Active-subspace learning reveals dominant parameter directions in nuclear observables
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Nuclear Theory. The day in nuclear theory: structure, reactions and the models behind them.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Active-subspace learning reveals dominant parameter directions in nuclear observables".
Mira: Here's the gist of the paper. Nuclear observables computed using ab initio methods can depend on many low-energy constants (LECs) of the underlying chiral effective field theory interactions, while being sensitive to only a few LEC combinations.
Kai: We use active-subspace learning to identify these dominant parameter directions directly from many-body calculations.
Mira: For ground-state energies and charge radii of oxygen isotopes, we find that more than 99% of the gradient-based variation is concentrated along a single direction for each observable in the 17-dimensional LEC space.
Kai: Comparing these directions across nuclei and observables quantifies similarities in their LEC dependence.
Paper summary: Kai: So, to recap this paper, they're showing that when you look at nuclear observables calculated using ab initio methods, they depend on many low-energy constants of the underlying chiral effective field theory interactions but are sensitive to only a few LEC combinations.
Mira: The thesis is that they use active-subspace learning to identify these dominant parameter directions directly from the many-body calculations. They find that for ground-state energies and charge radii of oxygen isotopes, more than ninety-nine percent of the gradient based variation is concentrated along a single direction for each observable in the seventeen dimensional LEC space.
Lev: That process involves taking those low energy constants, standardizing them, and then using an eigenvalue decomposition on a matrix that captures how those gradients relate to each other.
Kai: They found that they can partition the eigenvalues and eigenvectors into active and inactive subspaces where the function varies significantly in the active directions y in R n, while it varies only weakly in the inactive directions z in R d minus n.
Mira: The key insight is that you can consider these rotated coordinates, these active directions y, where things change a lot, instead of working with all seventeen original dimensions.
Lev: They even found that this rotation can improve the performance of Gaussian process emulators when using kernels with diagonal length scale matrices because it aligns the coordinates with the dominant directions from gradient based variation.
Kai: And they showed that this technique can yield derivative based global sensitivities without needing to build a global emulator, which saves you a lot of overhead on your side.
Conclusion: Mira: Looking at the title, "Active-subspace learning reveals dominant parameter directions in nuclear observables," it really highlights how high dimensional spaces can be simplified when you understand what's actually driving the physics.
Kai: It suggests that for complex interactions like chiral effective field theory, instead of needing to map out every single interaction parameter individually, you can focus your effort on these few dominant combinations.
Lev: From an engineering standpoint, this means if you are building a model or running simulations on real hardware, knowing those active directions lets you focus your computational budget on the most informative regions of the parameter space instead of wasting time exploring directions where nothing changes much.
Mira: And what's really interesting is that they showed these dominant directions for the oxygen isotopes are strongly aligned when you compare energies and radii across different isotopes.
Kai: That alignment suggests a universal correlation in how these nuclear forces behave, which connects back to some of those correlations we see in smaller few body systems.
Lev: So, the implication is that we might not need to treat every single LEC as an independent variable when trying to understand these kinds of many body problems; focusing on those active subspaces is a way to get the most information out efficiently.
T. Plies, M. Companys Franzke, A. Schwenk
Technische University of Darmstadt · ExtreMe Matter Institute EMMI · Max Planck Institute for Nuclear Physics
nucl-th
Submitted: 2026-11-02
Updated: 2026-11-02
Comments: 10 pages, 8 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 99/100
The gist: Nuclear observables computed using ab initio methods can depend on many low-energy constants (LECs) of the underlying chiral effective field theory interactions…
Key concepts
- Active Subspace Learning (ASL)
- ASL is a method used to identify the most important directions in a high-dimensional parameter space by analyzing the gradients of the function. It finds low-dimensional subspaces where the function's variation is most significant, effectively reducing complexity while retaining essential information about how observables change.
- Low-Energy Constants (LECs)
- LECs are fundamental parameters within chiral effective field theory that describe the interactions between nucleons at low energies. The paper investigates how nuclear observables depend on these many LECs, finding that they are governed by only a few dominant combinations rather than all of them.
- Gaussian Process (GP) Emulators
- A GP emulator is a statistical tool used to approximate complex functions based on training data. By rotating the input parameter space into the learned active subspace directions, the paper showed that these emulators become vastly more accurate, reducing error by orders of magnitude for predicting nuclear properties.
- Dominant Active Subspace Direction
- This is a specific direction in the high-dimensional LEC space where an observable's variation is most pronounced. The analysis found that for many observables (like oxygen isotope energies), the response is highly concentrated along just one such direction, indicating a simple structure to the parameter dependence.
Terminology
Summary
Nuclear observables computed using ab initio methods can depend on many low-energy constants (LECs) of the underlying chiral effective field theory interactions, while being sensitive to only a few LEC combinations. p. 0 We use active-subspace learning to identify these dominant parameter directions directly from many-body calculations. p. 0 For ground-state energies and charge radii of oxygen isotopes, we find that more than 99% of the gradient-based variation is concentrated along a single direction for each observable in the 17-dimensional LEC space. p. 0 Comparing these directions across nuclei and observables quantifies similarities in their LEC dependence. p. 0 The same analysis yields derivative-based global sensitivities without requiring a global emulator. p. 0 Active subspace directions can also serve as an efficient representation for emulation, greatly improving Gaussian-process emulators especially for broad LEC variations. p. 0
Identifying Dominant Parameter Structure
The core of the active subspace learning (ASL) method involves characterizing the response through gradients with respect to the input parameters and thereby identifying combinations of LECs that dominate an observable over the sampled full domain. The process begins by standardizing each LEC to zero mean and unit variance over the sampled ensemble, denoted by x. p. 0 The central element of ASL is the matrix C given by the outer product of gradients of f C = E[∇f(x)∇f(x)⊤], (1) where E denotes the expectation value over the considered parameter ensemble. p. 0 The eigenvalue decomposition of C is written as C = WΛW⊤, with Λ = diag(λ1, λ2,. p. 0, d). The eigenvectors W describe a rotation in R d whereas the eigenvalues λi measure the sensitivity of the function in each direction of W. p. 0
Quantifying Dominance and Universality
The analysis applied to ground-state energies and charge radii of oxygen isotopes revealed that each observable's response is concentrated primarily along a single observable-specific direction. Comparing these directions across nuclei and observables provides a direct measure of how similarly they probe the underlying nuclear forces. p. 0 Furthermore, the dominant active-subspace direction y1 for E(16O) shows that the ground-state energies for 16O, 18O, and 24O are shown as a function of this coordinate in Fig. p. 1 2. The data projected against the dominant active-subspace coordinate looks remarkably linear. p. 1 This indicates that the dominant directions of LEC dependence across the different isotopes are strongly aligned. p. 2
Enhancing Emulator Performance
The active subspace transformation substantially improves Gaussian-process (GP) emulation by rotating the LEC space to the active-subspace directions. This is demonstrated for Hartree-Fock calculations of 48Ca, where the emulator error is improved with active-subspace learning by more than two orders of magnitude. p. 0 The transformation aligns coordinates with the dominant directions from gradient-based variation, which can improve performance using kernels with diagonal length-scale matrices. p. 1
Sensitivity Analysis and Conclusion
Derivative-based global sensitivity measures (DGSM) can be computed directly from the calculated ensemble without requiring a global emulator. When the active subspace is effectively one-dimensional, the parameter sensitivities νi are approximately proportional to the squared components of the leading active direction w1, i. p. 2e., to each parameter’s contribution to the first active-subspace coordinate y1. p. 2 The results suggest that the dominant directions for understanding nuclear-observable dependencies may often be one dimensional LEC combinations rather than individual parameters. p. 4 The similar results found for oxygen isotopes are reminiscent of the universal correlations found in few-body systems. p. 4
Emulator Improvements with ASL
The application of ASL to GP emulators showed that the emulator error is reduced by more than two orders of magnitude for a fixed kernel choice when using the active subspace transformation. For E(16O), projecting the original 17-dimensional LEC space onto the dominant active coordinate yields a one-dimensional GP emulator with a validation NRMSE of 7. p. 46 × 10−3. The non-linearity also becomes evident in Fig. p. 7 6, where the ground-state energy and charge radius of 48Ca against the dominant active-subspace direction y1 follows closely the curve. p. 7 Training the emulator in the ASL transformed coordinates yields a factor 3 improvement for both kernels, decreasing the RMSE from 0. p. 912 MeV to 0. p. 1 p. 1 p. 104 MeV for the SE kernel. p. 3 p. 3 p. 3 The overall findings suggest that rotating the LEC space into activesubspace coordinates substantially improves GP emulation for broad LEC ranges. p. 3
Dominant Parameter Directions from ASL
The analysis of IMSRG results showed that the normalized first eigenvalue is λ1 = 0.998 at emax = 4 and λ1 = 0. p. 8 p. 8 p. 8995 at emax = 8, indicating an only slightly richer structure at larger emax. p. 8 p. 8 p. 8 Table I provides the components of the leading active-subspace direction w1 for the ground-state energy and charge radius of 16O at emax = 8. p. 8 The behavior of TABLE I shows that cD has a large component in both directions, suggesting its dominance. p. 8 p. 8 p. 8
The GP emulators profit strongly from a rotation of the input space for these large LEC ranges. When using the full ASL dimension, the RMSE decreases from 0. p. 312 MeV to 0.04 MeV for the SE kernel. This demonstrates that the observable’s dominant direction of variation is significantly misaligned with the original LEC coordinates. p. 7 p. 7 The same misalignment is present for the oxygen isotopes. p. 7 p. 7 p. 7 p. 7
Conclusions and Outlook
Using ASL, we have shown that the dependence of nuclear observables on a high-dimensional interaction parameter space is effectively governed by very few dominant combinations of LECs. We have shown that rotating the LEC space into activesubspace coordinates substantially improves GP emulation for the broad 48Ca HF benchmark, reducing the emulator error by more than two orders of magnitude for a fixed kernel choice. p. 3 Our results suggest that the dominant directions for understanding nuclear-observable dependencies may often be one dimensional LEC combinations rather than individual parameters. p. 4 p. 4 p. 4 Co-active-subspace methods [24] provide a natural next step for quantifying such relationships jointly. p. 4 p. 4 The learned low-dimensional directions can also guide many-body calculations and inference in the most informative regions of LEC space. p. 4 p. 4
Supplemental Material to “Active-subspace learning reveals dominant parameter directions in nuclear observables”
For a point xi ∈ R d and a nearby point xj, a first-order Taylor approximation gives f(xj) ≈ f(xi) + ∇xf(xi)⊤(xj − xi). Since ∇xf(xi) ∈ R d, a single neighbor does not uniquely determine the gradient. p. 6 We therefore use k neighbors with k ≥ d chosen as the k nearest neighbors of xi in parameter space, according to the Euclidean distance. p. 6 Stacking the k linear relations yields ∆Xi gi ≈ ∆fi, where gi ≈ ∇xf(xi) and ∆Xi = (xi,1 − xi)⊤. p. 6.. (xi,k − xi)⊤ ∈ R k×d.
Gaussian Process Emulators
A Gaussian process prior is placed on the latent emulator function f(x)θ ∼ GP (µ(x), K(x, x′; θ)), where µ is a mean function, K(x, x′; θ) is a covariance kernel, and θ are hyperparameters of the kernel that have to be learned from data. The observed training data is modeled as oi = f(xi) + εi, with εi ∼ N (0, σ2n), and normal distribution N and constant noise variance σ2n (in general we use a fixed σn = 10−4 MeV). p. 6 The Matérn-3/2 kernel is given by K3/2(x, x′; σ,L) = σ squared (1 + √3r(x, x′)) × exp − √3r(x, x′) i. p. 6
The input space is additionally rotated if one chooses instead a full matrix Lfull, which is particularly helpful if there are correlated directions in parameter space. This transformation is particularly useful when the full correlation length matrix can significantly improve emulator performance for low-dimensional input spaces. p. 7
The IMSRG results for the ground-state energies and charge radii of oxygen isotopes from Ref. [9] include 399 LEC samples at emax = 4 and 262 samples at emax = 8. p. 8 p. 8 p. 8 In Fig. 7, we show 200 samples of E(16O) at emax = 8 against the normalized first four dominant active-subspace coordinates respectively. p. 8 The behavior of FIG. p. 7 p. 7 7 shows a very strong correlation between the ground-state energy and the first coordinate y1. p. 8 p. 8
The training in the ASL transformed coordinates yields a factor 3 improvement for both kernels, decreasing the RMSE from 0.12 MeV to 0.04 MeV for the SE kernel. This demonstrates that the observable’s dominant direction of variation is significantly misaligned with the original LEC coordinates. The same misalignment is present for the oxygen isotopes.
The IMSRG results for the ground-state energies and charge radii of oxygen isotopes from Ref. [9] include 399 LEC samples at emax = 4 and 262 samples at emax = 8. We find λ1 = 0. p. 7 p. 7998 at emax = 4 and λ1 = 0.995 at emax = 8, indicating an only slightly richer structure at larger emax.
We find that the GP emulators profit strongly from a rotation of the input space for these large LEC ranges. We find that rotating the LEC space into activesubspace coordinates substantially improves GP emulation for broad LEC ranges. p. 3
Using ASL, we have shown that the dependence of nuclear observables on a high-dimensional interaction parameter space is effectively governed by very few dominant combinations of LECs. We have shown that rotating the LEC space into activesubspace coordinates substantially improves GP emulation for broad LEC ranges. p. 3 p. 3 p. 3 Our results suggest that the dominant directions for understanding nuclear-observable dependencies may often be one dimensional LEC combinations rather than individual parameters. We have shown that rotating the LEC space into activesubspace coordinates substantially improves GP emulation for broad LEC ranges. The same misalignment is present for the oxygen isotopes. The same gradient information yields derivative-based global sensitivity measures directly from the calculated ensemble, without requiring the construction of a global emulator. p. 3 p. 3
We pick a fixed number of nearest neighbors k that yields the best emulator performance in terms of the RMSE. We use this knn for all training set sizes and kernels. p. 7 The behavior of TABLE I shows that cD has a large component in both directions, suggesting its dominance. We treat σn as an additional hyperparameter that is being sampled. p. 8 The behavior of TABLE I shows that cD has a large component in both directions, suggesting its dominance.
Using ASL, we have shown that the dependence of nuclear observables on a high-dimensional interaction parameter space is effectively governed by very few dominant combinations of LECs. We have shown that rotating the LEC space into activesubspace coordinates substantially improves GP emulation for broad LEC ranges. Our results suggest that the dominant directions for understanding nuclear-observable dependencies may often be one dimensional LEC combinations rather than individual parameters. Co-active-subspace methods [24] provide a natural next step for quantifying such relationships jointly. The learned low-dimensional directions can also guide many-body calculations and inference in the most informative regions of LEC space.
We find that rotating the LEC space into activesubspace coordinates substantially improves GP emulation for broad LEC ranges. The same misalignment is present for the oxygen isotopes. The same gradient information yields derivative-based global sensitivity measures directly from the calculated ensemble, without requiring the construction of a global emulator.
The IMSRG results for the ground-state energies and charge radii of oxygen isotopes from Ref. [9] include 399 LEC samples at emax = 4 and 262 samples at emax = 8. We find λ1 = 0.998 at emax = 4 and λ1 = 0.995 at emax = 8, indicating an only slightly richer structure at larger emax. The behavior of FIG. 7 shows a very strong correlation between the ground-state energy and the first coordinate y1.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Identify dominant parameter directions directly from many-body calculations by using active subspace learning to characterize response through gradients with respect to input parameters. This allows for identifying
combinations of LECs that dominate an observable over the sampled full domain.
-
Improve Gaussian process emulators by rotating the input space to align coordinates with the dominant directions from gradient-based variation, which can
greatly improve Gaussian-process emulators especially for broad LEC variations.
-
Achieve a two orders of magnitude improvement in emulator error for Hartree-Fock calculations of 48Ca by applying active subspace learning, as shown in the paper:
we demonstrate this for Hartree-Fock calculations of 48Ca, where the emulator error is improved with active-subspace learning by more than two orders of magnitude.
-
Perform derivative based global sensitivity measures without requiring a global emulator, as the same gradient information yields these measures directly from the calculated ensemble. This provides a
straightforward, computationally efficient approach to estimate sensitivities directly from data.
-
Reduce the effective input dimension for Gaussian process emulators when using a diagonal correlation length matrix by projecting onto active subspace and truncating inactive directions. This allows for
reduction of the effective input dimension when using a diagonal correlation length matrix,
which is beneficial for higher-dimensional applications where full matrices are computationally challenging.
Sources
- Modern Theory of Nuclear Forces
- Chiral effective field theory and nuclear forces
- A Guided Tour of Ab Initio Nuclear Many-Body Theory
- Three-Nucleon Forces: Implementation and Applications to Atomic Nuclei and Dense Matter
- What is ab initio in nuclear theory?
- Global sensitivity analysis of bulk properties of an atomic nucleus
- Multiscale physics of atomic nuclei from first principles
- Global Framework for Emulation of Nuclear Calculations
- Eigenvector continuation with subspace learning
- Eigenvector Continuation as an Efficient and Accurate Emulator for Uncertainty Quantification
- Eigenvector Continuation and Projection-Based Emulators
- BUQEYE Guide to Projection-Based Emulators in Nuclear Physics
- Emulators for scarce and noisy data: application to auxiliary field diffusion Monte Carlo for the deuteron
- Inferring three-nucleon couplings from multi-messenger neutron-star observations
- An Efficient Learning Method to Connect Observables
- Hartree-Fock emulators for nuclei: Application to charge radii of $^{48,52}$Ca
- Quantum Monte Carlo Calculations of Light Nuclei with Fully Propagated Theoretical Uncertainties
- Linking Electromagnetic Moments to Nuclear Interactions with a Global Physics-Driven Machine-Learning Emulator
- Active subspace methods in theory and practice: applications to kriging surfaces
- Emulating \emph{ab initio} computations of infinite nucleonic matter
Related papers
- Constrained Hamiltonian Formulation of Kinetic Theory for Relativistic Spin- 1/2 Particles
- Long- and short-range longitudinal flow decorrelations in the quark-gluon plasma
- Recurrent Fourier Neural Operator for Event-by-Event Relativistic Hydrodynamics
- Bayesian Inference of the Landau Parameter G'0 from Joint Gamow-Teller Measurements
- UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations
- Transport simulations with a constrained momentum-dependent chiral mean-field equation of state at various isospin fractions