Search NASA⌕ Search

SEARCH · Search NASA

Results for “Gaussian Process Models”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,081 records · Page 60

Gravitational waves at interferometer scales and primordial black holes in axion inflation

We study the prospects of detection at terrestrial and space interferometers, as well as at pulsar timing array experiments, of a stochastic gravitational wave background which can be produced in models of axion inflation. This potential signal, and the development of these experiments, open a new window on inflation on scales much smaller than those currently probed with Cosmic Microwave Background and Large Scale Structure measurements. The sourced signal generated in axion inflation is an ideal candidate for such searches, since it naturally grows at small scales, and it has specific properties (chirality and non-gaussianity) that can distinguish it from an astrophysical background. We study under which conditions such a signal can be produced at an observable level, without the simultaneous overproduction of scalar perturbations in excess of what is allowed by the primordial black hole limits. We also explore the possibility that scalar perturbations generated in a modified version of this model may provide a distribution of primordial black holes compatible with the current bounds, that can act as a seeds of the present black holes in the universe.

79 ASTRONOMY AND ASTROPHYSICS↗

Evidence for X-Ray Emission in Excess to the Jet-afterglow Decay 3.5 yr after the Binary Neutron Star Merger GW 170817: A New Emission Component

For the first ~3 yrs after the binary neutron star merger event GW 170817, the radio and X-ray radiation has been dominated by emission from a structured relativistic off-axis jet propagating into a low-density medium with n < 0.01 cm -3 . We report on observational evidence for an excess of X-ray emission at δt > 900 days after the merger. With L x ≈ 5 × 10 38 erg s -1 at 1234 days, the recently detected X-ray emission represents a ≥ 3.2σ (Gaussian equivalent) deviation from the universal post-jet-break model that best fits the multiwavelength afterglow at earlier times. In the context of JetFit afterglow models, current data represent a departure with statistical significance ≥ 3.1σ, depending on the fireball collimation, with the most realistic models showing excesses at the level of ≥ 3.7σ. A lack of detectable 3 GHz radio emission suggests a harder broadband spectrum than the jet afterglow. These properties are consistent with the emergence of a new emission component such as synchrotron radiation from a mildly relativistic shock generated by the expanding merger ejecta, i.e., a kilonova afterglow. In this context, we present a set of ab initio numerical relativity binary neutron star (BNS) merger simulations that show that an X-ray excess supports the presence of a high-velocity tail in the merger ejecta, and argues against the prompt collapse of the merger remnant into a black hole. Radiation from accretion processes on the compact-object remnant represents a viable alternative. Neither a kilonova afterglow nor accretion-powered emission have been observed before, as detections of BNS mergers at this phase of evolution are unprecedented

79 ASTRONOMY AND ASTROPHYSICS↗

Intra-beam scattering and its application to ERL

Treatment of Coulomb collisions within the beam requires consideration of both large and small angle scattering. Such collisions lead to the Touschek effect and Intrabeam Scattering (IBS). The Touschek effect refers to particle loss as a result of a single collision, where only transfer from the transverse direction into longitudinal plays a role. It is important to consider this effect for ERL design to have an appropriate choice of collimation system. The IBS is a diffusion process which leads to changes of beam distribution but does not necessarily result in a beam loss. Evaluation of IBS in ERLs, where beam distribution is non-Gaussian, requires special treatment. Here we describe the IBS and Touschek effects with application to ERLs. In circular accelerators both the Touschek effect and IBS were found important. The generalized formulas for Touschek calculations are available and are already being used in advanced tracking simulations of several ERL-based projects. The IBS (which is diffusion due to multiple Coulomb scattering) is not expected to cause any significant effect on beam distribution in ERLs, unless one considers very long transport of high-brightness beams at low energies. Both large and small-angle Coulomb scattering can contribute to halo formation in future ERLs with high-brightness beams, as follows from simple order-of-magnitude estimates. In this report, a test comparison between 'local' and 'sliced' IBS models within the BET ACOOL code was presented for an illustrative ERL distribution. We also presented accumulated current loss distribution due to Touschek scattering for design parameters of ERL proposed for the eRHIC project, as well as scaling for multi-pass ERLs.

43 PARTICLE ACCELERATORS↗

Posterior predictive modeling using multi-scale stochastic inverse parameter estimates.

Multi-scale binary permeability field estimation from static and dynamic data is completed using Markov Chain Monte Carlo (MCMC) sampling. The binary permeability field is defined as high permeability inclusions within a lower permeability matrix. Static data are obtained as measurements of permeability with support consistent to the coarse scale discretization. Dynamic data are advective travel times along streamlines calculated through a fine-scale field and averaged for each observation point at the coarse scale. Parameters estimated at the coarse scale (30 x 20 grid) are the spatially varying proportion of the high permeability phase and the inclusion length and aspect ratio of the high permeability inclusions. From the non-parametric, posterior distributions estimated for these parameters, a recently developed sub-grid algorithm is employed to create an ensemble of realizations representing the fine-scale (3000 x 2000), binary permeability field. Each fine-scale ensemble member is instantiated by convolution of an uncorrelated multiGaussian random field with a Gaussian kernel defined by the estimated inclusion length and aspect ratio. Since the multiGaussian random field is itself a realization of a stochastic process, the procedure for generating fine-scale binary permeability field realizations is also stochastic. Two different methods are hypothesized to perform posterior predictive tests. Different mechanisms for combining multi Gaussian random fields with kernels defined from the MCMC sampling are examined. Posterior predictive accuracy of the estimated parameters is assessed against a simulated ground truth for predictions at both the coarse scale (effective permeabilities) and at the fine scale (advective travel time distributions). The two techniques for conducting posterior predictive tests are compared by their ability to recover the static and dynamic data. The skill of the inference and the method for generating fine-scale binary permeability fields are evaluated through flow calculations on the resulting fields using fine-scale realizations and comparing them against results obtained with the ground truth fine-scale and coarse-scale permeability fields.

58 GEOSCIENCES↗

Improving Boundary-layer Turbulence and Cloud Processes in CAM with a Higher-order Turbulence Closure Scheme and ASR Measurement

The intermediately-prognostic higher-order turbulence closure (IPHOC) introduces a joint double-Gaussian distribution of liquid water potential temperature (θ l ), total water mixing ratio (q t ), and vertical velocity (w) to represent any skewed turbulence circulation. The distribution is inferred from the first-, second-, and third-order moments of the variables given above, and is used to diagnose cloud fraction and gridmean liquid water mixing ratio, as well as the buoyancy term and fourth-order terms in the equations describing the evolution of the second- and third-order moments. Only three third-order moments, i.e., the triple moments of θ l , q t , and w, are predicted in IPHOC.

58 GEOSCIENCES↗

Optimization and Evaluation of Stochastic Unified Convection Using Single-Column Model Simulations at Multiple Observation Sites

We extend the previously developed stochastic unified convection scheme (UNICON) for shallow convection to deep convection by parameterizing the impact of mesoscale organized flow on updraft properties. The extended stochastic UNICON parameterizes thermodynamic properties of updrafts at the near-surface as a multivariate Gaussian distribution, where the variances of the distribution are the summation of variances from non-organized turbulence and mesoscale organized flow. The distribution of updraft radius is parameterized as a power-law distribution with a scale break which is parameterized as a linear function of the strength of mesoscale organized flow. The proposed parameterization is validated using a series of large-eddy simulations of deep convection. The free parameters introduced in the formulation of stochastic UNICON are optimized using 10 cases of single-column model simulations over the ocean. Stochastic UNICON with the optimized parameters significantly reduces the biases of thermodynamic profiles and surface precipitation rates simulated in the original UNICON for tropical convection cases. The simulation of the variation in anomalies of temperature and moisture associated with the Madden-Julian oscillation is also improved. The overall improvements in simulated thermodynamic profiles are found to be due to the increased heating and drying tendencies by convective processes in stochastic UNICON. An additional simulation of an idealized deep convection case shows that stochastic UNICON produces enhanced cloud variabilities with dependency on updraft radius, indicating its ability to represent the coexistence of shallow and deep convection.

54 ENVIRONMENTAL SCIENCES↗

Earth System Reanalysis in Support of Climate Model Improvements

Recent climate model developments, established through increased model resolution, have led to substantial improvements in model simulations of the time-evolving, coupled Earth system and its subcomponents. However, regardless of resolution, climate models will always produce climate features and variability that differ from the real world and will be prone to biases. This is due to many remaining uncertainties, such as in parametric and structural model uncertainty, in the initial conditions prescribed, and in the prescribed (scenario) forcing which varies on decadal to centennial timescales. Further model improvements are expected to arise specifically from improved representation of physical processes realized through model-data fusion. This will create an unprecedented opportunity to better exploit a large array of Earth observations, from in situ measurements to weather radars and satellite observations, as the resolved scales of the models approach those of the observations. For this, climate DA will be the central tool to bring models and observations into consistency, by improving initial conditions, inferring uncertain model parameters and structure, and quantifying uncertainty. Generally, there will be advantages and complementarities of adjoint-based smoother approaches, ensemble-based filter approaches, or new ML-inspired approaches. Yet, the ever-increasing model resolution will present growing challenges arising from computational cost, calling for new ways of performing data assimilation and model optimization. Using the complementarity in a hybrid approach, blending tools and concepts from variational, ensemble and ML methods might be what is required in the future. In this context ML could be important to handle non-linear responses, and to better approximate non-Gaussian distributions.

54 ENVIRONMENTAL SCIENCES↗

Improving qubit readout with hidden Markov models

We demonstrate the application of pattern recognition algorithms via hidden Markov models (HMM) for qubit readout. This scheme provides a state-path trajectory approach capable of detecting qubit-state transitions and makes for a robust classification scheme with higher starting-state assignment fidelity than when compared to a multivariate Gaussian or a support vector machine scheme. Therefore, the method also eliminates the qubit-dependent readout time optimization requirement in current schemes. Using a HMM state discriminator we estimate fidelities reaching the ideal limit. Unsupervised learning gives access to transition matrix, priors, and IQ distributions, providing a toolbox for studying qubit-state dynamics during strong projective readout.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

An Iterative Response-Surface-Based Approach for Chance-Constrained AC Optimal Power Flow Considering Dependent Uncertainty

A modern power system is characterized by a stochastic variation of the loads and an increasing penetration of renewable energy generation, which results in large uncertainties in its states. These uncertainties bring formidable challenges to the power system planning and operation process. To address these challenges, we propose a cost-effective, iterative response-surface-based approach for the chance-constrained AC optimal power-flow problem that aims to ensure the secure operation of the power systems considering dependent uncertainties. Starting from a stochastic-sampling-based framework, we first utilize the copula theory to simulate the dependence among multivariate uncertain inputs. Then, to reduce the prohibitive computational time required in the traditional Monte-Carlo method, we propose, instead of using the original complicated power-system model, to rely on a polynomial-chaos-based response surface. This response surface allows us to efficiently evaluate the time-consuming power-system model at arbitrary distributed sampled values with a negligible computational cost. This further enables us to efficiently conduct an online stochastic testing for the system states that not only screens out the statistical active constraints, but also assists in a better design of the tightened bounds without using any Gaussian or symmetric assumption. Finally, an iterative procedure is executed to fine-tune the optimal solution that better satisfies a predefined probability. The simulations conducted in multiple test systems demonstrate the excellent performance of the proposed method.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Intrinsic transverse momentum and parton correlations from dynamical chiral symmetry breaking

The dynamical breaking of chiral symmetry in QCD is caused by nonperturbative interactions on a distance scale rho ~ 0.3 fm, much smaller than the typical hadronic size R ~ 1 fm. These short-distance interactions influence the intrinsic transverse momentum distributions of partons and their correlations at a low normalization point. We study this phenomenon in an effective description of the low-energy dynamics in terms of chiral constituent quark degrees of freedom, which refers to the large-N_c limit of QCD. The nucleon is obtained as a system of constituent quarks and antiquarks moving in a self-consistent classical chiral field (relativistic mean-field approximation, or chiral quark-soliton model). The calculated transverse momentum distributions of constituent quarks and antiquarks are matched with QCD quarks, antiquarks and gluons at the chiral symmetry--breaking scale rho^{-2}. We find that the transverse momentum distribution of valence quarks is localized at p_T^2 ~ R^{-2} and roughly of Gaussian shape. The distribution of unpolarized sea quarks exhibits a would-be power-like tail ~1/p_T^2 extending up to the chiral symmetry-breaking scale. Similar behavior is observed in the flavor-nonsinglet polarized sea. The high-momentum tails are the result of short-range correlations between sea quarks in the nucleon&#x27;s light-cone wave function, which are analogous to short-range NN correlations in nuclei. We show that the nucleon&#x27;s light-cone wave function contains correlated pairs of transverse size rho << R with scalar-isoscalar (Sigma) and pseudoscalar-isovector (Pi) quantum numbers, whose internal wave functions have a distinctive spin structure and become identical at p_T^2 ~ rho^{-2} (restoration of chiral symmetry). These features are model-independent and represent an effect of dynamical chiral symmetry breaking on the nucleon&#x27;s partonic structure. Our results have numerous implications for the transverse momentum distributions of particles produced in hard scattering processes. Under certain conditions the nonperturbative parton correlations predicted here could be observed in particle correlations between the current and target fragmentation regions of deep-inelastic scattering.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Group Additivity Determination for Oxygenates, Oxonium Ions, and Oxygen-Containing Carbenium Ions

Bio-oil produced from biomass fast pyrolysis often requires catalytic upgrading to remove oxygen and acidic species over zeolite catalysts. The elementary reactions in the mechanism for this process involve carbenium and oxonium ions. In order to develop a detailed kinetic model for the catalytic upgrading of biomass, rate constants are required for these elementary reactions. The parameters in the Arrhenius equation can be related to thermodynamic properties through structure-reactivity relationships, such as the Evans-Polanyi relationship. For this relationship, enthalpies of formation of each species are required, which can be reasonably estimated using group additivity. However, the literature previously lacked group additivity values for oxygenates, oxonium ions, and oxygen-containing carbenium ions. In this work, 71 group additivity values for these types of groups were regressed, 65 of which had not been reported previously and six of which were newly estimated based on regression in the context of the 65 new groups. Heats of formation based on atomization enthalpy calculations for a set of reference molecules and isodesmic reactions for a small set of larger species for which experimental data was available were used to demonstrate the accuracy of the Gaussian-4 quantum mechanical method in estimating enthalpies of formation for species involving the moieties of interest. Isodesmic reactions for a total of 195 species were constructed from the reference molecules to calculate enthalpies of formation that were used to regress the group additivity values. The results showed an average deviation of 1.95 kcal/mol between the values calculated from Gaussian-4 and isodesmic reactions versus those calculated from the group additivity values that were newly regressed. Importantly, the new groups enhance the database for group additivity values, especially those involving oxonium ions.

09 BIOMASS FUELS↗

Bayesian Abel Inversion in Quantitative X-Ray Radiography

A common image formation process in high-energy X-ray radiography is to have a pulsed power source that emits X-rays through a scene, a scintillator that absorbs X-rays and uoresces in the visible spectrum in response to the absorbed photons, and a CCD camera that images the visible light emitted from the scintillator. The intensity image is related to areal density, and, for an object that is radially symmetric about a central axis, the Abel transform then gives the object's volumetric density. Two of the primary drawbacks to classical variational methods for Abel inversion are their sensitivity to the type and scale of regularization chosen and the lack of natural methods for quantifying the uncertainties associated with the reconstructions. In this work we cast the Abel inversion problem within a statistical framework in order to compute volumetric object densities from X-ray radiographs and to quantify uncertainties in the reconstruction. A hierarchical Bayesian model is developed with a likelihood based on a Gaussian noise model and with priors placed on the unknown density pro le, the data precision matrix, and two scale parameters. This allows the data to drive the localization of features in the reconstruction and results in a joint posterior distribution for the unknown density pro le, the prior parameters, and the spatial structure of the precision matrix. Results of the density reconstructions and pointwise uncertainty estimates are presented for both synthetic signals and real data from a U.S. Department of Energy X-ray imaging facility.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Particle filtering with path sampling and an application to a bimodal ocean current model

This paper introduces a recursive particle filtering algorithm designed to filter high dimensional systems with complicated non-linear and non-Gaussian effects. The method incorporates a parallel marginalization (PMMC) step in conjunction with the hybrid Monte Carlo (HMC) scheme to improve samples generated by standard particle filters. Parallel marginalization is an efficient Markov chain Monte Carlo (MCMC) strategy that uses lower dimensional approximate marginal distributions of the target distribution to accelerate equilibration. As a validation the algorithm is tested on a 2516 dimensional, bimodal, stochastic model motivated by the Kuroshio current that runs along the Japanese coast. The results of this test indicate that the method is an attractive alternative for problems that require the generality of a particle filter but have been inaccessible due to the limitations of standard particle filtering strategies.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Commissioning dose computation models for spot scanning proton beams in water for a commercially available treatment planning system

Purpose: To present our method and experience in commissioning dose models in water for spot scanning proton therapy in a commercial treatment planning system (TPS). Methods: The input data required by the TPS included in-air transverse profiles and integral depth doses (IDDs). All input data were obtained from Monte Carlo (MC) simulations that had been validated by measurements. MC-generated IDDs were converted to units of Gy mm{sup 2}/MU using the measured IDDs at a depth of 2 cm employing the largest commercially available parallel-plate ionization chamber. The sensitive area of the chamber was insufficient to fully encompass the entire lateral dose deposited at depth by a pencil beam (spot). To correct for the detector size, correction factors as a function of proton energy were defined and determined using MC. The fluence of individual spots was initially modeled as a single Gaussian (SG) function and later as a double Gaussian (DG) function. The DG fluence model was introduced to account for the spot fluence due to contributions of large angle scattering from the devices within the scanning nozzle, especially from the spot profile monitor. To validate the DG fluence model, we compared calculations and measurements, including doses at the center of spread out Bragg peaks (SOBPs) as a function of nominal field size, range, and SOBP width, lateral dose profiles, and depth doses for different widths of SOBP. Dose models were validated extensively with patient treatment field-specific measurements. Results: We demonstrated that the DG fluence model is necessary for predicting the field size dependence of dose distributions. With this model, the calculated doses at the center of SOBPs as a function of nominal field size, range, and SOBP width, lateral dose profiles and depth doses for rectangular target volumes agreed well with respective measured values. With the DG fluence model for our scanning proton beam line, we successfully treated more than 500 patients from March 2010 through June 2012 with acceptable agreement between TPS calculated and measured dose distributions. However, the current dose model still has limitations in predicting field size dependence of doses at some intermediate depths of proton beams with high energies. Conclusions: We have commissioned a DG fluence model for clinical use. It is demonstrated that the DG fluence model is significantly more accurate than the SG fluence model. However, some deficiencies in modeling the low-dose envelope in the current dose algorithm still exist. Further improvements to the current dose algorithm are needed. The method presented here should be useful for commissioning pencil beam dose algorithms in new versions of TPS in the future.

60 APPLIED LIFE SCIENCES↗

TU-AB-202-11: Tumor Segmentation by Fusion of Multi-Tracer PET Images Using Copula Based Statistical Methods

Purpose: Multi-tracer PET imaging is getting more attention in radiotherapy by providing additional tumor volume information such as glucose and oxygenation. However, automatic PET-based tumor segmentation is still a very challenging problem. We propose a statistical fusion approach to joint segment the sub-area of tumors from the two tracers FDG and FMISO PET images. Methods: Non-standardized Gamma distributions are convenient to model intensity distributions in PET. As a serious correlation exists in multi-tracer PET images, we proposed a new fusion method based on copula which is capable to represent dependency between different tracers. The Hidden Markov Field (HMF) model is used to represent spatial relationship between PET image voxels and statistical dynamics of intensities for each modality. Real PET images of five patients with FDG and FMISO are used to evaluate quantitatively and qualitatively our method. A comparison between individual and multi-tracer segmentations was conducted to show advantages of the proposed fusion method. Results: The segmentation results show that fusion with Gaussian copula can receive high Dice coefficient of 0.84 compared to that of 0.54 and 0.3 of monomodal segmentation results based on individual segmentation of FDG and FMISO PET images. In addition, high correlation coefficients (0.75 to 0.91) for the Gaussian copula for all five testing patients indicates the dependency between tumor regions in the multi-tracer PET images. Conclusion: This study shows that using multi-tracer PET imaging can efficiently improve the segmentation of tumor region where hypoxia and glucidic consumption are present at the same time. Introduction of copulas for modeling the dependency between two tracers can simultaneously take into account information from both tracers and deal with two pathological phenomena. Future work will be to consider other families of copula such as spherical and archimedian copulas, and to eliminate partial volume effect by considering dependency between neighboring voxels.

60 APPLIED LIFE SCIENCES↗

Kernel principal component analysis for stochastic input model generation

Highlights: {yields} KPCA is used to construct a reduced order stochastic model of permeability. {yields} A new approach is proposed to solve the pre-image problem in KPCA. {yields} Polynomial chaos is used to provide a parametric stochastic input model. {yields} Flow in porous media with channelized permeability is considered. - Abstract: Stochastic analysis of random heterogeneous media provides useful information only if realistic input models of the material property variations are used. These input models are often constructed from a set of experimental samples of the underlying random field. To this end, the Karhunen-Loeve (K-L) expansion, also known as principal component analysis (PCA), is the most popular model reduction method due to its uniform mean-square convergence. However, it only projects the samples onto an optimal linear subspace, which results in an unreasonable representation of the original data if they are non-linearly related to each other. In other words, it only preserves the first-order (mean) and second-order statistics (covariance) of a random field, which is insufficient for reproducing complex structures. This paper applies kernel principal component analysis (KPCA) to construct a reduced-order stochastic input model for the material property variation in heterogeneous media. KPCA can be considered as a nonlinear version of PCA. Through use of kernel functions, KPCA further enables the preservation of higher-order statistics of the random field, instead of just two-point statistics as in the standard Karhunen-Loeve (K-L) expansion. Thus, this method can model non-Gaussian, non-stationary random fields. In this work, we also propose a new approach to solve the pre-image problem involved in KPCA. In addition, polynomial chaos (PC) expansion is used to represent the random coefficients in KPCA which provides a parametric stochastic input model. Thus, realizations, which are statistically consistent with the experimental data, can be generated in an efficient way. We showcase the methodology by constructing a low-dimensional stochastic input model to represent channelized permeability in porous media.

97 MATHEMATICS AND COMPUTING↗

Alfvén wave propagation through a moderate-amplitude transverse inhomogeneity in a magnetized plasma

Parallel propagation of a plane Alfvén wave in a moderate-amplitude Gaussian-shaped transverse inhomogeneity is studied numerically using a fluid model retaining low-frequency kinetic effects. It is shown that in such a situation, common in the solar wind where elongated pressure-balanced structures are frequently observed, phase mixing competes with wave focusing, in contrast with coronal loops or auroral regions where sharp gradients present at the edges of the inhomogeneities make phase mixing dominant. Some understanding about this competition is provided by a model based on an envelope formalism. Depending on the magnitude of the Alfvén wavelength and of the inhomogeneity transverse scale relative to the ion inertial length, various regimes can develop, ranging from the formation of localized gradients when phase mixing dominates to the development of an intense magnetic filament when focusing is stronger, with a different efficiency for the generation of magnetosonic and kinetic Alfvén waves. Electron parallel heating and parallel electric field generation are also briefly discussed.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

Exact Gaussian processes for massive datasets via non-stationary sparsity-discovering kernels

Abstract A Gaussian Process (GP) is a prominent mathematical framework for stochastic function approximation in science and engineering applications. Its success is largely attributed to the GP’s analytical tractability, robustness, and natural inclusion of uncertainty quantification. Unfortunately, the use of exact GPs is prohibitively expensive for large datasets due to their unfavorable numerical complexity of $$O(N^3)$$ O ( N 3 ) in computation and $$O(N^2)$$ O ( N 2 ) in storage. All existing methods addressing this issue utilize some form of approximation—usually considering subsets of the full dataset or finding representative pseudo-points that render the covariance matrix well-structured and sparse. These approximate methods can lead to inaccuracies in function approximations and often limit the user’s flexibility in designing expressive kernels. Instead of inducing sparsity via data-point geometry and structure, we propose to take advantage of naturally-occurring sparsity by allowing the kernel to discover—instead of induce—sparse structure. The premise of this paper is that the data sets and physical processes modeled by GPs often exhibit natural or implicit sparsities, but commonly-used kernels do not allow us to exploit such sparsity. The core concept of exact, and at the same time sparse GPs relies on kernel definitions that provide enough flexibility to learn and encode not only non-zero but also zero covariances. This principle of ultra-flexible, compactly-supported, and non-stationary kernels, combined with HPC and constrained optimization, lets us scale exact GPs well beyond 5 million data points.

97 MATHEMATICS AND COMPUTING↗