LCD: Water Networks & Structural Systems

• Least-cost design (LCD) is an optimization problem configuring systems to minimize cost while meeting hydraulic or structural feasibility constraints.
• In water distribution networks, LCD couples discrete pipe choices with continuous hydraulic variables to enforce acyclic flows and consistent pressure drops.
• In structural systems, LCD strategically selects state variables at varied costs to ensure controllability or observability using graph-theoretic methods.

Least-Cost Design (LCD) denotes a class of optimization problems in which a feasible engineered system is configured to minimize a prescribed cost subject to domain-specific physical or structural constraints. In the arXiv literature represented here, the term appears most directly in water distribution networks (WDNs), where LCD is the problem of selecting pipe segments and flow patterns so that a cyclic network can be built at minimum capital cost while satisfying hydraulic feasibility in a single-source, multi-demand setting (Goyal et al., 2021). A related usage arises in large-scale linear structural systems, where least-cost design concerns minimum-cost input/output placement that guarantees structural controllability or observability (Pequito et al., 2014). The common feature is not a single universal model, but the combination of heterogeneous design costs with exact feasibility conditions that are strongly shaped by network or system structure.

1. Domain-specific meanings of least-cost design

In WDNs, LCD is formulated over an undirected graph $G(N,L)$ with a source $s$, demand values $D_i$, elevations $E_i$, and minimum residual pressure requirements $p_i$. Each link may be built from one or more commercially available pipe segments, and the objective is to minimize total construction cost while enforcing flow conservation and hydraulic feasibility. The decision space therefore couples discrete pipe-segment allocation with continuous hydraulic variables and unknown flow directions (Goyal et al., 2021).

In large-scale linear structural systems, the phrase refers to the selection of directly actuated or measured state variables with heterogeneous state-dependent costs. The design variables are binary placement patterns in an input or output matrix, and feasibility is defined by structural controllability or structural observability rather than by hydraulic equations. The cost is attached to state variables, not to the labels of inputs or outputs (Pequito et al., 2014).

These two research lines share the least-cost principle but differ fundamentally in what constitutes feasibility. In WDNs, feasibility is hydraulic and non-convex. In structural systems, feasibility is graph-theoretic and can be reduced to weighted matching under the assumptions stated in the problem formulation.

2. Formal problem statement in water distribution networks

For WDN LCD, the standard objective is to minimize total construction cost

$C(.)=\sum_{i \in L}\sum_{j \in P} l_{ij}*c_j,$

where $l_{ij}$ is the length of pipe type $j$ placed on link $i$, and $c_j$ is its unit cost (Goyal et al., 2021). The network is physically cyclic, but the actual flow pattern in a feasible water network must be acyclic and thus enforces an acyclic orientation. This is a central structural fact: the optimization must determine not only pipe choices but also a feasible orientation of flows.

Flow conservation is written with a signed incidence matrix $s$0: $s$1 with $s$2 the flow in link $s$3. The nonlinear hydraulic relation is the Hazen–Williams headloss equation. Because each link can carry flow in either direction, and the direction is not known in advance, the model must simultaneously decide flow directions, flow magnitudes, and pipe sizes or segments.

The refinement adopted in the cited work is to encode hydraulic consistency through cycle and path relations rather than by enumerating all possible cycles and source-to-node paths. Pressure drop around every cycle must be zero, but it is sufficient to impose this only on a cycle basis of size $s$4. Node head and residual-pressure constraints are written along a selected spanning-tree path from the source. This sharply reduces the number of cycle and path constraints that must be handled by the solver (Goyal et al., 2021).

3. Non-convexity and solver-oriented formulations

A core difficulty of WDN LCD is that “indeterminism of flow values and flow direction” leads to nonlinear constraints and hence a non-convex problem (Goyal et al., 2021). The physical source of the non-convexity is not merely the hydraulic law itself, but the coupling between unknown signs of flows, pressure constraints, and discrete pipe-segment decisions.

The baseline model is the Discrete Segment (DS) formulation, in which each link may consist of multiple discrete pipe segments: $s$5 This is closer to practical construction than a discrete-pipe model because real links may include multiple pipe diameters. The formulation is solved with generic nonlinear programming rather than with specialized problem-specific heuristics.

Two additional refinements are introduced to handle flow-direction ambiguity more directly. The first is orientation search / re-solving. Since water flows must be acyclic, a feasible orientation is searched for or reused, and the optimization is then re-solved under that fixed direction pattern. The graph-reduction and expansion procedure removes bridges, collapses degree-2 paths, removes self-loops, and collapses parallel edges in order to generate a smaller graph whose orientations can be enumerated or sampled. The second is the Parallel Link (PL) formulation, which replaces each physical link by two nonnegative directional flows $s$6 and $s$7 with the complementarity-style condition

$s$8

Flow conservation becomes

$s$9

and hydraulic terms are rewritten using the signed difference $D_i$0. The stated purpose is numerical: when flow in a link is near zero, the PL representation reduces solver oscillation around ambiguous sign changes (Goyal et al., 2021).

Formulation	Defining feature	Stated role
DS	Multiple discrete pipe segments per link	Baseline practical NLP formulation
Orientation search / re-solving	Fixes a feasible acyclic direction pattern	Reduces search space after graph reduction
PL	Two nonnegative directional variables per link	Improves tractability near zero flow

These formulations exemplify a solver-oriented view of LCD: the principal contribution is not a new meta-heuristic, but a re-expression of the same physical design problem in forms that expose more exploitable structure to generic NLP solvers.

4. Benchmark performance and tractability in WDN LCD

The WDN study explicitly positions itself against LP, NLP, dynamic programming, and a wide range of meta-heuristics such as GA, SA, ACO, PSO, and related hybrids. Its central claim is that recent advances in generic NLP solvers, together with model refinement, can compete with or outperform problem-specific solving techniques (Goyal et al., 2021).

On standard benchmark networks—Two-Loop, Taichung, Hanoi, Double Hanoi, and Triple Hanoi—the DS formulation matches the best least-cost solutions reported in the literature, often exactly matching known minima or being competitive with the best published values. This establishes that the discrete-segment model is LCD-capable without requiring a bespoke search heuristic.

The paper also introduces larger test networks from the HydroGen archive because existing benchmark networks are small in size. On these larger instances, the PL formulation outperforms DS, re-solving, and orientation search in both solution quality and tractability. Orientation search is faster on tiny networks, but it breaks down on larger ones because feasible orientations are too sparse, and random orientation sampling often yields no feasible solutions at all. Re-solving DS can improve average performance and standard deviation, but it does not improve the best objective values consistently. By contrast, PL gives smaller minimum costs on bigger instances, better average performance, and comparable or better runtimes than DS (Goyal et al., 2021).

The broader significance is methodological. The paper reframes least-cost water network design as a formulation problem: the challenge lies in representing cyclic physical topology, acyclic flow realization, and hydraulic feasibility in a way that modern solvers can exploit.

5. Least-cost input/output design in structural systems

A distinct least-cost design problem appears in structural control, where the objective is to choose directly actuated or measured state variables with heterogeneous costs so that the resulting placement guarantees structural controllability or structural observability (Pequito et al., 2014). The setting is a structural LTI system

$D_i$1

where only the zero/nonzero pattern of $D_i$2 is known.

The placement cost is

$D_i$3

with $D_i$4 counting the number of directly actuated state variables. Two optimization problems are studied. In $D_i$5, the goal is minimum cost among minimum-cardinality feasible placements. In $D_i$6, the minimum-cardinality restriction is removed, so a design may actuate more states if that lowers total cost. This distinction is operationally important because a “double-role” state can sometimes be replaced by two cheaper states.

Feasibility is characterized graph-theoretically. A pair $D_i$7 is structurally controllable iff there exists a maximum matching of the state bipartite graph such that the rows of $D_i$8 cover all right-unmatched vertices and at least one state in each non-top linked strongly connected component is directly actuated. The least-cost design problem is then reduced to weighted maximum matching on an augmented bipartite graph with slack vertices. Carefully chosen edge weights encode either the constrained problem $D_i$9 or the unconstrained problem $E_i$0, and both algorithms run in

$E_i$1

The 7-state example with cost vector

$E_i$2

illustrates the difference. For $E_i$3, the minimum-cardinality feasible dedicated placement has total cost $E_i$4. For $E_i$5, allowing an additional actuated state yields the final set $E_i$6 with total cost $E_i$7 (Pequito et al., 2014). The example shows that “least cost” can differ substantially from “least cardinality,” even when both satisfy the same structural controllability requirement.

6. Terminological ambiguity and disciplinary context

A persistent source of confusion is that LCD is heavily overloaded across disciplines. In coding theory, the acronym commonly means linear complementary dual codes, not least-cost design. The paper “Optimal Binary LCD Codes” states this explicitly: “In this paper, LCD means linear complementary dual codes, not ‘least-cost design’ or any other optimization acronym,” and defines the condition

$E_i$8

as the central object of study (Bouyuklieva, 2020).

This ambiguity matters in literature searches and citation practice. In optimization and infrastructure design, LCD can refer to minimum-cost network or placement design problems such as WDN construction or structural input/output placement [(Goyal et al., 2021); (Pequito et al., 2014)]. In coding theory, the same acronym refers to a completely different algebraic notion involving the hull of a code and the nonsingularity of $E_i$9 or related Gram matrices (Bouyuklieva, 2020). The field context therefore determines the meaning.

Within the optimization usage, the available arXiv evidence points to two distinct technical regimes. One regime is non-convex and hydraulically constrained, where tractability depends on formulation choices such as cycle-basis constraints, orientation handling, and parallel-link representations. The other is graph-theoretic and structurally constrained, where exact polynomial-time algorithms arise through weighted matching reductions. The shared label “least-cost design” thus denotes a family resemblance at the level of objective—cost minimization under feasibility—not a single canonical mathematical program.