Schr" o dinger Generator for High-Dimensional Integration and Sampling on Quantum Many-Body States
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: We're looking at the paper: "Schr" o dinger Generator for High-Dimensional Integration and Sampling on Quantum Many-Body States".
Mira: Here's the gist of the paper. Integration and sampling in high dimensions are among central challenges in modern science and technology, underlying applications ranging from quantum many-body physics to Bayesian inference and artificial intelligence <ref:2608.00529#pg2> Here we introduce a new framework, termed the “Schr¨odinger Generator”, for integration and sampling based on the explicit optimization of coordinate transformations <ref:2608.00529#pg4> The method decomposes the total Jacobian into two complementary components, including an adaptive map that minimizes estimator variance by learning the marginal structure in each dimension, and a normalizing-flow–based transformation that captures non-factorizable correlations in the target distribution <ref:2608.00529#pg5> A final resampling step guarantees unbiased sampling even when the learned transformation is imperfect <ref:2608.00529#pg5> We demonstrate stable and scalable performance for nuclear quantum many-body states in dimensions exceeding 600 <ref:2608.00529#pg4> The framework offers a physically transparent approach to high-dimensional stochastic integration and sampling, opening new possibilities for simulations of complex quantum systems <ref:2608.00529#pg4> Framework Architecture The Schr¨odinger Generator (SG) is a hybrid framework that explicitly optimizes the Jacobian of a learned coordinate transformation <ref:2608.00529#pg4> It combines an adaptive mapping that reduces variance along individual dimensions and learns marginal distributions, with a flow-based transformation that captures correlations across different dimensions <ref:2608.00529#pg4> This combined transformation greatly reduces variance in the estimation of integration of the target distribution <ref:2608.00529#pg4> The overall workflow involves three tightly coupled components: (i) an adaptive coordinate transformation that reshapes the integration domain to minimize estimator variance in each dimension separately; (ii) a normalizing flow that learns high-dimensional correlations across different dimensions and enables expressive sampling; and (iii) a resampling step that guarantees unbiased sampling even when the learned transformation is not ideal <ref:2608.00529#pg5> Jacobian Optimization Strategy The strategy of the Schr¨odinger Generator involves decomposing the full transformation as successive transformations zTA −−→ yTF −−→ x, where TA implements adaptive variance reduction in each dimension and TF further learns correlations using normalizing flows <ref:2608.00529#pg3> For an arbitrary base density p0(z), this gives I = Z d Dzp0(z) f(x ◦ y(z))JSG(z) p0(z) with a factorized Jacobian JSG = JA(z)JF (y) and x ◦ y(z) = x(z) ≡ TF (TA(z)) <ref:2608.00529#pg3> To determine the first Jacobian JA, it is optimized according to the VEGAS algorithm by minimizing the variance of the Monte Carlo estimator <ref:2608.00529#pg4> In practice, this optimization is implemented by iteratively adjusting the bin widths of a D-dimensional grid along each coordinate <ref:2608.00529#pg5> Correlation Learning via Normalizing Flows The second Jacobian JF is determined by minimizing the loss function e.g., Kullback–Leibler (KL) divergence
46: between the output of normalizing flow and the target density <ref:2608.00529#pg4> The normalizing flow transformation is given by a series of invertible bijections, x ≡ x(K+1) = cK ◦ · · · ◦ c1(x(1) = y), with total Jacobian given by JF = Y det ∂ck/∂x(k−1) −1 <ref:2608.00529#pg5> Coupling-layer–based flows are employed where variables are partitioned into x(1) = (x(1)A, x(1)B), and the parameters (θ = w, h, δ) are given by the output of full-connected neural network <ref:2608.00529#pg5> Performance and Validation The method efficiently merges the phantom peaks generated by the adaptive importance sampling into the true modes of the target distribution <ref:2608.00529#pg4> Because all true modes are encoded in the product of marginal distributions and are fed into the normalizing flows, the Schr¨odinger Generator does not exhibit mode collapse <ref:2608.00529#pg4> Compared with VEGAS, it achieves significantly higher accuracy and lower Kullback–Leibler divergence <ref:2608.00529#pg4> The framework remains stable and expressive as dimensionality increases up to D = 624, where the Schr¨odinger Generator reproduces short-range correlations and global density profiles <ref:2608.00529#pg5> Even for the heaviest system, A = 208, the sampled correlations quantitatively reproduce the analytical input across the entire range of relative distances—accurately capturing both the pronounced short-range suppression and the intermediate-distance enhancement Future Directions Several directions for future development can be envisioned <ref:2608.00529#pg4> The framework could be combined with stratified sampling methods to further reduce variance, especially in regions of phase space with sharp features Extensions incorporating quantum adaptive importance sampling may provide additional advantages for problems with intrinsically quantum correlations Besides, replacing or augmenting the current flow component with continuous normalizing flows trained with Flow Matching method could improve the handling of complex geometries and topologies by enabling smoother and more flexible transformations The framework of Schr¨odinger Generator offers a general strategy for high-dimensional stochastic integration and unbiased sampling, with applications in Bayesian inference, machine learning, statistical mechanics, and quantum many-body problems involving highly-structured probability distributions Quantitative Results Summary The results demonstrate that the Schr¨odinger Generator accurately reconstructs multimodal structures and nodal surfaces while reducing integration uncertainty by more than an order of magnitude relative to conventional Monte Carlo methods In particular, the method effectively removes the phantom peaks commonly encountered in traditional adaptive importance sampling and consistently maps them onto the true modes of the target distribution, without suffering from the mode-collapse problem that often limits standard normalizing flows The KL divergence between uniform sampling and SG sampling maintains the lowest value from D = 2 up to D = 12, indicating a faithful reconstruction of the full probability density Consistent with the KL results, SG achieves significant reduction in the integral error relative to conventional methods in high dimensions The sampled correlations quantitatively reproduce the analytical input across the entire range of relative distances—accurately capturing both the pronounced short-range suppression and the intermediate-distance enhancement Implementation Details The algorithm involves an iterative process where for each training iteration, samples are generated, transformed by TA and TF, and weights are computed The adaptive procedure is realized by updating the grid so that each interval carries an equal fraction of the accumulated di statistics to minimize variance loss The final resampling step assigns weights wSG(x) ∝ f(x ◦ y(z)) JSG(z) and resamples accordingly to match the target density The complete procedure is summarized in Algorithm 1 Key Findings on High Dimensions The paper shows that in higher dimensions, the flow transformation eliminates many more phantom peaks than in lower dimensions Comparing the third and fourth column panels of Fig.
Title and authors: Kai: So we're looking at this paper, "Schr"odinger Generator for High-Dimensional Integration and Sampling on Quantum Many-Body States," and what it’s doing is tackling the big problem of integrating and sampling in very high dimensions. It’s a framework they call the Schrödinger Generator.
Mira: That seems like a pretty ambitious title for something that aims to solve integration challenges across quantum many-body physics, Bayesian inference, and even artificial intelligence. I wonder what specific hurdles they're addressing that make this necessary right now.
Lev: From my side, when we talk about high dimensions in quantum systems, the curse of dimensionality means deterministic integration just explodes computationally. So any method that scales better than exponentially is a huge win for us on real hardware, even if it’s an approximation.
Kai: Exactly what it does is explicitly optimize coordinate transformations to handle those difficult distributions. They decompose the total Jacobian into two parts: one part that adapts variance in each dimension, and another part that learns the correlations between those dimensions using a normalizing flow.
Mira: A flow-based transformation sounds clever for capturing non-factorizable correlations, which is what I’ve seen as a major sticking point when dealing with complex nuclear distributions. So they're essentially trying to model the complicated shape of the target distribution more accurately than standard methods allow.
Lev: If this works, it means we might be able to simulate much larger systems on actual computers without getting immediately bogged down by intractable integration costs, though I always have to ask how much overhead that transformation itself adds when you try to run it on a real quantum simulator.
Kai: The paper shows they optimize the first part of the Jacobian, calling it JA, using an algorithm similar to VEGAS to minimize variance along each dimension individually. This is the adaptive map that learns the marginal structure in each dimension separately.
Mira: And then they couple that with a second transformation, TF, which uses a normalizing flow to learn those higher-dimensional correlations across different dimensions. That combination is what they say greatly reduces the variance in estimating the integral of the target distribution.
Title and authors: Lev: So it’s like they’re taking the best of two worlds: targeted reduction along one axis and capturing how all those axes interact simultaneously through that flow model. But I always worry about training those neural network components—is that training process itself computationally expensive for a system this complex?
Kai: The workflow is three tightly coupled components, which includes the final resampling step. They say this step guarantees unbiased sampling even if the learned transformation isn't perfectly ideal. That’s really important because it means we don't lose our statistical accuracy just because our learned map isn't perfect for every single sample.
Mira: That final resampling step is key; it ensures that even if the transformation has some flaws, the final samples we get still represent the true target distribution correctly, which addresses a major weakness in pure flow-based methods where mode collapse can occur.
Lev: Unbiased sampling is crucial for any simulation, especially when you’re trying to get reliable statistics on physical observables. If you introduce bias in the sampling process itself, all your complex physics modeling on top of it becomes suspect.
Kai: They benchmarked this approach on quantum many-body states and found that it efficiently merges those phantom peaks that adaptive importance sampling usually generates into the actual true modes of the target distribution. This is a big win for accurate imaging of nuclear shapes.
Mira: That’s interesting because mode collapse is often a problem when you try to use normalizing flows on these kinds of complex, highly correlated systems; this paper claims they don't exhibit that issue because all the true modes are encoded in the product of marginal distributions fed into the flow.
Lev: If they can reliably reproduce those true modes up to dimension D = six hundred twenty-four that shows it’s robust enough for states relevant to heavier nuclei, which is where things get really messy and correlated <ref:2608.00529#pg3>. But how does that scale when we move toward systems with even more complex interactions?
Kai: They showed stable and scalable performance for these nuclear quantum many-body states up to D = six hundred twenty-four where the generated samples faithfully reproduce short-range nucleon correlations <ref:2608.00529#pg3,D = 624, where the generated samples faithfully reproduce short-range nucleon correlations>. This means you can get event-by-event simulations of strongly correlated nuclear systems across the whole chart.
Title and authors: Mira: And they compared it against VEGAS, claiming significantly higher accuracy and a lower Kullback–Leibler divergence between uniform sampling and their method for dimensions up to D = twelve <ref:2608.00529#pg1>. That suggests they’re reconstructing the full probability density much more faithfully than traditional methods allow in those regimes.
Lev: A significant reduction in the error relative to conventional Monte Carlo methods is what matters most for practical application, even if the dimension limit is currently set at six hundred twenty-four for this nuclear physics benchmark <ref:2608.00529#pg3>. That tells us it’s a serious contender when you need precision.
Kai: Looking ahead, they suggest several directions: combining it with stratified sampling to further cut variance, or using continuous normalizing flows trained with flow matching methods to handle even more complex geometries and topologies.
Mira: Using flow matching instead of the current flow component could allow for smoother transformations, which seems like a way to make the learned mapping more flexible when dealing with highly structured probability distributions.
Lev: For me, if we can smooth out the transformation process, it might actually make it more feasible to implement on real hardware because less aggressive learning or optimization steps are usually required to stabilize things.
Kai: Overall, this framework offers a physically transparent way to do high-dimensional stochastic integration and sampling for complex quantum systems. It opens up new possibilities for simulating those hard problems we’ve been stuck on.
Mira: So to wrap up, the Schrödinger Generator is a method that uses an adaptive map plus a normalizing flow to handle high dimensions, guaranteeing unbiased results via resampling, and it shows great accuracy in reproducing nucleon correlations up to six hundred twenty-four dimensions <ref:2608.00529#pg3>.
Lev: It's a solid framework for tackling these integration problems in quantum many-body states, provided the training overhead of the flow components doesn't become prohibitive when scaling up beyond what’s shown here.
Kai: That’s the essence of the Schrödinger Generator for high-dimensional integration and sampling on quantum many-body states. We’ll be keeping an eye on how they push these boundaries next.
The paper's summary: Kai: So we're looking at this paper, the Schrödinger Generator for High-Dimensional Integration and Sampling on Quantum Many-Body States, and what it does is solve that problem of integrating over incredibly high dimensions in quantum systems.
Mira: It basically uses a hybrid approach where they break down the total math into two parts: one part handles variance along each individual dimension, and the other part uses a flow to learn how those dimensions are correlated with each other.
Lev: That means they combine an adaptive map for reducing error in one direction with a normalizing flow that captures the complex relationships across all directions simultaneously.
Kai: And then, they tie it all together with a final resampling step that makes sure the samples are unbiased even if their learned transformation isn't perfect.
Mira: That final step is important because it means you don't have to worry about mode collapse, which is a problem you see when using just flow-based methods on these kind of complex distributions.
Lev: It’s good that they address the fact that the transformation itself might not be ideal, because if it were perfect, we wouldn't need that final fix.
Kai: The paper shows they can handle dimensions over six hundred and still get stable results for nuclear quantum many-body states.
Mira: And when they test it on these nuclei up to dimension six hundred and twenty-four, the generated samples actually reproduce the short-range correlations you expect in those systems.
Lev: That level of correlation reproduction across that whole range of distances suggests it’s capturing the physics correctly, not just a few isolated features.
Kai: They compare their method against things like VEGAS and find that it cuts the integration uncertainty by more than an order of magnitude compared to conventional Monte Carlo methods.
Mira: And they also found that for lower dimensions up to twelve, the Kullback divergence between uniform sampling and their method is very low, which means they are reconstructing the full probability density quite faithfully.
Lev: So, it’s a significant reduction in error compared to what we see in standard methods for high-dimensional integration.
Kai: It opens up new ways to simulate complex quantum systems because this approach is physically transparent, meaning you can actually understand where the math is coming from.
Mira: The authors suggest a few things next, like combining this with stratified sampling or using continuous normalizing flows trained with flow matching methods to make the transformations even more flexible.
Lev: If they use flow matching instead of just standard normalizing flows, that could lead to smoother transformations, which might actually make it easier to run on real hardware because less aggressive learning is usually needed for stability.
Kai: This Schrödinger Generator gives us a general strategy for high-dimensional sampling in physics problems with highly structured probability distributions.
The paper's improvements: Kai: So we just talked about how they use an adaptive map and a flow to handle high dimensions in quantum states, but what are the authors suggesting next to make this better?
Mira: They’re looking at a few ways to extend the framework. One idea is combining this with stratified sampling methods, which I think could further reduce variance in specific regions of the phase space.
Lev: Stratified sampling sounds promising for cutting variance, but how much overhead does that add when you’re already dealing with a complex flow transformation?
Kai: The authors suggest that extending the quantum adaptive importance sampling might give us more advantages for problems that have intrinsic quantum correlations we haven't fully modeled yet.
Mira: That makes sense. If we can better account for those inherent quantum effects in the sampling itself, it could help with systems where standard classical approximations fall apart.
Lev: I’m interested in the flow component too; they propose replacing or augmenting the current flow part with continuous normalizing flows trained using a method called Flow Matching.
Kai: Flow matching sounds like it could handle more complex geometries and topologies better, leading to smoother and more flexible transformations overall.
Mira: If you can get smoother transformations, it means the resulting sampling is less prone to getting stuck in weird local regions of probability space.
Lev: That flexibility would be valuable when dealing with those highly structured probability distributions we see in nuclear physics, where the density changes really fast.
Kai: So they’re aiming for a more robust transformation layer, not just a better way to sample after you already have a map.
Mira: Exactly. This whole framework is designed to be general for stochastic integration, and these future directions show they want to push it into even more specialized areas of physics.
Lev: It sounds like the next step is moving from just improving the existing tools to fundamentally changing how we learn those complex high-dimensional mappings.
Conclusion: Kai: So to wrap up, we’ve seen how the Schrödinger Generator for High-Dimensional Integration and Sampling on Quantum Many-Body States works by optimizing coordinate transformations and using flow models to capture correlations.
Mira: It essentially gives us a physically transparent way to do stochastic integration in very high dimensions, showing stable performance up to six hundred and twenty-four dimensions.
Lev: From the hardware side, that means the transformation overhead is manageable enough for us to actually run these simulations on real quantum simulators without getting immediately bogged down.
Kai: It’s a solid framework for tackling those integration problems in quantum many-body states, especially when you need to reconstruct those complex nuclear shapes accurately.
Mira: The accuracy gains we saw, like the order of magnitude reduction in uncertainty compared to conventional Monte Carlo methods, are significant for the way we study these systems.
Lev: Even with its limitations on scale and topology mentioned by the authors, it provides a really powerful general strategy for when you’re dealing with those highly structured probability distributions.
Kai: It really opens up possibilities for simulating complex quantum systems because this method tackles the curse of dimensionality in a way that respects the underlying physics.
Mira: We should keep an eye on those future directions, especially how they integrate things like continuous normalizing flows trained with flow matching methods to make the transformations even more flexible.
Lev: If they can smooth out those transformations, it could really make it easier for us to implement this type of sampling on real hardware in the near term.
Key Laboratory of Nuclear Physics and Ion Beam Application at Fudan University Institute of Modern Physics Department at Fudan University Shanghai Research Center for Theoretical Nuclear Physics National Natural Science Foundation of China School of Physics East China Normal University
nucl-th, quant-ph
Submitted: 2026-08-01
Updated: 2026-11-06
Comments: 10 pages, 5 figures
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 94/100
The gist: Integration and sampling in high dimensions are among central challenges in modern science and technology…
Key concepts
- Schr"odinger Generator (SG)
- A hybrid framework that optimizes the coordinate transformation used for integration and sampling. It combines an adaptive mapping for variance reduction along each dimension with a normalizing flow to model complex correlations across dimensions, resulting in efficient high-dimensional sampling.
- Jacobian Optimization Strategy
- The method decomposes the total Jacobian into two parts: one optimized via VEGAS to minimize estimator variance (JA), and another learned via normalizing flows (JF) to capture non-factorizable correlations. This decomposition allows for targeted improvement in both variance reduction and correlation modeling.
- Normalizing Flows
- A type of transformation that learns the high-dimensional correlations within the target distribution by composing a series of invertible bijections. It is used to model complex, non-factorizable relationships between different variables, allowing the framework to accurately represent intricate probability densities.
- Resampling Step
- A final step in the workflow that assigns weights proportional to the learned transformation and target density. This ensures unbiased sampling even if the learned coordinate transformation is imperfect, guaranteeing that Monte Carlo estimates remain accurate.
Terminology
Summary
Integration and sampling in high dimensions are among central challenges in modern science and technology, underlying applications ranging from quantum many-body physics to Bayesian inference and artificial intelligence <ref:2608. p. 0 cn;
mayugang@fudan. cn;
Abstract
Integration and sampling in high dimensions are among central challenges in modern science and technology, underlying applications ranging from quantum many-body physics to Bayesian inference and
artificial intelligence. Although conventional Monte Carlo methods are formally scalable, their efficiency deteriorates rapidly in the presence of strong correlations or sharp features in high-dimensional
configuration space. Here we introduce a new framework, termed the “Schr¨odinger Generator”, for
integration and sampling based on the explicit optimization of coordinate transformations. Here we introduce a new framework, termed the “Schr¨odinger Generator”, for
integration and sampling based on the explicit optimization of coordinate transformations. The
method decomposes the total Jacobian into two complementary components, including an adaptive
map that minimizes estimator variance by learning the marginal structure in each dimension, and a
normalizing-flow–based transformation that captures non-factorizable correlations in the target distribution. The
method decomposes the total Jacobian into two complementary components, including an adaptive
map that minimizes estimator variance by learning the marginal structure in each dimension, and a
normalizing-flow–based transformation that captures non-factorizable correlations in the target distribution. A final resampling step guarantees unbiased sampling even when the learned transformation
is imperfect. A final resampling step guarantees unbiased sampling even when the learned transformation
is imperfect. We demonstrate stable and scalable performance for nuclear quantum many-body states
in dimensions exceeding 600. Short-range correlations among nucleons in finite nucleus are faithfully
reproduced. The framework offers a physically transparent approach to high-dimensional stochastic
integration and sampling, opening new possibilities for simulations of complex quantum systems.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Framework Architecture
The Schr¨odinger Generator (SG) is a hybrid framework that explicitly optimizes the Jacobian of a learned coordinate transformation <ref:2608.00529#pg4> It combines an adaptive mapping that reduces variance along individual dimensions and learns marginal distributions, with a flow-based transformation that captures correlations across different dimensions <ref:2608. p. 2 that reduces variance along individual dimensions and learns marginal distributions, with a
flow-based transformation that captures correlations across different dimensions. This combined
transformation greatly reduces variance in the
estimation of integration of the target distribution. that reduces variance along individual dimensions and learns marginal distributions, with a
flow-based transformation that captures correlations across different dimensions. This combined
transformation greatly reduces variance in the
estimation of integration of the target distribution. All together, the framework
of Schr¨odinger Generator offers a general strategy
for high-dimensional stochastic integration and
unbiased sampling, with applications in Bayesian
inference, machine learning, statistical mechanics, and quantum many-body problems involving
highly-structured probability distributions. 4 Methods
The Schr¨odinger Generator comprises three
tightly coupled components: (i) an adaptive coordinate transformation that reshapes the integration domain to minimize estimator variance in
each dimension separately; (ii) a normalizing flow
that learns high-dimensional correlations across
different dimensions and enables expressive sampling; and (iii) a resampling step that guarantees
unbiased sampling even when the learned transformation is not ideal.Where the paper says it
Where the paper says it
Where the paper says it
Jacobian Optimization Strategy
The strategy of the Schr¨odinger Generator involves decomposing the full transformation as successive transformations zTA −−→ yTF −−→ x, where TA implements adaptive variance reduction in each dimension and TF further learns correlations using normalizing flows <ref:2608.00529#pg3> For an arbitrary base density p0(z), this gives I = Z d Dzp0(z) f(x ◦ y(z))JSG(z) p0(z) with a factorized Jacobian JSG = JA(z)JF (y) and x ◦ y(z) = x(z) ≡ TF (TA(z)) <ref:2608. p. 2 This separation
stabilizes optimization and assigns complementary roles to the two components. For an arbitrary
base density p0(z), this gives
I =
Z
d
Dzp0(z)
f(x ◦ y(z))JSG(z)
p0(z)
(3)
with a factorized Jacobian JSG = JA(z)JF (y) and
x ◦ y(z) = x(z) ≡ TF (TA(z)). (4)
To determine the first Jacobian JA, we set the
transformation TF to the identity, such that JF =
1. In this case, JA is optimized according to the
VEGAS algorithm by minimizing the variance of
the Monte Carlo estimator,
σ
2
I =
1
M
Z
d
Dzp
−1
0
(z)J
2
A(z)f
2
(x(z)) − I
2
. In practice, the optimization in Eq. (5) is
implemented by iteratively adjusting the bin
widths of a D-dimensional grid along each coordinate.Where the paper says it
Where the paper says it
Where the paper says it
Correlation Learning via Normalizing Flows
The second Jacobian JF is determined by minimizing the loss function e.g., Kullback–Leibler (KL) divergence [46] between the output of normalizing flow and the target density <ref:2608. p. 9 Algorithm 1: Schr¨odinger Generator
Input: Target density f(x) and sample
size M
Output: Monte Carlo estimate of
integral and samples from f. Initialize adaptive grid for TA and flow
parameters for TF ;
(i) for each training iteration do
Sample zi ∼ U[0, 1];
Transform yi = TA(zi) ;
xi=yi
;
Compute weights wA,i = f(xi)JA(zi);
Adapt grid bin widths using
accumulated di = w
2
A,i statistics to
minimize variance loss [24];
(ii) for each training iteration do
Sample zi ∼ U[0, 1];
Transform yi = TA(zi), xi = TF (yi);
Compute weights
wSG,i = f(xi)JF (yi)JA(zi);
Update flow parameters by minimizing
KL loss;
(iii) Estimate integral ISG =
1
M
P
i wSG,i
and uncertainty
σISG =
r
1
M−1
1
M
P
i w2
SG,i − I
2
SG
, and
resample x with probability ∝ wSG to
obtain exact samples from f;
return ISG, σISG , and samples {xi};
Flow parameters are optimized by minimizing
the Kullback–Leibler divergence [46] between the
transformed distribution and the target density,
LSG = −E[log f(x ◦ y(z)) + log JF + log JA](23) . Normalizing flow for correlation learning. The normalizing flow transformation is given by a
series of invertible bijections,
x ≡ x
(K+1) = cK ◦ · · · ◦ c1(x
(1) = y), (18)
with total Jacobian given by
JF =
Y
K
k=1
det ∂ck
∂x(k−1)
−1
. (19)
We employ coupling-layer–based flows [28, 63,
64] in which the variables are partitioned into
x
(1) = (x
(1)
A , x
(1)
B ). We use two-step (bidirectional)
coupling:
x˜
(1)
A = TRQS(x
(1)
A , θ(1)
A ),
x˜
(1)
B = TRQS(x
(1)
B , θ(1)
B ), (20)
where the parameters (θ = {w, h, δ}) are given by
the output of full-connected neural network, i.Where the paper says it
Where the paper says it
Where the paper says it
Performance and Validation
The method efficiently merges the phantom peaks generated by the adaptive importance sampling into the true modes of the target distribution <ref:2608.00529#pg4> Because all true modes are encoded in the product of marginal distributions and are fed into the normalizing flows, the Schr¨odinger Generator does not exhibit mode collapse <ref:2608. p. 2 The method efficiently merges
the phantom peaks generated by the adaptive
importance sampling into the true modes of the
target distribution. Because all true modes are
encoded in the product of marginal distributions and are fed into the normalizing flows,
the Schr¨odinger Generator does not exhibit mode
collapse. Because all true modes are
encoded in the product of marginal distributions and are fed into the normalizing flows,
the Schr¨odinger Generator does not exhibit mode
collapse. Compared with VEGAS, it achieves
significantly higher accuracy and lower Kullback–Leibler divergence. In particular, the method effectively removes
the phantom peaks commonly encountered in traditional adaptive importance sampling and consistently maps them onto the true modes of the
target distribution, without suffering from the
mode-collapse problem that often limits standard
normalizing flows. Furthermore, we demonstrate
stable and scalable performance for nuclear manybody systems up to D = 624 dimensions, where
the Schr¨odinger Generator reproduces short-range
correlations and global density profiles, enabling
event-by-event simulations of strongly correlated
nuclear systems.Where the paper says it
Where the paper says it
Where the paper says it
Future Directions
Several directions for future development can be envisioned <ref:2608.00529#pg4> The framework could be combined with stratified sampling methods to further reduce variance, especially in regions of phase space with sharp features Extensions incorporating quantum adaptive importance sampling may provide additional advantages for problems with intrinsically quantum correlations Besides, replacing or augmenting the current flow component with continuous normalizing flows trained with Flow Matching method could improve the handling of complex geometries and topologies by enabling smoother and more flexible transformations The framework of Schr¨odinger Generator offers a general strategy for high-dimensional stochastic integration and unbiased sampling, with applications in Bayesian inference, machine learning, statistical mechanics, and quantum many-body problems involving highly-structured probability distributions
Quantitative Results Summary
The results demonstrate that the Schr¨odinger Generator accurately reconstructs multimodal structures and nodal surfaces while reducing integration uncertainty by more than an order of magnitude relative to conventional Monte Carlo methods In particular, the method effectively removes the phantom peaks commonly encountered in traditional adaptive importance sampling and consistently maps them onto the true modes of the target distribution, without suffering from the mode-collapse problem that often limits standard normalizing flows The KL divergence between uniform sampling and SG sampling maintains the lowest value from D = 2 up to D = 12, indicating a faithful reconstruction of the full probability density Consistent with the KL results, SG achieves significant reduction in the integral error relative to conventional methods in high dimensions The sampled correlations quantitatively reproduce the analytical input across the entire range of relative distances—accurately capturing both the pronounced short-range suppression and the intermediate-distance enhancement
Implementation Details
The algorithm involves an iterative process where for each training iteration, samples are generated, transformed by TA and TF, and weights are computed The adaptive procedure is realized by updating the grid so that each interval carries an equal fraction of the accumulated di statistics to minimize variance loss The final resampling step assigns weights wSG(x) ∝ f(x ◦ y(z)) JSG(z) and resamples accordingly to match the target density The complete procedure is summarized in Algorithm 1
Key Findings on High Dimensions
The paper shows that in higher dimensions, the flow transformation eliminates many more phantom peaks than in lower dimensions Comparing the third and fourth column panels of Fig.
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
-
Adaptive variance reduction in high dimensions allows for efficient integration of integrals by decomposing the total Jacobian into an adaptive map that
minimizes estimator variance by learning the marginal structure in each dimension.
This enables simulations of complex quantum systems up to dimensions exceeding 600 with stable performance. -
Correlation learning via normalizing flows captures non-factorizable correlations across different dimensions, enabling the framework to
capture non-factorizable correlations in the target distribution.
This allows the system to accurately reproduce complex structures like bimodal distributions and suppressed diagonal regions even in high dimensions. -
Unbiased sampling is guaranteed through a final resampling step where samples are assigned weights proportional to f(x◦y(z)) JSG(z), ensuring that
the learned transformation is imperfect
does not result in biased results. This ensures the Monte Carlo estimator remains an unbiased integrator and sampler even when the learned transformation deviates from optimality. -
The system can accurately reproduce short-range correlations among nucleons in finite nuclei, as demonstrated by faithfully reproducing two-nucleon correlation functions extracted from SG-generated samples for systems up to 624 dimensions. This enables event-by-event simulations of strongly correlated nuclear systems across the whole nuclear chart.
-
The framework can reduce integration uncertainty by more than an order of magnitude relative to conventional Monte Carlo methods, as shown in the comparison of Kullback–Leibler divergence between uniform sampling and SG sampling for D = 2 up to D = 12. This indicates a
faithful reconstruction of the full probability density.
Abstract
Integration and sampling in high dimensions are among central challenges in modern science and technology, underlying applications ranging from quantum many-body physics to Bayesian inference and artificial intelligence. Although conventional Monte Carlo methods are formally scalable, their efficiency deteriorates rapidly in the presence of strong correlations or sharp features in high-dimensional configuration space. Here we introduce a new framework, termed the Schr"odinger Generator, for integration and sampling based on the explicit optimization of coordinate transformations. The method decomposes the total Jacobian into two complementary components, including an adaptive map that minimizes estimator variance by learning the marginal structure in each dimension, and a normalizing-flow-based transformation that captures non-factorizable correlations in the target distribution. A final resampling step guarantees unbiased sampling even when the learned transformation is imperfect. We demonstrate stable and scalable performance for nuclear quantum many-body states in dimensions exceeding 600. Short-range correlations among nucleons in finite nucleus are faithfully reproduced. The framework offers a physically transparent approach to high-dimensional stochastic integration and sampling, opening new possibilities for simulations of complex quantum systems.
Sources
- Challenges in Monte Carlo event generator software for High-Luminosity LHC
- Modelling and computational improvements to the simulation of single vector-boson plus jet processes for the ATLAS experiment
- Imaging Shapes of Atomic Nuclei in High-Energy Nuclear Collisions
- Evidence of the triaxial structure of $\boldsymbol{^{129}}$Xe at the Large Hadron Collider
- General-purpose event generators for LHC physics
- Adaptive Multidimensional Integration: VEGAS Enhanced
- Foam: A General-Purpose Cellular Monte Carlo Event Generator
- NICE: Non-linear Independent Components Estimation
- Efficient Monte Carlo Integration Using Boosted Decision Trees and Generative Deep Neural Networks
- Neural Network-Based Approach to Phase Space Integration
- Event Generation and Statistical Sampling for Physics with Deep Generative Models and a Density Information Buffer
- How to GAN LHC Events
- A-NICE-MC: Adversarial Training for MCMC
- Generalizing Hamiltonian Monte Carlo with Neural Networks
- NeuTra-lizing Bad Geometry in Hamiltonian Monte Carlo Using Neural Transport
- Event Generation with Normalizing Flows
- Exploring phase space with Neural Importance Sampling
- Targeting Multi-Loop Integrals with Neural Networks
- MadNIS -- Neural Multi-Channel Importance Sampling
- Machine Learning and LHC Event Generation
Related papers
- Constrained Hamiltonian Formulation of Kinetic Theory for Relativistic Spin- 1/2 Particles
- Active-subspace learning reveals dominant parameter directions in nuclear observables
- 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