Mathematical Program Networks

• Mathematical Program Networks (MPNs) are a framework for modeling interconnected optimization problems using directed graphs that encode dependency structures and optimality conditions.
• They generalize classical formulations such as Nash equilibria, bilevel programming, and MPECs by integrating node-specific solution graphs within a unified structure.
• Acyclic Quadratic Program Networks and extensions to dynamic games demonstrate MPNs' practical value in ensuring tractable computations and handling asymmetric information.

Searching arXiv for papers on Mathematical Program Networks and closely related formulations. Mathematical Program Networks (MPNs) are a framework for representing a collection of interdependent Mathematical Programs (MPs) that are to be solved simultaneously while respecting a directed network that defines their relationships. In this formulation, the network is not merely an indexing device: it determines which decision variables each constituent program can influence directly, and which additional constraints arise from the optimal decisions of descendant nodes. The framework was introduced as a common language for problems that are otherwise presented as Nash equilibrium problems, multilevel optimization problems, and Equilibrium Programs with Equilibrium Constraints (EPECs), among others (Laine, 2024). Subsequent work has used the same formalism to encode dynamic games with arbitrary interleaved information structures and to model sequential platform–driver–passenger interactions in the gig economy (K et al., 19 Mar 2026, Koirala et al., 2024).

1. Core formalism

An MP is defined as a tuple $(f, C, J)$, where $f : \mathbb{R}^n \to \mathbb{R}$ is the objective, $C \subset \mathbb{R}^n$ is the feasible set, and $J \subset [n]$ indexes the endogenous decision variables controlled privately by that program. An MPN is then defined as a directed graph of such MPs. In the foundational formulation, an MPN is the tuple

$(\{f^i, C^i, J^i\}_{i \in [N]}, E),$

where $E \subset [N]\times[N]$ is the directed edge set and $\mathbf{x}\in\mathbb{R}^n$ is the shared decision vector. Node $i$ directly controls the subvector $\mathbf{x}^i := [x_j]_{j\in J^i}$ (Laine, 2024).

The edge set determines the dependency structure. If $(i,j)\in E$, then node $f : \mathbb{R}^n \to \mathbb{R}$0 is a child of node $f : \mathbb{R}^n \to \mathbb{R}$1. In the sequential-game interpretation, this means that node $f : \mathbb{R}^n \to \mathbb{R}$2 moves first and node $f : \mathbb{R}^n \to \mathbb{R}$3 responds rationally; if two nodes are not connected by directed paths, they are interpreted as moving simultaneously (Koirala et al., 2024). This suggests that MPNs generalize multilevel programming by replacing a fixed stack of levels with an arbitrary directed precedence structure.

To capture indirect dependence, the framework introduces the reachable transition relation

$f : \mathbb{R}^n \to \mathbb{R}$4

with $f : \mathbb{R}^n \to \mathbb{R}$5 iff there is a path from node $f : \mathbb{R}^n \to \mathbb{R}$6 to node $f : \mathbb{R}^n \to \mathbb{R}$7. From this, one defines

$f : \mathbb{R}^n \to \mathbb{R}$8

Thus $f : \mathbb{R}^n \to \mathbb{R}$9 contains node $C \subset \mathbb{R}^n$0 and all of its descendants, and $C \subset \mathbb{R}^n$1 collects the private variables of that set (Laine, 2024).

2. Solution graphs and equilibrium semantics

The central semantic object is the solution graph $C \subset \mathbb{R}^n$2 of node $C \subset \mathbb{R}^n$3. It is the set of global decisions $C \subset \mathbb{R}^n$4 for which node $C \subset \mathbb{R}^n$5's controlled variables are locally optimal, with the variables outside $C \subset \mathbb{R}^n$6 frozen, and with all child solution-graph constraints enforced. Formally,

$C \subset \mathbb{R}^n$7

The $C \subset \mathbb{R}^n$8 is taken in a local sense, which is important because the induced feasible regions can be nonconvex or disconnected (Laine, 2024).

An equilibrium of the MPN is a point that lies in every node’s solution graph: $C \subset \mathbb{R}^n$9 This provides a single equilibrium notion across networked optimization, hierarchy, and equilibrium-constrained formulations. Because a solution graph is always a subset of the corresponding feasible set, the framework also yields the inclusion $J \subset [n]$0 for every edge $J \subset [n]$1, and hence for every reachable pair. When a set of source nodes reaches all others, the equilibrium set is already determined by intersecting the solution graphs of the sources (Laine, 2024).

In later applications to dynamic games, the same equilibrium idea is retained. There, a node’s solution graph consists of all global decision vectors such that the subvector corresponding to its dependency set solves that node’s optimization problem given the decisions of nodes outside the dependency set and given that all child-node solution-graph conditions are satisfied. The equilibrium again has the form

$J \subset [n]$2

This continuity of semantics across domains is one of the defining features of the framework (K et al., 19 Mar 2026).

3. Relationship to classical problem classes

A principal motivation for MPNs is that they subsume several established formulations within one graph-based object. If the network has no edges, each node solves its own optimization problem independently, and equilibrium reduces to a simultaneous optimality condition of Nash or generalized Nash type. If the network is hierarchical, then upper-level problems include lower-level solution graphs as constraints, reproducing bilevel and multilevel optimization. If several nodes are constrained by equilibrium behavior of other nodes, the result is an MPN representation of EPECs; the foundational paper also explicitly lists Mathematical Programs with Equilibrium Constraints (MPECs), Multi-Leader-Multi-Follower games, and Feedback Nash equilibrium problems among the covered classes (Laine, 2024).

The sequential-game interpretation in the ridesharing application makes the same point in a different way. That paper defines an MPN as $J \subset [n]$3, emphasizes that MPNs generalize multilevel programming, and uses a four-node network to encode a game in which two platforms move simultaneously, drivers respond to both platforms, and passengers respond after drivers. The directed edges

$J \subset [n]$4

encode the order of moves and the conditional dependence of lower-level decisions on higher-level choices (Koirala et al., 2024).

A concise way to summarize the reduction pattern is the following.

Formulation	MPN network pattern	Equilibrium interpretation
Nash / generalized Nash	No edges	Simultaneous node-wise optimality
Bilevel / multilevel	Hierarchical directed graph	Upper nodes constrained by descendant solution graphs
EPEC / MPEC-related formulations	Coupled network of dependent programs	Simultaneous consistency of all node solution graphs

This suggests that the distinctive contribution of MPNs is not a new equilibrium concept in isolation, but a common graph semantics for problems that were previously described through separate formalisms.

4. Acyclicity, Quadratic Program Networks, and algorithms

The main computational development in the foundational paper is restricted to acyclic MPNs. Cyclic MPNs are treated as degenerate because the recursive definition of solution graphs becomes circular; the paper gives a corollary showing that in a fully connected cyclic subnetwork, singleton sets can serve as valid solution graphs in a trivial sense, so the concept loses meaningful content. For that reason, the substantive theory focuses on acyclic networks (Laine, 2024).

Within the acyclic setting, the paper develops the subclass of Quadratic Program Networks (QPNs). A QPN is an MPN in which each $J \subset [n]$5 is quadratic and convex with respect to $J \subset [n]$6, and each feasible region $J \subset [n]$7 is a convex polyhedral region. This yields a tractable structure: the solution graph of a parametric convex quadratic program is a union of polyhedra indexed by active-constraint sets. The paper states

$J \subset [n]$8

and then extends this polyhedral-union structure to entire acyclic QPNs (Laine, 2024).

A further technical step is the local representation of solution graphs. Near a given point, only the polyhedral pieces intersecting a small neighborhood matter, which allows local equilibrium verification and local graph construction. On that basis, the paper develops several algorithms: Check QP Solution, Generate Local Solution Graph for QP, Generate Local Solution Graph for a QPN Node, Compute Nash equilibrium among QPs via a large linear mixed complementarity problem, and QPN Equilibrium Search, which proceeds layer by layer in a topological depth ordering $J \subset [n]$9 (Laine, 2024).

The same paper is also explicit about current limitations. The present computational development is strongest for acyclic QPNs, and open difficulties include nonconvex node objectives, numerical conditioning, explosion in the number of local polyhedral components, and cyclic networks. A plausible implication is that the general MPN framework is broader than the currently mature solution technology.

5. Dynamic games and interleaved information structures

A major extension of the MPN framework models deterministic discrete-time dynamic games with arbitrary interleaved information structures. In that formulation, each agent-time pair $(\{f^i, C^i, J^i\}_{i \in [N]}, E),$0 is identified with an MPN node, and the network directly encodes informational dependence among optimization problems. The paper treats open-loop and feedback information structures as canonical extremes and then generalizes to arbitrary asymmetric observation patterns (K et al., 19 Mar 2026).

The game dynamics are

$(\{f^i, C^i, J^i\}_{i \in [N]}, E),$1

and agent $(\{f^i, C^i, J^i\}_{i \in [N]}, E),$2 minimizes

$(\{f^i, C^i, J^i\}_{i \in [N]}, E),$3

The main modeling move is that the information structure determines the edge set of the MPN. Temporal edges always connect $(\{f^i, C^i, J^i\}_{i \in [N]}, E),$4, while observation edges connect $(\{f^i, C^i, J^i\}_{i \in [N]}, E),$5 whenever agent $(\{f^i, C^i, J^i\}_{i \in [N]}, E),$6 observes agent $(\{f^i, C^i, J^i\}_{i \in [N]}, E),$7 at time $(\{f^i, C^i, J^i\}_{i \in [N]}, E),$8. The paper states that the graph is not just a computational device; it encodes who reasons about whom (K et al., 19 Mar 2026).

In open-loop play, each agent observes only the initial state and chooses controls of the form

$(\{f^i, C^i, J^i\}_{i \in [N]}, E),$9

The induced MPN has $E \subset [N]\times[N]$0 nodes per agent, temporal edges only, and no cross-agent informational edges. In feedback play, each agent observes the current state,

$E \subset [N]\times[N]$1

and the Bellman-style recursion introduces cross-agent dependence through feedback consistency constraints. General interleaved information structures interpolate between these two cases by superposing temporal and observation edges (K et al., 19 Mar 2026).

For linear-quadratic games,

$E \subset [N]\times[N]$2

with quadratic stage costs and conditions such as $E \subset [N]\times[N]$3 and $E \subset [N]\times[N]$4, the paper uses the MPN formulation to derive Riccati-like equations from KKT conditions. The procedure is: build the MPN from the information structure, write each node’s solution-graph optimization, form the Lagrangian, write KKT stationarity and primal feasibility, and eliminate multipliers backward in time. Because the constraints are affine and the costs are convex quadratic, the KKT conditions are not only necessary but also sufficient. The result is a coupled Riccati-like system whose coupling pattern mirrors the MPN graph (K et al., 19 Mar 2026).

The illustrative example is a three-agent, three-step LQ game with a cyclic information structure in which agent 1 observes agent 2, agent 2 observes agent 3, and agent 3 observes agent 1. The informational graph includes time-indexed edges such as

$E \subset [N]\times[N]$5

and similarly at later times. The example demonstrates that the framework can handle nontrivial asymmetric information, cyclic observation dependencies, and multi-agent, multi-time coupling beyond the open-loop and feedback extremes (K et al., 19 Mar 2026).

6. Applications, analogies, and common misunderstandings

One substantive application uses MPNs to model a ridesharing duopoly with two platforms, drivers, passengers, and an outside option interpreted as public transit. The MPN has four nodes $E \subset [N]\times[N]$6, $E \subset [N]\times[N]$7, $E \subset [N]\times[N]$8, and $E \subset [N]\times[N]$9, with edges $\mathbf{x}\in\mathbb{R}^n$0, $\mathbf{x}\in\mathbb{R}^n$1, and $\mathbf{x}\in\mathbb{R}^n$2. Platform nodes choose passenger prices and driver commissions, drivers choose allocations $\mathbf{x}\in\mathbb{R}^n$3, and passengers minimize a generalized cost that includes fares and waiting costs. The paper emphasizes that the model captures both cross-side externalities and same-side congestion, and that it supports both single-homing and multi-homing interpretations (Koirala et al., 2024).

Within that stylized model, the paper derives a strong equilibrium conclusion: a profitable non-trivial duopoly can exist only if the platforms engage in some form of tacit collusion, and the stable collusive outcome is

$\mathbf{x}\in\mathbb{R}^n$4

so drivers are paid only enough to cover gas cost. The same paper is careful to note a limitation: tacit collusion is inferred from the stylized equilibrium logic and is not empirically demonstrated. This is an instructive example of the distinction between what an MPN-based model establishes formally and what would require external empirical confirmation (Koirala et al., 2024).

A different but related line of work, “Multi-Model Probabilistic Programming,” is presented as a network-of-models framework for modular Stan programs. Its terminology is different, but the abstraction is described as highly analogous in spirit to a Mathematical Program Network: nodes are concretized probabilistic models, edges represent one-module changes, and the framework supports graph construction, neighborhood generation, traversal, and search over related programs. This suggests that the MPN viewpoint has a broader conceptual reach than optimization alone, even when the host formalism is domain-specific and not named an MPN in the source paper (Bernstein, 2022).

A recurrent misunderstanding concerns the acronym itself. In graph representation learning, “MPN” often refers informally to message-passing architectures, but “On the expressive power of message-passing neural networks as global feature map transformers” explicitly studies Message-Passing Neural Networks (MPNNs), not Mathematical Program Networks (Geerts et al., 2022). The two usages are unrelated: the former concerns interdependent optimization problems and equilibrium semantics, whereas the latter concerns graph neural architectures and global feature map transformers.

Overall, MPNs are best understood as a graph-theoretic semantics for coupled optimization problems. Their distinctive contribution lies in making dependence structure first-class: the edge set specifies precedence, informational dependence, and descendant optimality constraints, while equilibrium is defined uniformly as membership in the intersection of node-level solution graphs. The framework is already broad enough to capture Nash-type interaction, hierarchical control, equilibrium-constrained optimization, and dynamic games with asymmetric information, but the strongest computational theory currently remains concentrated in acyclic quadratic settings (Laine, 2024).