Closure, Symmetry Constraints, and Branch Selection

A closure replaces an infinite functional hierarchy by finitely many unknown functions. It is scientifically useful only when retained and omitted tensor structures are explicit, counterterms close in the chosen basis, symmetry identities are tested rather than invoked, and every numerical branch is labeled. Solver convergence measures how well the closed equations were solved; it does not measure the error made by closing them.

Required background. Schwinger–Dyson hierarchies and renormalization inputs separates exact identities from closure, while symmetry and regulator dependence in functional RG supplies the general symmetry-error logic. Helpful background. Slavnov–Taylor and Zinn–Justin identities supplies the non-Abelian constraints.

Write the finite problem before solving it

Suppose the exact hierarchy is written schematically as

$$\mathcal F_n[ G,\Gamma^{(3)},\Gamma^{(4)},\ldots ]=0, \qquad n=2,3,\ldots .$$

A finite closure chooses unknowns

$$\mathcal U =\left\{ G,\Gamma^{(3)}_{a_1\ldots a_r},\ldots \right\}_{\mathrm{retained}}$$

and a map for everything else,

$$\Gamma_{\mathrm{omitted}} \longmapsto \mathcal A_{\mathrm{omitted}}[\mathcal U;\alpha],$$

where $\alpha$ denotes ansatz parameters. A complete closure record states:

• the field, Lorentz, color, flavor, and momentum tensor basis retained at each vertex;
• which skeletons or vertex functions are omitted;
• the momentum configuration used to project each dressing;
• the regulator and counterterms needed in the retained basis;
• the expansion parameter, if one exists, and the expected order of the first omission;
• the symmetry identities imposed exactly and those only monitored as residuals.

“Dressed vertex” is not a specification. Two ansätze can have the same ultraviolet limit and very different transverse momentum dependence.

The chapter-wide functional-equation closure and validation map shows where this record enters. The functional-method validation comparison states the solver, benchmark, covariance, and reporting requirements.

Ward identity as a closure test

For an Abelian fermion propagator, write

$$S^{-1}(p) =A(p^2)p\!\!\!/-B(p^2).$$

The vector Ward–Takahashi identity is

$$q_\mu\Gamma^\mu(k,p) =S^{-1}(k)-S^{-1}(p), \qquad q=k-p.$$

Using a bare vertex $\Gamma^\mu=\gamma^\mu$ together with nonconstant $A$ or $B$ violates the identity. A longitudinal construction that satisfies it and has no kinematic singularity is

$$\begin{aligned} \Gamma^\mu_{\mathrm{BC}}(k,p) ={}& \frac{A(k^2)+A(p^2)}{2}\gamma^\mu \\ &+\frac{(k+p)^\mu}{k^2-p^2} \left[ \frac{A(k^2)-A(p^2)}{2} (k\!\!\!/+p\!\!\!/) -\bigl(B(k^2)-B(p^2)\bigr) \right]. \end{aligned}$$

Contracting with $q_\mu$ gives the difference of inverse propagators. The identity fixes only the longitudinal part; transverse tensors remain unconstrained by it and can materially affect a bound-state or spectral calculation. Ball and Chiu derive this decomposition and its kinematic regularity Ball and Chiu 1980, §§ II–III.

In a non-Abelian gauge theory the Slavnov–Taylor identity also contains ghost dressings and fermion–ghost scattering kernels. Reusing the Abelian longitudinal vertex without those factors is not “symmetry preserving”; the non-Abelian identity and its renormalization consequences are developed in Taylor 1971, pp. 437–443.

Branches are part of the result

Nonlinear integral equations may admit several solutions at the same renormalized parameters. Fixed-point iteration,

$$\mathcal U_{r+1} =\mathcal T[\mathcal U_r],$$

selects a basin of attraction. Underrelaxation or Newton methods can reach a different basin without changing the equations.

A branch record should include:

• the seed and continuation path in mass, coupling, temperature, or gauge parameter;
• turning points detected by a small Jacobian singular value;
• symmetry or positivity tests on each branch;
• an effective-action or free-energy comparison when that functional is consistently available;
• hysteresis checks under forward and reverse parameter continuation.

Continuity from a weak-coupling solution is useful but not universally decisive; a phase transition can change the physical branch. Conversely, the largest condensate or lowest iteration residual is not a branch-selection principle.

Separate six error channels

For an observable $O$, report at least

$$\begin{aligned} O=O_\star &\pm\Delta_{\mathrm{num}} \pm\Delta_{\mathrm{trunc}} \pm\Delta_{\mathrm{sym}} \\ &\pm\Delta_{\mathrm{branch}} \pm\Delta_{\mathrm{cont}} \pm\Delta_{\mathrm{param}}, \end{aligned}$$

with correlations retained when a shared input changes several terms together.

• $\Delta_{\mathrm{num}}$ comes from quadrature, grid, domain, interpolation, and solver tolerances.
• $\Delta_{\mathrm{trunc}}$ comes from omitted functions and tensor structures, estimated through nested closures or ansatz variation.
• $\Delta_{\mathrm{sym}}$ is tied to measured Ward, Slavnov–Taylor, crossing, or conservation residuals.
• $\Delta_{\mathrm{branch}}$ records competing admissible nonlinear solutions and ambiguity in parameter continuation.
• $\Delta_{\mathrm{cont}}$ records analytic continuation, extrapolation, or spectral-reconstruction ambiguity.
• $\Delta_{\mathrm{param}}$ propagates renormalized input covariance.

These quantities need not be independent or Gaussian. Quoting them as one root-sum-square number is justified only after their covariance and statistical interpretation are established.

Scalar closure as a minimal example

For the zero-dimensional quartic fixture, the exact first identity is

$$1=G+\langle x^4\rangle .$$

Writing

$$\langle x^4\rangle=3G^2+\kappa_4$$

shows the omitted structure exactly: Gaussian closure sets the connected cumulant $\kappa_4$ to zero. Solving gives the positive and negative branches. Positivity rejects the negative root, while exact quadrature measures the positive branch’s truncation error. The second exact identity is a held-out residual. This example is small enough that closure, branch, residual, and external benchmark cannot be confused.

Common pitfalls

Imposing one identity and declaring symmetry preservation. A longitudinal Ward identity leaves transverse tensors free. Non-Abelian identities also couple ghost and scattering kernels.

Varying only solver tolerance. A stable digit under tighter iteration says nothing about omitted vertices. Truncation variation must change the finite theory space.

Discarding branches that are hard to reach. Numerical attraction is algorithm-dependent. Physical admissibility requires independent criteria.

Exercises

1. Contract $\Gamma^\mu_{\mathrm{BC}}$ with $q_\mu=k_\mu-p_\mu$ and verify the Ward identity.

Solution

Use $q\cdot(k+p)=k^2-p^2$. The second line loses its denominator. Combining its $A(k^2)-A(p^2)$ term with the first line gives

$$A(k^2)k\!\!\!/-A(p^2)p\!\!\!/.$$

The remaining term is $-[B(k^2)-B(p^2)]$. Their sum is $S^{-1}(k)-S^{-1}(p)$.

2. Two branches have equation residuals $10^{-12}$ and $10^{-9}$, but only the second has a positive spectral density. Which residual selects the physical branch?

Solution

Neither residual alone selects it; both are numerically accurate solutions of the closed equations. If positivity is a valid physical requirement for the declared operator, the first branch is inadmissible despite its smaller residual. The second remains only a candidate until truncation and external benchmark tests are also passed.

Continue

2PI and nPI Effective Actions makes the retained skeleton set variationally explicit. Coupled Propagator, Vertex, and Bound-State Systems propagates closure and branch choices through a full dependency graph.

References

• Ball, James S., and Ting-Wai Chiu. “Analytic Properties of the Vertex Function in Gauge Theories. I.” Physical Review D 22 (1980): 2542–2549. DOI.
• Taylor, John C. “Ward Identities and Charge Renormalization of the Yang–Mills Field.” Nuclear Physics B 33 (1971): 436–444. DOI.