Multiple-Scale Spatial Marching Method

• Multiple-scale spatial marching is a computational template that separates global ordering from local refinement to regularize and efficiently solve complex spatial problems.
• It uses distinct coarse scales like signed-distance thresholds, edge counts, or slower progression variables alongside detailed local procedures such as primitive fitting, triangulation, or iterative sweeping.
• Applications include shape abstraction, subgrid mesh reconstruction, Eikonal equation solvers, and turbulent flow analysis, balancing computational speed with solution accuracy.

Multiple-scale spatial marching denotes a family of coarse-to-fine computational strategies in which propagation at one scale establishes admissible global structure, while a second, more local procedure resolves geometry, topology, or dynamics within that structure. In current arXiv literature, this pattern appears in at least four technically distinct settings: primitive abstraction from signed distance functions, subgrid surface reconstruction on tetrahedral grids, hybrid solvers for the Eikonal equation, and high-Reynolds-number spatial development of shear flows (Liu et al., 2023, Baktash et al., 30 May 2026, Chacon et al., 2011, Song et al., 5 Aug 2025). Despite the disparity of applications, the shared principle is to separate a global ordering variable—signed-distance level, cell priority, tetrahedral face consistency, or slow streamwise coordinate—from a local reconstruction or relaxation step that is computationally richer.

1. Common architectural pattern

A multiple-scale spatial marching method combines two forms of progression. One form establishes coarse causality or admissible connectivity. The other performs local inference, fitting, or relaxation inside the currently active region. In the signed-distance setting, the coarse variable is a descending sequence of negative thresholds, and the local stage is probabilistic primitive fitting in voxel space. In Subgrid Marching Tetrahedra, the coarse object is the combinatorial boundary information on tetrahedral faces induced by arbitrary edge-intersection counts, and the local stage is deterministic triangulation of the resulting loops. In Eikonal solvers, marching is done on coarse cells while sweeping is done on fine grids inside each cell. In curved-channel flow, the slow variable is the downstream coordinate $X$, while the local fast-scale problem is a periodic exact coherent structure posed in a local box (Liu et al., 2023, Baktash et al., 30 May 2026, Chacon et al., 2011, Song et al., 5 Aug 2025).

Setting	Coarse-scale march	Local refinement
Signed-distance abstraction	Descending negative SDF thresholds	EM-like primitive fitting on active voxels
Tetrahedral contouring	Facewise consistency from edge counts	Per-tet loop triangulation and subdivision
Eikonal equations	Coarse cell ordering by marching or cell heap	Sweeping or locking sweeping within cells
Curved-channel flow	Downstream march in slow coordinate $X$	Local ECS solve with Reynolds-stress feedback

This suggests that the phrase is best understood as a methodological template rather than a single canonical algorithm. The template is especially useful when a direct fine-scale solve is expensive or ambiguous, but a coarser ordering variable can regularize the problem before local computation begins.

2. Signed-distance-domain and voxel-space marching in primitive abstraction

In "Marching-Primitives" (Liu et al., 2023), the multiple-scale spatial marching method is defined on a signed distance function $f$, with the sign convention negative inside and positive outside. The operational interior sub-level sets are

$$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$

discretized on a voxel grid $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$ with scalar values $d(x_i)$. Connectivity is computed in the voxel graph using 26-connectivity,

$$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$

and connected components are retained only when their cardinality is at least $N_c$. Thresholds follow a geometric sequence,

$$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$

so marching proceeds from deep interior to near-surface regions.

The method alternates between signed-distance marching and spatial marching. At threshold $t^c_m$, connected interior volumes of interest act as seeds. For each such seed, a primitive is initialized from an ellipsoidal bounding-box estimate and then refined by an EM-like procedure. Superquadrics are the main primitive family, with implicit function

$X$0

where $X$1, $X$2, and $X$3. Because the exact SDF of superquadrics lacks a closed form, the method uses the radial approximation

$X$4

with truncated signed distances near zero.

The probabilistic extraction step assigns soft correspondences between voxels and a candidate primitive. For voxel $X$5 with observed target SDF $X$6, the posterior $X$7 weights a least-squares update

$X$8

where

$X$9

Interior voxels receive soft association, while exterior voxels always contribute by penalizing leakage. After fitting, voxels satisfying $f$0 and $f$1 are deactivated, and the threshold sweep continues.

The method is explicitly presented as a coarse-to-fine decomposition: prominent interior parts are recovered first, and finer features are captured near the zero level. The claimed computational profile separates connectivity marching, with cost $f$2, from EM-like primitive updates, with cost $f$3, where $f$4 is an active band near the primitive surface. A dense grid stores $f$5 SDF values, with the example of $f$6 MB at $f$7, whereas the final abstraction stores $f$8 parameter vectors, stated as $f$9D per superquadric and yielding multi-order compression such as $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$0–$$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$1 KB for complex shapes or scenes.

Empirically, on ShapeNet using $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$2 SDF grids over $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$3, Marching-Primitives with superquadrics achieved mean Chamfer-L1 $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$4 and IoU $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$5 across $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$6 categories, compared with a non-parametric baseline at $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$7 Chamfer-L1 and $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$8 IoU, and a learning-based superquadric method at $$\Omega(\tau)=\{x\in \mathbb{R}^3: f(x)\le \tau\}, \qquad \tau<0,$$9 Chamfer-L1 and $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$0 IoU. For chairs, the reported IoU is $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$1 versus $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$2 and $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$3. On a large real indoor scene, Reading Room, the scene was abstracted at $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$4 grid resolution, and the primitive set occupied $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$5 KB versus $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$6 MB for the SDF. The principal trade-off is representational: meshes can capture arbitrary detail, whereas a primitive vocabulary may miss thin or intricate topologies.

3. Subgrid-scale marching on tetrahedral grids

"Subgrid Marching Tetrahedra" (Baktash et al., 30 May 2026) formulates a multiple-scale spatial marching method for reconstructing a manifold, intersection-free triangle mesh from edge-surface intersections on a tetrahedral grid. The input is not a node-sampled sign field but, for every grid edge $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$7, the ordered set of surface intersection points and the count $V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$8. The global generalized edge-normal coordinate vector is

$V=\{x_i\in \mathbb{R}^3, i=1,\dots,N\}$9

For a tetrahedron with vertices $d(x_i)$0, the six edge counts are grouped as

$d(x_i)$1

The face-level admissibility conditions are purely combinatorial. On a face $d(x_i)$2, a collection of normal arcs can exist if and only if the edge counts satisfy an even-sum condition,

$d(x_i)$3

and the triangle inequalities,

$d(x_i)$4

When these hold, corner coordinates are

$d(x_i)$5

If parity is odd, the method deterministically subtracts $d(x_i)$6 from all three edge counts and recomputes corner data; if a triangle inequality is violated, the offending edge count is reduced and the residual is handled by “scoops.”

Per tetrahedron, the primal algorithm constructs boundary segments on each face, assembles closed boundary loops, classifies them as normal or non-normal, and triangulates them according to loop type. Corner triangles are emitted for $d(x_i)$7-segment loops. Quads are split along a consistent diagonal. Octagons use Steiner points placed along the segment between midpoints of the two opposite edges carrying $d(x_i)$8 intersections. Single long loops are triangulated by fanning to an interior point. Multiple long loops trigger recursive subdivision with an interior vertex $d(x_i)$9 and induced interior edge counts

$$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$0

with termination in $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$1 steps and no triggering below $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$2 total edge intersections.

The method’s multiple-scale character lies in its ability to represent arbitrarily many surface patches inside a single tetrahedron and to recover features thinner than the grid spacing. The paper states that the encoding sidesteps the usual Nyquist--Shannon limit because it records all edge intersections rather than a single sign change per edge. Reconstruction remains local to each tetrahedron, but face-level boundary-curve construction is deterministic and identical on shared faces, so neighboring tetrahedra glue without cracks.

The mathematical guarantees are stronger than those of classical marching cubes or marching tetrahedra. In the even-parity case, the primal output is an abstract simplicial $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$3-complex homeomorphic to a manifold, possibly with boundary. In odd-parity cases, the primal output is still edge-manifold and interior-vertex-manifold. The primal triangulation is globally embedded and intersection-free by construction, through convex-hull containment, nonintersection properties of corner triangles and smeared polygons, monotone Steiner ordering for octagons, and contraction arguments for non-normal diagonal-type loops. The method also supports surfaces with boundary and does not require a global inside/outside classification or consistently oriented input geometry.

Quantitatively, the paper reports large-scale tests on $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$4 models in which SMT achieves lower error for the same grid resolution and similar time, and often matches the error of classic marching at a much coarser grid, with the example that SMT at $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$5 is comparable to classic marching at $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$6. It further reports that an ablation called “subgrid mod 2,” which uses exact intersection positions but allows at most one polygon per tet, yields only minor improvements over classic marching and still fails to capture subgrid topology. The stated limitation is that SMT preserves all observed edge intersections, but global topology is not guaranteed to match the target surface, which is explicitly identified as the same limitation as classic marching.

4. Two-scale marching and sweeping for Eikonal equations

For Eikonal equations, the multiple-scale spatial marching method described by Chacon and Vladimirsky (Chacon et al., 2011) addresses the isotropic boundary-value problem

$$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$7

on a uniform Cartesian grid $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$8 with spacing $$N(v)=\{u\in V: u\neq v,\ \max(|u_x-v_x|,|u_y-v_y|,|u_z-v_z|)\le 1\},$$9. The monotone, causal discretization is the Godunov upwind scheme, and the standard local update at grid point $N_c$0 uses

$N_c$1

If $N_c$2, then

$N_c$3

otherwise

$N_c$4

The paper’s central observation is that Fast Marching Method and Fast Sweeping Method are complementary. Fast Sweeping is efficient when characteristics are largely straight, while Fast Marching is usually more efficient when characteristics change direction frequently or the domain geometry is complicated. The proposed hybridization uses marching on a coarser scale and sweeping on a finer scale. Three methods are defined.

The Fast Marching–Sweeping Method (FMSM) first runs Fast Marching on coarse cell centers to obtain an acceptance order of cells, then sweeps each accepted cell with a limited number of directional sweeps inferred from coarse causality. It is a one-pass method and is the fastest of the three, but it may introduce additional error. The Heap-Cell Method (HCM) is a cell-level label-correcting method: cells are keyed in a heap by a cell arrival time $N_c$5, neighbor cells are re-enqueued when interface data improve, and sweeping inside each cell is performed to convergence by Locking Sweeping. HCM is stated to converge to the exact discrete solution. The Fast Heap-Cell Method (FHCM) uses the same cell-heap coupling logic as HCM, but performs at most four sweeps per cell in preferred directions, using a monotonicity check on interface values to reduce sweep directions; it is faster than HCM but introduces small additional error.

For HCM and FHCM, the cell arrival time is computed from the largest current interface value on a processed cell and a one-step estimate into the neighbor: $N_c$6

$N_c$7

This is the coarse-scale marching quantity that schedules fine-scale sweeping. In HCM and FHCM, a neighbor cell is enqueued when an interface node in the processed cell decreases and becomes smaller than the corresponding node in the neighbor.

The paper organizes complexity around the number of fine-grid nodes $N_c$8, the number of cells $N_c$9, and the number of sweeps. FMM costs $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$0. FSM costs $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$1, with $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$2 the number of sweeps. FMSM adds coarse $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$3 marching to a few sweeps per cell. HCM and FHCM use a heap of cells, and for sufficiently small coarse cells the average heap removals per cell tends to $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$4. HCM preserves exactness; FHCM generally gives the best speed/accuracy trade-off.

The numerical results are reported extensively. On a $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$5 comb maze with $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$6 barriers, FMM took $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$7s, FSM $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$8s, and LSM $$t^c_1=\min_{x_i\in V} d(x_i), \qquad t^c_{m+1}=\alpha t^c_m, \quad \alpha\in(0,1),$$9s. Hybrid methods were faster: HCM about $t^c_m$0–$t^c_m$1s, FHCM $t^c_m$2–$t^c_m$3s, and FMSM $t^c_m$4–$t^c_m$5s, with $t^c_m$6. On a $t^c_m$7 $t^c_m$8 checkerboard, FMM took $t^c_m$9s, FSM $X$00s, and LSM $X$01s; FHCM took $X$02–$X$03s, HCM $X$04–$X$05s, and FMSM $X$06–$X$07s. For $X$08, where characteristics are straight, FMM took $X$09s, FSM $X$10s, LSM $X$11s, FMSM $X$12–$X$13s, FHCM $X$14–$X$15s, and HCM about $X$16–$X$17s, with no extra errors. These results support the paper’s core thesis that coarse causality and fine local relaxation can be separated profitably.

5. Slow–fast spatial marching in non-uniformly curved channel flow

In "Spatial marching with subgrid-scale local exact coherent structures in non-uniformly curved channel flow" (Song et al., 5 Aug 2025), the multiple-scale spatial marching method is formulated for high-Reynolds-number flows with slow streamwise variation. The geometry is a non-uniformly curved channel with dimensional half-width $X$18 and centreline radius of curvature $X$19, written in body-fitted orthogonal coordinates $X$20. Non-dimensionalization uses $X$21, $X$22, and fluid density, giving Reynolds number

$X$23

Weak curvature is imposed through the Dean-number scaling

$X$24

with slow streamwise variation on the coordinate $X$25.

The large-scale mean flow is governed by Boundary Region Equations obtained from the slow-variable expansion

$X$26

$X$27

At leading order,

$X$28

$X$29

Because streamwise diffusion is negligible at leading order, the BRE are parabolic in $X$30 and can be marched downstream without downstream boundary conditions.

The local fine scale is represented by a fast phase

$X$31

with local streamwise wavenumber $X$32, frequency $X$33, and phase speed $X$34. Fields are decomposed into a $X$35-averaged mean and a $X$36-periodic fluctuation. Under the paper’s assumptions, the mean equations become forced BRE,

$X$37

where the forcing is the $X$38-averaged Reynolds stress from the fast fluctuation,

$X$39

The fluctuation equations define a local exact coherent structure in a box $X$40, periodic in $X$41 and $X$42, no-slip at $X$43, and closed by the requirement $X$44.

The numerical procedure marches station by station in $X$45. At each step, the mean-field slow derivatives are discretized by an implicit backward difference, while the coupled mean and fluctuation potentials are solved simultaneously by Newton–Raphson. The phase condition $X$46 is enforced to $X$47 by line search on $X$48. If the fluctuation amplitude drops below threshold or Newton fails to converge to a nontrivial fluctuation, the method switches to BRE-only marching, recovering Dean vortices.

The inlet condition is a travelling-wave ECS of plane Poiseuille flow. For the MS-A case at $X$49, the inlet parameters are $X$50, $X$51, $X$52, $X$53, $X$54, and $X$55. For MS-A2, they are $X$56, $X$57, $X$58, $X$59, $X$60, and $X$61. The curvature profile is prescribed by

$X$62

with $X$63, $X$64, and $X$65, and a reversed profile $X$66 is used for decreasing-curvature runs.

The paper uses the streamwise pressure gradient $X$67 as the primary transport metric. For the z-averaged streamwise momentum equation,

$X$68

evaluation at the lower wall gives

$X$69

The paper reports that in the increasing-curvature MS-A run, fast-scale fluctuations persist downstream and the outlet solution approximates an annulus ECS with $X$70, $X$71 versus $X$72 for full Navier–Stokes. In MS-A2, the fast fluctuation disappears at $X$73, and the solution transitions to a BRE-only Dean vortex.

A second use of the marching framework is subcritical triggering. Starting from Dean vortices with decreasing curvature, the unforced BRE-only solution decays to laminar flow when curvature vanishes. If a small $X$74-periodic forcing is applied near the neutral point of the Dean-vortex secondary instability, nontrivial fluctuations can persist downstream even after the curvature vanishes. For $X$75, neutrality occurs at $X$76 with $X$77. The paper further reports a band of unstable frequencies and notes that at $X$78 marching stabilizes again and converges to a new ECS, labeled MS-A3. The method is explicitly contrasted with PSE: it enforces fast periodicity and slow dependence to remain asymptotically consistent over long distances.

6. Scope, limitations, and recurrent misconceptions

A recurring misconception is that multiple-scale spatial marching is synonymous with classical isosurface extraction. The cited literature shows otherwise. In Marching-Primitives, the goal is not a polygonal zero set but a sparse, analytic, continuous SDF approximation with part structure (Liu et al., 2023). In SMT, the method reconstructs a manifold, intersection-free triangle mesh from arbitrary edge intersections and can emit multiple disjoint disks within one tetrahedron, which lies beyond the finite lookup-table logic of classic marching cubes or marching tetrahedra (Baktash et al., 30 May 2026). In Eikonal solvers, marching is not a surface-extraction device at all; it is a causal scheduling mechanism for coarse blocks, combined with fine-scale sweeping (Chacon et al., 2011). In the channel-flow problem, spatial marching refers to downstream integration of a parabolized slow system coupled to fast local exact coherent structures (Song et al., 5 Aug 2025).

The principal limitations are domain-specific. Primitive abstraction can miss thin structures smaller than voxel spacing, can fail on highly non-convex shapes outside the convex superquadric regime, and can degrade when the truncation threshold $X$79 is overly small or when the SDF is extremely noisy (Liu et al., 2023). SMT depends on reliable one-dimensional root finding along edges; if intersections are missed, the method still produces valid manifold output but may not recover the full topology, and codimension-2 features that never intersect edges remain invisible (Baktash et al., 30 May 2026). The Eikonal hybrids depend strongly on cell size and alignment with speed discontinuities: poor alignment increases cyclic inter-cell dependencies, produces extra error in FMSM, and causes more re-enqueues in HCM and FHCM (Chacon et al., 2011). The curved-channel formulation assumes high Reynolds number, slow streamwise variation, weak curvature, and strict fast-scale periodicity; it also requires either an inlet ECS or a forcing scenario that can trigger one (Song et al., 5 Aug 2025).

A second misconception is that multiscale marching always sacrifices correctness for speed. The evidence is mixed rather than uniform. HCM is stated to converge to the exact discrete Eikonal solution, whereas FHCM and FMSM introduce small additional errors. SMT’s primal variant is designed to guarantee embeddedness and manifoldness, but not necessarily exact recovery of the target topology. Marching-Primitives improves compactness and part-level structure, but only within the expressiveness of its primitive family. The curved-channel method preserves asymptotic consistency under its scaling assumptions, but it is not a full direct numerical simulation.

Taken together, these works indicate that the defining feature of multiple-scale spatial marching is not a single data structure or PDE discretization. It is the deliberate separation of global admissibility from local resolution. This suggests why the approach recurs across computer vision, geometry processing, numerical Hamilton–Jacobi theory, and wall-bounded turbulence: each field contains problems in which coarse spatial order is available earlier or more robustly than full fine-scale structure, and marching becomes the mechanism that turns that order into an efficient computational pipeline.