Differentiable Principal-Value Inversion for Neural-Network Extraction of Generalized Parton Distributions

summary

The gist

The paper presents a machine–learning method for the nonparametric extraction of generalized parton distributions (GPDs) from Compton form factors (CFFs) constrained by experimental data It…

In short

This research develops a machine learning method to extract Generalized Parton Distributions (GPDs) from Compton form factors using a differentiable principal-value integral transform. The method embeds the complex QCD integral kernel directly into a neural network, enforcing physical constraints like exact oddness and endpoint suppression. This allows for a fully probabilistic extraction of GPDs with uncertainty quantification, offering a flexible alternative to traditional functional parametrizations.

Key concepts

Generalized Parton Distributions (GPDs)
GPDs are fundamental functions in quantum chromodynamics that describe the internal structure of hadrons. They provide a three-dimensional picture of quarks and gluons inside protons or neutrons, including information about their momentum and spatial distribution, which is crucial for understanding how quarks move within the nucleon.
Principal-Value (PV) Linear Integral Transform
This is a mathematical operation used to relate GPDs to measurable quantities like Compton form factors. Because the kernel has a singularity at certain points, this transform is difficult to invert directly. The paper solves this by creating a differentiable representation of this transform and making it part of the neural network, allowing for direct training.
Replica Ensemble
This technique is used for robust uncertainty quantification. Instead of one GPD prediction, the model trains an ensemble of GPD networks using multiple versions (replicas) of experimental data (Compton form factors). This ensemble provides a mean prediction and a band representing the statistical uncertainty across different experimental realizations.
Curvature-Based Regularization
This is a technique used during training to stabilize the extraction process. It penalizes high-frequency oscillations in the GPD by measuring its curvature (the second derivative) with respect to kinematic variables like x, ξ, and t. This ensures the resulting GPD surface is smooth and physically realistic.

Terminology used across episodes

This episode discusses

The paper

Differentiable Principal-Value Inversion for Neural-Network Extraction of Generalized Parton Distributions · Read on arXiv

Dima Watkins, *, I. P. Fernando, † Dustin Keller

Department of Physics, University of Virginia

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: "Differentiable Principal-Value Inversion for Neural-Network Extraction of Generalized Parton Distributions".

Mira: Here's the gist of the paper. The paper presents a machine–learning method for the nonparametric extraction of generalized parton distributions (GPDs) from Compton form factors (CFFs) constrained by experimental data It addresses the longstanding inverse problem posed by the principal-value (PV) linear integral transform with a singular kernel that relates the charge–conjugation–even (C-even) quark GPD H(+) to the real part of the deeply virtual Compton scattering (DVCS) amplitude Our approach constructs a differentiable representation of the quantum chromodynamics (QCD) PV kernel and embeds it as a fixed, physics-preserving layer inside a neural network that parameterizes the GPD H(+)(x, ξ, t,Q2) The model enforces exact oddness in x, implements endpoint suppression, and includes curvature-based regularization that stabilizes the inversion in kinematically ill-conditioned regions A Monte-Carlo ensemble of CFFs, obtained from a global neural-network fit to unpolarized DVCS measurements with propagated experimental uncertainties, serves as input to a replica ensemble of GPD networks, yielding a fully probabilistic extraction of H(+) over the phase space We demonstrate the method using a global determination of ReH for Jefferson Lab measurements and present a direct neural network reconstruction of three-dimensional GPD surfaces H(+)(x0, ξ, t,Q20) obtained from experimental CFF inputs.

Kai: Pruning redundant reasoning tokens The method introduces a new approach in which the QCD integral transform that maps H(+) to ReH is implemented as a known, fixed, differentiable integral operator inside a neural network This design has several key advantages including enforcing exact C-parity and endpoint suppression, preserving the analytic structure of QCD factorization, enabling gradient-based training directly on ReH, and providing a natural way to implement regularization in the GPD domain, such as curvature penalization in x, ξ, and t.

Mira: How the training signal is built The network outputs the GPD itself given CFF input information which enforces exact oddness H(+)(x) = −H(+)(-x) and the expected endpoint behavior through a compact envelope function (1−x)β The CFF (ReH) input information is provided by a DNN model continuous in a region of phase space defined by (xb, t,Q2), obtained from a global fit to experimental DVCS data using the BMK formalism.

Kai: The integral transform is computed on a high-resolution x grid in a manner that is fully differentiable enabling gradient-based training needed for our modern deep-learning approach.

Title and authors: Kai: So we're talking about this paper now. It’s called "Differentiable Principal-Value Inversion for Neural-Network Extraction of Generalized Parton Distributions." It tackles a really old problem in the field, which is how to get those generalized parton distributions, or GPDs, from the Compton form factors, or CFFs.

Mira: Right. This paper addresses that big inverse problem posed by the principal-value integral transform with its singular kernel that links the charge-conjugation even quark GPD H plus to the real part of deep virtual Compton scattering amplitude. It’s basically saying we can build a way to do this using machine learning.

Lev: And what's interesting is their core idea: they construct a differentiable representation of that quantum chromodynamics principal-value kernel and embed it as a fixed layer inside a neural network that parameterizes the GPD itself, H plus of x, xi, t and Q squared.

Kai: That means the network isn't just guessing; it has this built-in rule for how QCD should work in terms of that integral transform. They also enforce exact oddness in x and they implement endpoint suppression directly into the learning architecture.

Mira: And they use curvature based regularization, which they say stabilizes the inversion in those kinematically ill conditioned regions where things tend to get messy. It’s an approach that aims to make the inversion more robust than just a standard fitting procedure would allow.

Lev: They also use a Monte Carlo ensemble of CFFs, which they get from a global neural network fit to unpolarized deep virtual Compton scattering measurements with propagated experimental uncertainties, as input for a replica ensemble of GPD networks. This gives them a fully probabilistic extraction of H plus over the entire phase space.

Kai: So, what they do is show that you can get this direct neural network reconstruction of the three dimensional GPD surfaces H plus of x naught, xi t and Q squared just from experimental CFF inputs. That’s a pretty direct way to think about it.

Title and authors: Mira: The paper also talks about how they discretize the principal-value operator using symmetric trapezoidal weights and treat that singular point at x equal to plus xi or minus xi by excising a small region around those points, which is how they make the forward mapping differentiable.

Lev: And for training, they make sure backpropagation passes exactly through both the principal-value integrand and the deep neural network itself. That’s what enables this gradient based training needed for their modern deep learning approach.

Kai: The uncertainties are quantified by training a replica ensemble in PyTorch across many replicas of the CFFs obtained from that global fit, where local fits were used to construct a continuous mapping across phase space. This gives them both a mean prediction and a pointwise uncertainty band for H plus at any desired kinematic point.

Mira: They also introduce smoothness priors directly into the GPD domain to stabilize reconstruction by penalizing high frequency oscillations in x, xi, and t. Specifically they use curvature regularization on H plus with respect to x, denoted as Rx of H plus equals Z over one minus one dx partial two H plus theta over x squared, and kinematic smoothness terms for xi and t.

Lev: The scalar loss function they minimize during training is L equals X n real part of theta n xi n t Q squared minus real part of exp exponent sigma squared n plus lambda x Rx plus lambda xi R xi plus lambda t R t. That term controls the whole process, balancing the data fit against these smoothness penalties.

Kai: This method gives us a way to learn the GPD directly in terms of those kinematic variables, x, xi, t and Q squared. It’s a minimally parametrized alternative to things like doubledistribution ansätze or regge inspired parametrizations.

Mira: And they also show that you can embed both the real part and imaginary part of the CFF data into this single network by training it end-to-end on both real part and imaginary part data, which is something that’s not always easy to do with separate models.

Title and authors: Lev: The paper also tests its closure test by comparing the reconstructed GPD against a known analytic truth model to quantify the intrinsic methodological error of the principal value deep neural network inversion. This lets them tune their hyperparameters until this algorithmic error becomes smaller than the uncertainties expected from realistic CFF inputs.

Kai: So, to sum up, this paper presents a method for nonparametric extraction of generalized parton distributions from Compton form factors using a differentiable representation of the QCD principal-value kernel embedded in a neural network.

Mira: The main implications are that we can get these GPDs probabilistically and with quantified uncertainty by training on CFF data, rather than relying on fixed functional forms or truncated expansions.

Lev: For real hardware running this, you'd need to consider the overhead of running an ensemble of replica networks and the complexity of calculating that principal-value integral in a differentiable way across a high resolution x grid.

Kai: What this means for us is that we can move toward more flexible analyses, fitting to both fixed target and collider kinematics without assuming any specific functional form for the GPD itself.

Mira: It’s about getting a more direct link between what we measure in deep virtual Compton scattering and the underlying structure of the nucleon.

Lev: And this approach also allows for probing polynomiality violation by computing the normalized residual epsilon m of t and Q squared for moments one and three, which lets us compare baseline reconstructions against a soft polynomiality penalty without degrading CFF fidelity.

Kai: That’s what we have with "Differentiable Principal-Value Inversion for Neural-Network Extraction of Generalized Parton Distributions." It's a new way to tackle this inverse problem in QCD.

Mira: It’s about learning the GPD directly from the experimental data through a differentiable mapping, and quantifying the resulting uncertainty probabilistically.

Lev: We need to keep an eye on how well that replica ensemble performs when we try to run it on actual experimental data sets, because that's where we really find out if this method is ready for real hardware.

The paper's summary: Kai: So, to recap, this paper is about using machine learning to pull generalized parton distributions, or GPDs, out of Compton form factors from experimental data by building a neural network around the math of the principal-value integral transform.

Mira: Right. They're taking that really tricky inverse problem and making it solvable by baking the physics—the quantum chromodynamics kernel—directly into the neural network structure itself. It's not just a black box fitting something; it’s a constrained fit where the constraints are derived from how QCD is supposed to work mathematically.

Lev: And they enforce things like exact oddness in the distribution, which is a real physical requirement, and they suppress endpoint behavior right there in the learning process. That means you don't have to guess those features; the architecture forces them to be right based on symmetry and kinematics.

Kai: The big takeaway here is how they handle uncertainty. They use a replica ensemble of CFF data from a global neural network fit, which then feeds into another ensemble of GPD networks. This gives you not just one answer for the GPD, but a whole distribution of possible answers—a mean prediction and a band showing you where the error lives across all those thousands of different CFF possibilities.

Mira: That probabilistic extraction is huge because it moves us away from just getting one number and gives us a real measure of how much we can trust our result based on the input data quality. It’s not just a fit; it’s a statistical map of the possibilities.

Lev: From an engineering standpoint, running that replica ensemble takes some serious computational muscle. You have to manage an ensemble of networks and do that complex principal-value integration repeatedly for every CFF replica you're using as input. It's heavy overhead before you even look at the physics results.

Kai: But they show that this method works without needing any specific functional form for the GPD itself, which is a massive win for flexibility. You can use it for fixed target experiments and collider kinematics, just by giving it different experimental inputs.

Mira: It really connects the dots between what we measure in deep virtual Compton scattering and the fundamental structure of the nucleon in a way that bypasses those old, restrictive parametrizations you see in other literature.

Lev: And they also tested how robust this is by comparing their final result against a known analytic truth model. They tune their regularization terms, like the curvature penalty on x, xi, and t, until that algorithmic error gets smaller than the uncertainty expected from the actual CFF inputs. That's how you prove it's stable enough to matter.

Kai: So this isn't just a new way to fit data; it’s a new way to learn the GPD surface directly from experimental observables with built-in physical constraints and quantified risk. It opens up a lot more avenues for probing the internal structure of protons and neutrons.

Mira: Exactly. This makes the transition from theoretical QCD calculations to experimentally accessible observables much smoother, especially when you're dealing with those complex kinematic regions where things used to break down.

Lev: It’s a powerful tool, but we still need to figure out how fast we can run this on actual hardware with low noise and high fidelity. That’s the next step for us in quantum error correction and real-world machine learning deployment.

The paper's improvements: Kai: So, to recap, this paper isn't just proposing a new way to do the inversion; they are also suggesting specific ways to make that inversion even more robust and accurate using regularization and diagnostics.

Mira: Exactly. They aren't stopping at the basic differentiability; they’re adding curvature-based regularization directly into the GPD domain, which is where they penalize high frequency oscillations in x, xi, and t. That’s meant to stabilize the reconstruction when things get messy kinematically.

Lev: From an engineering side, that curvature penalty term equals Z over one minus one times the partial derivative of H plus with respect to x squared. It’s a direct way to suppress those spurious oscillations that the principal-value operator would otherwise amplify across the grid.

Kai: And they also introduced kinematic smoothness terms for xi and t, which basically keep things smooth in those other two dimensions as well. They're trying to control the noise in every single direction of the four dimensional phase space.

Mira: Plus, they have this closure test idea where you compare their reconstructed GPD against a known analytic truth model. This lets them tune the hyperparameters until that algorithmic error is smaller than the experimental uncertainty they expect from the CFF inputs.

Lev: That’s smart because it gives you a concrete benchmark for when to stop tweaking things and when you've hit a plateau where more tuning won't help much. It sets a realistic limit on how good the model can get before it starts just memorizing the input data noise instead of learning the physics.

Kai: And they mentioned probing polynomiality violation by looking at the normalized residual epsilon m for moments one and three. This is a diagnostic tool to see if their reconstruction is actually respecting fundamental QCD constraints, even when you're using this flexible neural network approach.

Mira: That’s important because it lets us check if the method is just fitting the data well, or if it’s learning something physically meaningful about how quarks inside the nucleon behave. It helps separate model bias from actual physics signals.

Lev: For running this on real hardware, those regularization terms add complexity to the training loop, but they are necessary because without them, you'd just get a wildly oscillating mess that won't converge properly in a deep network architecture.

Kai: So these improvements show they’re not just throwing layers at the problem; they’re actively shaping the learning process to respect QCD structure and stability. It’s about making sure the machine actually learns something useful, not just some statistical curve-fit.

Mira: It's about controlling the model's behavior so that its output is physically sensible across all those kinematic variables. That control is what makes this approach viable for real physics extraction.

Lev: We still need to figure out how to implement those curvature penalties efficiently in a production setting, especially when you have a massive Monte Carlo ensemble of CFFs to work with before you even start training the GPD network itself.

Conclusion: Kai: So, to wrap up, this paper is about taking generalized parton distributions or GPDs and extracting them from Compton form factors by embedding the principal-value kernel directly into a differentiable neural network structure.

Mira: It’s a method that uses machine learning to solve that tricky inverse problem by making the mathematical physics of quantum chromodynamics part of the network's architecture, rather than just treating it as an external constraint.

Lev: And they show you can get a probabilistic extraction, which means you get both a mean value and a band showing the uncertainty across all those different input possibilities.

Kai: It also shows that we can train this end-to-end with CFF inputs, without needing to guess the functional form of the GPD beforehand. That flexibility for fitting to fixed target or collider data is really significant for how we explore this nucleon structure.

Mira: It moves us toward a more direct link between what we measure in deep virtual Compton scattering and the fundamental structure of the nucleon itself, bypassing those older, restrictive ways of parameterizing the distributions.

Lev: From an engineering standpoint, it’s powerful because it handles complex kinematics well, but we still have to manage that replica ensemble overhead when you try to run this on real hardware for high fidelity measurements.

Kai: It’s a lot of work getting the initial CFF data right, though. The quality of the final GPD extraction is ultimately bounded by how accurate those input form factors are, because any systematic errors in the CFF model propagate into our final result.

Mira: That’s true. The paper also confirmed they can diagnose polynomiality violation using moments one and three to check if their reconstruction respects basic QCD rules, which is a good way to validate the physical meaning of the output.

Lev: I think the ability to use this for combined fits across different energy scales is what makes it most useful for real experimentalists right now. It’s not just theoretical work on a clean paper; it’s about making measurements more powerful on existing hardware.

More episodes

← Home