Home / Curated / Algorithms / Gaussian-process moment matching for Gaussian inputs

Gaussian-process moment matching for Gaussian inputs

Algorithms AIMech:GaussianProcessGaussianInputMomentMatching DRAFT

Aliases: Gaussian input Gaussian process moment matching

A Gaussian-process prediction procedure that analytically integrates a fixed squared-exponential posterior over an independently prescribed Gaussian test-input distribution and represents the resulting latent prediction by a Gaussian matching its exact first two moments.

Identity Notes & Concept Scope Compared all 442 pinned AIO IDs, labels and scoped synonyms and 540 native IDs, names and aliases against the proposed ID/name, descriptive alias, uncertain-input and GP-moment terminology: no normalized identity match. Ignored- and hidden-inclusive full-text searches in the recovery, older discovery and original Documents checkouts returned 60, 47 and 46 lines. All ten distinct matched text strings were read in full; they concern pending GP source leads, EP or DropConnect moment approximations, and neighboring derivative GP records. Read all 37 semantic AIO rows, 76 native name/definition neighbors, six full annotated AIO CLASS nodes with incident edges and generated records, and four complete contextual native records. The procedure integrates a prescribed Gaussian test-input law against a fixed squared-exponential GP posterior and matches the resulting first two moments. It is narrower than generic GP regression and is not EP site matching, DropConnect prediction, a derivative-observation method, or the earlier local Taylor approximation. Iterative forecasting uses the procedure with additional Gaussian and temporal-independence approximations. Absence is bounded to the inspected pin and record corpus; no universal uncertain-input-GP alias is asserted.

Claims & Grounded Evidence

5 Claims

Mechanism #mixture-moments

Let X have a prescribed Gaussian distribution independent of the latent GP posterior, conditional on the training data and fixed hyperparameters. Integrating p(f(x)|x,data) over this input law gives a generally non-Gaussian mixture. Its mean is E_X[m(X)] and its variance is E_X[v(X)]+Var_X[m(X)], where m and v are the fixed-input latent posterior mean and variance. Replacing this mixture by a Gaussian with those moments is a distributional approximation, even when both moments are evaluated exactly.

supports Source: Gaussian Process Priors with Uncertain Inputs Application to Multiple-Step Ahead Time Series Forecasting · Locator: Sections 3–3.1, Equations 5–7, PDF p. 3; exact moments, Equations 10–13, PDF p. 4
supports Source: On Simulation and Trajectory Prediction with Gaussian Process Dynamics · Locator: Sections 1–3, especially the dependence distinction on PDF pp. 1–4

Mechanism #squared-exponential-integrals

For the paper’s zero-mean, unit-amplitude squared-exponential GP with Gaussian observation noise, write beta=K_y^{-1}y, k_X for covariances between the training inputs and X, q=E[k_X], and Q=E[k_X k_X^T]. Then the latent predictive mean is q^T beta and variance is 1-tr(K_y^{-1}Q)+beta^T Q beta-(q^T beta)^2. Gaussian integration supplies closed forms for q and Q. For the one-dimensional unit-lengthscale kernel and X~N(mu,s), q_i=exp(-(mu-x_i)^2/(2(1+s)))/sqrt(1+s) and Q_ij=exp(-(x_i-x_j)^2/4-(mu-(x_i+x_j)/2)^2/(1+2s))/sqrt(1+2s), with s denoting input variance. The zero-input-variance limit recovers ordinary fixed-input prediction. These exact moment statements condition on the specified kernel, data and hyperparameters.

supports Source: Gaussian Process Priors with Uncertain Inputs Application to Multiple-Step Ahead Time Series Forecasting · Locator: Sections 2.1–2.2, Equations 1–4, PDF pp. 2–3; Section 3.1, Equations 10–13, PDF p. 4

Application #forecasting-extension

For iterative autoregressive forecasting, the source carries a joint Gaussian approximation to the lagged outputs, computes the next scalar output moments and its cross-covariances with the existing lags, and shifts the resulting covariance block. Full lag covariance is part of that construction; independently propagating only marginal variances is a different approximation. When the predicted quantity is a future noisy observation, add its independent observation-noise variance once to the new output diagonal. Gaussian moment closure and treating the current input distribution as independent of the uncertain GP function are additional assumptions; retained lag covariance alone does not make the complete shared-function trajectory posterior exact.

supports Source: Gaussian Process Priors with Uncertain Inputs Application to Multiple-Step Ahead Time Series Forecasting · Locator: Section 4, iterative forecasting and the unnumbered lagged covariance construction, PDF pp. 4–5
supports Source: On Simulation and Trajectory Prediction with Gaussian Process Dynamics · Locator: Introduction and Sections 2–3, PDF pp. 1–4

Scope #taylor-boundary

The earlier first-order posterior-mean and second-order posterior-variance Taylor construction approximates the same total-moment integrals locally; it is not the closed-form squared-exponential Gaussian integral. The technical report explicitly leaves exact evaluation of those moments to future work. Curator boundary calculations show why the distinction matters: with the unit kernel, an exact observation f(0)=0, input mean 1 and input variance 2, the local Taylor variance is 1-3/e<0, whereas the exact mixture variance is 1-exp(-1/5)/sqrt(5)>0. This deliberately nonlocal example is not a reproduction of the paper’s benchmark or a rejection of the approximation within its small-uncertainty regime.

supports Source: Gaussian Process Priors with Uncertain Inputs Application to Multiple-Step Ahead Time Series Forecasting · Locator: Section 3.1, Equations 8–9, PDF p. 3
supports Source: Gaussian Process priors with Uncertain Inputs: Multiple-Step-Ahead Prediction · Locator: Sections 3.3–3.4, Equations 16–22, PDF pp. 8–9; conclusions, PDF p. 17

Limitation #shared-function-limit

In a trajectory generated by a single uncertain GP dynamics function, later inputs generally depend on that same function. Repeated independent-input moment calculations do not preserve this dependence. Hewing and colleagues explicitly identify this issue in established GP trajectory prediction, including the Girard construction. A curator linear-kernel illustration uses f(x)=A*x with A~N(0,1) and x_0=1: reusing A gives x_2=A^2 with mean 1 and variance 2, while independently redrawing A at each step gives a product with mean 0 and variance 1. This illustrates a change in the modeled law, not a squared-exponential benchmark replication or a universal direction of variance error.

supports Source: On Simulation and Trajectory Prediction with Gaussian Process Dynamics · Locator: Introduction, persistent-versus-independent parameter examples and Gaussian trajectory discussion, PDF pp. 1–4

⚠️ Curated Limitations & Negative Results

Closed-form moments here concern the zero-mean squared-exponential GP and a prescribed Gaussian test-input distribution with fixed hyperparameters. Other kernels, non-Gaussian inputs, uncertain training locations or hyperparameter integration require further treatment. The method does not automatically estimate the input distribution or solve a general errors-in-variables problem.
Exact first and second moments do not make the mixture Gaussian or its Gaussian credible intervals calibrated. A unit-kernel f(0)=0 example in the curator calculator has a fourth moment different from that of the matched Gaussian. Input integration can also lower variance relative to evaluating the GP only at the input mean; uncertainty is not guaranteed to increase at every step.
Iterative lag covariance and observation noise must be handled consistently. A singular Gaussian input covariance can occur when initial lags are exact; Gaussian expectations have well-defined degenerate limits, but an implementation must not blindly invert that input covariance. Noise in training K_y and future observation noise have different roles.
The earlier Taylor report is contextual evidence for a distinct local approximation. Its notation changes the meaning of kernel width parameters between sections. The proceedings text also contains a lengthscale irrelevance statement inconsistent with its displayed kernel and ambiguous trace/parenthesis typography; the record uses the total-variance identity and directly checked scalar integrals instead of reproducing those ambiguities.
The source’s Mackey–Glass and pH illustrations do not establish universal forecasting superiority or calibration. The proceedings pH test count is printed as an unresolved placeholder; the earlier technical report has different preprocessing and results, so its count is not silently transferred. No comparative numerical performance claim is curated.
No author implementation, fitted GP, stochastic trajectory simulation or benchmark was executed. Curator analytic and numerical-quadrature checks establish only the displayed mathematical examples. The record remains DRAFT after author self-review; correlation-aware trajectory methods require their own identity and implementation-scope assessments.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:RegressionAnalysis` Regression Analysis	related_match	The algorithm integrates a fitted GP response prediction over uncertain inputs. RegressionAnalysis names the task context rather than this particular inference procedure; no exact identity or upstream subclass is asserted.
`AIO:TimeSeriesForecasting` Time Series Forecasting	related_match	The primary method is developed and evaluated as a component of iterative multiple-step time-series forecasting. This explicit application relationship does not identify all forecasting with the moment-matching procedure or turn the algorithm into a task.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`girard2002`	Gaussian Process Priors with Uncertain Inputs Application to Multiple-Step Ahead Time Series Forecasting Agathe Girard, Carl Edward Rasmussen, Joaquin Quiñonero Candela, Roderick Murray-Smith	primary paper full_text	ID: https://papers.nips.cc/paper_files/paper/2002/hash/f3ac63c91272f19ce97c7397825cc15f-Abstract.html sha256:b8b19f828d8d…
`girard-tr2002`	Gaussian Process priors with Uncertain Inputs: Multiple-Step-Ahead Prediction Agathe Girard, Carl Edward Rasmussen, Roderick Murray-Smith	author technical report full_text	ID: TR-2002-119 sha256:f2e78afd674d…
`hewing2020`	On Simulation and Trajectory Prediction with Gaussian Process Dynamics Lukas Hewing, Elena Arcari, Lukas P. Fröhlich, Melanie N. Zeilinger	primary paper full_text	ID: https://proceedings.mlr.press/v120/hewing20a.html sha256:e03bc1f12364…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-12T21:02:07Z

Updated: 2026-10-12T21:08:34Z

Source File: data/curated/GaussianProcessGaussianInputMomentMatching.yaml