When Is the Paraxial Model Safe in Long-Range Correlated Media?

Beams traveling through turbulent air, the ocean, or the subsurface are usually simulated with the paraxial wave equation. This is a one-way model that assumes the wave keeps moving forward at small angles to a main axis, which turns an expensive three-dimensional wave problem into cheap step-by-step marching. Practitioners decide whether this shortcut is safe by comparing the wavelength to the medium's correlation length, the distance over which fluctuations stay statistically related. That rule has no meaning for media with long-range correlations, where the covariance decays so slowly that the correlation length is infinite. Yet these are the media where recent theory predicts the most interesting behavior: a randomly wandering arrival time described by fractional Brownian motion, and intensity statistics with heavier tails than Gaussian or Kolmogorov models predict (Paraxial Wave Propagation in Random Media with Long-Range Correlations). That theory assumes the paraxial equation is exact. Nobody has checked when that assumption holds in this regime.

This note proposes replacing the correlation-length rule with a spectral one. Whether the paraxial model holds depends on how much fluctuation power the medium has at two specific scales. The first is the axial scale of half a wavelength, which controls backscattering. The second is the transverse scale that would deflect waves outside the small-angle cone. Both scales are measured relative to the low-wavenumber power, which produces only the harmless large-scale phase drift. Long-range correlations push power toward low wavenumbers. The proposal therefore makes a testable and somewhat counterintuitive prediction: at matched phase or travel-time variance, long-range correlated media should be more paraxial-friendly than short-range mixing media, not less. We would test this against full-wave simulations, and in the same runs we would give the first numerical check of the fractional-Brownian travel-time and heavy-tail predictions.

Core Claim

Take a stationary random index perturbation with variance $\sigma^2$ whose covariance decays like $r^{-\alpha}$. Its power spectrum then behaves like $|\kappa|^{\alpha-d}$ at small wavenumber $\kappa$ in $d$ dimensions, and $\alpha<1$ is the non-integrable, long-range case. To leading order in single scattering, the fraction of power reflected over a propagation distance $L$ at wavenumber $k$ is

$$ \eta_{\text{back}} \approx \frac{k^2}{4}\, L\, \hat{C}_z(2k), $$

where $\hat{C}_z$ is the axial power spectrum of the perturbation. By contrast, the variance of the axial travel time grows like $L^{2-\alpha}$ for $\alpha<1$, which is the fractional-Brownian scaling of the travel-time frame. This growth is driven by the spectrum near zero wavenumber.

The paraxial model is accurate when two ratios are small. The first is $\eta_{\text{back}}$ relative to unity. The second is $\eta_{\text{wide}}$, the scattered power beyond a paraxial angle $\theta_p$, normalized by the power that ends up in phase drift. Correlation length enters neither ratio. Our claim is that these two numbers, computed from the medium spectrum alone, predict the measured error of a one-way paraxial solver against a full Helmholtz reference across all values of $\alpha$, including $\alpha<1$. A length-based criterion either cannot be evaluated in that range or predicts wrongly.

Relation to Prior Work

The motivating result is Paraxial Wave Propagation in Random Media with Long-Range Correlations. It constructs a random travel-time frame in which the leading random phase is removed. It shows that the central travel time follows a fractional Brownian motion on a short scale, while the beam profile follows a Brownian-driven Schrödinger equation on a longer scale. The work is purely asymptotic, scalar and one-way, and it never asks whether the parabolic equation itself is valid. We keep its predictions as targets and test them against a model that allows reflection and wide angles.

Two related results suggest the low-wavenumber story is right. Wave propagation in random waveguides with long-range correlations proves that in a waveguide, long-range correlations produce one synchronized, frequency-independent fractional-Brownian phase across all modes, while mode coupling and forward-backward coupling vanish asymptotically. A general framework for waves in random media with long-range correlations proves for layered one-dimensional media that pulses arrive unblurred, with all the randomness pushed into a travel-time shift. Both results hint that long-range randomness mostly shifts phase rather than scattering energy. Both are limited, however: one by waveguide confinement or layering, and both by $\epsilon\to 0$ limits. Neither has been validated numerically, which the research file notes as an open gap.

The constraining work comes from the forward-backward literature in the white-noise regime. Coupled paraxial wave equations in random media in the white-noise regime couples forward and backward paraxial waves and derives an enhanced-backscattering factor of exactly 2 and beam spreading that scales as $L^{3/2}$. Its derivation needs delta-correlated axial noise and so fails under long-range axial memory. Polarization effects for electromagnetic wave propagation in random media shows that scalar paraxial models break down at wider forward angles through exchange between the two polarization modes (TE and TM), again for rapidly mixing media.

Gu and Bal show that standard central-limit and homogenization arguments fail once dependence is long-range (Fluctuations of Parabolic Equations with Large Random Potentials; An invariance principle for Brownian motion in random scenery). This rules out a shortcut: we cannot plug an effective white-noise strength into the Garnier–Sølna formulas.

The methods come from Coupled Parabolic Equations for Wave Propagation. That paper builds a deterministic, stabilized forward-backward parabolic sweep and validates it against Helmholtz on a domain 20 wavelengths wide by 200 wavelengths deep. We reuse its geometry and solver, but in stochastic long-range media and with ensemble statistics as outputs.

Research Program

We would build a reproducible testbed of synthetic random media generated by spectral synthesis. The covariance family is $(1+r^2/\ell^2)^{-\alpha/2}$ with $\alpha \in \{0.3, 0.6, 0.9\}$ for long-range media and $\{1.5, 3\}$ plus a Gaussian covariance for mixing media. Each medium has an explicit outer-scale cutoff, so that finite-domain effects are controlled and reported rather than hidden. Index variance is swept so that each $\alpha$ family can be matched across families on three quantities: local variance $\sigma^2$, travel-time variance at range $L$, and scintillation index at range $L$.

The main configuration is two-dimensional, with one transverse dimension, so that full Helmholtz references remain affordable over hundreds of wavelengths. It is solved with perfectly matched absorbing layers. A smaller three-dimensional set checks the trends, since the fractional-Brownian theory is stated with two transverse dimensions.

For every realization we run four solvers:

  • the full Helmholtz reference;
  • a standard narrow-angle split-step paraxial solver;
  • a wide-angle rational-approximation parabolic solver, a standard tool for which no paper in the file reports results;
  • the coupled forward-backward parabolic solver of Huang et al.

From each run we compute three groups of quantities:

  • Error decomposition. Phase-only error, measured after the arrival-time shift has been aligned. Angular error, measured as the angular spectrum outside $\theta_p$. Reflected power.
  • Ensemble statistics. Travel-time variance against $L$, fitted to $L^{2-\alpha}$. Mean-intensity tails. Scintillation index. The backscattering-cone enhancement factor.
  • Spectral indicators. $\eta_{\text{back}}$ and $\eta_{\text{wide}}$ as defined above.

The analysis fits the measured error of the paraxial solver as a function of the two spectral indicators and asks whether one curve collapses all $\alpha$ families together. A secondary analytic deliverable is a closed-form evaluation of $\eta_{\text{back}}$ and $\eta_{\text{wide}}$ for the power-law family, with inner and outer scales, so the criterion can be computed from a measured spectrum.

Baselines

The naive reference point is the standard narrow-angle split-step paraxial solver, judged safe by the usual correlation-length criterion. For long-range media we apply that criterion using the outer scale as the effective length, which is the most charitable version. It runs on identical realizations and grids. The proposal beats it if the spectral indicators predict which runs exceed a fixed error tolerance (5% relative $L^2$ field error against Helmholtz) with clearly higher accuracy than the length rule. We would measure this as classification accuracy over the full sweep, with a balanced split of passing and failing runs.

The strongest existing method for capturing what paraxial models miss is the coupled forward-backward parabolic solver of Coupled Parabolic Equations for Wave Propagation. That paper reports 15 iterations to converge, a runtime "approximately 15 times slower than a one-way PE," and results "orders of magnitude faster" than full Helmholtz on its 20×200-wavelength benchmark. It reports no field-error number in random long-range media; no reported result was found on Lacuna for that solver's accuracy under long-range correlations. We run it with the same convergence tolerance and wall-clock accounting as the other solvers. The question is whether our criterion correctly flags the cases where its 15-fold cost buys accuracy, and the cases where the plain one-way solver already matches it.

For the statistical predictions, the reference is the asymptotic theory of Paraxial Wave Propagation in Random Media with Long-Range Correlations. It gives scaling laws rather than numbers, so success means the measured travel-time exponent matches $2-\alpha$ within the confidence interval obtained by bootstrapping over realizations.

As a calibration of achievable numerical error in a related setting, Computational multiscale method for parabolic wave approximations in heterogeneous media reports a relative $L^2$ error of approximately $10^{-3}$ on a three-layer benchmark. Our discretization error must sit well below the paraxial modeling error we attribute to physics.

The key ablation isolates which ingredient of the criterion matters. We drop $\eta_{\text{back}}$, then $\eta_{\text{wide}}$, then replace both with total variance $\sigma^2$, and report how much predictive accuracy each removal costs.

Why not just…?

Why not just use the outer scale as the correlation length? The outer scale sets the travel-time variance but says almost nothing about power at the Bragg wavenumber $2k$, which controls reflection. Two media with the same outer scale but different $\alpha$ therefore differ in backscatter at fixed variance. The naive baseline above tests exactly this. If the outer-scale rule classifies runs as well as the spectral criterion, the core claim fails, and we would report that.

Why not just always run the forward-backward or wide-angle solver? This is a real alternative for anyone who can afford it. It costs about 15 times a one-way solve according to Huang et al., and that cost multiplies across the thousands of realizations needed for scintillation and tail statistics. It also does not tell the user why or when the cheaper model is enough. If the experiments show the one-way solver fails across most of the long-range regime, the practical recommendation becomes "always use coupled solvers," and the paper reports that.

Why not just trust the asymptotic theory? The theory is proved only as $\epsilon\to 0$ and assumes the paraxial equation from the start. The waveguide and layered-media results that support it are also unvalidated numerically. Intermediate ranges, where correlation scales are comparable to propagation distance, are exactly where the field notes that analysis gives no answer.

Why not just fit an effective white-noise model? Gu and Bal show that the central-limit structure behind white-noise moment equations breaks down under long-range dependence. As a concrete check, we fit the best effective white-noise model to each long-range ensemble and show whether it reproduces the measured $L^{2-\alpha}$ travel-time growth and the intensity tails. It should not be able to, since white noise gives linear growth.

Decisive Checks

Three results would decide the proposal.

The first is the collapse test. The paraxial-solver error, plotted against the two spectral indicators, should fall on a single curve for all $\alpha$. Plotted against variance or outer scale, it should not.

The second is the matched-phase comparison. At equal travel-time variance, reflected power and wide-angle power should be smaller for $\alpha<1$ than for Gaussian media. The opposite outcome would refute the central prediction.

The third is the theory check. The fitted travel-time exponent should match $2-\alpha$. The enhanced-backscattering factor should approach 2 in the mixing media but may depart from it in the long-range media, which would be the first evidence on backscattering behavior that no current theory covers.