DiffExp: Feynman Integrals & Diffusion Tuning

• DiffExp is a dual-purpose research tool offering a Mathematica package for multi-loop Feynman integral evaluation and an exploration strategy for diffusion model fine-tuning.
• Its Feynman integral variant employs power series, analytic continuation, and Jordan block analysis to enable high-precision collider computations.
• The diffusion model approach enhances reward fine-tuning by using dynamic classifier-free guidance scheduling and random prompt-weighting to boost sample diversity.

DiffExp is the designation for two distinct but widely referenced research tools in contemporary computational science: (1) a Mathematica package for the efficient solution of systems of differential equations governing multi-loop Feynman integrals, and (2) an exploration strategy for reward fine-tuning in text-to-image diffusion models. Both share an underpinning in high-dimensional optimization and efficient exploration or evaluation of large solution spaces, but they serve disparate domains: precision collider phenomenology and diffusion-based deep learning, respectively. Below, both branches are elucidated.

1. Differential Equation-Based Feynman Integral Evaluation with DiffExp

The "DiffExp" Mathematica package (Hidding, 2020) is a general-purpose computational tool for evaluating multi-loop Feynman master integrals. It constructs one-dimensional (possibly generalized) series solutions to systems of first-order linear differential equations derived via integration-by-parts (IBP) reduction. The package is particularly effective for high-loop, multi-scale processes where closed-form analytic results are unavailable or unwieldy, as in recent three-loop Standard Model calculations (Greub et al., 2024, Canko et al., 2024).

Mathematical and Algorithmic Foundation

DiffExp operates on differential equations of the form

$$\frac{d}{dx}\,\vec I(x,\epsilon) = \mathbf A_x(x,\epsilon)\,\vec I(x,\epsilon)$$

where $\vec I$ is the vector of master integrals, $x$ parameterizes a one-dimensional kinematic path, and $\epsilon$ is the dimensional regulator. The system is expanded order-by-order in $\epsilon$ after normalization to remove $1/\epsilon$ poles, resulting in a triangular, recursive structure for the $O(\epsilon^k)$ components.

Local solutions are constructed via power series, possibly including logarithmic and Puiseux terms when the system has nontrivial Jordan blocks or half-integer exponents. Singularities are crossed via analytic continuation, with path segmentation to ensure convergence and control truncation errors.

Boundary conditions are provided as $\epsilon$-series at specified points, typically derived either from fully analytic results at special kinematic configurations or via expansion-by-regions asymptotics.

Implementation Workflow

The DiffExp pipeline is as follows:

1. Preparation of Input: Differential equation systems and boundary values are expressed in Mathematica rule-sets; variables and expansion order in $\epsilon$ are specified.
2. Initialization: Families of integrals and their canonical master bases are registered in the DiffExp environment, allowing the system to construct the appropriate coupled ODEs.
3. Series Solution: Given a target kinematic point, the path from the boundary is subdivided at singularities; on each segment, local series (power or generalized) are generated and analytically continued.
4. Numerical Evaluation: Series are truncated at user-specified order based on desired precision and proximity to singularities, providing high-precision numerical results for each master integral up to $O(\epsilon^6)$ or higher.
5. Output and Postprocessing: Results are written as Mathematica rules, with automated LaTeX export for tabulation.

A typical input-output workflow is shown below:

<< DiffExp.m InitFamily["PL1", "Variables" -> {s12, s23, m3^2, m4^2}, ...]; sol = SeriesSolution["PL1", "TargetPoint" -> {s12 -> 3, s23 -> 1, ...}, "SeriesOrder" -> {ε, 0, 6}, "PrecisionGoal" -> 40]; Export["output_PL1.m", sol, "Package"];

Practical Applications

DiffExp has been employed for:

• Three-loop calculations in rare-decay amplitudes, such as $b \to s\gamma$ with full mass dependence (Greub et al., 2024).
• Determination of integrals for two off-shell vector boson production with general mass scales (ladder-box, tennis-court, and reducible families) (Canko et al., 2024).
• Multi-scale, multi-loop scalar Feynman diagram evaluations, including the five-propagator "banana" graphs.

In these applications, the package's automation of Fuchsian reduction, Jordan block analysis, and analytic continuation via segmentwise series expansions is central to overcoming previous performance and generality bottlenecks.

A summary of performance metrics from (Canko et al., 2024) is provided:

Family	# Masters	Timing (sij)	Timing (xyz)
RL₁	27	175 s	55 s
PL₁	189	2765 s	818 s
PT₄	150	4987 s	1478 s

Working in rationalized coordinates reduces segmentation and accelerates computation.

Error Control, Limitations, and Future Directions

The package estimates truncation errors via comparison of successive series orders and warns when the radius of convergence is approached. For massive phenomenological studies, precomputing a dense grid of boundary expansions is recommended. The Mathematica implementation is a limiting factor for very large problems, and a compiled backend is expected to yield substantial speedups. Currently, block-triangular structure is not fully exploited; more efficient solvers could further reduce computational overhead (Canko et al., 2024).

2. DiffExp: Efficient Exploration in Reward Fine-tuning for Diffusion Models

A second instance of "DiffExp" designates a lightweight exploration enhancement for reward optimization in text-to-image diffusion models (Chae et al., 19 Feb 2025). This "DiffusionExplore" strategy, intended for policy-gradient (DDPO) or direct backpropagation (AlignProp) fine-tuning, is characterized by two exploration levers: dynamic classifier-free guidance (CFG) scheduling, and random prompt-phrase weighting.

Methodology and Theoretical Basis

In reward-based fine-tuning of diffusion models, exploration is often limited by the sample inefficiency of online image generation. DiffExp addresses this by:

• Dynamic CFG scheduling: Rather than a static guidance scale $w$, DiffExp schedules $w(t)$ so that early denoising steps use minimal guidance ($w_l \approx 0.1$), maximizing latent mode diversity, before switching to high guidance ($w_h$) late in the chain to ensure sample quality. This increases the breadth of sampled image configurations capable of achieving high reward.
• Random prompt-phrase weighting: During sampling, a random token in the text prompt is up-weighted in its embedding, perturbing the model to more strongly condition on a random phrase and thereby increasing compositional variability.

These mechanisms are applied stochastically during fine-tuning but omitted at evaluation. The dynamic CFG introduces early-stage high-entropy exploration analogous to episodic RL, while prompt-phrase weighting traverses more of the text-conditioned manifold.

The general training loop is:

initialize θ (or LoRA) for iteration = 1 to max_iters: for each c_j in batch: pick random token i, sample w_p ∼ U(1.0,1.2) form c̃_j with up-weighted c_j[i] x₀^{(j)} ← DiffusionSample(θ, c̃_j, dynamic w(t)) r_j ← RewardModel(x₀^{(j)}, c_j) Update θ with policy gradient or reward backprop

Empirical Performance and Benchmarks

In experiments with Stable Diffusion v1.5, DiffExp demonstrates:

• For both DDPO and AlignProp optimizers, a ∼20% reduction in sample requirement for reaching target rewards (on-seen prompts) compared to fixed-guidance baselines.
• Generalization: On unseen prompts and challenging DrawBench tasks, DiffExp-augmented models achieve higher reward trajectories.
• When evaluated with a held-out reward function (ImageReward), image quality improvement translates to 10–15% higher scores.
• Ablations confirm both exploration mechanisms are beneficial independently, but combined they yield the most pronounced gains.

Implementation Specifics

Key hyperparameters and settings include:

• Inference steps: 50
• $w_l = 0.1$, $w_h = 5.0$ (DDPO) / $7.5$ (AlignProp)
• Prompt weight $w_p \sim \mathcal{U}(1.0, 1.2)$
• $t_{\rm thresh}=900$ (of $T=1000$ denoising steps)
• Batch size: 64
• Learning rates: $3 \times 10^{-4}$ (DDPO), $1 \times 10^{-3}$ (AlignProp)
• AdamW optimizer, $\beta_1=0.9$, $\beta_2=0.99$

Guidelines are given for tuning the exploration phase ($t_{\rm thresh}$), restricting random prompt weighting to training, and maintaining a low early guidance scale for maximal exploration (Chae et al., 19 Feb 2025).

Limitations and Outlook

The exploration techniques in DiffExp introduce transient variance that must be annealed for robust convergence, and there is potential for incorporating adaptive schedules or more semantically targeted prompt perturbations. The approach is agnostic to prompt semantics, suggesting further integration with uncertainty-based or entropy-regularized RL techniques might enhance control.

3. Case Studies in Collider and Multi-Loop Quantum Field Theory

DiffExp's Feynman integral variant has played a critical role in current-high precision collider calculations:

• In the computation of three-loop master integrals for $b \to s\gamma$, DiffExp was used to automate Fuchsification, Jordan decomposition, and to construct $z^n \log^k z$ expansions about singularities (e.g., $z=0$, $z=1/10$), ensuring analytic continuation and matching at switch-over points. Agreement with independent AMFlow-based results and with recent analytic asymptotics validated the approach (Greub et al., 2024).
• For cross-section calculations involving two off-shell vector bosons of differing masses, DiffExp enabled efficient high-precision evaluation of large master-integral families (e.g., 189-dimensional for ladder-box families), exploiting variable rationalization to reduce computational segmentation (Canko et al., 2024).

4. Package Configuration and Usage Details

DiffExp is distributed as a Mathematica package, requiring user-supplied differential-equation specifications, variable lists, and boundary data. Primary routines include:

• PrepareBoundaryConditions
• IntegrateSystem
• TransportTo
• SeriesSolution

Precision control is managed via options specifying truncation order, working precision, and series segmentation strategy. Output can be formatted for immediate numerical evaluation or for LaTeX typesetting.

Recommendations for large-scale practical use include:

• Rationalizing variables to minimize branch points and segments required.
• Precomputing base-point expansions for batch or grid-based evaluations.
• Parallelizing over large families or kinematic scan points.
• Employing Padé approximants for convergence acceleration, balanced by precision settings (Canko et al., 2024).

5. Comparative Impact and Distinctive Features

Both instantiations of DiffExp share a design philosophy centered on efficient exploration of high-dimensional solution spaces, either via analytic series expansion (Feynman integral package) or explicit stochastic strategies (diffusion model fine-tuning). The Feynman-integral DiffExp is notable for its automation of analytic continuation, Jordan block management, and error estimation without manual intervention, directly facilitating high-precision, multi-loop results. The diffusion model DiffExp is distinguished by its simplicity—two minimal, computationally free augmentations—yet yields significant improvements in reward-fine-tuning throughput and sample diversity, with plug-and-play compatibility for existing policy-gradient and direct-backprop pipelines (Chae et al., 19 Feb 2025).

Future development directions for both frameworks include expanded automation, block structure exploitation, and integration of more sophisticated exploration criteria, such as entropy regularization or significance-driven prompt weighting. In both the quantum field theory and deep learning contexts, DiffExp has become a key tool for scalable, high-fidelity computation.