Taming Symbolic IBP Reduction with Intermediate Bases

arXiv:2606.22500 · hep-ph · Submitted 2026-06-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

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

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Taming Symbolic IBP Reduction with Intermediate Bases".

Mira: Here's the gist of the paper. Reconstructing all reduction coefficients for Feynman integrals through a sequence of intermediate bases offers an efficient method for symbolic IBP reduction <ref:2606.22500#pg2>.

Kai: The resulting analytic reduction coefficients are products of a few analytic matrices, whose nonzero entries are simple rational polynomials <ref:2606.22500#pg2>.

Mira: The Core Idea The basic idea of this approach to express complicated Feynman integrals in terms of simpler ones, which we call an intermediate basis, by row reducing a smaller set of IBP relations <ref:2606.22500#pg2>.

Kai: These reduction rules are then inserted back into the original IBP system, so that the complicated integrals no longer appear <ref:2606.22500#pg2>.

Title and authors: Kai: So we're talking about the paper "Taming Symbolic IBP Reduction with Intermediate Bases" and its authors Qian Song from Ghent University. It sounds like they are tackling a really hard problem in how we reduce these Feynman integrals.

Mira: Yeah, it seems like they’re focusing on the fact that getting all those reduction coefficients—which are rational polynomials in kinematic variables and space-time dimension—is just too hard because there are so many unknowns in the general setup, leading to a linear system that's way too big to solve directly.

Lev: From an engineering side, that means we can't just write down a massive matrix and try to invert it; we need something much more structured for practical calculation on hardware.

Kai: The paper suggests they have found an algorithm using a sequence of intermediate bases to solve this, and the main result is that instead of one huge transformation matrix with complicated entries, all the reduction coefficients end up being a product of just a few sparse matrices whose entries are simple rational polynomials.

Mira: That decomposition is key because it means we can represent those complicated functions in a much more compact analytic form, which really helps when you need to do fast numerical evaluations because those complicated rational polynomials get broken down into simpler pieces.

Lev: If the resulting structure is a product of sparse matrices, that should mean the computational cost for reconstructing these coefficients drops significantly compared to whatever the naive approach requires.

Kai: They even show this efficiency with two examples, like a three-point massive box-triangle and a four-point massive pentagontriangle. For that box-triangle example, they say you only need three thousand two hundred eighty-nine or thirteen thousand thirteen numerical samplings to reconstruct everything, which is way fewer than the one million four hundred seven thousand four hundred six or twenty-one million six hundred thirty-eight thousand three hundred thirty-one unknowns in the general setup <ref:2606.22500#pg1>.

Mira: That comparison between the required samplings and the number of unknowns really highlights just how much simpler this approach makes it to actually get those results.

The paper's summary: Kai: So, looking at what Qian Song did in "Taming Symbolic IBP Reduction with Intermediate Bases," they are essentially using an iterative process where they derive reduction rules by row reduction, expressing a subset of complicated Feynman integrals in terms of simpler ones. They keep doing this until the final reduction rules only depend on the master integrals.

Mira: That iterative derivation is smart; it's like building up a simplified system step-by-step, moving from the most complex integrals down to just the fundamental master ones. They use a degree bound N bar to control complexity, making sure that no rational polynomial in those basis transformation matrices has a degree higher than that N bar.

Lev: So, they are systematically simplifying the problem by reducing relations until they only involve the master integrals, which is necessary for any practical reduction scheme anyway. What I want to know is how this iterative derivation translates into a concrete calculation when we move away from pure symbolic manipulation.

Kai: They admit that in practice, symbolic row reduction of a large matrix is too heavy, so they have to set the kinematic variables and the dimension D to random numerical values. The complexity of an integral is then determined by the degree of the most complicated rational polynomial in those reduction coefficients, which they call deg r.

Mira: That complexity metric tells us how hard it is for a given integral; for instance, they show that eqs. eight through twelve imply that one type of integral has a higher order than another. This ordering helps them guide the selection of intermediate bases to manage the difficulty effectively.

Lev: So, the paper isn't just about finding a symbolic trick; it's about creating a practical algorithm where we can use numerical values for kinematics to find those coefficients, and their complexity dictates how much sampling you actually need.

The paper's improvements: Kai: One of the big wins they point out is that the bound on the number of samplings required for reconstruction is determined by the most complicated entry in those basis transformation matrices, not by the most complicated entry in all those final product forms. That’s a huge relief.

Mira: And another major improvement is that this product form representation gives a compact analytic way to describe everything, and crucially, it’s well suited for fast numerical evaluation because those complicated rational polynomials are decomposed into simpler pieces before you try to evaluate them.

Lev: If you can decompose the complicated functions into simpler pieces that are easy to evaluate numerically, that drastically cuts down on the computational time needed for those evaluations. We're talking about moving from potentially years of solving a full linear system down to just tens of seconds for matrix products.

Kai: They demonstrate this efficiency with the box-triangle example again, showing that even with these improvements, you still only need three thousand two hundred eighty-nine samplings compared to the one million four hundred seven thousand four hundred six unknowns in the general ansatz <ref:2606.22500#pg1,unknowns in the general ansatz>. That shows a significant reduction in what we actually have to run on our systems.

Mira: So, they’ve essentially found a way to maintain the analytic structure of the problem while drastically improving both the required input data and the speed of processing those results numerically. It's about getting a tighter bound on complexity that directly translates into fewer necessary calculations.

Lev: The paper also points out that this whole process can be summarized as iteratively deriving reduction rules by row reduction, which they do symbolically, but then they have to switch to numerical values for the kinematics and dimension D because doing symbolic row reduction on large matrices is just too much overhead.

Conclusion: Kai: So, wrapping up on "Taming Symbolic IBP Reduction with Intermediate Bases," the main point is that this method allows us to reconstruct all reduction coefficients through a sequence of intermediate bases, where those coefficients are products of a few sparse matrices with simple rational polynomial entries.

Mira: The big implication for us is that we can get an analytic representation that is compact and suitable for fast numerical evaluation because the complicated rational polynomials are decomposed into simpler pieces before they are used.

Lev: For the real world, this means we can significantly reduce the sampling requirements, as the bound on those samplings is now tied to a specific matrix element rather than the worst-case complexity of all those products. It also promises much faster and more stable numerical evaluations for these reduction coefficients.

Kai: The efficiency is backed up by their examples, showing that for complex topologies like the box-triangle, you only need about three thousand two hundred eighty-nine or thirteen thousand thirteen samplings compared to the massive number of unknowns in the general ansatz <ref:2606.22500#pg1,number of unknowns in the general ansatz>.

Mira: It’s a very flexible method that lets us tame symbolic IBP reduction without getting bogged down by solving an impossibly large linear system directly. We expect this kind of approach to be very important for perusing analytical results for scattering amplitudes at higher loop orders.

Lev: For real hardware, if the overhead lands in the tens of seconds range for matrix products instead of years, that’s a massive practical win when you're trying to get high-precision results on your superconducting qubits.

Kai: So that’s what this paper is about: using intermediate bases to find a more efficient way to reconstruct those reduction coefficients. Thanks for listening.

Department of Physics and Astronomy, Ghent University

hep-ph

Submitted: 2026-06-21

Updated: 2026-11-08

Comments: 6 pages, 3 figures; v2: minor revisions to the text, matches the published version Phys.Rev.D 114 (2026) 9, 096004

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Reconstructing all reduction coefficients for Feynman integrals through a sequence of intermediate bases offers an efficient method for symbolic IBP reduction <ref:2606.22500#pg2>.

Key concepts

Intermediate Basis
A set of simpler Feynman integrals used to express more complicated ones. These bases are found by iteratively deriving reduction rules through row reduction on a smaller set of IBP relations. This process expresses complex integrals in terms of simpler ones, eventually relating everything back to master integrals.
Reduction Coefficients Representation
The coefficients that relate different integral representations are not stored as one large matrix but as a product of several sparse matrices. The entries in these matrices are simple rational polynomials. This product form is highly efficient for fast numerical evaluation because complicated polynomials are broken down into simpler components.
Complexity Metric (deg r)
The complexity of an integral's reduction coefficients is measured by the degree of the most complicated rational polynomial involved, denoted as deg r. This degree guides the ordering of integrals and controls the required number of numerical samplings needed to reconstruct the full set.

Terminology

Summary

Reconstructing all reduction coefficients for Feynman integrals through a sequence of intermediate bases offers an efficient method for symbolic IBP reduction <ref:2606. p. 322500#pg2>. The resulting analytic reduction coefficients are products of a few analytic matrices, whose nonzero entries are simple rational polynomials <ref:2606. p. 022500#pg2>.

The Core Idea

The basic idea of this approach to express complicated Feynman integrals in terms of simpler ones, which we call an intermediate basis, by row reducing a smaller set of IBP relations <ref:2606.22500#pg2>. These reduction rules are then inserted back into the original IBP system, so that the complicated integrals no longer appear <ref:2606. p. 022500#pg2>. Iterating this procedure eventually expresses all integrals in terms of the master integrals <ref:2606. p. 022500#pg2>. Instead of one transformation matrix with complicated entries, all reduction coefficients are represented as a product of a few sparse matrices whose entries are simple rational polynomials <ref:2606. p. 022500#pg2>.

Advantages of the Representation

The representation in eq. (1) has two advantages <ref:2606. p. 022500#pg2>. First, the number of samplings required for reconstructing M is significantly fewer <ref:2606. p. 022500#pg2>. It is bounded by the most complicated entry in the basis transformation matrices, rather than by the most complicated entry of the products <ref:2606. p. 022500#pg2>. Second, the product form gives a compact analytic representation that is well suited for fast numerical evaluation, since complicated rational polynomials are decomposed into simpler pieces <ref:2606. p. 022500#pg2>.

The Algorithm for Intermediate Bases

The algorithm starts with a complete set of IBP relations, the list of all Feynman integrals appearing in IBP relations, and a degree bound N¯ <ref:2606.22500#pg2>. The process involves iteratively deriving reduction rules by row reduction, expressing a subset of complicated Feynman integrals in terms of simpler ones, until the final reduction rules only depend on the master integrals <ref:2606. p. 222500#pg2>. The algorithm terminates when the number of integrals in the intermediate basis equals the number of the master integrals <ref:2606. p. 222500#pg2>.

Efficiency Demonstrated by Examples

The efficiency is demonstrated with two cutting edge examples: a three-point massive box-triangle and a four-point massive pentagontriangle <ref:2606.22500#pg2>. Reconstructing all reduction coefficients for the box-triangle (pentagon-triangle) requires 3289 (13013) numerical samplings, significantly fewer than the number of unknowns in the general ansatz, 1407406 (21638331) <ref:2606. p. 022500#pg2>.

Complexity Metrics

The complexity of a Feynman integral is determined by the degree of the most complicated rational polynomial in its reduction coefficients <ref:2606.22500#pg2>. The degree is denoted as deg r = deg(d1, d2) where d1(2) represents the degree of the numerator (denominator) polynomial <ref:2606. p. 122500#pg2>. This complexity is used to order Feynman integrals, such as I[2,2] > I[2,1] = I[1,2] > I[0,2] = I[2,0] <ref:2606. p. 222500#pg4>. The degree bound N¯ serves as an upper bound on the rational polynomials in basis transformation matrices and controls the samplings number required for reconstruction <ref:2606. p. 222500#pg2>.

Conclusion

In this paper, we present a new method for reconstructing the symbolic dependence of IBP reduction coefficients on the kinematic variables and the space-time dimension from numerical IBP reduction <ref:2606.22500#pg2>. The number of samplings needed for reconstruction is much smaller compared to the number of unknowns in a general ansatz <ref:2606. p. 322500#pg2>. The reduction coefficients are expressed as a product of several matrices, which enables fast and stable numerical evaluations <ref:2606. p. 322500#pg2>. This method is very efficient and flexible, and we expect it to be of crucial importance for perusing analytical result for scattering amplitude at higher loop orders <ref:2606. p. 322500#pg2>.

References

[1] K. G. Chetyrkin and F. p. 3 V. Tkachov, “Integration by parts: The algorithm to calculate β-functions in 4 loops,” Nucl. p. 3 Phys. B 192, 159–204 (1981) <ref:2606. p. 322500#pg2>

[2] S. Laporta, “High-precision calculation of multiloop Feynman integrals by difference equations,” Int. p. 3 J. Mod. Phys. A 15, 5087–5159 (2000), arXiv:hep-ph/0102033 <ref:2606. p. 322500#pg2>

[3] Charalampos Anastasiou and Achilleas Lazopoulos, “Automatic integral reduction for higher order perturbative calculations,” JHEP 07, 046 (2004), arXiv:hepph/0404258 <ref:2606.22500#pg2>

[4] R. N. Lee, “Presenting LiteRed: a tool for the Loop Integrals REDuction,” (2012), arXiv:1212. p. 42685 [hep-ph] <ref:2606. p. 022500#pg6>

[5] Roman N. Lee, “LiteRed 1. p. 44: a powerful tool for reduction of multiloop integrals,” J. p. 4 Phys. Conf. Ser. 523, 012059 (2014), arXiv:1310. p. 41145 [hep-ph]

[6] A. V. Smirnov, “Algorithm FIRE – Feynman Integral REDuction,” JHEP 10, 107 (2008), arXiv:0807. p. 43243 [hepph] <ref:2606. p. 422500#pg6>

[7] A. V. Smirnov and V. p. 4 p. 4 A. Smirnov, “FIRE4, LiteRed and accompanying tools to solve integration by parts relations,” Comput. p. 4 Phys. Commun. 184, 2820–2827 (2013), arXiv:1302. p. 45885 [hep-ph]

[8] Alexander V. Smirnov, “FIRE5: A C++ implementation of Feynman Integral REduction,” Comput. p. 4 Phys. Commun. 189, 182–191 (2015), arXiv:1408. p. 42372 [hep-ph]

[9] A. V. Smirnov and F. p. 4 S. Chukharev, “FIRE6: Feynman Integral REduction with modular arithmetic,” Comput. p. 4 Phys. Commun. 247, 106877 (2020), arXiv:1901. p. 407808 [hep-ph]

[10] C. Studerus, “Reduze – Feynman integral reduction in C++,” Comput. p. 4 Phys. Commun. 181, 1293–1300 (2010), arXiv:0912. p. 42546 [physics. p. 4comp-ph]

[11] A. von Manteuffel and C. p. 4 Studerus, “Reduze 2 - Distributed Feynman Integral Reduction,” (2012), arXiv:1201. p. 44330 [hep-ph]

[12] Philipp Maierh¨ofer, Johann Usovitsch, and Peter Uwer, “Kira—A Feynman integral reduction program,” Comput. Phys. Commun. 230, 99–112 (2018), arXiv:1705. p. 405610 [hep-ph]

[13] Jonas Klappert, Fabian Lange, Philipp Maierh¨ofer, and Johann Usovitsch, “Integral reduction with Kira 2.0 and finite field methods,” Comput. p. 4 Phys. Commun. 266, 108024 (2021), arXiv:2008. p. 406494 [hep-ph] <ref:2606. p. 022500#pg6>

[14] Zihao Wu, Janko Boehm, Rourou Ma, Hefeng Xu, and Yang Zhang, “NeatIBP 1.0, a package generating small-size integration-by-parts relations for Feynman integrals,” Comput. p. 4 Phys. Commun. 295, 108999 (2024), arXiv:2305. p. 408783 [hep-ph]

[15] Alexander V. Smirnov and Mao Zeng, “Feynman integral reduction: balanced reconstruction of sparse rational functions and implementation on supercomputers in a codesign approach,” Numerical Methods and Programming 25, 30–45 (2024), arXiv:2409. p. 419099 [hep-ph]

[16] Andreas von Manteuffel and Robert M. Schabinger, “A novel approach to integration by parts reduction,” Phys. p. 4 Lett. B 744, 101–104 (2015), arXiv:1406. p. 44513 [hep-ph]

[17] Tiziano Peraro, “Scattering amplitudes over finite fields and multivariate functional reconstruction,” JHEP 12, 030 (2016), arXiv:1608.01902 [hep-ph] <ref:2606. p. 022500#pg6>

[18] Jonas Klappert and Fabian Lange, “Reconstructing rational functions with FireFly,” Comput. Phys. Commun. 247, 106951 (2020), arXiv:1904. p. 400009 [cs.SC]

[19] Tiziano Peraro, “FiniteFlow: multivariate functional reconstruction using finite fields and dataflow graphs,” JHEP 7, 031 (2019), arXiv:1905.08019 [hep-ph] <ref:2606. p. 022500#pg6>

[20] Andreas Maier, “Scaling up to multivariate rational function reconstruction,” Comput. Phys. Commun. 317, 109827 (2025), arXiv:2409. p. 408757 [hep-ph]

[21] A. V. Belitsky, A. p. 4 V. Smirnov, and R. p. 4 V. Yakovlev, “Balancing act: Multivariate rational reconstruction for IBP,” Nucl. p. 4 Phys. B 993, 116253 (2023), arXiv:2303. p. 402511 [hep-ph]

[22] Michael Monagan, “Maximal quotient rational reconstruction: an almost optimal algorithm for rational reconstruction,” in Proceedings of the 2004 International Symposium on Symbolic and Algebraic Computation, ISSAC ’04 (Association for Computing Machinery, New York, NY, USA, 2004) p. 243–249 <ref:2606. p. 022500#pg6>

[23] Michael Ben-Or and Prasoon Tiwari, “A deterministic algorithm for sparse multivariate polynomial interpolation,” in Proceedings of the Twentieth Annual ACM Symposium on Theory of Computing, STOC ’88 (Association for Computing Machinery, New York, NY, USA, 1988) p. 301–309 <ref:2606. p. 022500#pg6>

[24] Richard Zippel, “Interpolating polynomials from their values,” (1989)

[25] Dima Grigoriev, Marek Karpinski, and Michael Singer, “Computational complexity of sparse rational interpolation,” SIAM Journal on Computing 23 (1997), 10.1137/S0097539791194069 <ref:2606. p. 422500#pg6>

[26] Erich Kaltofen, “Greatest common divisors of polynomials given by straight-line programs,” J. ACM 35, 231–264 (1988)

[27] Erich Kaltofen and Barry Trager, “Computing with polynomials given by black boxes for their evaluation: Greatest common divisors, factorization, separation of numerators and denominators,” (1988) pp. 296–305

[28] Erich Kaltofen and Zhengfeng Yang, “On exact and approximate interpolation of sparse rational functions,” in Proceedings of the 2007 International Symposium on Symbolic and Algebraic Computation, ISSAC ’07 (Association for Computing Machinery, New York, NY, USA, 2007) p. 203–210 <ref:2606. p. 022500#pg6>

[29] Jennifer de Kleine, Michael Monagan, and Allan Wittkopf, “Algorithms for the non-monic case of the sparse modular gcd algorithm,” in Proceedings of the 2005 International Symposium on Symbolic and Algebraic Computation, ISSAC ’05 (Association for Computing Machinery, New York, NY, USA, 2005) p. 124–131 <ref:2606. p. 022500#pg6>

[30] Angel D´ıaz and Erich Kaltofen, “Foxbox: a system for manipulating symbolic objects in black box representation,” in Proceedings of the 1998 International Symposium on Symbolic and Algebraic Computation, ISSAC ’98 (Association for Computing Machinery, New York, NY, USA, 1998) p. 30–37 <ref:2606. p. 022500#pg6>

[31] Erich Kaltofen, Wen-shin Lee, and Austin A. Lobo, “Early termination in ben-or/tiwari sparse interpolation and a hybrid of zippel’s algorithm,” in Proceedings of the 2000 International Symposium on Symbolic and Algebraic Computation, ISSAC ’00 (Association for Computing Machinery, New York, NY, USA, 2000) p. p. 4 192–201 <ref:2606. p. 022500#pg6>

[32] Qiao-Long Huang and Xiao-Shan Gao, “Sparse polynomial interpolation with finitely many values for the coefficients,” (2017), arXiv:1704.04359 [cs.SC]

[33] Jonas Klappert, Sven Yannick Klein, and Fabian Lange, “Interpolation of dense and sparse rational functions and other improvements in FireFly,” Comput. Phys. Commun. 264, 107968 (2021), arXiv:2004. p. 401463 [cs.MS] <ref:2606. p. 022500#pg6>

[34] Michael S. Floater and Kai Hormann, “Barycentric rational interpolation with no poles and high rates of approximation,” Numerische Mathematik 107, 315–331 (2007)

[35] Johannes M. Henn, “Multiloop integrals in dimensional regularization made simple,” Phys. p. 4 Rev. Lett. 110, 251601 (2013), arXiv:1304. p. 41806 [hep-th]

[36] Roman N. Lee, “Reducing differential equations for multiloop master integrals,” JHEP 04, 108 (2015), arXiv:1411. p. 40911 [hep-ph]

[37] Xiao Liu, “Reconstruction of rational functions made simple,” Phys. Lett. B 850, 138491 (2024), arXiv:2306. p. 512262 [hep-ph]

[38] A. V. Smirnov and V. A. Smirnov, “How to choose master integrals,” Nucl. p. 5 Phys. B 960, 115213 (2020), arXiv:2002. p. 508042 [hep-ph]

[39] Johann Usovitsch, “Factorization of denominators in integration-by-parts reductions,” (2020), arXiv:2605.13256 [hep-ph]

[40] S. Abreu, J. Dormans, F. p. 5 Febres Cordero, H. p. 5 Ita, and B. p. 5 Page, “Analytic Form of Planar Two-Loop Five-Gluon Scattering Amplitudes in QCD,” Phys. p. 5 Rev. Lett. 122, 082002 (2019), arXiv:1812. p. 504586 [hep-ph]

[41] Giuseppe De Laurentis and Ben Page, “Ans¨atze for scattering amplitudes from p-adic numbers and algebraic geometry,” JHEP 12, 140 (2022), arXiv:2203.04269 [hepth]

[45] Sid Smith and Mao Zeng, “Feynman integral reduction using syzygy-constrained symbolic reduction rules,” JHEP 01, 102 (2026), arXiv:2507.11140 [hep-th]

[46] Leonardo de la Cruz and David A. Kosower, “Seedless Reduction of Feynman Integrals,” (2026), arXiv:2602. p. 522111 [hep-ph] <ref:2606. p. 022500#pg6>

[47] Such an estimation is based on evaluation on 32-bit prime number with private linear solver

[48] Philipp Maierh¨ofer and Johann Usovitsch, “Kira 1.2 Release Notes,” (2018), arXiv:1812. p. 501491 [hep-ph]

[49] Fabian Lange, Johann Usovitsch, and Zihao Wu, “Kira 3: integral reduction with efficient seeding and optimized equation selection,” Comput. Phys. Commun. 322, 109999 (2026), arXiv:2505. p. 5

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

  1. Improved AI can reconstruct all reduction coefficients for complex Feynman integrals by using a sequence of intermediate bases, which are expressed as products of a few sparse matrices whose entries are simple rational polynomials. This allows for a compact analytic representation well suited for fast numerical evaluation because complicated rational polynomials are decomposed into simpler pieces.

  2. The improved system can significantly reduce the number of required samplings for reconstruction, as the bound is bounded by the most complicated entry in the basis transformation matrices, rather than by the most complicated entry of the products. This is demonstrated in examples where reconstruction requires significantly fewer samplings compared to solving a linear system based on a general ansatz.

  3. The AI can find intermediate bases through an iterative process motivated by row reduction, starting with an intuitive example, which allows it to express complicated Feynman integrals in terms of simpler ones. This method controls complexity by setting a degree bound N bar and ensures that the basis exchange matrix contains no rational polynomial whose degree is higher than N bar.

  4. The system can efficiently identify the most complicated reduction coefficients for high-order topologies, as shown by finding degrees like deg 35, 36 in five variables for the box-triangle example. This complexity information is used to guide the selection of intermediate bases, which are ordered according to their reconstruction complexity.

  5. The AI can handle large systems efficiently by replacing a large linear system with a product form, such as the representation in equation (1). This product form enables fast and stable numerical evaluations and allows for the evaluation of matrix products to take only 10s of seconds, compared to potentially years required for solving the full linear system.

Sources

Related papers