Independent-Crossing Approximation

• Independent-Crossing Approximation is a reduction principle that treats correlated crossing events as statistically independent, simplifying multi-stage dynamics.
• In stochastic processes, ICA employs Rice’s formula to model successive threshold crossings, yielding exponential laws for first-passage times and excursion intervals.
• In neutrino oscillations, ICA factorizes the three-flavor matter interaction into sequential two-level transitions corresponding to high- and low-density resonances.

Independent-Crossing Approximation (ICA) denotes a family of approximations in which distinct crossing events are treated as effectively independent, so that a correlated multi-stage dynamics is reduced to a Poisson, renewal, or factorized transition model. In the arXiv literature, the term has two principal technical meanings. In level-crossing and excursion theory, ICA approximates successive threshold crossings of a stochastic process by approximately independent events in time, leading to exponential or nonhomogeneous exponential laws for first-passage and excursion times (Malinovskii, 2018, Bengtsson et al., 2024). In three-flavor neutrino oscillations in matter, ICA factorizes the full three-level MSW evolution into two sequential, effectively two-level nonadiabatic transitions, typically associated with the high-density and low-density resonances (Yamamoto, 2010). A separate graph-theoretic phrase, “independent crossings,” is not an approximation at all, but a drawing constraint requiring that crossings occur only between non-adjacent edges and that each pair of independent edges crosses at most once (Pach et al., 2018).

1. Conceptual scope

The shared idea behind ICA is a decoupling ansatz: one replaces a globally coupled crossing process by locally defined crossing mechanisms whose interactions are neglected, averaged out, or assumed weak. In stochastic-process applications, the relevant objects are upcrossings, excursion intervals, and first-passage times. In neutrino oscillations, the relevant objects are nonadiabatic transitions at separate matter resonances. In both settings, the approximation becomes credible only when a separation principle holds: rare-event decorrelation in one case, or resonance separation and weak off-subspace coupling in the other [(Malinovskii, 2018); (Yamamoto, 2010)].

Domain	Object approximated	Independence statement
Level crossing	First-passage or excursion times	Successive crossings or intervals are treated as approximately independent
Neutrino oscillations	Three-flavor MSW evolution	H and L resonances are treated as two separate two-level jumps
Graph drawing terminology	Edge-crossing structure	Not an approximation; crossings are restricted combinatorially

A common misconception is that ICA names a single universal formalism. The literature instead uses the same phrase for structurally similar but mathematically distinct approximations. This suggests that ICA is best understood as a reduction principle whose implementation depends on the crossing object being modeled.

2. Stochastic level crossing and first-passage times

In the level-crossing problem, ICA approximates the sequence of upcrossings of a fixed level $u$ by treating successive crossings as approximately independent in time. When crossings are rare and correlations decay fast, the point process of upcrossings is well approximated by a Poisson process. If $T_u$ denotes the time to the first upcrossing, ICA yields, for a stationary process with constant upcrossing rate $\nu_u^{+}$,

$$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$

For nonstationary settings, the constant hazard is replaced by a time-dependent hazard $h(t)$, giving

$$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$

with $h(t) \approx \nu_u^{+}(t)$ or, for a time-varying mean $m(t)$, $h(t)=\nu_{u(t)}^{+}$ with $u(t)=u-m(t)$ (Malinovskii, 2018).

The central input is the upcrossing intensity, typically obtained from Rice’s formula. For a differentiable stationary process,

$T_u$0

For a zero-mean stationary Gaussian process with variance $T_u$1 and derivative variance $T_u$2,

$T_u$3

where $T_u$4 and $T_u$5; equivalently, with spectral density $T_u$6,

$T_u$7

The approximation is therefore operationally simple: compute $T_u$8 from covariance or spectrum, then insert it into a homogeneous or nonhomogeneous exponential survival law (Malinovskii, 2018).

The technical conditions are stricter than those required by some competing methods. ICA in this setting assumes stationarity or local stationarity, mean-square differentiability and smooth sample paths, mixing or short-range dependence, and regularity sufficient for Rice’s formula, typically $T_u$9 finite. These assumptions support the “smooth crossing” paradigm: finite crossing intensity, weak inter-crossing dependence, and a rare-event regime at high levels. By contrast, nonsmooth processes such as standard Brownian motion lack a finite upcrossing rate in the Rice sense, so ICA does not apply directly (Malinovskii, 2018).

Initial conditions matter. If the process starts strictly below the level, the stationary constant-rate model can be inaccurate near $\nu_u^{+}$0, because the stationary hazard presumes the process already visits neighborhoods of $\nu_u^{+}$1. Practical fixes include a hazard that ramps up over one correlation time or a short entrance delay $\nu_u^{+}$2, after which the standard ICA form is used. Diagnostics are equally explicit: under a Poisson crossing model, $\nu_u^{+}$3; substantial overdispersion indicates clustering and hence a failure of the independence ansatz (Malinovskii, 2018).

3. Excursion intervals and the Slepian-model formulation

A more recent development places ICA on a probabilistic foundation through a Slepian-model-based independent interval approximation for excursion-time distributions of smooth stationary Gaussian processes. Here the objects are not only first-passage times but also the lengths $\nu_u^{+}$4 and $\nu_u^{+}$5 of above-level and below-level excursion intervals, with densities $\nu_u^{+}$6 and $\nu_u^{+}$7, survivals $\nu_u^{+}$8 and $\nu_u^{+}$9, and, for $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$0, a generally asymmetric pair of distributions (Bengtsson et al., 2024).

For a zero-mean stationary Gaussian process $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$1 with autocorrelation $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$2, normalized covariance $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$3, and $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$4, the crossing rate is

$$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$5

The Slepian process at level $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$6, describing the law of the process around an up-crossing of $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$7, is

$$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$8

where $$\mathbb{P}(T_u > t) \approx e^{-\nu_u^{+} t}, \qquad f_{T_u}(t) \approx \nu_u^{+} e^{-\nu_u^{+} t}.$$9 is a standard Rayleigh variable and $h(t)$0 is a non-stationary Gaussian residual process with covariance

$h(t)$1

Clipping this process at level $h(t)$2 produces binary processes $h(t)$3 and $h(t)$4, whose expectations $h(t)$5 and $h(t)$6 are matched to the expected value functions of a non-stationary switch process with independent intervals (Bengtsson et al., 2024).

The core identification is written at the level of Laplace transforms: $h(t)$7 Here $h(t)$8 are the Laplace transforms of the approximating excursion-time densities above and below $h(t)$9. Under the monotonicity conditions $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$0 and $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$1, the approximation yields geometric-divisible representations: $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$2 where $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$3 and $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$4 are geometric with parameters $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$5 and $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$6, and the densities of $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$7 and $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$8 are obtained directly from $$\mathbb{P}(T_u > t) \approx \exp\!\Big(-\int_0^t h(s)\, ds\Big), \qquad f_{T_u}(t) \approx h(t)\,\exp\!\Big(-\int_0^t h(s)\, ds\Big),$$9 (Bengtsson et al., 2024).

At zero level, the Slepian-based formulation reproduces the classical stationary ICA/IIA. Two identities are central. First, the clipped stationary process has covariance

$h(t) \approx \nu_u^{+}(t)$0

Second, the clipped Slepian expectation simplifies to

$h(t) \approx \nu_u^{+}(t)$1

The 2024 analysis shows that the Slepian-based method is equivalent to the ordinary stationary IIA/ICA for $h(t) \approx \nu_u^{+}(t)$2, but not for non-zero excursions. For $h(t) \approx \nu_u^{+}(t)$3, covariance matching alone does not separate the above- and below-level interval laws, whereas the Slepian construction directly produces two distinct transforms $h(t) \approx \nu_u^{+}(t)$4 and $h(t) \approx \nu_u^{+}(t)$5 (Bengtsson et al., 2024).

The paper’s numerical example, the two-dimensional Gaussian diffusion with $h(t) \approx \nu_u^{+}(t)$6, illustrates both the mechanics and the accuracy. Persistence coefficients estimated from $h(t) \approx \nu_u^{+}(t)$7 samples generated by the geometric-sum representation closely match coefficients estimated from $h(t) \approx \nu_u^{+}(t)$8 long simulated trajectories. At $h(t) \approx \nu_u^{+}(t)$9, $m(t)$0 from Slepian-IIA versus $m(t)$1 from simulation, with exact value $m(t)$2. At $m(t)$3, $m(t)$4 versus $m(t)$5, and $m(t)$6 versus $m(t)$7. These data show that the approximation captures both exponential tails and the strong asymmetry between above- and below-level excursions at positive thresholds (Bengtsson et al., 2024).

4. Three-flavor neutrino oscillations in matter

In three-flavor neutrino oscillations through dense matter, ICA is a reduction of the full time-dependent three-level problem to two sequential, effectively two-level nonadiabatic transitions. In dense matter such as supernova envelopes, a neutrino can encounter two Mikheyev–Smirnov–Wolfenstein resonances as the density decreases: a high-density $m(t)$8 resonance, governed primarily by $m(t)$9 and $h(t)=\nu_{u(t)}^{+}$0, and a low-density $h(t)=\nu_{u(t)}^{+}$1 resonance, governed by $h(t)=\nu_{u(t)}^{+}$2 and $h(t)=\nu_{u(t)}^{+}$3 (Yamamoto, 2010).

The flavor-basis Hamiltonian is

$h(t)=\nu_{u(t)}^{+}$4

with matter potential

$h(t)=\nu_{u(t)}^{+}$5

For a given $h(t)=\nu_{u(t)}^{+}$6–$h(t)=\nu_{u(t)}^{+}$7 pair, the two-flavor resonance condition is

$h(t)=\nu_{u(t)}^{+}$8

and the Landau–Zener jump probability for a locally linear profile is

$h(t)=\nu_{u(t)}^{+}$9

where large $u(t)=u-m(t)$0 implies adiabatic evolution and small $u(t)=u-m(t)$1 implies nonadiabatic jumping (Yamamoto, 2010).

ICA posits that the total evolution operator factorizes as

$u(t)=u-m(t)$2

with $u(t)=u-m(t)$3 acting in the $u(t)=u-m(t)$4–$u(t)=u-m(t)$5 subspace and $u(t)=u-m(t)$6 acting in the $u(t)=u-m(t)$7–$u(t)=u-m(t)$8 subspace, after neglecting phases and interference. For normal ordering and neutrinos, if $u(t)=u-m(t)$9 and $T_u$00 are the jump probabilities at the $T_u$01 and $T_u$02 resonances, then after phase averaging the electron-neutrino survival probability is

$T_u$03

Equivalently, ICA composes the jump matrices

$T_u$04

so that the net jump matrix is approximately $T_u$05 (Yamamoto, 2010).

The physical interpretation depends on ordering and on whether one considers neutrinos or antineutrinos. In normal ordering, neutrinos can experience both $T_u$06 and $T_u$07 resonances. In normal ordering for antineutrinos, there is no $T_u$08 resonance in the standard MSW case. In inverted ordering, the $T_u$09 resonance shifts to antineutrinos, and for neutrinos ICA is trivially satisfied when only the $T_u$10 crossing occurs (Yamamoto, 2010).

5. Validity regimes and breakdown mechanisms

In stochastic level-crossing theory, ICA is most accurate for high levels, short correlation time, smooth sample paths, and stationary Gaussian-like processes with rapidly decaying correlation. Its principal failure modes are long-range dependence, clumping of crossings, heavy-tailed inter-crossing times, and nonsmooth sample paths. These are not minor corrections: clustering produces overdispersion relative to a Poisson model, and the absence of a finite Rice upcrossing rate prevents direct application altogether. The Slepian-based formulation refines the validity question by imposing monotonicity of the expected clipped Slepian functions $T_u$11, which is the condition used to guarantee valid Laplace transforms and geometric-divisible excursion laws (Malinovskii, 2018, Bengtsson et al., 2024).

In three-flavor neutrino oscillations, the corresponding validity criteria are geometric separation of the nonadiabatic layers, decoupling of the third state during each crossing, and phase decoherence between crossings. The resonance-separation criterion is

$T_u$12

Yamamoto also introduces contamination parameters

$T_u$13

and ICA requires $T_u$14. The paper’s central empirical point is that geometric overlap of the resonance peaks is not sufficient for breakdown: even with very large $T_u$15, ICA can remain valid if the mass hierarchy is strong and the contamination parameters remain small (Yamamoto, 2010).

The numerical analysis uses

$T_u$16

with an exponential electron-density profile scaled by a factor of $T_u$17 relative to the solar model. Numerical tests range from $T_u$18 MeV to $T_u$19 TeV. For realistic oscillation parameters, roughly $T_u$20 and small $T_u$21, ICA agrees with full three-flavor numerical solutions to within numerical tolerance, about $T_u$22 in survival probability. Breakdown appears only for unrealistic parameter ranges, specifically when the hierarchy weakens to $T_u$23 and $T_u$24 (Yamamoto, 2010).

A second misconception is therefore field-specific: overlap does not by itself invalidate neutrino ICA, just as nonzero threshold asymmetry does not by itself invalidate excursion-time ICA. In both literatures, the decisive issue is not visual overlap of crossings but the size of the residual coupling neglected by the approximation.

6. Related terminology and neighboring approximations

The level-crossing literature uses “Independent-Interval Approximation” (IIA) and “Independent-Crossing Approximation” largely interchangeably to denote renewal-type approximations in which successive excursion intervals are treated as independent. For smooth stationary Gaussian processes at $T_u$25, the Slepian-based method and the classical stationary IIA/ICA coincide. For $T_u$26, they diverge: covariance matching alone is underdetermined because it cannot separate the two Laplace transforms $T_u$27 and $T_u$28, whereas the Slepian-based construction produces both explicitly through expectation matching (Bengtsson et al., 2024).

ICA also occupies a specific place relative to other first-passage approximations. The overview of the level-crossing problem contrasts it with normal, diffusion, corrected diffusion, and inverse Gaussian approximations. The inverse Gaussian method discussed there is derived by a new method, is valid under mild regularity conditions, and carries an explicit uniform error of order $T_u$29 in renewal/risk settings. ICA, by contrast, is preferred for stationary, differentiable Gaussian-like processes at high levels with short-range dependence, where Rice’s upcrossing intensity makes the approximation easy to implement (Malinovskii, 2018).

The phrase “independent crossings” in graph theory should not be conflated with ICA. In branching topological multigraphs, conditions (ii)–(iii) require that adjacent edges do not cross and that independent edges cross at most once. These are structural hypotheses used in a crossing lemma for multigraphs; they are not probabilistic or asymptotic independence assumptions. The distinction matters because the terminology is close while the mathematical content is entirely different (Pach et al., 2018).

Taken together, these usages show that ICA is best viewed as a domain-specific decoupling principle. In stochastic crossing theory it converts correlated threshold dynamics into Poisson or renewal structure; in neutrino oscillation theory it converts a three-level matter problem into sequential Landau–Zener transitions; and in adjacent graph-theoretic terminology the same vocabulary of “independent crossings” designates a combinatorial crossing constraint rather than an approximation.