Functional Gradient Ascent: Theory & Applications

• Functional Gradient Ascent (FGA) is an optimization method extending gradient ascent to infinite-dimensional function spaces, enabling sophisticated control and learning.
• It employs inner product-based differentiation and basis expansion techniques to update functionals in applications such as quantum control and minimax problems.
• Empirical studies show FGA achieves low quantum gate infidelities and enhances performance in overparameterized neural networks through effective functional updates.

Functional Gradient Ascent (FGA) refers to a class of optimization algorithms in which ascent steps are taken in a functional (infinite-dimensional) space rather than in simple parameter or finite-dimensional vector spaces. FGA methods are motivated by control theory, statistical learning, and the analysis of overparameterized neural networks, where optimization must be performed over functions, distributions, or other objects in high- or infinite-dimensional domains. This article provides an in-depth exposition of the theory, methodology, variants, and applications of FGA in quantum control, nonconvex learning, and minimax optimization.

1. Mathematical Foundations of Functional Gradient Ascent

FGA generalizes the notion of gradient ascent from finite-dimensional parameter spaces to spaces of functions. Given a real-valued functional $J[f]$ defined on a function space (e.g., $f: X \to \mathbb{R}^K$ or quantum control fields $\epsilon_j(t)$), the directional derivative in a direction $\delta f$ is

$$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$

If there exists a function $g(\cdot)$ such that $$dJ[f; \delta f] = \langle g, \delta f \rangle_{\mathcal{H}}$$ for all admissible $\delta f$ (with $\langle \cdot, \cdot \rangle_{\mathcal{H}}$ an inner product), then $g$ is called the functional gradient $f: X \to \mathbb{R}^K$0. This serves as the direct infinite-dimensional analogue of gradients in parameter optimization (Johnson et al., 2020).

In practical implementations, FGA involves defining an appropriate inner product, performing chain-rule differentiation through any function parameterizations (e.g., neural network weights or functional basis coefficients), and updating the function in the direction of ascent.

2. Functional Gradient Ascent in Quantum Optimal Control

A canonical realization of FGA appears in quantum optimal control, specifically in the GRAFS method for synthesizing quantum gates through shaped control fields. In this setting, the optimization target is a functional of the quantum evolution operator $f: X \to \mathbb{R}^K$1: $f: X \to \mathbb{R}^K$2 where $f: X \to \mathbb{R}^K$3 is the drift Hamiltonian, $f: X \to \mathbb{R}^K$4 are control Hamiltonians, and $f: X \to \mathbb{R}^K$5 are time-dependent control fields. The phase-invariant fidelity objective is

$f: X \to \mathbb{R}^K$6

Controls are constrained to be band-limited and of finite amplitude, incorporating physical hardware limits (Lucarelli, 2016).

To ensure constraints and efficient parameterization, each $f: X \to \mathbb{R}^K$7 is represented as a linear combination of Slepian sequences (discrete prolate spheroidal functions), leading to a finite basis expansion

$f: X \to \mathbb{R}^K$8

with $f: X \to \mathbb{R}^K$9 for time-steps $\epsilon_j(t)$0 and half-bandwidth $\epsilon_j(t)$1. The gradient of the fidelity with respect to the basis coefficients $\epsilon_j(t)$2 is computed via product rule and chain rule, yielding

$\epsilon_j(t)$3

and the update is performed as $\epsilon_j(t)$4, with post-update projection to amplitude bounds if needed (Lucarelli, 2016).

3. FGA in Nonconvex and Infinite-Dimensional Learning Problems

In machine learning, FGA is leveraged for training nonconvex models and for minimax problems defined over infinite-dimensional function classes. In such applications, functionals $\epsilon_j(t)$5 represent risks or losses over predictors $\epsilon_j(t)$6. The FGA algorithm generalizes stochastic (mirror) descent to function spaces:

• Compute the functional gradient $\epsilon_j(t)$7 at the current iterate $\epsilon_j(t)$8.
• Take a functional-mirror-descent or preconditioned step:

$\epsilon_j(t)$9

where $\delta f$0 is a Bregman divergence from a convex function $\delta f$1, and $\delta f$2 is the pointwise loss. For $\delta f$3, this reduces to $\delta f$4.

• Parameter updates are performed by aligning parametric $\delta f$5 to $\delta f$6 under the chosen divergence, typically via stochastic gradient descent (Johnson et al., 2020).

In minimax optimization, such as those encountered in conditional expectation estimation or adversarial scenarios, FGA is combined with gradient ascent/descent on saddle-point objectives in function space. For two-layer neural networks, this leads to mean-field dynamics that can be interpreted as Wasserstein gradient flows in the infinite-width limit (Zhu et al., 2024).

4. Convergence Properties and Quantum Speed Limits

Convergence guarantees for FGA depend on both the structure of the function space and objective regularity. In quantum control, the time-bandwidth quantum speed limit (QSL) constrains reachable fidelities as a function of available bandwidth. For a target infidelity $\delta f$7, the minimal pulse duration $\delta f$8 achievable using a band-limited control is shown to scale as $\delta f$9, with GRAFS numerically attaining this bound for entangling gates (Lucarelli, 2016).

For mirror-descent variants in nonconvex risk minimization, theoretical results establish monotonic decrease of the risk, with

$$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$0

implying convergence to stationary points (Johnson et al., 2020).

In neural minimax optimization, mean-field FGA corresponds to Wasserstein gradient flows for parameter distribution measures. Global convergence to stationary points at rate $$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$1 is established under boundedness and regularity, and a sublinear $$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$2 rate is proven under strong convexity in the regularizer (Zhu et al., 2024). The evolution of the distribution of neural network features under FGA is controlled, with the 2-Wasserstein distance to initialization bounded by $$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$3.

5. Algorithmic Summaries and Implementation

The GRAFS algorithm in quantum control proceeds as follows (Lucarelli, 2016):

1. Initialize basis coefficients $$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$4.
2. Compute control fields via basis expansion.
3. Evaluate the fidelity $$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$5 by propagating the quantum evolution.
4. Calculate the gradient w.r.t. coefficients via backpropagation of the matrix exponential derivatives.
5. Update coefficients using the gradient and project to amplitude bounds.
6. Repeat until fidelity or gradient norms meet stopping criteria.

For nonconvex learning, typical implementation alternates between functional guide steps and parameter updates, with hyperparameters such as outer stages, inner SGD batch size, momentum, and step size. Practical experiments show that FGA achieves consistent generalization improvements over standard stochastic gradient descent and self-distillation across multiple vision and text domains (Johnson et al., 2020).

In minimax mean-field learning, the discrete-time functional GDA (with overparameterized two-layer nets) converges to infinite-width Wasserstein gradient flow PDEs. The convergence rates and representation shifts depend on scaling parameters and network width (Zhu et al., 2024).

6. Empirical Performance and Benchmarks

FGA and its variants demonstrate empirical success across several domains:

• In quantum control, infidelities as low as $$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$6 are achieved on three-qubit Toffoli gates in $$dJ[f; \delta f] = \lim_{\epsilon \to 0} \frac{J[f + \epsilon \delta f] - J[f]}{\epsilon}.$$7 GRAFS iterations. Minimal time to reach target fidelity is shown to obey the predicted inverse-bandwidth scaling (Lucarelli, 2016).
• For machine learning, FGA-trainings exhibit "smooth path" dynamics where intermediate iterates surpass base-model generalization. On datasets such as CIFAR100 and ImageNet, test error reductions of 1-2\% absolute over strong SGD baselines are reported. The performance also surpasses that of deeper standard-trained architectures when FGA is applied to shallower models (Johnson et al., 2020).
• In mean-field minimax neural optimization, FGA's global convergence and representation learning effects are rigorously characterized; applications include policy evaluation, IV regression, asset pricing, and adversarial Riesz estimation (Zhu et al., 2024).

7. Variants, Generalizations, and Significance

FGA encompasses a diversity of algorithmic variants:

• Basis-expansion-based (e.g., Slepian sequence parameterizations in GRAFS).
• Mirror-descent approaches in function spaces using general convex divergences.
• Wasserstein flows for parameter distributions in overparameterized neural nets.
• Successive functional gradient steps with adaptive or Newton-like preconditioning.

A key significance of FGA is efficient exploitation of function space structure, either to encode physical constraints (bandwidth, amplitude in quantum control) or to realize smooth interpolation between models in learning applications. The flexibility of FGA to various application domains, its compatibility with theoretical convergence rates, and empirical demonstration of generalization and control performance reinforce its practical and conceptual importance (Lucarelli, 2016, Johnson et al., 2020, Zhu et al., 2024).