Integration-by-Parts Identities and Master Integrals

Integration-by-parts identities turn the vanishing integral of a total derivative into linear relations among integrals with shifted propagator powers and numerators. For the standard dimensionally regulated families used in perturbation theory, the resulting quotient over rational functions of dimension and kinematics is finite-dimensional and can be represented by master integrals. That finiteness is a nontrivial structural result, not a consequence of writing one total derivative, and is not proved here. The identities determine a basis representation; they do not evaluate the masters.

Required background. Dimensional Regularization as an Amplitude Tool supplies the analytically continued, translation-invariant setting in which total derivatives can be defined consistently.

Helpful background. Tensor Reduction shows how numerator scalar products are expressed in an integral family before or alongside IBP reduction.

The total-derivative identity

For denominators $D_j$ and indices $a_j\in\mathbb Z$, define

$$I(\mathbf a)=\mu^{2L\epsilon} \int\prod_{r=1}^{L}\frac{\mathrm d^d\ell_r}{(2\pi)^d} \frac{1}{D_1^{a_1}\cdots D_n^{a_n}}.$$

The denominator list must be complete enough to span every reducible loop scalar product; any remaining irreducible scalar products are retained as explicit numerators or represented by auxiliary denominator indices. Choose a loop momentum $\ell_r$ and a vector $v^\mu$ built from loop and external momenta. Dimensional regularization gives

$$0=\mu^{2L\epsilon}\int\prod_s\frac{\mathrm d^d\ell_s}{(2\pi)^d}\, \frac{\partial}{\partial\ell_r^\mu} \left( \frac{v^\mu}{D_1^{a_1}\cdots D_n^{a_n}} \right).$$

After differentiating, scalar products in the numerator are rewritten in terms of denominators and irreducible scalar products. The result is a linear relation among nearby lattice points $I(\mathbf a+\boldsymbol\delta)$. Lorentz-invariance and symmetry relations can add further equations.

The absence of a boundary term is not an assertion about an ordinary convergent surface integral in exactly four dimensions. It follows by establishing the identity in a convergence domain, or with auxiliary analytic regulators, and continuing it. The continuation argument and IBP construction are given in Abreu, Britto, and Duhr 2022, §§2.1–2.2, pp. 8–11.

A tadpole recurrence

Let

$$T_a(m^2)=\mu^{2\epsilon}\int\frac{\mathrm d^d\ell}{(2\pi)^d} \frac{1}{(\ell^2-m^2+i0)^a}.$$

Use $v^\mu=\ell^\mu$. Differentiation gives

$$0=d\,T_a-2a\mu^{2\epsilon}\int\frac{\mathrm d^d\ell}{(2\pi)^d} \frac{\ell^2}{(\ell^2-m^2+i0)^{a+1}}.$$

Writing $\ell^2=(\ell^2-m^2+i0)+m^2-i0$ and taking the regulated limit yields

$$(d-2a)T_a-2am^2T_{a+1}=0.$$

The gamma-function formula confirms the recurrence:

$$\frac{T_{a+1}}{T_a}=\frac{d-2a}{2am^2}.$$

This normalization check is sensitive to the sign of the Minkowski denominator. With a Euclidean denominator $\ell_E^2+m^2$, the recurrence is instead

$$(d-2a)T_a^E+2am^2T_{a+1}^E=0,$$

so the ratio has the opposite sign after analytic continuation. Positivity of the uncontinued Euclidean integrand must not be mixed with the dimensionally continued value outside its convergence domain.

An equal-mass bubble reduction

To see a genuine top sector reduce to masters, work in Euclidean kinematics with

$$D_1=k^2+m^2, \qquad D_2=(k-p)^2+m^2, \qquad p^2=Q^2>0,$$

and define $B_{ab}=\mu^{2\epsilon}\int \mathrm d^dk/(2\pi)^dD_1^{-a}D_2^{-b}$. Reflection $k\mapsto p-k$ gives $B_{ab}=B_{ba}$, and the pinched sectors are tadpoles $T_a$. Applying the total derivative with $v^\mu=k^\mu$ at $a=b=1$ gives

$$(Q^2+4m^2)B_{12}=T_2-(d-3)B_{11}.$$

The Euclidean tadpole recurrence is $T_2=-(d-2)T_1/(2m^2)$, so

$$\boxed{ B_{12} =-\frac{(d-3)B_{11}+\dfrac{d-2}{2m^2}T_1} {Q^2+4m^2} }.$$

Thus the doubled bubble reduces to one top-sector master $B_{11}$ and one subsector master $T_1$; no master has been evaluated. The result has dimension $-2$, as $B_{12}$ must. It also passes the independent mass-derivative check

$$B_{12}=-\frac12\frac{\partial B_{11}}{\partial m^2},$$

where the factor $1/2$ appears because the derivative acts on both equal-mass denominators. At $d=3$ and $m=Q=1$, $T_2=1/(8\pi)$ and the reduction gives $B_{12}=1/(40\pi)$, in agreement with direct differentiation of $B_{11}=\arctan(Q/2m)/(4\pi Q)$ Abreu, Britto, and Duhr 2022, §§ 2.3–2.4, pp. 12–15.

Sectors, ordering, and masters

The sign pattern of $\mathbf a$ defines a sector: positive entries are present propagators, zero entries are pinches, and negative entries are numerator factors. A reduction algorithm orders integrals and solves IBP equations so that more complicated elements are expressed through simpler ones. Integrals not eliminated by the chosen complete relation set form a master basis:

$$I(\mathbf a)=\sum_{k=1}^{N_{\mathrm M}} c_k(\mathbf a;d,\{s\},\{m^2\})M_k.$$

The coefficients are rational functions of $d$, masses, and invariants for the usual algebraic setup. A different ordering or basis can change the list of masters without changing the vector space they span. “Master” therefore means irreducible relative to the relation set and coefficient field, not a unique or intrinsically simplest integral.

The reduction problem, sector ordering, and a massive one-loop bubble example are worked through in Abreu, Britto, and Duhr 2022, §§2.3–2.4, pp. 12–15. A broad algorithmic account appears in Weinzierl 2022, §6.1, pp. 157–162.

What a reduction must preserve

An accepted reduction records the denominator definitions, loop routing, index convention, dimension, generic-kinematics assumptions, and basis normalization. It should also be checked by substituting the reduced expressions back into a sample of IBP equations and by testing symmetry-related integrals.

Special kinematics require care. A coefficient can contain a denominator that vanishes at a threshold or Gram-degenerate point even when the original integral has a smooth limit. Such a denominator can also coincide with a genuine singular locus, so smoothness must be checked rather than assumed. Take a regular limit in a basis adapted to that point or derive a local expansion; do not merely substitute into a generic rational reduction.

Exercises

1. Put $a=1$ and $d=4-2\epsilon$ in the tadpole recurrence. It predicts $T_2=(1-\epsilon)T_1/m^2$, which agrees with differentiating the gamma-function form with respect to $m^2$.
2. Why do IBP identities not supply numerical master values? They are homogeneous linear relations. Boundary data or direct evaluation is needed to choose a particular solution.

Where the master basis leads

• Differential Equations for Master Integrals turns kinematic derivatives of the chosen basis into a finite boundary-value system.
• Numerical Evaluation and Validation of Loop Integrals supplies independent values and residual checks for the reduced representation.

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.
• Weinzierl, Stefan. Feynman Integrals. Cham: Springer, 2022. doi:10.1007/978-3-030-99558-4. Open PDF.