🔢 mathematics

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

This paper demonstrates that for shot noise with state-dependent amplitude, the Marcus prescription is the physically correct interpretation because it uniquely preserves the system's invariant barriers and zero-crossings of the gain function, whereas the Itô and Stratonovich (Heun) approaches fail to do so by either crossing these barriers or introducing systematic biases that cannot be corrected by simple drift renormalization.

Original authors: Marco Bianucci, Riccardo Mannella

Published 2026-10-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marco Bianucci, Riccardo Mannella

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of physics, many systems are subject to random jolts. A neuron firing in the brain, a climate system reacting to sudden weather shifts, or a crack growing in a piece of metal are all driven by unpredictable bursts of energy. For decades, scientists have used a standard mathematical tool to describe how these systems evolve when hit by such random noise. This tool relies on a choice: when a system is hit, do we calculate the effect based on the state of the system just before the hit, or do we calculate it as if the system were moving smoothly through the moment of impact? For very small, continuous jitters, this choice matters only slightly, creating a tiny, accumulated error that can be corrected with a simple adjustment. However, when the jolts are large, sudden spikes rather than a gentle hum, this small correction is no longer enough. The choice between these two ways of calculating becomes a decision between two completely different realities, where the system might behave in ways that are impossible under the other method.

This paper investigates what happens when a system is driven by these large, sudden spikes, specifically when the strength of the spike depends on where the system currently is. The researchers, Marco Bianucci and Riccardo Mannella, focused on a specific type of mathematical rule that describes how the system changes. They discovered that for these large spikes, the standard method used by most scientists leads to a fundamental error. This error is not just a small miscalculation; it is a qualitative difference that changes the very nature of the system's behavior. In their study, they showed that one method allows the system to cross a specific boundary that should be impenetrable, while the other method correctly keeps the system on its side of the barrier.

The researchers studied a system where the random spikes have a finite size, meaning they are not infinitesimally small but are distinct, measurable events. They found that when the system approaches a point where the spike's effect naturally drops to zero, the two mathematical approaches diverge sharply. One approach, which they identify as the physically correct one for this type of noise, treats the spike as a smooth flow that the system travels along during the event. This method ensures that the system can never cross a specific barrier where the spike's influence vanishes. It is like a river flowing toward a waterfall; the water can get close to the edge, but the flow itself prevents it from jumping over the lip in a way that defies the laws of the river's path. The other approach, which is more commonly used, treats the spike as an instantaneous jump based on the system's position right before the event. This method fails to respect the barrier. If the spike is large enough, this approach allows the system to teleport across the barrier, landing on the other side, or even collapse the entire system into a single point.

The authors demonstrated that this difference is not a matter of degree but of kind. For certain types of random jumps, the standard method allows the system to leak across the barrier with a finite probability, no matter how small the parameters are set. In contrast, the correct method, which they call the Marcus prescription, guarantees that the barrier remains intact. The system can approach the barrier, but it cannot cross it via a spike. This distinction is crucial because it changes the long-term behavior of the system. Under the correct method, the system's density vanishes rapidly as it approaches the barrier, effectively keeping the system confined to a safe region. Under the incorrect method, the system spreads out and leaks into forbidden regions, creating a false picture of how the system behaves.

To test these ideas, the researchers ran extensive computer simulations using different types of random jumps, including simple two-state switches, uniform distributions, and Gaussian distributions. They compared the results of the standard method against the correct flow-based method. The simulations confirmed their theoretical predictions. When the jump size was small, both methods gave similar results, but as the jump size increased, the standard method began to fail dramatically. In cases where the jump size crossed a specific threshold, the standard method caused the system to collapse or flip across the barrier, while the flow-based method kept the system stable and confined. The researchers also found that a common numerical technique used to simulate these systems, known as the Heun method, was insufficient. While this method accidentally preserved the barrier in some cases due to a mathematical coincidence, it consistently shortened the jumps, leading to significant errors in the system's behavior, particularly in the regions near the barrier.

The paper concludes that for systems driven by large, state-dependent spikes, the standard mathematical tools are inadequate. The correct approach requires treating the spike not as an instantaneous jump but as a continuous flow that the system follows for the duration of the event. This ensures that the system respects the physical constraints imposed by the barrier. The authors emphasize that this is not just a theoretical nuance but a practical necessity for accurately modeling real-world phenomena like synaptic inputs in the brain or impulsive forces in climate models. By using the correct flow-based method, scientists can avoid the artificial leakage and collapse that plague the standard approach, leading to a more accurate understanding of how complex systems respond to sudden, powerful changes. The study provides a clear path forward, offering a precise mathematical recipe for integrating these systems without introducing artificial errors, ensuring that the models reflect the true physical reality of the systems they describe.

Technical Summary: Itô versus Stratonovich for Shot Noise with State-Dependent Amplitude

Problem Statement
The paper addresses the Itô–Stratonovich dilemma in the context of multiplicative noise driven by renewal shot noise (a train of Dirac spikes with finite random amplitudes) rather than Gaussian white noise. While the standard dilemma for Gaussian white noise involves a difference of order $O(dt)$ per time step that accumulates into a "spurious drift" ( $\frac{1}{2}D I I'$ ), the authors demonstrate that for shot noise with state-dependent gain $I(x)$ , the ambiguity is not infinitesimal. Instead, the choice of prescription (Itô vs. Stratonovich/Marcus) results in two distinct, finitely separated maps of the state space. Crucially, where the gain function $I(x)$ vanishes (at a point $x^*$ ), the difference between the prescriptions becomes qualitative, altering the topological support of the probability density.

Methodology
The authors analyze the stochastic differential equation (SDE) $\dot{x} = -C(x) - I(x)\xi[t]$ , where $\xi[t]$ is a renewal train of spikes.

Analytical Derivation: They regularize the Dirac spikes as smooth pulses of width $\epsilon$ $ϵ$ and take the limit $\epsilon \to 0$ $ϵ \to 0$ .
- Marcus Prescription: They derive the "Marcus kick" as the flow map of the autonomous equation $dx/ds = I(x)$ over a virtual time equal to the jump amplitude $-\xi$ . This yields a map $\phi_{-\xi}(x) = Y^{-1}(Y(x) - \xi)$ , where $Y(x) = \int^x dz/I(z)$ .
- Itô Prescription: They define the Itô kick as the pre-point evaluation $x^+ = x^- - \xi I(x^-)$ .
Comparison: The authors compare the algebraic properties of these maps, specifically their group structure, invertibility, and behavior near zeros of $I(x)$ . They expand the Marcus map in powers of the jump amplitude to show it corresponds to an infinite series, whereas the Itô map corresponds only to the first-order truncation.
Master Equations: They formulate the exact non-local master equations for both prescriptions. They highlight that the Marcus operator is generated by a single operator ( $\partial_x I(x)$ ), allowing the jump law to enter only via its characteristic function, whereas the Itô operator lacks this single-generator structure.
Numerical Experiments: The authors perform Monte Carlo simulations for various jump distributions (dichotomous, uniform, Gaussian, two-sided exponential) and compare three integration schemes:
- Exact Marcus flow map.
- Standard Heun scheme (a second-order truncation of the Marcus map).
- Exact Itô map.

Key Contributions and Results

Finite Discrepancy and Topological Differences: Unlike the Gaussian case, the difference between Itô and Marcus prescriptions for shot noise is finite at every event. When $I(x)$ has a zero at $x^*$ , the Marcus prescription acts as a barrier: no kick can cross $x^*$ regardless of jump amplitude or statistics. In contrast, the Itô prescription allows trajectories to cross $x^*$ if the jump amplitude exceeds a sharp threshold ( $\xi \beta > 1$ for linear $I$ ).
The Collapse and Threshold: For $I(x) = 1 + \beta x$ , the Itô map exhibits a singularity at $\xi \beta = 1$ , where it collapses the entire real line onto $x^*$ . Above this threshold, the map reverses orientation, transporting probability across the barrier. The Marcus map remains a diffeomorphism for all parameters.
Bounded vs. Unbounded Distributions:
- For bounded jump distributions, the two prescriptions agree on the dynamics' topology below the threshold but disagree above it.
- For unbounded distributions (e.g., Gaussian), the Itô prescription leaks across the barrier with finite probability for any parameter set, whereas the Marcus prescription strictly forbids crossing.
Master Equation Structure: The authors confirm that the exact master equation derived in their previous work (Bianucci, Bologna, and Mannella, 2026) is intrinsically a Marcus equation because its generator depends on the jump law only through its characteristic function. This structure is lost in the Itô formulation.
Numerical Artifacts of the Heun Scheme: The standard Heun integrator, often used for Stratonovich SDEs, is shown to be a second-order truncation of the Marcus map.
- While it accidentally preserves the barrier for linear $I$ (due to the even parity of the truncation), it systematically shortens jump amplitudes.
- This shortening biases second moments (underestimating variance) by a relative amount proportional to the squared jump amplitude and the kurtosis of the jump law ( $\sim -\kappa z^2/3$ ).
- The Heun scheme significantly depletes the probability mass near the barrier compared to the exact Marcus flow, even at moderate jump amplitudes.

Significance and Claims
The paper claims that for state-dependent shot noise, the choice between Itô and Marcus is not merely a matter of renormalizing the drift or a perturbative correction; it is a non-perturbative, topological distinction.

Physical Interpretation: The authors argue that the Marcus prescription is the physically correct limit for impulses of short but non-zero duration where the gain varies continuously with the state (e.g., conductance-based synaptic input). The Itô prescription is appropriate only for truly atomic events where the magnitude is determined by the pre-event state (e.g., discrete packet arrival).
Numerical Necessity: The paper asserts that standard integration schemes like Heun are insufficient for accurate integration of such SDEs. To avoid artificial depletion of the barrier region and biased moments, numerical algorithms must employ the exact flow map $\phi_{-\xi}(x)$ rather than polynomial truncations.
Theoretical Framework: The work establishes that the "Stratonovich" intuition (smooth limit) for jump processes is correctly captured by the Marcus canonical equation, not the Stratonovich integral itself, and that this distinction is critical for models involving zeros in the multiplicative noise coefficient.

Technical Summary: Itô versus Stratonovich for Shot Noise with State-Dependent Amplitude

More like this