Mass Assignment Problem Overview

• Mass Assignment Problem is a family of constrained allocation issues that spans discrete optimization, optimal transport, cybersecurity vulnerabilities, and topological measure assignments.
• In discrete optimization, it encompasses one-to-one matching, conflict-constrained, capacitated, and multidimensional assignments, with solutions like the Hungarian and auction algorithms.
• Its applications include personnel scheduling, beam shaping, API security analysis, and topological partitions, highlighting trade-offs in feasibility, optimality, and computational complexity.

“Mass Assignment Problem” is not a single universally fixed term. In current technical usage it denotes several non-equivalent problems: a discrete assignment or generalized allocation problem in operations research; a discretized form of mass transport in optimal transport; a security vulnerability in REST APIs caused by unintended attribute binding; and, in combinatorial topology and discrete geometry, a partition problem for continuously varying measures on bundles and Grassmannians. The common thread is constrained allocation, but the mathematical objects differ substantially: permutation matrices and matchings, transport couplings, HTTP request attributes, or fiberwise measures.

1. Discrete assignment and matching core

In its most classical optimization sense, the problem is a one-to-one assignment between two finite sets, typically agents and tasks, represented by a cost or weight matrix. The standard linear assignment formulation uses binary variables $x_{ij}$ and enforces exactly one assignment per row and per column: $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$ subject to

$$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$

In the optimal-transport interpretation, this is the discrete Monge problem with unit masses and no splitting, so feasible solutions are permutation matrices (Seyedi et al., 4 Sep 2025).

A max-weight variant appears in personnel scheduling, where workers and tasks form the two sides of a weighted bipartite graph $G=(X\cup Y,E)$, and the edge weight $w(x,y)$ quantifies efficiency or productivity. The objective is

$\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$

over matchings $M\subseteq E$. The paper “The Application of Bipartite Matching in Assignment Problem” presents this as the “Ultimate Hungarian Algorithm,” i.e., a Hungarian/Kuhn–Munkres style method for maximum-weight bipartite matching. Its core notions are feasible labels

$l(x)+l(y)\ge w(x,y),$

the equality graph

$E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$

augmenting paths, and label updates by minimum slack

$$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$

The paper states the usual polynomial complexity $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$0, contrasted with brute force $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$1, and its optimality proof is the standard feasible-label duality argument for perfect matchings in the equality graph (Chen et al., 2019).

2. From assignment to optimal transport

A second major meaning treats assignment as a special case of mass transport. In this view, Monge’s problem seeks a transport map $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$2 moving a source measure $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$3 to a target measure $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$4, while Kantorovich’s relaxation replaces the deterministic map by a coupling $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$5 or transport plan $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$6 satisfying prescribed marginals: $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$7 The classical assignment problem is then recovered at the extreme points of the transport polytope when source and target supports have equal cardinality and uniform masses (Seyedi et al., 4 Sep 2025).

This reduction is explicit in freeform optical design. In “Designing double freeform surfaces for collimated beam shaping with optimal mass transportation and linear assignment problems,” the source beam cross-section $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$8 with irradiance $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$9 and the target cross-section $$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$0 with irradiance $$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$1 satisfy

$$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$2

and a ray map $$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$3 must preserve flux on every measurable subset. The continuous Monge–Kantorovich problem is discretized by partitioning $$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$4 and $$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$5 into $$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$6 equal-flux cells, turning transport into a linear assignment problem with cost matrix

$$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$7

The optical cost is not quadratic but

$$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$8

which encodes physical realizability of the refractive design. For the main examples, $$\sum_j x[i][j]=1,\qquad \sum_i x[i][j]=1,\qquad x[i][j]\in\{0,1\}.$$9, and the linear assignment problem was solved with an auction algorithm in under 3 minutes on a standard PC (Doskolovich et al., 2018).

The same transport logic appears in dynamic multi-agent control. “Dynamic multi-agent assignment via discrete optimal transport” defines pairwise assignment costs by solving the underlying one-agent/one-target optimal control problem, then uses those values as the ground metric in a discrete OT problem. With equal cardinalities and uniform masses, the Kantorovich LP is tight and returns a permutation matrix; the paper further argues that the optimal dynamics-aware assignment is constant over time, so one assignment computation suffices over the operating lifetime. In the reported simulations, dynamics-aware assignment achieved almost $G=(X\cup Y,E)$0 cost reduction compared with distance-based metrics (Kachar et al., 2019).

3. Generalized, constrained, and large-scale allocation

Beyond one-to-one matching, the term often denotes large discrete allocation models with additional feasibility structure. One such extension is the Assignment Problem with Conflicts (APC), defined on a complete bipartite graph with edge costs $G=(X\cup Y,E)$1 and a conflict set $G=(X\cup Y,E)$2 on pairs of assignment edges. The feasible solution remains a perfect matching, but every conflict pair imposes

$G=(X\cup Y,E)$3

The resulting model is a minimum-cost perfect matching intersected with a stable-set condition in the conflict graph on assignment edges. A 2025 study solved the standard APC MILP directly with Google OR-Tools CP-SAT 9.12 on 130 benchmark instances; $G=(X\cup Y,E)$4 ranged from $G=(X\cup Y,E)$5 to $G=(X\cup Y,E)$6, and the number of conflicts reached values as large as $G=(X\cup Y,E)$7 in the reported tables. The paper states that CP-SAT found all optimal solutions within a 3600-second limit and reported average times of $G=(X\cup Y,E)$8 seconds to the best solution and $G=(X\cup Y,E)$9 seconds to prove optimality (Montemanni et al., 3 Jun 2025).

A structurally different variant arises in hierarchies. “Task Assignment in Tree-Like Hierarchical Structures” defines Maximum Weight Tree Matching (MWTM), where assignable agents are nodes of a rooted tree and assigning a node disables all its ancestors and descendants. The ILP retains one-node-per-task and one-task-per-node constraints but adds, for every root-to-leaf path $w(x,y)$0,

$w(x,y)$1

The paper proves NP-hardness, gives a feasibility characterization—an instance has a feasible solution iff the number of leaves is at least the number of tasks—and proposes the BOA heuristic, an iterative LP-relaxation method that is guaranteed to return a feasible solution whenever one exists (Evrendilek et al., 2014).

Capacitated many-to-many assignment with dynamic state appears in reinforcement learning form in “Reinforcement Learning for Assignment Problem with Time Constraints.” There, a worker may perform multiple tasks until its time budget is exhausted, so the model is closer to a generalized assignment problem than to classical one-to-one assignment. A reconstructed formulation uses

$w(x,y)$2

with task efforts $w(x,y)$3, worker capacities $w(x,y)$4, and eligibility mask $w(x,y)$5. The proposed solver is a PPO-based masked actor–critic that chooses workers sequentially for current tasks, uses reward equal to negative assignment cost, and reuses the trained policy under changed dynamic states without retraining; the paper explicitly does not provide optimality guarantees (Pathan et al., 2021).

Two large-scale formulations emphasize industrial scale. “Approximation Algorithm for Generalized Budgeted Assignment Problems and Applications in Transportation Systems” introduces the Generalized Budgeted Assignment Problem (GBAP), with bin-opening variables $w(x,y)$6, assignment variables $w(x,y)$7, multidimensional capacities $w(x,y)$8, binary consecutive-ones resource vectors $w(x,y)$9, and budget

$\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$0

Its main algorithm is configuration-LP-based randomized rounding with a constant expected approximation guarantee: at least $\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$1 in general, and

$\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$2

when $\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$3 (Jiang et al., 2022). At still larger scale, “A Practical Distributed ADMM Solver for Billion-Scale Generalized Assignment Problems” studies convex relaxations of assignment with objective $\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$4, global owner-level constraints

$\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$5

and item-wise convex constraints $\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$6. Using BADMM to recover separability, the paper reports a $\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$7-variable instance with $\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$8 items and $\max_M \; w(M), \qquad w(M)=\sum_{e\in M} w(e),$9 owners, solved on 2000 CPUs with an average time of about 100 seconds per iteration (Zhou et al., 2022).

4. Multidimensional and multipartite assignment

In another established sense, the mass assignment problem is the multidimensional assignment problem (MAP), also written $M\subseteq E$0-AP. Here one selects $M\subseteq E$1 feasible $M\subseteq E$2-tuples from

$M\subseteq E$3

such that no coordinate value is repeated in any dimension. If $M\subseteq E$4 is a full assignment,

$M\subseteq E$5

and the objective is to minimize $M\subseteq E$6. The case $M\subseteq E$7 is the classical assignment problem; for $M\subseteq E$8, the problem is NP-hard (0906.2960).

The construction heuristics literature focuses on rapid feasible initialization. An empirical study compared Greedy, Max-Regret, ROM, and Shift-ROM. It also proposed memory-access optimizations that produced speedups of roughly $M\subseteq E$9 to $l(x)+l(y)\ge w(x,y),$0 times by favoring successive scans and reducing rescanning of the $l(x)+l(y)\ge w(x,y),$1-sized weight tensor. Across benchmark families, Shift-ROM was the best all-purpose construction heuristic, whereas ROM was the strongest lightweight speed-quality compromise (0906.2960).

Multipartite entity resolution gives a distinct but related MAP formulation. “Multidimensional Assignment Problem for multipartite entity resolution” models a linked entity as an $l(x)+l(y)\ge w(x,y),$2-tuple

$l(x)+l(y)\ge w(x,y),$3

where each source may contribute either a real record or a dummy $l(x)+l(y)\ge w(x,y),$4. Binary variables

$l(x)+l(y)\ge w(x,y),$5

select tuples subject to one-to-one usage of every real record. The cost is derived from a maximum-posterior partition formulation and can be interpreted through pairwise similarity aggregation

$l(x)+l(y)\ge w(x,y),$6

Because MAP is NP-hard for $l(x)+l(y)\ge w(x,y),$7, the paper studies Greedy and very large-scale neighborhood search (VLSN), and finds that VLSN—especially with multi-start—outperforms simple Greedy, while a deterministic design-based multi-start can be more efficient than random-sampling multi-start as database size grows (Kammerdiner et al., 2021).

5. Cybersecurity meaning: mass assignment vulnerabilities

In software security, “mass assignment” has a wholly different meaning. It denotes a vulnerability in which a REST API unintentionally permits modification of protected attributes because incoming request fields are automatically bound to internal objects. The relevant setting is not combinatorial optimization but specification mining and vulnerability analysis.

“Mining REST APIs for Potential Mass Assignment Vulnerabilities” introduces LightMass, a static method that examines an OpenAPI specification and identifies operations prone to this flaw. The central heuristic compares the response attributes of a GET operation with the request attributes of a related POST, PUT, or PATCH operation. Relatedness is inferred via Jaccard similarity: $l(x)+l(y)\ge w(x,y),$8 If the similarity is at least $l(x)+l(y)\ge w(x,y),$9 and the GET response contains more attributes than the write request, then the extra GET-only attributes are flagged as candidate read-only fields that could become mass-assignment targets if the backend accepts them anyway. On 100 APIs, the study identified 25 potentially vulnerable APIs, comprising 115 endpoints, 133 operations, and 495 candidate attributes; in six locally validated open-source APIs, it confirmed nine real vulnerable operations. The paper also reports one documented false positive in VAmPI, where the flagged attribute owner was protected by server-side restrictions (Mazidi et al., 2024).

This cybersecurity usage preserves the allocation metaphor only loosely: the “mass” is not physical or combinatorial flow, but the uncontrolled assignment of request fields to server-side state.

6. Topological and geometric mass assignments

A third mathematical meaning arises in combinatorial geometry and topology, where a mass assignment is a continuous family of measures attached to fibers of a vector bundle. If $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$0 is a Euclidean vector bundle, a mass assignment is a section

$E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$1

so that each base point $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$2 carries a measure $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$3 on the fiber $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$4. This replaces the marginal of a fixed measure by a parameterized measure field (Camarena et al., 2024).

One line of work studies center-transversal theorems in this setting. “A Center Transversal Theorem for mass assignments” considers mass assignments on the flag manifold

$E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$5

whose points are flags $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$6 with $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$7 and $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$8. The main theorem states that if $E_l=\{(x,y)\in E: l(x)+l(y)=w(x,y)\},$9 and

$$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$0

then for every collection of $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$1 mass assignments on $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$2, there exist $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$3 and $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$4 such that every assigned measure on $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$5 has depth at least

$$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$6

The proof is obstruction-theoretic and uses Stiefel–Whitney classes on the flag manifold (Camarena et al., 2024).

A broader bundle-theoretic framework appears in “Many partitions of mass assignments.” There, one studies equipartitions of continuously varying masses or functions on the sphere bundle $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$7 of a Euclidean vector bundle $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$8, by arrangements of hyperplanes in fibers $$\delta_l=\min\{l(u)+l(v)-w(u,v)\mid u\in S,\; v\notin T\}.$$9. The main criterion is cohomological: if

$$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$00

then $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$01, meaning that every $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$02-tuple of continuous functions on $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$03 admits a $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$04-hyperplane equipartition in some fiber. The same framework yields constrained and orthogonal variants, as well as extensions to polynomial partitions, caps, and wedges (Blagojevic et al., 2023).

A closely related Grassmannian formulation extends the Grünbaum–Hadwiger–Ramos problem. “Topology of the Grünbaum–Hadwiger–Ramos problem for mass assignments” defines a mass assignment on $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$05 as a continuous choice $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$06 of a mass on each $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$07-plane $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$08. It proves that for $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$09 such assignments and $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$10, there exist $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$11 and $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$12 affine hyperplanes in $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$13 that equipart all $$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$14 assigned masses whenever

$$\text{Minimize } \sum_{i=1}^n \sum_{j=1}^n C[i][j]\cdot x[i][j]$$15

The proof uses Fadell–Husseini index theory and a parametrized configuration-space/test-map scheme (Blagojević et al., 2022).

7. Conceptual unification and scope

Across these literatures, the expression “Mass Assignment Problem” designates a family of allocation problems rather than a single model. In discrete optimization, it ranges from one-to-one matching to conflict-constrained, capacitated, hierarchical, multidimensional, and billion-scale generalized assignment. In optimal transport, it marks the passage from rigid permutation assignments to transport couplings, Wasserstein distances, and structural matchings such as Gromov–Wasserstein. In security, it names a specification/implementation mismatch in attribute binding. In topology and discrete geometry, it refers to the partition of continuously varying measures over bundles and Grassmannians.

A plausible implication is that the term is best interpreted contextually. When the underlying object is a binary assignment matrix or matching, the relevant theory is combinatorial optimization; when the object is a coupling with prescribed marginals, the relevant theory is optimal transport; when the object is an OpenAPI schema, the relevant theory is program analysis and security testing; and when the object is a section of a measure bundle, the relevant theory is equivariant topology. The shared vocabulary of “assignment” and “mass” is real, but the governing notions of feasibility, optimality, and complexity differ fundamentally across these domains.