betaFeedback

WorksPhysics and Astronomy

Itô versus Stratonovich for shot noise with state-dependent amplitude: why the answer is Marcus

Abstract

For Gaussian white noise the Itô and Stratonovich prescriptions differ by O(dt)O(dt) per time step, which accumulates into the familiar spurious drift 12D II′\tfrac12 D\,II'. We show that for renewal shot noise with state-dependent gain, x˙=−C(x)−I(x)ξ[t]\dot x=-C(x)-I(x)ξ[t] with ξ[t]ξ[t] a train of Dirac spikes of finite random amplitude, the two prescriptions differ by a finite amount at every event, and that where II vanishes the difference can become qualitative. The Marcus kick, the jump-process counterpart of the Stratonovich prescription, is the flow map of dx/ds=I(x)dx/ds=I(x): no kick can cross a zero x∗x^\ast of II, for any jump and waiting-time statistics. The Itô kick is only the tangent map of that flow. For I(x)=1+βxI(x)=1+βx it crosses x∗=−1/βx^\ast=-1/β if and only if ξβ>1ξβ>1 and collapses the whole line onto x∗x^\ast at ξβ=1ξβ=1; below threshold it is itself a Marcus kick with a remapped jump law. Bounded jump laws thus have a threshold, while unbounded ones leak across the barrier for any parameters. The discrepancy cannot be absorbed into a renormalised drift beyond lowest order, and only the Marcus kick is generated by a single operator, which makes the master equation of Bianucci, Bologna and Mannella (2026) intrinsically Marcus. Numerically, the standard Heun kick is a second-order truncation of the Marcus map that preserves the barrier only because its order is even; it shortens every jump, biasing second moments in proportion to the squared jump amplitude and the jump kurtosis, and depleting the region near the barrier. We give the exact-flow integrator that replaces it.

Introduction

The Itô–Stratonovich dilemma is among the oldest and most discussed questions in the theory of multiplicative noise [1, 2]. For the Langevin equation x˙=−C(x)−I(x)ζ(t)\dot{x}=-C(x)-I(x)\zeta(t) driven by Gaussian white noise with ⟨ζ(t)ζ(t′)⟩=Dδ(t−t′)\langle\zeta(t)\zeta(t')\rangle=D\delta(t-t'), the two prescriptions are related by the spurious drift 12DI(x)I′(x)\frac{1}{2}D I(x)I'(x). The drift is finite, but it builds up from a difference of order dtdt per time step: at each step the two readings differ only infinitesimally, because the increments of the Wiener process are infinitesimal. Which of the two is physically correct is then a question about the vanishing correlation time of the real physical fluctuation, and the standard answer – Stratonovich when the white noise is the idealisation of a coloured one, Itô when the increments are genuinely non-anticipating – is by now textbook material. The state of that discussion was surveyed in this journal in Ref. 3; the present paper may be read as a sequel to it, asking what survives when the idealised white noise is replaced by a train of spikes of finite amplitude, i.e. a renewal train of spikes,

ξ(t)=∑q=0∞ξqδ(t−t0−∑k=0qθk),(1)\xi(t) = \sum_{q=0}^{\infty} \xi_q \delta\left(t-t_0-\sum_{k=0}^{q}\theta_k\right), \tag*{(1)}

with waiting times θk\theta_k drawn from ψ(θ)\psi(\theta) and amplitudes ξq\xi_q from p(ξ)p(\xi). Such shot-noise forcing is the standard driving of continuous-time random walk (CTRW) models and appears among others, with state-dependent gain, in synaptic input in neuroscience, in impulsive forcing of slow climate modes, and in intermittent crack growth (see Ref. 4 for more examples and related literature). Here the amplitude ξq\xi_q of a single event is a finite random number, not an infinitesimal increment. The ambiguity in evaluating I(x)I(x) “during” a spike therefore does not produce a small correction: it produces two different, finitely separated, maps of the line onto itself.

The main purpose of this paper is to make that statement precise and to show that its consequences are not merely quantitative. For noise with jumps, the natural counterpart of the Stratonovich prescription is not the Stratonovich integral itself but the canonical equation introduced by Marcus [5, 6], in which each jump is realised by flowing along the vector field of the noise for a virtual time equal to the jump amplitude. We call it the Marcus prescription, and discuss its relation to the Stratonovich integral in § 2.3. Our main result is that the Marcus prescription endows the kick with the structure of a flow – hence with invertibility, a group law, and invariant sets – while the Itô prescription retains only its linearisation and in general loses all three. When II has a zero, the difference becomes a difference in the support of the probability density, which is the most visible thing a stochastic model can have. A second result of this paper is to pinpoint the appropriate numerical scheme to accurately integrate the SDE, under the Marcus prescription in the presence of delta kicks.

The two kick maps

The Marcus (flow-map) kick

Consider the SDE

x˙=−C(x)−I(x) ξ[t](2)\dot{x}=-C(x)-I(x)\,\xi[t] \tag*{(2)}

and regularise the qq-th spike as a smooth pulse of unit area and width ϵ\epsilon, δϵ(t−tq)→δ(t−tq)\delta_\epsilon(t-t_q) \to\delta(t-t_q). During the pulse the drift CC, being bounded, contributes O(ϵ)O(\epsilon) and may be dropped, so that

dxdt=−ξqI(x)δϵ(t−tq).(3)\frac{\mathrm{d}x}{\mathrm{d}t}=-\xi_q I(x)\delta_{\epsilon}(t-t_q). \tag*{(3)}

Introducing the virtual time s(t)=−ξq∫tq−ϵtδϵ(u−tq) dus(t)=-\xi_q\int_{t_q-\epsilon}^{t}\delta_{\epsilon}(u-t_q)\,\mathrm{d}u, which runs from 00 to −ξq-\xi_q irrespective of the pulse shape, the equation becomes autonomous,

dxds=I(x),(4)\frac{\mathrm{d}x}{\mathrm{d}s}=I(x), \tag*{(4)}

and the kick is the time-(−ξq)(-\xi_q) map of its flow:

x(tq+)=ϕ−ξq(x(tq−)),ϕs(x)=Y−1(Y(x)+s),Y(x):=∫xdzI(z)(5)x(t_q^+)=\phi_{-\xi_q}\left(x(t_q^-)\right),\qquad\phi_s(x)=Y^{-1}(Y(x)+s),\qquad Y(x):=\int^x\frac{\mathrm{d}z}{I(z)} \tag*{(5)}

The simplest case when I(x)I(x) has a zero is I(x)=1+βxI(x)=1+\beta x: then the map reads

u↦ue−ξβ,u:=x−x∗,x∗=−1β,(6)u\mapsto u e^{-\xi\beta},\qquad u:=x-x^*,\qquad x^*=-\frac{1}{\beta}, \tag*{(6)}

Proposition 1 (Shape independence). The map (5) is independent of the regularising pulse profile δϵ\delta_\epsilon and of the limit ϵ→0\epsilon\to0, provided δϵ\delta_\epsilon has unit area, CC is bounded on the relevant interval, and the flow of Eq. (4) is complete (for instance, II of at most linear growth).

Proof. Since Eq. (4) is autonomous, x(t)=ϕs(t)(x(tq−ϵ))x(t)=\phi_{s(t)}(x(t_q-\epsilon)) for any path s(t)s(t), monotone or not, and only its end value s=−ξqs=-\xi_q matters. The neglected drift contributes O(ϵsup⁡∣C∣)O(\epsilon\sup|C|).

Proposition 1 is the exact analogue, for spikes of finite area, of the Wong–Zakai theorem [7]. It is what makes the Marcus prescription the natural one within a stochastic Liouville description along single trajectories: the continuity equation holds for each realisation, and the kick is a mass-conserving transport along Eq. (4).

The Itô kick

The Itô (pre-point) prescription evaluates II at the state immediately before the event,

x(tq+)=x(tq−)−ξqI(x(tq−))=:mξq(x(tq−)).(7)x(t_q^+)=x(t_q^-)-\xi_q I\left(x(t_q^-)\right)=:m_{\xi_q}\left(x(t_q^-)\right). \tag*{(7)}

Expanding (5) in powers of the jump amplitude,

ϕ−ξ(x)=x−ξI+ξ22II′−ξ36I(II′′+I′2)+O(ξ4),(8)\phi_{-\xi}(x)=x-\xi I+\frac{\xi^2}{2}II'-\frac{\xi^3}{6}I\left(II''+I'^2\right)+O(\xi^4), \tag*{(8)}

shows that (7) reproduces only the first-order term. The familiar Stratonovich–Itô drift 12II′\frac{1}{2}II' is recognisable as the O(ξ2)O(\xi^2) coefficient; what is new here is that ξ\xi is a finite random variable, so the series is not a small correction but an expansion that may not even converge usefully.

For I(x)=1+βxI(x)=1+\beta x, the Itô map reads

u↦u(1−ξβ)u\mapsto u(1-\xi\beta)

with uu as in Eq. (6): the truncation at first order of the exponential, a sensible approximation only for ∣ξ∣β≪1|\xi|\beta\ll1.

Remark 2 (Post-point prescription). The post-point prescription [8, 9] x+=x−−ξI(x+)x^{+}=x^{-}-\xi I(x^{+}) gives u↦u/(1+ξβ)u \mapsto u/(1+\xi\beta), i.e. the [0/1][0/1] Padé approximant of e−ξβe^{-\xi\beta}. It has a mirrored threshold: the map is undefined at ξβ=−1\xi\beta=-1 and, for ξβ<−1\xi\beta<-1, reverses orientation and carries trajectories across x∗x^{*}.

Reference 4 notes in passing that calculi other than Stratonovich “can be readily studied, both numerically and theoretically, renormalizing C(x)C(x)”, citing the white-noise analysis of Ref. 3. For Gaussian white noise driving this is exact: the whole Itô–Stratonovich difference is the drift 12DII′\frac{1}{2}DII' and nothing else. For spike noise of finite amplitude it holds only to lowest order. Indeed, from (8) the two kick maps differ by

ϕ−ξ(x)−mξ(x)=ξ22II′−ξ36I(II′′+I′2)+O(ξ4),(9)\phi_{-\xi}(x)-m_{\xi}(x)=\frac{\xi^{2}}{2}II'-\frac{\xi^{3}}{6}I\left(II''+I'^{2}\right)+O(\xi^{4}), \tag*{(9)}

and only the first term, averaged over p(ξ)p(\xi) and over the event rate R(t)R(t), is equivalent to the drift renormalisation ΔC=−12R(t)ξ2‾II′\Delta C=-\frac{1}{2}R(t)\overline{\xi^{2}}II'. The remainder is a genuine difference between two bounded jump operators, and a bounded jump operator is not a first-order differential operator: no choice of ΔC\Delta C reproduces it. The obstruction is not merely one of accuracy. Since I(x∗)=0I(x^{*})=0, both kick maps fix x∗x^{*}, and every term of Eq. (9) carries a factor II and vanishes there; so does any drift ΔC\Delta C assembled from them. A renormalised drift therefore leaves the velocity at x∗x^{*} unchanged: whatever transport across x∗x^{*} the original drift CC produces, the renormalised one produces identically, and nothing more. The Itô kick, by contrast, transports probability across x∗x^{*} above the threshold of Proposition 7, on top of whatever CC does, and no choice of ΔC\Delta C can reproduce that. We return to this in § 3.

The maps of Eq. (5) and Eq. (7) differ in an algebraic property that matters for renewal dynamics.

Proposition 3 (Group law). The Marcus kicks form a one-parameter group, ϕs1∘ϕs2=ϕs1+s2\phi_{s_{1}}\circ\phi_{s_{2}}=\phi_{s_{1}+s_{2}}, so that nn events of amplitudes ξ1,…,ξn\xi_{1},\ldots,\xi_{n} occurring at coincident times are equivalent to a single event of amplitude ∑iξi\sum_i \xi_i. The Itô maps do not: mξ1∘mξ2≠mξ1+ξ2m_{\xi_{1}}\circ m_{\xi_{2}}\ne m_{\xi_{1}+\xi_{2}} unless II is constant, the discrepancy being ξ1ξ2II′+O(ξ3)\xi_{1}\xi_{2}II'+O(\xi^{3}).

A practical consequence is that the sub-stepping used in SDE integrators, where several events falling in a single time step are accumulated into one kick of amplitude A=∑iξiA=\sum_i \xi_i, is exact under Marcus and introduces an O(ξ2)O(\xi^{2}) error per coincidence under Itô.

Proposition 3 is not, however, what makes an exact master equation possible. Let L0P=∂x(CP)\mathcal{L}_{0}P=\partial_x(CP) be the drift operator, K=∫p(ξ)Kξ dξ\mathcal{K}=\int p(\xi)\mathcal{K}_{\xi}\,\mathrm{d}\xi the averaged kick operator on densities, with Kξ\mathcal{K}_{\xi} the push-forward under either kick map, and Ψ(t)=∫t∞ψ(t′) dt′\Psi(t)=\int_t^{\infty}\psi(t')\,\mathrm{d}t' the survival function. Conditionally on the event times the density evolves by alternating eL0θe^{\mathcal{L}_{0}\theta} and K\mathcal{K}, and for independent, identically distributed waiting times the renewal series sums in Laplace space,

P^(s)=Φ^Ψ(s)[1−K^ψ^(s)]−1P0,Φ^f(s)=∫0∞f(t)e−steL0t dt(10)\hat{P}(s)=\hat{\Phi}_{\Psi}(s)\left[1-\hat{K}\hat{\psi}(s)\right]^{-1}P_0,\qquad\hat{\Phi}_{f}(s)=\int_{0}^{\infty}f(t)e^{-st}e^{\mathcal{L}_0t}\,\mathrm{d}t \tag*{(10)}

for any kick operator, whether or not the kicks commute and whether or not the kick map is invertible. An exact, non-local evolution equation therefore exists under both prescriptions. What distinguishes Marcus is its form. The Marcus push-forward is Kξ=eξ∂xI(x)K_{\xi}=e^{\xi\partial_x I(x)}, the exponential of a single operator independent of ξ\xi, so that K=p^(−i∂xI(x))\mathcal{K}=\hat{p}(-i\partial_x I(x)): the jump law enters only through its characteristic function, and its moments multiply powers of one and the same operator. This is the structure on which the cumulant expansion and the mean-field local equation of Ref. 4 rest. For a general II the Itô push-forwards MξM_{\xi} admit no such common generator, and the averaged Itô kick is not a function of one operator (the linear case is special; see Remark 8).

The block structure of the nn-time correlation functions of (1), in which a block of mm coincident times contributes ξm‾\overline{\xi^m}, is a separate matter: it arises because the same spike is counted mm times, and it is a property of the noise alone, independent of any kick map. The single-generator form of K\mathcal{K} is what allows it to be carried over to the evolution of PP. In this precise sense the exact theory of Ref. 4 is a Marcus theory.

Relation to the Stratonovich integral

The map (5) is not a notational variant of the Stratonovich integral. For driving processes with jumps the Fisk–Stratonovich integral and the flow-map prescription are different objects, and the one defined by (5) is the Marcus canonical (or simply canonical) equation, introduced by Marcus [5, 6] and developed in generality by Kurtz, Pardoux and Protter [10].

The reason the distinction matters is that the two properties for which physicists prefer Stratonovich in the Brownian case – the ordinary chain rule, and the fact that it arises as the small-correlation-time limit of a smooth approximation – are retained in the jump case by Marcus and not by the Stratonovich integral proper. This comparison is made explicitly by Chechkin and Pavlyukevich [11], who derive the Marcus equation as the limit of Ornstein–Uhlenbeck coloured jump noise of vanishing correlation time. Proposition 1 is the one-dimensional specialisation of that statement to a single spike of finite area, and Proposition 3 is the semigroup property of the Marcus flow. Neither is claimed here as new; they are recalled because the rest of the paper rests on them.

What is specific to the present work is what the flow structure implies once II has a zero – the one-way, repelling barrier of Proposition 6 and the equilibrium law Eq. (15) – together with the sharp threshold, the collapse and the orientation reversal of the Itô map (Proposition 7), the resulting bounded/unbounded classification of Table 1, and the analysis of the Heun kick in § 5.

jump PDFItô crossing conditionper-event crossing probability
dichotomous ξ=±a\xi=\pm aaβ>1a\beta>112Θ(aβ−1)\frac{1}{2}\Theta(a\beta-1)
uniform on [−3a,3a][-\sqrt{3}a,\sqrt{3}a]3aβ>1\sqrt{3}a\beta>112(1−13aβ)+\frac{1}{2}\left(1-\frac{1}{\sqrt{3}a\beta}\right)_{+}
Gaussian, variance a2a^2always12erfc⁡(12aβ)>0\frac{1}{2}\operatorname{erfc}\left(\frac{1}{\sqrt{2}a\beta}\right)>0
two-sided exponential, scale a/2a/\sqrt{2}always12e−2/(aβ)>0\frac{1}{2}e^{-\sqrt{2}/(a\beta)}>0

Table 1. Itô crossing condition and per-event crossing probability for the jump PDF’s considered.

Remark 4 (Relation to the physicists’ usage). Much of the physics literature on delta-correlated noise, including Ref. 4, uses “Stratonovich” to mean the smooth-pulse (Wong–Zakai) limit. Under that reading the two names denote the same object here, and no result of Ref. 4 is affected. We nonetheless use “Marcus” throughout, because the point of this paper is precisely that the tangent-map and flow-map prescriptions are inequivalent, and that distinction is invisible if both are called Stratonovich.

The series and its resummation

For Poisson white noise, Di Paola and Falsone showed that the correction relating the Itô and the Stratonovich prescriptions is not a single term but an infinite series in the jump amplitude [12, 13]. That is the series side of our Eq. (9), and it already contains the statement that no finite drift renormalisation can account for the difference.

Two things are added here. The first is the resummation: the series is the Taylor expansion of the flow map ϕ−ξ\phi_{-\xi}, so the correction has a closed form, and for I=1+βxI = 1 + \beta x that closed form is the elementary factor e−ξβe^{-\xi\beta} of Eq. (6). The second, and the more consequential, concerns the barrier. Whether x∗x^* is preserved depends on the sign of a multiplicative factor, not on the size of a correction, and it is not settled order by order. For I=1+βxI = 1 + \beta x the truncation of e−ze^{-z}, z=ξβz = \xi\beta, at order nn is Tn(z)=∑k=0n(−z)k/k!T_n(z) = \sum_{k=0}^{n}(-z)^k/k!. For odd nn it has exactly one real root, which is positive: z=1z = 1 at first order—the Itô threshold of Proposition 7—z≃1.596z \simeq1.596 at third order, z≃2.181z \simeq2.181 at fifth. For even nn it has none, like the exact factor. (Since Tn′=−Tn−1T'_n = -T_{n-1}, at a critical point of an even-order truncation Tn(z0)=z0n/n!>0T_n(z_0) = z_0^n/n! > 0; an odd-order one is therefore strictly decreasing, and positive for z≤0z \le0.) Truncations of odd order thus all cross the barrier, each at a different threshold, while those of even order preserve it, the second-order one included. Preservation of the barrier alternates with the parity of the truncation instead of converging to the exact behaviour; only the resummed map guarantees it.

Remark 5 (The Itô master equation). Equation (18) is an instance of the generalised Fokker–Planck equation derived, in the Itô interpretation and for multiplicative noise with arbitrary increment distribution, by Denisov, Horsthemke and Hänggi [14]. The integral defined by Kanazawa, Sagawa and Hayakawa [15] as the white-noise limit of coloured non-Gaussian noise is neither Itô nor Fisk–Stratonovich and is closely related to the flow map used here.

Zeros of II: a barrier in one prescription, a threshold in the other

The main result of the paper is most clearly stated against the white-noise case. With Gaussian white noise, a zero x∗x^* of II behaves in the same way under both prescriptions: the paths are continuous, the noise vanishes at x∗x^*, and so does the spurious drift 12DII′\frac{1}{2}DII' that distinguishes the two readings. In both, the velocity at x∗x^* is −C(x∗)-C(x^*), and the prescriptions agree on which side of x∗x^* a trajectory can reach; they differ quantitatively, not in the topology of the dynamics. With spikes of finite amplitude this is no longer so. A finite jump can carry a trajectory across a point that no continuous path can cross, and whether it does depends on the prescription: the flow-map kick never does, the Itô kick does above a sharp threshold. This is the concrete content of the statement that the difference is finite rather than infinitesimal.

Marcus: invariant sectors

Proposition 6 (Barrier to kicks). Let II be locally Lipschitz, with a complete flow, and let I(x∗)=0I(x^*) = 0. Then ∣Y(x)∣→∞|Y(x)| \to\infty as x→x∗x \to x^*, and consequently ϕs\phi_s maps each maximal interval on which II has constant sign into itself, for every finite ss. Under the Marcus prescription no kick can therefore carry a trajectory across x∗x^*, whatever the jump amplitudes, the jump distribution, or the waiting-time statistics.

Proof. Local Lipschitz continuity gives ∣I(z)∣≤L∣z−x∗∣|I(z)| \le L|z-x^*| near x∗x^*, so ∫dz/∣I(z)∣≥∫dz/(L∣z−x∗∣)\int dz/|I(z)| \ge\int dz/(L|z-x^*|) and YY diverges at least logarithmically. Adding a finite ss to YY in (5) cannot leave the image of the interval.

Proposition 6 concerns the kicks. The drift, being a continuous vector field, does cross x∗x^* whenever C(x∗)≠0C(x^*) \ne0, but only in the direction of its velocity there. For the case of interest, C(x)=γxC(x) = \gamma x, that velocity is −C(x∗)=γ/β>0-C(x^*) = \gamma/\beta> 0, directed into the sector u>0u > 0. That sector is therefore forward-invariant under the full Marcus dynamics, whereas the sector u<0u < 0 is not: a trajectory started below x∗x^* is carried across it by the drift and never returns. The barrier is one-way. Since all our initial conditions lie above x∗x^*, the Marcus and Heun densities carry no mass below it at any time, and any mass found there under Itô can only have been placed by a kick.

The Marcus master equation already acknowledges this. For I(x)=1+βxI(x) = 1 + \beta x and dichotomous kicks of amplitude aa, the operators e±a∂xIe^{\pm a\partial_x I} act as dilations in y=1+βxy = 1 + \beta x. For exponential waiting times, ψ(t)=e−t/τ/τ\psi(t) = e^{-t/\tau}/\tau, a drift C(x)=γxC(x) = \gamma x, and with z≡aβz \equiv a\beta, the stationary density P(y)P(y) satisfies

0=∂y[γ(y−1)P(y)]+1τ[12e−zP(e−zy)+12ezP(ezy)−P(y)].(11)0 = \partial_y[\gamma(y - 1)P(y)] + \frac{1}{\tau}\left[\frac{1}{2}e^{-z}P(e^{-z}y) + \frac{1}{2}e^zP(e^z y) - P(y)\right]. \tag*{(11)}

For y→0y \to0, the dominant balance is between the drift and the inflow from larger yy:

γP′(y)≃ez2τP(ezy).(12)\gamma P'(y) \simeq\frac{e^z}{2\tau}P(e^z y). \tag*{(12)}

We set ℓ≡ln⁡(1/y)\ell\equiv\ln(1/y), P=eS(ℓ)P = e^{S(\ell)}, and ν(ℓ)=−S′(ℓ)\nu(\ell) = -S'(\ell), so that ν\nu is the local exponent dln⁡P/dln⁡yd\ln P/d\ln y. Expanding S(ℓ−z)≃S(ℓ)+zν−z22ν′S(\ell-z) \simeq S(\ell) + z\nu- \frac{z^2}{2}\nu' with ν′≃1/z\nu' \simeq1/z, (12) becomes

zν=ℓ+ln⁡ν+ln⁡(2γτ)−z2.(13)z\nu= \ell+ \ln\nu+ \ln(2\gamma\tau) - \frac{z}{2}. \tag*{(13)}

Iterating once, with L≡ln⁡(2γτ)L \equiv\ln(2\gamma\tau) and c≡L−z/2c \equiv L-z/2, we get

ν(ℓ)≃1z[ℓ+c+ln⁡ℓ+Lz],(14)\nu(\ell) \simeq\frac{1}{z}\left[\ell+c+\ln\frac{\ell+L}{z}\right], \tag*{(14)}

and integrating gives

P(y)≃Aexp⁡{−1z[ℓ22+cℓ+(ℓ+L)(ln⁡ℓ+Lz−1)]}.(15)P(y) \simeq A\exp\left\{-\frac{1}{z}\left[\frac{\ell^2}{2}+c\ell+(\ell+L)\left(\ln\frac{\ell+L}{z}-1\right)\right]\right\}. \tag*{(15)}

When y→0+y\to0^{+}, since ν∼ℓ/z→∞\nu\sim\ell/z\to\infty, the density vanishes faster than any power of yy. This is confirmed by simulations, see Figure 1 and discussion in § 5. The parameter z=aβz=a\beta can also be written as z=a/∣x∗∣z=a/|x^{*}|: it measures the typical jump against the distance from the origin to the zero of II.

Figure 1. Probability density P(y)P(y) from Monte Carlo integration of (2) with the Marcus kick, at t=128t=128 (when stationarity is well established), compared with the asymptotic law (15) for z=0.5z=0.5 (left), z=1.0z=1.0 (middle) and z=1.5z=1.5 (right); dichotomous jumps, β=0.5\beta=0.5, γτ=4\gamma\tau=4. The dashed line marks y=e−zy=e^{-z}, below which Eq. (15) applies (see Sec. 5).

Itô: a sharp threshold and a collapse

Take I(x)=1+βxI(x)=1+\beta x, β>0\beta>0, x∗=−1/βx^{*}=-1/\beta, and an initial condition x>x∗x>x^{*}, i.e. u=x−x∗>0u=x-x^{*}>0. The Itô map gives u↦u(1−ξβ)u\mapsto u(1-\xi\beta), so

Proposition 7 (Itô crossing threshold). Under the Itô prescription the point x∗x^{*} is crossed at an event of amplitude ξ\xi if and only if ξβ>1\xi\beta>1. Moreover:

(i) for ξβ<1\xi\beta<1 the map is orientation preserving and the sector u>0u>0 is invariant, as in the Marcus case;

(ii) at ξβ=1\xi\beta=1 the map mξm_{\xi} is not injective: it collapses the entire real line onto the single point x∗x^{*}, so that a single event turns any density into δ(x−x∗)\delta(x-x^{*});

(iii) for ξβ>1\xi\beta>1 the map is orientation reversing and every point of the sector is transported across x∗x^{*}.

Proof. Immediate from u↦u(1−ξβ)u\mapsto u(1-\xi\beta); for (ii) note that mξ(x)=x(1−ξβ)−ξ≡−1/βm_{\xi}(x)=x(1-\xi\beta)-\xi\equiv-1/\beta for all xx when ξβ=1\xi\beta=1.

Item (ii) deserves emphasis: the Itô prescription is not merely a different model, it is singular on a codimension-one set of parameters. Nothing of the kind happens under Marcus, where ϕ−ξ\phi_{-\xi} is a diffeomorphism for every ξ\xi.

Remark 8 (Below threshold, Itô is Marcus with a remapped jump law). For I=1+βxI=1+\beta x the Itô maps u↦(1−ξβ)uu\mapsto(1-\xi\beta)u commute with one another, although they do not satisfy the additive group law of Proposition 3. Whenever ξβ<1\xi\beta<1 one can write

1−ξβ=e−ξ′β,ξ′=−ln⁡(1−ξβ)β,(16)1-\xi\beta=e^{-\xi'\beta},\qquad\xi'=-\frac{\ln(1-\xi\beta)}{\beta}, \tag*{(16)}

so each Itô kick coincides exactly with the Marcus kick of amplitude ξ′\xi'. If p(ξ)p(\xi) is supported in ξβ<1\xi\beta<1, the Itô model is therefore a Marcus model with the jump law p′p' obtained by transporting pp through Eq. (16), and the exact theory of Ref. 4 applies to it with p^\hat p replaced by p^′\hat p'. The remapped law is biased even when pp is symmetric: since −ln⁡(1−w)=w+w2/2+⋯-\ln(1-w)=w+w^2/2+\cdots, one has βξˉ′=ξ2ˉβ2/2+O(β4ξ4ˉ)>0\beta\bar{\xi}'=\bar{\xi^2}\beta^2/2+O(\beta^4\bar{\xi^4})>0. For dichotomous jumps at z=aβ=0.5z=a\beta=0.5 the amplitudes ±a\pm a become +1.386 a+1.386\,a and −0.811 a-0.811\,a. Below threshold the two prescriptions thus differ by a reparametrisation of the noise – a quantitative difference, although not a small one.

The threshold acquires a transparent meaning. As ξβ→1−\xi\beta\to1^{-}, ξ′→+∞\xi'\to+\infty: the Itô kick becomes a Marcus kick of infinite amplitude, which sends every point onto x∗x^{*} – the collapse of Proposition 7(ii). For ξβ>1\xi\beta>1 the factor 1−ξβ1-\xi\beta is negative and no real ξ′\xi' exists, because no flow map reverses orientation. The topological transition is the point at which the remapped jump law ceases to exist.

Bounded versus unbounded jump distributions

Combining Proposition 7 with some specific jump laws gives a clean classification. All distributions of Table 1 have zero average and variance a2a^2, but the expected behaviour is very different. Qualitatively:

  • For bounded jump distributions the two prescriptions (Marcus vs Itô) agree on the topology of the dynamics below a sharp threshold in aβa\beta and disagree above it (see also Figure 2 and Figure 3).

  • For unbounded jump distributions they disagree for every choice of parameters, because the barrier leaks under Itô with finite probability per event, while no Marcus kick can cross it (see Figure 4).

Figure 2. P(x,t)P(x,t) at t=128t = 128 from Monte Carlo integration of (2): purple is Itô, blue is Marcus, green is Heun, for β=0.5\beta= 0.5, dichotomous jumps. Left, aβ=0.5a\beta= 0.5 (below threshold), middle, aβ=1a\beta= 1 (at threshold), right, aβ=1.5a\beta= 1.5 (above threshold). Only the Itô density, and only in the right panel, leaks below x∗=−2x^* = -2 (dashed line). See the text for a description of the insets.

Figure 3. P(x,t)P(x,t) at t=128t = 128 from Monte Carlo integration of (2): purple is Itô, blue is Marcus, green is Heun, for β=0.5\beta= 0.5, uniform jumps. Left, aβ=0.5a\beta= 0.5 (below threshold), middle, aβ=1/3≈0.577a\beta= 1/\sqrt{3} \approx0.577 (at threshold), right, aβ=1a\beta= 1 (above threshold). Only the Itô density, and only in the right panel, leaks below x∗=−2x^* = -2 (dashed line). See the text for a description of the insets.

Figure 4. P(x,t)P(x,t) at t=128t = 128 from Monte Carlo integration of (2): purple is Itô, blue is Marcus, green is Heun, for β=0.5\beta= 0.5, a=1a = 1, and Gaussian jumps (left) or two-sided exponential jumps (right). Unlike the dichotomous case, the Itô density leaks below x∗=−2x^* = -2 for any aa, as expected for unbounded jumps. Insets: mass below x∗x^* as a function of time.

We note that the multiplicative examples of Ref. 4 use β=0.5\beta=0.5 with dichotomous jumps of unit variance, i.e. aβ=0.5a\beta=0.5, which is below the threshold of Proposition 7. The effect discussed here is therefore invisible in those figures, making the present paper particularly relevant, given that we will choose aβa\beta on both sides of unity here.

The two master equations

Under Marcus the kick operator is the push-forward along the flow, whose generator is ∂xI(x)\partial_x I(x), giving the operator of Ref. 4,

∂tP=∂xC(x)P+R(t)[p^(−i∂xI(x))−1]P,p^(−i∂xI)=∫p(ξ)eξ∂xI(x)dξ.(17)\partial_t P=\partial_x C(x)P+R(t)\left[\hat{p}(-i\partial_x I(x))-1\right]P,\qquad\hat{p}(-i\partial_x I)=\int p(\xi)e^{\xi\partial_x I(x)}d\xi. \tag*{(17)}

Under Itô the kick is the push-forward along mξm_\xi, which is not the exponential of any first-order operator, and the master equation reads

∂tP=∂xC(x)P+R(t)[∫p(ξ)Mξdξ−1]P,(MξP)(x)=P(mξ−1(x))∣mξ′(mξ−1(x))∣.(18)\partial_t P=\partial_x C(x)P+R(t)\left[\int p(\xi)M_\xi d\xi-1\right]P,\qquad(M_\xi P)(x)=\frac{P\left(m_\xi^{-1}(x)\right)}{\left|m_\xi'\left(m_\xi^{-1}(x)\right)\right|}. \tag*{(18)}

where the push-forward is understood as a sum over preimages. For I=1+βxI=1+\beta x each point has a single preimage when ξβ≠1\xi\beta\ne1, while at ξβ=1\xi\beta=1 the map is constant and MξP=δ(x−x∗)M_\xi P=\delta(x-x^*) – the analytic counterpart of Proposition 7(ii). With the rate R(t)R(t), (17) is the mean-field local equation of Ref. 4: it is exact only for exponential waiting times, R=1/τR=1/\tau, and (18) is the corresponding Itô form. For any other waiting-time law the exact evolution, under either prescription, is the non-local (10).

Remark 9. Both operators are Markov (positivity- and mass-preserving), so the two equations are equally “legitimate” as evolution equations; the point is that only (17) is generated by a flow – its kick is invertible for every ξ\xi – and is a function of the single operator ∂xI(x)\partial_x I(x), so that the jump law enters only through p^\hat{p}.

Numerical experiments

Numerical simulations were done with β=0.5\beta=0.5 and γ=1\gamma=1, for the four jump laws of Table 1, all with zero mean and variance a2a^2, so that z=aβz=a\beta. Waiting times were drawn from ψ(θ)=(μ−1)T0−1(1+θ/T0)−μ\psi(\theta)=(\mu-1)T_0^{-1}(1+\theta/T_0)^{-\mu} with T0=1T_0=1 and μ=1.5\mu=1.5 for all figures except Figure 1, for which ψ(θ)=e−θ/τ/τ\psi(\theta)=e^{-\theta/\tau}/\tau with γτ=4\gamma\tau=4 was used. Between events the drift was integrated exactly, x↦xe−γθx \mapsto x e^{-\gamma\theta}. Trajectories, of order 1.6×1081.6 \times10^{8} per case, started from a Gaussian of zero mean and standard deviation 0.3 and were followed to t=128t = 128 (to t=256t = 256 for Table 3, whose averages are taken over 128≤t≤256128 \leq t \leq256). The Itô, Heun and exact Marcus kicks were driven by the same random numbers (identical event times and amplitudes), so that their differences are measured with a variance much smaller than that of independent ensembles.

The Heun kick integrator

In Ref. 4 the SDE was integrated with Heun’s scheme, the standard algorithm for Stratonovich SDEs driven by a Wiener process [16, 17]. The Heun scheme for the delta kick integration was enforced as

x~=x−ξI(x),x+=x−ξ2[I(x)+I(x~)],(19)\tilde{x} = x - \xi I(x), \qquad x^{+} = x - \frac{\xi}{2}\left[I(x) + I(\tilde{x})\right], \tag*{(19)}

which however reproduces (8) through O(ξ2)O(\xi^{2}) only and therefore implements Marcus not exactly but only to second order in the jump amplitude.

The exact-flow kick integrator

On the other hand, at a δ\delta kick the correct Marcus update is the exact flow map (5), which for the cases of interest is available in closed form,

I(x)=1  ⇒  ϕ−ξ(x)=x−ξ,I(x)=1+βx  ⇒  ϕ−ξ(x)=−1β+(x+1β)e−ξβ,(20)I(x) = 1 \;\Rightarrow\; \phi_{-\xi}(x) = x - \xi, \qquad I(x) = 1 + \beta x \;\Rightarrow\; \phi_{-\xi}(x) = -\frac{1}{\beta} + \left(x + \frac{1}{\beta}\right)e^{-\xi\beta}, \tag*{(20)}

and otherwise requires one evaluation of YY and one inversion.

For I=1+βxI = 1 + \beta x the Heun step acts as u↦u(1−z+z22)u \mapsto u\left(1 - z + \frac{z^{2}}{2}\right) with z=ξβz = \xi\beta, i.e. the second-order Taylor truncation of the exact factor e−ze^{-z}. This has two consequences:

  • the polynomial 1−z+z2/21 - z + z^{2}/2 is the second-order truncation of e−ze^{-z}, and, like every truncation of even order, has no real root (§ 2.4): the Heun kick preserves the barrier of Proposition 6 because of the parity of the truncation, not because of its accuracy – the third-order kick would cross it at z≃1.596z \simeq1.596;

  • the magnitude is badly wrong once zz is not small, see Table 2. The factor 1−z+z2/21 - z + z^2/2 is smaller than one only for 0<z<20 < z < 2; for z>2z > 2 a kick that should contract towards the barrier pushes the trajectory away from it.

z=ξβz = \xi\betae−ze^{-z}Heunrel. error
0.250.77880.7813+0.3%
0.500.60650.6250+3.0%
1.000.36790.5000+36%
2.000.13531.0000+639%
3.000.04982.5000+4921%

Table 2. Contraction factor of the exact flow kick and of the Heun kick.

This is a concrete, quantified reason to replace the Heun kick with the exact flow map whenever II is non-constant.

Figure 1 compares the equilibrium density obtained with the Marcus kick and dichotomous jumps with the asymptotic law (15). The data show a kink at y=e−zy = e^{-z}, the image of the peak at y=1y = 1 under a single contracting kick; below it the density is built by repeated contracting kicks, which is exactly the regime of Eq. (15). With AA in Eq. (15) fixed, for each zz, by the 60th percentile of the ratio between simulated and theoretical density over the range y<e−zy < e^{-z}, the agreement there is very good, confirming that under the Marcus prescription the density vanishes faster than any power of yy as y→0y \to0.

Figure 2–Figure 4 illustrate Proposition 7. Figure 2 shows P(x,t)P(x,t) at t=128t = 128 for dichotomous jumps. For all values of aa considered, neither the Marcus nor the Heun kick leaks below x∗=−2x^* = -2. The Itô kick does not leak either as long as aβ≤1a\beta\le1, but for aβ>1a\beta> 1 (right panel) it clearly does. The inset in the right panel shows the mass below x∗x^* as a function of time; it decreases because the drift carries trajectories back towards x=0x = 0. The inset in the middle panel shows the mass in the histogram bin containing x∗x^*, i.e. within dr/2=0.005dr/2 = 0.005 of the barrier: these are Itô trajectories just collapsed onto x∗x^* by kicks with ξβ=1\xi\beta= 1 (Proposition 7(ii)) and carried back into u>0u > 0 by the drift. No mass lies strictly below x∗x^*. Figure 3 shows the same behaviour for uniform jumps, except that at threshold the Itô kick brings trajectories very close to x∗x^* (Proposition 7(i)) rather than exactly onto it. Figure 4 shows P(x,t)P(x,t) at t=128t = 128 for Gaussian (left) and two-sided exponential (right) jumps. Although aβ<1a\beta< 1 in both cases, the Itô density leaks markedly below x∗x^*, as expected for unbounded jumps; again, neither the Marcus nor the Heun kick leaks.

Figure 5 compares the measured per-event crossing probabilities with the predictions of Table 1. The agreement is excellent.

Figure 5. Per-event crossing probability versus zz, for the four jump laws. Symbols are the result of numerical simulations, solid lines are the corresponding theoretical predictions (see Table 1).

Figure 2–Figure 4 also show that the Heun density decays towards the barrier faster than the exact Marcus one. The reason is not a one-sided bias of the Heun factor (1−z+z2/2)(1-z+z^2/2) exceeds e−ze^{-z} for z>0z>0 but falls short of it for z<0z<0 but the fact that it differs from e−ze^{-z} by z3/6+O(z4)z^3/6+O(z^4) with the sign of zz. Every Heun jump is therefore shorter than the exact one, by a relative amount z2/6z^2/6, in both directions. Shorter jumps produce a narrower density, and the deficit is largest in the tails, where the mass is built by rare sequences of large kicks. This is particularly evident in Figure 2, where the kinks in the distributions result from a single dichotomous kick applied to trajectories starting at the peak x=0x=0. Consequently, the distance between the kinks and x=0x=0 measures the size of the elementary kick under each scheme. In the left panel (z=0.5z=0.5), kinks are located at x=±1x=\pm1 for the Itô scheme, at x=−0.79/+1.30x=-0.79/+1.30 for Marcus, and at x=−0.75/+1.25x=-0.75/+1.25 for Heun. In the middle panel, the Itô distribution collapses onto x∗x^* and features a kink at x=2x=2, whereas the kinks appear at x=−1.26/+3.44x=-1.26/+3.44 for Marcus and x=−1.0/+3.0x=-1.0/+3.0 for Heun. Finally, in the right panel, the Itô kinks are at x=−3.0x=-3.0 (outside the plot) and x=+3.0x=+3.0, the Marcus kinks are at x=−1.55x=-1.55 and x=6.96x=6.96 (outside the plot), and the Heun kinks are at x=−0.75x=-0.75 and x=3.25x=3.25 (outside the plot).

Accuracy of the Heun kick

The shortening of the jumps has a simple quantitative consequence. Writing ζ=ξβ\zeta=\xi\beta, the squared jump of uu is u2ζ2(1−ζ+7ζ2/12)u^2\zeta^2(1-\zeta+7\zeta^2/12) for the exact kick and u2ζ2(1−ζ+ζ2/4)u^2\zeta^2(1-\zeta+\zeta^2/4) for Heun, so the Heun jump variance is low by u2ζ4‾/3u^2\overline{\zeta^4}/3. Relative to the leading term this is

Δ(δu)2‾(δu)2‾≃−ζ4‾3ζ2‾=−κz23,κ=ξ4‾ξ2‾2,(21)\frac{\overline{\Delta(\delta u)^2}}{\overline{(\delta u)^2}} \simeq-\frac{\overline{\zeta^4}}{3\overline{\zeta^2}}=-\frac{\kappa z^2}{3}, \qquad\kappa=\frac{\overline{\xi^4}}{\overline{\xi^2}^2}, \tag*{(21)}

with κ=1\kappa=1, 1.81.8, 33 and 66 for the dichotomous, uniform, Gaussian and two-sided exponential laws. The error is O(z2)O(z^2) in relative terms, one order lower than the O(z3)O(z^3) error of a single kick in Table 2, and it grows with the kurtosis of the jump law. Table 3 compares Eq. (21) with the measured bias of ⟨x2⟩\langle x^2\rangle, together with the mass in the region x<−1x<-1, half-way to the barrier.

Two features stand out. First, Eq. (21) is accurate up to z≈0.5z\approx0.5 for the dichotomous, uniform and Gaussian laws, so the bias of the moments can be predicted before any simulation is run; for the two-sided exponential law the estimate at z=0.5z=0.5 is not converged, for the reason given in the caption of Table 3. Second, the near-barrier region – the subject of this paper – is where the Heun kick is worst: for Gaussian jumps at z=0.5z=0.5 it underestimates the mass in x<−1x<-1 by a factor of more than four, while the per-kick error of Table 2 is only 3%. The simulations of Ref. 4 used dichotomous jumps at z=0.5z=0.5, where the second-moment bias is about 8%.

Which prescription does an experiment realise?

The answer is set by the comparison of two time scales and one amplitude. Let ϵ\epsilon be the true duration of the physical impulse and TrelT_{\rm rel} the relaxation time of xx under

jump lawzzΔ⟨x2⟩/⟨x2⟩\Delta\langle x^{2}\rangle/\langle x^{2}\rangle−κz2/3-\kappa z^{2/3}MHeun/MexactM_{\mathrm{Heun}}/M_{\mathrm{exact}}
dichotomous0.3−2.99%−3.0%0.750
dichotomous0.5−8.34%−8.3%0.719
dichotomous1.0−44.3%a−33.3%0.214
Gaussian0.3−8.98%−9.0%0.366
Gaussian0.5−24.3%−25.0%0.211
Gaussian1.0−80.9%a−0.081
Uniform0.3−5.32%−5.4%0.667
Uniform0.5−14.6%−15.0%0.377
Uniform1.0−61.7%a−0.100
two-sided exp.0.3−17.1%−18.0%0.172
two-sided exp.0.5−46.1%a−50.0%0.131
two-sided exp.1.0−99.7%b−0.074

caption: Heun kick versus exact Marcus kick, β=0.5\beta= 0.5, μ=1.5\mu= 1.5, same noise realisations; second moments and masses averaged over 128≤t≤256128 \le t \le256. Relative bias of ⟨x2⟩\langle x^{2}\rangle, measured and from Eq. (21) (leading order; “—” where it does not apply), and ratio of the probability masses in x<−1x < -1 (Heun/exact). a Estimate dominated by rare large excursions, because the kick factor has a large or infinite fourth moment: independent runs differ by several percent. b The exact ⟨x2⟩\langle x^{2}\rangle is infinite, since ⟨e2βξ⟩\langle e^{2\beta\xi}\rangle diverges for this law when z≥1/2z \ge1/\sqrt{2}.

:::

the drift. If ϵ≪Trel\epsilon\ll T_{\mathrm{rel}} the spike idealisation is justified, and Proposition 1 then forces the Marcus prescription regardless of how small ϵ\epsilon is: the limit is singular in ϵ\epsilon but not in the prescription. The Itô prescription is appropriate only when the event is genuinely atomic and its effect is specified externally – for instance a counter increment, a discrete packet arrival whose size is fixed before the system state is consulted, or a model in which I(x)I(x) represents a measured gain sampled before the event. In other words:

  • physical impulse of short but nonzero duration, gain varying continuously with the state ⇒\Rightarrow Marcus;

  • discrete, externally specified event whose magnitude is determined by the pre-event state ⇒\Rightarrow Itô.

A concrete instance of the first class is conductance-based synaptic input. A synaptic event of strength aa opens a brief conductance pulse, V˙=g(t)(E−V)\dot{V} = g(t)(E - V) with g(t)=a δ(t−tq)g(t) = a\,\delta(t - t_q), so that the gain I(V)∝E−VI(V) \propto E - V vanishes at the reversal potential EE. The flow-map kick V↦E+(V−E)e−aV \mapsto E + (V - E)e^{-a} is the exact solution during a short conductance pulse [18], and the membrane potential approaches its reversal potential without ever crossing it. The Itô kick V↦V+a(E−V)V \mapsto V + a(E - V) overshoots the reversal potential whenever a>1a > 1, which is physically impossible: here the physics settles the prescription, and it is Marcus.

When the pulse duration and the relaxation time of the system are comparable, neither idealisation applies. For systems with inertia driven by compound Poisson coloured noise, Li, Min and Wang showed that adiabatic elimination yields the Itô or the Marcus form, or an intermediate one, depending on the ratio between the relaxation time and the correlation time of the noise [19].

Where II has a zero, the two are not interchangeable at any order, and the choice must be made on physical grounds before any fitting is done.

Conclusions

In contrast to Gaussian white noise, where the Itô and Stratonovich prescriptions differ by O(dt)O(\mathrm{d}t) per time step and the difference accumulates into a finite spurious drift, state-dependent renewal shot noise with finite random jump amplitudes ξq\xi_q produces a finite, non-negligible discrepancy at every single kick event. This discrepancy cannot be re-absorbed into a renormalized drift operator ΔC(x)\Delta C(x) beyond lowest order in the jump amplitude. Every term of that series vanishes where II does, so a renormalised drift cannot alter the transport across a zero of II, whereas the Itô kick can carry probability across it. Consequently, wherever II vanishes the choice between Itô and Marcus is non-perturbative: for linear II, below a threshold it amounts to a remapping of the jump law, above it to a change in the topology of the dynamics.

Structurally, the Marcus kick represents the exact time-(−ξq-\xi_q) map of the autonomous flow dx/ds=I(x)\mathrm{d}x/\mathrm{d}s = I(x). It acts as a diffeomorphism on the real line that preserves each maximal sign interval of I(x)I(x) and maintains the group property ϕs1∘ϕs2=ϕs1+s2\phi_{s_1} \circ\phi_{s_2} = \phi_{s_1+s_2}. Zeros x∗x^* of I(x)I(x) therefore cannot be crossed by any Marcus kick, whatever the jump amplitudes or waiting-time statistics. Combined with a drift directed into the physical sector, as for C(x)=γxC(x) = \gamma x, they act as one-way barriers; in the stationary state, computed here for exponential waiting times, the density vanishes there faster than any power of 1+βx1+\beta x, Peq(x∗)=0P_{\mathrm{eq}}(x^*) = 0. In contrast, the Itô kick represents only the tangent approximation mξ(x)m_{\xi}(x) of this flow and lacks both invertibility and the group structure. For gain functions such as I(x)=1+βxI(x) = 1+\beta x, the Itô prescription exhibits a sharp threshold at ξβ=1\xi\beta= 1: below threshold it preserves orientation, at threshold it collapses the entire line onto x∗x^*, and above threshold it reverses orientation, allowing trajectories to cross x∗x^*. As a result, for unbounded jump distributions (such as Gaussian or exponential), Itô dynamics leaks across the barrier with finite probability per event across all parameter ranges.

An exact non-local master equation exists under either prescription, but only the Marcus kick is generated by a single operator; the closed-form master equation of Ref. 4, whose generator depends on the jump law only through its characteristic function, is therefore intrinsically a Marcus equation. Finally, from a numerical perspective, standard white-noise integration schemes such as the Heun kick method only match the Marcus flow to second order O(ξ2)O(\xi^2). While Heun accidentally preserves the barrier for I(x)=1+βxI(x) = 1+\beta x, because it is a truncation of even order, it shortens every jump: second moments are biased by a relative kξ2/3k\xi^2/3, and the mass near the barrier is depleted already at moderate zz – by a factor of four for Gaussian jumps at z=0.5z = 0.5. To integrate stochastic differential equations driven by state-dependent shot noise accurately without artificial numerical artifacts, algorithms must employ the exact flow map ϕ−ξ(x)\phi_{-\xi}(x) in place of traditional white-noise approximations.

Acknowledgments

We thank the Green Data Center of University of Pisa for providing the computational power needed for the present paper. During the preparation of this manuscript, the authors used Claude (https://claude.ai) for the purposes of improving the English and readability of some of the textual content. The authors have reviewed and edited the output and take full responsibility for the content of this publication. Data are available upon reasonable request.

References

  1. [1]N. G. van Kampen, Itô versus Stratonovich, J. Stat. Phys. 24, 175 (1981); doi:10.1007/BF01007642.DOI
  2. [2]H. Risken, The Fokker–Planck Equation, 2nd ed. (Springer, 1996); doi:10.1007/978-3-642-61544-3.
  3. [5]S. I. Marcus, Modeling and analysis of stochastic differential equations driven by point processes, IEEE Trans. Inform. Theory 24, 164 (1978); doi:10.1109/TIT.1978.1055857.DOI
  4. [6]S. I. Marcus, Modeling and approximation of stochastic differential equations driven by semimartingales, Stochastics 4, 223 (1981); doi:10.1080/17442508108833165.DOI
  5. [7]E. Wong and M. Zakai, On the convergence of ordinary integrals to stochastic integrals, Ann. Math. Statist. 36, 1560 (1965); doi:10.1214/aoms/1177699916.
  6. [8]Y. L. Klimontovich, Ito, Stratonovich and kinetic forms of stochastic equations, Physica A 163, 515 (1990); doi:10.1016/0378-4371(90)90142-F.DOI
  7. [9]P. Ao, Potential in stochastic differential equations: Novel construction, J. Phys. A: Math. Gen. 37, L25 (2004); doi:10.1088/0305-4470/37/3/L01.DOI
  8. [10]T. G. Kurtz, É. Pardoux and P. Protter, Stratonovich stochastic differential equations driven by general semimartingales, Ann. Inst. Henri Poincaré Probab. Statist. 31, 351 (1995).
  9. [11]A. Chechkin and I. Pavlyukevich, Marcus versus Stratonovich for systems with jump noise, J. Phys. A: Math. Theor. 47, 342001 (2014); doi:10.1088/1751-8113/47/34/342001.DOI
  10. [12]M. Di Paola and G. Falsone, Itô and Stratonovich integrals for delta-correlated processes, Probab. Eng. Mech. 8, 197 (1993); doi:10.1016/0266-8920(93)90015-N.DOI
  11. [13]M. Di Paola and G. Falsone, Stochastic dynamics of nonlinear systems driven by non-normal delta-correlated processes, J. Appl. Mech. 60, 141 (1993); doi:10.1115/1.2900736.DOI
  12. [14]S. I. Denisov, W. Horsthemke and P. Hänggi, Generalized Fokker–Planck equation: derivation and exact solutions, Eur. Phys. J. B 68, 567 (2009); doi:10.1140/epjb/e2009-00126-3.DOI
  13. [15]K. Kanazawa, T. Sagawa and H. Hayakawa, Stochastic energetics for non-Gaussian processes, Phys. Rev. Lett. 108, 210601 (2012); doi:10.1103/PhysRevLett.108.210601.DOI
  14. [16]R. Mannella, Integration of stochastic differential equations on a computer, Int. J. Mod. Phys. C 13, 1177 (2002); doi:10.1142/S0129183102004042.DOI
  15. [17]R. Mannella, Numerical Integration of Stochastic Differential Equations: The Heun Algorithm Revisited and the Itô–Stratonovich Calculus, Entropy 27, 910 (2025); doi:10.3390/e27090910.DOI
  16. [18]M. J. E. Richardson and W. Gerstner, Synaptic shot noise and conductance fluctuations affect the membrane voltage with equal significance, Neural Comput. 17, 923 (2005); doi:10.1162/0899766053429444.DOI
  17. [19]T. Li, B. Min and Z. Wang, Adiabatic elimination for systems with inertia driven by compound Poisson colored noise, Phys. Rev. E 89, 022144 (2014); doi:10.1103/PhysRevE.89.022144.DOI

Paper details

vlnc:W_uj3ub2gz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Claude

View statement
Looking for related papers…
Contents