Two-loop leading-color QCD corrections
for Higgs plus two-jet in the heavy-top limit

Giuseppe De Laurentis
CERN - TH

two-loop leading-color $pp\rightarrow Hjj$ amplitudes in heavy-top limit

arXiv:2605.04009 / JHEP07(2026)251 with H. Ita, V. Kuschke, M. Ruf, and V. Sotnikov

see also:
one-loop $pp\rightarrow Hjj$, $pp\rightarrow HHj$, $gg\rightarrow HHH$, $q\bar q\rightarrow t\bar tH$, amplitudes in the SM
arXiv:2002.04018, 2408.12686, 2507.19313, 2504.19909, with K. Ellis, J. Campbell et al.

The 23rd Workshop of the LHC Higgs Working Group

$\;$


Find these slides at gdelaurentis.github.io/slides/higgsnov2026

Introduction

Representative Two-Loop Hjj Diagrams
Process
$N_c^2$
$N_cN_f$
$N_f^2$
$gg\to ggH$
$u\bar u\to ggH$
$u\bar u\to d\bar dH$
Generalised leading-colour components: $\mathcal{A}_{ggggH}^{(2)} = N_c^2 A^{\{2,0\}}_{ggggH} \,+\, N_c N_f A^{\{1,1\}}_{ggggH} \,+\, N_f^2 A^{\{0,2\}}_{ggggH} \,+\, $ subleading
Number of propagators 3 4 5 6 7 8
$A^{\{2,0\}}_{ggggH}$ 9 99 342 486 300 66
$A^{\{1,1\}}_{ggggH}$ 9 211 626 698 340 60
Number of cut diagrams (vertices are trees, tree counts as 1 diagram)

Motivation (Theoretical & Phenomenological)

$\circ\,$ Theoretical motivation for $H+nj$ in the HTL
$\;\small\rhd\,$ Two-loop \(Hjj\) amplitudes enter NNLO \(Hjj\),
$\;\phantom{\small\rhd}\,$ N³LO \(Hj\), and N⁴LO inclusive \(H\) production;
$\;\small\rhd\,$ Limited knowledge at higher loops/points;
$\;\small\rhd\,$ Phenomenology can be hindered by complexity
$\;\phantom{\small\rhd}\,$ of results, especially if IR cancellations are needed.
$\circ\,$ Phenomenological motivation for ggF $Hjj$
$\;\small\rhd\,$ is an important irreducible background to VBF;
$\;\small\rhd\,$ is a probe of the CP structure of Higgs couplings;
$\;\small\rhd\,$ Precision requires higher-order QCD corrections.
The HTL enables higher-order precision, but finite-$m_t$ effects become essential at high energies.

Amplitude Computation

Partial Amplitudes & Finite Remainders

$\circ$ Amplitude (integrands) can be written (for a suitable choice of master integrals) as

$$ \displaystyle A(\lambda, \tilde\lambda, \ell) = \sum_{\substack{\Gamma,\\ i \in M_\Gamma \cup S_\Gamma}} \, c_{\,\Gamma,i}(\lambda, \tilde\lambda, \epsilon) \, \frac{m_{\Gamma,i}(\lambda\tilde\lambda, \ell)}{\textstyle \prod_{j} \rho_{\,\Gamma,j}(\lambda\tilde\lambda, \ell)} \;\; \xrightarrow[]{\int d^D\ell} \;\; \sum_{\substack{\Gamma,\\ i \in M_\Gamma}} \frac{ \sum_{k=0}^{\text{finite}} \, {\color{red}c^{(k)}_{\,\Gamma, i}}(\lambda, \tilde\lambda) \, \epsilon^k}{\prod_j (\epsilon - a_{ij})} \, {\color{orange}I_{\Gamma, i}}(\lambda\tilde\lambda, \epsilon) $$
$\circ$ $\Gamma$: topologies $\quad\circ$ $M_\Gamma$: master integrands $\quad\circ$ $S_\Gamma$: surface terms
Abreu, Dormans,
Febres Cordero, Ita
Kraus, Page, Pascual,
Ruf, Sotnikov ('20)
$\circ$ All physical information is contained in the finite remainders, at two loops
Weinzierl ('11)
$$ \underbrace{\mathcal{R}^{(2)}}_{\text{finite remainder}} = \mathcal{A}^{(2)}_R \underbrace{- \quad I^{(1)}\mathcal{A}^{(1)}_R \quad - \quad I^{(2)}\mathcal{A}^{(0)}_R}_{\text{divergent + convention-dependent finite part}} + \mathcal{O}(\epsilon) $$
Catani ('98) Becher, Neubert ('09) Gardi, Magnea ('09)
$\phantom{\circ}$ $\mathcal{A}^{(1)}_R$ to order $\epsilon^2$ is still needed to build $\mathcal{R}^{(2)}$, but there is no real physical reason to reconstruct it.
$\circ$ Finite remainder as a weighted sum of pentagon functions + Zoia ('21) Chicherin, Sotnikov ('20) + Abreu, Ita, Page, Tschernow ('23)
$$ \textstyle \mathcal{R}(\lambda, \tilde\lambda) = \sum_i \color{red}{r_{i}(\lambda,\tilde\lambda)} \, \color{orange}{G_i(\lambda\tilde\lambda)} $$

Numerical Computation

$\circ\,$ Direct analytic computations are unfeasible, complexity explodes at intermediate steps.
$\circ\,$ Floating-point evaluations can take hours per point, and lose all precision due to cancellations.
$\circ\,$ The solution is to compute the rational coefficients using finite fields:
$\phantom{\circ}\,$ ordinary arithmetic, but modulo a prime number $p$.
$$ \mathbb{F}_p = \{0,1,\ldots,p-1\}, \qquad \underbrace{\frac{1}{2} = 4 \pmod 7}_{2\times4=8\equiv1\pmod7} $$
Exact arithmetic, no rounding errors! Fixed size, no running out of memory!
However, these are not quite the numbers you would want in your Monte Carlo...
$\circ\,$ We have to reconstruct the analytic rational functions from such numerical samples,
$\mathbb{F}_p$: von Manteuffel, Schabinger ('14); Peraro ('16); $\mathbb{C}$: GDL, Maître ('19); $\mathbb{Q}_p$: GDL, Page ('22)
$$ \text{Black box} \quad\rightarrow\quad r_i\big|_{\mathbb{F}_p} \quad\xrightarrow{\text{reconstruct}}\quad r_i(\lambda,\tilde\lambda) \quad\rightarrow\quad \text{phenomenology} $$
Goal: compact analytic expressions that are fast and stable to evaluate (in $\mathbb{C}$).

Rational functions of what exactly? A toy example

$\circ$ Manifesting amplitude's gauge and Lorentz invariance, and spin and little-group covariance requires
variables subject to constraints.
$\circ$ Consider polynomials $f, g, h$ in two variables $x, y$. They live in a polynomial ring:
$$ \displaystyle f(x,y), g(x, y), h(x, y) \in \mathbb{Q}[x, y] \, . $$
$\circ$ Now, restrict them, e.g. on the unit circle $(x^2+y^2-1)$
$$ \displaystyle f(x,y) \approx g(x, y) + h(x, y) (x^2+y^2-1) \, , $$
$\phantom{\circ}$ we should consider $f(x,y)$ and $g(x, y)$ as equivalent, for any $h(x,y)$.
$\circ$ The structure is that of a polynomial quotient ring
$$ \displaystyle \mathbb{Q}[x, y] \big/ \big\langle x^2+y^2-1 \big\rangle \\[2mm] $$
$\phantom{\circ}$ its elements are equivalence classes of polynomials.
$\circ$ $\big\langle x^2+y^2-1 \big\rangle \subset \mathbb{Q}[x, y]$ is an example of an ideal, the infinite set of polynomials
$\phantom{\circ}$ $h(x, y) (x^2+y^2-1)$ that vanishes on the unit circle.

Spinor Component and Invariant Rings

$\circ$ There are two sources of redundancies: kinematic constraints & Schouten/Gram identities.
$\circ$ For $n$-point massless scattering, the covariant quotient ring is
$$ \displaystyle \kern10mm R_{n} = \mathbb{F}\Big[|1⟩_{\alpha}, [1|_{\dot\alpha}, \dots, |n⟩_{\alpha}, [n|_{\dot\alpha} \Big] \Big/ \Big\langle {\textstyle \sum_{i=1}^n} |i\rangle[ i | \Big\rangle $$
$\phantom{\circ}$ The space has $4n$ dimension subject to $4$ momentum-conservation constraints.
$\circ$ With a single massive leg, e.g. $pp \rightarrow Hjj$, we have $R_{n-1}$ without momentum conservation.
$$ \displaystyle \kern10mm R_{Hjj} = \mathbb{F}\big[|1⟩_{\alpha}, [1|_{\dot\alpha}, |2⟩_{\alpha}, [2|_{\dot\alpha}, |3⟩_{\alpha}, [3|_{\dot\alpha}, |4⟩_{\alpha}, [4|_{\dot\alpha} \big] $$
$\circ$ Lorentz invariance restricts us to the invariant subring, generated by spinor brackets:
$$ \displaystyle \displaystyle R_{Hjj}\supset R_{Hjj}^{SL(2)_L\times SL(2)_R} \cong \mathcal{R}_{Hjj} = \frac{ \mathbb{F}[\langle ij\rangle,[ij] \;:\; 1\leq i\lt j\leq4] }{ \big\langle \langle12\rangle\langle34\rangle-\langle13\rangle\langle24\rangle+\langle14\rangle\langle23\rangle, \;[12][34]-[13][24]+[14][23] \big\rangle } $$
$\circ$ The rational functions $r_i$ belong to the field of fractions of $R_n$ (or $\mathcal{R}_n \cong R_n^{SL(2)_L\times SL(2)_R}$),
$$ \displaystyle r_i(|i\rangle,[i|) = \frac{\mathcal{N}(|i\rangle,[i|)}{\mathcal{D}(|i\rangle,[i|)} \, , \quad r_i(|i\rangle,[i|) \in \text{Frac}(R_n) $$

A Tree Example and Denominator Factors

$\circ$ To not make this too abstract, here is the most complex tree amplitude
$$ {A}_g^{(0)}(1^{+}_{g}, 2^{+}_g, 3^{+}_g, 4^{-}_g, \boldsymbol{5}_H) = \Bigg[ \frac{[23]⟨4|2+3|1]^2}{⟨23⟩⟨34⟩[34]s_{234}}+ \frac{[13]⟨23⟩[23]²⟨24⟩⟨34⟩}{⟨12⟩⟨14⟩[14]⟨23⟩⟨34⟩[34]}\Bigg] + \Bigg[ \; \Bigg]_{123456\rightarrow 321456}+\\ \qquad\quad\frac{[13]²⟨4|1+3|2]²}{⟨14⟩[14]⟨34⟩[34]s_{134}}+\frac{[13]²(s_{13}-2s_{123})}{⟨12⟩[14]⟨23⟩[34]} $$
$\phantom{\circ}$ We need analogous expressions for the MI coefficients.
$\circ$ How do we get them from $\mathbb{F}_p$ evaluations only ? We first need the possible denominators
$\phantom{\circ}$ They are conjectured to be (mostly) related to the letters of the symbol alphabet
Abreu, Dormans, Febres Cordero, Ita, Page ('18)
$\circ$ Surprisingly, they are unchanged from the planar two-loop one-mass amplitudes
$$ \displaystyle \mathcal{D}_{Hjj} = \mathcal{D}_{Vjj} \subset \kern-3mm \bigcup_{\sigma \; \in \; \text{Aut}(R_6)} \sigma \circ \big\{ \langle 12 \rangle, \langle 1|2+3|1], \langle 1|2+3|4], s_{123}, \Delta_{12|34|56}, \underbrace{⟨3|2|5+6|4|3]-⟨2|1|5+6|4|2]}_{\normalsize\text{only new one at two loops!}} \big\} $$
$\phantom{\circ}$ where $s_{ij} = \langle ij \rangle [ji] $, $s_{ijk} = s_{ij}+s_{ik}+s_{jk} $, $\Delta_{ij|kl|mn}$ is quadratic in the Mandelstams, etc.






ANALYTIC RECONSTRUCTION



See antares, lips, linac, syngular, pyadic.




Reduced Denominators
$\circ\,$ We can now determine the least common denominators (LCDs),
$$ \displaystyle \mathcal{D} = \prod_j \mathcal{D}_j^{q_{ij}}(|i\rangle,[i|) \, . $$
$\phantom{\circ}\,$ Obtain the $q_{ij}$ from a univariate slice $\vec\lambda(t)$, i.e. a 1D curve.
$\circ$ The curve must intersect all varieties $V(\langle \mathcal{D}_j \rangle)$, e.g.
$$ \displaystyle |i\rangle \rightarrow |i\rangle + t a_i |\eta\rangle, \quad [i| \rightarrow [i| + t b_i [\eta| $$
$\phantom{\circ}\,$ Solve for $a_i, b_i$ such that constraints are satisfied.
Space has dimension $4n-4$,
$\mathcal{D}_j = 0$ have dimension $4n-5$,
$\vec\lambda(t)$'s have dimension 1.
Poles & Zeros $\;\Leftrightarrow\;$ Irreducible Varieties $\;\Leftrightarrow\;$ Prime Ideals
Physics $\kern24mm$ Geometry $\kern24mm$ Algebra
$\phantom{\circ}\,$ For $Hjj$ the most complicated function had a mass dim. ($\approx$ poly. degree) of 168 $\Rightarrow$ 80M parameters

Effect of Basis Change on LCDs

$\circ\,$ Change basis from a subset of the pentagon coefficients $r_{i \in \mathcal{B}}$ to $\mathbb{Q}$-linear combinations $\tilde r$,

$$ R = r_j h_j = r_{i\in \mathcal{B}} M_{ij} h_j = \tilde{r}_{i} \, O_{ii'}M_{i'j} \, h_j \, , \qquad O_{ii'}, M_{i'j}\in \mathbb{Q} $$
[6] Abreu, Febres Cordero, Ita, Klinkert, Page, Sotnikov '21
$\circ\,$ For the recent Hjj computation

Basis Change from Laurent Coefficients

$\circ\,$ By Gaussian elimination, partition the space (abusing notation for residue):

$$ \text{span}(r_{i \in \mathcal{B}}) = \underbrace{\text{column}(\text{Res}(r_{i \in \mathcal{B}}, \mathcal{D}_k^m))}_{\text{functions with the singularity}} \;\;\; \oplus \, \underbrace{\text{null}(\text{Res}(r_{i \in \mathcal{B}}, \mathcal{D}_k^m))}_{\text{functions without the singularity}} $$
$\circ\,$ Search for linear combinations that remove as many singularities as possible (while not dropping rank)

$\phantom{\circ}\,$ $N^{(\text{pole order})}_{\text{denominator factor}}$ denote null-spaces, and arrows denote intersections. Greatest observed breadth $\mathcal{O}(100k)$.
mPFD as Ideal Membership
GDL, Maître ('19) GDL, Page ('22)
$\circ$ The basis change is still not sufficient. Without knowing $\mathcal{N}$ analytically, we need to answer
$$ \frac{\mathcal{N}}{\mathcal{D}_1\mathcal{D}_2} \stackrel{?}{=} \frac{\mathcal{N}_2}{\mathcal{D}_1} + \frac{\mathcal{N}_1}{\mathcal{D}_2} $$
$\circ$ Multivariate partial fraction decompositions follow from varieties where pairs of denominator factors vanish
$$ \frac{\mathcal{N}}{\mathcal{D}_1\mathcal{D}_2} \stackrel{?}{=} \frac{\mathcal{N}_2}{\mathcal{D}_1} + \frac{\mathcal{N}_1}{\mathcal{D}_2} \; \Longleftrightarrow \; \mathcal{N} \stackrel{?}{\in} \big\langle \mathcal{D}_1, \mathcal{D}_2 \big\rangle \, \text{ i.e. } \; \mathcal{N} \stackrel{?}{=} \mathcal{N}_1 \mathcal{D}_1 + \mathcal{N}_2 \mathcal{D}_2 $$
$\cap$
$=$
$$ \langle xy^2 + y^3 - z^2 \rangle + \langle x^3 + y^3 - z^2 \rangle = \langle xy^2 + y^3 - z^2, x^3 + y^3 - z^2 \rangle = \langle 2y^3-z^2, x-y \rangle \cap \langle y^3-z^2, x \rangle \cap \langle z^2, x+y \rangle $$
$\phantom{\circ}$ This is a primary decomposition. If $\mathcal{N}$ vanishes on all branches, then the partial fraction decomposition exists.
$\circ$ In the end, with a partial fraction decomposition, the largest system we need to solve has $\boldsymbol{\approx 12k}$ unknowns


Conclusions
&
Outlook

Spinor-Helicity Hjj Remainders

$\circ$ The $pp\rightarrow Hjj$ coefficient functions are 1.5 MB of plain-text LaTeX, fast and stable to evaluate.
$\phantom{\circ}$ Matrices $M_{ij}$ account for another 22 MB. Transcendental basis at PentagonFunctions++.
$\quad\small\rhd$ The size split is: 4gH ppmm 31%; pppm 22%, pppp 5%; 2q2gH pmpm 27%, pmpp 13%, 4qH pmpm 2%.
$\quad\small\rhd$ The largest (rational number) numerator (denominator) in the functions has 8 digits (6 digits);
$\quad\small\phantom{\rhd}$ while in the rational matrices they are 28 and 23 digits respectively.
$\quad\small\rhd$ Pheno ready results for the hard functions are available at FivePointAmplitudes.
A Numerical CAS for Computations in Q-Rings

(partially work in progress)

$\circ$ antares (automated numerical to analytical reconstruction software)
$\rightarrow$ Univariate slicing, LCD determination, basis change, multivariate partial fractioning strategies,
$\phantom{\rightarrow}$ constraining of numerators, Ansatz generation and fitting strategies
$\rightarrow$ Limit analytic manipulations as much as possible, mostly relies on numerical evaluations.
$\circ$ antares-results (human readable exprs in docs) with CI tests for coefficients and/or full amplitudes
In [1]: from antares_results.Hjj.remainders import remainder
In [2]: from antares_results.Hjj.momenta import oPs
In [3]: complex(remainder("ggggH", "pmpm_2L_Nc2_Nf0", oPs))
Out[3]: (-37.04012190864545-74.17026277643683j)
In [1]: import syngular
In [2]: from antares_results.Hjj.HTL.ggggH.ppmm import lTerms
In [3]: syngular.USE_ELLIPSIS_FOR_PRINT = True
In [4]: print(lTerms[:2], lTerms[-2:])
Out[4]: [Terms("""+(1⟨3|4⟩²)/(⟨1|2⟩²)"""), ('3412𝟓', True)] [Terms("""
+(-994/9⟨1|2⟩³⟨1|3⟩⟨2|4⟩[1|2]³+...⟪45 terms⟫...+45⟨1|2⟩⟨2|4⟩⟨3|4⟩³[2|4][3|4]²)/(⟨1|2⟩³⟨2|3⟩[2|3][3|4])
+(-45/2⟨1|2⟩⁴⟨1|3⟩²[1|2]⁶[1|3]+...⟪102 terms⟫...+45/2⟨1|2⟩⟨2|3⟩⟨3|4⟩⁴[1|3]²[2|3][2|4]³[3|4])/(⟨1|2⟩⟨1|3⟩[1|3]⟨2|3⟩[2|3][2|4][3|4]⟨2|1+3|2]²)
+(-172/3⟨1|2⟩³⟨1|4⟩[1|2]⁴[1|3][2|4]+...⟪37 terms⟫...+45/2⟨1|3⟩⟨3|4⟩³[1|2][1|3][2|3][3|4]³)/(⟨1|2⟩[1|3][2|4][3|4]⟨1|2+4|1]⟨1|2+4|3])
+(45/2⟨1|2⟩³⟨1|3⟩⟨1|4⟩²[1|2]³[1|3]+...⟪76 terms⟫...-1/3⟨1|2⟩⟨1|4⟩⟨3|4⟩⁴[2|3][3|4]³)/(⟨1|2⟩²⟨2|4⟩[3|4]⟨1|3+4|2]⟨1|2+4|3])
+⟨1|3⟩(-1/3⟨1|2⟩²⟨1|4⟩³[1|2]⁴+...⟪34 terms⟫...-1/3⟨1|4⟩⟨3|4⟩⁴[2|3]²[3|4]²)/(⟨1|2⟩⟨2|4⟩⟨1|3+4|2]²⟨1|2+4|3])
+(68/3⟨1|2⟩²⟨1|4⟩³[1|2]³[1|3]²[2|4]²+...⟪51 terms⟫...-2⟨1|4⟩⟨2|4⟩⟨3|4⟩³[1|2][2|3]²[2|4][3|4]³)/([1|3]⟨2|4⟩[2|4][3|4]⟨1|3+4|2]⟨1|2+4|3]²)
+(56/3⟨1|2⟩²⟨1|4⟩²[1|2]²[1|3][2|3][2|4]²+...⟪23 terms⟫...+8⟨1|3⟩⟨3|4⟩³[2|3]³[3|4]³)/([2|4][3|4]⟨1|2+4|3]³)
+(s_124-s_234)[1|2](2063/48⟨1|2⟩³⟨1|4⟩²[1|2]³[1|4]+...⟪84 terms⟫...-6293/72⟨1|3⟩⟨1|4⟩²⟨2|4⟩⟨3|4⟩[1|2][1|4][3|4]²)/(⟨1|2+4|3]²⟨2|1+3|4]Δ_13|24|𝟓)
+(s_124-s_234)(s_123-s_134)⟨1|3⟩²⟨2|4⟩[1|2][1|3](-23/16⟨1|2⟩⟨1|3⟩[1|2]²[1|3]+...⟪10 terms⟫...-69/16⟨2|4⟩⟨3|4⟩[1|2][2|4][3|4])/(⟨1|2+4|3]⟨2|1+3|4]Δ_13|24|𝟓²)
+[1|2](-253/2⟨1|2⟩³⟨1|3⟩⟨2|4⟩[1|2]⁴+...⟪16 terms⟫...-161/2⟨1|4⟩⁴⟨3|4⟩[1|4]⁴)/(⟨1|2+4|3]⟨2|1+3|4]Δ_13|24|𝟓)
+('2143𝟓', False, '+')
+('3412𝟓', True, '+')
+('4321𝟓', True, '+')
"""), ('1243𝟓', False)]
Thank you
for your attention!


These slides are powered by:
markdown, html, revealjs, hugo, mathjax, github
Backup slides.

Discontinuity in the second derivative GDL, Ita, Kuschke, Ruf, Sotnikov

$\circ$ First (?) fully massless pseudo-threshold(?). Cross $\Sigma_5^{(3)}=0 \text{ at } t=0$ ,
$$ \Sigma_5^{(3)} \, = \, \big(s_{12}s_{24}+s_{13}(s_{14}+s_{24})+s_{123}(s_{12}+s_{14})\big)^2-4s_{13}s_{14}s_{123}s_{124} \, . $$
Bivariate Slice mPFD Algorithm
A partial work-around
$\circ$ Issues: 1) get prim. dec.; 2) sample at codim. two; 3) still need to intersect ideals
$\circ$ Build a bivariate slice (curve) that intersects all varieties $V(\langle \mathcal{D}_i, \mathcal{D}_j \rangle)$, e.g.
$$ \displaystyle |i\rangle \rightarrow |i\rangle + t_1 a_i |\eta\rangle + t_2 b_i |\theta\rangle, \quad [i| \rightarrow [i| + t_1 c_i [\eta| + t_2 d_i [\theta| $$
$\phantom{\circ}$ Such that for all $t_1, t_2$ momentum is conserved: gives 20 eqs. defining a codim.-10 variety in 24 dims
$\phantom{\circ}$ Pick sol. with syngular.Ideal.points_on_variety, impl. see lips.Particles.bivariate_slice
$\circ$ Reconstruct (Newton-interpolate) the numerators $\mathcal{N}(t_1, t_2)$; this requires $\text{deg}(\mathcal{N}+1)^2$ points.
$\circ$ For each $\mathcal{J}_{ij}^{a_i,a_j} = \big\langle \mathcal{D}_i(t_1, t_2)^{a_i}, \mathcal{D}_j(t_1, t_2)^{a_j} \big\rangle\;$ test $\;\mathcal{N}_i(t_1, t_2) \in \mathcal{J}_{ij}^{a_i,a_j}$
$\circ$ If the test passes, construct the simplified intersection from least common multiples (lcm's)
$$ \begin{equation} \begin{aligned}\label{eq:RS_merge_generalised} \mathcal{I}_{\rm s}'&= \mathcal{I}_{\rm s} \cap_{s} \mathcal{J}_{ij}^{a_ia_j}:= \big\langle \mathrm{lcm}(m_1,n_1), \mathrm{lcm}(m_1,n_2), \ldots, \mathrm{lcm}(m_r,n_1), \mathrm{lcm}(m_r,n_2) \big\rangle \\ & \kern45mm \mbox{with}\quad n_1=\mathcal{D}_i(u,v)^{a_i} \quad\mbox{and}\quad n_2=\mathcal{D}_j(u,v)^{a_j}\,, \end{aligned} \end{equation} $$
$\phantom{\circ}$ Where $\mathcal{I}_{\rm s}$ starts as the unit ideal, and is otherwise generated by $m_1, \dots, m_r$.
$\circ$ Test $\;\mathcal{N}_i(t_1, t_2) \in \mathcal{I}_{\rm s}'$, if it passes set $\; \mathcal{I}_{\rm s} = \mathcal{I}_{\rm s}'$
Bivariate Slice mPFD Interpretation
$\circ$ In the end we are left with an ideal
$$ \begin{equation} \mathcal{I}_{\rm s}=\langle m_1(\mathcal{D}),\ldots,m_r(\mathcal{D})\rangle \quad \mbox{with} \quad \mathcal{N} \in \mathcal{I}_{\rm s} \,. \end{equation} $$
$\phantom{\circ}$ whose genreators $m_k$ are monomials in the denominator factors $\mathcal{D}_i$,
$$ \begin{align} m_k(\mathcal{D}) = \prod_j \mathcal{D}_j^{\hat q_{jk}} \in \mathcal{I}_{\rm s} \quad\mbox{with}\quad k=1,\ldots, r \end{align} $$
$\phantom{\circ}$ with powers not exceeding those of the LCD, i.e. no spurious singularities by construction.
$\circ$ We interpret this back in the full multivariate setting as
$$ \mathcal{N} = \sum_{k=1}^{k_{\rm max}} \mathcal{N}_k \, m_k(\mathcal{D} ) \, $$
$\phantom{\circ}$ and thus the (tentative) multivariate decomposition (Question: how to detect failures before fitting?)
$$ \begin{equation} \tilde r = \frac{\mathcal{N}}{\prod_j \mathcal{D}_j^{q_j}} = \sum_{k=1}^{r} \frac{\mathcal{N}_k }{\prod_j \mathcal{D}_j^{q_{jk}}} \quad\mbox{with} \quad \prod_j \mathcal{D}_j^{q_{jk}} = \frac{\prod_j \mathcal{D}_j^{q_{j}}}{m_k(\mathcal{D})} \, . \end{equation} $$
$\circ$ Verify by fitting the ansatz. If it fails, remove some ideals from the intersection.

The Numerator Ansatz

$\circ\,$ The numerator Ansatz takes the form (massless case)
GDL, Maître ('19)
$\displaystyle \text{Num. poly} = \sum_{\vec \alpha, \vec \beta} c_{(\vec\alpha,\vec\beta)} \prod_{j=1}^n\prod_{i=1}^{j-1} \langle ij\rangle^{\alpha_{ij}} [ij]^{\beta_{ij}}$
$\phantom{\circ}$ subject to constraints on $\vec\alpha,\vec\beta$ due to: 1) mass dimension; 2) little group; 3) linear independence.

$\circ\,$ Construct the Ansatz via the algorithm from Section 2.2 of GDL, Page ('22)
Linear independence = irreducibility by the Gröbner basis of a specific ideal.
$\circ\,$ Efficient implementation using open-source software only

Gröbner bases $\rightarrow$ constrain $\vec\alpha,\vec\beta$
Decker, Greuel, Pfister, Schönemann

Integer programming $\rightarrow$ enumerate sols. $\vec\alpha,\vec\beta$
Perron and Furnon (Google optimization team)





$\circ\,$ Linear systems solved w/ CUDA over $\mathbb{F}_{2^{31}-1}$ ($t_{\text{solving}} \ll t_{\text{sampling}}$) w/ linac (coming soon)
Preview of Linac

work in collaboration with Jack Franklin, avaiable now! @ github.com/GDeLaurentis/linac

cuda_row_reduce(A, field_characteristic=primes[0], verbose=False)  # A is a 2D numpy.ndarray
$\circ\,$ Performance on a laptop GPU of approx. 60 CPU cores
$\circ\,$ Performance on a workstation GPU of approx. 600 CPU cores
$\circ\,$ Tested on systems up to 100k equations and unknowns (takes 45 minutes).