Multi-Objective Global Descent Function

• Multi-objective global descent functions are auxiliary constructions that generate a common descent direction for all objectives without relying on predefined scalarization weights.
• They detect first-order Pareto criticality by certifying the absence of a descent direction and are formulated using convex QP, LP, or worst-case quadratic models.
• Variants like incremental, stochastic, and fractional approaches, as well as filled-function methods, improve computational efficiency and help escape local weakly efficient basins.

A multi-objective global descent function is an auxiliary construction for multi-objective optimization that is designed to produce a single update mechanism with simultaneous descent across several objectives, or to certify first-order Pareto criticality when such a direction does not exist. In the cited literature, the term is used in several related ways: as a direction-generating map obtained from a convex QP or LP, as a worst-case quadratic model such as $Q(x;d)=\max_j Q_j(x;d)$, and as a vector-valued filled function used to leave a local weakly efficient basin in non-convex problems. Across these variants, the defining role is the same: avoid predefined scalarization weights or ordering information, and organize descent directly in the multi-objective geometry (Xu et al., 2024, Ansary, 2022, Adhikary et al., 30 Jul 2025).

1. Problem setting and criticality notions

The basic setting is the multi-objective problem

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$

or, in the smooth unconstrained case,

$$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$

Several papers treat composite structure, writing each component as a smooth part plus a nonsmooth part, such as $F_j(x)=f_j(x)+g_j(x)$ or $F_j(x)=s_j(x)+r_j(x)$. The order relation is componentwise: $f(y)\le f(x)$ means $f_i(y)\le f_i(x)$ for all $i$ (Ansary, 2022, Yang, 2024, Ansary, 2024).

First-order criticality is expressed in closely related forms. A point is Pareto critical if there is no direction $d$ such that $\langle \nabla f_i(x),d\rangle<0$ for all objectives. In composite problems, a critical point is equivalently described by the weak Pareto condition

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$0

In smooth non-convex settings, weak Pareto criticality is also written through multipliers: $\min_x F(x)=(F_1(x),\dots,F_m(x)),$1 A recurrent structural fact is that the descent construction itself detects criticality: in different formulations, $\min_x F(x)=(F_1(x),\dots,F_m(x)),$2, $\min_x F(x)=(F_1(x),\dots,F_m(x)),$3, or $\min_x F(x)=(F_1(x),\dots,F_m(x)),$4 is equivalent to first-order Pareto criticality (Ansary, 2022, Yang, 2024, Adhikary et al., 31 Mar 2026).

2. Direction-generating formulations in differentiable optimization

For smooth unconstrained problems, a multi-objective global descent function is often a map $\min_x F(x)=(F_1(x),\dots,F_m(x)),$5 or $\min_x F(x)=(F_1(x),\dots,F_m(x)),$6 defined as the solution of a local optimization problem. These constructions differ in geometry and computational cost, but they share the objective of finding a common descent direction for all objectives.

Formulation	Defining subproblem	Reported property
Global descent direction $\min_x F(x)=(F_1(x),\dots,F_m(x)),$7	$\min_x F(x)=(F_1(x),\dots,F_m(x)),$8	If $\min_x F(x)=(F_1(x),\dots,F_m(x)),$9 is not Pareto-stationary, then $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$0 for all $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$1 (Xu et al., 2024)
Central descent direction $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$2	$$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$3 s.t. $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$4	Invariant under monotone re-scaling $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$5; admits simultaneous descent for sufficiently small $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$6 under $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$7-Lipschitz gradients (Oliveira et al., 2021)
LP-based direction $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$8	LP with normalized-gradient constraints $$\min_{x\in\mathbb R^n} f(x)=(f_1(x),\dots,f_m(x))^\top.$$9, box bounds on $F_j(x)=f_j(x)+g_j(x)$0, and $F_j(x)=f_j(x)+g_j(x)$1	Returns a shared strict descent direction when one exists, otherwise a shared non-ascent direction (Santa, 2024)
GBBN direction $F_j(x)=f_j(x)+g_j(x)$2	$F_j(x)=f_j(x)+g_j(x)$3	$F_j(x)=f_j(x)+g_j(x)$4, and $F_j(x)=f_j(x)+g_j(x)$5 is not Pareto critical (Yang, 2024)

The steepest-descent template underlying several methods is the minimax quadratic model

$F_j(x)=f_j(x)+g_j(x)$6

whose dual representation yields

$F_j(x)=f_j(x)+g_j(x)$7

PSMGD adopts the same minimal-norm weighted-gradient principle but computes the weights only periodically, reusing them over short intervals. The incremental method based on the central descent direction instead stores gradients and refreshes only one or two per iteration, while still solving the full central-descent QP (Xu et al., 2024, Oliveira et al., 2021).

A notable distinction lies in normalization. GBBN replaces $F_j(x)=f_j(x)+g_j(x)$8 by

$F_j(x)=f_j(x)+g_j(x)$9

with the stated aim of avoiding unbalanced gradients and enlarging steps. The LP-based method also normalizes gradients, using $F_j(x)=s_j(x)+r_j(x)$0, but couples this with a linear rather than quadratic subproblem. This suggests that “global descent function” is not tied to one canonical subproblem; rather, it denotes a family of common-descent generators with different trade-offs between robustness, invariance, and computational expense (Yang, 2024, Santa, 2024).

3. Quadratic model-based global descent for composite objectives

In convex composite multi-objective optimization, the global descent function is often the worst-case value of a local quadratic model. For

$F_j(x)=s_j(x)+r_j(x)$1

with $F_j(x)=s_j(x)+r_j(x)$2 convex, twice continuously differentiable, $F_j(x)=s_j(x)+r_j(x)$3 Lipschitz continuous, and $F_j(x)=s_j(x)+r_j(x)$4, the Newton-type proximal-gradient method defines

$F_j(x)=s_j(x)+r_j(x)$5

and then sets

$F_j(x)=s_j(x)+r_j(x)$6

Because $F_j(x)=s_j(x)+r_j(x)$7 is $F_j(x)=s_j(x)+r_j(x)$8-strongly convex, it has a unique minimizer

$F_j(x)=s_j(x)+r_j(x)$9

The method computes $f(y)\le f(x)$0, performs an Armijo type line search, and updates $f(y)\le f(x)$1. Under bounded level sets and the stated regularity assumptions, every limit point is a critical point, with the descent estimate

$f(y)\le f(x)$2

playing the central role (Ansary, 2022).

The trust-region proximal-gradient method uses the same max-over-objectives principle but adds a trust-region constraint. With

$f(y)\le f(x)$3

it solves

$f(y)\le f(x)$4

and denotes the unique minimizer by $f(y)\le f(x)$5, with $f(y)\le f(x)$6. The paper derives

$f(y)\le f(x)$7

defines the predicted reduction by $f(y)\le f(x)$8, the actual reduction by

$f(y)\le f(x)$9

and updates the trust-region radius through the ratio $f_i(y)\le f_i(x)$0. The same paper states that the method is free from any kind of priori chosen parameters or ordering information of objective functions, and that every accumulation point is a critical point under mild assumptions (Ansary, 2024).

These model-based constructions make “global descent function” a scalarized worst-case model rather than a scalarization of the original objectives. The max operator preserves simultaneity: the model is only favorable when every objective admits decrease along the chosen step (Ansary, 2022, Ansary, 2024).

4. Filled-function constructions for non-convex basin escape

In non-convex multi-objective optimization, the term also denotes a vector-valued auxiliary function whose purpose is not merely local descent but escape from a current local weakly efficient basin. This line of work is explicitly modeled on filled-function methods from single-objective optimization.

One formulation starts from a current local weakly efficient solution $f_i(y)\le f_i(x)$1 and defines the one-parameter filled function

$f_i(y)\le f_i(x)$2

where

$f_i(y)\le f_i(x)$3

The paper establishes three properties: $f_i(y)\le f_i(x)$4 is a local weakly efficient solution of $f_i(y)\le f_i(x)$5; there are no Fritz–John or KKT critical points of $f_i(y)\le f_i(x)$6 in a set $f_i(y)\le f_i(x)$7 describing points that do not improve all components; and if there exists a different local front with a point strictly better in all objectives, then for small enough $f_i(y)\le f_i(x)$8 the filled function has a new weakly efficient solution near that better front. Under the assumptions of finitely many disjoint local weak Pareto fronts and interior global efficient points, alternating a local phase and a fill phase yields finite convergence to global weak efficiency (Adhikary et al., 31 Mar 2026).

A related construction defines, at a local weakly efficient point $f_i(y)\le f_i(x)$9, a vector-valued function

$i$0

with $i$1. The defining conditions are: $i$2 is a strict local weak-efficient solution of $i$3; $i$4 has no Fritz John stationary points in the basin

$i$5

and whenever there exists a strictly better local weakly efficient solution $i$6, the auxiliary problem admits a local weakly efficient solution $i$7 near $i$8 with

$i$9

Under compactness, Lipschitz gradients, finiteness of local weakly efficient values, and interior global solutions, the method is stated to generate the entire global Pareto front (Adhikary et al., 30 Jul 2025).

These non-convex formulations clarify a central ambiguity of the phrase “global descent.” Here, “global” does not denote one-step global optimization. It denotes an auxiliary function engineered so that local search on the auxiliary landscape can leave a local Pareto trap of the original problem and transition to an improved local weakly efficient basin (Adhikary et al., 31 Mar 2026, Adhikary et al., 30 Jul 2025).

5. Computational variants: incremental, stochastic, normalized, and fractional

A substantial part of the literature modifies the global descent construction to improve per-iteration cost, step-size behavior, or geometric robustness. The incremental descent method replaces the requirement of computing all $d$0 gradients at every iteration by a round-robin strategy that refreshes only a constant number of gradients while still solving the central-descent QP with stored gradients. Under $d$1-Lipschitz gradients and lower bounded objectives, it retains the standard stationarity rate

$d$2

but reduces the total query complexity from $d$3 to $d$4, independent of $d$5 (Oliveira et al., 2021).

PSMGD uses the minimal-norm weighted-gradient global descent direction but recomputes the simplex weights only every $d$6 iterations. The stochastic update is

$d$7

with periodic solutions of

$d$8

The paper introduces backpropagation complexity, defined as the total number of backpropagations required to reach a target accuracy $d$9, and shows that if $\langle \nabla f_i(x),d\rangle<0$0, the backpropagation complexity becomes objective-independent: $\langle \nabla f_i(x),d\rangle<0$1 for strongly convex objectives, and $\langle \nabla f_i(x),d\rangle<0$2 for convex and non-convex objectives (Xu et al., 2024).

GBBN addresses a different issue: the tendency of steepest descent to generate small stepsizes. It couples gradient normalization with Barzilai–Borwein step rules and a nonmonotone Armijo-type line search. Under bounded lower level sets and Lipschitz gradients, every accumulation point is Pareto critical; under strong convexity and an angle condition, the paper establishes R-linear convergence (Yang, 2024).

The Caputo-fractional variant replaces ordinary gradients by adaptive-order Caputo fractional gradients,

$\langle \nabla f_i(x),d\rangle<0$3

and computes the descent direction from the convex quadratic program

$\langle \nabla f_i(x),d\rangle<0$4

Its KKT system yields

$\langle \nabla f_i(x),d\rangle<0$5

and the Armijo rule is used to ensure $\langle \nabla f_i(x),d\rangle<0$6 for all $\langle \nabla f_i(x),d\rangle<0$7. The convergence result is stated for Tikhonov-regularized solutions, with a linear-rate bound in the quadratic case and convergence to an integer-order Pareto-critical point as the regularization vanishes (Shaw et al., 10 Jul 2025).

6. Constrained formulations and transfer to single-objective escape mechanisms

The global descent idea extends naturally to constrained multi-objective optimization. In the SQP framework for inequality constraints, one defines the penalty

$\langle \nabla f_i(x),d\rangle<0$8

the merit functions

$\langle \nabla f_i(x),d\rangle<0$9

and the directional model

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$00

The search direction comes from the convex QP

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$01

subject to the linearized objective and constraint bounds

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$02

The subproblem is always feasible, and if $\min_x F(x)=(F_1(x),\dots,F_m(x)),$03, then for sufficiently large $\min_x F(x)=(F_1(x),\dots,F_m(x)),$04,

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$05

Under Lipschitz gradients and the Mangasarian–Fromovitz constraint qualification, every accumulation point is either a strong or weak MOP-critical point (Ansary et al., 2018).

The same descent logic has also been transferred back to single-objective multimodal optimization through multiobjectivization. A single objective $\min_x F(x)=(F_1(x),\dots,F_m(x)),$06 is lifted to the bi-objective problem

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$07

with normalized gradients

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$08

and multi-objective gradient

$\min_x F(x)=(F_1(x),\dots,F_m(x)),$09

At a strict local minimizer $\min_x F(x)=(F_1(x),\dots,F_m(x)),$10 of $\min_x F(x)=(F_1(x),\dots,F_m(x)),$11 with $\min_x F(x)=(F_1(x),\dots,F_m(x)),$12, one has $\min_x F(x)=(F_1(x),\dots,F_m(x)),$13 but $\min_x F(x)=(F_1(x),\dots,F_m(x)),$14, so the multi-objective gradient does not vanish and the trajectory leaves the single-objective trap. This provides a concrete example in which a multi-objective descent construction is used not to approximate a Pareto set, but to connect basins of attraction in a multimodal single-objective landscape (Steinhoff et al., 2020).

Taken together, these formulations show that a multi-objective global descent function is not a single mathematical object. In the cited papers it may be a direction map $\min_x F(x)=(F_1(x),\dots,F_m(x)),$15, a model value $\min_x F(x)=(F_1(x),\dots,F_m(x)),$16, a worst-case quadratic function $\min_x F(x)=(F_1(x),\dots,F_m(x)),$17, or a vector-valued auxiliary landscape $\min_x F(x)=(F_1(x),\dots,F_m(x)),$18. The unifying criterion is operational rather than formal: each construction is designed to enforce simultaneous descent, detect Pareto criticality, or escape a local weakly efficient basin without relying on predefined scalarization weights or preference ordering (Santa, 2024, Ansary, 2022, Adhikary et al., 30 Jul 2025).