Epstein–Glaser Renormalization on Curved Spacetime

In position-space perturbation theory, renormalization begins with a precise extension problem. Causal factorization determines a time-ordered distribution whenever its arguments do not all coincide; ultraviolet ambiguity enters only when that distribution is extended to the total diagonal. On curved spacetime, finite scaling degree guarantees an extension, while locality, covariance, microlocal spectrum, scaling, unitarity, and field independence restrict its finite ambiguity to local covariant counterterms.

Required background. Flat-Spacetime Quantum Field Theory supplies Wick products and perturbative time ordering; Partial Differential Equations and Green Functions provides distributional kernels and causal support.

Helpful background. Causality and Global Structure establishes the globally hyperbolic causal ordering used in factorization; Quantum Fields in Curved Spacetime supplies Hadamard wave-front sets and local covariance.

Locally, a distribution is a continuous linear functional on compactly supported test functions. It can be multiplied by a smooth function without difficulty, but multiplying singular distributions requires a wave-front-set condition. Hadamard singularities satisfy the condition needed for Wick products away from coincident configurations; Epstein–Glaser renormalization asks how to cross the remaining coincidence set without introducing nonlocal data.

Causal recursion away from the total diagonal

Assume renormalized products $T_k$ are already known for $k<n$. For $n$ local fields, causal factorization determines $T_n(x_1,\ldots,x_n)$ on $M^n\setminus\Delta_n$, where

$$\Delta_n=\{(x,\ldots,x):x\in M\}$$

is the total diagonal. Indeed, near any configuration outside $\Delta_n$, the points can be divided into nonempty sets $I$ and $I^c$ so that no point in $I$ lies causally before a point in $I^c$. In that neighborhood,

$$T_n\!\left(\prod_{j=1}^n \mathcal O_j(x_j)\right) =T_{|I|}\!\left(\prod_{i\in I}\mathcal O_i(x_i)\right) \star T_{|I^c|}\!\left(\prod_{j\in I^c}\mathcal O_j(x_j)\right).$$

Compatibility on overlaps follows from lower-order causal factorization. Thus the inductive step is not an arbitrary regularization of the whole kernel; it is an extension from a punctured configuration space. Brunetti and Fredenhagen make this causal induction explicit and then classify extensions by microlocal scaling degree Brunetti–Fredenhagen 2000, §§4–6.

The microlocal spectrum condition also controls products appearing in the decomposition into Wick monomials and numerical distributions. It replaces global momentum-space spectral support, which is unavailable on a generic curved background. The result is local in spacetime and does not require a Fourier transform or a preferred vacuum.

Scaling degree counts the finite ambiguity

For $t_0\in\mathcal D'(\mathbb R^d\setminus\{0\})$, define its scaling degree at the origin by

$$\operatorname{sd}(t_0)=\inf\left\{\delta:\lim_{\lambda\to0^+}\lambda^\delta t_0(\lambda\,\cdot)=0\right\}.$$

The limit is in the distribution topology: $t_0(\lambda\,\cdot)$ acts on $f$ as $\lambda^{-d}\langle t_0,f(\cdot/\lambda)\rangle$. For a finite scaling degree, if $\operatorname{sd}(t_0)\lt d$, there is a unique extension to $\mathbb R^d$ with the same scaling degree. If $d\le\operatorname{sd}(t_0)\lt\infty$, such extensions still exist, but two choices can differ by

$$\sum_{|\alpha|\le \lfloor\operatorname{sd}(t_0)-d\rfloor} c_\alpha\,\partial^\alpha\delta.$$

The number $\operatorname{sd}(t_0)-d$ is the superficial degree of divergence in this position-space statement. For the total diagonal of an $n$-point kernel in four-dimensional spacetime, the relative-coordinate dimension is $d=4(n-1)$, not four independent of $n$. The coefficients are finite renormalization parameters; the ambiguity is supported exactly where the original distribution was undefined.

As a model, take $t_0(x)=|x|^{-4}$ on $\mathbb R^4\setminus\{0\}$. It has scaling degree four. Choose a compactly supported smooth $\chi$ equal to one near the origin and define

$$\langle t_\chi,f\rangle= \int_{\mathbb R^4}\frac{f(x)-f(0)\chi(x)}{|x|^4}\,d^4x.$$

Near the origin $f(x)-f(0)\chi(x)=O(|x|)$, so the radial absolute integral scales as $\int_0^\epsilon r^3r^{-4}r\,dr$, which converges. Subtracting this zeroth Taylor coefficient is enough for the logarithmic case; a stronger singularity requires a longer Taylor jet. With “new minus old” understood, replacing $\chi$ by $\chi'$ changes the answer by

$$\langle t_{\chi'}-t_\chi,f\rangle =f(0)\int_{\mathbb R^4}\frac{\chi(x)-\chi'(x)}{|x|^4}\,d^4x =c\,\langle\delta,f\rangle,$$

exactly the one-parameter ambiguity predicted by scaling degree. This Euclidean radial kernel is only a transparent model of the extension count; Lorentzian time-ordered kernels carry wave-front and causal prescriptions in addition.

Curvature enters through a local scaling expansion

Near a point $x$, choose a convex normal neighborhood and use the exponential map to describe the other arguments by relative tangent vectors. A locally covariant numerical coefficient $t^0[g](x,\cdot)$ known away from coincidence admits a finite scaling expansion of the form

$$t^0[g](x,\xi)=\sum_{k=0}^{N}C_k[g](x)\,u_k(\xi)+r_N[g](x,\xi).$$

Here the $C_k[g](x)$ are polynomials in curvature, its covariant derivatives, mass parameters, and the metric; the $u_k$ are Lorentz-invariant distributions in tangent-space relative coordinates; and $N$ can be chosen so the remainder has scaling degree below the codimension of the diagonal. The tangent-space pieces can then be extended by the flat distribution theorem, while the remainder extends uniquely by continuity.

Hollands–Wald 2002, §4.1, Theorem 4.1 supplies the detailed scaling-expansion properties, and §§4.2–4.3 use them to extend the distribution and verify the required conditions. The theorem is not a claim that curvature disappears in the ultraviolet. Its leading singular terms resemble Minkowski distributions, while subleading divergences multiply local curvature tensors.

Local covariance restricts counterterms

Scaling degree alone would allow the coefficients $c_\alpha$ to vary with the base point. Local covariance narrows them to natural tensors built locally from the background fields. For a four-dimensional scalar theory, finite redefinitions may therefore involve terms such as

$$\phi^4,\qquad (\nabla\phi)^2,\qquad m^2\phi^2, \qquad R\phi^2,$$

together with appropriate identity-valued curvature polynomials. Engineering dimension, tensor type, symmetries, reality, and field independence further restrict the list. The new term $R\phi^2$ is not an arbitrary curved-space decoration: it is a local covariant possibility of the correct dimension that flat spacetime cannot distinguish.

By contrast, a prescription that adds $c(x)\phi^2(x)$ for a fixed coordinate function such as $c(x)=x^0$ is forbidden. Under a change of chart or a causality-preserving isometric embedding, $x^0$ is not determined from the metric and other background fields, so the prescription does not transform naturally. This gives a direct test of why “a finite function of position” is not an admissible renormalization freedom.

Unitarity constrains adjoints of time-ordered products, and field independence controls differentiation with respect to the basic field. These requirements act together: a term can have acceptable scaling dimension yet fail covariance, reality, or the variational identities.

The extension theorem and its boundary

The hypotheses are that the distributions are known off the total diagonal, have finite scaling degree, obey the microlocal spectrum condition there, and arise in a causal induction satisfying locality, covariance, scaling, unitarity, and field independence. The proof architecture has three steps: causal factorization supplies the off-diagonal distribution; the local scaling expansion separates finitely many tangent-space singular terms from a softer remainder; and extension theorems plus covariance classify the diagonal-supported freedom.

Dropping finite scaling degree is a genuine failure. A punctured distribution with growth modeled by $e^{1/|x|^2}$ has no finite scaling degree, so no finite jet subtraction produces the finite counterterm classification above. Dropping covariance gives the coordinate-dependent counterexample just described. Allowing a timelike boundary without specifying boundary conditions and boundary-local counterterms also lies outside this globally hyperbolic, boundary-free statement.

Finally, existence of time-ordered products is coefficientwise in a formal perturbative expansion. It proves neither convergence of that expansion nor a nonperturbative interacting measure. It also does not choose a state: the algebraic construction and the selection of a physically appropriate interacting state remain distinct questions.

Solved diagnostic: classify a singular extension

Let $t_0(x)=|x|^{-6}$ on $\mathbb R^4\setminus\{0\}$. Determine the most general distribution supported at the origin that can relate two extensions with unchanged scaling degree. Then impose rotational invariance and identify what remains.

Solution

The kernel scales as $t_0(\lambda x)=\lambda^{-6}t_0(x)$, so its scaling degree is six and its degree of divergence is two. Two extensions may therefore differ by derivatives of the delta distribution through order two: a scalar multiple of $\delta$, a vector combination $c^\mu\partial_\mu\delta$, and a symmetric tensor combination $c^{\mu\nu}\partial_\mu\partial_\nu\delta$. Rotational invariance eliminates the vector and forces the symmetric tensor to be proportional to the Euclidean metric. The surviving ambiguity is thus $c_0\delta+c_2\Delta\delta$. This check shows separately what scaling permits and what symmetry removes.

References

• Brunetti, Romeo, and Klaus Fredenhagen. “Microlocal Analysis and Interacting Quantum Field Theories: Renormalization on Physical Backgrounds.” Communications in Mathematical Physics 208 (2000): 623–661. DOI.
• Hollands, Stefan, and Robert M. Wald. “Existence of Local Covariant Time Ordered Products of Quantum Fields in Curved Spacetime.” Communications in Mathematical Physics 231 (2002): 309–345. DOI.