Two-Stage Stochastic Equilibrium Investment Model

• The paper introduces a two-stage stochastic equilibrium investment model that integrates pre-uncertainty capacity decisions with scenario-specific operational responses under endogenous market feedback.
• It employs formulations like stochastic Nash equilibrium, variational inequalities, and linear complementarity systems to capture investment dynamics under uncertainty.
• The analysis establishes conditions for existence, uniqueness, and computational stability, guiding equilibrium selection and practical implementation in complex markets.

Searching arXiv for the cited papers to ground the article in current records. A two-stage stochastic equilibrium investment model couples a nonanticipative investment or capacity decision with scenario-contingent recourse decisions and closes the system through an equilibrium condition on prices, quantities, or shadow values. In the literal two-stage formulations, the first stage is a here-and-now vector $x$, the second stage is a wait-and-see recourse vector $y(\xi)$, and equilibrium is represented as a stochastic Nash equilibrium, a stochastic variational inequality, or a linear complementarity system (Jiang et al., 2019, Chen et al., 2017). Closely related literatures replace the literal two-stage structure by continuous-time mean-field equilibrium, competitive equilibrium with irreversible investment, or subgame-perfect equilibrium in timing and time-inconsistent control problems, but preserve the same backbone: investment under uncertainty, endogenous aggregate feedback, and an equilibrium consistency requirement (Calvia et al., 2024, Kardaras et al., 3 Dec 2025, Dahl et al., 2019).

1. Conceptual and economic structure

In the formulations synthesized here, the first stage is a pre-uncertainty commitment and the second stage is a scenario-dependent operational response. In the two-stage noncooperative framework, each agent $j$ chooses $x_j$ before uncertainty is revealed and then chooses $y_j(\xi)$ after realization of $\xi$. Strategic interaction enters through market price, which depends on total second-stage quantity. In the production-and-supply application, the inverse demand is

$$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$

and the linking constraint $y_j(\xi)\le x_j$ is the direct analogue of an operational output constraint under installed capacity (Jiang et al., 2019).

The same architecture appears in the duopoly capacity game formulated as a two-stage stochastic linear complementarity problem. There, stage 1 is capacity $x_i$, stage 2 is scenario-wise output $y_i(\xi)$, and the capacity shadow value $y(\xi)$0 links current investment incentives to future equilibrium scarcity. The first-stage condition

$y(\xi)$1

has the standard investment interpretation: capacity is chosen until expected marginal scarcity rent equals marginal capital cost (Chen et al., 2017).

This structure distinguishes the topic from a purely centralized stochastic program. The defining feature is not merely recourse, but recourse embedded in decentralized equilibrium: each agent optimizes against endogenous market feedback, and the aggregate outcome must be internally consistent with those optimal responses.

2. Canonical mathematical formulations

A generic two-stage stochastic equilibrium problem can be written as a two-stage stochastic Nash system. For each agent $y(\xi)$2,

$y(\xi)$3

The induced first-stage problem can be expressed through the recourse value function

$y(\xi)$4

so that the first-stage decision internalizes expected equilibrium recourse value (Jiang et al., 2019).

For Cournot-Nash settings under uncertainty, the equilibrium can be written as a two-stage stochastic complementarity system. In the formulation with first-stage commitments $y(\xi)$5, second-stage outputs $y(\xi)$6, and multipliers $y(\xi)$7,

$y(\xi)$8

Here $y(\xi)$9 is the multiplier on $j$0, and therefore the scenario-wise scarcity value of installed capacity (Jiang et al., 2019).

A more general two-stage stochastic linear complementarity problem is

$j$1

This form is especially important because it accommodates equilibrium recourse directly, rather than deriving it only as an optimization subproblem. In the duopoly investment application, the stage-2 Nash-Cournot game is embedded in the second-stage complementarity block, and its expectation feeds the first-stage investment condition (Chen et al., 2017).

These formulations make precise what “equilibrium investment” means in the two-stage setting: stage-1 commitments and stage-2 operations are jointly determined by KKT conditions, complementarity, or variational inequalities, rather than by a single planner’s objective.

3. Existence, uniqueness, and equilibrium selection

The solvability theory of two-stage stochastic equilibrium investment models is driven by monotonicity and recourse regularity. In the two-stage SLCP framework, the central structural condition is the block positive-definiteness assumption

$j$2

for a positive continuous $j$3 with $j$4. Under this assumption, the two-stage SLCP has a unique solution $j$5; the second-stage solution map $j$6 is globally Lipschitz in $j$7; the reduced first-stage mapping is Lipschitz; and all matrices in the Clarke generalized Jacobian of the reduced map are nonsingular and uniformly inverse-bounded (Chen et al., 2017).

The two-stage SVI analysis for Cournot-Nash equilibrium under uncertainty establishes sufficient conditions for existence of solutions and proves relatively complete recourse. For fixed $j$8, the second-stage equilibrium is unique by Rosen’s theorem for concave $j$9-person games. However, the multiplier $x_j$0 associated with $x_j$1 may fail to be unique even when $x_j$2 is unique. This nonuniqueness is economically meaningful, because $x_j$3 enters the first-stage condition and therefore affects equilibrium investment incentives (Jiang et al., 2019).

To select equilibria and restore numerical stability, the multiplier block can be regularized: $x_j$4 The regularized system is strongly monotone and admits a unique solution $x_j$5. As $x_j$6, any accumulation point is a solution of the original problem, and for fixed $x_j$7 the regularization selects the unique least $x_j$8-norm solution of the unregularized second-stage LCP (Jiang et al., 2019).

A recurring implication is that uniqueness of the stage-2 primal equilibrium is not, by itself, sufficient for clean stage-1 comparative statics. The equilibrium object relevant for investment is the entire first-stage mapping, including expected shadow values.

4. Continuous-time and mean-field generalizations

Several papers do not formulate a literal two-stage stochastic program, but they retain the same equilibrium backbone in continuous time. In the stochastic mean-field game of optimal investment, the representative firm controls production capacity $x_j$9 via

$y_j(\xi)$0

with operating profit $y_j(\xi)$1, quadratic investment cost $y_j(\xi)$2, and equilibrium condition

$y_j(\xi)$3

For a given average-capacity path $y_j(\xi)$4, the unique optimal control is deterministic,

$y_j(\xi)$5

The equilibrium reduces to a nonlinear integral equation in expected capacity. Existence and uniqueness hold on finite horizons among equilibria with $y_j(\xi)$6; on the infinite horizon they hold under the restriction $y_j(\xi)$7; and the infinite-horizon model admits the unique constant solution

$y_j(\xi)$8

Because interaction occurs only through expected capacity, $y_j(\xi)$9 drops out of the aggregate equilibrium characterization (Calvia et al., 2024).

A distinct continuous-time generalization is the competitive equilibrium model of irreversible capacity investment with stochastic demand and heterogeneous producers. Producers are price takers, choose output $\xi$0 and irreversible capability expansion $\xi$1, and face isoelastic demand

$\xi$2

After optimizing output pointwise,

$\xi$3

and market clearing implies

$\xi$4

The equilibrium price is therefore a nonlinear functional of the transformed demand state $\xi$5 and the endogenous aggregate capability distribution. The producer problem becomes a three-dimensional singular stochastic control problem, and the paper derives an explicit HJB solution together with a closed-form free-boundary surface separating investment and waiting regions (Kardaras et al., 3 Dec 2025).

These models are continuous-time, not literal two-stage programs. This suggests that the two-stage stochastic equilibrium investment model is best understood not as a single formalism, but as a family of equilibrium structures in which investment under uncertainty is linked to endogenous price or aggregate-state feedback.

5. Strategic timing and alternative equilibrium notions

The topic also includes models in which the two-stage structure is generated by timing, information revelation, or dynamic inconsistency rather than by an explicit recourse matrix. In the duopoly preemption game with two alternative stochastic investment choices, stage 1 is the strategic waiting problem before any firm invests, and stage 2 begins once a leader invests and the follower observes the realized profitability of the leader’s project. The follower then chooses whether to copy the leader’s technology at reduced cost or innovate independently at full cost. The resulting stochastic stopping game exhibits two disjoint preemption regions and, for a specific choice of parameters, no pure symmetric subgame perfect Nash equilibrium; an asymmetric equilibrium is characterized instead (Dahl et al., 2019).

A different extension arises in continuous-time portfolio choice with dynamic preference uncertainty. The control is the risky-asset share $\xi$6, the terminal preference state $\xi$7 plays the role of the second-stage scenario, and the objective is

$\xi$8

The inner object is a scenario-wise conditional expected utility, transformed by a scenario-specific certainty-equivalent map and then aggregated across future preference states. Because this nested structure is nonlinear, the dynamic programming principle fails, and the model is solved through a subgame-perfect equilibrium notion: $\xi$9 In the CRRA specialization with arithmetic Brownian preference factor, the equilibrium policy decomposes into a myopic demand term and a preference-hedging demand term (Aquino et al., 24 Dec 2025).

These papers clarify an important boundary condition. “Equilibrium” in the broader investment literature need not mean only market clearing among firms. It may instead denote a stopping-game equilibrium among strategic investors or an intra-personal subgame-perfect equilibrium among successive selves. What remains common is that future contingencies are valued through a scenario-dependent continuation problem, and current investment is chosen in anticipation of that continuation structure.

6. Discretization, robustness, computation, and domain boundaries

When uncertainty is continuously distributed, two-stage stochastic equilibrium investment models require an approximation layer. In the SLCP framework, the support $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$0 is partitioned into cells $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$1, conditional averages are formed within each cell, and the resulting discretized problem is a finite-dimensional deterministic LCP: $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$2 Under Lipschitz continuity of $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$3, the approximation error is first order in the maximal cell diameter, and Voronoi partitions are proposed for higher-dimensional uncertainty. The discretized SLCP may then be solved by the progressive hedging method, which relaxes nonanticipativity, solves scenario blocks in parallel, and restores first-stage consensus by averaging (Chen et al., 2017).

The two-stage SVI literature uses a different approximation route: regularized sample average approximation. With iid samples $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$4,

$$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$5

so the stochastic equilibrium is replaced by a deterministic large-scale LCP. The paper proves strong consistency of the first-stage regularized SAA solution for fixed $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$6, and then shows that the combined limits $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$7 and $$p\big(\xi,T(y(\xi))\big)=p_0(\xi)-\gamma(\xi)T(y(\xi)), \qquad T(y(\xi))=\sum_{j=1}^J y_j(\xi),$$8 recover the original solution set. Computation is organized through Progressive Hedging Method decomposition and scenario-wise smoothing Newton solves (Jiang et al., 2019).

The same section of the literature also extends the known-distribution model to a distributionally robust setting. The ambiguity set is defined by moment constraints, and the distributionally robust first-stage equilibrium becomes a semi-infinite complementarity system after dualization. This is an ex post robust equilibrium notion: stage-1 investment must be viable for every distribution in the ambiguity set, while stage-2 equilibrium remains scenario-wise (Chen et al., 2017).

A common misconception is to equate any adaptive two-stage stochastic investment model with a two-stage stochastic equilibrium investment model. The adaptive grid-hardening framework with undergrounding and vegetation management is explicitly a centralized two-stage stochastic optimization model, not an equilibrium model: it has no market-clearing equilibrium, no Nash game among utilities, no complementarity system, and no decentralized strategic interaction. Its value for the equilibrium literature lies instead in its clear separation of long-lived and short-lived resilience actions, its scenario-tree representation of multi-hazard uncertainty, and its inter-technology coupling constraints (Chowdhury et al., 29 Jan 2026).

In this literature, the most stable definition is therefore structural rather than notational. A two-stage stochastic equilibrium investment model is one in which pre-uncertainty investment or capacity choice is linked to post-uncertainty recourse behavior through equilibrium conditions, with current decisions disciplined by expected or robustified future scarcity values, prices, or continuation payoffs. Literal two-stage SVI and SLCP models provide the clearest mathematical realization of that idea, while continuous-time mean-field, competitive-equilibrium, timing-game, and time-inconsistent-control models show how the same logic persists once the stage structure is generalized.