Functional-Method Validation and Error Control

A robust functional result is not defined by a small equation residual. It survives a battery that tests the exact identity, renormalization, closure, symmetry, branch, numerical solution, analytic continuation, and an external observable. These error channels must be reported separately because they arise from different assumptions and are often correlated. Agreement between two calculations that share the same ansatz or matching input is not independent evidence.

Required background. Closure, symmetry, and branch selection supplies the finite system, while coupled propagator, vertex, and bound-state systems supplies shared dependencies and covariance. Helpful background. Correlated evidence and triangulation supplies the independence test.

Minimum validation battery

Before interpreting a functional observable $O$, record:

1. Exact identity: which regulated Schwinger–Dyson, nPI stationarity, Bethe–Salpeter pole, or functional-RG equation is the starting point?
2. Inputs: which regulator, counterterms, renormalized parameters, boundary data, gauge prescription, and positivity class define the problem?
3. Closure: which functions, skeletons, tensor structures, and momentum dependences are retained and omitted?
4. Projection: at which fields and momenta are dressings extracted?
5. Branch: how was the solution continued, and what competing branches were found?
6. Symmetry: what Ward, Slavnov–Taylor, crossing, conservation, or sum-rule residuals were measured?
7. Numerics: what grid, domain, quadrature, interpolation, eigenvalue, and solver refinements were performed?
8. Truncation variation: what nested closure or basis sequence changes the theory-space content?
9. Continuation: what analytic assumptions connect Euclidean data to the claimed pole or real-time observable?
10. External benchmark: what test does not reuse the decisive closure assumption?
11. Covariance and digits: how are shared inputs propagated, and how many digits survive the largest relevant uncertainty?

This record is displayed compactly in the functional-method validation comparison. The preceding functional-equation closure and validation map shows why each stage is a separate inference. The exact-versus-projected distinction for functional RG begins with the unprojected flow in Wetterich 1993, Eqs. (1)–(7).

Six errors that must not be collapsed

Write the reported result as

$$O =O_\star +\delta_{\mathrm{trunc}} +\delta_{\mathrm{branch}} +\delta_{\mathrm{sym}} +\delta_{\mathrm{cont}} +\delta_{\mathrm{param}} +\delta_{\mathrm{num}}.$$

These symbols describe sources, not automatically independent random variables.

• Truncation error is caused by omitted vertices, skeletons, operators, or momentum structures. Nested closures and basis variation can estimate it.
• Branch error records multiple admissible nonlinear solutions or an unresolved parameter-continuation path.
• Symmetry error is tied to measured identity residuals and the observable’s sensitivity to them.
• Continuation error arises when poles, cuts, or real-time response are inferred from Euclidean data.
• Parameter error propagates the joint covariance of renormalized inputs and fitted ansatz parameters.
• Numerical error covers finite grids, domains, quadrature, interpolation, linear algebra, and stopping tolerances.

A one-number total is appropriate only after a joint statistical model or a conservative combination rule has been justified. A useful default is to publish the components and the response matrix.

Worked validation of a derived bound-state mass

Consider a dressed propagator $S_\alpha$, a related kernel

$$K_\alpha=-\frac{\delta\Sigma_\alpha}{\delta S},$$

and a Bethe–Salpeter eigenvalue $\lambda_\alpha(P^2)$. The mass satisfies

$$\lambda_\alpha(-M^2)=1.$$

The kernel–self-energy relation that preserves the axial identity is derived in Munczek 1995, §§ II–IV.

A defensible validation sequence is:

Identity and renormalization

Check the propagator equation residual on and between grid nodes, reproduce the imposed subtraction conditions, and measure the axial or vector Ward-identity residual using the same $S_\alpha$ and $K_\alpha$.

Branch and solver

Continue from a controlled parameter point in both directions, repeat with distinct seeds and solvers, and monitor the smallest Jacobian singular value. An eigenpair residual is checked separately from the propagator–vertex residual.

Truncation

Add transverse vertex tensors or the next skeleton class, reconstruct $K$ coherently, and re-solve. Varying only a fitted parameter inside one ansatz is sensitivity analysis, not a complete truncation sequence.

Continuation

Demonstrate access to $P^2=-M^2$ with the constituent singularities and integration contour controlled. If the result is extrapolated from spacelike eigenvalues, report the fit family, domain, mock-function recovery, and spread as a continuation error. The ill conditioning of Euclidean spectral inversion is analyzed in Jarrell and Gubernatis 1996, §§ II–IV.

External benchmark

Compare a dimensionless mass ratio, decay amplitude, or finite-volume observable with a method whose decisive assumptions differ. If the comparison uses the same perturbative matching coefficient or fitted propagator input, record that covariance.

The evidence triangulation graph makes those shared dependencies explicit. A loop of mutually calibrated functional truncations is useful cross-checking, but it is not three independent confirmations.

From residuals to observable errors

An identity residual $r_{\mathrm{WTI}}(p)$ is not itself an uncertainty on $M$. Measure sensitivity by introducing a controlled deformation along the violating tensor direction:

$$\Gamma^\mu \longmapsto \Gamma^\mu+\epsilon\,\Delta\Gamma^\mu,$$

and compute

$$\frac{\mathrm dM}{\mathrm d\epsilon}.$$

Then a bounded residual can be translated, cautiously, into an observable shift when the deformation spans the relevant violation. Without this response, quoting “WTI violation below one percent” beside a mass does not show a one-percent mass error.

For continuous parameters with covariance $C_\alpha$,

$$C_O =J C_\alpha J^{\mathsf T}, \qquad J_i=\frac{\mathrm dO}{\mathrm d\alpha_i},$$

where each derivative re-solves the coupled equations on the same branch. Truncation and continuation alternatives usually remain separate model discrepancies.

Acceptance decisions

Use bounded statuses rather than a universal score:

• Supported in the declared closure: identities, branch, solver, variation, and at least one external benchmark pass at the precision claimed.
• Qualified: the result is useful, but one named error channel dominates or only a bound is justified.
• Unresolved: multiple branches, continuation models, or closure sequences give materially different answers without a selection criterion.
• Rejected: the candidate violates a required identity, positivity condition for its operator, renormalization input, or held-out exact benchmark.

Internal convergence alone cannot produce the first status. The zero-dimensional fixture on the hierarchy page illustrates this: the Gaussian branch solves its closed equation exactly but misses quadrature and a held-out identity.

Justified digits

Suppose

$$M=1.237846$$

in internal units, while closure variation changes it by $0.018$, continuation choices by $0.006$, and numerical refinement by $0.0002$. The solver supports many digits, but the scientific result is at best

$$M\simeq1.24$$

with the separate error components stated. Printing $1.237846(2)$ would mislabel numerical repeatability as total accuracy.

Common pitfalls

Treating two related truncations as independent evidence. Shared vertices, renormalization inputs, or continuation priors create correlated errors.

Converting every variation into a Gaussian standard deviation. A few closure alternatives do not define a probability distribution without an additional model.

Reporting residual percentages without observable sensitivity. A small identity residual can strongly affect a protected channel, while a larger residual elsewhere may be irrelevant.

Exercises

1. A calculation has $\Delta_{\mathrm{num}}=0.001$, $\Delta_{\mathrm{trunc}}=0.04$, and two admissible branches separated by $0.10$. What should dominate the status?

Solution

The unresolved branch ambiguity dominates. A small solver error and moderate truncation spread do not select between branches separated by $0.10$. Report both branch results or a union interval until an independent physical criterion chooses one.

2. Two benchmark calculations use the same fitted propagator. What must be propagated when their derived masses are compared?

Solution

The shared propagator-input covariance contributes to both masses and their covariance. The uncertainty of the difference is

$$\operatorname{Var}(M_1-M_2) =\operatorname{Var}(M_1) +\operatorname{Var}(M_2) -2\operatorname{Cov}(M_1,M_2).$$

Ignoring the covariance can make agreement look more or less significant than it is.

Continue

Gauge Fixing, BRST Constraints, and the Gribov Problem applies the battery to gauge-fixed correlators. Use the evidence triangulation graph whenever several methods support the same reader-facing claim.

References

• Jarrell, Mark, and J. E. Gubernatis. “Bayesian Inference and the Analytic Continuation of Imaginary-Time Quantum Monte Carlo Data.” Physics Reports 269 (1996): 133–195. DOI.
• Munczek, H. J. “Dynamical Chiral Symmetry Breaking, Goldstone’s Theorem, and the Consistency of the Schwinger–Dyson and Bethe–Salpeter Equations.” Physical Review D 52 (1995): 4736–4740. DOI.
• Wetterich, Christof. “Exact Evolution Equation for the Effective Potential.” Physics Letters B 301 (1993): 90–94. DOI.