Constructive renormalization

Renormalization asks how the description of a system changes when short-distance variables are integrated out. Constructive renormalization adds a demanding question: can that procedure be controlled as an actual integration, with estimates strong enough to remove the cutoff?

Prerequisites: continuum limits and uniform estimates.

After this page: explain a scale decomposition, identify the estimate needed to accumulate many small errors, and distinguish ultraviolet stability from a constructed physical spectrum.

Scope: the calculations below are finite-dimensional illustrations of integration and error control. They are not a construction of a gauge field measure.

Start with an integration you can do

Let $x$ denote a retained variable and $y$ a variable to be eliminated. Consider

$$S(x,y)=\frac12 ax^2+bxy+\frac12 cy^2, \qquad c>0,\quad a>\frac{b^2}{c}.$$

Completing the square gives

$$S(x,y)=\frac c2\left(y+\frac bc x\right)^2+ \frac12\left(a-\frac{b^2}{c}\right)x^2.$$

Consequently,

$$\int_{\mathbb R} e^{-S(x,y)}\,dy =\sqrt{\frac{2\pi}{c}}\, \exp\left[-\frac12\left(a-\frac{b^2}{c}\right)x^2\right].$$

The effective coefficient is $a_{\mathrm{eff}}=a-b^2/c$. Eliminating a variable changes the remaining action, including its normalization. Dropping the normalization loses free-energy information; dropping the induced term changes even the remaining probability distribution.

Field-theory scale decompositions retain this basic idea while replacing one Gaussian integral by many coupled fluctuation integrals. Interactions generate more terms than were present in the starting action. Gauge covariance and the treatment of large fields constrain how the integration is organized. A formal diagram expansion describes candidate coefficients; it does not automatically bound the remainder.

Why uniformity is the central issue

Here is an independent error-budget model. Suppose an approximate integration step differs from the exact one by a relative factor $1+\varepsilon_j$, with $|\varepsilon_j|\leq 1/2$. After $K$ steps, the multiplicative discrepancy is

$$R_K=\prod_{j=1}^K(1+\varepsilon_j).$$

The elementary estimate $|\log(1+u)|\leq2|u|$ for $|u|\leq1/2$ implies

$$|\log R_K|\leq 2\sum_{j=1}^K|\varepsilon_j|.$$

If $|\varepsilon_j|\leq C2^{-j}$, the right-hand side stays bounded as $K$ grows. A fixed bound $|\varepsilon_j|\leq0.01$ gives only a bound proportional to $K$; it does not control arbitrarily many steps. Actual constructive arguments use more refined norms, decompositions and rescalings, but the question is the same: what accumulates, and what remains uniform?

A bounded partition-function quantity is also different from convergence of every required renormalized observable. An existence argument must identify the limiting object, its symmetry and positivity properties, and its relation to the intended gauge theory.

Where the historical result belongs

Bałaban’s Large field renormalization II, journal version, CMP 122 (1989), Theorem 1, page 355, gives a conditional result: every effective coupling $g_k$ must remain in $(0,\gamma]$, for sufficiently small $\gamma$. Then the effective densities obey the representation, conditions and bounds in Convergent renormalization expansions, §2. This package includes the large-field contributions and inherited regularity restrictions, not just a bound on one coefficient.

The underlying setting is four-dimensional pure gauge theory on refined periodic lattices, with compact semisimple matrix group $G\subset U(N)$. The dossier traces that setup to the earlier papers and records the precise scope of the selected-section reading.

Immediately after Theorem 1, the 1989 article says removing its coupling assumption depends on a calculation whose proof had not yet been published. That is a historical qualification, not a claim about its publication status today. Conditional ultraviolet control does not itself verify the required coupling trajectory or construct the physical spectrum.

The official target is more extensive: Jaffe–Witten, sections 4–5 connect the intended local fields and short-distance behavior to a nontrivial theory with a positive physical mass gap.

A source-reading workflow

For a constructive estimate, write down the quantity being bounded before writing “stability.” Then record the regulator, finite or infinite volume, allowed group, coupling trajectory, and variables in which the constants are uniform. Locate how observables enter the estimate. Finally, ask which limit the estimate actually supports.

The limits and uniformity guide explains why taking subsequences, identifying the same limiting theory and controlling physical units are distinct tasks. The open-obstacle index connects them to the atlas without measuring percentage progress.

Exercise: small at every step

Suppose $|\varepsilon_j|\leq1/(10j)$ instead of $C2^{-j}$. Does the displayed bound prove uniform control of $R_K$? What if $|\varepsilon_j|\leq1/(10j^2)$?

Hint: compare the sums with an integral.

Solution: the harmonic sum grows like $\log K$, so the first estimate does not give a uniform bound. This does not prove that the actual error diverges; signs or further structure might help. The second series converges, so the logarithmic bound is uniform. The distinction is between failure of an available estimate and failure of the mathematical object itself.

Continue to renormalized observables or compare a different strategy in rigorous lattice methods.