Towards a robust criterion of anomalous diffusion

The study leverages single-particle tracking (SPT) to interrogate complex systems where anomalous diffusion is frequently observed. Anomalous diffusion is typically characterized via the power-law scaling of the mean-squared displacement (MSD) ⟨X²⟩ ~ t^α, with α = 1 indicating Brownian motion, 0 < α < 1 subdiffusion, and α > 1 superdiffusion. Estimating α reliably from finite, noisy trajectories is challenging: static and dynamic localization errors bias MSD-based inference; trajectories often exhibit heterogeneous or time-varying scaling; and when α is close to 1, distinguishing anomalous from normal diffusion is difficult. The research question is to establish a robust, easy-to-implement criterion to decide whether single-trajectory dynamics are anomalous or Brownian, and to determine the type of anomaly, even in the presence of unavoidable localization errors. The authors focus on fractional Brownian motion (FBM) as a representative Gaussian anomalous-diffusion model (with Hurst index H and anomalous exponent α = 2H) and investigate a power spectral, single-trajectory-based criterion resilient to measurement noise.

The paper surveys advances in SPT-enabled biophysics and soft-matter research, highlighting numerous systems where anomalous diffusion was reported (e.g., intracellular transport, membrane proteins, chromatin motion, mRNA, viruses, organelles, and movement ecology). It reviews analytical approaches for anomalous diffusion characterization: mean-maximal excursion statistics; Bayesian maximum-likelihood and nested-sampling schemes; deep learning and feature-based classifiers; and challenges in MSD fitting due to noise and heterogeneity. Localization errors are categorized as static (finite photon count/statistical localization) and dynamic (finite exposure-time motion blur), known to bias MSD and correlation analyses. While classical power spectra assume infinite-time/large-ensemble limits, recent single-trajectory power spectral density (PSD) methods allow inference from individual trajectories. Prior work established limiting values of the coefficient of variation γ of the single-trajectory PSD for pure FBM at high frequency, proposing γ as a diagnostic for diffusion regimes.

Core idea: analyze single-trajectory power spectral density (PSD) S(f, T) = (1/T) ∫₀^T dt₁ ∫₀^T dt₂ cos[f(t₁ − t₂)] X_{t1} X_{t2}. For Gaussian processes, the PSD distribution is characterized by its mean μ(f, T) and variance σ²(f, T). Define the coefficient of variation γ(f, T) = σ(f, T) / μ(f, T). For pure FBM, γ tends to distinct high-frequency limits: γ → 1 (subdiffusion), γ → √5/2 (normal), γ → √2 (superdiffusion). Measurement error modeling: - Static localization error modeled as additive OU process ε_t with relaxation time τ₀ ≪ Δt and variance 2Dτ₀; observed process Y_t = X_t + ε_t. - Dynamic localization error modeled as finite exposure-time averaging: measured position is time-averaged over τ_e, effectively low-pass filtering the trajectory. Analytical treatment: For Y_t = X_t + ε_t (static error only), derive the relation for γ of Y in terms of constituents’ moments and variances: γ_Y = (σ_X² + σ_ε² + 2 μ_X μ_ε) / (μ_X + μ_ε)². Use known high-frequency asymptotics of μ_X and σ_X for FBM and for OU noise to obtain limiting γ corrections as functions of H, D, τ₀, T, and f. Dynamic error is addressed via numerical simulations due to analytical complexity. Simulations: - 1D FBM in Python (fbm package): trajectories with various H (1/3, 1/2, 2/3), Δt = 1, N up to 2¹⁴–2²³; dynamic error implemented by averaging over n_τ = τ_e/Δt substeps; static error added via OU with τ₀ = Δt and varying noise amplitude. - 2D FBM in Matlab (wfbm): M = 100 trajectories, N = 2500, Δt = 0.1 s, step size 0.01 μm; emulate imaging with Gaussian PSF sampling to generate dominant static or dynamic error regimes by varying detected photon counts (n_τ). Experimental datasets: - Subdiffusive: telomeres in U2OS cell nuclei and P-granules in C. elegans embryos; trajectories pre-normalized by RMS step to reduce diffusivity variability. - Brownian: 1.2 μm polystyrene beads in aqueous buffer imaged at 100 fps. - Superdiffusive: cytoskeleton fluctuations via RGD-coated beads attached to A549 cells at multiple temperatures (5 Hz sampling); nanoparticles (100 nm) in hMSCs tracked after cytoplasmic microinjection. For each dataset, compute γ(f, T) versus fT and compare to theoretical limits and noise-induced corrections.

• For pure FBM, the high-frequency limiting values of γ(f, T) are: subdiffusion (H < 1/2): γ → 1; normal diffusion (H = 1/2): γ → √5/2 ≈ 1.1180; superdiffusion (H > 1/2): γ → √2 ≈ 1.4142. - With static localization error (additive OU): • Subdiffusion: γ remains 1 at high frequency, independent of static noise (robust). • Normal diffusion: γ decreases with increasing static noise intensity σ²/D, approaching 1 for large noise; analytical correction quantified by γ̃(f, T) ≈ 1 + (1 + σ²/(2D))^{-2}. • Superdiffusion: γ decreases with static noise, with dependence on measurement time T; for long T the effect vanishes and γ approaches √2; for finite T, γ drops towards 1 as σ²/(2D) increases, per γ̃(f, T) ≈ 1 + (1 + σ²/(2D)) T^{2H−1} − 2 (asymptotic form). - With dynamic localization error (finite exposure τ_e): • Subdiffusion: γ remains ≈ 1 (robust). • Normal diffusion: γ increases above √5/2 at high frequencies; larger τ_e causes deviations to appear at lower fT. • Superdiffusion: γ essentially unaffected by dynamic error; trends match pure FBM. - Simulations (1D and 2D) confirm analytic predictions: • For H = 1/3: γ → 1 regardless of static or dynamic noise. • For H = 1/2: static noise lowers γ towards 1; dynamic noise raises γ above √5/2. • For H = 2/3: static noise lowers γ from √2 towards 1; dynamic noise has negligible effect; longer T mitigates static-noise impact. - Experimental validations: • Subdiffusive telomeres and P-granules: γ → 1 despite noticeable static localization errors. • Brownian beads in water: baseline shows increased γ at high f due to dynamic error; artificially adding static error drives γ downward toward 1, matching predictions. • Cytoskeleton fluctuations: superdiffusive behavior (2H ≈ 1.5–1.7) at 29–37 °C with γ ≈ √2 across timescales; at 13–21 °C, γ indicates subdiffusion (γ ≈ 1) at high frequencies with crossover to superdiffusion at longer times, capturing temperature-dependent motor activity. • Nanoparticles in hMSCs: despite MSD being affected by static error, γ cleanly converges to √2, confirming superdiffusion (α ≈ 1.23). Overall: γ as a single-trajectory PSD-based metric robustly distinguishes diffusion regimes and is particularly reliable for subdiffusion and for superdiffusion in the presence of dynamic error; normal diffusion is most sensitive to both noise types.

The proposed criterion based on the coefficient of variation of the single-trajectory PSD addresses the central challenge of reliably identifying anomalous versus Brownian diffusion from finite, noisy SPT data. Unlike MSD fitting, which is sensitive to static/dynamic localization errors and trajectory heterogeneity, γ exhibits regime-specific, noise-robust limiting values. For subdiffusion, γ’s invariance to both static and dynamic errors enables confident detection even when α is close to 1. For superdiffusion, γ is largely unaffected by dynamic (motion-blur) errors and only transiently impacted by static noise, with longer measurement times restoring the expected limit. For normal diffusion, γ is most susceptible to both error sources, decreasing with static noise and increasing with dynamic noise. The method also sensitively captures crossovers between regimes, as demonstrated in temperature-dependent cytoskeletal dynamics. These properties enhance interpretability of diffusion mechanisms in complex environments and provide a complementary tool alongside MSD-based analyses. Moreover, the predictable trends of γ under controlled additions of synthetic static/dynamic noise can be used as a post-processing strategy to probe measurement-error effects and approach limiting values for classification.

The study establishes the coefficient of variation γ of single-trajectory PSDs as a robust and practical criterion to diagnose anomalous diffusion from SPT data, even under realistic measurement errors. Analytical results (for static error) and extensive simulations (including dynamic error) show: (i) subdiffusion is robustly identified (γ → 1) irrespective of localization noise; (ii) superdiffusion remains robust against dynamic error and only transiently sensitive to static error; (iii) normal diffusion is sensitive to both noise types, exhibiting characteristic deviations in γ. Experimental validations across diverse systems (subdiffusive telomeres and P-granules, Brownian beads, superdiffusive cytoskeleton fluctuations and intracellular nanoparticles) corroborate these predictions. The approach can be used standalone or integrated into decision trees and feature-based or neural network pipelines to improve anomalous-diffusion classification. Future work may extend the analytical treatment to dynamic localization error, generalize beyond FBM to other anomalous processes, and exploit γ’s sensitivity to detect and quantify dynamic crossovers and ageing effects.

• Analytical treatment explicitly addresses static localization error; dynamic error effects are characterized numerically, lacking closed-form corrections. - Normal diffusion is least robust: γ is affected by both static and dynamic errors, complicating classification near α ≈ 1. - For superdiffusion, finite measurement time T and strong static noise can transiently bias γ toward 1; long trajectories are required to recover √2. - Results assume Gaussian FBM dynamics; while expected to generalize, performance on non-Gaussian or non-stationary processes may differ. - Accurate knowledge of experimental noise parameters (e.g., τ₀, σ², τ_e) aids interpretation; misestimation can affect correction and classification. - High-frequency and large-T limits underpin the asymptotic behaviors; short or low-sampling trajectories may show slower convergence and larger variance.