Dual Volume Packing: Theory & Applications

• Dual volume packing is a framework that leverages dual linear programming, combinatorial decompositions, and bipartite structures to derive rigorous bounds and efficient algorithms.
• In operations research, it underpins compact MILP models and EPTAS for splittable bin packing under cardinality constraints, achieving near-optimal performance.
• In 3D vision, the approach enables part-level object generation by partitioning contact graphs into dual volumes, ensuring non-collision and enhanced semantic coherence.

Dual volume packing encompasses a set of rigorous mathematical and algorithmic frameworks leveraging dual structures in packing-type problems across combinatorial optimization, convex geometry, and generative modeling for 3D vision. The term most prominently appears in three distinct, yet structurally related, lines of research: (1) dual volume packing in high-dimensional geometry and sphere packing via dual linear programming; (2) dual splittable bin packing under cardinality constraints in operations research; and (3) dual volume packing as a part-level organization and latent representation mechanism in 3D object generation from images. Each context utilizes a conceptually dual construct—whether in the sense of LP duality, graph-theoretic bipartition, or combinatorial decompositions—to yield efficient algorithms or tight bounds.

1. Dual Volume Packing in High-Dimensional Geometry

In the geometric context, dual volume packing refers to the derivation of density upper bounds for sphere (or unit ball) packings in Euclidean space $\mathbb{R}^d$ using dual linear programming (LP) formulations. The classical Cohn–Elkies framework expresses the problem as an infinite-dimensional LP: given a rapidly decreasing, radial function $f:\mathbb{R}^d\to\mathbb{R}$ with certain positivity and support conditions on $f$ and its Fourier transform $\widehat f$, one obtains an upper bound on the optimal packing density $\delta_d$ via

$$\delta_d \leq \left(\tfrac{r}{2}\right)^d \frac{f(0)}{\widehat f(0)}.$$

The optimal value $CE_d$ of the Cohn–Elkies LP is in general not tight for $d > 2$, except in exceptional dimensions ($d=8,24$).

The dual LP, living on the space of Radon measures, assigns mass to constraints in the primal (vanishing outside a radius, nonnegativity of the Fourier transform), yielding a dual objective that, for feasible dual variables, certifies rigorous upper bounds. However, the infinite-dimensional character precludes computer solution. Recent advances discretize this LP—by sampling on the integer lattice $\mathbb{Z}^d$ and modding out by $m$—to produce a finite-dimensional dual LP whose optimal solution $D_{d,m,r}$ retains the rigorous upper bound property: $\delta_d \leq CE_d \geq \max_{m, r} D_{d,m,r},$ where feasible duals $\mu,\lambda$ satisfy

$\max(\mathrm{Dual}_{d,m,r}) = m^{-d/2}\lambda(0)$

subject to symmetry-adapted linear constraints. Numerical optimization and post-hoc rationalization of dual solutions certify that $D_{d} > \delta_d$ for $3\leq d\leq 13$ (excluding $d=8$), hence proving non-sharpness of $CE_d$ in these cases (Li, 2022).

2. Dual Splittable Bin Packing With Cardinality Constraints

In combinatorial optimization, dual volume packing arises as the dual of the splittable bin packing problem under cardinality constraints (DSBP) (Jaykrishnan et al., 2022). Given $n$ items $I=\{1, \dots, n\}$, each with size $s_i>0$, and $m$ bins $B$, the goal is to assign possibly fractional parts $x_{i,b}$ of each item $i$ to bins, subject to:

• Completeness: $\sum_{b\in B} x_{i,b} = 1$ for $i=1,\dots, n$,
• Per-bin load constraint: $\sum_{i\in I} x_{i,b}s_i \leq L$ (makespan $L$ to minimize),
• Per-bin cardinality: At most $k$ distinct items per bin.

The problem admits the following compact MILP model: $$\begin{aligned} & \text{Minimize } L \ \text{subject to}\quad & \sum_{b=1}^m x_{i,b} = 1,\quad i=1,\dots, n \ & \sum_{i=1}^n s_i x_{i,b} \leq L, \quad b=1,\dots, m \ & x_{i,b} \leq y_b^i,\quad i=1,\dots, n;\ b=1,\dots, m \ & \sum_{i=1}^n y_b^i \leq k,\quad b=1,\dots, m \ & x_{i,b} \geq 0,\ y_b^i \in\{0,1\}. \end{aligned}$$ An efficient polynomial time approximation scheme (EPTAS) addresses this dual problem by rounding item sizes, classifying large/small items, enumerating patterns/configurations, and solving a small MILP. The EPTAS achieves a $(1+\epsilon)$-approximation in time $f(1/\epsilon, k)\cdot \mathrm{poly}(n)$ (Jaykrishnan et al., 2022).

3. Dual Volume Packing for Part-Level 3D Object Generation

In 3D vision, dual volume packing denotes a reconstruction and representation strategy enabling efficient part-level object synthesis from a single RGB image (Tang et al., 11 Jun 2025). The methodology proceeds as follows:

• Given a 3D mesh with $K$ ground-truth parts $\{P_1, ..., P_K\}$, construct a contact graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$, with $\mathcal{V} = \{1, ..., K\}$ and $(i,j) \in \mathcal{E}$ iff $P_i$ and $P_j$ collide.
• Minimize the set of edges $\mathcal{E}_c \subseteq \mathcal{E}$ to produce $$\mathcal{G}' = (\mathcal{V}, \mathcal{E}\setminus\mathcal{E}_c)$$ bipartite; then 2-color the graph so parts decode into two non-contacting volumes.
• Pack all parts into two signed-distance-field (SDF) volumes $$V_1, V_2 \in \mathbb{R}^{H \times W \times D \times 1}$$. Voxels are marked occupied if $|V_k(x)| < \tau$.
• A VecSet-based VAE encodes each packed volume to latent vectors $z_1, z_2$; a diffusion model denoises their joint latent code conditioned on image features. After decoding, disjoint part meshes are retrieved via Marching Cubes and assembled.

This dual organization (bipartition of the contact graph) ensures a fixed-size latent code for any number of parts, streamlines generative modeling, and alleviates part collision (Tang et al., 11 Jun 2025). Empirical results show improved editing and semantic coherence compared to generation schemes that fuse all parts as a monolithic mesh.

4. Mathematical and Algorithmic Formulations

All three domains use mathematically formal dual constructs:

• Linear Programming Duality: For sphere packing, explicit duality in LPs enables rigorous certification of upper bounds on the optimal packing density $\delta_d$ by solving a corresponding dual over discretized measure variables (Li, 2022).
• Mixed Integer Formulation: For DSBP, rigorous constraints (cardinality, load, completeness) are imposed through integer and continuous decision variables, and structure-preserving rounding/classification achieves EPTAS guarantees (Jaykrishnan et al., 2022).
• Graph Bipartition and Latent Decomposition: In 3D object generation, a combinatorial 2-coloring of the part contact graph supports dual-volume latent encoding, enabling stochastic denoising and mesh extraction for arbitrary part cardinality (Tang et al., 11 Jun 2025).

Key methodological patterns include discretization/reduction, exploitation of problem symmetries, bounded enumeration of patterns/configurations, and compositionality via dual structures.

5. Practical Applications and Empirical Results

• Geometric Packing: The finite dual LP bounds provide computer-verifiable certificates for the non-sharpness of the Cohn–Elkies LP in all dimensions $3 \leq d \leq 13$ except $d=8$, with explicit tabulated $D_d$ values strictly exceeding known densest packings (Li, 2022).
• Combinatorial Optimization: The DSBP EPTAS solves instances with makespan within $(1+O(\epsilon))$ of optimal and running time $f(1/\epsilon, k)\cdot \mathrm{poly}(n)$, employing a sequence of rounding, classification, MILP solving, and integrality-restoring matchings (Jaykrishnan et al., 2022).
• 3D Vision: Dual volume packing in image-to-3D frameworks enables part-aware synthesis within $\sim$30 seconds per shape. Comparative experiments show higher mesh quality, semantic part coherence, and faster inference than segmentation-plus-completion pipelines (which scale with number of parts) (Tang et al., 11 Jun 2025). Implementation relies on large-scale datasets ($\sim$254K bipartite-filtered meshes), substantial compute (up to 256 A100s), high-resolution SDF volumes, and progressive training with VAE and rectified flow objectives.

6. Limitations and Open Challenges

Dual volume packing schemes are subject to inherent representational and computational limitations:

• High-Dimensional Packing: The discretized dual LP approach cannot provide sharp Cohn–Elkies bounds outside exceptional dimensions; for all other $d$, the bound exceeds the best-known packing (Li, 2022).
• Bin Packing with Cardinality: EPTAS runtime scales singly-exponentially in $1/\epsilon^2$; practical solvability is limited for very small $\epsilon$ or large $k$ (Jaykrishnan et al., 2022).
• 3D Object Synthesis: The dual volume decomposition cannot model mutual contact among more than two parts (chromatic number $>2$), is limited by dataset-driven part granularity, and may break down for highly entangled part-contact graphs that resist complete bipartitioning. The number of volumes is fixed to two, and part-level semantic granularity is not user-controllable (Tang et al., 11 Jun 2025).

7. Interrelations and Theoretical Significance

A recurring theme is the operational benefit of dualizing: efficiently certifying bounds, shrinking solution search spaces, and enabling tractable model architectures in high-combinatoriality domains. In all cases, dual volume packing leverages a combinatorial or functional decomposition (LP duality, graph bipartitioning, MILP enumeration) to advance both theoretical limits and empirical performance. The dual perspectives inform algorithmic design (as in EPTAS or part-level latent diffusion), and provide rigorous certificates in classical open problems of geometry and complexity.

Dual volume packing thus serves as a bridge between mathematical optimization, geometric analysis, and deep generative modeling, with active research focused on extending duals to richer structures (e.g., higher chromatic numbers, multi-volume decompositions, and beyond bipartite formulations).