True Self-Avoiding Walk (TSAW) Overview

• True Self-Avoiding Walk (TSAW) is a non-Markovian process where transition probabilities are adjusted based on accumulated site or edge visits to enforce self-repulsion.
• The model is formulated via lattice occupation times or edge local times, linking discrete dynamics to continuum self-repelling motion and revealing rich geometric phase transitions.
• TSAW exhibits standard diffusive scaling in high dimensions and super-diffusive t^(2/3) behavior in one dimension, underpinning both theoretical insights and algorithmic sampling applications.

True self-avoiding walk (TSAW), also called the myopic or “true” self-avoiding random walk, is a non-Markovian nearest-neighbor walk whose transition mechanism depends on the accumulated occupation profile of the path, so that motion is biased toward less visited regions. In lattice formulations this bias is expressed through site local times or edge traversal counts; in continuum limits it is related to true self-repelling motion; and in algorithmic settings it appears as an irreversible, self-repelling sampling dynamics. The rigorous picture is strongly dimension-dependent: in $d\ge 3$ one has diffusion for broad classes of self-interaction functions and a Brownian finite-dimensional scaling limit in a restricted quartic case, whereas in $d=1$ the process is super-diffusive with $t^{2/3}$ scaling and converges, after rescaling, to the Tóth–Werner true self-repelling motion (Horvath et al., 2010).

1. Definitions and principal formulations

In the lattice model studied by Horváth, Tóth, and Vető, $X(t)\in \mathbb Z^d$ is a continuous-time nearest-neighbor jump process with occupation-time field

$L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$

A smooth rate function $w:\mathbb R\to(0,\infty)$ is fixed, with ellipticity $\inf_{u\in\mathbb R} w(u)>0$, and decomposed into even and odd parts

$$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$

Conditionally on $X(t)=x$, the walker jumps to a nearest neighbor $y$ at rate $d=1$0. In the continuum heuristic, the drift is proportional to the negative gradient of the occupation measure, so the walk is repelled from frequently visited sites (Horvath et al., 2010).

The one-dimensional literature also uses a site-local-time formulation in which the jump rates out of $d=1$1 are

$d=1$2

or, in discrete time,

$d=1$3

where $d=1$4 is the visit count at $d=1$5. In the limit $d=1$6, this becomes a strictly self-avoiding walk in which only new sites are entered when possible (Maggs, 2024).

A distinct one-dimensional formulation uses edge local times. Writing $d=1$7 for the number of right- and left-directed traversals of adjacent edges, the transition probabilities are weighted by past edge usage, so that heavily used edges become less attractive. A related edge-repulsion model on infinite locally finite trees assigns weight $d=1$8 to an edge after $d=1$9 traversals, and from a current vertex $t^{2/3}$0 the next step to a neighbor $t^{2/3}$1 is chosen with probability proportional to $t^{2/3}$2 (Kosygina et al., 16 Feb 2025).

2. Environment process and stationary Gibbs structure

A central device in the rigorous analysis is the environment as seen from the walker. Define

$t^{2/3}$3

Then $t^{2/3}$4 is itself a Markov process on a configuration space $t^{2/3}$5, and its generator acts on smooth cylinder functions $t^{2/3}$6 by

$t^{2/3}$7

where $t^{2/3}$8. This converts the original non-Markovian walk into a Markovian environment process, at the price of enlarging the state space (Horvath et al., 2010).

The invariant law is a translation-invariant Gibbs measure $t^{2/3}$9. Writing

$X(t)\in \mathbb Z^d$0

the finite-volume conditional density of $X(t)\in \mathbb Z^d$1 on $X(t)\in \mathbb Z^d$2 given $X(t)\in \mathbb Z^d$3 is proportional to an exponential involving nearest-neighbor differences $X(t)\in \mathbb Z^d$4. In $X(t)\in \mathbb Z^d$5, strict convexity of $X(t)\in \mathbb Z^d$6, expressed in the cited work by $X(t)\in \mathbb Z^d$7, guarantees existence and uniqueness of $X(t)\in \mathbb Z^d$8, and $X(t)\in \mathbb Z^d$9 is stationary and ergodic for the environment process (Horvath et al., 2010).

The special linear case $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$0, for which $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$1, is structurally distinguished. Then $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$2 becomes the massless free Gaussian field on $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$3, with covariance

$L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$4

This Gaussian setting is the one in which the most complete central limit theorem is obtained (Horvath et al., 2010).

3. Diffusion in $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$5

For the non-recurrent dimensions $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$6, the stationary environment immediately yields zero asymptotic speed: for $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$7-almost every initial environment, almost surely

$L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$8

The displacement nevertheless has diffusive order. For every unit vector $L(t,x)=L(0,x)+\int_0^t \mathbf 1_{\{X(s)=x\}}\,ds.$9, there exist constants $w:\mathbb R\to(0,\infty)$0 such that

$w:\mathbb R\to(0,\infty)$1

The proof proceeds by decomposing the coordinate process into orthogonal martingales and compensators, then controlling the compensators through variational estimates, infrared bounds, and Brascamp–Lieb inequalities (Horvath et al., 2010).

A more explicit hypothesis class is given by the conditions

$w:\mathbb R\to(0,\infty)$2

together with an entire-function growth condition on $w:\mathbb R\to(0,\infty)$3. Under these assumptions, the cited work proves finite, positive upper and lower diffusive bounds for every coordinate direction (Horvath et al., 2010).

The full finite-dimensional central limit theorem is proved in a restricted quartic case: $w:\mathbb R\to(0,\infty)$4 with $w:\mathbb R\to(0,\infty)$5 sufficiently small. Then the asymptotic covariance matrix

$w:\mathbb R\to(0,\infty)$6

exists and is non-degenerate, and the rescaled process

$w:\mathbb R\to(0,\infty)$7

has finite-dimensional distributions converging to those of a $w:\mathbb R\to(0,\infty)$8-dimensional Brownian motion with covariance $w:\mathbb R\to(0,\infty)$9 (Horvath et al., 2010).

These results rigorously confirm part of the renormalization-group predictions of Amit, Parisi, and Peliti for $\inf_{u\in\mathbb R} w(u)>0$0: standard diffusive scaling, Gaussian finite-dimensional limits in the quartic subcase, and $\inf_{u\in\mathbb R} w(u)>0$1-linear growth of the mean-square displacement under mild assumptions on the rate function (Horvath et al., 2010).

4. Non-reversible Kipnis–Varadhan theory and the graded-sector method

The rigorous route to the central limit theorem is non-reversible Kipnis–Varadhan theory applied to the environment process. In $\inf_{u\in\mathbb R} w(u)>0$2, the generator is split as

$\inf_{u\in\mathbb R} w(u)>0$3

so that $\inf_{u\in\mathbb R} w(u)>0$4, with $\inf_{u\in\mathbb R} w(u)>0$5 self-adjoint and $\inf_{u\in\mathbb R} w(u)>0$6 skew-adjoint. The reversible part is carried by $\inf_{u\in\mathbb R} w(u)>0$7, while the asymmetry of TSAW is encoded by $\inf_{u\in\mathbb R} w(u)>0$8 (Horvath et al., 2010).

For a centered observable $\inf_{u\in\mathbb R} w(u)>0$9, let $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$0. The resolvent form of Kipnis–Varadhan theory requires

$$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$1

When these conditions hold, the additive functional $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$2 admits an efficient martingale approximation, and a central limit theorem follows (Horvath et al., 2010).

The main technical difficulty is verifying the resolvent criteria in a non-reversible, graded state space. The cited analysis decomposes $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$3 according to the degree of the environment observable and introduces a diagonal operator $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$4 that acts approximately as $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$5 on $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$6. The graded-sector condition then requires operator-norm control on grade-preserving and grade-changing components such as

$$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$7

Horváth, Tóth, and Vető slightly weaken the original Sethuraman–Varadhan–Yau graded-sector condition, and in the Gaussian subcase explicit Fock-space operator estimates make the program verifiable (Horvath et al., 2010).

This framework is significant because it isolates the essential analytic structure of TSAW in high dimension: the process is non-reversible, but it is still sufficiently controlled by the reversible part of the environment dynamics to permit a Brownian limit theorem.

5. One-dimensional scaling and the true self-repelling motion

In one dimension, TSAW is super-diffusive rather than diffusive. Rigorous results summarized by Maggs state that as $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$8, the displacement $$s(u)=\tfrac12[w(u)+w(-u)],\qquad r(u)=\tfrac12[w(u)-w(-u)].$$9 and endpoint local time $X(t)=x$0 satisfy

$X(t)=x$1

$X(t)=x$2

where $X(t)=x$3 and $X(t)=x$4 have explicit Airy-type integral representations. In particular,

$X(t)=x$5

and

$X(t)=x$6

with densities $X(t)=x$7 and $X(t)=x$8. In the associated dynamic-exponent notation $X(t)=x$9, one has $y$0. The cited discussion places the model in a super-diffusive universality class with several KPZ-like features, while emphasizing that its scaling functions differ from the GUE Tracy–Widom form (Maggs, 2024).

A complete functional scaling limit is now available. Kosygina and Peterson prove that

$y$1

converge jointly in $y$2, with the Skorokhod $y$3-topology, to the Tóth–Werner true self-repelling motion together with its local time at the current position. Their proof is organized around a joint generalized Ray–Knight theorem for rescaled local-time curves, absorption points, and merge points, followed by tightness and finite-dimensional convergence arguments (Kosygina et al., 16 Feb 2025).

This one-dimensional theory shows that the phrase “diffusive limit” is dimension-sensitive for TSAW. In $y$4, the correct scaling is $y$5, the limit is non-Gaussian, and the local-time field is part of the limiting object rather than a secondary observable.

6. Dimensional dependence, geometric phase structure, and rare-event geometry

The asymptotic regime depends sharply on the underlying geometry.

Regime	Main behavior	Source
$y$6, $y$7	Diffusive bounds; Brownian finite-dimensional limit in a restricted quartic case	(Horvath et al., 2010)
$y$8	Expected $y$9; $d=1$00 does not exist without an infrared cutoff; only partial non-rigorous and anisotropic results are known	(Horvath et al., 2010)
$d=1$01	Super-diffusive $d=1$02 scaling; convergence to true self-repelling motion after rescaling	(Kosygina et al., 16 Feb 2025)
Infinite locally finite trees	Recurrent if $d=1$03, transient if $d=1$04	(Nguyen, 27 Apr 2026)

For trees, the asymptotic criterion is formulated through the branching-ruin number

$d=1$05

where $d=1$06 is the collection of minimal cut-sets. The boundary $d=1$07, equipped with the metric

$d=1$08

has Hausdorff dimension equal to $d=1$09. The cited theorem shows a sharp phase transition: the TSAW on $d=1$10 is recurrent when $d=1$11 and transient when $d=1$12. The proof uses Rubin’s construction, one-dimensional ruin probabilities with asymptotic $d=1$13, and a quasi-independent ruin percolation argument in the sense of Lyons (Nguyen, 27 Apr 2026).

Two-dimensional rare-event geometry reveals an additional distinction between bulk and tail behavior. Schawe and Hartmann study the convex-hull area and perimeter of a TSAW on the square lattice with transition probabilities

$d=1$14

Their simulations confirm bulk scaling with

$d=1$15

but the large-deviation rate function for the convex-hull area behaves as

$d=1$16

which is inconsistent with the naïve prediction $d=1$17 and instead agrees with $d=1$18, the value associated in the paper with the Smart Kinetic SAW exponent $d=1$19. The qualitative explanation offered there is that in the far-right tail, traps and loops are strongly suppressed, so the walk behaves effectively like a loop-free growth process even though the bulk is governed by the TSAW exponent $d=1$20 (Schawe et al., 2018).

7. Algorithmic realizations and sampling applications

TSAW has also become a mechanism for irreversible sampling algorithms. In the one-dimensional event-chain Monte Carlo (ECMC) setting, Maggs and collaborators identify a microscopic dynamics whose large-scale law coincides with the TSAW. The algorithm uses factorized Metropolis updates: one active particle moves continuously to the right at unit speed until it collides with a neighbor, collision times are determined by

$d=1$21

and activity is then passed to the neighbor. No rejections occur, detailed balance is broken, and only global balance is preserved. Adding a factor field

$d=1$22

cancels the mean drift and exposes the TSAW scaling. For harmonic chains at zero net tension, the resulting dynamics reproduces the $d=1$23 and $d=1$24 laws for displacement and endpoint local time, and yields an autocorrelation decay exponent $d=1$25, compared in the cited summary with $d=1$26 for classic ECMC and $d=1$27 for reversible Monte Carlo (Maggs, 2024).

A different algorithmic extension places TSAW on a finite state space $d=1$28 with an irreducible base kernel $d=1$29 and stationary distribution $d=1$30. If

$d=1$31

and

$d=1$32

then the adaptive transition rule is

$d=1$33

For every state $d=1$34 and every edge $d=1$35 with $d=1$36, the empirical counts satisfy

$d=1$37

Consequently, for every bounded $d=1$38,

$d=1$39

The cited paper contrasts this with the usual $d=1$40 error scaling for empirical averages under standard random-walk-based MCMC and attributes the improvement to the self-avoidance mechanism, which suppresses local over-sampling and enforces near-balancing of empirical flow along each row. It also notes the practical cost of storing the full matrix of transition counts and identifies extensions to continuous or very large state spaces as open (Qinghua et al., 28 May 2026).

Across these settings, TSAW functions both as a probabilistic model of long-memory self-repulsion and as a design principle for non-reversible algorithms. The shared structural theme is the same: empirical overuse of sites or edges feeds back into future motion, but the resulting asymptotics range from Brownian diffusion in high-dimensional lattices to non-Gaussian $d=1$41 scaling in one dimension, geometry-controlled phase transitions on trees, and sharpened empirical-integration bounds on finite state spaces.