---
title: 'Graph GP-UCB: Manifold Bayesian Optimization'
url: https://www.emergentmind.com/topics/graph-gaussian-process-ucb-ggp-ucb
type: topic
---

# Graph GP-UCB: Manifold Bayesian Optimization

Graph Gaussian Process Upper Confidence Bound (GGP-UCB) is a Bayesian optimization framework designed for maximizing expensive black-box functions defined on unknown compact manifolds. GGP-UCB operates under the minimal informational assumption that only a point-cloud sample from the manifold is available. The method constructs a surrogate Gaussian process model using a graph built over these samples and utilizes a UCB acquisition criterion that corrects for geometric misspecification, offering both theoretical guarantees and empirical robustness across synthetic and inverse-problem benchmarks [2210.10962].

## 1. Problem Setting and Motivation

GGP-UCB addresses the scenario of optimizing a function $f$ defined on an $m$-dimensional compact manifold $M \subset \mathbb{R}^d$ when only a finite set $M_N = \{x_1, \ldots, x_N\} \subset M$ is accessible—the full analytic structure of $M$ is unknown. Evaluations of $f$ are costly, motivating the need for query-efficient, geometry-adaptive Bayesian optimization. The framework leverages manifold learning principles to construct a statistical surrogate for $f$ that respects intrinsic geometric constraints, thus ameliorating the limitations of standard Euclidean Gaussian process (GP) surrogates that ignore manifold structure.

## 2. Graph Construction and Spectral Kernel Approximation

The surrogate model is defined on a weighted graph constructed over $M_N$:

- The weight matrix $W \in \mathbb{R}^{N \times N}$ is defined by
  $$
  W_{ij} =
  C_m h_N^{-(m+2)} \mathbf{1}_{\|x_i - x_j\| < h_N}
  $$
  where $C_m = 2(m+2) / \operatorname{Vol}(B_1)$ normalizes by the unit $m$-ball's volume, and $h_N$ is a connectivity radius.
- The (unnormalized) graph Laplacian is $\Delta_N = D - W$, with $D$ the diagonal degree matrix.
- Spectral decomposition yields eigenpairs $\{\psi_{N,\ell}, \lambda_{N,\ell}\}_{\ell=1}^N$, $0 = \lambda_{N,1} \leq \ldots$.

Graph-based approximations to manifold kernels are formed via spectral filtering:
- **Graph Matérn kernel** (with smoothness $s>m/2$ and inverse length-scale $\kappa$):
  $$
  k^{(N)}_M(x_i, x_j) = \kappa^{2s-m} \sum_{\ell=1}^{k_N} (\kappa^2 + \lambda_{N,\ell})^{-s} \psi_{N,\ell}(x_i)\psi_{N,\ell}(x_j)
  $$
- **Graph squared-exponential (SE or heat kernel)** (with length-scale $\tau$):
  $$
  k^{(N)}_{SE}(x_i, x_j) = \tau^{m/2} \sum_{\ell=1}^{k_N} \exp(-\tau \lambda_{N,\ell}) \psi_{N,\ell}(x_i) \psi_{N,\ell}(x_j)
  $$
Truncation parameter $k_N$ controls the spectral resolution; appropriate scaling of $h_N$ and $k_N$ guarantees uniform convergence to the true manifold covariance functions as $N \to \infty$.

## 3. Gaussian Process Surrogate and Bayesian Inference

A zero-mean GP prior is placed on $u_N = (u_N(x_1), \ldots, u_N(x_N))^T$ with covariance $K_{ij} = k^{(N)}(x_i, x_j)$. After $t-1$ noisy observations $y_{1:t-1} = (y_1, \ldots, y_{t-1})^T$ at queries $z_{1:t-1}$, the GP posterior at $x \in M_N$ is:

- Posterior mean:
  $$
  \mu_{t-1}(x) = k(x,Z_{1:t-1})^T [K_{Z,Z}+\sigma^2 I]^{-1} y_{1:t-1}
  $$
- Posterior variance:
  $$
  \sigma^2_{t-1}(x) = k(x,x) - k(x,Z_{1:t-1})^T [K_{Z,Z}+\sigma^2 I]^{-1} k(Z_{1:t-1},x)
  $$
where $K_{Z,Z}$ is the covariance submatrix for the chosen queries.

## 4. Upper Confidence Bound (UCB) Acquisition and Misspecification Correction

GGP-UCB selects the next query $z_t$ by maximizing the acquisition function
$$
\alpha_t(x) = \mu_{t-1}(x) + \sqrt{\beta_t}\,\sigma_{t-1}(x)
$$
over $x \in M_N$. The UCB parameter $\beta_t$ is calibrated to ensure high-probability coverage of the true function under surrogate misspecification. For a graph GP prior approximating the manifold GP with sup-norm error $\epsilon_N$,
$$
\beta_t = 2\log\left(\frac{\pi^2 t^2 N}{6\delta}\right) + \left( \frac{\epsilon_N \sqrt{t-1}}{\delta \sigma} \right)^2
$$
guarantees with probability at least $1-\delta$ that the UCB envelope covers both the surrogate and true values.

## 5. Algorithmic Workflow and Computational Costs

The GGP-UCB procedure iterates as follows:
1. Input: point cloud $M_N$, graph parameters ($h_N$, $k_N$, kernel type, smoothness/length-scale), noise $\sigma$, confidence $\delta$, query budget $L$.
2. Initialization: $\mu_0 \equiv 0$, $\sigma_0(x)=k(x,x)$.
3. For $t=1,\ldots,L$,
   - Compute $\beta_t$ as above.
   - Choose $z_t = \arg\max_{x \in M_N} [\mu_{t-1}(x) + \sqrt{\beta_t}\, \sigma_{t-1}(x)]$.
   - Query $y_t = f(z_t) + \eta_t$, $\eta_t \sim \mathcal{N}(0,\sigma^2)$.
   - Update posterior.
4. Output $z_L^* = \arg\max_{z \in \{z_1, \ldots, z_L\}} \mu_L(z)$ (alternatively, report the query with maximal $y_t$ for small $\sigma$).

Computational steps involve:
- Building $\Delta_N$: $O(N^2)$ (sparsifiable).
- Graph Laplacian eigendecomposition: $O(N^3)$, or $O(Nk_N^2)$ for partial spectra.
- GP update (naive): inverting a $t \times t$ matrix per step, for total $O(L^3)$; low-rank GP acceleration is feasible for small $k_N$.

## 6. Theoretical Guarantees and Key Assumptions

Under Assumption 2.1 (manifold is a smooth compact boundaryless $m$-manifold, $M_N$ is an i.i.d. uniform sample), the following are established:
- The sup-norm surrogate error is $\epsilon_N = O(N^{-\gamma}\,\mathrm{polylog}\,N)$ where $\gamma>0$ is determined by $m$, kernel smoothness; for SE kernels, $\gamma \approx 1/(2\alpha_m)$, $\alpha_m \approx m+4$; for Matérn kernels with $s>(3/2)m-1/2$ similar rates hold.
- The main regret bound (Theorem 3.1): with probability at least $1-2\delta-O(N^{-c})$,
  $$
  r_{N,L} = f(z^*) - f(z_L^*) \leq C \left[ \left( \sqrt{2\log(\pi^2L^2N/(6\delta))}/\sqrt{L} + \epsilon_N/(\delta \sigma) \right) \sqrt{k_N \log L} \right]
  $$
- For $L \ll N$ and slowly growing $k_N$ (e.g., $k_N = N^{\xi}\mathrm{polylog}N$), as $L \to \infty$ the simple regret $r_{N,L} \to 0$.

Key assumptions include:
- The true $f$ is a manifold GP restricted to $M_N$, with graph GP surrogate incurring sup-norm error $\epsilon_N$.
- Spectral convergence and associated parameter scaling for $h_N$ and $k_N$.
- $\beta_t$ is properly corrected for surrogate error.

## 7. Empirical Evaluation and Practical Performance

GGP-UCB has been benchmarked across synthetic and applied settings:

- **Unit circle ($m=1$):** GGP-UCB was compared against “oracle” manifold-GP-UCB and a version using marginal-likelihood-estimated (ML) kernel parameters. All variants locate the maximizer using $L \ll N$ queries, with GGP-UCB closely matching oracle performance even when $f$ is not observed from a GP.
- **Artificial 2D surface (“cow”):** On point clouds in $\mathbb{R}^3$ ($N=2000$), GGP-UCB achieves lower simple regret than standard Euclidean GP-UCB, reflecting sensitivity to underlying manifold geometry.
- **Heat-source inverse problem on the sphere ($m=2$):** Given noisy observations of a diffused heat signal over the sphere, GGP-UCB rapidly recovers the source location with high accuracy and few queries, outperforming Euclidean baselines.

These illustrations confirm that GGP-UCB consistently attains vanishing regret with a number of queries far smaller than the sample size $N$, robustly handling both exact and approximate manifold function settings [2210.10962].

Source: https://www.emergentmind.com/topics/graph-gaussian-process-ucb-ggp-ucb