Facility Location Function Overview

• The facility location function is a mathematical objective used to optimize facility placement by balancing opening and connection costs.
• It encompasses classical models with fixed costs, submodular cost structures, and choice-based stochastic frameworks, enabling scalable and reliable optimization.
• It informs applications in logistics, network design, and resource allocation, utilizing methods like sketching, LP relaxations, and convex programming.

A facility location function defines the objective in facility location problems, which aim to select sites for facilities to optimize specific system-wide criteria, typically involving the costs of opening facilities and serving clients, and often incorporating stochasticity, agent preferences, or submodular cost structures. The facility location function is central to a broad class of operations research, mechanism design, and combinatorial optimization problems, including both classical formulations and modern generalizations with submodular, randomized, or game-theoretic objectives.

1. Classical and Graph-Based Facility Location Objectives

A canonical facility location problem specifies a set of clients $C$ and candidate facilities $F$, with the assignment of each client to a facility incurring a connection (service) cost, and an opening cost associated with each facility. In the classical (uncapacitated) model, the facility location function is

$$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$

where $f_i$ is the opening cost of facility $i$, and $d(i,j)$ is the cost (possibly the shortest-path distance in a graph) between facility $i$ and client $j$ (Garimella et al., 2015). This objective seeks a subset $S$ of open facilities minimizing the sum of opening and connection costs, with each client assigned to the nearest open facility.

In the graph formulation, $G=(V,E,w)$ provides the edge-weighted network, enabling the use of shortest-path distances between facilities and clients. Algorithms for massive networks employ sketching, parallel computation, and maximal independent set computations to obtain scalable (approximate) solutions with proven guarantees (Garimella et al., 2015).

2. Submodular Facility Location and Generalized Cost Structures

The submodular facility location (SFL) framework extends the classical cost model by introducing a facility opening cost that is a monotone submodular set function $F$0 of the assigned clients: $F$1 where $F$2 is the set of clients assigned to facility $F$3, and $F$4 models economies of scale, risk pooling, or stochastic activation via submodularity (i.e., $F$5) (Abbasi et al., 2022).

The SFL captures the standard facility location function as the special case $F$6 for some constant $F$7, as well as more general scenarios with concave-in-size costs, stochastic client activations, or hybrid additive/multiplicative opening costs. Submodularity ensures that greedy and LP-based approximations achieve strong theoretical guarantees, with $F$8-approximation algorithms established for the general SFL model (Abbasi et al., 2022).

3. Stochastic and Choice-Based Facility Location Functions

Recent models accommodate stochastic or agent-based decision processes where facility usage and market share are defined via random utility maximization (RUM) or logit models. For a firm selecting facilities $F$9 (with $$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$0 existing sites), the expected market share is

$$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$1

where $$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$2 are random utilities depending on observable and unobservable attributes (Legault et al., 2022). Sample-average approximations and deterministic equivalent reformulations yield a weighted maximum coverage objective

$$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$3

which is provably submodular and directly tractable for Benders-type decomposition via submodular inequalities. The submodular structure ensures nondecreasing returns and supports efficient optimization (Legault et al., 2022).

4. Mechanism Design and Welfare Objectives

The facility location function can also encode social welfare, fairness, and individual incentives. When agent locations are private inputs, the Nash social welfare (NSW) objective

$$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$4

balances efficiency and fairness in placement decisions (Lam et al., 2023). Maximizing Nash welfare is single-peaked, location-invariant, and lacks closed-form solutions for $$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$5 (roots of degree-$$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$6 polynomials). Polynomial-time additive $$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$7-approximation algorithms exist via convex optimization (Lam et al., 2023).

Strategy-proofness is not compatible with constant-factor approximations to Nash welfare: any deterministic, strategy-proof mechanism has exponential loss in NSW. The MidOrNearest rule (a “phantom median” with $$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$8 phantoms at $$\min_{S\subseteq F}\;\Bigl(\sum_{i\in S}f_i\;+\;\sum_{j\in C}\min_{i\in S}d(i,j)\Bigr)$$9) is strategyproof, anonymous, Pareto-optimal, and achieves tight approximation bounds for utilitarian and egalitarian objectives, but only exponential approximation for Nash welfare (Lam et al., 2023).

5. Distributed Mechanisms and Distortion Bounds

Distributed facility location models partition agents into districts and sequentially aggregate information, introducing an inherent loss quantified as distortion. The worst-case ratio of distributed mechanism cost to the centralized optimum depends on both the aggregation method and the set of allowed facility locations:

• With discrete facility sets and non-strategyproof mechanisms (e.g., minimum-district-sum), tight distortion bound is $f_i$0;
• For strategyproofness or ordinal mechanisms, the lower and upper bound is $f_i$1 (attained by the district-median rule);
• In the continuous model (facility can be anywhere), continuous median-of-medians achieves strategyproofness and distortion $f_i$2 (Filos-Ratsikas et al., 2020).

These results formalize the price of decentralization and the trade-offs among optimality, incentive-compatibility, and informational structure in facility location (Filos-Ratsikas et al., 2020).

6. Optimization Methods and Relaxations

Multiple algorithmic paradigms support the efficient optimization of facility location functions:

• Large-scale combinatorial and graph-based settings use all-distances sketching, shared-nothing parallel computation, and randomized maximal independent set selection to achieve $f_i$3-approximation and linear scaling (Garimella et al., 2015).
• Submodular cost models benefit from configuration-LP relaxations with variables for facility–client set assignments and the Lovász extension for convexification (Abbasi et al., 2022).
• Choice-based objectives under RUM/logit models leverage sample-average approximation, submodular function maximization, and partial Benders decomposition modulated by entropy measures of preference profile distributions (Legault et al., 2022).
• Convex–programming (e.g., ellipsoid or interior-point) is used for Nash welfare maximization on line intervals (Lam et al., 2023).

These methodologies are tailored to the structural properties of the facility location function, particularly submodularity, convexity, or the separability of the cost/objective function across decomposition schemes.

7. Generalizations and Contextual Implications

The facility location function unifies classical and modern paradigms in combinatorial and stochastic optimization. Submodularity, monotonicity, and normalization enable inclusion of diverse operational settings such as risk-averse location, stochastic/universal covering, and assortment optimization. Theoretical developments on approximation bounds, mechanism design constraints, and scalable algorithms provide a rigorous foundation for application domains involving logistics, resource allocation, market competition, and network design (Abbasi et al., 2022, Lam et al., 2023, Legault et al., 2022, Garimella et al., 2015, Filos-Ratsikas et al., 2020).

A plausible implication is that the evolution of facility location functions continues to be driven by both structural generality—submodular and utility-based modeling—and by computational tractability, with modern methods exploiting problem-specific relaxations, stochasticity, and parallelism.