--- title: Multi-Objective Global Descent Function url: https://www.emergentmind.com/topics/multi-objective-global-descent-function type: topic --- # Multi-Objective Global Descent Function 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 [2412.10961] [2205.04200] [2507.22390]. ## 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 \(i\) [2205.04200] [2403.05070] [2410.19502]. 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 \[ \max_{j=1,\dots,m} F_j'(x;d)\ge 0 \quad \text{for all } d\in\mathbb R^n. \] In smooth non-convex settings, weak Pareto criticality is also written through multipliers: \[ \sum_{j=1}^m \lambda_j \nabla f_j(x^*)=0,\qquad \lambda\in\mathbb R_+^m,\ \lambda\neq 0. \] A recurrent structural fact is that the descent construction itself detects criticality: in different formulations, \(p(x)=0\), \(d(x)=0\), or \(t(x)=0\) is equivalent to first-order Pareto criticality [2205.04200] [2403.05070] [2603.29338]. ## 2. Direction-generating formulations in differentiable optimization For smooth unconstrained problems, a multi-objective global descent function is often a map \(x\mapsto d(x)\) or \(x\mapsto p(x)\) 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 \(p(x)\) | \(\min_{w\in\Delta_m}\left\|\sum_i w_i \nabla f_i(x)\right\|^2\) | If \(x\) is not Pareto-stationary, then \(\nabla f_i(x)^\top p(x)<0\) for all \(i\) [2412.10961] | | Central descent direction \(V_c(x)\) | \(\min_V \tfrac12\|V\|^2\) s.t. \(\nabla f_i(x)^\top V\le -\|\nabla f_i(x)\|\) | Invariant under monotone re-scaling \(f_i\mapsto g_i\circ f_i\); admits simultaneous descent for sufficiently small \(\alpha\) under \(L\)-Lipschitz gradients [2105.11845] | | LP-based direction \(D(x)=p^*(x)\) | LP with normalized-gradient constraints \(\bar g_i(x)^\top p\le \beta\), box bounds on \(p\), and \(\beta\le 0\) | Returns a shared strict descent direction when one exists, otherwise a shared non-ascent direction [2406.08147] | | GBBN direction \(d_E(x)\) | \(\min_d \max_i\{\langle \nabla \hat f_i(x),d\rangle+\tfrac12\|d\|^2\}\) | \(\theta_E(x)=-\tfrac12\|d_E(x)\|^2\le 0\), and \(d_E(x)\neq 0\iff x\) is not Pareto critical [2403.05070] | The steepest-descent template underlying several methods is the minimax quadratic model \[ d(x)=\arg\min_d \max_i \left\{\langle \nabla f_i(x),d\rangle+\tfrac12\|d\|^2\right\}, \] whose dual representation yields \[ d(x)=-\sum_{i=1}^m \lambda_i(x)\nabla f_i(x),\qquad \lambda(x)\in\Delta_m. \] 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 [2412.10961] [2105.11845]. A notable distinction lies in normalization. GBBN replaces \(\nabla f_i(x)\) by \[ \nabla \hat f_i(x)=\frac{\nabla f_i(x)}{\|\nabla f_i(x)\|+\eta}, \] with the stated aim of avoiding unbalanced gradients and enlarging steps. The LP-based method also normalizes gradients, using \(\bar g_i(x)=g_i(x)/\|g_i(x)\|_2\), 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 [2403.05070] [2406.08147]. ## 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 \[ F_j(x)=f_j(x)+g_j(x), \] with \(f_j\) convex, twice continuously differentiable, \(\nabla f_j\) Lipschitz continuous, and \(\nabla^2 f_j(x)\succeq \sigma I\), the Newton-type proximal-gradient method defines \[ Q_j(x;d)=\nabla f_j(x)^\top d+\frac12 d^\top \nabla^2 f_j(x)d+\bigl[g_j(x+d)-g_j(x)\bigr], \] and then sets \[ Q(x;d)=\max_{j=1,\dots,m} Q_j(x;d). \] Because \(Q(x;\cdot)\) is \(\sigma\)-strongly convex, it has a unique minimizer \[ d(x)=\arg\min_d Q(x;d),\qquad t(x)=Q(x;d(x))\le 0. \] The method computes \(d(x)\), performs an Armijo type line search, and updates \(x^{k+1}=x^k+\alpha_k d^k\). Under bounded level sets and the stated regularity assumptions, every limit point is a critical point, with the descent estimate \[ t(x)\le -\frac{\sigma}{2}\|d(x)\|^2 \] playing the central role [2205.04200]. The trust-region proximal-gradient method uses the same max-over-objectives principle but adds a trust-region constraint. With \[ Q_j(x^k,d)=\langle \nabla s_j(x^k),d\rangle+\frac12 d^\top B_j^k d+r_j(x^k+d)-r_j(x^k), \] it solves \[ \min_{\|d\|\le \Delta_k} Q(x^k,d),\qquad Q(x^k,d)=\max_j Q_j(x^k,d), \] and denotes the unique minimizer by \(d^k\), with \(t^k=Q(x^k,d^k)\le 0\). The paper derives \[ t^k\le -(m/2+\mu)\|d^k\|^2<0 \quad \text{whenever } d^k\neq 0, \] defines the predicted reduction by \(\mathrm{Pred}(d^k)=-t^k\), the actual reduction by \[ \mathrm{Ared}(d^k)=\min_j [F_j(x^k)-F_j(x^k+d^k)], \] and updates the trust-region radius through the ratio \(\rho^k=\mathrm{Ared}(d^k)/\mathrm{Pred}(d^k)\). 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 [2410.19502]. 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 [2205.04200] [2410.19502]. ## 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 \(x^k\in \operatorname{int}\mathcal X\) and defines the one-parameter filled function \[ F_{j,x^k,\mu}(x)= -\|x-x^k\|^2+\varphi_\mu\bigl(f_j(x)-f_j(x^k)\bigr), \] where \[ \varphi_\mu(t)= \begin{cases} -\mu t^3, & t\ge 0,\\[1ex] -\dfrac{t^2}{\mu}, & t<0. \end{cases} \] The paper establishes three properties: \(x^k\) is a local weakly efficient solution of \(\min (-F_{x^k,\mu}(x))\); there are no Fritz–John or KKT critical points of \(F_{x^k,\mu}\) in a set \(U_1\) 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 \(\mu\) 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 [2603.29338]. A related construction defines, at a local weakly efficient point \(\bar x\), a vector-valued function \[ G_{\mu,\alpha,\tau}^{(\bar x)}(z)=\bigl(G_1(z),\dots,G_m(z)\bigr)^\top,\qquad G_j(z)=A_\tau(f_j(z)-f_j(\bar x))-\mu\|z-\bar x\|, \] with \(A_\tau(y)=yV_\tau(y)\). The defining conditions are: \(\bar x\) is a strict local weak-efficient solution of \(\min -G_{\mu,\alpha,\tau}(z)\); \(G\) has no Fritz John stationary points in the basin \[ \hat X(\bar x)=\{z\in \operatorname{int}X:\ z\neq \bar x,\ f_j(z)\ge f_j(\bar x)\ \text{for at least one }j\}; \] and whenever there exists a strictly better local weakly efficient solution \(z^{**}\), the auxiliary problem admits a local weakly efficient solution \(z'\) near \(z^{**}\) with \[ f_j(z')