Multi-Objective Geometric Programming

• MOGP is an optimization framework that extends classical geometric programming by handling multiple, competing posynomial objectives and recovering Pareto-optimal solutions.
• Scalarization methods, including weighted-mean and lexicographic approaches, enable efficient exploration of trade-offs among objectives.
• The dual formulation and convex transformation techniques facilitate practical solution procedures for engineering design problems with continuous cost coefficients.

Multi-objective geometric programming (MOGP) extends the classical geometric programming (GP) paradigm to address problems with multiple, typically competing, objective functions—frequently encountered in engineering design and management scenarios. MOGP handles vector-valued posynomial objectives, subject to posynomial constraints, with solutions aiming to recover Pareto-optimal or non-inferior points that characterize trade-offs among objectives. Canonical approaches include scalarization via weighted mean methods and lexicographic orderings. Recent contributions further generalize the standard model by introducing cost coefficients as continuous functions, enabling explicit modeling of time-varying or parameter-dependent objective terms (Ojha et al., 2010, Ojha et al., 2010, 0912.1832).

1. Mathematical Structure of MOGP

Let $x = (x_1, …, x_n)^\top > 0$ denote the vector of positive decision variables. A multi-objective geometric programming problem is typically defined as:

$$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$

where for each $k = 1,...,p$,

$$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$

with $C_{k0t} > 0$ and $g_k(t) > 0$ a continuous function, allowing model coefficients to vary smoothly (e.g., as explicit functions of time or another system parameter), and $a_{k0t\,j} \in \mathbb{R}$.

Constraints have the standard posynomial form: $$h_i(x) = \sum_{t=1}^{T_i} C_{it} \prod_{j=1}^n x_j^{a_{it\,j}} \leq 1$$ with $C_{it} > 0$ and $a_{it\,j} \in \mathbb{R}$.

This structure supports efficient transformation to convex form under logarithmic change of variables ($$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$0), provided all coefficients are positive. The degree of difficulty (DoD) of a GP instance, crucial in duality theory and solution properties, is defined as: $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$1 (Ojha et al., 2010).

2. Scalarization via Weighted-Mean Methods

The principal technique for extracting Pareto-optimal solutions in MOGP is scalarization—constructing a single, aggregated posynomial objective as a convex combination of all $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$2 objectives. For weights $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$3 with $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$4, define: $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$5 which remains a posynomial for each $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$6. The resulting scalar GP is solved subject to the original constraints. Varying $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$7 over the unit simplex traces out the convex hull of the Pareto front. Each solution corresponds to a noninferior (efficient) design; only the convex front can be attained via linear scalarization; nonconvex regions require alternative schemes (e.g., $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$8-constraint methods) (Ojha et al., 2010, Ojha et al., 2010).

Optionally, each $$\begin{aligned} &\min \ F(x) = (f_1(x), ..., f_p(x)) \ &\text{subject to} \quad h_i(x) \leq 1, \quad i = 1,...,m \ &\qquad\qquad\ x_j > 0,\quad j = 1,...,n \ \end{aligned}$$9 may be normalized to a common scale, but in many engineering applications, the weights themselves encode the required trade-off.

3. Dual Formulation and Solution Procedure

The scalarized single-objective GP (for a fixed $k = 1,...,p$0) is efficiently solved via its Duffin–Peterson–Zener dual. Let $k = 1,...,p$1 denote the total number of monomial terms in the aggregate objective. Associated dual variables are $k = 1,...,p$2 for the objective, and $k = 1,...,p$3 for each constraint term. The dual program is:

$k = 1,...,p$4

$k = 1,...,p$5 is concave; the constraints are affine, yielding a convex optimization problem solvable by interior-point or KKT methods. Once the dual optimum $k = 1,...,p$6 is recovered, primal variables are given by:

$k = 1,...,p$7

subject to normalization satisfying the affine dual constraints (Ojha et al., 2010).

The core algorithm for scalarized MOGP is summarized as:

Step	Description
1	Select grid of weight vectors $k = 1,...,p$8
2	Form aggregated posynomial coefficients for each $k = 1,...,p$9
3	Solve dual program for $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$0 to recover $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$1
4	Recover $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$2 and evaluate $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$3
5	Collect $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$4 as a Pareto point

(Ojha et al., 2010, Ojha et al., 2010).

4. Lexicographic Multi-Objective Geometric Programming

An alternative approach is lexicographic optimization, in which objectives are hierarchically prioritized: $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$5. The lex-minimal solution first minimizes $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$6, then subsequently minimizes $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$7 subject to the minimum value of $$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$8 already achieved, and so forth. The formal structure is:

$$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t)\,C_{k0t} \prod_{j=1}^n x_j^{a_{k0t\,j}}$$9

The dual of each subproblem matches the standard GP dual, with additional constraints for previously fixed objective values. Existence and uniqueness of solutions are guaranteed if the exponent matrix formed by stacking all relevant exponent rows has full rank $C_{k0t} > 0$0. Lexicographic duality theory and a stepwise procedure are developed in (0912.1832).

Key steps:

• Rank objectives by strict priority.
• Sequentially solve each subproblem, each time constraining all higher-priority objectives to their optimum achieved in preceding stages.
• Recover final lexicographically minimal solution $C_{k0t} > 0$1 and $C_{k0t} > 0$2.

Lexicographic GP is suited to settings where strict ranking of objectives is justified. However, globally Pareto-optimal tradeoffs are not necessarily obtained, as later objectives may suffer after higher-priority objectives are fixed (0912.1832).

5. Theoretical Properties, Existence, and Complexity

Fundamental properties of MOGP formulations include:

• Existence of optima: For feasible posynomial constraints, every scalarized GP ($C_{k0t} > 0$3) has at least one global minimizer.
• Solution uniqueness: If DoD (degree of difficulty) is zero, the dual optimum is unique and so is the primal; for DoD $C_{k0t} > 0$4, primal solutions may not be unique, but all deliver identical objective values.
• Pareto front coverage: The weighted-sum scalarization technique yields only the convex hull of the true Pareto set; nonconvex efficient points are unreachable by linear weighting.
• Computational complexity: Each dual problem is a convex (log-concave) maximization under affine constraints and can be solved in polynomial time for fixed numbers of monomial terms and constraints, with the total number of dual variables growing linearly in the number of monomial terms (Ojha et al., 2010, Ojha et al., 2010).

A plausible implication is that for large, high-dimensional MOGPs or those with highly nonconvex objective regions, alternative formulations (e.g., $C_{k0t} > 0$5-constraint, fuzzy methods) or decomposition approaches may be required.

6. Illustrative Examples

Examples in the literature demonstrate the stepwise application of both weighted-mean and lexicographic approaches.

• In (Ojha et al., 2010), a two-objective GP with continuous cost coefficients,

$C_{k0t} > 0$6

with $C_{k0t} > 0$7, $C_{k0t} > 0$8, and a single posynomial constraint, is scalarized and solved. For $C_{k0t} > 0$9, the solution $g_k(t) > 0$0 and $g_k(t) > 0$1 are found as a Pareto point.

• A four-variable, two-objective instance is detailed in (Ojha et al., 2010), where varying the weights across the simplex generates the full (in that case, linear) Pareto front.
• (0912.1832) presents a lexicographic MOGP where the first objective, $g_k(t) > 0$2, is minimized subject to posynomial constraints, then $g_k(t) > 0$3 minimized subject to $g_k(t) > 0$4, and so forth, producing an explicitly lex-minimal solution.

These examples illustrate the generality and computational feasibility of the proposed dual-based procedures for both continuous-cost and standard coefficient regimes.

7. Comparative Discussion and Limitations

Weighted-mean scalarization is efficient, tractable, and compatible with standard GP solvers, with full theoretical guarantees on optimality and convex Pareto-front coverage in convex regions. However, it is fundamentally limited to convex hulls in objective space; for truly nonconvex Pareto sets, some non-inferior solutions cannot be recovered, necessitating alternative methods such as the $g_k(t) > 0$5-constraint or fuzzy programming approaches (Ojha et al., 2010, Ojha et al., 2010).

The lexicographic approach is advantageous in scenarios demanding strict, hierarchical prioritization, guaranteeing best-possible achievement of higher-ranked objectives before improving subordinate ones. Its limitations lie in the arbitrariness of objective ranking for large $g_k(t) > 0$6, potential inefficiency in lower-priority objectives, and growth of problem size with additional constraints at each priority level (0912.1832).

In summary, multi-objective geometric programming, supported by duality and scalarization or prioritization approaches, is a mature and extensible methodology for optimizing design and engineering problems involving posynomial models and vector-valued objectives, with contemporary research extending its applicability through continuous coefficients and computational advances (Ojha et al., 2010, Ojha et al., 2010, 0912.1832).