Continuum scaling and scale setting

Prerequisites: the continuum-limit lesson and glueball spectroscopy.

You will learn to construct a units-consistent scaling table, choose and state a reference scale, and interpret an extrapolation with its uncertainties.

Dimensionless measurements need a reference

A lattice calculation directly gives quantities such as $am$, $a^2\sigma$ and $L/a$. Here $a$ is a spacing, $m$ a mass, $\sigma$ a string tension and $L$ a physical length. To compare two lattice spacings, use a physical reference and follow the same theory.

A ratio

$$\frac{m}{\sqrt{\sigma}}=\frac{am}{a\sqrt{\sigma}}$$

does not require assigning either factor a value in GeV. In pure gauge theory such a dimensional assignment is an extra scale-setting convention; it is not a direct identification with an experimental resonance in QCD with dynamical quarks.

Different scale definitions agree only to the accuracy and in the limits established for their relation. A reference observable also has statistical and systematic uncertainty.

A synthetic scaling table

The following exact teaching model sets a reference length $r=1$ and a physical mass $m=2/r$. It contains no measured Yang–Mills data. Keep $L=4r$ while refining:

$a/r$	$am$	$\xi/a=1/(am)$	$N_s=L/a$	$Lm$
$1/4$	$1/2$	$2$	$16$	$8$
$1/8$	$1/4$	$4$	$32$	$8$
$1/16$	$1/8$	$8$	$64$	$8$

The mass in lattice units decreases even though the physical mass does not. Resolution improves, while the physical box and its finite-volume exposure remain unchanged. To test volume effects at fixed spacing, vary $N_s$ independently.

The dimensionless combination $Lm$ is useful because it counts correlation lengths across the box. Its being large can motivate a finite-volume study, but no universal numerical cutoff for “safe volume” follows without a model and error target.

Define a flow scale before using it

A common reference is a flow time $t_0$ defined implicitly by

$$t^2\langle E(t)\rangle\big|_{t=t_0}=0.3.$$

Here flow time has length-squared units and the energy density has mass dimension four. The defining number and operator discretization are part of the prescription. Lüscher, arXiv:1006.4518v3, §3.4, equation (3.3) uses this definition to examine cutoff effects. Its corrected v3 should not silently be mixed with older tabulated values.

The ratio $m\sqrt{t_0}$ can be obtained from $(am)\sqrt{t_0/a^2}$. A stable ratio is evidence about the observables used. It does not establish every correlation function, reconstruction axiom or physical spectral sector.

What a continuum fit assumes

Suppose a dimensionless ratio $R(a)$ is modeled over a selected range by

$$R(a)=R_0+c(a/r)^2+d(a/r)^4.$$

This is a fit ansatz, not an exact identity for all spacings. Its suitability depends on the action, observable, symmetries, improvement, logarithmic corrections and available range. A credible analysis checks sensitivity to omitting coarse points and changing admissible fit forms, alongside correlated statistical errors.

As an exact synthetic illustration, take $R(a)=3+2(a/r)^2$. The values at $a/r=1/4,1/8,1/16$ are $3.125$, $3.03125$ and $3.0078125$. A linear fit against $(a/r)^2$ recovers three. Real data have noise and unknown subleading effects; reproducing this synthetic intercept does not validate a physical extrapolation.

Athenodorou–Teper, arXiv:2106.00364v3, §§2.1, 5.3 and 5.5 provides a Wilson-action spectroscopy example with continuum fits. Read its fit ranges and caveats alongside the benchmark dossier, rather than treating the word “continuum” as a rigorous existence theorem.

Exercise: propagate the shared scale

A simulation gives $am=0.40\pm0.02$ and $a/r=0.10\pm0.01$. Compute the central value of $mr$. If their errors were independent, estimate its uncertainty. Why might independence fail?

Solution. $mr=(am)/(a/r)=4$. Independent small-error propagation gives

$$\delta(mr)\approx4\sqrt{(0.02/0.40)^2+(0.01/0.10)^2}\approx0.45.$$

Both quantities may come from the same ensembles and fluctuate together, so their covariance can matter. A correlated resampling analysis preserves that information. The example’s uncertainties are invented for arithmetic, not source benchmarks.

Numerical evidence and mathematical control

A fit estimates a limiting value from finite data and a model of corrections. A constructive theorem must prove convergence and required uniform estimates for the relevant observable family. Both are valuable; they answer different questions. Continue to limits and uniformity or vary the units directly in the continuum lab.