Differential Equations for Master Integrals

Master integrals depend on masses and kinematic invariants, so differentiating them produces integrals in the same family. Integration-by-parts reduction closes those derivatives into a finite matrix system. Solving that system transports boundary data through kinematic space; it does not create the boundary constants or choose a branch automatically.

Required background. Integration-by-Parts Identities and Master Integrals supplies the finite basis and the reduction of differentiated integrals back to it.

Helpful background. Holomorphic Functions and Cauchy Theory supplies path continuation, regular singular points, and monodromy concepts used below.

Closing derivatives on the master basis

Let $\mathbf I(x,\epsilon)$ be a vector of masters depending on a dimensionless ratio $x$, with all other independent invariants held fixed. Differentiation acts on denominators and prefactors. After IBP reduction,

$$\frac{\mathrm d}{\mathrm d x}\mathbf I(x,\epsilon) =A(x,\epsilon)\mathbf I(x,\epsilon),$$

where $A$ is rational or algebraic in $x$ and rational in $\epsilon$ in a standard basis. For several variables $x_i$,

$$\mathrm d\mathbf I=\Omega\mathbf I, \qquad \Omega=\sum_i A_i\,\mathrm d x_i.$$

Consistency requires a flat connection away from singular loci:

$$\mathrm d\Omega-\Omega\wedge\Omega=0.$$

Equivalently, $\partial_iA_j-\partial_jA_i-[A_i,A_j]=0$. This provides a stringent check on a multivariable reduction. The derivation, basis changes, boundary-value problem, and one-loop bubble example are presented in Abreu, Britto, and Duhr 2022, §§3.1–3.2, pp. 19–26.

Boundary data determine the solution

In one variable the formal solution is a path-ordered exponential,

$$\mathbf I(x)= \mathcal P\exp\!\left(\int_{x_0}^{x}A(t,\epsilon)\,\mathrm d t\right) \mathbf I(x_0).$$

$\mathbf I(x_0)$ must come from a regular limit, a simpler integral, a direct parameter evaluation, a symmetry condition, or another physical boundary condition. Regularity can relate constants but should not be imposed across a genuine threshold singularity.

Choose $x_0$ in a Euclidean region whenever possible. Continue along a specified path that respects the Feynman boundary value. Paths passing on opposite sides of a singular point can differ by monodromy; writing only the endpoint $x$ is insufficient above a cut.

The equal-mass bubble as a master system

For the Euclidean family on the IBP page, choose the masters

$$B(Q^2,m^2)=\mu^{2\epsilon}\int\frac{\mathrm d^dk}{(2\pi)^d} \frac1{D_1D_2}, \qquad T_2(m^2)=\mu^{2\epsilon}\int\frac{\mathrm d^dk}{(2\pi)^d} \frac1{(k^2+m^2)^2},$$

and set $x=Q^2/m^2$. Homogeneity, common-mass differentiation, and the reduction of $B_{12}$ close the system:

$$\frac{\mathrm d}{\mathrm dx} \begin{pmatrix}B\\T_2\end{pmatrix} = \begin{pmatrix} \dfrac{(d-4)x-4}{2x(x+4)}&\dfrac{2}{x(x+4)}\\[7pt] 0&0 \end{pmatrix} \begin{pmatrix}B\\T_2\end{pmatrix}.$$

At $x=0$ the propagators coincide, so regularity selects $B(0,m^2)=T_2(m^2)$. The normalized solution is

$$\frac{B(x)}{T_2} =\int_0^1\mathrm du\,[1+xu(1-u)]^{d/2-2} ={}_2F_1\!\left(1,2-\frac d2;\frac32;-\frac x4\right).$$

Expanding the parameter representation gives $B/T_2=1+(d-4)x/12+O(x^2)$ and directly verifies the differential equation. The coefficient matrix is singular at $x=0$, but the selected solution is regular; this is an apparent system singularity, not a physical one.

Continuing with $x=-s/m^2-i0$ passes below the singular point $x=-4$ and fixes the threshold sheet. In four dimensions the finite subtracted function

$$F(z)=-\int_0^1\mathrm du\, \log[1-zu(1-u)-i0], \qquad z=\frac{s}{m^2},$$

obeys the scalar equation

$$z(4-z)F'(z)+2F(z)-z=0, \qquad F(0)=0.$$

For $z>4$, the chosen path gives $\operatorname{Im}F=+\pi\sqrt{1-4/z}$. A rational differential matrix alone cannot choose that sign; it comes from the Euclidean anchor and continuation path.

Canonical form and iterated integrals

A basis change $\mathbf J=T(x,\epsilon)\mathbf I$ gives

$$\mathrm d\mathbf J=\left(\mathrm dT\,T^{-1}+T\Omega T^{-1}\right)\mathbf J.$$

For many polylogarithmic families one can seek the canonical form

$$\mathrm d\mathbf J =\epsilon\sum_r C_r\,\mathrm d\log W_r(x)\,\mathbf J,$$

with constant matrices $C_r$. The functions $W_r$ are letters whose zeros and poles mark possible singular loci of the differential system. Iterating in $\epsilon$ produces length-$n$ iterated integrals at order $\epsilon^n$; uniform transcendental weight additionally requires suitably normalized boundary data. Henn introduced this basis criterion and its connection with leading singularities in Henn 2013, preprint pp. 1–2, PDF; a detailed review appears in Abreu, Britto, and Duhr 2022, §3.3, pp. 27–33.

Canonical form is a simplification, not a universal theorem. Elliptic or more general families may require non-polylogarithmic kernels, and algebraic changes of variables can introduce their own branch choices.

A one-letter normalization check

The scalar equation

$$\frac{\mathrm dJ}{\mathrm d x}=\frac{\epsilon}{x}J$$

has solution

$$\operatorname{Log}_{\gamma}\!\frac{x}{x_0} \equiv\int_{\gamma:x_0\to x}\frac{\mathrm d t}{t}, \qquad J(x)=J(x_0)e^{\epsilon\operatorname{Log}_{\gamma}(x/x_0)},$$

and hence

$$J(x)=J(x_0)\left[1+\epsilon\operatorname{Log}_{\gamma}\!\frac{x}{x_0} +\frac{\epsilon^2}{2}\operatorname{Log}_{\gamma}^2\!\frac{x}{x_0}+\cdots\right].$$

Differentiating verifies every coefficient and fixes the factorials. Taking a loop around $x=0$ multiplies the exact solution by $e^{2\pi i\epsilon}$, displaying how the differential equation retains branch information only when the continuation path is specified.

Validation checks

A trustworthy solution should satisfy all of the following:

• substitute into the original matrix equation through the requested order in $\epsilon$;
• reproduce the independently computed boundary value;
• respect the expected mass dimension and symmetry under invariant permutations;
• have singular points compatible with, but not automatically equal to, the physical Landau loci;
• agree with a direct parameter or numerical evaluation at points on each relevant sheet.

Apparent poles of $A$ can be basis artifacts. Conversely, a regular matrix entry does not prove the integrated amplitude is regular if the boundary constants or continuation path are singular.

Exercises

1. With $x_0=1$ and the branch fixed by $\gamma$, verify that $J=e^{\epsilon\operatorname{Log}_\gamma x}$ solves the one-letter equation and that circling $x=0$ once counterclockwise gives $e^{2\pi i\epsilon}J$.
2. Why is the flatness condition automatic in one variable? Every two-form is zero there; in several variables it becomes a nontrivial compatibility test.

Where the differential system leads

• Numerical Evaluation and Validation of Loop Integrals tests boundary constants, continuation paths, and differential-equation residuals at concrete kinematic points.
• Leading Singularities and Integrand Geometry explains why constant maximal residues can suggest, but do not prove, a canonical master basis.

References

• Abreu, Samuel, Ruth Britto, and Claude Duhr. “The SAGEX Review on Scattering Amplitudes, Chapter 3: Mathematical Structures in Feynman Integrals.” Journal of Physics A: Mathematical and Theoretical 55 (2022): 443004. doi:10.1088/1751-8121/ac87de.
• Henn, Johannes M. “Multiloop Integrals in Dimensional Regularization Made Simple.” Physical Review Letters 110 (2013): 251601. doi:10.1103/PhysRevLett.110.251601. Open PDF.