Ray-based Optimization Schemes

• Ray-based optimization schemes are computational methods that exploit the geometric and variational properties of rays to reduce complexity in wave propagation, imaging, and inverse design.
• They employ techniques from geometric optics, discrete ordinates, and optimal transport to achieve speedups and mitigate artifacts across multiple applications.
• Practical implementations such as Successive Ray Refinement in sparse regression and adaptive ray grouping in tomography demonstrate substantial efficiency gains and accuracy improvements.

Ray-based optimization schemes are a class of computational techniques that exploit the structure and propagation of rays—formal, geometric, or computational entities—to accelerate, regularize, or otherwise improve algorithms in wave propagation, imaging, radiative transfer, network optimization, and inverse design. These methods leverage the mathematical, algorithmic, and physical properties of ray tracing and ray grouping to reduce complexity, mitigate artifacts, and enable task-specific or physically informed optimization. Ray-based optimization appears in diverse application domains, including high-frequency wave solvers, radiative transfer, optimal transport in optics, coordinate descent in machine learning, acquisition design in tomographic imaging, and more.

1. Mathematical Foundations of Ray-based Optimization

Ray-based optimization is grounded in the geometric and variational structure of rays as solution characteristics for high-frequency wave equations, radiative transport, or beam path problems. The techniques abstract from the underlying PDEs or physical models by focusing on rays as the organizing principle of computation or optimization.

• Geometric Optics and Eikonal Structure: In high-frequency regimes, the solution amplitude and phase are governed by transport along characteristics—rays—derived from eikonal equations or Hamilton-Jacobi PDEs. Ray-based optimization schemes often exploit these laws directly (e.g., ray theory for dispersion minimization (Stolk, 2015), ray-mapping for freeform optics (Bösel et al., 2015)).
• Discrete Ray Structures in Computational Algorithms: In radiative transport, the discretization of the angular domain leads to rays as fundamental computational objects, as in the S$_N$ discrete ordinates method, where angular quadratures are interpreted as rays (Frank et al., 2019, Tanaka et al., 2014).
• Variational and LP Approaches: Optimal transport formulations in reflector and lens design cast the beam-shaping problem as a constrained minimization (often a Monge–Kantorovich LP in continuous or discretized form), with rays encoded via mass-transfer between directions and targets (Glimm et al., 2011, Bösel et al., 2015).
• Gradient-based Optimization with Rays: In differentiable ray tracing, the end-to-end sensitivity of ray-parameterized models to scene geometry or material properties enables the use of automatic differentiation and smoothing for non-smooth ray-based objectives (Eertmans et al., 2024).
• Ray-Continuation in Iterative Algebraic Methods: In coordinate descent for Lasso optimization, the empirical observation that iterates lie along “rays” motivates Successive Ray Refinement (SRR), introducing one-dimensional minimizations along directions informed by previous steps (Liu et al., 2015).

2. Key Methodological Implementations

A wide variety of algorithmic instantiations implement ray-based optimization, tailored to the particulars of the physical problem or computational context:

• Discrete Ordinates with Ray Effect Mitigation: The as-S$_N$ method augments the S$_N$ scheme for radiative transfer by adding a forward-peaked artificial scattering term, which introduces angular diffusion along rays. The parameters controlling angular smoothing are scaled to vanish with grid refinement, reminiscent of artificial viscosity in spatial discretization (Frank et al., 2019). This method reduces nonphysical ray artifacts while preserving convergence.
• Parallel Ray Tracing and Grouping: Highly parallel 3D radiative transfer utilizes conflict-free batching of rays into color groups, matching the architecture of GPUs and CPU clusters, and employs grouping strategies to avoid write conflicts and optimize memory locality. The computational cost is analyzed as $O(N_m^{5/3})$ for $N_m$ mesh cells, balancing angular resolution and performance (Tanaka et al., 2014).
• Successive Ray Refinement (SRR): In coordinate descent for sparse regression, SRR exploits the empirical “ray continuation property” that iterates are nearly collinear, optimizing over a one-dimensional affine subspace between the last two iterates. Two schemes—SRR-Chain and SRR-Triangle—select the ray baseline differently, and optimal relaxation parameter $\alpha^*$ is computed using either closed-form (for $\lambda=0$) or piecewise-convex minimization (for $\lambda>0$), yielding monotonic objective reductions and substantial empirical speedups in iteration count for small regularization (Liu et al., 2015).
• Ray Mapping via Optimal Mass Transport (OMT): Freeform surface design reduces to a ray mapping problem, where the integrability condition for constructing a continuous surface is approximately satisfied by curl-free OMT maps. The remaining surface geometry is solved via a linear advection equation with characteristics aligned to the mapped rays. This decouples ray assignment (OMT) from surface solving, providing computationally efficient and physically consistent designs (Bösel et al., 2015).
• Sparse and Hierarchical Ray Data Structures: Non-LTE radiative transfer leverages mipmapping and sparse voxel grids to reduce the effective computational resolution where ray-path quantities (emissivity, opacity) are smooth or negligible. A hierarchical digital differential analyzer (HDDA) marches rays at the maximum allowed block size, as determined by a variance-limited metric, yielding up to 8x speedup with negligible physical error (Osborne, 11 Nov 2025).

3. Task-specific and Problem-driven Ray Selection

Ray-based optimization is also applied in experimental design and data acquisition where the physical or information-theoretic content of each candidate ray is assessed relative to task-specific objectives.

• Acquisition Optimization in Tomography: In anisotropic X-ray dark-field tomography (AXDT), task-driven acquisition is formulated by maximizing a detectability index (e.g., NPWM observer $d'$) across possible measurement (ray) orientations. An accelerated greedy algorithm with batched selection iteratively adds rays that most enhance the detectability metric, achieving near-optimal ROI sensitivity with significantly reduced acquisition load compared to uniform designs (Cheslerean-Boghiu et al., 2022).
• Adaptive Ray Pool Construction: The problem is formulated over a pool $\mathcal{G}$ of candidate ray poses, with constraints on number, geometry, and coverage. Extensions include online adaptation, replacement of the forward model or observer, and incorporation of non-spatial task templates.
• Reflector and Lens Design via Iterative Ray-based LPs: Dual LP formulations for beam shaping identify active constraints (which correspond to rays mapping input-to-output apertures) and prune the candidate set via proximity to constraint saturation, enabling tractable optimization at mesh sizes unattainable by global discretization, with sublinear (in constraint set size) memory growth (Glimm et al., 2011).

4. Artifact Mitigation and Algorithmic Acceleration

A central motivation for ray-based optimization is artifact reduction and efficiency improvement in high-frequency or high-dimensional problems.

• Ray Effect Suppression in Discrete Ordinates: Artificial angular diffusion in as-S$_N$0 reduces spurious ray imprinting and enables prescribed accuracy with lower memory and ordinate count, significantly improving computational feasibility for target error tolerances (Frank et al., 2019).
• Ray Grouping for Conflict-free Parallelism: Partitioning rays into conflict-free batches enables high parallelism without atomic operations, nearly linear scaling with hardware resources, and optimized bandwidth utilization. Cost analysis of ray/group counts provides scalability forecasts and informs optimal sampling parameter choices (Tanaka et al., 2014).
• Task-driven Ray Set Construction: By aligning acquisition schemes with the spatial/frequency structure hinted by pilot data, task-specific optimization avoids unnecessary measurements and maintains high fidelity for the features of interest, as demonstrated in both simulated and experimental AXDT data (Cheslerean-Boghiu et al., 2022).
• Discontinuity Smoothing for Differentiable Ray Tracing: Smoothing geometric discontinuities with parametrized approximations (e.g., tanh-based transitions) ensures that ray-tracing-based loss functions are differentiable everywhere, circumventing zero-gradient regions and enabling provable success in gradient-based optimization (Eertmans et al., 2024).

5. Generalization and Applicability Across Domains

Ray-based optimization frameworks exhibit broad generalizability due to their abstraction over domain-specific details:

• Physics-independent Ray Abstractions: The separation between ray generation, validity/weight assignment, and aggregation allows methods to be ported between radio, optical, acoustic, or neutron transport problems by replacing loss/propagation laws and geometric operators.
• LP and Variational Ray-based Methods: The duality-based LP framework is readily adapted to lens, kinoform, or refractive surfaces by changing the cost kernel, while the iterative pruning procedure remains robust across geometry and intensity distributions (Glimm et al., 2011).
• Ray-theoretic Stencil Optimization: The ray-theory-based tuning of discretization parameters in wave solvers is extendable to finite-difference, finite-element, spectral, and DG schemes for a wide class of linear hyperbolic PDEs by matching phase and amplitude response along rays (Stolk, 2015). The procedural steps—formulating the discrete dispersion relation, parameterizing coefficients, and minimizing phase slowness error—are invariant across these cases.
• Adaptive Smoothing and Batching: Differentiable ray tracing extends to any scenario with non-smooth optimization landscapes determined by ray/geometry interaction, with smoothing/annealing controlling the bias-variance tradeoff in gradient information (Eertmans et al., 2024).

6. Performance, Scaling, and Limitations

Empirical results across ray-based optimization schemes consistently report substantial efficiency gains, error reductions, or memory savings.

Method	Domain	Key Performance Metric	Reported Gain
as-S$_N$1 (Frank et al., 2019)	RTE/Discrete Ordinates	L$_N$2 error at fixed memory	Up to 3x fewer ordinates
Parallel RT (Tanaka et al., 2014)	3D Diffuse RT	Scaling on CPUs/GPUs	Near-linear, 8x+ cores
AXDT AGS (Cheslerean-Boghiu et al., 2022)	Tomography	Task detectability (ROI, EM, RMSE)	30–50% lower RMSE
SRR-CD (Liu et al., 2015)	Lasso Optimization	Pass count for small $_N$3	4–12x fewer iterations
VLM+HDDA (Osborne, 11 Nov 2025)	Non-LTE RT	Iteration time, 99.9th-% population error	8x speedup, <0.5% error

Limitations include tuning hyperparameters (e.g., diffusion width, smoothing thresholds), potential oversmoothing/overaveraging in highly anisotropic or sharply heterogeneous domains, absence of global optimality guarantees in greedy selection techniques, and additional complexity in code architecture or data management due to hierarchical or nonstandard data structures.

7. Future Directions and Extensions

Recent research highlights several open avenues for ray-based optimization:

• Integration with Learning-based Models: Deep learning for ray field extraction or proxy generation (e.g., in high-frequency Helmholtz solvers (Yeung et al., 2021)) promises further acceleration and adaptivity by combining ray-based models with data-driven ray inference.
• Adaptive and Online Strategies: Extending batching, smoothing, or active-set pruning to online and adaptive frameworks, dynamically adjusting ray pools or discretization to evolving objectives or data distributions.
• Cross-domain Portability: Formalization of ray abstraction APIs (e.g., in DiffeRT2d (Eertmans et al., 2024) or DexRT (Osborne, 11 Nov 2025)) to generalize optimizers across radio, optical, thermal, and quantum domains, leveraging domain-specific physical models while reusing ray-based algorithmic infrastructure.
• Rigorous Analysis: Theoretical study of performance guarantees, error propagation under adaptive smoothing and multi-resolution, and stability in iterative/pruned LP frameworks.
• Hardware-aware Implementations: Further optimization for emerging GPU, many-core, and custom accelerator architectures, with focus on persistent memory, data layout, and communication patterns as shaped by ray grouping and batching.

Ray-based optimization schemes thus constitute a versatile, physically grounded, and algorithmically rich toolbox for high-dimensional, multi-physics, or computation-constrained optimization tasks across scientific computing, imaging, and inverse design.