Mean-Return-Time Phase

• Mean-Return-Time phase is a phase concept for stochastic oscillators defined by constant mean return times to isochrons after a full rotation.
• It is derived using backward Kolmogorov equations with periodic-plus-jump conditions, which clarify how noise and geometry interact in the phase formulation.
• The framework extends to random walks on graphs, where mean return constraints relate local structure to global dynamics via established formulas like Kac's.

Mean-Return-Time phase denotes, in its most precise arXiv usage, a phase notion for stochastic oscillators in which phase lines or isochrons are defined by a return-time criterion: starting from any point on such a line, the mean time to return to the same line after one full rotation is independent of the starting point and equals the oscillator’s mean period. In the broader return-time literature, the same wording is also used more loosely for regimes organized by mean recurrence constraints, such as long-time mean-field tails in random walks, infinite-mean recurrence regimes, or topologically quantized mean detected return times. The term is therefore formal in the stochastic-oscillator literature and analogical in several adjacent settings (Cao et al., 2019, Holzhausen et al., 2021, Hormann et al., 2024).

1. Formal definition for stochastic oscillators

For planar stochastic oscillators, the Mean-Return-Time (MRT) phase was introduced through the criterion of Schwabedal and Pikovsky and later given a PDE-based formulation. A curve in phase space is an MRT isochron if the mean time to return to that same curve after one full oscillation is constant along the curve. The associated phase is then obtained from a mean-return-time function $T$ by a linear rescaling, for example

$$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$

where $\overline{T}$ is the mean rotation period and $\Theta_0$ fixes the phase origin (Cao et al., 2019).

In the isotropic setting, the oscillator is written in polar coordinates as

$$\dot{\rho} = g(\rho) + q_{\rho}(\rho)\,\xi_{\rho}(t), \qquad \dot{\phi} = f(\rho) + q_{\phi}(\rho)\,\xi_{\phi}(t),$$

with independent white Gaussian noises and reflecting boundaries on an annulus $\rho_-<\rho<\rho_+$. Rotational invariance implies that the isochrons can be written as

$\phi = \phi_{\mathrm I}(\rho) + \psi,$

so the geometric problem reduces to determining the radial phase profile $\phi_{\mathrm I}(\rho)$ (Holzhausen et al., 2021).

This formulation makes the MRT phase a probabilistic replacement for deterministic asymptotic phase. Deterministic phase is defined by asymptotic convergence to a periodic orbit; MRT phase is defined instead by equal mean return times after one completed cycle. A recurrent misconception is that these constructions are interchangeable. The oscillator literature treats them as distinct, even when they coincide in special cases such as the isotropic Stuart–Landau oscillator (Cao et al., 2019).

2. Backward-equation formulation and analytic structure

The central mathematical object is the mean-first-passage or mean-return-time function $T$, which solves a backward Kolmogorov equation. In the planar oscillator setting this takes the form

$\mathcal{L}^\dagger T=-1,$

supplemented by reflecting boundary conditions on the radial boundaries and a periodic-plus-jump condition in the angular direction. On a reference strip, the jump condition can be written as

$$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$0

The jump is essential: without unwrapping the angular variable and enforcing a one-rotation shift, the return time to an isochron would be trivially zero because the starting point is already on the set (Cao et al., 2019).

Under smoothness, periodicity, strong ellipticity, existence of a unique stationary density, and nonzero mean flux assumptions, the MRT isochron function exists and is unique up to an additive constant. In this sense the phase is defined canonically modulo the usual phase-offset freedom (Cao et al., 2019).

For isotropic stochastic oscillators, the rotational symmetry allows separation of variables. The solution can be written as

$$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$1

reducing the PDE to an ODE in $$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$2. The paper then derives an explicit quadrature formula for $$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$3. Two structural consequences are emphasized. First, $$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$4 does not enter the isochron shape in this isotropic setting: phase diffusion affects fluctuations but not MRT isochron geometry. Second, for finite annular domains, sufficiently strong radial noise flattens the isochrons into radial spokes, whereas for unbounded domains this need not occur (Holzhausen et al., 2021).

The same analytic machinery also clarifies that the MRT criterion singles out level sets of the first cumulant only. That point becomes important when comparing MRT isochrons with other return-time-based geometric constructions.

3. Relation to stochastic asymptotic phase and geometric phase

A major later development is the comparison between MRT phase and stochastic asymptotic phase. The latter is defined spectrally from the backward Kolmogorov operator: if $$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$5 is the nontrivial eigenvalue with least negative real part and corresponding eigenfunction is written as

$$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$6

then $$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$7 is the stochastic asymptotic phase. MRT phase $$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$8, by contrast, is constructed from the return-time PDE and the level curves of $$\Theta(x,y)=\Theta_0-T(x,y)\frac{2\pi}{\overline{T}},$$9 (Du, 13 Sep 2025).

The quantitative bridge between the two is expressed by

$\overline{T}$0

where

$\overline{T}$1

In the paper’s interpretation, $\overline{T}$2 reflects the different reference frames used by the two phase definitions, while $\overline{T}$3 is the nontrivial geometric correction generated by the diffusion structure. Through the generalized Doob’s $\overline{T}$4-transform with $\overline{T}$5, this correction becomes the stochastic analogue of a Berry-like geometric term (Du, 13 Sep 2025).

This geometric viewpoint also sharpens a second common misconception: constancy of mean return time does not imply constancy of higher return-time cumulants. Using the same formalism, iso-variance curves can be constructed from the variance $\overline{T}$6 of the return time, but these curves generally differ from the MRT isochrons. In the Newby–Schwemmer oscillator, the iso-variance line is significantly less twisted than the MRT isochron; along the former the return-time variance is approximately constant, whereas along the latter the mean return time is nearly constant (Holzhausen et al., 2021).

4. Mean-return-time regimes in random walks on graphs

Outside stochastic oscillators, “Mean-Return-Time phase” is often best understood as a regime organized by the mean return time rather than as a formal phase variable. This is explicit in the sparse-network random-walk literature, where the paper does not introduce a separate MRT phase as a formal term. Instead, it splits the approximation of the first return time distribution into a short-time, local-structure regime and a long-time, mean-field regime. If the phrase is used there, it refers to the asymptotic regime in which the distribution is forced to have the correct mean return time (Hormann et al., 2024).

For a simple random walk on a finite, connected, undirected network, the first return time distribution $\overline{T}$7 is linked to the mean first return time by

$\overline{T}$8

with $\overline{T}$9 the degree of node $\Theta_0$0 and $\Theta_0$1 the total number of edges. This Kac formula is a central anchor: the mean depends only on $\Theta_0$2 and $\Theta_0$3, not on detailed global structure. A related electrical-network proof gives the weighted generalization

$\Theta_0$4

which reduces to $\Theta_0$5 in the unweighted case (Markowsky, 2016).

The sparse-network approximation combines two layers. A message-passing or tree approximation captures short-time return behavior and local geometry. A mean-field geometric tail then corrects the long-time behavior, enforces normalization, and restores the exact finite-network mean return time. In that sense, the asymptotic regime is mean-return-time dominated: local structure controls short returns, while the long-time tail is constrained chiefly by the exact mean $\Theta_0$6 and by the total number of edges (m