Regionally Optimal Contour Deformations with Neural Networks

summary

The gist

In this article, we explore the use of contour deformation for the numerical evaluation of Feynman integrals after sector decomposition This work introduces a method for choosing the contour…

In short

The work introduces a method for choosing a single contour deformation that works well across an entire phase-space region using only an initial training step. This results in a lower variance integrand compared to heuristic methods, suggesting that optimizing the contour for QMC samples is ill-defined. This knowledge can speed up conventional integration or allow neural networks to avoid retraining per point.

Key concepts

Contour Deformation
This technique involves strategically deforming the path of integration in complex mathematical spaces. The goal is to change the integration path so that the resulting function being integrated has lower variance, which improves numerical accuracy when estimating values for Feynman integrals.
Neural Network Training Signal
A specific loss function is used to train a neural network. This training involves minimizing a loss that balances two terms: one related to variance reduction and another related to matching the network's approximation of the integrand, allowing it to learn optimal deformation parameters efficiently.
Free Deformation Flexibility
The method allows the contour deformation parameters ($\lambda_i$) to depend on both the sampling point ($x_1, ..., x_N$) and other variables. This flexibility enables the system to choose a small deformation in one area to reduce variance while choosing a large deformation elsewhere to avoid singularities.
Variance Reduction
The primary outcome is that the resulting integrand has lower variance than what is achieved using simple heuristic methods. Lower variance directly translates into better performance for standard Monte Carlo integration, meaning fewer samples are needed to achieve a desired level of accuracy.

Terminology used across episodes

This episode discusses

The paper

Regionally Optimal Contour Deformations with Neural Networks · Read on arXiv

S. Jones, D. Maètre, A. Olssona

Institute for Particle Physics Phenomenology, Durham University

DOI: 10.1007/JHEP11(2026)090

Transcript

Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.

Kai: We're looking at the paper: "Regionally Optimal Contour Deformations with Neural Networks".

Mira: Here's the gist of the paper. In this article, we explore the use of contour deformation for the numerical evaluation of Feynman integrals after sector decomposition This work introduces a method for choosing the contour deformation for an entire phase-space region using only an initial sampling or training step The resulting integrand has a lower variance than that obtained with heuristic methods and shows that optimising a contour to reduce the estimated error of a Randomised Quasi-Monte Carlo sample is an ill-defined problem The a priori knowledge of the integration path obtained in this work can be used to improve the speed of conventional integration methods or be leveraged for integration using neural networks, where, crucially, it removes the need to retrain the neural network for each phase-space point Contour Deformation Methodology The paper details a method for obtaining valid values of the deformation parameters for a range of kinematic invariants that also results in a low variance of the resulting integrand by training a neural network on the integrand both as a function of the Feynman parameters and the external kinematic invariants This approach avoids repeating expensive neural network training for each choice of kinematic parameters The technique can be applied to select an integration contour (and therefore fix the integrand) before using any of these integration improvement methods Training Signal Construction The training is organized in two phases, where the loss function is given by L = Lvar + C Z dx ds ReLu (ImF(z; s))2 In the first phase, the value of δ is progressively reduced by a constant factor while the learning rate is kept constant Once the regulator δ has reached a small value, the second phase starts where the learning rate is reduced by a factor when the validation loss reaches a plateau The approximation m(s) of ˆf(s) used to encourage the network to learn the locally best value of λ is also allowed to span several orders of magnitude by fitting its modulus and phase Performance Gains Analysis The comparison between the variance of the globally determined contour and one optimized on a phase-space point-by-phase-space point basis shows that the free deformation achieves a lower variance The improvement in the variance of the integrand translates directly to a better performance for a standard Monte Carlo integration However, this improvement does not always translate into an improvement of the QMC performance Free Deformation Flexibility The free deformation allows the deformation to depend also on the sampling point x1,..., xN, such that λi:= λi(x1,.,., xN; s1,.,., sM) This flexibility can negotiate situations where one would like a small λ to reduce the variance of the integrand in one part of the integration region, but a large value of λ to avoid large values of the integrand close to a pole The ability to determine contours in this manner removes the need to find valid contours on a phase-space-by-phase-space basis, potentially improving the performance of existing tools for the evaluation of Feynman integrals in the physical Minkowski regime The use of the technique is only beneficial if the gains above are amortising the computing time needed to train the networks Pruning redundant reasoning tokens In this article, we explore the use of contour deformation for the numerical evaluation of Feynman integrals after sector decomposition In existing codes, the contour of integration is determined heuristically for each phase-space point by sampling the integrand In this work, we introduce a method for choosing the contour deformation for an entire phase-space region using only an initial sampling or training step We demonstrate that the resulting integrand has a lower variance than that obtained with heuristic methods and show that optimising a contour to reduce the estimated error of a Randomised Quasi-Monte Carlo sample is an ill-defined problem The a priori knowledge of the integration path obtained in this work can be used to improve the speed of conventional integration methods or be leveraged for integration using neural networks, where, crucially, it removes the need to retrain the neural network for each phase-space point The techniques described in this work can be adapted to other problems where a non-trivial integration path has to be chosen subject to a set of constraints.

Kai: How the training signal is built The loss function Lvar = (¯f − m)2 + Z dx ds fλ(x; s) − ¯f − (m(s) − m) squared The relative weight between the two terms could be tuned, but given that the network can easily adjust the second term to vanish by an appropriate setting of the bias in its last layer, it is not an important consideration in practice.

Mira: Where the gains come from The method is useful because it removes the need to find valid contours on a phase-space-by-phase-space basis The guided NN is a lot easier to integrate into existing tools as only the new values of λi have to be passed The free deformation requires providing both the deformed contour and the Jacobian for each lattice point evaluated.

Paper summary: Kai: So, we're looking at this paper today called "Regionally Optimal Contour Deformations with Neural Networks". It deals with using contour deformation for numerical evaluation of Feynman integrals after sector decomposition.

Mira: Basically, the authors are introducing a way to pick the contour deformation for an entire phase-space region using just an initial sampling or training step. They claim that this results in a lower variance integrand compared to what you get from heuristic methods.

Lev: That’s significant because it suggests that trying to optimize a contour just to reduce the error of a randomized quasi monte carlo sample is actually an ill-defined problem, according to the paper > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg1>.

Kai: It means they found a better way than guessing where the integration path should go. And this knowledge about the integration path can actually help speed up conventional methods or even neural network integrations without needing to retrain for every single point in phase space.

Mira: They detail a method for getting valid deformation parameters for a range of kinematic invariants that also yields low variance by training a neural network on the integrand as a function of Feynman parameters and external kinematic invariants. This avoids repeating expensive neural network training for each set of kinematic parameters.

Lev: How does the training signal get built? I'm looking at how they define the loss function, which is L equals Lvar plus C times Z times dx ds times ReLU of ImF(z; s) squared. This looks like a combination of variance reduction and some regularization term > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg4>.

Kai: The first phase involves reducing the regulator delta by a constant factor while keeping the learning rate steady, and once delta is small, they switch to a second phase where the learning rate reduces when validation loss plateaus. They also let their approximation of that function m of s span several orders of magnitude by fitting its modulus and phase > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg4>.

Mira: The paper compares the variance from their globally determined contour against one optimized on a point-by-point basis, and the free deformation actually achieves a lower variance, which translates to better performance for standard monte carlo integration.

Lev: But they also admit that this improvement in integrand variance doesn't always translate into an improvement for quasi monte carlo methods. They have to be careful about what kind of integration tool you're using > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg6>.

Kai: The paper talks about free deformation flexibility, where the deformation can depend on the sampling point x one through x n, meaning lambda i is a function of those points and the Feynman parameters > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg2>.

Mira: This flexibility is useful because you can have a small lambda to reduce variance in one part of the region while needing a large value of lambda elsewhere to avoid huge values near poles > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg2>.

Lev: So, what does this mean for running this on real hardware? If we use this technique, we still need to provide both the deformed contour and the Jacobian for every lattice point we evaluate > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg6>.

Kai: It removes that need to find valid contours on a phase-space-by-phase-space basis, which is a big deal because it simplifies how we set up things in the physical Minkowski regime.

Mira: The authors suggest this can improve the speed of conventional integration methods or be used for neural network integration where you don't have to retrain the neural network for each phase space point > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg1>.

Lev: Amortizing the computing time needed to train these networks is what matters here, so if the gains in variance reduction don't outweigh that training cost, it’s not a practical improvement for us on real hardware.

Kai: So, in short, this paper shows a way to use a trained network to guide the integration path better than just random sampling or simple heuristics.

Mira: The title "Regionally Optimal Contour Deformations with Neural Networks" highlights that they are optimizing the contour deformation locally within phase space using an AI component.

Lev: It’s about moving from finding a single good contour globally to letting the network determine how that contour should change based on where you are in the integration region > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg4>.

Kai: This gives us a tool we can use to evaluate integrals more efficiently, provided we can handle the training overhead > Regionally Optimal Contour Deformations with Neural Networks <ref:2601.11448#pg3>.

Conclusion: Kai: So, this paper by Regionally Optimal Contour Deformations with Neural Networks is about using neural networks to figure out the best way to draw integration contours for Feynman integrals after sector decomposition.

Mira: It seems like they're taking a lot of guesswork out of picking those paths across a whole region of phase space.

Kai: Exactly. They show that by training this network on just an initial sample, you can get an integrand with lower variance than you’d get from the usual trial and error or heuristic guesses.

Lev: That’s interesting because finding a valid contour on a point-by-point basis is usually what people do, but this method lets them skip that whole tedious process.

Mira: The authors argue that trying to optimize the contour just to lower the error of a randomized quasi monte carlo sample isn't really an ill-defined problem anymore.

Kai: They claim this new knowledge about the integration path can help speed up old conventional integration methods, or it can be used with neural networks without needing to retrain for every single phase space point.

Lev: But that still means you need a reliable way to figure out what the deformation parameters are for each lattice point you evaluate.

Mira: The core idea is that this method removes the need to find those valid contours on a phase space-by-phase space basis, which should improve how we tackle these integrals in the physical Minkowski regime.

Kai: It's about getting a better map of where the physics is happening during integration, guided by data from training.

Mira: The real question for me is whether this gain in variance reduction actually pays off when you factor in the cost of training all those neural networks.

Lev: Right, if the time spent training the network doesn't get outweighed by how much faster or more accurate the final integration becomes, then it’s not a practical improvement for running on real hardware.

More episodes

← Home