Fractional Motion Estimation

• Fractional Motion Estimation (FME) is a collection of techniques for estimating parameters in systems with fractional, non-integer order dynamics characterized by long-range dependence and self-similarity.
• It encompasses methods such as power variation, regression of increment-type statistics, and nonparametric approaches to derive critical parameters like the Hurst exponent and stability indices.
• FME integrates theoretical foundations with practical implementations, offering robust solutions for both statistical modeling of time series and real-time applications in video and 3D point cloud compression.

Fractional Motion Estimation (FME) refers to a class of methodologies and algorithms for estimating parameters, typically pertaining to long-range dependence and self-similar scaling, in processes or systems that exhibit fractional (non-integer order) behavior. Applications range from statistical modeling of time series—such as fractional Brownian motion (fBm), operator fractional Brownian motion (OFBM), and fractional stable motions—to practical engineering challenges in video coding and 3D point cloud compression, where motion estimation at sub-pixel or sub-voxel precision is critical for optimal performance.

1. Theoretical Foundations of Fractional Motion

Fractional-motion models underpin a variety of stochastic and signal-processing contexts:

• Fractional Brownian Motion (fBm) is a mean-zero Gaussian process with parameter $H\in(0,1)$ (the Hurst exponent) governing long-memory and self-similarity.
• Operator Fractional Brownian Motion (OFBM) generalizes fBm to multivariate processes scaling via a matrix $H$, with self-similarity expressed as $X(ct)\overset{d}{=}c^H X(t)$; each eigenvalue of $H$ interpretable as a "coordinate" Hurst exponent (Abry et al., 2015).
• Linear Fractional Stable Motion (LFSM) further generalizes to non-Gaussian $\alpha$-stable processes with three structural parameters: scale $\sigma>0$, self-similarity $H\in(0,1)$, and stability index $\alpha\in(0,2)$ (1802.06373).
• Normalized Integrated fBm (nifBm) models data that have been averaged over time, such as in environmental or financial time series, and retains key features such as stationary increments and scaling, though it is no longer strictly a fBm (Mastrogiovanni et al., 22 Sep 2025).
• In applied domains, FME also refers to algorithms for estimating optimal motion vectors at sub-pixel or sub-voxel resolution in visual data (Chen et al., 2023, Hong et al., 2022).

Key properties of these fractional models—such as stationary increments and long-range dependence—are central to both theoretical and applied FME.

2. Estimation Methodologies in Fractional Motion

A diverse set of statistical estimation strategies have been developed under FME, each tailored to process structure and observation regime.

2.1. Power Variation and Empirical Characteristic Function

For LFSM, the estimation of the triplet $(\sigma, \alpha, H)$ proceeds via negative- and positive-power variation statistics and the empirical characteristic function (ECF):

• Power Variation: Observed increments $\Delta_i^n X$ (on either high or low-frequency grids) are raised to powers $p\in(-1/2,1/2)\setminus\{0\}$, yielding $V_n(p)=\sum_{i=1}^n|\Delta_i^n X|^p$; normalization ensures a nondegenerate limit under scaling.
• Estimation of $H$: The ratio of power variations at different lags, $R_n(p)$—for instance, using second-order differences—converges in probability to a known functional $2^{pH}$, enabling a plug-in or regression estimator for $H$.
• ECF-Based Estimation of $(\alpha, \sigma)$: The real part of the ECF of rescaled increments admits explicit nonlinear equations relating it to the structural parameters. These are inverted using two frequencies $t_1, t_2$ and yield consistent estimators, with negative-power approaches extending utility in heavy-tailed regimes (1802.06373).

2.2. Regression of Increment-Type Statistics

For nifBm and OFBM, estimation hinges on functional relationships between statistics computed over different dyadic scales or blockings:

• nifBm (averaged fBm): Dyadic quadratic increments across scales are averaged and fitted to closed-form equations involving $H$ and scale parameters. In multi-component settings, a system of equations is solved to resolve multiple Hurst indices and scales (Mastrogiovanni et al., 22 Sep 2025).
• OFBM (multivariate): The empirical wavelet spectrum matrix is diagonalized at each scale; the eigenvalue exponents are then regressed linearly against the scale index to yield all Hurst eigenvalues, disambiguating mixtures of power-law behaviors inherent in multivariate scaling (Abry et al., 2015).

2.3. Increments Ratio-Based Statistics for fBm

Second-order increments and functionals of their ratios allow for robust, low-complexity estimation of the Hurst index in fBm:

• The statistic $r_i = d_{i+1}/d_i$ (where $d_i$ are second-order increments) and its nonlinear transforms $h(r_i)$ are aggregated and inverted through a deterministic transformation $\varphi$ to estimate $H$.
• Asymptotic properties, including strong consistency and normality, are obtained via ergodic and central limit theorems for stationary Gaussian sequences (Kubilius et al., 2016).

2.4. Nonparametric Estimation in Mixed Fractional Models

In SDEs driven by mixed fractional Brownian motion, nonparametric estimation of time-varying drift is accomplished by transforming the process to an equivalent semimartingale via an integral operator and applying kernel smoothing:

• Rates of mean-square convergence are directly tied to the Hölder regularity of the effective drift, with the optimal bandwidth scaling as $\epsilon^{2/(2\gamma+1)}$ in the small-noise regime (Rao, 2019).

3. Fractional Motion Estimation in Video and Point Cloud Coding

In high-throughput video and point cloud compression, FME designates a class of interpolation or interpolation-free sub-pixel/voxel search algorithms:

3.1. Interpolation-Free Error Surface FME (VVC)

• The rate–distortion (R–D) cost function around the integer motion vector (IMV) is modeled as a quadratic surface fit to nine integer-location costs.
• The extremum of this surface is computed in closed form, yielding sub-pixel coordinates directly, eliminating the need for iterative interpolation and evaluation at multiple candidate positions.
• Hardware implementations achieve dramatic reductions in area and power at minimal coding efficiency loss (e.g., 192k gates, 12.6 mW for 4K@30fps and a 0.47% BDBR loss compared to VTM 16.0) (Chen et al., 2023).

3.2. Fractional-Voxel Motion Estimation in Point Clouds

• Fractional-voxel ME addresses irregular voxel occupancy in 3D blocks by super-resolving reference frames with half-voxel insertions only between adjacent occupied voxels.
• Fractional motion vectors are searched locally across a finite grid of displacements, with color attributes propagated via nearest-neighbor correspondence and residuals coded via (Region-Adaptive) Graph Fourier Transform.
• Substantial bitrate savings and PSNR gains are documented relative to integer-only methods (e.g., BD-rate savings up to 81% vs. all-intra RAHT) (Hong et al., 2022).

4. Asymptotic Theory: Consistency and Normality

Across stochastic process settings, strong consistency and asymptotic normality theorems underpin FME methodologies:

• For power-variation/ECF approaches in LFSM, $\sqrt{n}$-consistency (or $\sqrt{n}/\log n$ in high-frequency) is attained for parameter estimators, with explicit forms for asymptotic covariances (1802.06373).
• Wavelet-eigenstructure methods for OFBM achieve asymptotically normal estimators of all Hurst eigenvalues via central limit theorems for eigenstructure regression (Abry et al., 2015).
• In the nifBm setup, CLTs and delta-method results yield explicit formulas for the variance of estimated $(H, a^2)$, and similar logic extends to two-component models (Mastrogiovanni et al., 22 Sep 2025).
• Ratio-based fBm estimators leveraging second-order increments and ergodicity yield simple linear complexity algorithms with CLT-based error quantification (Kubilius et al., 2016).

5. Practical Guidelines and Applications

Selection of FME methodology is functionally dependent on:

• Data type (continuous vs. discrete, multivariate vs. univariate, presence of averaging or drift)
• Modeling context (stationary increments, self-similarity, noise structure, memory characteristics)
• Application domain (statistical inference, real-time hardware encoding, 3D point cloud streaming)
• Resource constraints (sample size, computational complexity, hardware area/power budgets)

In averaged data contexts (e.g., finance, environmental monitoring), modeling with nifBm and estimation using quadratic increment statistics ensures consistency and efficiency. In resource-constrained, real-time video applications, interpolation-free FME with quadratic error surface fitting stands out for minimal complexity and high accuracy. For time-series models with mixed or long-range dependent noise, kernel-based smoothing on martingalized transforms achieves optimal rates given regularity constraints.

6. Implementation Considerations and Performance Results

A comparative table of representative FME methods is given below:

Application Domain	Method/Statistic	Consistency/Complexity	Notable Results
fBm time-series	2nd-order increment ratio	Strong, $O(n)$	CLT, tunable $h$ (Kubilius et al., 2016)
Multivariate OFBM	Wavelet eigenstructure	$\sqrt{n}$, multi-H	Bias $\to$ 0 for $n\gtrsim 2^{14}$ (Abry et al., 2015)
Averaged data (nifBm)	Quadratic increments (dyadic regression)	$\sqrt{N}$, $O(N)$	Asymptotic normality, large $N$ (Mastrogiovanni et al., 22 Sep 2025)
Video coding (VVC)	Error surface, quadratic fit	Hardware, minimal gates	$<$0.5% BDBR loss, 192k gates (Chen et al., 2023)
Point cloud coding	Fractional-voxel search, super-resolved	$O(27)$ over IvME	Up to 81% BD-rate gain (Hong et al., 2022)

Empirical findings consistently validate theoretical predictions: in regime-appropriate applications, these FME methods achieve high accuracy, efficient computation, and, for hardware targets, outstanding energy/performance tradeoffs.

7. Extensions, Challenges, and Future Directions

The spectrum of FME research now encompasses non-Gaussian stable processes, mixed-noise SDEs, multivariate/multifractal scaling, and real-time high-dimensional encoding/decoding. Key research frontiers include:

• Extension to non-ergodic, non-stationary, or multifractal processes;
• Rate-optimal and provably robust FME algorithms for irregular, missing, or quantized data;
• Joint parameter estimation in high-dimensional mixtures and coupled source processes;
• Architecture-specific FME acceleration in next-generation immersive media and sensor networks.

The interplay between theoretical FME methods and their hardware/software realizations will likely continue to drive both statistical methodology and engineering application in systems exhibiting intrinsic fractional dynamics.