] ,
( H − 1 S ) h y p o ( h 0 ) , t , x ( u , v ) : = E u [ ( H − 1 S ) t , x − τ ( B ( τ ) , v ) ; τ < ∞ ] . (\mathrm{H}^{-1}\mathrm{S})^{\mathrm{hypo}(h_0),t,x}(u,v):=\mathbb{E}_u\left[(\mathrm{H}^{-1}\mathrm{S})^{t,x-\tau}(B(\tau),v);\ \tau<\infty\right]. ( H − 1 S ) hypo ( h 0 ) , t , x ( u , v ) := E u [ ( H − 1 S ) t , x − τ ( B ( τ ) , v ) ; τ < ∞ ] .
Let U s = e s D U_s=e^{sD} U s = e sD , so ( U s ϕ ) ( u ) = ϕ ( u + s ) (U_s\phi)(u)=\phi(u+s) ( U s ϕ ) ( u ) = ϕ ( u + s ) . Set
F − j = S t , − y j U − s j , F j = − S H t , y j U − s j , F_{-j}=\mathrm{S}^{t,-y_j}U_{-s_j},\qquad F_j=-\mathrm{S}_{\mathrm{H}}^{t,y_j}U_{-s_j}, F − j = S t , − y j U − s j , F j = − S H t , y j U − s j ,
G − j = − ( H − 1 S ) h y p o ( h 0 ) , t , − y j U − s j , G j = S h y p o ( h 0 ) , t , y j U − s j . G_{-j}=-(\mathrm{H}^{-1}\mathrm{S})^{\mathrm{hypo}(h_0),t,-y_j}U_{-s_j},\qquad
G_j=\mathrm{S}^{\mathrm{hypo}(h_0),t,y_j}U_{-s_j}. G − j = − ( H − 1 S ) hypo ( h 0 ) , t , − y j U − s j , G j = S hypo ( h 0 ) , t , y j U − s j .
and collect the four factors in
S j : = ( F − j F j G − j G j ) . \mathcal{S}_j:=\begin{pmatrix} F_{-j} & F_j \\ G_{-j} & G_j \end{pmatrix}. S j := ( F − j G − j F j G j ) .
The translations move the height thresholds s j s_j s j to zero. The complete limiting kernel is
K i j f p = R i j f p + K ~ i j f p , K ~ i j f p = J 0 S i ∗ J 0 ( I H 0 I ) S j , K_{ij}^{\mathrm{fp}}=R_{ij}^{\mathrm{fp}}+\widetilde{K}_{ij}^{\mathrm{fp}},\qquad\widetilde{K}_{ij}^{\mathrm{fp}}=J_0S_i^*J_0\begin{pmatrix} I & H \\ 0 & I \end{pmatrix}\mathcal{S}_j, K ij fp = R ij fp + K ij fp , K ij fp = J 0 S i ∗ J 0 ( I 0 H I ) S j ,
where
R i j f p = ( 1 { j < i } U s i e ( y i − y j ) D 2 U − s j − U s i e y i D 2 H e y j D 2 U − s j 0 1 { i < j } U s i e ( y j − y i ) D 2 U − s j ) . R_{ij}^{\mathrm{fp}}=\begin{pmatrix} \mathbb{1}_{\{j<i\}}U_{s_i}e^{(y_i-y_j)D^2}U_{-s_j} & -U_{s_i}e^{y_iD^2}He^{y_jD^2}U_{-s_j} \\ 0 & \mathbb{1}_{\{i<j\}}U_{s_i}e^{(y_j-y_i)D^2}U_{-s_j} \end{pmatrix}. R ij fp = ( 1 { j < i } U s i e ( y i − y j ) D 2 U − s j 0 − U s i e y i D 2 H e y j D 2 U − s j 1 { i < j } U s i e ( y j − y i ) D 2 U − s j ) .
The adjoint includes the height variable and internal matrix indices; the final height variables lie in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . With J m J_m J m acting by J 0 J_0 J 0 at each observation-height pair, the transition formula is
P ( h ρ ( t , y j ; h 0 ) ≤ s j , 1 ≤ j ≤ m ) = Pf ( J m + J m K f p ) L 2 ( { 1 , … , m } × [ 0 , ∞ ) ; C 2 ) . \mathbb{P}\bigl(h^\rho(t,y_j;h_0)\leq s_j,\ 1\leq j\leq m\bigr)=\operatorname{Pf}\bigl(J_m+J_mK^{\mathrm{fp}}\bigr)_{L^2(\{1,\ldots,m\}\times[0,\infty);\mathbb{C}^2)}. P ( h ρ ( t , y j ; h 0 ) ≤ s j , 1 ≤ j ≤ m ) = Pf ( J m + J m K fp ) L 2 ({ 1 , … , m } × [ 0 , ∞ ) ; C 2 ) .
We use Zhang’s Pfaffian transition laws [32 ] together with their realization as the continuous Markov process h ρ h^\rho h ρ in [17 Theorem 4.11] .
Lemma 3.4 (Uniform Airy-factor limits). Put n δ = ⌊ 2 T ⌋ n_\delta=\lfloor2T\rfloor n δ = ⌊ 2 T ⌋ and, for u , v ∈ R u,v\in\mathbb{R} u , v ∈ R , put u δ = ⌊ u / δ ⌋ u_\delta=\lfloor u/\delta\rfloor u δ = ⌊ u / δ ⌋ and v δ = n δ + ⌊ v / δ ⌋ v_\delta=n_\delta+\lfloor v/\delta\rfloor v δ = n δ + ⌊ v / δ ⌋ . For x x x in a fixed compact set, the following limits are locally uniform:
δ − 1 S T , x δ − 2 ( u δ , v δ ) ⟶ S t , x ( u , v ) , ( H c δ S T , x δ − 2 ) ( u δ , v δ ) ⟶ S H t , x ( u , v ) , δ − 2 ( H c δ − 1 S T , x δ − 2 ) ( u δ , v δ ) ⟶ ( H − 1 S ) t , x ( u , v ) . (68) \begin{aligned}
\delta^{-1}S_{T,x\delta^{-2}}(u_\delta,v_\delta)&\longrightarrow S^{t,x}(u,v),\\
(H_{c_\delta}S_{T,x\delta^{-2}})(u_\delta,v_\delta)&\longrightarrow S_{\mathrm{H}}^{t,x}(u,v),\\
\delta^{-2}(H_{c_\delta}^{-1}S_{T,x\delta^{-2}})(u_\delta,v_\delta)&\longrightarrow(H^{-1}S)^{t,x}(u,v). \tag*{(68)}
\end{aligned} δ − 1 S T , x δ − 2 ( u δ , v δ ) ( H c δ S T , x δ − 2 ) ( u δ , v δ ) δ − 2 ( H c δ − 1 S T , x δ − 2 ) ( u δ , v δ ) ⟶ S t , x ( u , v ) , ⟶ S H t , x ( u , v ) , ⟶ ( H − 1 S ) t , x ( u , v ) . (68) If the terminal variable is restricted to v ≥ r v\geq r v ≥ r , then each free or inverse factor, multiplied on the right by e η v e^{\eta v} e η v , has row L 2 L^2 L 2 norm at most C a e a u C_a e^{au} C a e a u , for every sufficiently large fixed a > η a>\eta a > η . The factor involving the boundary kernel multiplied on the right by e − θ v e^{-\theta v} e − θ v has row norm at most C a ( e a u + e ϱ u ) C_a(e^{au}+e^{\varrho u}) C a ( e a u + e ϱ u ) when θ > ∣ ϱ ∣ \theta>|\varrho| θ > ∣ ϱ ∣ . These bounds are uniform for sufficiently small δ \delta δ , for the limits, and for r r r in a compact set. The corresponding weighted rows converge in L 2 L^2 L 2 for each fixed u u u .
Proof. The Laurent symbol of S T , X S_{T,X} S T , X is exp { T ( z − z − 1 ) + X ( z + z − 1 − 2 ) } \exp\{T(z-z^{-1})+X(z+z^{-1}-2)\} exp { T ( z − z − 1 ) + X ( z + z − 1 − 2 )} . On the circle z = e δ ( a + i ξ ) z=e^{\delta(a+i\xi)} z = e δ ( a + i ξ ) , ∣ ξ ∣ ≤ π / δ |\xi|\leq\pi/\delta ∣ ξ ∣ ≤ π / δ , its centered exponent is
E δ ( w ) = 2 T sinh ( δ w ) + 2 x δ − 2 ( cosh ( δ w ) − 1 ) + ( u δ − v δ ) δ w . E_\delta(w)=2T\sinh(\delta w)+2x\delta^{-2}(\cosh(\delta w)-1)+(u_\delta-v_\delta)\delta w. E δ ( w ) = 2 T sinh ( δ w ) + 2 x δ − 2 ( cosh ( δ w ) − 1 ) + ( u δ − v δ ) δ w .
It converges locally to t w 3 / 3 + x w 2 + ( u − v ) w tw^3/3+xw^2+(u-v)w t w 3 /3 + x w 2 + ( u − v ) w . Moreover,
Re ( E δ ( a + i ξ ) − E δ ( a ) ) = − 2 [ T sinh ( δ a ) + x δ − 2 cosh ( δ a ) ] ( 1 − cos ( δ ξ ) ) ≤ − κ ξ 2 \operatorname{Re}(E_\delta(a+i\xi)-E_\delta(a))=-2\left[T\sinh(\delta a)+x\delta^{-2}\cosh(\delta a)\right](1-\cos(\delta\xi))\leq-\kappa\xi^2 Re ( E δ ( a + i ξ ) − E δ ( a )) = − 2 [ T sinh ( δ a ) + x δ − 2 cosh ( δ a ) ] ( 1 − cos ( δ ξ )) ≤ − κ ξ 2
for a a a sufficiently large, with κ > 0 \kappa>0 κ > 0 independent of small δ \delta δ . Also Re E δ ( a ) ≤ C a + a ( u − v ) \operatorname{Re}E_\delta(a)\leq C_a+a(u-v) Re E δ ( a ) ≤ C a + a ( u − v ) . The factor δ − 1 \delta^{-1} δ − 1 in (3.6) cancels the Jacobian d z / z = δ d w dz/z=\delta\,dw d z / z = δ d w . Dominated convergence proves the first limit and the bound C a e a ( u − v ) C_a e^{a(u-v)} C a e a ( u − v ) .
The symbols of the boundary operator and its lower-triangular inverse are, respectively,
h c ( z ) = z − z − 1 2 ( 1 − c z ) ( 1 − c z − 1 ) , u c ( z ) = 2 z − 1 ( 1 − c z ) ( 1 − c z − 1 ) 1 − z − 2 . h_c(z)=\frac{z-z^{-1}}{2(1-cz)(1-cz^{-1})},\qquad u_c(z)=\frac{2z^{-1}(1-cz)(1-cz^{-1})}{1-z^{-2}}. h c ( z ) = 2 ( 1 − cz ) ( 1 − c z − 1 ) z − z − 1 , u c ( z ) = 1 − z − 2 2 z − 1 ( 1 − cz ) ( 1 − c z − 1 ) .
On the same contour,
δ h c δ ( e δ w ) ⟶ ϱ 2 − w 2 ϱ 2 − w 2 , δ − 1 u c δ ( e δ w ) ⟶ ϱ 2 − w 2 ϱ 2 − w 2 . \delta h_{c_\delta}(e^{\delta w})\longrightarrow\frac{\varrho^2-w^2}{\varrho^2-w^2},\qquad\delta^{-1}u_{c_\delta}(e^{\delta w})\longrightarrow\frac{\varrho^2-w^2}{\varrho^2-w^2}. δ h c δ ( e δ w ) ⟶ ϱ 2 − w 2 ϱ 2 − w 2 , δ − 1 u c δ ( e δ w ) ⟶ ϱ 2 − w 2 ϱ 2 − w 2 .
For a > ∣ ϱ ∣ a>|\varrho| a > ∣ ϱ ∣ , the first multiplier is bounded uniformly on the contour. The second is bounded by C ( 1 + ∣ ξ ∣ 2 ) C(1+|\xi|^2) C ( 1 + ∣ ξ ∣ 2 ) : on ∣ δ ξ ∣ ≤ π / 2 |\delta\xi|\leq\pi/2 ∣ δ ξ ∣ ≤ π /2 this follows by expanding the factors at 1 1 1 ; on the complementary arc the denominator has modulus at least 1 − e − 2 δ a 1-e^{-2\delta a} 1 − e − 2 δ a and the bound is absorbed by 1 + ∣ ξ ∣ 2 1+|\xi|^2 1 + ∣ ξ ∣ 2 . The Gaussian bound therefore applies to both integrals.
For c ≠ 0 c\neq0 c = 0 , put
F δ ( c ) : = e T ( c − 1 − c ) + X ( c + c − 1 − 2 ) c v δ − u δ , X = x δ − 2 . F_\delta(c):=e^{T(c^{-1}-c)+X(c+c^{-1}-2)c^{v_\delta-u_\delta}},\qquad X=x\delta^{-2}. F δ ( c ) := e T ( c − 1 − c ) + X ( c + c − 1 − 2 ) c v δ − u δ , X = x δ − 2 .
For 0 < c < 1 0<c<1 0 < c < 1 , the Laurent integrand at the selected pole satisfies
Res z = c − 1 [ h c ( z ) e T ( z − z − 1 ) + X ( z + z − 1 − 2 ) z v δ − u δ 1 z ] = − F δ ( c ) 2 c . \mathop{\operatorname{Res}}_{z=c^{-1}}\left[h_c(z)e^{T(z-z^{-1})+X(z+z^{-1}-2)}z^{v_\delta-u_\delta}\frac{1}{z}\right]=-\frac{F_\delta(c)}{2c}. Res z = c − 1 [ h c ( z ) e T ( z − z − 1 ) + X ( z + z − 1 − 2 ) z v δ − u δ z 1 ] = − 2 c F δ ( c ) .
Moving the positively oriented circle past this pole therefore adds F δ ( c ) / ( 2 c ) F_\delta(c)/(2c) F δ ( c ) / ( 2 c ) to the centered boundary factor. For fixed T , X T,X T , X , expansion of S T , X = e − 2 X e ( T + X ) τ + + e ( X − T ) τ − S_{T,X}=e^{-2X}e^{(T+X)\tau_++e^{(X-T)\tau_-}} S T , X = e − 2 X e ( T + X ) τ + + e ( X − T ) τ − gives ∑ n ∣ S T , X ( 0 , n ) ∣ R ∣ n ∣ < ∞ \sum_n|S_{T,X}(0,n)|R^{|n|}<\infty ∑ n ∣ S T , X ( 0 , n ) ∣ R ∣ n ∣ < ∞ for every R ≥ 1 R\geq1 R ≥ 1 . Thus the explicit convolution H c S T , X H_cS_{T,X} H c S T , X is entrywise entire in c c c , with normal convergence on compact c c c -sets. For the fixed contour radius R = e a d > 1 R=e^{ad}>1 R = e a d > 1 , the contour integral plus the displayed residue term is holomorphic on R − 1 < ∣ c ∣ < R R^{-1}<|c|<R R − 1 < ∣ c ∣ < R . Equality for R − 1 < c < 1 R^{-1}<c<1 R − 1 < c < 1 extends to c = c δ c=c_\delta c = c δ when a > ∣ ϱ ∣ a>|\varrho| a > ∣ ϱ ∣ , including c = 1 c=1 c = 1 . This is a fixed-δ \delta δ continuation; the preceding contour estimates supply the bounds uniform in small δ \delta δ . At c = c δ c=c_\delta c = c δ , F δ ( c δ ) = e E δ ( ϱ ) F_\delta(c_\delta)=e^{E_\delta(\varrho)} F δ ( c δ ) = e E δ ( ϱ ) . The added term therefore converges to the last term of (3.5); its absolute value is bounded by C e ϱ ( u − v ) Ce^{\varrho(u-v)} C e ϱ ( u − v ) . This proves the second limit, including the residue of the boundary pole. Integration against the indicated exponential weights gives the row bounds and, by dominated convergence, weighted row convergence.
First-hitting measures and uniform estimates. The initial profile enters through the joint time and height of first contact. We prove convergence of these hitting measures under local hypograph convergence and check the walk’s state at a contact occurring at a jump. Weighted row estimates at arbitrary hitting positions remove compact cutoffs using the common linear upper bound. Write M β M_\beta M β for multiplication by e β u e^{\beta u} e β u and b ± = max ( ± b , 0 ) b^{\pm}=\max(\pm b,0) b ± = max ( ± b , 0 ) for the positive and negative parts.
Lemma 3.5 (Uniform kernel estimates at arbitrary hitting positions). Fix t > 0 t>0 t > 0 , a compact set of x x x values, a lower bound r r r for the terminal height, and a finite C ≥ 0 C\geq0 C ≥ 0 . Put n δ = ⌊ 2 T ⌋ n_\delta=\lfloor2T\rfloor n δ = ⌊ 2 T ⌋ , b δ = ⌊ b / δ ⌋ b_\delta=\lfloor b/\delta\rfloor b δ = ⌊ b / δ ⌋ , and v δ = n δ + ⌊ v / δ ⌋ v_\delta=n_\delta+\lfloor v/\delta\rfloor v δ = n δ + ⌊ v / δ ⌋ . For 0 ≤ s ≤ t δ − 1 + x 0\leq s\leq t\delta^{-1}+x 0 ≤ s ≤ t δ − 1 + x , define
g δ , + x ( s , b ; v ) = δ − 1 S T , ( x − s ) δ − 2 ( b δ , v δ ) , g δ , − x ( s , b ; v ) = δ − 2 ( H c δ , − − 1 S T , ( x − s ) δ − 2 ) ( b δ , v δ ) . g^{x}_{\delta,+}(s,b;v)=\delta^{-1}S_{T,(x-s)\delta^{-2}}(b_\delta,v_\delta),\qquad
g^{x}_{\delta,-}(s,b;v)=\delta^{-2}(H^{-1}_{c_\delta,-}S_{T,(x-s)\delta^{-2}})(b_\delta,v_\delta). g δ , + x ( s , b ; v ) = δ − 1 S T , ( x − s ) δ − 2 ( b δ , v δ ) , g δ , − x ( s , b ; v ) = δ − 2 ( H c δ , − − 1 S T , ( x − s ) δ − 2 ) ( b δ , v δ ) . Extend these rows by zero when s > t δ − 1 + x s>t\delta^{-1}+x s > t δ − 1 + x . For every μ > 0 \mu>0 μ > 0 and every fixed terminal exponential weight, there are c , K , δ 0 > 0 c,K,\delta_0>0 c , K , δ 0 > 0 such that both prescribed weighted row L 2 ( [ r , ∞ ) ) L^2([r,\infty)) L 2 ([ r , ∞ )) norms satisfy
∥ g δ , ± x ( s , b ; ⋅ ) M ± θ ∥ 2 ≤ K e − c s 3 − μ b − , b ≤ C ( 1 + s ) , 0 < δ < δ 0 . (69) \lVert g^{x}_{\delta,\pm}(s,b;\cdot)M_{\pm\theta}\rVert_2\leq Ke^{-cs^3-\mu b^{-}},\qquad b\leq C(1+s),\quad0<\delta<\delta_0. \tag*{(69)} ∥ g δ , ± x ( s , b ; ⋅ ) M ± θ ∥ 2 ≤ K e − c s 3 − μ b − , b ≤ C ( 1 + s ) , 0 < δ < δ 0 . (69) Either sign of a fixed weight can be used for either row. The limiting rows S t , x − s ( b , ⋅ ) S^{t,x-s}(b,\cdot) S t , x − s ( b , ⋅ ) and ( H − 1 S ) t , x − s ( b , ⋅ ) (H^{-1}S)^{t,x-s}(b,\cdot) ( H − 1 S ) t , x − s ( b , ⋅ ) satisfy the same bound. On compact sets of ( s , b ) (s,b) ( s , b ) the weighted microscopic rows converge uniformly in L 2 L^2 L 2 to these limiting rows.
Proof. Choose κ \kappa κ larger than μ \mu μ , all absolute terminal weight parameters, ∣ ρ ∣ |\rho| ∣ ρ ∣ , and 1. The centered Laurent phase on z = e η + i δ ξ z=e^{\eta+i\delta\xi} z = e η + i δ ξ , where a = η / δ a=\eta/\delta a = η / δ , is
E δ ( a + i ξ ) = 2 t δ − 3 { sinh ( δ ( a + i ξ ) ) − δ ( a + i ξ ) } + 2 ( x − s ) δ − 2 { cosh ( δ ( a + i ξ ) ) − 1 } + ( b − v + γ δ δ ) ( a + i ξ ) , ∣ γ δ ∣ ≤ 3. E_\delta(a+i\xi)=2t\delta^{-3}\{\sinh(\delta(a+i\xi))-\delta(a+i\xi)\}+2(x-s)\delta^{-2}\{\cosh(\delta(a+i\xi))-1\}+(b-v+\gamma_\delta\delta)(a+i\xi),\qquad|\gamma_\delta|\leq3. E δ ( a + i ξ ) = 2 t δ − 3 { sinh ( δ ( a + i ξ )) − δ ( a + i ξ )} + 2 ( x − s ) δ − 2 { cosh ( δ ( a + i ξ )) − 1 } + ( b − v + γ δ δ ) ( a + i ξ ) , ∣ γ δ ∣ ≤ 3.
The bounded correction γ δ \gamma_\delta γ δ records the height rounding; its contribution at the positive real point is O ( η ) O(\eta) O ( η ) . Put S = s − x S=s-x S = s − x . If S ≥ 0 S\geq0 S ≥ 0 and r δ = δ S / t < 1 r_\delta=\delta S/t<1 r δ = δ S / t < 1 , choose
η = max { κ δ , 2 arctanh r δ } . (70) \eta=\max\{\kappa\delta,2\operatorname{arctanh}r_\delta\}. \tag*{(70)} η = max { κ δ , 2 arctanh r δ } . (70)
If S < 0 S<0 S < 0 , choose η = κ δ \eta=\kappa\delta η = κ δ . On the adaptive part of the contour one has the exact identities
2 t δ − 3 ( sinh η − η ) − 2 S δ − 2 ( cosh η − 1 ) = 4 t δ − 3 ( r δ − arctanh r δ ) ≤ − 4 S 3 3 t 2 , (71) 2t\delta^{-3}(\sinh\eta-\eta)-2S\delta^{-2}(\cosh\eta-1)=4t\delta^{-3}(r_\delta-\operatorname{arctanh}r_\delta)\leq-\frac{4S^3}{3t^2}, \tag*{(71)} 2 t δ − 3 ( sinh η − η ) − 2 S δ − 2 ( cosh η − 1 ) = 4 t δ − 3 ( r δ − arctanh r δ ) ≤ − 3 t 2 4 S 3 , (71)
t δ − 3 sinh η − S δ − 2 cosh η = S δ − 2 . (72) t\delta^{-3}\sinh\eta-S\delta^{-2}\cosh\eta=S\delta^{-2}. \tag*{(72)} t δ − 3 sinh η − S δ − 2 cosh η = S δ − 2 . (72)
On the part where η = κ δ \eta=\kappa\delta η = κ δ , the choice in (3.8) implies S ≤ t δ − 1 tanh ( κ δ / 2 ) S\leq t\delta^{-1}\tanh(\kappa\delta/2) S ≤ t δ − 1 tanh ( κ δ /2 ) when S ≥ 0 S\geq0 S ≥ 0 . Consequently the coefficient on the left of (3.10) is at least t δ − 3 tanh ( κ δ / 2 ) t\delta^{-3}\tanh(\kappa\delta/2) t δ − 3 tanh ( κ δ /2 ) . It has a uniform positive multiple of δ − 2 \delta^{-2} δ − 2 as a lower bound in both cases. Thus, for ∣ ξ ∣ ≤ π / δ |\xi|\leq\pi/\delta ∣ ξ ∣ ≤ π / δ ,
Re ( E δ ( a + i ξ ) − E δ ( a ) ) = − 2 { t δ − 3 sinh η − S δ − 2 cosh η } ( 1 − cos δ ξ ) ≤ − k ξ 2 (73) \operatorname{Re}(E_\delta(a+i\xi)-E_\delta(a))=-2\{t\delta^{-3}\sinh\eta-S\delta^{-2}\cosh\eta\}(1-\cos\delta\xi)\leq-k\xi^2 \tag*{(73)} Re ( E δ ( a + i ξ ) − E δ ( a )) = − 2 { t δ − 3 sinh η − S δ − 2 cosh η } ( 1 − cos δ ξ ) ≤ − k ξ 2 (73)
with a fixed k > 0 k>0 k > 0 .
The linear upper bound is absorbed by the cubic decay, including near the causal endpoint. Since a ≥ κ a\geq\kappa a ≥ κ , b ≤ C ( 1 + s ) b\leq C(1+s) b ≤ C ( 1 + s ) and v ≥ r v\geq r v ≥ r give
a ( b − v ) ≤ a { C ( 1 + s ) + ∣ r − ∣ } − μ b − − κ v . a(b-v)\leq a\{C(1+s)+|r^{-}|\}-\mu b^{-}-\kappa v. a ( b − v ) ≤ a { C ( 1 + s ) + ∣ r − ∣ } − μ b − − κ v .
For r δ ≤ 1 / 2 r_\delta\leq1/2 r δ ≤ 1/2 one has a ≤ K ( 1 + s ) a\leq K(1+s) a ≤ K ( 1 + s ) , so the additional positive term is quadratic in s s s and is absorbed by the negative cubic in (3.9); the region where a = κ a=\kappa a = κ is bounded in s s s . For r δ > 1 / 2 r_\delta>1/2 r δ > 1/2 , there is an absolute k 1 > 0 k_1>0 k 1 > 0 such that arctanh r δ − r δ ≥ k 1 arctanh r δ \operatorname{arctanh}r_\delta-r_\delta\geq k_1\operatorname{arctanh}r_\delta arctanh r δ − r δ ≥ k 1 arctanh r δ . Here s ≤ t δ − 1 + x s\leq t\delta^{-1}+x s ≤ t δ − 1 + x and a ≥ 2 δ − 1 arctanh r δ a\geq2\delta^{-1}\operatorname{arctanh}r_\delta a ≥ 2 δ − 1 arctanh r δ . The contribution from the linear upper bound and the rounding term are bounded by K δ − 2 arctanh r δ K\delta^{-2}\operatorname{arctanh}r_\delta K δ − 2 arctanh r δ . For small δ \delta δ they absorb at most half of the negative 4 k 1 t δ − 3 arctanh r δ 4k_1t\delta^{-3}\operatorname{arctanh}r_\delta 4 k 1 t δ − 3 arctanh r δ . Since s = O ( δ − 1 ) s=O(\delta^{-1}) s = O ( δ − 1 ) in this region, the remainder bounds a negative fixed multiple of s 3 s^3 s 3 . These observations show that
Re E δ ( a ) ≤ K − c s 3 − μ b − − κ v . (74) \operatorname{Re}E_\delta(a)\leq K-cs^3-\mu b^{-}-\kappa v. \tag*{(74)} Re E δ ( a ) ≤ K − c s 3 − μ b − − κ v . (74)
In the region r δ > 1 / 2 r_\delta>1/2 r δ > 1/2 , split the remaining negative phase into two equal parts before obtaining this estimate. One part bounds − c s 3 -cs^3 − c s 3 ; the other retains the stronger additional term − k 2 δ − 3 arctanh r δ -k_2\delta^{-3}\operatorname{arctanh}r_\delta − k 2 δ − 3 arctanh r δ .
The contour integral for the rescaled S S S kernel has multiplier 1. For the scaled lower inverse, the multiplier is δ − 1 u c δ ( z ) \delta^{-1}u_{c_\delta}(z) δ − 1 u c δ ( z ) , with u c u_c u c as in the Airy-factor lemma. Uniformly on these contours, ∣ δ − 1 u c δ ( e η + i δ ξ ) ∣ ≤ K ( a + ξ 2 ) |\delta^{-1}u_{c_\delta}(e^{\eta+i\delta\xi})|\leq K(a+\xi^2) ∣ δ − 1 u c δ ( e η + i δ ξ ) ∣ ≤ K ( a + ξ 2 ) . Indeed, for η ≤ 1 \eta\leq1 η ≤ 1 , the numerator is bounded by K ( η 2 + min ( δ ∣ ξ ∣ , 1 ) 2 ) K(\eta^2+\min(\delta|\xi|,1)^2) K ( η 2 + min ( δ ∣ ξ ∣ , 1 ) 2 ) and ∣ 1 − z − 1 ∣ ≥ κ η |1-z^{-1}|\geq\kappa\eta ∣1 − z − 1 ∣ ≥ κ η ; division by δ \delta δ gives K ( a + ξ 2 / a ) K(a+\xi^2/a) K ( a + ξ 2 / a ) . For η ≥ 1 \eta\geq1 η ≥ 1 the unscaled multiplier is bounded, and δ − 1 ≤ a \delta^{-1}\leq a δ − 1 ≤ a . Gaussian integration of (3.11) therefore gives a prefactor at most K ( 1 + a ) K(1+a) K ( 1 + a ) in front of (3.12). When r δ ≤ 1 / 2 r_\delta\leq1/2 r δ ≤ 1/2 this is polynomial in s s s ; when r δ > 1 / 2 r_\delta>1/2 r δ > 1/2 its logarithm is absorbed by the unused negative k δ − 3 arctanh r δ k\delta^{-3}\operatorname{arctanh}r_\delta k δ − 3 arctanh r δ term. Decreasing c c c once gives the uniform pointwise bound
∣ g δ , ± x ( s , b ; v ) ∣ ≤ K e − c s 3 − μ b − − κ v . |g^{x}_{\delta,\pm}(s,b;v)|\leq Ke^{-cs^3-\mu b^{-}-\kappa v}. ∣ g δ , ± x ( s , b ; v ) ∣ ≤ K e − c s 3 − μ b − − κ v .
Integration against the squared terminal weights proves (3.7).
At s = t δ − 1 + x s=t\delta^{-1}+x s = t δ − 1 + x , the free operator is S T , − T = e 2 T ( I − T ) S_{T,-T}=e^{2T(I-T)} S T , − T = e 2 T ( I − T ) , which is lower triangular. The lower inverse is also lower triangular. The common linear upper bound gives b δ ≤ K δ − 2 b_\delta\leq K\delta^{-2} b δ ≤ K δ − 2 for a constant K K K independent of δ \delta δ , whereas the terminal height is 2 t δ − 3 + O ( δ − 1 ) 2t\delta^{-3}+O(\delta^{-1}) 2 t δ − 3 + O ( δ − 1 ) . Both rows therefore vanish for sufficiently small δ \delta δ . This proves the bound also at the causal endpoint without an infinite contour radius.
For the continuum rows, choose a = max { κ , 2 ( s − x ) + / t } a=\max\{\kappa,2(s-x)^{+}/t\} a = max { κ , 2 ( s − x ) + / t } . The real cubic phase has the same negative cubic bound after the linear upper bound is included, and its off-real part is − [ t a − ( s − x ) ] ξ 2 -[ta-(s-x)]\xi^2 − [ t a − ( s − x )] ξ 2 , with a fixed positive coefficient. This proves the limiting bound. On compact ( s , b ) (s,b) ( s , b ) sets, the Airy contour argument applies uniformly, and the terminal exponential bound permits uniform removal of the v v v tail. This proves the stated weighted row convergence.
Lemma 3.6 (Convergence of first-hitting measures). Let X δ ( s ) = δ X ( s δ − 2 ) X_\delta(s)=\delta X(s\delta^{-2}) X δ ( s ) = δ X ( s δ − 2 ) be the rescaled walk, started at δ ⌊ u / δ ⌋ \delta\lfloor u/\delta\rfloor δ ⌊ u / δ ⌋ , and let
τ δ = inf { s ≥ 0 : X δ ( s ) ≤ f δ ( s ) } , ν δ , u ( d s , d b ) = P u ( τ δ ∈ d s , X δ ( τ δ ) ∈ d b , τ δ < ∞ ) . \tau_\delta=\inf\{s\geq0:X_\delta(s)\leq f_\delta(s)\},\qquad\nu_{\delta,u}(ds,db)=\mathbb{P}_u(\tau_\delta\in ds,X_\delta(\tau_\delta)\in db,\tau_\delta<\infty). τ δ = inf { s ≥ 0 : X δ ( s ) ≤ f δ ( s )} , ν δ , u ( d s , d b ) = P u ( τ δ ∈ d s , X δ ( τ δ ) ∈ d b , τ δ < ∞ ) . Define τ \tau τ and ν u \nu_u ν u analogously for variance-2 Brownian motion and the limiting barrier f f f . Under the hypotheses of Theorem 3.3, ν δ , u → ν u \nu_{\delta,u}\to\nu_u ν δ , u → ν u vaguely for Lebesgue-almost every u u u . For every c > 0 c>0 c > 0 , μ > 0 \mu>0 μ > 0 , and 0 < λ < μ 0<\lambda<\mu 0 < λ < μ ,
∫ e − c s 3 − μ b − ν δ , u ( d s , d b ) ≤ K e − λ ∣ u ∣ (75) \int e^{-cs^3-\mu b^-}\,\nu_{\delta,u}(ds,db)\leq Ke^{-\lambda|u|} \tag*{(75)} ∫ e − c s 3 − μ b − ν δ , u ( d s , d b ) ≤ K e − λ ∣ u ∣ (75) uniformly for sufficiently small δ \delta δ . The same bound holds for ν u \nu_u ν u .
Proof. The lower-semicontinuous path and upper-semicontinuous barrier give a closed set of hitting times, so every finite first hit is attained. The first hitting times are stopping times for the completed natural filtrations. Along any sequence δ ↓ 0 \delta\downarrow0 δ ↓ 0 , the invariance principle couples the centered walk increments X δ ∘ : = X δ − X δ ( 0 ) X^\circ_\delta:=X_\delta-X_\delta(0) X δ ∘ := X δ − X δ ( 0 ) and a centered variance-2 Brownian motion B ∘ B^\circ B ∘ so that X δ ∘ → B ∘ X^\circ_\delta\to B^\circ X δ ∘ → B ∘ uniformly on compact intervals almost surely. Set B ( s ) : = u + B ∘ ( s ) B(s):=u+B^\circ(s) B ( s ) := u + B ∘ ( s ) for the Brownian motion started at u u u . Replacing the walk by a continuous interpolation changes its value by at most δ \delta δ . Choose a deterministic countable dense set T \mathcal{T} T of positive continuity points of the extended-real upper-semicontinuous function f f f . Such points are dense because f f f is a Baire-one function (one can first truncate it below). Local hypograph convergence then also holds after restriction to [ 0 , q ] [0,q] [ 0 , q ] , for every q ∈ T q\in\mathcal{T} q ∈ T . For each such q q q , the common compact upper bound and hypograph convergence imply
A δ ( q ) : = sup 0 ≤ s ≤ q { f δ ( s ) − X δ ∘ ( s ) } ⟶ A ( q ) : = sup 0 ≤ s ≤ q { f ( s ) − B ∘ ( s ) } . A_\delta(q):=\sup_{0\leq s\leq q}\{f_\delta(s)-X^\circ_\delta(s)\}\longrightarrow A(q):=\sup_{0\leq s\leq q}\{f(s)-B^\circ(s)\}. A δ ( q ) := 0 ≤ s ≤ q sup { f δ ( s ) − X δ ∘ ( s )} ⟶ A ( q ) := 0 ≤ s ≤ q sup { f ( s ) − B ∘ ( s )} .
For the upper bound, take a convergent subsequence of approximate maximizers and use the hypograph limsup inequality; if A ( q ) = − ∞ A(q)=-\infty A ( q ) = − ∞ , this rules out every finite lower bound. Hypograph recovery sequences give the reverse inequality. The event { τ δ ≤ q } \{\tau_\delta\leq q\} { τ δ ≤ q } is the height comparison { δ ⌊ u / δ ⌋ ≤ A δ ( q ) } \{\delta\lfloor u/\delta\rfloor\leq A_\delta(q)\} { δ ⌊ u / δ ⌋ ≤ A δ ( q )} . For each Brownian sample, exclude the countably many real values A ( q ) A(q) A ( q ) , q ∈ T q\in\mathcal{T} q ∈ T . For every remaining u u u these height comparisons converge at all q ∈ T q\in\mathcal{T} q ∈ T . Fubini’s theorem therefore gives this conclusion almost surely for almost every deterministic u u u . The exceptional deterministic heights can be chosen to be the atoms of the countably many random variables A ( q ) A(q) A ( q ) , so they depend only on the limiting Brownian law and not on the chosen sequence. Monotonicity of the hitting events and density of T \mathcal{T} T imply τ δ → τ \tau_\delta\to\tau τ δ → τ , including the value ∞ \infty ∞ . At a finite limiting time, compact uniform convergence of the paths gives X δ ( τ δ ) → B ( τ ) X_\delta(\tau_\delta)\to B(\tau) X δ ( τ δ ) → B ( τ ) . Thus bounded continuous functions with compact support of the hitting pair converge in expectation, proving vague convergence. The lower-semicontinuous jump convention changes interpolated values by at most δ \delta δ and does not change this argument.
For the estimate, the exponential martingales of the walk have compensator ψ δ ( ℓ ) = 2 δ − 2 ( cosh ( δ ℓ ) − 1 ) \psi_\delta(\ell)=2\delta^{-2}(\cosh(\delta\ell)-1) ψ δ ( ℓ ) = 2 δ − 2 ( cosh ( δ ℓ ) − 1 ) . For any fixed ℓ \ell ℓ , ψ δ ( ℓ ) \psi_\delta(\ell) ψ δ ( ℓ ) is uniformly bounded and converges to ℓ 2 \ell^2 ℓ 2 . Optional stopping, first at bounded times and then using Fatou’s lemma, gives
E u [ e ± ℓ X δ ( τ δ ) − ψ δ ( ℓ ) τ δ ; τ δ < ∞ ] ≤ K e ± ℓ u . \mathbb{E}_u[e^{\pm\ell X_\delta(\tau_\delta)-\psi_\delta(\ell)\tau_\delta};\tau_\delta<\infty]\leq Ke^{\pm\ell u}. E u [ e ± ℓ X δ ( τ δ ) − ψ δ ( ℓ ) τ δ ; τ δ < ∞ ] ≤ K e ± ℓ u .
A first hit at an upward jump from height a ∈ δ Z a\in\delta\mathbb{Z} a ∈ δ Z would require its time to be a left-isolated point of the deterministic closed set { f δ ≥ a } \{f_\delta\geq a\} { f δ ≥ a } . These points, over all such a a a , are countable and are almost surely avoided by the jump times. Thus at every finite first hit the lower-semicontinuous path agrees with the right-continuous Markov state; at a downward jump both use the post-jump value. Initial height rounding contributes at most e ℓ δ e^{\ell\delta} e ℓ δ to the martingale bound. At a hit, b ≤ C ( 1 + s ) b\leq C(1+s) b ≤ C ( 1 + s ) by (3.2). If u ≤ 0 u\leq0 u ≤ 0 , use e − μ b − ≤ e ℓ b e^{-\mu b^-}\leq e^{\ell b} e − μ b − ≤ e ℓ b and sup s ≥ 0 e − c s 3 + ψ δ ( λ ) s < K \sup_{s\geq0}e^{-cs^3+\psi_\delta(\lambda)s}<K sup s ≥ 0 e − c s 3 + ψ δ ( λ ) s < K . The positive exponential-martingale bound gives K e λ u Ke^{\lambda u} K e λ u . If u ≥ 0 u\geq0 u ≥ 0 , use
e − c s 3 − μ b − ≤ e − λ b − ψ δ ( λ ) s e − c s 3 + λ C ( 1 + s ) + ψ δ ( λ ) s , e^{-cs^3-\mu b^-}\leq e^{-\lambda b-\psi_\delta(\lambda)s}e^{-cs^3+\lambda C(1+s)+\psi_\delta(\lambda)s}, e − c s 3 − μ b − ≤ e − λb − ψ δ ( λ ) s e − c s 3 + λ C ( 1 + s ) + ψ δ ( λ ) s ,
whose second factor is uniformly bounded. The negative exponential-martingale bound gives K e − λ u Ke^{-\lambda u} K e − λ u . Brownian motion has compensator ℓ 2 \ell^2 ℓ 2 , so the same proof applies verbatim to its hitting measure. □
Convergence of the Fredholm Pfaffians. The 11 and 22 blocks converge in trace norm, while the 12 and 21 blocks require Hilbert–Schmidt convergence. We prove these mixed estimates and construct the normalized Pfaffian as an analytic square root, fixing its branch and justifying the conjugations below. Apply the transition formula [10, Theorem 1.1] with thresholds [ 2 T + δ − 1 r i ] [2T+\delta^{-1}r_i] [ 2 T + δ − 1 r i ] . Embed the height lattice into L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) by constant functions on cells of length δ \delta δ , center heights by [ 2 T ] [2T] [ 2 T ] , and use the internal conjugation C δ = diag ( δ − 1 / 2 , δ − 1 / 2 ) C_\delta=\operatorname{diag}(\delta^{-1/2},\delta^{-1/2}) C δ = diag ( δ − 1/2 , δ − 1/2 ) . At the kernel level the rescaling is
K ^ δ ( u , v ) : = δ − 1 C δ K ( [ 2 T ] + ⌊ u / δ ⌋ , [ 2 T ] + ⌊ v / δ ⌋ ) C δ − 1 . \widehat{K}_\delta(u,v):=\delta^{-1}C_\delta K([2T]+\lfloor u/\delta\rfloor,[2T]+\lfloor v/\delta\rfloor)C_\delta^{-1}. K δ ( u , v ) := δ − 1 C δ K ([ 2 T ] + ⌊ u / δ ⌋ , [ 2 T ] + ⌊ v / δ ⌋) C δ − 1 .
Thus the four entries of the rescaled kernel carry the factors δ − 1 \delta^{-1} δ − 1 , 1 1 1 , − δ − 2 -\delta^{-2} − δ − 2 , δ − 1 \delta^{-1} δ − 1 , respectively. Indeed, the isometric embedding sends ( v n ) (v_n) ( v n ) to δ − 1 / 2 v n \delta^{-1/2}v_n δ − 1/2 v n on [ δ n , δ ( n + 1 ) ) [\delta n,\delta(n+1)) [ δ n , δ ( n + 1 )) , giving the density factor δ − 1 \delta^{-1} δ − 1 . On finite height cutoffs using the same projection in both components, C δ T J 0 C δ = J 0 C_\delta^{T}J_0C_\delta=J_0 C δ T J 0 C δ = J 0 and det C δ = 1 \det C_\delta=1 det C δ = 1 ; the congruence on the skew kernel is J m K ↦ C δ − T ( J m K ) C δ − 1 J_mK\mapsto C_\delta^{-T}(J_mK)C_\delta^{-1} J m K ↦ C δ − T ( J m K ) C δ − 1 , which preserves the normalized Fredholm Pfaffian.
Choose ∣ ρ ∣ < θ 1 < ⋯ < θ m |\rho|<\theta_1<\cdots<\theta_m ∣ ρ ∣ < θ 1 < ⋯ < θ m , and conjugate the two components at observation i i i by e − θ i v e^{-\theta_i v} e − θ i v and e θ i v e^{\theta_i v} e θ i v , respectively. A superscript δ \delta δ on a kernel below denotes this centered, embedded and conjugated version. Write R = ( R i j ) R=(R_{ij}) R = ( R ij ) for the free block matrix. Its diagonal blocks are heat kernels between distinct observation positions. Write p ℓ δ ( u , v ) p_\ell^\delta(u,v) p ℓ δ ( u , v ) for the embedded cell kernel at microscopic time ℓ δ − 2 \ell\delta^{-2} ℓ δ − 2 and
p ℓ ( z ) = 1 4 π ℓ e − z 2 / ( 4 ℓ ) , ℓ > 0 , p_\ell(z)=\frac{1}{\sqrt{4\pi\ell}}e^{-z^2/(4\ell)},\qquad\ell>0, p ℓ ( z ) = 4 π ℓ 1 e − z 2 / ( 4 ℓ ) , ℓ > 0 ,
for its Gaussian limit. The local central limit theorem and the same shifted-circle estimate as above give, for every fixed b > 0 b>0 b > 0 , ∣ p ℓ δ ( u , v ) ∣ ≤ C b e − b ∣ u − v ∣ |p_\ell^\delta(u,v)|\le C_b e^{-b|u-v|} ∣ p ℓ δ ( u , v ) ∣ ≤ C b e − b ∣ u − v ∣ , uniformly for small δ \delta δ . This is δ − 1 \delta^{-1} δ − 1 times the lattice heat kernel at ( ⌊ u / δ ⌋ , ⌊ v / δ ⌋ ) (\lfloor u/\delta\rfloor,\lfloor v/\delta\rfloor) (⌊ u / δ ⌋ , ⌊ v / δ ⌋) ; its semigroup identity holds on the embedded cell subspace. The Gaussian limit and this bound give convergence in every weighted difference-variable L 2 L^2 L 2 norm, with cell-rounding errors absorbed in the same bound. To obtain trace convergence for the 11 block with i > j i>j i > j , put ℓ = x i − x j > 0 \ell=x_i-x_j>0 ℓ = x i − x j > 0 and choose θ j < η < θ i \theta_j<\eta<\theta_i θ j < η < θ i . Factor the conjugated heat block as
( 1 > r i M − θ i p ℓ / 2 δ M η ) ( M − η p ℓ / 2 δ M θ j 1 > r j ) . (1_{>r_i}M_{-\theta_i}p_{\ell/2}^\delta M_\eta)(M_{-\eta}p_{\ell/2}^\delta M_{\theta_j}1_{>r_j}). ( 1 > r i M − θ i p ℓ /2 δ M η ) ( M − η p ℓ /2 δ M θ j 1 > r j ) .
The two factors converge in Hilbert–Schmidt norm: integrating out the difference variable leaves, respectively, the integrable weights e 2 ( η − θ i ) u e^{2(\eta-\theta_i)u} e 2 ( η − θ i ) u on u > r i u>r_i u > r i and e 2 ( θ j − η ) v e^{2(\theta_j-\eta)v} e 2 ( θ j − η ) v on v > r j v>r_j v > r j . Their products therefore converge in trace norm. The 22 blocks follow by transposition. The O ( δ ) O(\delta) O ( δ ) changes of the scaled thresholds caused by lattice rounding vanish under the same bounds. The remaining block of R R R is the boundary kernel convolved with heat kernels. The random-walk exponential moments and the explicit formula for H c H_c H c give, after the stated conjugation,
∣ R 12 δ ( i , u ; j , v ) ∣ ≤ C e − θ i u − θ j v + ∣ ρ ∣ ∣ u − v ∣ . |R_{12}^{\delta}(i,u;j,v)|\le Ce^{-\theta_i u-\theta_j v+|\rho||u-v|}. ∣ R 12 δ ( i , u ; j , v ) ∣ ≤ C e − θ i u − θ j v + ∣ ρ ∣∣ u − v ∣ .
on the projected half-lines. This bound is square-integrable there because θ i , θ j > ∣ ρ ∣ \theta_i,\theta_j>|\rho| θ i , θ j > ∣ ρ ∣ . These kernels converge almost everywhere by the Gaussian limit for the heat factors and the pointwise limit of H c H_c H c . This bound proves Hilbert–Schmidt convergence, including x i = x j = 0 x_i=x_j=0 x i = x j = 0 . For operators, write ∥ ⋅ ∥ 1 \|\cdot\|_1 ∥ ⋅ ∥ 1 for the trace norm and ∥ ⋅ ∥ 2 \|\cdot\|_2 ∥ ⋅ ∥ 2 for the Hilbert–Schmidt norm. Thus the blocks of R R R converge in the corresponding components of the mixed norm
∥ A ∥ 1 ∣ 2 = ∥ A 11 ∥ 1 + ∥ A 22 ∥ 1 + ∥ A 12 ∥ 2 + ∥ A 21 ∥ 2 . \|A\|_{1|2}=\|A_{11}\|_1+\|A_{22}\|_1+\|A_{12}\|_2+\|A_{21}\|_2. ∥ A ∥ 1∣2 = ∥ A 11 ∥ 1 + ∥ A 22 ∥ 1 + ∥ A 12 ∥ 2 + ∥ A 21 ∥ 2 .
Write J = J m J=J_m J = J m on the continuum observation-height space ⨁ i = 1 m L 2 ( R ) ⊗ C 2 \bigoplus_{i=1}^m L^2(\mathbb{R})\otimes\mathbb{C}^2 ⨁ i = 1 m L 2 ( R ) ⊗ C 2 . For kernels in this mixed Schatten class such that ( J K ) T = − J K (JK)^T=-JK ( J K ) T = − J K , the generalized normalized Pfaffian is constructed as follows. Choose real finite-rank orthogonal projections P N ↑ I P_N\uparrow I P N ↑ I on the scalar observation-height space and put Π N = P N ⊗ I 2 \Pi_N=P_N\otimes I_2 Π N = P N ⊗ I 2 , K N = Π N K Π N K_N=\Pi_NK\Pi_N K N = Π N K Π N . The normalized finite-dimensional Pfaffian p N p_N p N satisfies
p N ( 0 ) = 1 , p N ( z ) 2 = det 2 ( I + z K N ) e z τ N , τ N = tr ( K N ) 11 + tr ( K N ) 22 . p_N(0)=1,\qquad p_N(z)^2=\det{}_2(I+zK_N)e^{z\tau_N},\qquad\tau_N=\operatorname{tr}(K_N)_{11}+\operatorname{tr}(K_N)_{22}. p N ( 0 ) = 1 , p N ( z ) 2 = det 2 ( I + z K N ) e z τ N , τ N = tr ( K N ) 11 + tr ( K N ) 22 .
Here det 2 ( I + A ) = det ( ( I + A ) e − A ) \det{}_2(I+A)=\det((I+A)e^{-A}) det 2 ( I + A ) = det (( I + A ) e − A ) is the regularized Fredholm determinant for Hilbert–Schmidt A A A . Mixed-norm convergence of K N K_N K N gives locally uniform convergence of these squares. The finite-dimensional inequality ∣ ( 1 + w ) e − w ∣ ≤ e ∣ w ∣ 2 / 2 |(1+w)e^{-w}|\le e^{|w|^2/2} ∣ ( 1 + w ) e − w ∣ ≤ e ∣ w ∣ 2 /2 and Schur triangularization give
∣ p N ( z ) ∣ ≤ exp { 1 4 ∣ z ∣ 2 ∥ K ∥ 2 2 + 1 2 ∣ z ∣ ( ∥ K 11 ∥ 1 + ∥ K 22 ∥ 1 ) } . |p_N(z)|\le\exp\left\{\frac{1}{4}|z|^2\|K\|_2^2+\frac{1}{2}|z|(\|K_{11}\|_1+\|K_{22}\|_1)\right\}. ∣ p N ( z ) ∣ ≤ exp { 4 1 ∣ z ∣ 2 ∥ K ∥ 2 2 + 2 1 ∣ z ∣ ( ∥ K 11 ∥ 1 + ∥ K 22 ∥ 1 ) } .
Montel’s theorem gives an entire limiting square root, unique by its zero value at 1 and hence unique everywhere. Thus the full sequence converges locally uniformly to a projection-independent p K p_K p K ; define Pf ( J + J K ) : = p K ( 1 ) \operatorname{Pf}(J+JK):=p_K(1) Pf ( J + J K ) := p K ( 1 ) . The same argument proves continuity under ∥ ⋅ ∥ 1 ∣ 2 \|\cdot\|_{1|2} ∥ ⋅ ∥ 1∣2 convergence. It agrees with the ordinary normalized Pfaffian for trace-class kernels.
For a bounded invertible component-diagonal D D D with D T J D = J D^TJD=J D T J D = J , similarity preserves det 2 \det{}_2 det 2 and each diagonal trace, and J D K D − 1 = D − T ( J K ) D − 1 JDKD^{-1}=D^{-T}(JK)D^{-1} J D K D − 1 = D − T ( J K ) D − 1 remains skew. Uniqueness of the normalized square root gives p D K D − 1 = p K p_{DKD^{-1}}=p_K p D K D − 1 = p K . Apply this on bounded height intervals and then remove the height cutoffs in mixed norm to justify the exponential conjugations whenever the resulting kernels lie in the stated mixed class.
Proposition 3.7 (Identification of the limiting transition kernel). For an upper-semicontinuous f ≠ − ∞ f\ne-\infty f = − ∞ with a linear upper bound, the limiting projected kernel has Fredholm Pfaffian
Pf ( h ρ ( t , x i ; f ) ≤ r i , 1 ≤ i ≤ m ) . \operatorname{Pf}\bigl(\mathfrak{h}^{\rho}(t,x_i;f)\le r_i,\ 1\le i\le m\bigr). Pf ( h ρ ( t , x i ; f ) ≤ r i , 1 ≤ i ≤ m ) .
Proof. For sufficiently small δ \delta δ , every T ± X i T\pm X_i T ± X i is positive. Substitute W p , q = T q − T p ∗ + T p ∗ H c T q W_{p,q}=T_q-T_p^*+T_p^*H_cT_q W p , q = T q − T p ∗ + T p ∗ H c T q in the symmetric kernel (3.4). Its columns involve only S T , − X i S_{T,-X_i} S T , − X i , H c S T , X j H_cS_{T,X_j} H c S T , X j , and the two direct hitting compositions T j − , T j + T_j^-,T_j^+ T j − , T j + . Replace these columns by the Airy and Brownian first-hitting factors displayed above. They give the four factors F ~ − , j , F ~ j , G ~ − , j , G ~ j \widetilde F_{-,j},\widetilde F_j,\widetilde G_{-,j},\widetilde G_j F − , j , F j , G − , j , G j in the limiting kernel, after the internal conjugation and the threshold translations U s j U_{s_j} U s j . Their signs follow from H c ∗ = − H c H_c^*=-H_c H c ∗ = − H c and the internal conjugation by C δ C_\delta C δ . Expanding the two-by-two product therefore gives
K ~ i j f p = J 0 S j ∗ J 0 ( I H H I ) S j , \widetilde K_{ij}^{\mathrm{fp}}=J_0S_{j}^{*}J_0
\begin{pmatrix}
I & H\\
H & I
\end{pmatrix}
S_j, K ij fp = J 0 S j ∗ J 0 ( I H H I ) S j ,
and the free block gives R i j f p R_{ij}^{\mathrm{fp}} R ij fp . Thus the complete limit is the displayed K f p K^{\mathrm{fp}} K fp , with y j = x j y_j=x_j y j = x j , s j = r j s_j=r_j s j = r j , h 0 = f h_0=f h 0 = f , and ϱ = − ρ \varrho=-\rho ϱ = − ρ . The contour prescription, including the half residue at zero, agrees with the boundary Airy factor (3.5). □
Proof of Theorem 3.3. The transition formula applies to each h 0 , δ h_{0,\delta} h 0 , δ . Since f ≠ − ∞ f\ne-\infty f = − ∞ , a bounded scaled interval contains a finite initial point for small δ \delta δ , making every observed height finite.
The direct hitting compositions have rescaled horizons
δ 2 ( T ± X i ) = t / δ + x i ⟶ ∞ . \delta^2(T\pm X_i)=t/\delta+x_i\longrightarrow\infty. δ 2 ( T ± X i ) = t / δ + x i ⟶ ∞.
The strong Markov decomposition at the first hitting time gives, with q = T ± X i q=T\pm X_i q = T ± X i ,
P 0 , q h 0 , δ e − q Δ S T , ± X i ( u , v ) = E u [ S T , ± X i − τ ( X ( τ ) , v ) ; τ ≤ q ] . P_{0,q}^{h_0,\delta}e^{-q\Delta}S_{T,\pm X_i}(u,v)=\mathbb{E}_u[S_{T,\pm X_i-\tau}(X(\tau),v);\tau\le q]. P 0 , q h 0 , δ e − q Δ S T , ± X i ( u , v ) = E u [ S T , ± X i − τ ( X ( τ ) , v ) ; τ ≤ q ] .
The same identity holds with H c δ , − − 1 H_{c_{\delta,-}}^{-1} H c δ , − − 1 inserted, since this convolution commutes with the free operators. For a fixed δ \delta δ , these equalities can first be tested on finitely supported vectors in the height space with squared norm ∑ n e 2 η n ∣ v n ∣ 2 \sum_n e^{2\eta n}|v_n|^2 ∑ n e 2 η n ∣ v n ∣ 2 , η < 0 \eta<0 η < 0 . The lower-triangular branch of the inverse is bounded there, its geometric shift series converging in operator norm; the free operators are bounded there as well. The hitting operator is dominated entrywise by the heat kernel and hence is bounded on the same space. The strong Markov identity and these weighted bounds permit passage to the completion, justifying the interchanges with the inverse. The argument for the convention at jumps in Lemma 3.6 justifies using the right-continuous Markov state at the first hit.
For the plus composition T i + T_i^+ T i + , the cancellation H c , − − 1 H c = I H_{c,-}^{-1}H_c=I H c , − − 1 H c = I is first used for 0 ≤ c < 1 0\le c<1 0 ≤ c < 1 on a compatible exponentially weighted space. For c ≥ 1 c\ge1 c ≥ 1 , the complete grouped identity is its analytic continuation, justified by [10 Theorem 1.1 and Appendix A] .
Consequently, δ − 2 T i − \delta^{-2}T_i^- δ − 2 T i − and δ − 1 T i + \delta^{-1}T_i^+ δ − 1 T i + are integrals of the rows in Lemma 3.5 against ν δ , u \nu_{\delta,u} ν δ , u . On a compact set of hitting pairs, vague convergence in Lemma 3.6 and uniform L 2 L^2 L 2 row convergence give convergence of these integrals in the weighted row L 2 L^2 L 2 norm for almost every u u u . A continuous compact cutoff in ( s , b ) (s,b) ( s , b ) justifies this Hilbert-space-valued statement by uniform approximation by finite sums of scalar continuous functions times fixed row vectors. The cutoff can be removed uniformly: the row bound tends uniformly to zero when s → ∞ s\to\infty s → ∞ or b − → ∞ b^-\to\infty b − → ∞ , while b + ≤ C ( 1 + s ) b^+\le C(1+s) b + ≤ C ( 1 + s ) on the support. The measures have total mass at most one. Moreover, (3.13) gives the uniform row bound K e − λ ∣ u ∣ Ke^{-\lambda|u|} K e − λ ∣ u ∣ , with λ \lambda λ as large as required by first choosing μ > λ \mu>\lambda μ > λ .
Apply the transition formula, lattice embedding, and symplectic conjugation above. For linear scattering products, the rank-one trace norm is dominated by K ( e a u + e ∣ ρ ∣ ∣ u ∣ ) e − λ ∣ u ∣ K(e^{au}+e^{|\rho||u|})e^{-\lambda|u|} K ( e a u + e ∣ ρ ∣∣ u ∣ ) e − λ ∣ u ∣ for λ > max ( a , ∣ ρ ∣ ) \lambda>\max(a,|\rho|) λ > max ( a , ∣ ρ ∣ ) . For quadratic products it is dominated by K e ∣ ρ ∣ ( ∣ u ∣ + ∣ v ∣ ) e − λ ( ∣ u ∣ + ∣ v ∣ ) Ke^{|\rho|(|u|+|v|)}e^{-\lambda(|u|+|v|)} K e ∣ ρ ∣ ( ∣ u ∣ + ∣ v ∣ ) e − λ ( ∣ u ∣ + ∣ v ∣ ) for λ > ∣ ρ ∣ \lambda>|\rho| λ > ∣ ρ ∣ . These bounds are integrable on the full internal height line(s). Dominated convergence therefore gives trace-norm convergence of all scattering blocks along the given hypograph-convergent sequence; the domination depends only on the common linear upper bound. Together with the free-block estimates, this gives convergence of the full kernel in ∥ ⋅ ∥ 1 ∣ 2 \lVert\cdot\rVert_{1|2} ∥ ⋅ ∥ 1∣2 .
Proposition 3.7 identifies the limiting kernel. Mixed-norm Pfaffian continuity gives convergence of the joint distribution functions to the KPZ transition law. The same weighted bounds, with thresholds in a compact set, prove continuity in the limiting thresholds, completing the weak convergence. □
For 0 ≤ a 1 < ⋯ < a k 0\le a_1<\cdots<a_k 0 ≤ a 1 < ⋯ < a k and b 1 , … , b k ∈ R b_1,\ldots,b_k\in\mathbb{R} b 1 , … , b k ∈ R , set
f ( x ) = { b j , x = a j , − ∞ , x ∉ { a 1 , … , a k } , h 0 , δ ( y ) = { ⌊ b j / δ ⌋ , y = a j δ − 2 , − ∞ , otherwise . (76) f(x)=
\begin{cases}
b_j, & x=a_j,\\
-\infty, & x\notin\{a_1,\ldots,a_k\},
\end{cases}
\qquad
h_{0,\delta}(y)=
\begin{cases}
\lfloor b_j/\delta\rfloor, & y=a_j\delta^{-2},\\
-\infty, & \text{otherwise}.
\end{cases} \tag*{(76)} f ( x ) = { b j , − ∞ , x = a j , x ∈ / { a 1 , … , a k } , h 0 , δ ( y ) = { ⌊ b j / δ ⌋ , − ∞ , y = a j δ − 2 , otherwise . (76)
The profiles in (76) converge locally in hypograph with a common linear upper bound, so Theorem 3.3 gives their finite-wedge transition limit.
Corollary 3.8 (Independent random initial data). The conclusion of Theorem 3.3 also holds when the initial profiles are random and independent of the PNG environment, provided f δ ⇒ f f_\delta\Rightarrow f f δ ⇒ f locally in hypograph, f ≠ − ∞ f\ne-\infty f = − ∞ almost surely, and their global linear upper bounds have tight random constants, namely, assume
lim A → ∞ lim δ ↓ 0 P ( ∃ s ≥ 0 : f δ ( s ) > A ( 1 + s ) ) = 0. \lim_{A\to\infty}\lim_{\delta\downarrow0}\mathbb{P}\bigl(\exists s\ge0:f_\delta(s)>A(1+s)\bigr)=0. A → ∞ lim δ ↓ 0 lim P ( ∃ s ≥ 0 : f δ ( s ) > A ( 1 + s ) ) = 0. The limiting fixed point is started from the random f f f independently of its arriving directed landscape.
Proof. Under a deterministic linear upper bound, the theorem gives sequential convergence of the conditional transition probabilities. A Skorokhod coupling of the initial profiles and bounded convergence therefore give joint convergence with the terminal height vector. For upper bounds with general tight random constants, use
f δ A ( s ) = min { f δ ( s ) , δ [ A ( 1 + s ) / δ ] } . f_\delta^A(s)=\min\{f_\delta(s),\delta[A(1+s)/\delta]\}. f δ A ( s ) = min { f δ ( s ) , δ [ A ( 1 + s ) / δ ]} .
At finite values this differs by at most δ \delta δ from clipping by the continuous ceiling, and both clips equal − ∞ -\infty − ∞ where f δ = − ∞ f_\delta=-\infty f δ = − ∞ . Thus it converges locally in hypograph to min { f , A ( 1 + ⋅ ) } \min\{f,A(1+\cdot)\} min { f , A ( 1 + ⋅ )} . The underlying integer cutoff is the floor of an increasing continuous affine function, so it is upper-semicontinuous and has finitely many jumps on compacts, preserving the allowed initial-data class. The clipped profiles differ from the originals with probability tending uniformly to zero as A → ∞ A\to\infty A → ∞ . The same tight-envelope property passes to the limiting profile, so its clipping error also tends to zero as A → ∞ A\to\infty A → ∞ . Letting A → ∞ A\to\infty A → ∞ proves the claim. □
Subcritical finite-wedge transitions
Fix c ⋆ ∈ [ 0 , 1 ) c_\star\in[0,1) c ⋆ ∈ [ 0 , 1 ) before sending δ \delta δ to zero. The limiting scattering form contains a derivative and, when an initial wedge is at zero, a jump contribution. We first treat finite wedges at positive terminal positions; stationary comparison will include the terminal wall. The boundary factors require the internal conjugation below, with d = 1 − c ⋆ d=1-c_\star d = 1 − c ⋆ :
C δ , ⋆ = diag ( i d δ − 1 / 2 , − i δ 1 / 2 / d ) . (77) C_{\delta,\star}=\operatorname{diag}(i d\delta^{-1/2},-i\delta^{1/2}/d). \tag*{(77)} C δ , ⋆ = diag ( i d δ − 1/2 , − i δ 1/2 / d ) . (77)
Together with the lattice density factor δ − 1 \delta^{-1} δ − 1 , it multiplies the four entries by δ − 1 \delta^{-1} δ − 1 , − d 2 δ − 2 -d^2\delta^{-2} − d 2 δ − 2 , − d − 2 -d^{-2} − d − 2 , and δ − 1 \delta^{-1} δ − 1 , respectively. Its determinant is one and it preserves J 0 J_0 J 0 .
Proposition 3.9 (Subcritical finite-wedge limit away from the terminal wall). For the finite-wedge initial data in (3.14), any fixed c ⋆ ∈ [ 0 , 1 ) c_{\star} \in[0,1) c ⋆ ∈ [ 0 , 1 ) , t > 0 t>0 t > 0 , and 0 < x 1 < ⋯ < x m 0<x_1<\cdots<x_m 0 < x 1 < ⋯ < x m , one has
( δ [ h ( c ⋆ ) ( t δ − 3 , x i δ − 2 ; h 0 , δ ) − 2 t δ − 3 ] ) i = 1 m ⟹ ( h − ∞ ( t , x i ; f ) ) i = 1 m . \left(\delta\left[h^{(c_{\star})}(t\delta^{-3},x_i\delta^{-2};h_{0,\delta})-2t\delta^{-3}\right]\right)_{i=1}^{m}\Longrightarrow\left(h^{-\infty}(t,x_i;f)\right)_{i=1}^{m}. ( δ [ h ( c ⋆ ) ( t δ − 3 , x i δ − 2 ; h 0 , δ ) − 2 t δ − 3 ] ) i = 1 m ⟹ ( h − ∞ ( t , x i ; f ) ) i = 1 m . Here the initial wedges may include a wedge at the spatial origin.
For finite wedges the random-walk hitting operator can be reduced to finitely many observation times. Set A j = a j δ − 2 A_j=a_j\delta^{-2} A j = a j δ − 2 and b j , δ l a t = ⌊ b j / δ ⌋ b_{j,\delta}^{\mathrm{lat}}=\lfloor b_j/\delta\rfloor b j , δ lat = ⌊ b j / δ ⌋ . For all sufficiently small δ \delta δ , every endpoint T ± X i T\pm X_i T ± X i exceeds A k A_k A k . With A 0 = 0 A_0=0 A 0 = 0 , put
D j , δ = ∏ ℓ = 1 j − 1 [ e ( A ℓ − A ℓ − 1 ) Δ 1 1 > b ℓ , δ l a t + e ( A j − A j − 1 ) Δ 1 1 ≤ b j , δ l a t ] . D_{j,\delta}=\prod_{\ell=1}^{j-1}\left[e^{(A_\ell-A_{\ell-1})\Delta_1}\mathbf{1}_{>b_{\ell,\delta}^{\mathrm{lat}}}+e^{(A_j-A_{j-1})\Delta_1}\mathbf{1}_{\leq b_{j,\delta}^{\mathrm{lat}}}\right]. D j , δ = ℓ = 1 ∏ j − 1 [ e ( A ℓ − A ℓ − 1 ) Δ 1 1 > b ℓ , δ lat + e ( A j − A j − 1 ) Δ 1 1 ≤ b j , δ lat ] .
The product is ordered by increasing ℓ \ell ℓ from left to right; an empty product is I I I . The first-hitting decomposition is the finite identity
P 0 , q h 0 , δ e − q Δ = ∑ j = 1 k D j , δ e − A j Δ , q ≥ A k . (78) \mathrm{P}_{0,q}^{h_{0,\delta}}e^{-q\Delta}=\sum_{j=1}^{k}D_{j,\delta}e^{-A_j\Delta},\qquad q\geq A_k. \tag*{(78)} P 0 , q h 0 , δ e − q Δ = j = 1 ∑ k D j , δ e − A j Δ , q ≥ A k . (78)
This conditions on the first successful height test, including time zero.
We interpret the limiting derivative operator as follows. For row kernels f , g f,g f , g , write ( f ⊗ g ) ( v , w ) = f ( v ) g ( w ) (f\otimes g)(v,w)=f(v)g(w) ( f ⊗ g ) ( v , w ) = f ( v ) g ( w ) for their rank-one tensor kernel, with component indices included in v , w v,w v , w . Let F , G F,G F , G be vector-valued row functions with exponentially integrable row norms and first derivatives. They may have one common jump at a height b b b . Define the following trace-class bilinear expression
q D ( F , G ) = 1 2 ∫ R ∖ { b } ( F ( u ) ⊗ G ′ ( u ) − F ′ ( u ) ⊗ G ( u ) ) d u + 1 2 ( F ( b − ) ⊗ G ( b + ) − F ( b + ) ⊗ G ( b − ) ) . (79) q_D(F,G)=\frac{1}{2}\int_{\mathbb{R}\setminus\{b\}}\left(F(u)\otimes G'(u)-F'(u)\otimes G(u)\right)\,du+\frac{1}{2}\left(F(b-)\otimes G(b+)-F(b+)\otimes G(b-)\right). \tag*{(79)} q D ( F , G ) = 2 1 ∫ R ∖ { b } ( F ( u ) ⊗ G ′ ( u ) − F ′ ( u ) ⊗ G ( u ) ) d u + 2 1 ( F ( b − ) ⊗ G ( b + ) − F ( b + ) ⊗ G ( b − ) ) . (79)
Omit the jump term for continuous rows. This midpoint convention is needed when the first initial wedge is at zero.
Lemma 3.10 (Subcritical operator limits). Use the height coordinates of Lemma 3.4 . Then, locally uniformly,
δ − 1 S T , x δ − 2 ⟶ S t , x , \delta^{-1}S_{T,x\delta^{-2}}\longrightarrow S^{t,x}, δ − 1 S T , x δ − 2 ⟶ S t , x , d 2 δ − 2 H c ⋆ S T , x δ − 2 ⟶ D S t , x , d^2\delta^{-2}H_{c_{\star}}S_{T,x\delta^{-2}}\longrightarrow DS^{t,x}, d 2 δ − 2 H c ⋆ S T , x δ − 2 ⟶ D S t , x , d − 2 H c ⋆ − 1 S T , x δ − 2 ⟶ D − 1 S t , x . (80) d^{-2}H_{c_{\star}}^{-1}S_{T,x\delta^{-2}}\longrightarrow D^{-1}S^{t,x}. \tag*{(80)} d − 2 H c ⋆ − 1 S T , x δ − 2 ⟶ D − 1 S t , x . (80) Here D − 1 D^{-1} D − 1 is the lower-triangular branch used above. With terminal weights e η v e^{\eta v} e η v or e − θ v e^{-\theta v} e − θ v on v > r v>r v > r , all three rows have norm at most C a e a u C_{a}e^{au} C a e a u for a sufficiently large fixed a a a ; the same statement holds with any fixed number of derivatives in either height variable. For embedded microscopic rows, these derivatives mean scaled finite differences, not derivatives of the piecewise constant embedding. The corresponding scaled hitting rows converge to
G i , + ⋆ ( u , v ) = E u [ S t , x i − τ ( B ( τ ) , v ) 1 { τ < ∞ } ] , G i , − ⋆ ( u , v ) = E u [ ( D − 1 S ) t , − x i − τ ( B ( τ ) , v ) 1 { τ < ∞ } ] , G_{i,+}^{\star}(u,v)=\mathbb{E}_{u}\left[S_{t,x_i-\tau}(B(\tau),v)\mathbf{1}_{\{\tau<\infty\}}\right],\qquad G_{i,-}^{\star}(u,v)=\mathbb{E}_{u}\left[(D^{-1}S)_{t,-x_i-\tau}(B(\tau),v)\mathbf{1}_{\{\tau<\infty\}}\right], G i , + ⋆ ( u , v ) = E u [ S t , x i − τ ( B ( τ ) , v ) 1 { τ < ∞ } ] , G i , − ⋆ ( u , v ) = E u [ ( D − 1 S ) t , − x i − τ ( B ( τ ) , v ) 1 { τ < ∞ } ] , where τ = min { a j : B ( a j ) ≤ b j } \tau=\min\{a_j:B(a_j)\leq b_j\} τ = min { a j : B ( a j ) ≤ b j } , with min ∅ = ∞ \min\varnothing=\infty min ∅ = ∞ , using the microscopic normalizations δ − 1 T i + \delta^{-1}T_i^{+} δ − 1 T i + and d − 2 T i − d^{-2}T_i^{-} d − 2 T i − . They and their first derivatives in the starting height obey a bound
Q 1 ( u ) = C a , b ( e a u 1 { u < 0 } + e − b u 1 { u ≥ 0 } ) , (81) Q_1(u)=C_{a,b}\left(e^{au}\mathbf{1}_{\{u<0\}}+e^{-bu}\mathbf{1}_{\{u\geq0\}}\right), \tag*{(81)} Q 1 ( u ) = C a , b ( e a u 1 { u < 0 } + e − b u 1 { u ≥ 0 } ) , (81) for arbitrarily large fixed a , b > 0 a,b>0 a , b > 0 , away from the sole possible jump at u = b 1 u=b_1 u = b 1 when a 1 = 0 a_1=0 a 1 = 0 . At that jump both one-sided row norms are bounded. Discrete first differences divided by δ \delta δ obey the same estimate on each side of the jump.
If F δ , G δ F_{\delta},G_{\delta} F δ , G δ are any two such hitting rows, then
d 2 2 ∑ n ∈ Z ∑ k ≥ 1 c ⋆ k − 1 [ F δ ( δ n ) ⊗ G δ ( δ ( n + k ) ) − F δ ( δ ( n + k ) ) ⊗ G δ ( δ n ) ] ⟶ q D ( F , G ) (82) \frac{d^2}{2}\sum_{n\in\mathbb{Z}}\sum_{k\geq1}c_{\star}^{k-1}\left[F_{\delta}(\delta n)\otimes G_{\delta}(\delta(n+k))-F_{\delta}(\delta(n+k))\otimes G_{\delta}(\delta n)\right]\longrightarrow q_D(F,G) \tag*{(82)} 2 d 2 n ∈ Z ∑ k ≥ 1 ∑ c ⋆ k − 1 [ F δ ( δ n ) ⊗ G δ ( δ ( n + k )) − F δ ( δ ( n + k )) ⊗ G δ ( δ n ) ] ⟶ q D ( F , G ) (82) in trace norm. At c ⋆ = 0 c_{\star}=0 c ⋆ = 0 only the term k = 1 k=1 k = 1 is retained.
Proof. On z = e δ ( u + i ξ ) z=e^{\delta(u+i\xi)} z = e δ ( u + i ξ ) the centered Airy exponent has the same uniform Gaussian estimate as in Lemma 3.4. For sufficiently small δ \delta δ the entire contour lies in the annulus c ⋆ < ∣ z ∣ < c ⋆ − 1 c_{\star}<|z|<c_{\star}^{-1} c ⋆ < ∣ z ∣ < c ⋆ − 1 when 0 < c ⋆ < 1 0<c_{\star}<1 0 < c ⋆ < 1 . At c ⋆ = 0 c_{\star}=0 c ⋆ = 0 the Laurent contour is unrestricted. In both cases no boundary pole is crossed. The symbols satisfy
d 2 δ h c ⋆ ( e δ w ) ⟶ w , δ d 2 u c ⋆ ( e δ w ) ⟶ 1 w . \frac{d^2}{\delta}\,h_{c_{\star}}(e^{\delta w})\longrightarrow w,\qquad\frac{\delta}{d^2}\,u_{c_{\star}}(e^{\delta w})\longrightarrow\frac{1}{w}. δ d 2 h c ⋆ ( e δ w ) ⟶ w , d 2 δ u c ⋆ ( e δ w ) ⟶ w 1 . On the whole contour the first normalized multiplier is bounded by C ( 1 + ∣ ξ ∣ ) C(1+|\xi|) C ( 1 + ∣ ξ ∣ ) . The second is bounded by a constant depending on a a a and c ⋆ c_{\star} c ⋆ : near either of the possible zeros z = ± 1 z=\pm1 z = ± 1 , use ∣ 1 − z − 2 ∣ ≥ 1 − e − 2 a δ |1-z^{-2}|\geq1-e^{-2a\delta} ∣1 − z − 2 ∣ ≥ 1 − e − 2 a δ : the numerator is bounded and the factor δ \delta δ cancels this denominator bound. Away from those two arcs the denominator is bounded away from zero. The uniform Gaussian exponent bound also suppresses the arc near − 1 -1 − 1 . Multiplying by powers of w w w supplies the asserted derivative estimates. Gaussian domination proves the factor limits and weighted row convergence.
In the finite first-hitting identity (3.16), the heat tests converge by the local central limit theorem. Positive-time Brownian heights have continuous densities; at time zero exclude u = b 1 u=b_1 u = b 1 . The uniform moment formula
E e η δ [ X ( a j δ − 2 ) − X ( 0 ) ] = e 2 a j δ − 2 ( cosh ( η δ ) − 1 ) \mathbb{E}e^{\eta\delta[X(a_j\delta^{-2})-X(0)]}=e^{2a_j\delta^{-2}(\cosh(\eta\delta)-1)} E e η δ [ X ( a j δ − 2 ) − X ( 0 )] = e 2 a j δ − 2 ( c o s h ( η δ ) − 1 )
and the preceding row bounds justify truncating hit heights and passing to weighted row limits. For the starting-height derivative, suppose first that a 1 > 0 a_1 > 0 a 1 > 0 . In each first-hitting term its first factor is the heat kernel at time a 1 a_1 a 1 . Its embedded first difference divided by δ \delta δ converges to the derivative of the variance-two Gaussian kernel and has uniform exponentially weighted bounds. One can see this directly from its Fourier integral by inserting ( e δ w − 1 ) / δ (e^{\delta w}-1)/\delta ( e δ w − 1 ) / δ and using the heat exponent’s Gaussian bound. Integrating this estimate against the remaining finite heat tests and the terminal Airy kernel gives (3.19). Its positive-height tail follows from exponential tilting together with the event that one of the finitely many tested heights is at most max j b j \max_j b_j max j b j ; its negative-height tail follows from the Airy-kernel row estimate. These are the same exponential-moment bounds used for the undifferentiated rows. If a 1 = 0 a_1 = 0 a 1 = 0 , split the starting height at b 1 b_1 b 1 . On the lower side the row is the corresponding Airy row. On the upper side the first possible positive hit time is a 2 > 0 a_2 > 0 a 2 > 0 , so the preceding heat argument applies; if there is no second wedge, the upper row is zero. The bond n = ⌊ b 1 / δ ⌋ n = \lfloor b_1/\delta\rfloor n = ⌊ b 1 / δ ⌋ crossing the jump is handled in the crossing sum below. This proves the estimates and one-sided convergence.
For the last assertion, use skew-symmetry before taking absolute values. Away from the possible jump the tensor difference for an increment k δ k\delta k δ has trace norm bounded by C k δ Ck\delta C k δ times the product of the bounds in (3.19). Summation over n n n therefore leaves an integrable bound of order C k Ck C k , uniformly in small δ \delta δ . If the increment crosses the jump there are at most k k k starting lattice sites; their tensors have uniformly bounded row norms, with at most a factor e C k δ e^{Ck\delta} e C k δ . The sums in k k k are uniformly summable because c ⋆ < 1 c_\star< 1 c ⋆ < 1 . For each fixed k k k , the ordinary Riemann sum off the jump converges to k 2 ( F ⊗ G ′ − F ′ ⊗ G ) \frac{k}{2}(F \otimes G' - F' \otimes G) 2 k ( F ⊗ G ′ − F ′ ⊗ G ) , and the k k k crossing terms converge to k 2 ( F ( b − ) ⊗ G ( b + ) − F ( b + ) ⊗ G ( b − ) ) \frac{k}{2}(F(b-) \otimes G(b+) - F(b+) \otimes G(b-)) 2 k ( F ( b − ) ⊗ G ( b + ) − F ( b + ) ⊗ G ( b − )) . The coefficient identity d 2 ∑ k ≥ 1 k c ⋆ k − 1 = 1 d^2 \sum_{k\ge1} kc_\star^{k-1} = 1 d 2 ∑ k ≥ 1 k c ⋆ k − 1 = 1 gives (3.20). All these statements hold in trace norm, since ∥ f ⊗ g ∥ 1 = ∥ f ∥ 2 ∥ g ∥ 2 \|f \otimes g\|_1 = \|f\|_2\|g\|_2 ∥ f ⊗ g ∥ 1 = ∥ f ∥ 2 ∥ g ∥ 2 .
Proof of Proposition 3.9. Apply the transition formula with (3.15). Choose 0 < θ 1 < ⋯ < θ m 0 < \theta_1 < \cdots< \theta_m 0 < θ 1 < ⋯ < θ m and use the same reciprocal exponential weights in the two components as in the finite-parameter proof. Linear scattering products converge in trace norm by Lemma 3.10; the row norms of the Airy and hitting factors have integrable products. For the quadratic scattering products retain their antisymmetrized sums and use (3.20). The diagonal heat-kernel blocks converge in trace norm by the two-Hilbert–Schmidt-factor argument already given. Finally, the boundary block converges in Hilbert–Schmidt norm to − D e ( x i + x j ) D 2 -D e^{(x_i+x_j)D^2} − D e ( x i + x j ) D 2 . Indeed, x i + x j > 0 x_i+x_j>0 x i + x j > 0 supplies a Gaussian Fourier bound, the normalized boundary multiplier converges to w w w , and the terminal exponential weights make the projected kernel square integrable. Thus the whole kernel converges in the mixed norm ∥ ⋅ ∥ 1 ∣ 2 \|\cdot\|_{1|2} ∥ ⋅ ∥ 1∣2 to a kernel K ⋆ K^\star K ⋆ determined by the three limits in (3.18) and the form (3.17).
To identify the limiting Pfaffian, take the critical-window kernel with ρ = − R \rho=-R ρ = − R , so its source parameter is ϱ = R \varrho=R ϱ = R , and apply the symplectic similarity multiplying its 12 12 12 entry by R 2 R^2 R 2 and its 21 21 21 entry by R − 2 R^{-2} R − 2 . Fix 0 < θ 1 < ⋯ < θ m 0<\theta_1<\cdots<\theta_m 0 < θ 1 < ⋯ < θ m , independently of R R R , and a sufficiently large fixed a > θ m a>\theta_m a > θ m . For R > 2 a R>2a R > 2 a , the boundary convolution on a + i R a+i\mathbb{R} a + i R has multiplier w / ( R 2 − w 2 ) w/(R^2-w^2) w / ( R 2 − w 2 ) , without a residue term.
∣ R 2 w R 2 − w 2 ∣ ≤ 2 ∣ w ∣ , R − 2 H R − 1 S = D − 1 S − R − 2 D S , \left|\frac{R^2w}{R^2-w^2}\right|\le2|w|,\qquad R^{-2}H_R^{-1}S=D^{-1}S-R^{-2}DS, R 2 − w 2 R 2 w ≤ 2∣ w ∣ , R − 2 H R − 1 S = D − 1 S − R − 2 D S ,
Gaussian contour domination gives
R 2 ( H R S ) ⟶ D S , R − 2 ( H R − 1 S ) ⟶ D − 1 S . R^2(H_RS)\longrightarrow DS,\qquad R^{-2}(H_R^{-1}S)\longrightarrow D^{-1}S. R 2 ( H R S ) ⟶ D S , R − 2 ( H R − 1 S ) ⟶ D − 1 S .
The same estimates and finite heat tests give the one-sided derivative bounds above uniformly in large R R R . These fixed exponential conjugations preserve the Fredholm Pfaffians. For the quadratic term the exact antisymmetrized expression is
R 2 2 ∫ 0 ∞ e − R s ∫ R [ F ( u ) ⊗ G ( u + s ) − F ( u + s ) ⊗ G ( u ) ] d u d s . \frac{R^2}{2}\int_0^\infty e^{-Rs}\int_{\mathbb{R}}\left[F(u)\otimes G(u+s)-F(u+s)\otimes G(u)\right]\,du\,ds. 2 R 2 ∫ 0 ∞ e − R s ∫ R [ F ( u ) ⊗ G ( u + s ) − F ( u + s ) ⊗ G ( u ) ] d u d s .
Off the jump, divide the bracket by s s s and use the derivative bounds. The crossing interval has length s s s , giving the jump term in (3.17). Since R 2 ∫ 0 ∞ s e − R s d s = 1 R^2\int_0^\infty se^{-Rs}\,ds=1 R 2 ∫ 0 ∞ s e − R s d s = 1 , dominated convergence gives that form in trace norm. Positive terminal positions smooth the boundary block. Hence the continuum kernels converge in the same fixed mixed norm to K ⋆ K^\star K ⋆ .
The proof of [17 Corollary 5.2] gives finite-dimensional convergence L − n → L − ∞ \mathcal{L}^{-n}\to\mathcal{L}^{-\infty} L − n → L − ∞ on rational endpoints as n → ∞ n\to\infty n → ∞ . The uniform modulus in [17 Proposition 4.10(1)] extends this convergence to the fixed real endpoints used here. Taking maxima over the fixed finite initial support identifies the weak limit of the corresponding fixed-point vectors as h − ∞ ( t , x j ; f ) h^{-\infty}(t,x_j;f) h − ∞ ( t , x j ; f ) . Mixed-norm Pfaffian continuity identifies its joint CDF with the Pfaffian of K ⋆ K^\star K ⋆ . The weighted trace and Hilbert–Schmidt bounds give threshold continuity, so Pfaffian convergence proves the microscopic limit at all fixed thresholds.
Remark 3.11 (Wall observations). At x i = x j = 0 x_i=x_j=0 x i = x j = 0 the boundary block tends to − D -D − D , a derivative of a delta distribution, and has no individual Hilbert–Schmidt limit. Lemma 3.17 supplies wall observations by stationary comparison and the one-point wall law.
Stationary comparison and wall observations
We now include terminal observations at the wall and obtain spatial comparison for tightness. We use exact stationarity, the one-parameter functional limit of Section 3.1, the critical transition limit, and Poisson passage bounds.
G c ( z , s ; y , t ) : = L P N G ( c ) ( z , s ; y , t ) , G c ( T ) : = G c ( 0 , 0 ; 0 , T ) , G_c(z,s;y,t):=\mathcal{L}^{(c)}_{\mathrm{PNG}}(z,s;y,t),\qquad G_c(T):=G_c(0,0;0,T), G c ( z , s ; y , t ) := L PNG ( c ) ( z , s ; y , t ) , G c ( T ) := G c ( 0 , 0 ; 0 , T ) ,
for the unscaled passage metric and its wall-to-wall specialization.
Lemma 3.12 (A wall lower tail). Let G c ( T ) G_c(T) G c ( T ) be the half-space PNG passage value from ( 0 , 0 ) (0,0) ( 0 , 0 ) to ( 0 , T ) (0,T) ( 0 , T ) , with bulk nucleation intensity 2 2 2 and boundary intensity c c c . There are absolute C , c 0 > 0 C,c_0>0 C , c 0 > 0 such that, for T ≥ 1 T\ge1 T ≥ 1 , a ≥ 1 a\ge1 a ≥ 1 , and every c ≥ 0 c\ge0 c ≥ 0 ,
P { G c ( T ) − 2 T ≤ − a T 1 / 3 } ≤ C exp ( − c 0 a 3 ) . \mathbb{P}\{G_c(T)-2T\le-aT^{1/3}\}\le C\exp(-c_0a^3). P { G c ( T ) − 2 T ≤ − a T 1/3 } ≤ C exp ( − c 0 a 3 ) .
Proof. It suffices to take c = 0 c=0 c = 0 by Poisson thinning. The coordinates u = t + x u=t+x u = t + x , v = t − x v=t-x v = t − x turn the bulk process into intensity-one Poisson points in 0 ≤ v < u ≤ T 0\le v<u\le T 0 ≤ v < u ≤ T . Reflecting these points about u = v u=v u = v gives the Poissonized fixed-point-free involution model. Any increasing path can be folded into u ≥ v u\ge v u ≥ v without changing its number of points: each excursion below the diagonal is reflected separately, and a point and its reflected partner are incomparable. Thus its longest increasing subsequence equals G 0 ( T ) G_0(T) G 0 ( T ) .
The triangle contains Poisson(T 2 / 2 T^2/2 T 2 /2 ) points and has no diagonal points, so P j ( T ; 0 ) : = P ( G 0 ( T ) ≤ j ) P_j(T;0):=\mathbb{P}(G_0(T)\le j) P j ( T ; 0 ) := P ( G 0 ( T ) ≤ j ) is the CDF in [4 (4.7)] . By [4 Theorem 4.2, (4.21)] , for l ≥ 1 l\ge1 l ≥ 1 ,
P 2 l ( T ; 0 ) = 1 2 e − T 2 / 2 ( D l − − ( T ) + D l − 1 + + ( T ) ) . P_{2l}(T;0)=\frac{1}{2}e^{-T^2/2}\left(D_l^{--}(T)+D_{l-1}^{++}(T)\right). P 2 l ( T ; 0 ) = 2 1 e − T 2 /2 ( D l −− ( T ) + D l − 1 ++ ( T ) ) .
In that paper, π n ( z ; T ) \pi_n(z;T) π n ( z ; T ) is the monic orthogonal polynomial on the unit circle for e 2 T cos θ d θ / ( 2 π ) e^{2T\cos\theta}\,d\theta/(2\pi) e 2 T c o s θ d θ / ( 2 π ) , and N n ( T ) N_n(T) N n ( T ) is its squared norm for this measure [4 (4.10)] . The determinant products in [4 (4.17)–(4.18)] are
e − T 2 / 2 D l − − ( T ) = ∏ j ≥ l N 2 j + 2 ( T ) − 1 ∏ j ≥ l ( 1 + π 2 j + 2 ( 0 ; T ) ) , e^{-T^2/2}D_l^{--}(T)=\prod_{j\ge l}N_{2j+2}(T)^{-1}\prod_{j\ge l}(1+\pi_{2j+2}(0;T)), e − T 2 /2 D l −− ( T ) = j ≥ l ∏ N 2 j + 2 ( T ) − 1 j ≥ l ∏ ( 1 + π 2 j + 2 ( 0 ; T )) ,
and the corresponding product with the minus sign gives e − T 2 / 2 D l + + ( T ) e^{-T^2/2}D_l^{++}(T) e − T 2 /2 D l ++ ( T ) . Set x = 2 ( l − T ) l − 1 / 3 x=2(l-T)l^{-1/3} x = 2 ( l − T ) l − 1/3 . The bounds [4 Corollary 7.6, (7.24)–(7.26)] give, for x ≤ − M x\le-M x ≤ − M ,
e − T 2 / 2 D l − − ( T ) ≤ C exp { − c ∣ x ∣ 3 + C ∣ x ∣ 3 / 2 } . e^{-T^2/2}D_l^{--}(T)\le C\exp\{-c|x|^3+C|x|^{3/2}\}. e − T 2 /2 D l −− ( T ) ≤ C exp { − c ∣ x ∣ 3 + C ∣ x ∣ 3/2 } .
The same upper bound, with l l l replaced by l − 1 l-1 l − 1 , applies to e − T 2 / 2 D l − 1 + + ( T ) e^{-T^2/2}D_{l-1}^{++}(T) e − T 2 /2 D l − 1 ++ ( T ) . For T / 2 ≤ l ≤ T T/2\le l\le T T /2 ≤ l ≤ T , rounding the threshold up to an even integer gives ∣ x ∣ ≥ a / 2 |x|\ge a/2 ∣ x ∣ ≥ a /2 for large a a a , hence the claimed cubic bound; enlarged constants cover bounded a , T a,T a , T . For l < T / 2 l<T/2 l < T /2 , monotonicity bounds the probability by C e − c T 2 Ce^{-cT^2} C e − c T 2 at l = ⌈ T / 2 ⌉ l=\lceil T/2\rceil l = ⌈ T /2 ⌉ . A nonempty event has a ≤ 2 T 2 / 3 a\le2T^{2/3} a ≤ 2 T 2/3 , so this also implies C e − c 0 a 3 Ce^{-c_0a^3} C e − c 0 a 3 .
Lemma 3.13 (Full-space compact and diagonal-block moments). Let L f u l l L^{\mathrm{full}} L full be the full-space Poisson passage metric with bulk intensity 2 2 2 in PNG coordinates, and define its centered rescaling by
L δ f u l l ( x , s ; y , t ) : = δ [ L f u l l ( x δ − 2 , s δ − 3 ; y δ − 2 , t δ − 3 ) − 2 ( t − s ) δ − 3 ] . L_\delta^{\mathrm{full}}(x,s;y,t):=\delta\left[L^{\mathrm{full}}(x\delta^{-2},s\delta^{-3};y\delta^{-2},t\delta^{-3})-2(t-s)\delta^{-3}\right]. L δ full ( x , s ; y , t ) := δ [ L full ( x δ − 2 , s δ − 3 ; y δ − 2 , t δ − 3 ) − 2 ( t − s ) δ − 3 ] . For every compact set K K K of endpoint pairs with time separation bounded away from zero and every p > 0 p>0 p > 0 ,
lim sup δ ↓ 0 E [ ( sup K ∣ L δ f u l l ∣ ) p ] < ∞ . \limsup_{\delta\downarrow0}\mathbb{E}\left[\left(\sup_K|L_\delta^{\mathrm{full}}| \right)^p\right]<\infty. δ ↓ 0 lim sup E [ ( K sup ∣ L δ full ∣ ) p ] < ∞. In particular there are C p , T p C_p,T_p C p , T p such that
E [ sup 0 ≤ y ≤ T 2 / 3 ∣ L f u l l ( y , 0 ; y , T ) − 2 T T 1 / 3 ∣ p ] ≤ C p , T ≥ T p . \mathbb{E}\left[\sup_{0\le y\le T^{2/3}}\left|\frac{L^{\mathrm{full}}(y,0;y,T)-2T}{T^{1/3}}\right|^p\right]\le C_p,\qquad T\ge T_p. E [ 0 ≤ y ≤ T 2/3 sup T 1/3 L full ( y , 0 ; y , T ) − 2 T p ] ≤ C p , T ≥ T p . Consequently, for T ≥ T p T\ge T_p T ≥ T p , R ≥ 0 R\ge0 R ≥ 0 , and a ≥ 1 a\ge1 a ≥ 1 ,
sup y ∈ R P ( L f u l l ( y , 0 ; y , T ) − 2 T > a T 1 / 3 ) ≤ C p a − p , \sup_{y\in R}\mathbb{P}\left(L^{\mathrm{full}}(y,0;y,T)-2T>aT^{1/3}\right)\le C_pa^{-p}, y ∈ R sup P ( L full ( y , 0 ; y , T ) − 2 T > a T 1/3 ) ≤ C p a − p , P { inf 0 ≤ y ≤ R [ L f u l l ( y , 0 ; y , T ) − 2 T ] < − a T 1 / 3 } ≤ C p ( 1 + R / T 2 / 3 ) a − p . \mathbb{P}\left\{\inf_{0\le y\le R}\left[L^{\mathrm{full}}(y,0;y,T)-2T\right]<-aT^{1/3}\right\}\le C_p(1+R/T^{2/3})a^{-p}. P { 0 ≤ y ≤ R inf [ L full ( y , 0 ; y , T ) − 2 T ] < − a T 1/3 } ≤ C p ( 1 + R / T 2/3 ) a − p .
Proof. The proof of [14 Theorem 13.8(3)] establishes compact upper and lower exponential tightness for planar Poisson passage. For the lower bound it applies Proposition 11.3 and Corollary 11.4, as in Theorem 12.7, using the one-point input (13.12). For the upper bound it transfers Corollary 10.8 to planar Poisson passage by vertical compression and then follows Theorem 12.6. These bounds imply all polynomial compact-supremum moments.
To check the normalization, the light-cone coordinates ( u , v ) = ( t + x , t − x ) (u,v)=(t+x,t-x) ( u , v ) = ( t + x , t − x ) send the bulk intensity 2 d t d x 2\,dt\,dx 2 d t d x to d u d v du\,dv d u d v . In [14 Corollary 13.6] , the scale parameters are τ T = 2 T 2 / 3 \tau_T=2T^{2/3} τ T = 2 T 2/3 and χ T = T 1 / 3 \chi_T=T^{1/3} χ T = T 1/3 , with centering h ( u , v ) = u + v h(u,v)=u+v h ( u , v ) = u + v . The cited coordinate map and ours have matrices
M T D V = ( 2 T 2 / 3 T 0 T ) , M T P N G = ( T 2 / 3 T − T 2 / 3 T ) , M_T^{\mathrm{DV}}=\begin{pmatrix}2T^{2/3}&T\\0&T\end{pmatrix},\qquad
M_T^{\mathrm{PNG}}=\begin{pmatrix}T^{2/3}&T\\-T^{2/3}&T\end{pmatrix}, M T DV = ( 2 T 2/3 0 T T ) , M T PNG = ( T 2/3 − T 2/3 T T ) ,
respectively, so
( M T D V ) − 1 M T P N G = ( 1 0 − T − 1 / 3 1 ) ⟶ I . \left(M_T^{\mathrm{DV}}\right)^{-1}M_T^{\mathrm{PNG}}
=\begin{pmatrix}1&0\\-T^{-1/3}&1\end{pmatrix}\longrightarrow I. ( M T DV ) − 1 M T PNG = ( 1 − T − 1/3 0 1 ) ⟶ I .
The images of any fixed compact set therefore stay in a common compact set with positive time separation, and h ∘ M T P N G ( x , t ) = 2 T t h\circ M_T^{\mathrm{PNG}}(x,t)=2Tt h ∘ M T PNG ( x , t ) = 2 T t . The cited estimates give the first assertion in our normalization. Take δ = T − 1 / 3 \delta=T^{-1/3} δ = T − 1/3 and restrict to { ( y , 0 ; y , 1 ) : 0 ≤ y ≤ 1 } \{(y,0;y,1):0\le y\le1\} {( y , 0 ; y , 1 ) : 0 ≤ y ≤ 1 } for the diagonal-block assertion. Translation invariance and Markov’s inequality give the one-point upper bound. A union bound over at most 1 + ⌊ R / T 2 / 3 ⌋ 1+\lfloor R/T^{2/3}\rfloor 1 + ⌊ R / T 2/3 ⌋ translated diagonal blocks gives the last bound, including R = 0 R=0 R = 0 . ∎
Wall passage tightness. For R < ∞ R < \infty R < ∞ and 0 < t 0 < t 1 < ∞ 0 < t_0 < t_1 < \infty 0 < t 0 < t 1 < ∞ , the family
δ [ G exp ( ρ δ ) ( t δ − 3 ) − 2 t δ − 3 ] , ∣ ρ ∣ ≤ R , t ∈ [ t 0 , t 1 ] , δ ↓ 0 , \delta\left[G_{\exp(\rho\delta)}(t\delta^{-3})-2t\delta^{-3}\right],\qquad|\rho|\le R,\quad t\in[t_0,t_1],\quad\delta\downarrow0, δ [ G e x p ( ρ δ ) ( t δ − 3 ) − 2 t δ − 3 ] , ∣ ρ ∣ ≤ R , t ∈ [ t 0 , t 1 ] , δ ↓ 0 ,
is tight. The same holds with boundary intensity zero. It also holds with any fixed boundary intensity c ⋆ ∈ [ 0 , 1 ) c_\star\in[0,1) c ⋆ ∈ [ 0 , 1 ) . Indeed, put η = ( t δ − 3 ) − 1 / 3 \eta=(t\delta^{-3})^{-1/3} η = ( t δ − 3 ) − 1/3 . The first displayed random variable is t 1 / 3 t^{1/3} t 1/3 times η [ G exp ( κ η ) ( η − 3 ) − 2 η − 3 ] \eta[G_{\exp(\kappa\eta)}(\eta^{-3})-2\eta^{-3}] η [ G e x p ( κ η ) ( η − 3 ) − 2 η − 3 ] , where κ = ρ t 1 / 3 \kappa=\rho t^{1/3} κ = ρ t 1/3 . Couple by thinning and bound this quantity between those with parameters κ = ± R t 1 1 / 3 \kappa=\pm Rt_1^{1/3} κ = ± R t 1 1/3 . Theorem 3.3, with one wedge at the wall and time 1, gives tightness of both bounding variables. At intensity zero, Lemma 3.12 controls the lower tail, while G 0 ≤ G 1 G_0\le G_1 G 0 ≤ G 1 and the critical limit control the upper tail. For fixed c ⋆ < 1 c_\star<1 c ⋆ < 1 , the thinning bounds G 0 ≤ G c ⋆ ≤ G 1 G_0\le G_{c_\star}\le G_1 G 0 ≤ G c ⋆ ≤ G 1 give the fixed-subcritical claim.
Lemma 3.14 (Comparison with full-space Poisson passage). Fix a time slab [ s , t ] [s,t] [ s , t ] . Couple G c G_c G c and G 0 G_0 G 0 by retaining the same bulk points, and extend the bulk environment independently to the negative half-line. Denote the resulting full-space passage metric by L f u l l L^{\mathrm{full}} L full . Put W c = G c ( 0 , s ; 0 , t ) W_c=G_c(0,s;0,t) W c = G c ( 0 , s ; 0 , t ) and W 0 = G 0 ( 0 , s ; 0 , t ) W_0=G_0(0,s;0,t) W 0 = G 0 ( 0 , s ; 0 , t ) . For every causally reachable z , y ≥ 0 z,y\ge0 z , y ≥ 0 , simultaneously,
L f u l l ( z , s ; y , t ) − L f u l l ( 0 , s ; 0 , t ) + W 0 ≤ G 0 ( z , s ; y , t ) ≤ G c ( z , s ; y , t ) ≤ L f u l l ( z , s ; y , t ) + W c − W 0 . L^{\mathrm{full}}(z,s;y,t)-L^{\mathrm{full}}(0,s;0,t)+W_0\le G_0(z,s;y,t)\le G_c(z,s;y,t)\le L^{\mathrm{full}}(z,s;y,t)+W_c-W_0. L full ( z , s ; y , t ) − L full ( 0 , s ; 0 , t ) + W 0 ≤ G 0 ( z , s ; y , t ) ≤ G c ( z , s ; y , t ) ≤ L full ( z , s ; y , t ) + W c − W 0 .
Proof. Take an optimizing wall-to-wall path in the half-space environment with no boundary atoms. Restrict first to the finite causal cone of the two endpoint pairs, which contains finitely many Poisson atoms. Any path from ( z , s ) (z,s) ( z , s ) to ( y , t ) (y,t) ( y , t ) that visits the wall, or leaves the half-space, must meet this path before and after each such excursion. Exchange the two paths between their first and last intersections. For comparison of G c G_c G c and G 0 G_0 G 0 , this puts all modified boundary points into the wall-to-wall path, and gives G 0 ( 0 , s ; 0 , t ) + G c ( z , s ; y , t ) ≤ G c ( 0 , s ; 0 , t ) + G 0 ( z , s ; y , t ) G_0(0,s;0,t)+G_c(z,s;y,t)\le G_c(0,s;0,t)+G_0(z,s;y,t) G 0 ( 0 , s ; 0 , t ) + G c ( z , s ; y , t ) ≤ G c ( 0 , s ; 0 , t ) + G 0 ( z , s ; y , t ) . For comparison of the restricted metric G 0 G_0 G 0 and L f u l l L^{\mathrm{full}} L full , the same exchange gives G 0 ( 0 , s ; 0 , t ) + L f u l l ( z , s ; y , t ) ≤ L f u l l ( 0 , s ; 0 , t ) + G 0 ( z , s ; y , t ) G_0(0,s;0,t)+L^{\mathrm{full}}(z,s;y,t)\le L^{\mathrm{full}}(0,s;0,t)+G_0(z,s;y,t) G 0 ( 0 , s ; 0 , t ) + L full ( z , s ; y , t ) ≤ L full ( 0 , s ; 0 , t ) + G 0 ( z , s ; y , t ) . Finally G 0 ≤ L f u l l G_0\le L^{\mathrm{full}} G 0 ≤ L full and G 0 ≤ G c G_0\le G_c G 0 ≤ G c . Path intersections can be included as vertices of weight zero, so this is precisely the finite directed-graph argument of [17 Proposition 3.10 and Corollary 3.11] . □
Lemma 3.15 (Maximizing starting points at large stationary slope). Fix t > 0 t>0 t > 0 and b < ∞ b<\infty b < ∞ . Let S δ , λ S_{\delta,\lambda} S δ , λ be the pinned stationary PNG profile with spectral parameter g = e λ δ / 2 g=e^{\lambda\delta/2} g = e λ δ /2 and boundary intensity e ρ δ e^{\rho\delta} e ρ δ , independent of the future environment, and write
S ^ δ , λ ( z ) = δ S δ , λ ( z δ − 2 ) , L δ ( z ; y ) = δ [ G e ρ δ ( z δ − 2 , 0 ; y δ − 2 , t δ − 3 ) − 2 t δ − 3 ] . \widehat S_{\delta,\lambda}(z)=\delta S_{\delta,\lambda}(z\delta^{-2}),\qquad L_\delta(z;y)=\delta\left[G_{e^{\rho\delta}}(z\delta^{-2},0;y\delta^{-2},t\delta^{-3})-2t\delta^{-3}\right]. S δ , λ ( z ) = δ S δ , λ ( z δ − 2 ) , L δ ( z ; y ) = δ [ G e ρ δ ( z δ − 2 , 0 ; y δ − 2 , t δ − 3 ) − 2 t δ − 3 ] . Let A δ , λ A_{\delta,\lambda} A δ , λ be the event that no maximizing starting point for S ^ δ , λ ( z ) + L δ ( z ; 0 ) \widehat S_{\delta,\lambda}(z)+L_\delta(z;0) S δ , λ ( z ) + L δ ( z ; 0 ) lies in [ 0 , b ] [0,b] [ 0 , b ] . Then, for λ > max ( 0 , 2 ρ ) \lambda>\max(0,2\rho) λ > max ( 0 , 2 ρ ) ,
lim λ → ∞ lim δ ↓ 0 P ( A δ , λ ) = 0. \lim_{\lambda\to\infty}\lim_{\delta\downarrow0}\mathbb{P}(A_{\delta,\lambda})=0. λ → ∞ lim δ ↓ 0 lim P ( A δ , λ ) = 0.
Proof. Set q λ = λ t / 2 q_\lambda=\lambda t/2 q λ = λ t /2 . The sandwich in Lemma 3.14 compares the metric on [ 0 , b ] [0,b] [ 0 , b ] and at q λ q_\lambda q λ with full-space passage, up to tight boundary-to-boundary passage terms. The compact full-space supremum is tight, and its one-point limit at q λ q_\lambda q λ , plus q λ 2 / t q_\lambda^2/t q λ 2 / t , has a law independent of q λ q_\lambda q λ .
The one-parameter stationary limit is the two-line Brownian passage construction with drifts ( λ , − λ ) (\lambda,-\lambda) ( λ , − λ ) and bridge E λ ∼ Exp ( λ / 2 − ρ ) E_\lambda\sim\operatorname{Exp}(\lambda/2-\rho) E λ ∼ Exp ( λ /2 − ρ ) . The elementary Lipschitz bound for this construction gives, for L < ∞ L<\infty L < ∞ ,
sup 0 ≤ z ≤ L ∣ H ρ ( λ , z ) − λ z ∣ ≤ 2 sup 0 ≤ z ≤ L ∣ B 1 ( z ) ∣ + 2 sup 0 ≤ z ≤ L ∣ B 2 ( z ) ∣ + 2 E λ , \sup_{0\le z\le L}|H^\rho(\lambda,z)-\lambda z|
\le2\sup_{0\le z\le L}|B_1(z)|+2\sup_{0\le z\le L}|B_2(z)|+2E_\lambda, 0 ≤ z ≤ L sup ∣ H ρ ( λ , z ) − λ z ∣ ≤ 2 0 ≤ z ≤ L sup ∣ B 1 ( z ) ∣ + 2 0 ≤ z ≤ L sup ∣ B 2 ( z ) ∣ + 2 E λ ,
where the B i B_i B i are centered Brownian motions of variance 2. At L = q λ ∨ b L=q_\lambda\vee b L = q λ ∨ b this error is O P ( λ ) O_{\mathbb{P}}(\sqrt{\lambda}) O P ( λ ) . Consequently, after first sending δ \delta δ to zero, the difference between the value at q λ q_\lambda q λ and the supremum over [ 0 , b ] [0,b] [ 0 , b ] is bounded from below by
λ 2 t / 4 − λ b − O P ( λ ) . \lambda^2t/4-\lambda b-O_{\mathbb{P}}(\sqrt{\lambda}). λ 2 t /4 − λb − O P ( λ ) .
It is positive with probability tending to one as λ → ∞ \lambda\to\infty λ → ∞ . Any maximizing starting point must then exceed b b b . □
Lemma 3.16 (One-sided spatial comparison with stationarity). Fix s < t s<t s < t , b < ∞ b<\infty b < ∞ , and a critical-window or fixed-subcritical boundary regime. For each sufficiently large positive λ \lambda λ , one can construct a height-scaled stationary output profile H ε , λ , t H_{\varepsilon,\lambda,t} H ε , λ , t and an event A ε , λ A_{\varepsilon,\lambda} A ε , λ such that
lim λ → ∞ lim sup ε ↓ 0 P ( A ε , λ c ) = 0 , \lim_{\lambda\to\infty}\limsup_{\varepsilon\downarrow0}\mathbb{P}(A_{\varepsilon,\lambda}^{c})=0, λ → ∞ lim ε ↓ 0 lim sup P ( A ε , λ c ) = 0 , and, on that event, simultaneously for 0 ≤ x ≤ b 0\le x\le b 0 ≤ x ≤ b and 0 ≤ y 1 ≤ y 2 ≤ b 0\le y_1\le y_2\le b 0 ≤ y 1 ≤ y 2 ≤ b ,
L ε ρ ( x , s ; y 2 , t ) − L ε ρ ( x , s ; y 1 , t ) ≤ H ε , λ , t ( y 2 ) − H ε , λ , t ( y 1 ) . (83) L^\rho_\varepsilon(x,s;y_2,t)-L^\rho_\varepsilon(x,s;y_1,t)\le H_{\varepsilon,\lambda,t}(y_2)-H_{\varepsilon,\lambda,t}(y_1). \tag*{(83)} L ε ρ ( x , s ; y 2 , t ) − L ε ρ ( x , s ; y 1 , t ) ≤ H ε , λ , t ( y 2 ) − H ε , λ , t ( y 1 ) . (83) For fixed λ \lambda λ , the output profiles are asymptotically spatially equicontinuous. An analogous inequality controls positive increments in x x x after time reversal.
Proof. Time translation allows s = 0 s=0 s = 0 . Independently of future atoms, take the stationary initial law with g ε = e λ ε 1 / 2 / 2 g_\varepsilon=e^{\lambda\varepsilon^{1/2}/2} g ε = e λ ε 1/2 /2 and the same boundary rate. Let H ε , λ , t H_{\varepsilon,\lambda,t} H ε , λ , t be its height-scaled output recentered at the wall, and let A ε , λ A_{\varepsilon,\lambda} A ε , λ be the event that no maximizing starting point for the output at 0 lies in [ 0 , b ] [0,b] [ 0 , b ] . Lemma 3.15 gives the probability bound. Its proof applies also at a fixed subcritical boundary rate: wall passages are bounded above by any critical-window passage and below by the passage at boundary rate zero, and the stationary horizon limit at each fixed positive slope is already proved in that regime.
Path exchange makes rightmost maximizing starting points nondecreasing in the target. On A ε , λ A_{\varepsilon,\lambda} A ε , λ , a maximizing path for stationary initial data to y 1 y_1 y 1 therefore starts to the right of b b b and crosses an optimizer from x ≤ b x \le b x ≤ b to y 2 ≥ y 1 y_2 \ge y_1 y 2 ≥ y 1 . Exchanging continuations proves (3.21). The recentered output has the stationary initial law. Asymptotic equicontinuity of its uninterpolated profiles follows from the uniform coupling in the proof of Proposition 3.1: the upper-semicontinuous profile and its right-continuous representative differ by at most one, hence by at most ε 1 / 2 \varepsilon^{1/2} ε 1/2 after scaling. Reflecting time on a fixed deterministic slab and reversing passage paths exchanges their endpoints and preserves the bulk and boundary Poisson laws. The conventions ( s , t ] (s,t] ( s , t ] and [ s , t ) [s,t) [ s , t ) give the same scores almost surely, since the deterministic endpoint slices contain no atoms. This gives the comparison in the initial coordinate. □
Lemma 3.17 (The subcritical finite-wedge limit at the wall). Fix c ⋆ ∈ [ 0 , 1 ) c_{\star} \in[0,1) c ⋆ ∈ [ 0 , 1 ) and a nonempty finite initial support P ⊂ R ≥ 0 P \subset\mathbb{R}_{\ge0} P ⊂ R ≥ 0 , with fixed initial heights a z a_z a z , z ∈ P z \in P z ∈ P . The subcritical finite-wedge convergence proved at positive terminal positions extends to finite terminal collections that include 0 0 0 .
Proof. Fix t > 0 t > 0 t > 0 and b > max P b > \max P b > max P . Write X ε , z = L ε − ∞ ( z , 0 ; 0 , t ) X_{\varepsilon,z}=L_{\varepsilon}^{-\infty}(z,0;0,t) X ε , z = L ε − ∞ ( z , 0 ; 0 , t ) and Y ε , z ( h ) = L ε − ∞ ( z , 0 ; h , t ) Y_{\varepsilon,z}(h)=L_{\varepsilon}^{-\infty}(z,0;h,t) Y ε , z ( h ) = L ε − ∞ ( z , 0 ; h , t ) . Lemma 3.16 gives, simultaneously for z ∈ P z \in P z ∈ P and 0 < h < b 0<h<b 0 < h < b ,
X ε , z ≥ Y ε , z ( h ) − Ω ε , λ ( h ) X_{\varepsilon,z} \ge Y_{\varepsilon,z}(h)-\Omega_{\varepsilon,\lambda}(h) X ε , z ≥ Y ε , z ( h ) − Ω ε , λ ( h )
on its comparison event, whose exceptional probability vanishes as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 and then λ → ∞ \lambda\to\infty λ → ∞ . Here
Ω ε , λ ( h ) : = sup u , v ∈ [ 0 , h ] ∣ H ε , λ , t ( u ) − H ε , λ , t ( v ) ∣ . \Omega_{\varepsilon,\lambda}(h):=\sup_{u,v\in[0,h]}\left|H_{\varepsilon,\lambda,t}(u)-H_{\varepsilon,\lambda,t}(v)\right|. Ω ε , λ ( h ) := u , v ∈ [ 0 , h ] sup ∣ H ε , λ , t ( u ) − H ε , λ , t ( v ) ∣ .
is the oscillation of the stationary output over [ 0 , h ] [0,h] [ 0 , h ] ; for fixed λ \lambda λ it tends to zero in probability as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 and then h ↓ 0 h \downarrow0 h ↓ 0 .
For each n ∈ N n \in\mathbb{N} n ∈ N , boundary thinning also gives X ε , z ≤ L ε − n ( z , 0 ; 0 , t ) X_{\varepsilon,z}\le L_{\varepsilon}^{-n}(z,0;0,t) X ε , z ≤ L ε − n ( z , 0 ; 0 , t ) for small ε \varepsilon ε , since c ⋆ < e − n ε 1 / 2 c_{\star}<e^{-n\varepsilon^{1/2}} c ⋆ < e − n ε 1/2 . Theorem 3.3 identifies the limit of this upper bound as L − n ( z , 0 ; 0 , t ) L^{-n}(z,0;0,t) L − n ( z , 0 ; 0 , t ) . The monotone degeneration in the proof of [17 Corollary 5.2] , extended from rational endpoints by its uniform modulus bound in Proposition 4.10(1), sends this law to that of L − ∞ ( z , 0 ; 0 , t ) L^{-\infty}(z,0;0,t) L − ∞ ( z , 0 ; 0 , t ) as n → ∞ n \to\infty n → ∞ . The lower bound has the same limiting law: by Proposition 3.9, Y ε , z ( h ) ⇒ L − ∞ ( z , 0 ; h , t ) Y_{\varepsilon,z}(h) \Rightarrow L^{-\infty}(z,0;h,t) Y ε , z ( h ) ⇒ L − ∞ ( z , 0 ; h , t ) , and target continuity permits h ↓ 0 h \downarrow0 h ↓ 0 . Applying bounded increasing Lipschitz functions to these bounds, first letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , then h ↓ 0 h \downarrow0 h ↓ 0 and λ → ∞ \lambda\to\infty λ → ∞ in the lower bound, and n → ∞ n \to\infty n → ∞ in the upper bound, proves X ε , z ⇒ L − ∞ ( z , 0 ; 0 , t ) X_{\varepsilon,z} \Rightarrow L^{-\infty}(z,0;0,t) X ε , z ⇒ L − ∞ ( z , 0 ; 0 , t ) , including z = 0 z=0 z = 0 .
These marginal limits and the one-sided comparison imply
lim h ↓ 0 lim sup ε ↓ 0 P ( max z ∈ P ∣ X ε , z − Y ε , z ( h ) ∣ > η ) = 0. (84) \lim_{h\downarrow0}\limsup_{\varepsilon\downarrow0}\mathbb{P}\left(\max_{z\in P}\left|X_{\varepsilon,z}-Y_{\varepsilon,z}(h)\right|>\eta\right)=0. \tag*{(84)} h ↓ 0 lim ε ↓ 0 lim sup P ( z ∈ P max ∣ X ε , z − Y ε , z ( h ) ∣ > η ) = 0. (84)
Indeed, choose a bounded strictly increasing Lipschitz function. The expected difference of its values at the two tight marginals tends to zero as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 and then h ↓ 0 h \downarrow0 h ↓ 0 . The negative part is bounded by the stationary modulus and the exit exception. On a compact height interval, a positive difference exceeding η \eta η forces a fixed positive function difference. For each fixed height cutoff, let λ → ∞ \lambda\to\infty λ → ∞ and then remove the cutoff to obtain absolute proximity; a finite union gives (3.22).
Maxima over P P P preserve this approximation. Apply the positive-terminal finite-wedge limit jointly at h h h and the other terminal positions, then let h ↓ 0 h \downarrow0 h ↓ 0 . Continuity of the limiting finite maximum max z ∈ P { a z + L − ∞ ( z , 0 ; y , t ) } \max_{z\in P}\{a_z+L^{-\infty}(z,0;y,t)\} max z ∈ P { a z + L − ∞ ( z , 0 ; y , t )} proves the claim. □
Short-time bounds and localization
Short vertical-passage lower bounds and stationary upper bounds will control nearby observation times and localize the variational supremum, allowing the passage from fixed-time laws to space–time convergence.
Lemma 3.18 (Lower bounds for short vertical passages). Fix b ≥ 2 b \ge2 b ≥ 2 and α ∈ ( 0 , 1 / 3 ) \alpha\in(0,1/3) α ∈ ( 0 , 1/3 ) , and choose p p p so that γ : = p ( 1 / 3 − α ) − 5 / 3 > 0 \gamma:=p(1/3-\alpha)-5/3>0 γ := p ( 1/3 − α ) − 5/3 > 0 . There is a constant C α C_{\alpha} C α such that, with
E ε t i m e ( b , q ) = { inf 0 ≤ x ≤ b t − s ≤ q − b ≤ s ≤ t ≤ b L ε ρ ( x , s ; x , t ) ≥ − C α q α } , \mathcal{E}_{\varepsilon}^{\mathrm{time}}(b,q)=\left\{\inf_{\substack{0\le x\le b\\ t-s\le q\\ -b\le s\le t\le b}}L_{\varepsilon}^{\rho}(x,s;x,t)\ge-C_{\alpha}q^{\alpha}\right\}, E ε time ( b , q ) = ⎩ ⎨ ⎧ 0 ≤ x ≤ b t − s ≤ q − b ≤ s ≤ t ≤ b inf L ε ρ ( x , s ; x , t ) ≥ − C α q α ⎭ ⎬ ⎫ , one has
lim q ↓ 0 lim inf ε ↓ 0 P ( E ε t i m e ( b , q ) ) = 1. \lim_{q\downarrow0}\liminf_{\varepsilon\downarrow0}\mathbb{P}\left(\mathcal{E}_{\varepsilon}^{\mathrm{time}}(b,q)\right)=1. q ↓ 0 lim ε ↓ 0 lim inf P ( E ε time ( b , q ) ) = 1. The conclusion is uniform over all nonnegative boundary rates.
Proof. First fix a duration r r r , put T = r ε − 3 / 2 T=r\varepsilon^{-3/2} T = r ε − 3/2 , and use the full-space extension in Lemma 3.14. That comparison gives, with Y = y ε − 1 Y=y\varepsilon^{-1} Y = y ε − 1 and S = s ε − 3 / 2 S=s\varepsilon^{-3/2} S = s ε − 3/2 ,
G c ( Y , S ; Y , S + T ) ≥ L f u l l ( Y , S ; Y , S + T ) − L f u l l ( 0 , S ; 0 , S + T ) + G 0 ( 0 , S ; 0 , S + T ) . G_c(Y,S;Y,S+T)\ge L^{\mathrm{full}}(Y,S;Y,S+T)-L^{\mathrm{full}}(0,S;0,S+T)+G_0(0,S;0,S+T). G c ( Y , S ; Y , S + T ) ≥ L full ( Y , S ; Y , S + T ) − L full ( 0 , S ; 0 , S + T ) + G 0 ( 0 , S ; 0 , S + T ) .
Split a lower deviation into three fixed fractions and apply Lemmas 3.13 and 3.12. Since b ε − 1 T − 2 / 3 = b r − 2 / 3 b\varepsilon^{-1}T^{-2/3}=br^{-2/3} b ε − 1 T − 2/3 = b r − 2/3 and ε 1 / 2 T 1 / 3 = r 1 / 3 \varepsilon^{1/2}T^{1/3}=r^{1/3} ε 1/2 T 1/3 = r 1/3 , for T ≥ T p T\ge T_p T ≥ T p and a ≥ 1 a\ge1 a ≥ 1 ,
P ( inf 0 ≤ y ≤ b L ε ρ ( y , s ; y , s + r ) < − a r 1 / 3 ) ≤ C p ( 1 + b r − 2 / 3 ) a − p . (85) \mathbb{P}\left(\inf_{0\le y\le b}L_{\varepsilon}^{\rho}(y,s;y,s+r)<-ar^{1/3}\right)\le C_p(1+br^{-2/3})a^{-p}. \tag*{(85)} P ( 0 ≤ y ≤ b inf L ε ρ ( y , s ; y , s + r ) < − a r 1/3 ) ≤ C p ( 1 + b r − 2/3 ) a − p . (85)
The bound is uniform in s s s and the boundary rate. Since G c ≥ 0 G_c\ge0 G c ≥ 0 , the event is empty if a ≥ 2 T 2 / 3 a\ge2T^{2/3} a ≥ 2 T 2/3 . Fix 0 < q < 1 0<q<1 0 < q < 1 and use the nested dyadic grids with spacings q 2 − j q2^{-j} q 2 − j . At level j ≥ 0 j\geq0 j ≥ 0 , use the grid of intervals of length r j = q 2 − j r_j=q2^{-j} r j = q 2 − j contained in [ − b − 1 , b + 1 ] [-b-1,b+1] [ − b − 1 , b + 1 ] . There are at most C b r j − 1 Cb r_j^{-1} C b r j − 1 such intervals. Apply (3.23) with threshold r j α r_j^\alpha r j α , that is, a j = r j α − 1 / 3 a_j=r_j^{\alpha-1/3} a j = r j α − 1/3 . For all sufficiently small q q q , a j ≥ 1 a_j\geq1 a j ≥ 1 . The union bound gives
P ( ∃ j , I : inf 0 ≤ x ≤ b L ε ρ ( x , inf I ; x , sup I ) < − r j α ) ≤ C b ∑ j ≥ 0 r j − 1 ( 1 + b r j − 2 / 3 ) r j p ( 1 / 3 − α ) ≤ C b , p q γ . (86) \mathbb{P}\left(\exists j,I:\inf_{0\leq x\leq b}L^\rho_\varepsilon(x,\inf I;x,\sup I)<-r_j^\alpha\right)\leq Cb\sum_{j\geq0}r_j^{-1}\left(1+br_j^{-2/3}\right)r_j^{p(1/3-\alpha)}\leq C_{b,p}q^\gamma. \tag*{(86)} P ( ∃ j , I : 0 ≤ x ≤ b inf L ε ρ ( x , inf I ; x , sup I ) < − r j α ) ≤ C b j ≥ 0 ∑ r j − 1 ( 1 + b r j − 2/3 ) r j p ( 1/3 − α ) ≤ C b , p q γ . (86)
The estimate applies to every nontrivial event in this sum once ε \varepsilon ε is sufficiently small. Indeed, if 2 r j ε − 1 ≤ r j α 2r_j\varepsilon^{-1}\leq r_j^\alpha 2 r j ε − 1 ≤ r j α , the event is empty. Otherwise
r j > ( ε / 2 ) 1 / ( 1 − α ) , T j = r j ε − 3 / 2 > 2 − 1 / ( 1 − α ) ε 1 / ( 1 − α ) − 3 / 2 ⟶ ∞ , r_j>(\varepsilon/2)^{1/(1-\alpha)},\qquad T_j=r_j\varepsilon^{-3/2}>2^{-1/(1-\alpha)}\varepsilon^{1/(1-\alpha)-3/2}\longrightarrow\infty, r j > ( ε /2 ) 1/ ( 1 − α ) , T j = r j ε − 3/2 > 2 − 1/ ( 1 − α ) ε 1/ ( 1 − α ) − 3/2 ⟶ ∞ ,
where the exponent is negative because α < 1 / 3 \alpha<1/3 α < 1/3 . Moreover a j < 2 T j 2 / 3 a_j<2T_j^{2/3} a j < 2 T j 2/3 in exactly this nontrivial range. Hence (3.24) applies uniformly to all nonempty events, and its right side tends to zero as q ↓ 0 q\downarrow0 q ↓ 0 .
On the complementary event, decompose any interval [ s , t ] ⊂ [ − b , b ] [s,t]\subset[-b,b] [ s , t ] ⊂ [ − b , b ] of length at most q q q into maximal dyadic subintervals. There are at most two of each length q 2 − j q2^{-j} q 2 − j . Centered superadditivity holds exactly with the half-open convention ( s , t ] (s,t] ( s , t ] , including at random composition times. Finite compositions therefore give the lower bound − 2 ∑ j ≥ 0 ( q 2 − j ) α = − 2 q α 1 − 2 − α -2\sum_{j\geq0}(q2^{-j})^\alpha=-\frac{2q^\alpha}{1-2^{-\alpha}} − 2 ∑ j ≥ 0 ( q 2 − j ) α = − 1 − 2 − α 2 q α . For arbitrary real endpoints, first use a finite inner dyadic approximation. The two unfilled endpoint intervals have total length at most 2 q 2 − n 2q2^{-n} 2 q 2 − n ; their centered passage values are bounded below by − 4 q 2 − n ε − 1 -4q2^{-n}\varepsilon^{-1} − 4 q 2 − n ε − 1 . Sending n → ∞ n\to\infty n → ∞ at fixed ε \varepsilon ε removes these remainders. This proves the assertion with C α = 2 / ( 1 − 2 − α ) C_\alpha=2/(1-2^{-\alpha}) C α = 2/ ( 1 − 2 − α ) . □
Write U ε = h ε ρ U_\varepsilon=h^\rho_\varepsilon U ε = h ε ρ for the centered height evolution. The functions f ε f_\varepsilon f ε are rescalings of PNG initial profiles in X a d m \mathcal{X}_{\mathrm{adm}} X adm .
Lemma 3.19 (Stationary upper bounds). Fix a critical-window or fixed-subcritical boundary regime, and suppose f ε ( z ) ≤ C ( 1 + z ) f_\varepsilon(z)\leq C(1+z) f ε ( z ) ≤ C ( 1 + z ) for one finite C ≥ 0 C\geq0 C ≥ 0 . For every fixed t > 0 t>0 t > 0 ,
lim A → ∞ lim sup ε ↓ 0 P ( ∃ z ≥ 0 : U ε ( t , z ) > A ( 1 + z ) ) = 0. (87) \lim_{A\to\infty}\limsup_{\varepsilon\downarrow0}\mathbb{P}\left(\exists z\geq0:U_\varepsilon(t,z)>A(1+z)\right)=0. \tag*{(87)} A → ∞ lim ε ↓ 0 lim sup P ( ∃ z ≥ 0 : U ε ( t , z ) > A ( 1 + z ) ) = 0. (87) Moreover, the centered evolution of a dominating stationary profile has a tight upper supremum on every compact subset of ( 0 , ∞ ) × [ 0 , ∞ ) (0,\infty)\times[0,\infty) ( 0 , ∞ ) × [ 0 , ∞ ) .
Proof. Put ρ ˉ = ρ \bar\rho=\rho ρ ˉ = ρ in the critical window and ρ ˉ = 0 \bar\rho=0 ρ ˉ = 0 for a fixed subcritical rate. The boundary rate is then at most c ˉ ε = e ρ ˉ ε 1 / 2 \bar c_\varepsilon=e^{\bar\rho\varepsilon^{1/2}} c ˉ ε = e ρ ˉ ε 1/2 for small ε \varepsilon ε . Choose λ > max { C , 0 , 2 ρ ˉ } \lambda>\max\{C,0,2\bar\rho\} λ > max { C , 0 , 2 ρ ˉ } and let H ε H_\varepsilon H ε be the scaled, pinned stationary profile for c ˉ ε \bar c_\varepsilon c ˉ ε and g ε = e λ ε 1 / 2 / 2 g_\varepsilon=e^{\lambda\varepsilon^{1/2}/2} g ε = e λ ε 1/2 /2 , independent of the future. There are deterministic C < d < λ C<d<\lambda C < d < λ and D < ∞ D<\infty D < ∞ , and tight nonnegative random variables K ε K_\varepsilon K ε , such that
d z − K ε ≤ H ε ( z ) ≤ D z + K ε , z ≥ 0. (88) dz-K_\varepsilon\leq H_\varepsilon(z)\leq Dz+K_\varepsilon,\qquad z\geq0. \tag*{(88)} d z − K ε ≤ H ε ( z ) ≤ D z + K ε , z ≥ 0. (88)
Indeed, write δ = ε 1 / 2 \delta=\varepsilon^{1/2} δ = ε 1/2 . The two-line formula gives H ε ( z ) ≥ δ S 1 ( z δ − 2 ) − δ Y H_\varepsilon(z)\geq\delta S_1(z\delta^{-2})-\delta Y H ε ( z ) ≥ δ S 1 ( z δ − 2 ) − δ Y , with Y ∼ Geo ( e ( ρ ˉ − λ / 2 ) δ ) Y\sim\operatorname{Geo}(e^{(\bar\rho-\lambda/2)\delta}) Y ∼ Geo ( e ( ρ ˉ − λ /2 ) δ ) . For a sufficiently small fixed β > 0 \beta>0 β > 0 , the cumulant per unit distance of − β [ δ S 1 ( z δ − 2 ) − δ z ] -\beta[\delta S_1(z\delta^{-2})-\delta z] − β [ δ S 1 ( z δ − 2 ) − δ z ] equals − β ( λ − d ) + β 2 + O ( δ ) < 0 -\beta(\lambda-d)+\beta^2+O(\delta)<0 − β ( λ − d ) + β 2 + O ( δ ) < 0 . Its exponential maximal inequality and the uniform exponential tail of δ Y \delta Y δ Y prove the lower bound. Applying the same inequality with both signs to each of the two walks bounds their absolute values by a deterministic linear function plus tight constants. Substitution in the running-maximum formula proves the upper bound.
Let Z ε Z_\varepsilon Z ε be the centered evolution from H ε H_\varepsilon H ε with boundary rate c ˉ ε \bar c_\varepsilon c ˉ ε . The one-parameter stationary limit, (3.26), and Corollary 3.8 give tightness of Z ε ( t , 0 ) Z_\varepsilon(t,0) Z ε ( t , 0 ) for each fixed t > 0 t>0 t > 0 . Stationarity gives
Z ε ( t , ⋅ ) − Z ε ( t , 0 ) = d H ε . Z_\varepsilon(t,\cdot)-Z_\varepsilon(t,0)\overset{\mathrm d}{=}H_\varepsilon. Z ε ( t , ⋅ ) − Z ε ( t , 0 ) = d H ε .
Since f ε ≤ H ε + C + K ε f_\varepsilon\leq H_\varepsilon+C+K_\varepsilon f ε ≤ H ε + C + K ε , the variational formula and boundary thinning imply U ε ( t , z ) ≤ Z ε ( t , z ) + C + K ε U_\varepsilon(t,z)\leq Z_\varepsilon(t,z)+C+K_\varepsilon U ε ( t , z ) ≤ Z ε ( t , z ) + C + K ε . The stationary upper bound and fixed-time tightness at zero prove (3.25).
For the last assertion, cover a compact time interval by a finite time grid of mesh q q q , extending it one step to the right. On the event in Lemma 3.18, if t y t_y t y is the next grid time after t t t , composition gives, simultaneously on a compact spatial interval, Z ε ( t , y ) ≤ Z ε ( t y , y ) + C q α Z_\varepsilon(t,y)\leq Z_\varepsilon(t_y,y)+Cq^\alpha Z ε ( t , y ) ≤ Z ε ( t y , y ) + C q α . At the finitely many grid times, the height at zero is tight and the stationary spatial profiles satisfy (3.26). First choose q q q so that the exceptional probability is small, then bound these finitely many variables. This proves the compact upper bound. □
Lemma 3.20 (Localization under a common linear upper bound). Fix a critical-window or fixed-subcritical regime. Suppose f ε → f ≢ − ∞ f_\varepsilon\to f\not\equiv-\infty f ε → f ≡ − ∞ locally in hypograph and f ε ( z ) ≤ C ( 1 + z ) f_\varepsilon(z)\leq C(1+z) f ε ( z ) ≤ C ( 1 + z ) . For every compact K ⊂ ( 0 , ∞ ) × [ 0 , ∞ ) \mathcal{K}\subset(0,\infty)\times[0,\infty) K ⊂ ( 0 , ∞ ) × [ 0 , ∞ ) ,
lim R → ∞ lim sup ε ↓ 0 P ( sup ( z > R ) { f ε ( z ) + L ε ρ ( z , 0 ; y , t ) } > sup 0 ≤ z ≤ R { f ε ( z ) + L ε ρ ( z , 0 ; y , t ) } for some ( t , y ) ∈ K ) = 0. \lim_{R\to\infty}\limsup_{\varepsilon\downarrow0}\mathbb{P}\left(\sup_{(z>R)}\{f_\varepsilon(z)+L^\rho_\varepsilon(z,0;y,t)\}>\sup_{0\leq z\leq R}\{f_\varepsilon(z)+L^\rho_\varepsilon(z,0;y,t)\}\text{ for some }(t,y)\in\mathcal{K}\right)=0. R → ∞ lim ε ↓ 0 lim sup P ( ( z > R ) sup { f ε ( z ) + L ε ρ ( z , 0 ; y , t )} > 0 ≤ z ≤ R sup { f ε ( z ) + L ε ρ ( z , 0 ; y , t )} for some ( t , y ) ∈ K ) = 0.
Proof. Use the dominating critical-window stationary profile from the preceding lemma. Its lower bound gives, simultaneously for z > R z>R z > R , H ε ( z ) − f ε ( z ) ≥ ( d − C ) R − C − K ε H_\varepsilon(z)-f_\varepsilon(z)\geq(d-C)R-C-K_\varepsilon H ε ( z ) − f ε ( z ) ≥ ( d − C ) R − C − K ε . By the variational formula and boundary thinning, the tail contribution is therefore at most Z ˉ ε ( t , y ) − ( d − C ) R + C + K ε \bar Z_\varepsilon(t,y)-(d-C)R+C+K_\varepsilon Z ˉ ε ( t , y ) − ( d − C ) R + C + K ε . Its supremum over K \mathcal{K} K tends to − ∞ -\infty − ∞ in probability as R → ∞ R\to\infty R → ∞ , uniformly for small ε \varepsilon ε .
Choose z 0 z_0 z 0 with f ( z 0 ) > − ∞ f(z_0)>-\infty f ( z 0 ) > − ∞ and a hypograph recovery sequence z ε → z 0 z_\varepsilon\to z_0 z ε → z 0 with f ε ( z ε ) → f ( z 0 ) f_\varepsilon(z_\varepsilon)\to f(z_0) f ε ( z ε ) → f ( z 0 ) . The comparison with full-space passage gives
L ε ρ ( z ε , 0 ; y , t ) ≥ L ε f u l l ( z ε , 0 ; y , t ) − L ε f u l l ( 0 , 0 ; 0 , t ) + W 0 , ε ( t ) , (89) L^\rho_\varepsilon(z_\varepsilon,0;y,t)\geq L^{\mathrm{full}}_{\sqrt{\varepsilon}}(z_\varepsilon,0;y,t)-L^{\mathrm{full}}_{\sqrt{\varepsilon}}(0,0;0,t)+W_{0,\varepsilon}(t), \tag*{(89)} L ε ρ ( z ε , 0 ; y , t ) ≥ L ε full ( z ε , 0 ; y , t ) − L ε full ( 0 , 0 ; 0 , t ) + W 0 , ε ( t ) , (89)
@page 42/48 starts_mid=1 ends_mid=1 where W 0 , ε ( t ) W_{0,\varepsilon}(t) W 0 , ε ( t ) is the centered wall-to-wall passage with no boundary atoms. The two full-space terms have compact upper and lower bounds by Lemma 3.13. The wall lower-tail estimate bounds W 0 , ε W_{0,\varepsilon} W 0 , ε below at finitely many positive grid times. Between consecutive grid times, composition and Lemma 3.18, applied at boundary rate zero, give the same lower bound up to C q α Cq^{\alpha} C q α . Thus this one-source contribution has a tight lower infimum over K K K . For large R R R the source lies in [ 0 , R ] [0,R] [ 0 , R ] , whereas the tail tends to − ∞ -\infty − ∞ . This proves the assertion. □
Proposition 3.21 (Fixed-time transition limit in the subcritical regime). The conclusion of Theorem 3.3 also holds for a fixed boundary rate c ⋆ ∈ [ 0 , 1 ) c_{\star} \in[0,1) c ⋆ ∈ [ 0 , 1 ) , with limiting parameter ρ = − ∞ \rho=-\infty ρ = − ∞ .
Proof. Fix t > 0 t>0 t > 0 and finitely many terminal positions y i ≥ 0 y_i \ge0 y i ≥ 0 . Write L δ ( z ; y ) = L δ 2 − ∞ ( z , 0 ; y , t ) L_\delta(z;y)=L^{-\infty}_{\delta^2}(z,0;y,t) L δ ( z ; y ) = L δ 2 − ∞ ( z , 0 ; y , t ) . Lemma 3.20 reduces the variational supremum to [ 0 , R ] [0,R] [ 0 , R ] , with an error probability vanishing as R → ∞ R\to\infty R → ∞ , uniformly for small δ \delta δ . Choose R R R large enough to contain a point where f f f is finite, and to be a continuity point of f f f in the extended-real topology.
Take a partition π : 0 = a 0 < a 1 < ⋯ < a N = R \pi:0=a_0<a_1<\cdots<a_N=R π : 0 = a 0 < a 1 < ⋯ < a N = R whose positive endpoints are continuity points of f f f , and put
∣ π ∣ : = max 0 ≤ j < N ( a j + 1 − a j ) , |\pi|:=\max_{0\leq j<N}(a_{j+1}-a_j), ∣ π ∣ := 0 ≤ j < N max ( a j + 1 − a j ) ,
M j , δ : = sup [ a j , a j + 1 ] f δ , M j : = sup [ a j , a j + 1 ] f , 0 ≤ j < N . M_{j,\delta}:=\sup_{[a_j,a_{j+1}]}f_\delta,\qquad M_j:=\sup_{[a_j,a_{j+1}]}f,\qquad0\leq j<N. M j , δ := [ a j , a j + 1 ] sup f δ , M j := [ a j , a j + 1 ] sup f , 0 ≤ j < N .
Local hypograph convergence gives M j , δ → M j M_{j,\delta}\to M_j M j , δ → M j in R ∪ { − ∞ } \mathbb{R}\cup\{-\infty\} R ∪ { − ∞ } . The upper bound follows by compactness and the hypograph limsup inequality. For the lower bound, approximate an endpoint maximizer from inside the cell using continuity, and then use a recovery sequence at an interior point. At the endpoint 0 0 0 , a recovery sequence already lies in the first cell. Such partitions can have arbitrarily small mesh, since the continuity points of an upper-semicontinuous function are dense after a bounded continuous change of its range.
Apply the initial-coordinate version of Lemma 3.16 on a compact interval containing [ 0 , R ] [0,R] [ 0 , R ] and all y i y_i y i . Outside an event whose probability vanishes as the auxiliary stationary parameter λ → ∞ \lambda\to\infty λ → ∞ , it gives, simultaneously in i i i ,
max j < N { M j , δ + L δ ( a j + 1 ; y i ) } − Ω δ , λ ( ∣ π ∣ ) ≤ sup 0 ≤ z ≤ R { f δ ( z ) + L δ ( z ; y i ) } ≤ max j < N { M j , δ + L δ ( a j ; y i ) } + Ω δ , λ ( ∣ π ∣ ) . \max_{j<N}\{M_{j,\delta}+L_\delta(a_{j+1};y_i)\}-\Omega_{\delta,\lambda}(|\pi|)
\leq\sup_{0\leq z\leq R}\{f_\delta(z)+L_\delta(z;y_i)\}
\leq\max_{j<N}\{M_{j,\delta}+L_\delta(a_j;y_i)\}+\Omega_{\delta,\lambda}(|\pi|). j < N max { M j , δ + L δ ( a j + 1 ; y i )} − Ω δ , λ ( ∣ π ∣ ) ≤ 0 ≤ z ≤ R sup { f δ ( z ) + L δ ( z ; y i )} ≤ j < N max { M j , δ + L δ ( a j ; y i )} + Ω δ , λ ( ∣ π ∣ ) .
Here Ω δ , λ ( h ) \Omega_{\delta,\lambda}(h) Ω δ , λ ( h ) denotes the maximum of the stationary spatial moduli supplied by the initial-coordinate comparison, over the finitely many terminal positions on the chosen compact interval. For each fixed λ \lambda λ , it satisfies
lim h ↓ 0 lim sup δ ↓ 0 P { Ω δ , λ ( h ) > η } = 0 ( η > 0 ) . \lim_{h\downarrow0}\limsup_{\delta\downarrow0}\mathbb{P}\{\Omega_{\delta,\lambda}(h)>\eta\}=0\qquad(\eta>0). h ↓ 0 lim δ ↓ 0 lim sup P { Ω δ , λ ( h ) > η } = 0 ( η > 0 ) .
For a fixed partition, both bounding vectors converge by the finite-wedge limits, including terminal positions at zero. Varying finite heights M j , δ → M j M_{j,\delta}\to M_j M j , δ → M j do not change this conclusion: the max-plus map is Lipschitz in the finitely many initial heights, and replacing M j M_j M j by δ ⌊ M j / δ ⌋ \delta\lfloor M_j/\delta\rfloor δ ⌊ M j / δ ⌋ costs at most δ \delta δ . Cells with M j = − ∞ M_j=-\infty M j = − ∞ may be omitted, since their heights tend to − ∞ -\infty − ∞ and the finitely many single-source passage values are tight.
Let f R = f f^R=f f R = f on [ 0 , R ] [0,R] [ 0 , R ] and f R = − ∞ f^R=-\infty f R = − ∞ elsewhere. The limiting left- and right-endpoint wedge profiles, with heights M j M_j M j , both converge to f R f^R f R in hypograph as ∣ π ∣ → 0 |\pi|\to0 ∣ π ∣ → 0 . Indeed, each cell maximizer is within ∣ π ∣ |\pi| ∣ π ∣ of its assigned endpoint, while upper semicontinuity supplies the limsup inequality. They have a common linear upper bound, for example 2 C ( 1 + z ) 2C(1+z) 2 C ( 1 + z ) when ∣ π ∣ ≤ 1 |\pi|\leq1 ∣ π ∣ ≤ 1 . Continuity of the target fixed-point semigroup [17 Theorem 4.11] therefore makes both limiting vectors converge to ( h − ∞ ( t , y i ; f R ) ) i (h^{-\infty}(t,y_i;f^R))_i ( h − ∞ ( t , y i ; f R ) ) i . Apply the coordinatewise brackets at continuity points of its joint distribution function, taking δ ↓ 0 \delta\downarrow0 δ ↓ 0 , then ∣ π ∣ ↓ 0 |\pi|\downarrow0 ∣ π ∣ ↓ 0 at fixed λ \lambda λ , followed by λ → ∞ \lambda\to\infty λ → ∞ . This proves the transition limit for the truncated supremum.
Finally, f R → f f^R\to f f R → f locally in hypograph with the same linear upper bound. Target semigroup continuity and the prelimiting localization lemma allow R → ∞ R\to\infty R → ∞ , proving the assertion. □
Spatial tightness and the Markov limit
Stationary comparison and localization upgrade fixed-time vector laws to spatial profile laws. Conditioning on the actual PNG profiles and using independent future Poisson slabs then gives joint laws at finitely many times. Short vertical-passage bounds control the intervals between them and complete the positive-time field limit.
Lemma 3.22 (Spatial profile convergence). Under the hypotheses of Lemma 3.20, for each t > 0 t>0 t > 0 ,
U ε ( t , ⋅ ) ⟹ h ρ ( t , ⋅ ; f ) locally uniformly in space, U_\varepsilon(t,\cdot)\Longrightarrow h^\rho(t,\cdot;f)\quad\text{locally uniformly in space,} U ε ( t , ⋅ ) ⟹ h ρ ( t , ⋅ ; f ) locally uniformly in space, with linear spatial interpolation on a mesh tending to zero. The interpolation error tends to zero in probability on compact intervals.
Proof. Theorem 3.3 and Proposition 3.21 give the finite-dimensional limits in the two boundary regimes. Fix R o b s < ∞ R_{\mathrm{obs}}<\infty R obs < ∞ . Localization allows us, with arbitrarily high probability, to restrict every maximizing source for y ∈ [ 0 , R o b s ] y\in[0,R_{\mathrm{obs}}] y ∈ [ 0 , R obs ] to [ 0 , R ] [0,R] [ 0 , R ] , for a fixed R R R . Apply Lemma 3.16 with b ≥ R ∨ R o b s b\geq R\vee R_{\mathrm{obs}} b ≥ R ∨ R obs . Taking the supremum over the same initial sources preserves its inequality: on its comparison event,
U ε ( t , y 2 ) − U ε ( t , y 1 ) ≤ H ε , λ ( y 2 ) − H ε , λ ( y 1 ) , 0 ≤ y 1 ≤ y 2 ≤ R o b s . U_\varepsilon(t,y_2)-U_\varepsilon(t,y_1)\leq H_{\varepsilon,\lambda}(y_2)-H_{\varepsilon,\lambda}(y_1),\qquad0\leq y_1\leq y_2\leq R_{\mathrm{obs}}. U ε ( t , y 2 ) − U ε ( t , y 1 ) ≤ H ε , λ ( y 2 ) − H ε , λ ( y 1 ) , 0 ≤ y 1 ≤ y 2 ≤ R obs .
Let
Ω ε , λ ( h ) : = sup u , v ∈ [ 0 , R o b s + 1 ] ∣ u − v ∣ ≤ h ∣ H ε , λ ( u ) − H ε , λ ( v ) ∣ \Omega_{\varepsilon,\lambda}(h):=
\sup_{\substack{u,v\in[0,R_{\mathrm{obs}}+1]\\ |u-v|\leq h}}
\left|H_{\varepsilon,\lambda}(u)-H_{\varepsilon,\lambda}(v)\right| Ω ε , λ ( h ) := u , v ∈ [ 0 , R obs + 1 ] ∣ u − v ∣ ≤ h sup ∣ H ε , λ ( u ) − H ε , λ ( v ) ∣
be the spatial modulus of the stationary profile. For neighboring mesh points y − ≤ y ≤ y + y^{-} \le y \le y^{+} y − ≤ y ≤ y + at distance at most h h h , this gives
U ε ( t , y + ) − Ω ε , λ ( h ) ≤ U ε ( t , y ) ≤ U ε ( t , y − ) + Ω ε , λ ( h ) . U_{\varepsilon}(t,y^{+})-\Omega_{\varepsilon,\lambda}(h) \le U_{\varepsilon}(t,y) \le U_{\varepsilon}(t,y^{-})+\Omega_{\varepsilon,\lambda}(h). U ε ( t , y + ) − Ω ε , λ ( h ) ≤ U ε ( t , y ) ≤ U ε ( t , y − ) + Ω ε , λ ( h ) .
For fixed h h h , the finite mesh vector converges to that of the continuous fixed-point profile. At fixed λ \lambda λ , the stationary modulus tends to zero in probability as h ↓ 0 h \downarrow0 h ↓ 0 , by the one-parameter stationary limit. First take ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , then h ↓ 0 h \downarrow0 h ↓ 0 , then λ → ∞ \lambda\to\infty λ → ∞ , and finally remove the localization error by increasing R R R . These brackets prove asymptotic spatial equicontinuity, including negative increments. Tightness identifies the spatial law, and the same modulus bounds the interpolation error on any deterministic vanishing mesh.
In C l o c C_{\mathrm{loc}} C loc statements, interpolate on expanding spatial intervals, for example [ 0 , ε − 1 / 4 ] [0,\varepsilon^{-1/4}] [ 0 , ε − 1/4 ] , replacing any − ∞ -\infty − ∞ mesh value by zero, and extend constantly beyond the right endpoint. A recovered finite initial source makes every fixed positive-time spatial compact causally reachable for small ε \varepsilon ε , so these extensions have no effect on the local limits. Markov conditioning always uses the actual upper-semicontinuous PNG profiles.
Under a common linear upper bound, the target transition is continuous for local hypograph convergence. Indeed, the parabolic shape bound in [17 Proposition 4.10(2)] makes the source contribution tend to − ∞ -\infty − ∞ uniformly on positive-time terminal compacts, while a recovery sequence at one finite initial value gives a uniform lower candidate. All maximizing sources therefore lie in one random compact set. Hypograph convergence and continuity of the landscape then give uniform convergence of the variational maxima.
Proposition 3.23 (Convergence at finitely many observation times). Under the same initial-data hypotheses, for every 0 < t 1 < ⋯ < t k 0 < t_{1} < \cdots< t_{k} 0 < t 1 < ⋯ < t k ,
( U ε ( t j , ⋅ ) ) j = 1 k ⟹ ( h ρ ( t j , ⋅ ; f ) ) j = 1 k in C l o c ( [ 0 , ∞ ) ) k , \left(U_{\varepsilon}(t_{j},\cdot)\right)_{j=1}^{k} \Longrightarrow\left(h^{\rho}(t_{j},\cdot;f)\right)_{j=1}^{k}\quad\text{in }C_{\mathrm{loc}}([0,\infty))^{k}, ( U ε ( t j , ⋅ ) ) j = 1 k ⟹ ( h ρ ( t j , ⋅ ; f ) ) j = 1 k in C loc ([ 0 , ∞ ) ) k , with the spatial interpolation just described.
Proof. The preceding lemma gives sequential convergence of the spatial transition laws whenever deterministic initial profiles converge locally in hypograph under a common linear upper bound. For a random current profile g ε g_{\varepsilon} g ε , put A ε = max { 0 , sup z ≥ 0 g ε ( z ) / ( 1 + z ) } A_{\varepsilon}=\max\{0,\sup_{z\ge0}g_{\varepsilon}(z)/(1+z)\} A ε = max { 0 , sup z ≥ 0 g ε ( z ) / ( 1 + z )} in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] ; its sublevel sets are closed in local hypograph topology. Suppose the current profile and the past tuple converge jointly, the limiting profile is nontrivial almost surely, and the coefficients A ε A_{\varepsilon} A ε have tight tails. Pass to a subsequence on which the profiles, past tuple, and coefficients converge jointly, and take a Skorokhod coupling. The limiting coefficient is finite almost surely, so the coupled coefficients are eventually bounded. Deterministic sequential convergence therefore applies along each coupled sample. Independence of the future slab from the entire past identifies its conditional transition kernel. Bounded convergence, after multiplication by bounded continuous functions of the past tuple, proves joint convergence with the next profile.
Lemma 3.19 gives tight global linear bounds at each positive observation time, as required for the Markov iteration in [23 Lemma 3.10 and Theorem 3.13] . Induction over t 1 , … , t k t_{1},\ldots,t_{k} t 1 , … , t k , using the Markov property and independent Poisson increments, gives the fixed-point transition from each current profile and hence the claimed joint law. At an intermediate time s s s , restart with the integer-valued profile h ( s ε − 3 / 2 , ⋅ ) − ⌊ 2 s ε − 3 / 2 ⌋ h(s\varepsilon^{-3/2},\cdot)-\lfloor2s\varepsilon^{-3/2}\rfloor h ( s ε − 3/2 , ⋅ ) − ⌊ 2 s ε − 3/2 ⌋ . Its rescaling differs from U ε ( s , ⋅ ) U_{\varepsilon}(s,\cdot) U ε ( s , ⋅ ) by a single constant of absolute value at most ε 1 / 2 \varepsilon^{1/2} ε 1/2 . Vertical shift covariance makes the finitely many resulting errors vanish.
Proof of Theorem 1.11. Write U ε = h ε ρ U_{\varepsilon}=h_{\varepsilon}^{\rho} U ε = h ε ρ . Fix a compact rectangle [ a , b ] × [ 0 , R o b s ] [a,b]\times[0,R_{\mathrm{obs}}] [ a , b ] × [ 0 , R obs ] with a > 0 a>0 a > 0 and a rational time grid of mesh at most q q q , extending slightly beyond [ a , b ] [a,b] [ a , b ] . On the event in Lemma 3.18, composition gives, for consecutive grid times t j ≤ t ≤ t j + 1 t_{j}\le t\le t_{j+1} t j ≤ t ≤ t j + 1 ,
U ε ( t j , y ) − C α q α ≤ U ε ( t , y ) ≤ U ε ( t j + 1 , y ) + C α q α , 0 ≤ y ≤ R o b s . U_{\varepsilon}(t_{j},y)-C_{\alpha}q^{\alpha}\le U_{\varepsilon}(t,y)\le U_{\varepsilon}(t_{j+1},y)+C_{\alpha}q^{\alpha},\qquad0\le y\le R_{\mathrm{obs}}. U ε ( t j , y ) − C α q α ≤ U ε ( t , y ) ≤ U ε ( t j + 1 , y ) + C α q α , 0 ≤ y ≤ R obs .
For fixed q q q , let V ε q V_{\varepsilon}^{q} V ε q linearly interpolate in time the spatially interpolated grid profiles. Let D ε ( q ) D_{\varepsilon}(q) D ε ( q ) be the largest uniform difference of neighboring grid profiles on [ 0 , R o b s ] [0,R_{\mathrm{obs}}] [ 0 , R obs ] , and E ε ( q ) E_{\varepsilon}(q) E ε ( q ) their largest spatial interpolation error. On the buffer event the brackets give ∥ U ε − V ε q ∥ [ a , b ] × [ 0 , R o b s ] ≤ D ε ( q ) + C α q α + E ε ( q ) \lVert U_{\varepsilon}-V_{\varepsilon}^{q}\rVert_{[a,b]\times[0,R_{\mathrm{obs}}]}\le D_{\varepsilon}(q)+C_{\alpha}q^{\alpha}+E_{\varepsilon}(q) ∥ U ε − V ε q ∥ [ a , b ] × [ 0 , R obs ] ≤ D ε ( q ) + C α q α + E ε ( q ) . Proposition 3.23 gives joint convergence of the grid profiles and E ε ( q ) → 0 E_{\varepsilon}(q)\to0 E ε ( q ) → 0 in probability at fixed q q q . The limiting neighboring differences tend to zero as q ↓ 0 q\downarrow0 q ↓ 0 by the space-time continuity in [17 Theorem 4.11] , and the limiting grid interpolants converge uniformly to the target field. The buffer exception vanishes in the same order, first ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 and then q ↓ 0 q\downarrow0 q ↓ 0 . These approximations give space-time tightness and identify the locally uniform limit. For continuous representatives, use rectangular meshes of diameter at most ε \varepsilon ε on [ ε 1 / 8 , ε − 1 / 8 ] × [ 0 , ε − 1 / 4 ] [\varepsilon^{1/8},\varepsilon^{-1/8}]\times[0,\varepsilon^{-1/4}] [ ε 1/8 , ε − 1/8 ] × [ 0 , ε − 1/4 ] , replace any − ∞ -\infty − ∞ nodal value by zero, interpolate bilinearly, and extend by clamping each coordinate to this rectangle. The interpolants are Borel C l o c C_{\mathrm{loc}} C loc -valued random elements. A bounded recovered source z ε z_{\varepsilon} z ε reaches the whole rectangle for small ε \varepsilon ε , since ε 1 / 2 ( ε − 1 / 4 + z ε ) < ε 1 / 8 \varepsilon^{1/2}(\varepsilon^{-1/4}+z_{\varepsilon})<\varepsilon^{1/8} ε 1/2 ( ε − 1/4 + z ε ) < ε 1/8 eventually. On each fixed compact, the interpolation error is bounded by the raw oscillation modulus on a slightly larger positive-time compact at the mesh diameter, and hence tends to zero in probability. These rectangles exhaust the domain, giving exactly the interpolation convention in the introduction.
Convergence to the half-space directed landscape
The fixed-point limit supplies finite-wedge laws on rational time slabs. Together with superadditivity and independence of disjoint Poisson slabs, these identify the rational passage field through [17 Theorem 5.1 and Corollary 5.2] . Stationary comparison in both spatial endpoints and short vertical passages then give four-variable tightness and locally uniform metric convergence. When ρ = − ∞ \rho=-\infty ρ = − ∞ , fix c ∗ ∈ [ 0 , 1 ) c_{*}\in[0,1) c ∗ ∈ [ 0 , 1 ) in (1.7).
Identification on rational endpoints
Set
D Q : = H ↑ 2 ∩ Q 4 . D_{\mathbb{Q}} := \mathbb{H}_{\uparrow}^{2} \cap\mathbb{Q}^{4}. D Q := H ↑ 2 ∩ Q 4 .
The raw metric L ε ρ L_{\varepsilon}^{\rho} L ε ρ is − ∞ -\infty − ∞ at noncausal endpoint pairs. For the countable product space, use the finite-valued completion
L ε , Q ρ ( x , s ; y , t ) : = { L ε ρ ( x , s ; y , t ) , ε 1 / 2 ∣ x − y ∣ ≤ t − s , 0 , otherwise , ( x , s ; y , t ) ∈ D Q . L_{\varepsilon,\mathbb{Q}}^{\rho}(x,s;y,t) :=
\begin{cases}
L_{\varepsilon}^{\rho}(x,s;y,t), & \varepsilon^{1/2}|x-y| \le t-s,\\
0, & \text{otherwise},
\end{cases}
\qquad(x,s;y,t) \in D_{\mathbb{Q}}. L ε , Q ρ ( x , s ; y , t ) := { L ε ρ ( x , s ; y , t ) , 0 , ε 1/2 ∣ x − y ∣ ≤ t − s , otherwise , ( x , s ; y , t ) ∈ D Q .
Every fixed compact subset of H ↑ 2 \mathbb{H}_{\uparrow}^{2} H ↑ 2 is causally reachable for all sufficiently small ε \varepsilon ε , so this completion does not affect any of the limits below. We retain the raw metric in all exact variational and composition identities.
For a nonempty finite P ⊂ Q ≥ 0 P \subset\mathbb{Q}_{\ge0} P ⊂ Q ≥ 0 and initial heights a : P → Q a:P\to\mathbb{Q} a : P → Q , put
f P , a ( x ) : = { a ( x ) , x ∈ P , − ∞ , x ∉ P . f_{P,a}(x) :=
\begin{cases}
a(x), & x\in P,\\
-\infty, & x\notin P.
\end{cases} f P , a ( x ) := { a ( x ) , − ∞ , x ∈ P , x ∈ / P .
Theorem 1.11, applied to these finite-wedge data, gives, for rational τ > 0 \tau>0 τ > 0 and finite Y ⊂ Q ≥ 0 Y\subset\mathbb{Q}_{\ge0} Y ⊂ Q ≥ 0 ,
( max x ∈ P { a ( x ) + L ε ρ ( x , 0 ; y , τ ) } ) y ∈ Y ⟹ ( h ρ ( τ , y ; f P , a ) ) y ∈ Y . (90) \left(\max_{x\in P}\{a(x)+L_{\varepsilon}^{\rho}(x,0;y,\tau)\}\right)_{y\in Y}
\Longrightarrow
\left(h^{\rho}(\tau,y;f_{P,a})\right)_{y\in Y}. \tag*{(90)} ( x ∈ P max { a ( x ) + L ε ρ ( x , 0 ; y , τ )} ) y ∈ Y ⟹ ( h ρ ( τ , y ; f P , a ) ) y ∈ Y . (90)
Rounding the microscopic initial heights costs at most ε 1 / 2 \varepsilon^{1/2} ε 1/2 . Time translation gives the same conclusion on every rational time slab.
Theorem 3.24 (Identification on a countable dense set). Fix ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R}\cup\{-\infty\} ρ ∈ R ∪ { − ∞ } . Then
L ε , Q ρ ⟹ L ρ ∣ D Q (91) L_{\varepsilon,\mathbb{Q}}^{\rho}\Longrightarrow L^{\rho}\big|_{D_{\mathbb{Q}}} \tag*{(91)} L ε , Q ρ ⟹ L ρ D Q (91) in R D Q \mathbb{R}^{D_{\mathbb{Q}}} R D Q with its countable product topology. In the subcritical case this holds for every fixed c ⋆ ∈ [ 0 , 1 ) c_{\star}\in[0,1) c ⋆ ∈ [ 0 , 1 ) .
Proof. Fix an arbitrary sequence ε n ↓ 0 \varepsilon_n\downarrow0 ε n ↓ 0 and enumerate D Q = { u 1 , u 2 , … } D_{\mathbb{Q}}=\{u_1,u_2,\ldots\} D Q = { u 1 , u 2 , … } . The one-source case of (3.27) gives convergence of each coordinate, hence tightness of its whole sequence of laws. For η > 0 \eta>0 η > 0 , choose R j < ∞ R_j<\infty R j < ∞ with
sup n P ( ∣ L ε n , Q ρ ( u j ) ∣ > R j ) ≤ η 2 − j . \sup_n \mathbb{P}\left(\left|L_{\varepsilon_n,\mathbb{Q}}^{\rho}(u_j)\right|>R_j\right)\le\eta2^{-j}. n sup P ( L ε n , Q ρ ( u j ) > R j ) ≤ η 2 − j .
The product ∏ j [ − R j , R j ] \prod_j[-R_j,R_j] ∏ j [ − R j , R j ] is compact in the product topology and has probability at least 1 − η 1-\eta 1 − η . Thus the completed fields are tight.
Let M ρ M^{\rho} M ρ be any subsequential limit. We verify precisely the three conditions in [17 Definition 4.7] . First, centered superadditivity gives
L ε ρ ( x , s ; z , r ) + L ε ρ ( z , r ; y , t ) ≤ L ε ρ ( x , s ; y , t ) , s < r < t . L_{\varepsilon}^{\rho}(x,s;z,r)+L_{\varepsilon}^{\rho}(z,r;y,t)\le L_{\varepsilon}^{\rho}(x,s;y,t),\qquad s<r<t. L ε ρ ( x , s ; z , r ) + L ε ρ ( z , r ; y , t ) ≤ L ε ρ ( x , s ; y , t ) , s < r < t .
Each rational triple is eventually causal, and the deterministic centerings add. The inequality therefore passes to the limit simultaneously for all rational triples.
Second, the fields on disjoint half-open Poisson time slabs are independent. The deterministic noncausal completion preserves this property. Independence passes to joint limits of finite collections of coordinates; a monotone-class argument then gives independence of the countable cylinder sigma-fields of the limiting slab fields.
Third, finite maxima are continuous functions of finitely many metric entries. The continuous mapping theorem and (3.27) give, for rational s < t s<t s < t ,
( max x ∈ P { a ( x ) + M ρ ( x , s ; y , t ) } ) y ∈ Y = d ( h ρ ( t − s , y ; f P , a ) ) y ∈ Y . \left(\max_{x\in P}\{a(x)+M^{\rho}(x,s;y,t)\}\right)_{y\in Y}
\overset{\mathrm{d}}{=}
\left(h^{\rho}(t-s,y;f_{P,a})\right)_{y\in Y}. ( x ∈ P max { a ( x ) + M ρ ( x , s ; y , t )} ) y ∈ Y = d ( h ρ ( t − s , y ; f P , a ) ) y ∈ Y .
Thus M ρ M^{\rho} M ρ satisfies the required finite-wedge tests and is a half-space pre-landscape. For finite ρ \rho ρ , [17 Theorem 5.1] identifies its law as L ρ ∣ D Q L^{\rho}\big|_{D_{\mathbb{Q}}} L ρ D Q .
For ρ = − ∞ \rho=-\infty ρ = − ∞ , [17 Lemma 4.9] first gives a continuous extension of M − ∞ M^{-\infty} M − ∞ to H ↑ 2 \mathbb{H}_{\uparrow}^{2} H ↑ 2 . Fix ρ ′ ∈ R \rho'\in\mathbb{R} ρ ′ ∈ R . Boundary thinning couples the prelimit fields so that L ε n , Q − ∞ ≤ L ε n , Q ρ ′ L_{\varepsilon_n,\mathbb{Q}}^{-\infty}\le L_{\varepsilon_n,\mathbb{Q}}^{\rho'} L ε n , Q − ∞ ≤ L ε n , Q ρ ′ for all sufficiently large n n n , since c ⋆ < e ρ ′ ε n 1 / 2 c_{\star}<e^{\rho'\varepsilon_n^{1/2}} c ⋆ < e ρ ′ ε n 1/2 . Pass to a further joint subsequence. The second field has limiting law L ρ ′ ∣ D Q L^{\rho'}\big|_{D_{\mathbb{Q}}} L ρ ′ D Q by the finite-parameter case, and the coordinatewise inequality passes to the limit. The unique continuous extensions preserve this inequality by density. Consequently the extension of M − ∞ M^{-\infty} M − ∞ is stochastically dominated by L ρ ′ L^{\rho'} L ρ ′ for each finite ρ ′ \rho' ρ ′ . Separate couplings for different ρ ′ \rho' ρ ′ suffice here. The one-source tests already give the subcritical narrow-wedge marginals at rational endpoint pairs. These are the two hypotheses of [17 Corollary 5.2] , which identifies the law as L − ∞ L^{-\infty} L − ∞ .
The asserted law is therefore the unique subsequential limit in both boundary regimes. □ \square □
Local uniform metric convergence
For b ≥ 2 b\ge2 b ≥ 2 , write
K b = { ( x , s ; y , t ) : 0 ≤ x , y ≤ b , ∣ s ∣ , ∣ t ∣ ≤ b , t − s ≥ b − 1 } . K_b=\{(x,s;y,t):0\le x,y\le b,\ |s|,|t|\le b,\ t-s\ge b^{-1}\}. K b = {( x , s ; y , t ) : 0 ≤ x , y ≤ b , ∣ s ∣ , ∣ t ∣ ≤ b , t − s ≥ b − 1 } .
These compact sets exhaust H ↑ 2 \mathbb{H}_{\uparrow}^{2} H ↑ 2 . For a real-valued field G G G on K b K_b K b , define
ω b ( G , r ) : = sup u , v ∈ K b ∥ u − v ∥ ∞ ≤ r ∣ G ( u ) − G ( v ) ∣ , ∥ u − v ∥ ∞ : = max 1 ≤ j ≤ 4 ∣ u j − v j ∣ . \omega_b(G,r):=\sup_{\substack{u,v\in K_b\\ \lVert u-v\rVert_{\infty}\le r}}|G(u)-G(v)|,
\qquad
\lVert u-v\rVert_{\infty}:=\max_{1\le j\le4}|u_j-v_j|. ω b ( G , r ) := u , v ∈ K b ∥ u − v ∥ ∞ ≤ r sup ∣ G ( u ) − G ( v ) ∣ , ∥ u − v ∥ ∞ := 1 ≤ j ≤ 4 max ∣ u j − v j ∣.
For a field F F F on a spatial square [ 0 , b ] 2 [0,b]^2 [ 0 , b ] 2 , write instead
ω b s p ( F , r ) : = sup x , y ∈ [ 0 , b ] 2 ∥ x − y ∥ ∞ ≤ r ∣ F ( x ) − F ( y ) ∣ . \omega_b^{\mathrm{sp}}(F,r) := \sup_{\substack{x,y\in[0,b]^2\\ \lVert x-y\rVert_\infty\le r}} |F(x)-F(y)|. ω b sp ( F , r ) := x , y ∈ [ 0 , b ] 2 ∥ x − y ∥ ∞ ≤ r sup ∣ F ( x ) − F ( y ) ∣.
Lemma 3.25 (Spatial tightness on a rational time slab). Fix rational s < t s<t s < t , 0 ≤ b < ∞ 0\le b<\infty 0 ≤ b < ∞ , and either boundary regime. For F ε ( x , y ) : = L ε ρ ( x , s ; y , t ) F_\varepsilon(x,y):=L^\rho_\varepsilon(x,s;y,t) F ε ( x , y ) := L ε ρ ( x , s ; y , t ) ,
lim r ↓ 0 lim sup ε ↓ 0 P ( ω b s p ( F ε , r ) > η ) = 0 , η > 0. \lim_{r\downarrow0}\limsup_{\varepsilon\downarrow0}\mathbb{P}\left(\omega_b^{\mathrm{sp}}(F_\varepsilon,r)>\eta\right)=0,\qquad\eta>0. r ↓ 0 lim ε ↓ 0 lim sup P ( ω b sp ( F ε , r ) > η ) = 0 , η > 0. Moreover, for every fixed finite collection of rational slabs, their spatial fields converge jointly, locally uniformly in ( x , y ) (x,y) ( x , y ) , to the corresponding restrictions of the same L ρ L^\rho L ρ , with continuous spatial interpolation.
Proof. Choose rational R c m p > b R_{\mathrm{cmp}}>b R cmp > b . Apply Lemma 3.16 in both spatial coordinates, using time reversal for the initial coordinate. The two stationary comparison profiles S ε , λ ( 1 ) S^{(1)}_{\varepsilon,\lambda} S ε , λ ( 1 ) and S ε , λ ( 2 ) S^{(2)}_{\varepsilon,\lambda} S ε , λ ( 2 ) need not be independent. Let Ω ε , λ ( r ) \Omega_{\varepsilon,\lambda}(r) Ω ε , λ ( r ) be the larger of their spatial moduli on [ 0 , R c m p ] [0,R_{\mathrm{cmp}}] [ 0 , R cmp ] . For fixed λ \lambda λ it vanishes in probability as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 and then r ↓ 0 r\downarrow0 r ↓ 0 . The union of the two comparison exceptions has probability tending to zero as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 and then λ → ∞ \lambda\to\infty λ → ∞ .
Take the finite mesh r N 0 2 ∩ [ 0 , R c m p ] 2 r\mathbb{N}_0^2\cap[0,R_{\mathrm{cmp}}]^2 r N 0 2 ∩ [ 0 , R cmp ] 2 , with positive rational spacing r r r small enough that the corners of every mesh cell meeting [ 0 , b ] 2 [0,b]^2 [ 0 , b ] 2 lie in [ 0 , R c m p ] 2 [0,R_{\mathrm{cmp}}]^2 [ 0 , R cmp ] 2 . On the comparison event, the two one-sided increment bounds give
F ε ( x + , y + ) − 2 Ω ε , λ ( r ) ≤ F ε ( x , y ) ≤ F ε ( x − , y − ) + 2 Ω ε , λ ( r ) , F_\varepsilon(x^+,y^+)-2\Omega_{\varepsilon,\lambda}(r)\le F_\varepsilon(x,y)\le F_\varepsilon(x^-,y^-)+2\Omega_{\varepsilon,\lambda}(r), F ε ( x + , y + ) − 2 Ω ε , λ ( r ) ≤ F ε ( x , y ) ≤ F ε ( x − , y − ) + 2 Ω ε , λ ( r ) , where x − ≤ x ≤ x + x^-\le x\le x^+ x − ≤ x ≤ x + and y − ≤ y ≤ y + y^-\le y\le y^+ y − ≤ y ≤ y + are the neighboring mesh coordinates. Consequently
ω b s p ( F ε , r ) ≤ max p , p ′ ∈ r N 0 2 ∩ [ 0 , R c m p ] 2 ∥ p − p ′ ∥ ∞ ≤ 3 r ∣ F ε ( p ) − F ε ( p ′ ) ∣ + 4 Ω ε , λ ( r ) . \omega_b^{\mathrm{sp}}(F_\varepsilon,r)\le
\max_{\substack{p,p'\in r\mathbb{N}_0^2\cap[0,R_{\mathrm{cmp}}]^2\\ \lVert p-p'\rVert_\infty\le3r}}
|F_\varepsilon(p)-F_\varepsilon(p')|+4\Omega_{\varepsilon,\lambda}(r). ω b sp ( F ε , r ) ≤ p , p ′ ∈ r N 0 2 ∩ [ 0 , R cmp ] 2 ∥ p − p ′ ∥ ∞ ≤ 3 r max ∣ F ε ( p ) − F ε ( p ′ ) ∣ + 4 Ω ε , λ ( r ) . For fixed r r r , Theorem 3.24 gives joint convergence of the finite mesh vector. Its limiting mesh oscillation tends to zero as r ↓ 0 r\downarrow0 r ↓ 0 by continuity of L ρ L^\rho L ρ . Taking ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , then r ↓ 0 r\downarrow0 r ↓ 0 at fixed λ \lambda λ , and finally λ → ∞ \lambda\to\infty λ → ∞ proves the modulus assertion. Use L ρ , q ε L^\varepsilon_{\rho,q} L ρ , q ε at rational mesh vertices for the spatial interpolants; the completion is inactive on every fixed spatial compact for small ε \varepsilon ε . One-point tightness and this modulus give tightness of bilinear spatial interpolations and a vanishing uniform interpolation error.
For finitely many rational slabs, marginal tightness gives product tightness. Every product subsequential limit has the joint rational spatial evaluations prescribed by Theorem 3.24. Evaluations on a countable dense set determine the joint law of continuous fields. This proves joint convergence to the restrictions of the same landscape. □
Lemma 3.26 (Local uniform tightness of the directed metric). For every ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R}\cup\{-\infty\} ρ ∈ R ∪ { − ∞ } , b ≥ 2 b\ge2 b ≥ 2 , and η > 0 \eta>0 η > 0 ,
lim r ↓ 0 lim sup ε ↓ 0 P ( ω b ( L ε ρ , r ) > η ) = 0. \lim_{r\downarrow0}\limsup_{\varepsilon\downarrow0}\mathbb{P}\left(\omega_b(L^\rho_\varepsilon,r)>\eta\right)=0. r ↓ 0 lim ε ↓ 0 lim sup P ( ω b ( L ε ρ , r ) > η ) = 0. There are Borel continuous mesh interpolations L ~ ε ρ \widetilde L^\rho_\varepsilon L ε ρ that are tight in C l o c ( H 2 2 ) C_{\mathrm{loc}}(\mathbb{H}_2^2) C loc ( H 2 2 ) and satisfy
∥ L ~ ε ρ − L ε ρ ∥ K b ⟶ 0 in probability . \lVert\widetilde L^\rho_\varepsilon-L^\rho_\varepsilon\rVert_{K_b}\longrightarrow0
\qquad\text{in probability}. ∥ L ε ρ − L ε ρ ∥ K b ⟶ 0 in probability . Proof. Fix b ≥ 2 b\ge2 b ≥ 2 and a positive rational mesh q < ( 8 b ) − 1 q<(8b)^{-1} q < ( 8 b ) − 1 . Let T b , q \mathcal{T}_{b,q} T b , q be the finite set of slabs ( a , d ) (a,d) ( a , d ) with a , d ∈ q Z ∩ [ − b − 1 , b + 1 ] a,d\in q\mathbb{Z}\cap[-b-1,b+1] a , d ∈ q Z ∩ [ − b − 1 , b + 1 ] and d − a ≥ ( 2 b ) − 1 d-a\ge(2b)^{-1} d − a ≥ ( 2 b ) − 1 . For u = ( x , s ; y , t ) u=(x,s;y,t) u = ( x , s ; y , t ) and v = ( x ′ , s ′ ; y ′ , t ′ ) v=(x',s';y',t') v = ( x ′ , s ′ ; y ′ , t ′ ) in K b K_b K b with ∥ u − v ∥ ∞ ≤ q \lVert u-v\rVert_\infty\le q ∥ u − v ∥ ∞ ≤ q , choose common grid times
a − = q ⌊ min ( s , s ′ ) / q ⌋ , a + = q ⌈ max ( s , s ′ ) / q ⌉ , a^- = q\lfloor\min(s,s')/q\rfloor,\qquad a^+ = q\lceil\max(s,s')/q\rceil, a − = q ⌊ min ( s , s ′ ) / q ⌋ , a + = q ⌈ max ( s , s ′ ) / q ⌉ , d − = q ⌊ min ( t , t ′ ) / q ⌋ , d + = q ⌈ max ( t , t ′ ) / q ⌉ . d^- = q\lfloor\min(t,t')/q\rfloor,\qquad d^+ = q\lceil\max(t,t')/q\rceil. d − = q ⌊ min ( t , t ′ ) / q ⌋ , d + = q ⌈ max ( t , t ′ ) / q ⌉ . The added initial and terminal intervals have length at most 2 q 2q 2 q . Also d − − a + ≥ b − 1 − 4 q ≥ ( 2 b ) − 1 d^- - a^+\ge b^{-1}-4q\ge(2b)^{-1} d − − a + ≥ b − 1 − 4 q ≥ ( 2 b ) − 1 , so both ( a + , d − ) (a^+,d^-) ( a + , d − ) and ( a − , d + ) (a^-,d^+) ( a − , d + ) lie in T b , q \mathcal{T}_{b,q} T b , q . Both resulting long-slab metric arguments lie in K 2 b + 2 K_{2b+2} K 2 b + 2 .
Put e q = 2 C α ( 2 q ) α e_q=2C_\alpha(2q)^\alpha e q = 2 C α ( 2 q ) α , with α ∈ ( 0 , 1 / 3 ) \alpha\in(0,1/3) α ∈ ( 0 , 1/3 ) as in Lemma 3.18. On E ε t i m e ( b + 1 , 2 q ) \mathcal{E}^{\mathrm{time}}_\varepsilon(b+1,2q) E ε time ( b + 1 , 2 q ) , centered superadditivity gives
L ε ρ ( x , a + ; y , d − ) − e q ≤ L ε ρ ( u ) ≤ L ε ρ ( x , a − ; y , d + ) + e q , L^\rho_\varepsilon(x,a^+;y,d^-)-e_q\le L^\rho_\varepsilon(u)\le L^\rho_\varepsilon(x,a^-;y,d^+)+e_q, L ε ρ ( x , a + ; y , d − ) − e q ≤ L ε ρ ( u ) ≤ L ε ρ ( x , a − ; y , d + ) + e q , and the same bounds hold for v v v . Zero-duration vertical passages are omitted. These exact comparisons use the raw metric.
Define the finite-slab oscillation
A ε , b ( q ) : = max ( a , d ) , ( a ′ , d ′ ) ∈ T b , q ∣ a − a ′ ∣ ∨ ∣ d − d ′ ∣ ≤ 2 q sup x , x ′ , y , y ′ ∈ [ 0 , b ] ∣ x − x ′ ∣ ∨ ∣ y − y ′ ∣ ≤ q ∣ L ε ρ ( x , a ; y , d ) − L ε ρ ( x ′ , a ′ ; y ′ , d ′ ) ∣ . A_{\varepsilon,b}(q):=
\max_{\substack{(a,d),(a',d')\in\mathcal{T}_{b,q}\\ |a-a'|\vee|d-d'|\le2q}}
\ \sup_{\substack{x,x',y,y'\in[0,b]\\ |x-x'|\vee|y-y'|\le q}}
\left|L^\rho_\varepsilon(x,a;y,d)-L^\rho_\varepsilon(x',a';y',d')\right|. A ε , b ( q ) := ( a , d ) , ( a ′ , d ′ ) ∈ T b , q ∣ a − a ′ ∣ ∨ ∣ d − d ′ ∣ ≤ 2 q max x , x ′ , y , y ′ ∈ [ 0 , b ] ∣ x − x ′ ∣ ∨ ∣ y − y ′ ∣ ≤ q sup ∣ L ε ρ ( x , a ; y , d ) − L ε ρ ( x ′ , a ′ ; y ′ , d ′ ) ∣ . The preceding brackets imply
ω b ( L ε ρ , q ) ≤ A ε , b ( q ) + 2 e q \omega_b(L^\rho_\varepsilon,q)\le A_{\varepsilon,b}(q)+2e_q ω b ( L ε ρ , q ) ≤ A ε , b ( q ) + 2 e q on the time-buffer event. For fixed q q q , Lemma 3.25 gives joint uniform convergence on the finitely many slabs in T b , q \mathcal{T}_{b,q} T b , q . Thus A ε , b ( q ) A_{\varepsilon,b}(q) A ε , b ( q ) converges in law to the same functional of L ρ L^\rho L ρ . Its limit is bounded by ω 2 b + 2 ( L ρ , 2 q ) \omega_{2b+2}(L^\rho,2q) ω 2 b + 2 ( L ρ , 2 q ) , which tends to zero almost surely as q ↓ 0 q\downarrow0 q ↓ 0 . The buffer exceptional probability tends to zero by Lemma 3.18. This proves the asserted modulus estimate.
Choose deterministic rational spacings η ε ↓ 0 \eta_\varepsilon\downarrow0 η ε ↓ 0 on the rectangular grid in ( x , s , y , t ) (x,s,y,t) ( x , s , y , t ) , with x , y ≥ 0 x,y \ge0 x , y ≥ 0 . At grid vertices use L ~ ε ρ \widetilde{L}^{\rho}_\varepsilon L ε ρ when t > s t>s t > s and the endpoints are causal, and zero otherwise; interpolate multilinearly between vertices. The resulting L ~ ε ρ \widetilde{L}^{\rho}_\varepsilon L ε ρ is continuous and is Borel measurable as a C l o c ( H ↑ 2 ) C_{\mathrm{loc}}(\mathbb{H}^2_{\uparrow}) C loc ( H ↑ 2 ) -valued random element, since it is determined by countably many measurable nodal values. For each K b K_b K b , all cells meeting it have causal vertices for small ε \varepsilon ε and lie in K 2 b + 2 K_{2b+2} K 2 b + 2 . The uniform interpolation error is then at most ω 2 b + 2 ( L ε ρ , η ε ) \omega_{2b+2}(L^\rho_\varepsilon,\eta_\varepsilon) ω 2 b + 2 ( L ε ρ , η ε ) and tends to zero in probability. One-point tightness, the modulus estimate, and Arzelà–Ascoli give tightness on every K b K_b K b ; a diagonal argument gives tightness in C l o c ( H ↑ 2 ) C_{\mathrm{loc}}(\mathbb{H}^2_{\uparrow}) C loc ( H ↑ 2 ) . □ \square □
Proof of Theorem 1.13. Lemma 3.26 gives tight continuous interpolations with vanishing local uniform error. Theorem 3.24 identifies the joint rational evaluations of every continuous subsequential limit as those of L ρ L^\rho L ρ . These evaluations determine the Borel law on C l o c ( H ↑ 2 ) C_{\mathrm{loc}}(\mathbb{H}^2_{\uparrow}) C loc ( H ↑ 2 ) . Hence the interpolations converge to L ρ L^\rho L ρ , and their vanishing error proves the asserted convergence of the raw metric. Both arguments apply to every fixed c ⋆ ∈ [ 0 , 1 ) c_\star\in[0,1) c ⋆ ∈ [ 0 , 1 ) when ρ = − ∞ \rho=-\infty ρ = − ∞ . □ \square □
Joint convergence from several initial profiles
Joint tightness retains the coupling between the metric and finitely many height evolutions. Localization and the exact variational formula identify every limiting height as a function of the same landscape.
Corollary 3.27 (Joint convergence from several initial profiles). Fix ρ ∈ R ∪ { − ∞ } \rho\in\mathbb{R} \cup\{-\infty\} ρ ∈ R ∪ { − ∞ } and k ∈ N k \in\mathbb{N} k ∈ N . For 1 ≤ j ≤ k 1 \le j \le k 1 ≤ j ≤ k , take deterministic PNG initial profiles h 0 , ε , j ∈ X a d m h_{0,\varepsilon,j} \in\mathcal{X}_{\mathrm{adm}} h 0 , ε , j ∈ X adm and put
f ε , j ( z ) : = ε 1 / 2 h 0 , ε , j ( ε − 1 z ) . f_{\varepsilon,j}(z) := \varepsilon^{1/2}h_{0,\varepsilon,j}(\varepsilon^{-1}z). f ε , j ( z ) := ε 1/2 h 0 , ε , j ( ε − 1 z ) . Suppose f ε , j → f j ∈ U C R + f_{\varepsilon,j} \to f_j \in\mathrm{UCR}^{+} f ε , j → f j ∈ UCR + locally in hypograph and f ε , j ( z ) ≤ C ( 1 + z ) f_{\varepsilon,j}(z) \le C(1+z) f ε , j ( z ) ≤ C ( 1 + z ) for one finite C C C . For the same PNG directed metric, put
U ε , j ( t , y ) : = sup z ≥ 0 { f ε , j ( z ) + L ε ρ ( z , 0 ; y , t ) } . U_{\varepsilon,j}(t,y) := \sup_{z \ge0}\{f_{\varepsilon,j}(z) + L^\rho_\varepsilon(z,0;y,t)\}. U ε , j ( t , y ) := z ≥ 0 sup { f ε , j ( z ) + L ε ρ ( z , 0 ; y , t )} . Then
( L ε ρ , U ε , 1 , … , U ε , k ) ⟹ ( L ρ , h ρ ( ⋅ , ⋅ ; f 1 ) , … , h ρ ( ⋅ , ⋅ ; f k ) ) (L^\rho_\varepsilon,U_{\varepsilon,1},\ldots,U_{\varepsilon,k}) \Longrightarrow(L^\rho,h^\rho(\mathord{\cdot},\mathord{\cdot};f_1),\ldots,h^\rho(\mathord{\cdot},\mathord{\cdot};f_k)) ( L ε ρ , U ε , 1 , … , U ε , k ) ⟹ ( L ρ , h ρ ( ⋅ , ⋅ ; f 1 ) , … , h ρ ( ⋅ , ⋅ ; f k )) in
C l o c ( H ↑ 2 ) × C l o c ( ( 0 , ∞ ) × [ 0 , ∞ ) ) k . C_{\mathrm{loc}}(\mathbb{H}^2_{\uparrow}) \times C_{\mathrm{loc}}((0,\infty)\times[0,\infty))^k. C loc ( H ↑ 2 ) × C loc (( 0 , ∞ ) × [ 0 , ∞ ) ) k . using continuous interpolations with vanishing local uniform errors. All limiting evolutions use the same landscape through (10) . For ρ = − ∞ \rho=-\infty ρ = − ∞ , the assertion holds for each fixed c ⋆ ∈ [ 0 , 1 ) c_\star\in[0,1) c ⋆ ∈ [ 0 , 1 ) .
Proof. Theorems 1.11 and 1.13 give joint tightness of the continuous metric and height interpolations L ~ ε ρ \widetilde{L}^{\rho}_\varepsilon L ε ρ and U ~ ε , j \widetilde{U}_{\varepsilon,j} U ε , j in this finite product. Consider any joint subsequential limit ( L ρ , V 1 , … , V k ) (L^\rho,V_1,\ldots,V_k) ( L ρ , V 1 , … , V k ) and use Skorokhod representation for these continuous fields. Include their interpolation errors on a countable compact exhaustion in the coupling; these errors converge to zero in probability and, on a further subsequence, almost surely.
Fix j j j , ( t , y ) ∈ Q > 0 × Q ≥ 0 (t,y) \in\mathbb{Q}_{>0} \times\mathbb{Q}_{\ge0} ( t , y ) ∈ Q > 0 × Q ≥ 0 , and an integer R ≥ 1 R \ge1 R ≥ 1 . Uniform metric convergence and the hypograph limsup inequality give
lim sup ε ↓ 0 sup 0 ≤ z ≤ R { f ε , j ( z ) + L ~ ε ρ ( z , 0 ; y , t ) } ≤ sup 0 ≤ z ≤ R { f j ( z ) + L ρ ( z , 0 ; y , t ) } . \limsup_{\varepsilon\downarrow0}\ \sup_{0\le z\le R}\{f_{\varepsilon,j}(z)+\widetilde{L}^{\rho}_\varepsilon(z,0;y,t)\}
\le
\sup_{0\le z\le R}\{f_j(z)+L^\rho(z,0;y,t)\}. ε ↓ 0 lim sup 0 ≤ z ≤ R sup { f ε , j ( z ) + L ε ρ ( z , 0 ; y , t )} ≤ 0 ≤ z ≤ R sup { f j ( z ) + L ρ ( z , 0 ; y , t )} .
Indeed, approximate maximizers have a convergent subsequence in [ 0 , R ] [0,R] [ 0 , R ] . For the reverse bound, every z ∈ [ 0 , R ] z \in[0,R] z ∈ [ 0 , R ] with f j ( z ) > − ∞ f_j(z)>-\infty f j ( z ) > − ∞ has a hypograph recovery sequence z ε → z z_\varepsilon\to z z ε → z , contained in [ 0 , 2 R ] [0,2R] [ 0 , 2 R ] for small ε \varepsilon ε . Hence
lim inf ε ↓ 0 sup 0 ≤ z ≤ 2 R { f ε , j ( z ) + L ~ ε ρ ( z , 0 ; y , t ) } ≥ sup 0 ≤ z ≤ R { f j ( z ) + L ρ ( z , 0 ; y , t ) } . \liminf_{\varepsilon\downarrow0}\ \sup_{0\le z\le2R}\{f_{\varepsilon,j}(z)+\widetilde{L}^{\rho}_\varepsilon(z,0;y,t)\}
\ge
\sup_{0\le z\le R}\{f_j(z)+L^\rho(z,0;y,t)\}. ε ↓ 0 lim inf 0 ≤ z ≤ 2 R sup { f ε , j ( z ) + L ε ρ ( z , 0 ; y , t )} ≥ 0 ≤ z ≤ R sup { f j ( z ) + L ρ ( z , 0 ; y , t )} .
Choose a finite sum e ε ( R ) e_\varepsilon(R) e ε ( R ) of the coupled exhaustion errors that bounds the sum of the height interpolation error at ( t , y ) (t,y) ( t , y ) and the metric interpolation error on { ( z , 0 ; y , t ) : 0 ≤ z ≤ 2 R } \{(z,0;y,t):0\le z\le2R\} {( z , 0 ; y , t ) : 0 ≤ z ≤ 2 R } . This sum tends to zero almost surely. Only the tail where this compact is causal is relevant. The exact prelimit variational formula gives
U ~ ε , j ( t , y ) ≥ sup 0 ≤ z ≤ 2 R { f ε , j ( z ) + L ~ ε ρ ( z , 0 ; y , t ) } − e ε ( R ) . \widetilde{U}_{\varepsilon,j}(t,y)\ge
\sup_{0\le z\le2R}\{f_{\varepsilon,j}(z)+\widetilde{L}^{\rho}_\varepsilon(z,0;y,t)\}-e_\varepsilon(R). U ε , j ( t , y ) ≥ 0 ≤ z ≤ 2 R sup { f ε , j ( z ) + L ε ρ ( z , 0 ; y , t )} − e ε ( R ) .
This inequality concerns only the represented fields and error variables, so it remains valid in the coupling. The second compact-source bound, followed by R → ∞ R\to\infty R → ∞ , implies V j ( t , y ) ≥ h ρ ( t , y ; f j ) V_j(t,y)\ge h^\rho(t,y;f_j) V j ( t , y ) ≥ h ρ ( t , y ; f j ) almost surely.
For the upper bound, the same variational formula gives, on the original probability space,
P ( U ~ ε , j ( t , y ) > sup 0 ≤ z ≤ R { f ε , j ( z ) + L ~ ε ρ ( z , 0 ; y , t ) } + e ε ( R ) ) ≤ P ( localization to [ 0 , R ] fails ) . \mathbb{P}\left(\widetilde{U}_{\varepsilon,j}(t,y)>
\sup_{0\le z\le R}\{f_{\varepsilon,j}(z)+\widetilde{L}^{\rho}_\varepsilon(z,0;y,t)\}+e_\varepsilon(R)\right)
\le
\mathbb{P}(\text{localization to }[0,R]\text{ fails}). P ( U ε , j ( t , y ) > 0 ≤ z ≤ R sup { f ε , j ( z ) + L ε ρ ( z , 0 ; y , t )} + e ε ( R ) ) ≤ P ( localization to [ 0 , R ] fails ) .
The event on the left is measurable in the represented variables, so its probability is unchanged by the coupling. The first compact-source bound implies that { V j ( t , y ) > h ρ ( t , y ; f j ) + η } \{V_j(t,y)>h^\rho(t,y;f_j)+\eta\} { V j ( t , y ) > h ρ ( t , y ; f j ) + η } is contained in the limit inferior of these comparison events. Fatou’s lemma gives
P ( V j ( t , y ) > h ρ ( t , y ; f j ) + η ) ≤ lim sup ε ↓ 0 P ( localization to [ 0 , R ] fails ) . \mathbb{P}(V_j(t,y)>h^\rho(t,y;f_j)+\eta)
\le
\limsup_{\varepsilon\downarrow0}\mathbb{P}(\text{localization to }[0,R]\text{ fails}). P ( V j ( t , y ) > h ρ ( t , y ; f j ) + η ) ≤ ε ↓ 0 lim sup P ( localization to [ 0 , R ] fails ) .
Lemma 3.20 makes the right-hand side tend to zero as R → ∞ R\to\infty R → ∞ . Then let η ↓ 0 \eta\downarrow0 η ↓ 0 . Thus equality holds almost surely at every rational terminal point, simultaneously for the finitely many j j j . Continuity of both limiting height fields extends the equality everywhere. Every joint subsequential limit is consequently the claimed common-landscape coupling, which proves the corollary. □ \square □
References