Pareto Navigation Gradient Descent

• Pareto Navigation Gradient Descent is a first-order optimization method that finds Pareto optimal solutions by balancing objective stationarity with a reference criterion.
• It separates the challenge of reaching the Pareto set from optimizing a user-defined or diversity-oriented target within that set.
• The method efficiently solves a constrained quadratic subproblem for parameter updates, achieving significant speed-ups over second-order approaches in deep learning settings.

Searching arXiv for the core PNGD paper and closely related preference-guided Pareto-set optimization works. Pareto Navigation Gradient Descent (PNG) is a first-order method for optimization in Pareto set (OPT-in-Pareto): given multiple objective functions, it seeks parameters that are Pareto optimal with respect to those objectives while simultaneously optimizing an additional reference criterion within the Pareto set. In the formulation introduced in "Pareto Navigation Gradient Descent: a First-Order Algorithm for Optimization in Pareto Set," the reference criterion can encode either a user preference or a diversity measure, and the method is designed for large-scale, non-convex, non-linear settings in which a naive Riemannian manifold gradient descent on the Pareto set is computationally expensive because it requires eigen-calculation of Hessian matrices (Ye et al., 2021).

1. Problem formulation

In multi-objective learning, one considers several possibly conflicting losses, $\ell_1(\theta),\ldots,\ell_m(\theta)$. The Pareto set collects models that cannot be strictly improved in one objective without worsening another. PNG is built around the observation that the Pareto set alone does not specify which Pareto model should be returned in practice; the missing ingredient is an extra criterion $F(\theta)$ to be optimized within the Pareto set (Ye et al., 2021).

The central object is the Pareto stationary set

$$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$

where $C^m$ is the probability simplex. The OPT-in-Pareto problem is then written as

$\min_{\theta \in P^*} F(\theta),$

with $P^*$ denoting the set of Pareto local optima. The restriction to local optima reflects the nonconvex, high-dimensional regime emphasized in the original formulation (Ye et al., 2021).

This formulation separates two tasks that are often conflated. The first is reaching the Pareto set, expressed through the stationarity measure $g(\theta)$. The second is navigating within that set according to $F(\theta)$. In the examples given for PNG, $F$ may represent a specific task preference, a reference-based criterion, or a diversity-inducing objective over several Pareto points (Ye et al., 2021).

2. Core optimization mechanism

PNG updates parameters by

$\theta_{k+1} = \theta_k - \xi v_k,$

where the search direction $F(\theta)$0 is defined as the solution of

$F(\theta)$1

The direction is therefore the closest feasible vector to the steepest descent direction of $F(\theta)$2, under simultaneous constraints on all task losses (Ye et al., 2021).

The switching quantity $F(\theta)$3 determines whether the iteration emphasizes Pareto improvement or criterion optimization: $F(\theta)$4 When $F(\theta)$5, the method is explicitly driven toward the Pareto set. When $F(\theta)$6, the constraints are effectively relaxed and the method behaves as unconstrained gradient descent on $F(\theta)$7 (Ye et al., 2021).

The constrained subproblem admits a dual form. The resulting direction can be written as

$F(\theta)$8

where the multipliers solve

$F(\theta)$9

Because this is a convex quadratic program in the number of objectives rather than the ambient parameter dimension, the method remains compatible with settings in which $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$0 is large but $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$1 is moderate (Ye et al., 2021).

Operationally, the procedure is: compute $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$2 and the task gradients, evaluate $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$3, set $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$4, solve the dual quadratic program for the multipliers, construct $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$5, and update $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$6. This is the basic mechanism by which PNG alternates between attraction to the Pareto set and motion along it (Ye et al., 2021).

3. Geometric interpretation and relation to multi-gradient descent

Geometrically, PNG does not attempt to parameterize the Pareto set explicitly. Instead, it uses first-order information to bias updates so that they either decrease all objectives or, once sufficiently close to Pareto stationarity, optimize the reference criterion while remaining in the relevant neighborhood of the Pareto set (Ye et al., 2021).

This makes PNG distinct from a naive manifold or Riemannian gradient descent on the Pareto set. In that second-order view, one projects the descent of $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$7 onto a tangent space of the Pareto manifold, which requires Hessian information and, in the presentation of PNG, eigen-calculation of Hessian matrices. PNG replaces that explicit tangent-space computation with a first-order constrained direction-finding problem (Ye et al., 2021).

The method is also closely related to the broader family of multi-gradient methods. Classical multiple-gradient descent seeks a common descent direction through a minimum-norm convex combination of objective gradients; the standard form is

$$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$8

and it is widely used to identify Pareto descent directions (Ip et al., 10 Feb 2025). PNG can be viewed as augmenting that common-descent logic with an explicit target criterion $$P := \left\{ \theta : g(\theta) = 0 \right\}, \qquad g(\theta) := \min_{\omega\in C^m} \left\| \sum_{i=1}^m \omega_i \nabla \ell_i(\theta) \right\|^2,$$9, thereby turning Pareto criticality from the endpoint of optimization into a feasible set for a second objective (Ye et al., 2021).

This suggests a useful distinction. Standard MGDA-type methods principally address how to approach Pareto stationary points. PNG addresses how to optimize once those points matter as a set. Later preference-guided methods preserve this separation, even when they use different technical devices such as constraints, penalties, or predictor-corrector constructions.

4. Convergence properties and empirical behavior

The theoretical guarantees given for PNG have two parts. Away from the Pareto set, when $C^m$0, the continuous-time analysis yields

$C^m$1

so every objective decreases and the trajectory reaches an $C^m$2-stationary Pareto point in finite time (Ye et al., 2021).

Near the Pareto set, when $C^m$3, the method follows unconstrained gradient descent on $C^m$4. In this regime, fixed points correspond to local minima of $C^m$5 subject to Pareto constraints. The method therefore combines two guarantees: convergence to a neighborhood of Pareto stationarity, and optimization of the chosen reference criterion inside that neighborhood (Ye et al., 2021).

The empirical studies reported for PNG cover several different uses of $C^m$6. For task-preference selection, the method recovers Pareto solutions with specific loss ratios. For general criteria, it is used with nonlinear, nonconvex user-specified functions of the objectives. For diversity, the authors optimize

$C^m$7

which is reported to yield solutions uniformly spread on the Pareto set asymptotically. The diversity experiments are evaluated with HV and IGD+ metrics, and large-scale demonstrations include NYUv2 three-task learning and a domain-generalization setting based on JiGen (Ye et al., 2021).

The reported runtime comparison is also central to the method’s positioning: PNG is described as over 50 times faster than recent second-order local expansion methods in deep network scenarios, specifically on Multi-MNIST. This claim is tied directly to the first-order construction and the avoidance of explicit Hessian-based geometry (Ye et al., 2021).

5. Subsequent developments and related preference-guided variants

Later work generalizes the idea of preference-aware navigation on or near the Pareto set, but with different formulations.

Method	Main mechanism	Relation to PNG
PMGDA (Zhang et al., 2024)	Predict-and-correct framework with a preference constraint $C^m$8	Targets a Pareto solution that fits a user preference or ROI
FOOPS (Chen et al., 26 Mar 2025)	Penalty reformulation using a merit function $C^m$9	Optimizes a preference function subject to weak Pareto optimality
PPE (Amakor et al., 17 Jun 2026)	Predictor steps tangential to the Pareto manifold, followed by MGDA correction	Interactive Pareto navigation with Hessian-free Krylov solves

PMGDA seeks a Pareto solution satisfying a user-specific preference encoded by a constraint function $\min_{\theta \in P^*} F(\theta),$0. It distinguishes a prediction step, which keeps motion tangent to the constraint surface through

$\min_{\theta \in P^*} F(\theta),$1

from a correction step, which explicitly reduces $\min_{\theta \in P^*} F(\theta),$2 while controlling degradation of the objectives. The formulation supports both “exact” preference alignment and region-of-interest constraints, and the reported experiments cover standard benchmarks, multi-task learning, and multi-objective reinforcement learning with more than thousands of decision variables (Zhang et al., 2024).

FOOPS reformulates optimization on the Pareto set as a semivectorial bilevel problem and replaces the weak-Pareto constraint by a smooth merit function,

$\min_{\theta \in P^*} F(\theta),$3

then solves a penalty objective of the form

$\min_{\theta \in P^*} F(\theta),$4

The paper states that this yields a first-order method with an easy-to-evaluate gradient and nonasymptotic convergence guarantees, while directly targeting a user-specified preference function over the weak Pareto set (Chen et al., 26 Mar 2025).

PPE moves closer to explicit manifold navigation. It performs predictor steps tangent to the manifold of Pareto-optimal solutions and corrector steps that restore Pareto criticality. To avoid explicit Hessian computations, it uses a Krylov subspace method based on matrix-vector products obtained by automatic differentiation. This makes the method interactive: a decision maker provides preferences during exploration, and each new preference induces a new local motion along the Pareto set (Amakor et al., 17 Jun 2026).

Taken together, these later methods preserve the central PNG theme—optimization guided by preference or auxiliary criteria with respect to Pareto-optimality—while altering the machinery used to represent or enforce Pareto structure.

6. Scope, limitations, and terminology

PNG is not a universal template for every multi-objective setting. In particular, constrained multi-objective optimization motivates additional machinery beyond standard MGDA- or PNGD-style updates. A later two-stage framework for constrained multi-gradient descent argues that MGDA may struggle with constrained problems, whether constraints are incorporated into the gradient hull or handled via projection onto the feasible region, and proposes a min-max first stage for weak Pareto stationarity followed by a min-min refinement stage for full Pareto stationarity (Lei et al., 19 Feb 2025).

Other nearby developments modify the descent mechanism rather than the preference formulation. Multiobjective memory gradient methods combine current and past directions and prove sufficient descent and global convergence to Pareto critical points (Chen et al., 2022). Inertial and Nesterov-accelerated multiobjective methods derive accelerated variants of Pareto navigation or multiobjective steepest descent, including an algorithm with an improved convergence rate compared to the plain multiobjective steepest descent method and a greedy accelerated variant that avoids solving a quadratic subproblem in each step (Sonntag et al., 2022). A linear-programming-based alternative computes shared descent directions through an LP and introduces a backtracking strategy intended to improve exploration of the Pareto set and the Pareto front (Santa, 2024).

A common terminological confusion arises because “Pareto-optimal” is also used in an unrelated complexity-theoretic sense. "Gradient Descent is Pareto-Optimal in the Oracle Complexity and Memory Tradeoff for Feasibility Problems" studies separation-oracle feasibility and proves lower bounds showing that gradient descent is Pareto-optimal in an oracle-complexity versus memory tradeoff. That result concerns feasibility problems, oracle queries, and memory usage, not optimization within a multi-objective Pareto set (Blanchard, 2024).

Within multi-objective optimization proper, Pareto Navigation Gradient Descent therefore denotes a specific first-order strategy for OPT-in-Pareto: approach Pareto stationarity using common-descent constraints, then optimize a preference or diversity criterion on the resulting set. Its main significance lies in making that program feasible without second-order manifold computations, while subsequent work shows that the same agenda can also be pursued through predict-and-correct constraints, merit-function penalties, and interactive tangent-space navigation (Ye et al., 2021).