Home·The theorem·The proof in one reading·Architecture·Units·Manuscript·Glossary·FAQ·Work items·Data·Method·Provenance·Changes
← III.5 Cluster expansion of the polymer gas (notes)Part III·Unit outlineIII.7 Matching the Wilson energy to fourth order (notes) →

III.6. The stationary branch: background field of the effective action — detailed notes

Contents

1. What this unit proves 2

2. What it uses 9

3. Strategy 11

4. Steps 11

4.1 The second-order jet (F02, F03) 12

4.2 The gap (F01) 13

4.3 The branch (F04) 13

4.4 Independence of the forest (F05) 14

4.5 The volume-independent branch and its jets 15

4.6 One level of the iteration 17

4.7 Remaining packages (statements only) 18

5. Where the difficulty lies 19

6. Notation 20

7. What remains unexplained 21

The full account of unit III.6: every proposition with its proof sketch, checked against the Lean in two rounds and edited. The condensed outline and the map of the Lean are on the unit page.

Unit page · All units · Overview

Part: III, Ultraviolet construction (renormalization group). Lean: the work items F01–F05 are stated in YM/Bid/Station.lean and F06–F09 in YM/Bid/Craft.lean. Their proofs are in YM/Work/F01–F09. The other packages are: the uniform branch YM/Push/STATIONARY_UNIFORM/; the jet port YM/Planned/STATIONARY_JET_PORT/; and the iteration step YM/Extra/ITERATION_NO_LOSS/. Also included are YM/Planned/PRIMARY_INTEGRAL_295/, YM/Further/PHYSICAL_PERIOD_LIMIT_F933/, YM/Planned/ATOM_REGULAR_ESTIMATE/, YM/Next/ROTATIONS/ and YM/Next/WARD_BIN/. Size: 21,329 lines in the main-theorem closure. Of these:

1. What this unit proves

Setting. The definitions are those of III.1. All lattice sums use counting measure.

Lattices, group, distance.

The step datum \(S\). It consists of:

\(P\) denotes the derivative of the predictor at the identity. It acts on fine \(\mathfrak{g}\)-valued \(1\)-forms; it is the linearized block average plus a gradient correction at the block roots. \(\bar{Q}\) denotes the linearized block average alone.

"Larger". \(S\) is larger (Larger) if three conditions hold:

Forests and the axis chart.

\(\displaystyle U^H (x, a)_e = e^{a_e} \hspace{0.17em} \mathcal{R} (e^x)_e \quad (e \in H^{\text{fr}}), \qquad U^H (x, a)_e = 1 \quad (e \text{a tree bond}) .\)

Clip and path datum.

The tested action.

\(\displaystyle \mathcal{M}_p (W ; U) = \frac{1}{|\mathcal{S}|} \sum_{s \in \mathcal{S}} \left[ \sum_{\pi} \left( c_p \hspace{0.17em} \text{price} ((s^{\ast} U)_{\pi}) + \tfrac{1}{2} | \ell_W (s^{\ast} U)_{\pi} |^2 \right) + \sum_x \sum_i (\text{cub}_i + \text{quart}_i) (s^{\ast} U)_x \right] .\)
\(\displaystyle \mathcal{T}^H_k (x, a) =\mathcal{M}_p (W ; U^H (x, a)) + \kappa \sum_b \text{price} \left( e^{x_b} \hspace{0.17em} Q_k (U^H (x, a))_b^{- 1} \right),\)

where \(b\) runs over the coarse bonds. The second term is the block-spin penalty.

Beds, towers and the quadratic identities.

\(\displaystyle K_B \mathcal{W}= \bar{Q}^{\top} K_C, \qquad \bar{Q} \mathcal{W}= 1 - \kappa^{- 1} K_C .\)

Critical branch. A critical branch for \((H, p, W)\) (Critical) is a map \(\rho\) from coarse fields to \(\mathfrak{g}^{H^{\text{fr}}}\) such that:

Chart ratio and thermal function.

\(\displaystyle \Theta^H_k (x) = \text{loop}_W (U^H (x, \rho (x))) + \tfrac{1}{2} \log \det \text{Hess}_a \mathcal{T}^H_k (x, \rho (x)) - \sum_{e \in H^{\text{fr}}} \log J_{\mathscr{J}} (\rho (x)_e) .\)

The Hessian is taken in the orthonormal colour basis. For a branch with a positive vertical gap, the determinant and chart densities are positive near \(0\).

Proposition III.6.1 (second-order jet of the tested action on the tree slice; F03). Let \(S\) be larger, \(H\) a forest and \(B\) a bed with a floor \((a, A)\) on \(\Lambda_f\) for the clip \(p\). Let \(W\) be a path datum with bounds \((p, \sigma, C, \Lambda)\), attached to \(B\), with \(B\) sworn on \(\Lambda_f\). Assume that the main term of \(W\) is analytic along analytic families through the identity (true by III.6.2). Then for every tag \(k\), \(\mathcal{T}^H_k\) is real-analytic at \(0\), with \(\mathcal{T}^H_k (0) = 0\) and \(D\mathcal{T}^H_k (0) = 0\). Moreover, for all \((x, a), (x', a')\),

\(\displaystyle D^2 \mathcal{T}^H_k (0) [(x, a), (x', a')] = \langle ds, \hspace{0.17em} M_B \hspace{0.17em} ds' \rangle + \kappa \hspace{0.17em} \langle x - Ps, \hspace{0.27em} x' - Ps' \rangle, \qquad s = \iota x + a, s' = \iota x' + a' .\)

Here \(a\) is extended by \(0\) to tree bonds, and the pairings are summed over plaquettes or coarse bonds and over colours. Status: theorem. Only the well-posedness clause of "larger" is used. Lean: Bid.F03, with conclusions Prismatic and Comptroller; proof Work.F03.answer.

Proposition III.6.2 (analyticity of the main term; F02). Let \(W\) have bounds \((p, \sigma, C, \Lambda)\). Let \(X\) be a finite-dimensional real normed space and \(x_0 \in X\). Let \(y \mapsto A (y)\) be a family of fine configurations with \(A (x_0) = 1\) whose link matrices are real-analytic at \(x_0\). Then \(y \mapsto \mathcal{M}_p (W ; A (y))\) is real-analytic at \(x_0\). The proof does not use \(A (x_0) = 1\): the main term is analytic along analytic families at every point. Status: theorem. It supplies the hypothesis Fires of III.6.1. Lean: Bid.F02, Work.F02.answer.

Proposition III.6.3 (gap of the vertical Hessian; F01). Assume the shell comparison E02 (§2). Fix \(G\), \(L\), \(J\), kernel choices \(K\) with penalty \(\kappa\), and \(a > 0\). Then there is \(\eta > 0\) with the following property. Take sides \(c, f\); a larger step datum \(S\) with kernel choices \(K\), block factor \(L\) and order \(J\); a forest \(H\); a clip \(p\); a bed \(B\) with floor \((a, A)\) for some \(A\); and a path datum \(W\) such that every \(\mathcal{T}^H_k\) is analytic at \(0\) with the value, gradient and Hessian specified in III.6.1. Then for every tag \(r\) and every \(v \in \mathfrak{g}^{H^{\text{fr}}}\),

\(\displaystyle \eta \sum_{e \in H^{\text{fr}}} |v_e |^2 \leqslant D_a^2 \mathcal{T}^H_r (0, 0) [v, v] .\)

One may take

\(\displaystyle \eta = \frac{\min (a, \kappa)}{1 + \beta + \gamma}, \qquad \beta = 2 C_{\text{E02}}^2 \hspace{0.17em} L^4 \cdot 4 \cdot 6 \hspace{0.17em} |X|^2, \qquad \gamma = 2 L^4,\)

where:

So \(\eta\) depends only on \((G, L, J, K, a)\). It does not depend on \(c\), \(f\), \(H\), \(B\) beyond its floor, or \(W\). Status: theorem, conditional on E02 in its statement. E02 is proved in III.10 (Work.E02.answer), and the combination is unconditional in III.6.6. Lean: Bid.F01, with conclusion PriceGap; proof Work.F01.answer.

Proposition III.6.4 (existence, tangent and value of the critical branch; F04). Let \(S\) be larger, \(H\) a forest, \(B, C\) beds, \(p\) a clip, \(W\) a path datum, \(a > 0\), and \(T\) towers for \((S, B, C)\) satisfying the diagonal identities. Assume:

Then there is a critical branch \(\rho\) for \((H, p, W)\) such that:

The floor of \(C\) is unused in the proof. Of the diagonal identities, only the clause \(K_B \mathcal{W}= \bar{Q}^{\top} K_C\), \(\bar{Q} \mathcal{W}= 1 - \kappa^{- 1} K_C\) is used. Status: theorem. It gives a germ only: no control of the size of the neighbourhood. Lean: Bid.F04, with conclusions Critical, Guessed and HeatG; proof Work.F04.answer.

Proposition III.6.5 (independence of the gauge-fixing forest; F05). Let \(S\) be larger, \(H, H'\) forests, and \(W\) a path datum with bounds \((p, \sigma, C, \Lambda)\). Let \(\rho, \rho'\) be critical branches for \(H\) and \(H'\). Assume positive gap constants \(a, b\) for \(H, H'\), respectively, and let \(\mathscr{J}\) be a local chart ratio. Then for every tag \(k\), for \(x\) near \(0\):

The gap for \(H\) is unused in the proof. The statement of (ii) allows an additive constant; the proof gives \(0\). Status: theorem. Lean: Bid.F05, Work.F05.answer, Work.F05.same_thermal.

Proposition III.6.6 (the stationary branch from F01–F04). For every \(G\), \(L\), \(J\), kernel choices \(K\) and \(a > 0\) there is \(\eta > 0\) with the following property. Take a larger \(S\) with these parameters; a forest \(H\); beds \(B, C\); clips \(p, q\); a path datum \(W\); and towers \(T\). Assume:

Then the gap of III.6.3 holds with constant \(\eta\), and there is a critical branch with properties (ii)–(iii) of III.6.4. Status: theorem. Bid.cospons derives it from E02 and F01–F04. Integrated.stationary_scene instantiates it with the proofs of E02 and F01–F04, so no work-item assumption remains. The geometric, floor and path hypotheses above are retained. Lean: Bid.Scene.

Proposition III.6.7 (the branch on a volume-independent ball).

(a) Uniform implicit function theorem. Let \(\iota, \tau\) be finite sets and let \(\ell : \iota \sqcup \tau \to \Lambda\) assign their coordinates to sites of a four-dimensional torus, with at most \(K_0\) coordinates per site. Equip the coordinate spaces with the sup norm. Fix \(C_0, a_0, r, m > 0\) and an integer \(n \geqslant 3\). Suppose \(f : \mathbb{R}^{\iota \sqcup \tau} \to \mathbb{R}\) is real-analytic at every point \(z\) of \(\{\|z\|_{\infty} < r\}\) and, for \(1 \leqslant k \leqslant n\), every coordinate tuple \(q_1, \ldots, q_k\) and every \(j \leqslant k\),

\(\displaystyle | \partial_{q_1} \cdots \partial_{q_k} f (z) | \leqslant C_0 \exp \hspace{-0.17em} \left( - a_0 \sum_{i = 1}^k \text{dist} (\ell (q_j), \ell (q_i)) \right) .\)

Here distance is the periodic max-distance. Assume \(\nabla_{\tau} f (0) = 0\) and \(D^2_{\tau \tau} f (0) \geqslant mI\) in counting coordinates. Then there are positive radii \(r_{\text{hor}} \leqslant r_{\text{zone}} \leqslant r_{\text{safe}} < r\), depending only on \((K_0, C_0, a_0, r, m)\), and a map \(g : \mathbb{R}^{\iota} \to \mathbb{R}^{\tau}\) with \(g (0) = 0\), such that whenever \(\|x\|_{\infty} < r_{\text{hor}}\):

(b) Application to the tested action. Fix \(G\), real constants \(\mu, C\), a patch order \(J\), kernel choices \(K\), and an integer \(n \geqslant 3\). There is a choice of auxiliary constants, made before \(L\), with the following property. For every clip \(p\), real \(\Lambda\), larger step \(S\) of order \(J\) with kernel choices \(K\), forest \(H\) and path datum \(W\) with bounds \((p, \mu, C, \Lambda)\), part (a) applies to the tested action in individual real colour components. It gives a critical branch \(g\) common to all tags: for each tag \(t\) and \(\|x\|_{\infty} < r_{\text{hor}}\), \(g (x)\) is the unique critical point of \(\mathcal{T}^H_t (x, \cdot)\) in the closed sup-ball of radius \(r_{\text{zone}}\). The radii depend on the auxiliary constants, \(p\) and \(L\), but not on \(c, f, H, \Lambda\) or \(W\) beyond its bounds. Coarse coordinates are located at their fine block roots when measuring decay. Status: theorem. Lean: (a) Receipt.throughout with hypotheses Receipt and radii reserveX. (b) Before.graph, Before.Service, pre_order.

Proposition III.6.8 (derivatives along the branch). Fix an integer \(N \geqslant 0\), the data and hypotheses of III.6.7(b), auxiliary constants chosen at order \(N + 3\), and any branch \(g\) satisfying its analytic, size and uniqueness conclusions. Then for every tag the following hold, with constants depending only on the corresponding bounds and radii.

For a matrix function \(A (z)\), localization of its coefficients means a bound

\(\displaystyle | \partial_{q_1} \cdots \partial_{q_k} A_{ij} (z) | \leqslant C_k \exp \hspace{-0.17em} \left[ - a_k \left( \text{dist} (i, j) + \sum_{\nu} \text{dist} (i, q_{\nu}) \right) \right],\)

with coordinates replaced by their assigned sites. For positive-order derivatives of a map such as \(g\), omit \(j\); for a scalar function use the bound in III.6.7(a). No volume-independent bound on the value of the classical action or of \(\log \det\) is asserted.

Status: theorem under the hypotheses of III.6.7(b). Lean: Counter.service, Counter, BlockPack.

Proposition III.6.9 (one level of the iteration). Assume the following.

Let \(m < N\) and let \(T\) be a path datum funded at level \(m + 1\) (Work item). That is, the rows of \(T\) enumerate the root-kernel coefficients of \(M_{N, m + 1}\) using the fixed coordinate paths to each offset; \(T\) has the level-\((m + 1)\) bounds; and the sums of absolute cubic and quartic coefficients weighted by exponential path length are each at most the level cap. This row convention is stronger than attachment alone. Then there exist a path datum \(R\) funded at level \(m\), branches \(\rho_{p, F}\) for every period \(p\) and every forest \(F\), and a number \(\alpha\) such that:

\(\displaystyle | D^2 \Theta^F_k (0) [w_1, w_2] - D^2 \left( \alpha \hspace{0.17em} \mathcal{M}^{(2)}_q (R) + \text{loop}_R \right) (0) [w_1, w_2] | \leqslant | \Lambda_c | \hspace{0.17em} \Gamma \hspace{0.17em} \lambda_{m + 1} \hspace{0.17em} e^{- \gamma c} .\)

Here \(\mathcal{M}^{(2)}_q (R ; x)\) denotes the symmetrized Wilson-plus-row-square term evaluated on \(e^x\):

\(\displaystyle \mathcal{M}^{(2)}_q (R ; x) = \frac{1}{|\mathcal{S}|} \sum_{s \in \mathcal{S}} \sum_{\pi} \left[ c_q \hspace{0.17em} \text{price} ((s^{\ast} e^x)_{\pi}) + \tfrac{1}{2} | \ell_R (s^{\ast} e^x)_{\pi} |^2 \right] .\)

It includes all higher-order dependence on \(x\) in these terms; it is not their degree-two Taylor polynomial. The constants \(\Gamma, \gamma > 0\) are chosen after \(L\). All conclusions hold for every tag \(k\), and the same \(R, \alpha\) serve every period and forest.

A companion statement, Iteration.peak, gives a funded path datum at the cutoff level \(N\) whose main term has the jets of order \(\leqslant 4\) of the Wilson energy. Status: theorem under the stated schedule, horizon and fresh-data hypotheses. Their discharge is discussed in §7. Lean: Iteration.extend, Move, ClassicalMatch, FiniteBound.

Further results attached to this unit (statements only; see §4.7).

2. What it uses

\(\displaystyle \text{chord} (U_e \mathcal{R}(V)_e^{- 1}) \leqslant C_{\text{E02}} \hspace{0.17em} \text{port}_i (U) +\boldsymbol{1}_{e \text{crossing}} \hspace{0.17em} \text{wire} (Q_k (U), V)_{(i, \mu)} .\)

Here \(\text{wire} (Q, V)_b = \text{chord} (V_b Q_b^{- 1})\), and \(\text{port}_i (U) \leqslant \sum_{z \in X} \sum_{\pi} \text{chord} (U_{\pi})\), summed over plaquettes \(\pi\) at \(\text{root}_i + z\). The estimate is proved in Work.E02.answer. The combination Bid.cospons and the horizon construction Bid.pinning are filed under III.10.

Imported constants and assumptions.

3. Strategy

At fixed coarse field, the renormalization step expands the fine-field integral around a stationary background. Gauge fixing removes the blockwise gauge directions. A positive vertical Hessian then gives an analytic branch, and the tower identities identify its tangent and the quadratic part of its classical value. The one-loop contribution is the Hessian determinant corrected by the Haar density and the loop atoms.

The construction uses the penalty \(\kappa \sum_b \text{price} (e^{x_b} Q_k (U)_b^{- 1})\) and fixes the fine field on a rooted forest inside each block. The shell comparison bounds the free links by curvature and predictor response. After linearization and summation, it gives a positive vertical Hessian with a lower bound independent of the torus volume. The analytic implicit function theorem produces the branch germ. A change of forest preserves both the stationary value and the thermal function: the squared Jacobian in the Hessian determinant cancels the Haar-density ratio.

For a branch on a ball independent of the volume, exponential decay of the inverse Hessian gives a uniform sup-norm bound. A Newton contraction produces the branch on that ball; differentiating the stationarity and resolvent identities controls its localized derivatives. Finally, the iteration combines inherited atoms with fresh classical and one-loop data supplied by the hypotheses of III.6.9. This is a penalized background-field construction in the spirit of Bałaban's method. It establishes local stationarity and quantitative control near the identity.

4. Steps

Lemmas 4.1–4.4 use the hypotheses of III.6.1, Lemmas 4.5–4.7 those of III.6.3, Lemmas 4.8–4.12 those of III.6.4, and Lemmas 4.13–4.17 those of III.6.5. Lemmas 4.19–4.20 use III.6.7(a)–(b), respectively, and Lemmas 4.22–4.24 use III.6.9.

4.1 The second-order jet (F02, F03)

Lemma 4.1 (holomorphic extension of the main term). Let \(W\) have bounds \((p, \sigma, C, \Lambda)\) and let \(U\) be any configuration on a finite torus. The \(\ell_W\)-series and the sums of cubic and quartic atoms have holomorphic extensions near the matrix pair \((U, U^{\ast})\). Their summands are uniformly summable on the ambient ball of radius \(R_{\sigma} = e^{\sigma / 2} - 1\), using the maximum of the linkwise Frobenius norms. Sketch: Treat the link and inverse-link matrices as independent complex variables. A word of length \(l\) has norm at most \((1 + R_{\sigma})^l \|1\|\). A slot of length \(l\) uses words of length at most \(2 l\), so its norm is at most \(C_G e^{\sigma l}\), with \(C_G\) independent of \(l\). The coefficient bounds therefore give normal convergence; a Cauchy estimate on a smaller ball controls the derivative series. The Wilson term is polynomial and symmetrization is finite. Composition with \(y \mapsto (A (y), A (y)^{\ast})\), followed by restriction to real scalars, proves III.6.2.

Lean: F02.analytic_series, rowExt_series, atomExt_series, mainTerm_contDiffAt.

Lemma 4.2 (the chart and its tangent). The map \((x, a) \mapsto U^H (x, a)\) is analytic near \(0\), with \(U^H (0, 0) = 1\). The derivative of \(t \mapsto U^H (t (x, a))_e\) at \(0\) is \(s_e\), where \(s = \iota x + a\). On tree bonds \(s_e = 0\), because tree bonds are internal and \(\iota x\) vanishes there. Consequently \(\text{incl} (x, a) = (x, \log U^H (x, a))\) is analytic near \(0\), with \(\exp \circ \log U^H = U^H\) near \(0\) and \(D \hspace{0.17em} \text{incl} (0) (x, a) = (x, s)\). Lean: F03.live_rate, inclusion_works, inclusion_deriv.

Lemma 4.3 (flat jets). Let \(\mathcal{F}_k (x, A) =\mathcal{M}_p (W ; e^A) + \kappa \sum_b \text{price} (e^{x_b} Q_k (e^A)_b^{- 1})\) be the action in logarithmic coordinates on all fine bonds. Under the hypotheses of III.6.1:

Sketch: The computation rests on an abstract principle: if \(f\) is \(C^2\) at \(0\) and, for every \(z\), \(f (tz) / t^2 \to \tfrac{1}{2} Q (z, z)\) for a symmetric bilinear \(Q\), then \(f (0) = 0\), \(Df (0) = 0\) and \(D^2 f (0) = Q\). The limits are:

Lean: F03.hessian_of_limit, quadratic_limit, main_limit, noise_limit, flat_jets, metric_pair. (proof read; the metric identity and cubic/quartic limit: statement only)

Lemma 4.4 (transport to the tree slice). Near \(0\), \(\mathcal{T}^H_k =\mathcal{F}_k \circ \text{incl}\). Since \(D\mathcal{F}_k (0) = 0\), \(D^2 \mathcal{T}^H_k (0) [u, u'] = D^2 \mathcal{F}_k (0) [D \text{incl} (0) u, D \text{incl} (0) u']\); no term with \(D^2 \text{incl}\) survives. With Lemmas 4.2–4.3 this gives III.6.1. Lean: F03.from_flat, hessian_comp.

4.2 The gap (F01)

Lemma 4.5 (link bound). Let \(v \in \mathfrak{g}^{H^{\text{fr}}}\) and \(y = s (0, v)\), i.e. \(v\) on free bonds and \(0\) on tree bonds. Assume E02. Then for every fine bond \(e\), in block \(i\) and direction \(\mu\),

\(\displaystyle |y_e | \leqslant C_{\text{E02}} \sum_{z \in X} \sum_{\pi} | dy_{(\text{root}_i + z, \pi)} | + | (Py)_{(i, \mu)} | .\)

Sketch: Apply E02 with \(V = 1\) to \(U_t = e^{ty}\). The configuration lies in the slice, and for small \(t\) the selected frames have positive mass. One obtains \(\text{chord} (U_{t, e}) \leqslant C \hspace{0.17em} \text{port}_i (U_t) + \text{chord} (Q_k (U_t)_{(i, \mu)})\). Divide by \(|t|\) and let \(t \to 0\), using \(\text{chord} (\gamma (t)) / |t| \to | \gamma' (0) |\) along curves through \(1\). Lean: F01.link_bound, chord_limit.

Lemma 4.6 (summation). For the extension \(y\) of Lemma 4.5, \(\sum_e |y_e |^2 \leqslant \beta \sum |dy|^2 + \gamma \sum |Py|^2\), with \(\beta, \gamma\) as in III.6.3. Sketch:

Lean: F01.sum_estimate, sum_ports_sq.

Lemma 4.7 (the gap). By III.6.1 on vertical vectors (where \(x = 0\), so the penalty argument is \(- Py\)) and by the floor,

\(\displaystyle D^2_a \mathcal{T}^H_r (0) [v, v] = \langle dy, M_B dy \rangle + \kappa |Py|^2 \geqslant a |dy|^2 + \kappa |Py|^2 .\)

Hence \(\eta |v|^2 \leqslant \eta (\beta |dy|^2 + \gamma |Py|^2) \leqslant a|dy|^2 + \kappa |Py|^2\). This proves III.6.3. Lean: F01.hess_lower, curl_floor, vertical_hessian.

4.3 The branch (F04)

Lemma 4.8 (tags agree near the identity). If \(S\) is well posed, the predictors \(Q_k\) for different tags coincide on a neighbourhood of the identity configuration. Hence the functions \(\mathcal{T}^H_k\) coincide near \(0\). Near the identity, the frame and connector rules select their unique minimizers, so the tags do not affect the predictor. Local logarithms transfer this equality to arbitrary nearby configurations and then to the axis chart. Lean: F04.prediction_germ, tested_germ.

Lemma 4.9 (existence). The gap makes the vertical Hessian at \(0\) injective, hence invertible in these finite-dimensional spaces. The analytic implicit function theorem, applied to \(D_a \mathcal{T}^H_{k_0} (x, a) = 0\) for one tag \(k_0\), gives an analytic \(\rho\) with \(\rho (0) = 0\). By Lemma 4.8 it is critical for every tag. Lean: F04.invertible, exists_branch.

Lemma 4.10 (test-pair identity). Assume the tower identities and the second jet of III.6.1. Then for every \(z = (y, b)\) and coarse \(x\),

\(\displaystyle D^2 \mathcal{T}^H_k (0) [z, (x, \text{rel}_H x)] = \langle dy, M_C \hspace{0.17em} dx \rangle .\)

Sketch:

When \(M_B\) is positive, the tower identities also imply that \(A =\mathcal{W}x\) minimizes \(\tfrac{1}{2} \langle dA, M_B dA \rangle + \tfrac{\kappa}{2} |x - \bar{Q} A|^2\), with minimum value \(\tfrac{1}{2} \langle dx, M_C dx \rangle\). Lean: F04.test_pair.

Lemma 4.11 (tangent). Differentiating \(D_a \mathcal{T} (x, \rho (x)) = 0\) at \(0\) gives \(D^2 \mathcal{T} (0) [(x, D \rho (0) x), (0, v)] = 0\). Lemma 4.10 with \(z = (0, v)\), together with the symmetry of \(D^2\), gives the same identity with \(\text{rel}_H x\) in place of \(D \rho (0) x\). Subtracting and inverting the vertical Hessian gives \(D \rho (0) = \text{rel}_H\). Lean: F04.tangent.

Lemma 4.12 (value). Let \(\beta (x) = (x, \rho (x))\). Since \(D\mathcal{T} (0) = 0\),

\(\displaystyle D^2 (\mathcal{T} \circ \beta) (0) [x, x'] = D^2 \mathcal{T} (0) [(x, \text{rel}_H x), (x', \text{rel}_H x')] = \langle dx, M_C dx' \rangle,\)

by Lemma 4.10. This is III.6.4(iii). Lean: F04.stationary_value.

4.4 Independence of the forest (F05)

Lemma 4.13 (change of forest is a gauge transformation). For \((x, a)\) near \(0\) define \(\text{sw} (x, a)_{e'} = \log \left( \tilde{U}_{e'} \hspace{0.17em} \mathcal{R}(e^x)_{e'}^{- 1} \right)\) for \(e' {\in H'}^{\text{fr}}\). Here \(\tilde{U}\) is the gauge transform of \(U^H (x, a)\) into the gauge of \(H'\), by a gauge function equal to \(1\) at the block roots. Then near \(0\):

Gauge invariance holds for each term:

Lean: F05.tested_switch, loop_switch, mainTerm_gauge, root_noise, switch.

Lemma 4.14 (the two graphs correspond). Near \(0\), \(\text{sw} (x, \rho (x)) = \rho' (x)\). Indeed, \(\text{sw} (x, \rho (x))\) is a critical point of \(\mathcal{T}^{H'}_k (x, \cdot)\), by the chain rule through the local inverse of \(\text{sw} (x, \cdot)\), and both branches tend to \(0\). Critical branches are unique near \(0\) when the vertical Hessian at \(0\) is non-degenerate: this is the inverse function theorem for \((x, a) \mapsto (x, D_a \mathcal{T}(x, a))\), using the gap of \(H'\). With Lemma 4.13 this gives III.6.5(i). Lean: F05.graph_switch, station_unique.

Lemma 4.15 (Hessian determinants). Write \(\text{sw} (x, \cdot) = \pi_{\epsilon} \circ \text{mv} (x, \cdot)\), where \(\pi_{\epsilon}\) relabels coordinates along a bijection \(\epsilon : H^{\text{fr}} {\to H'}^{\text{fr}}\); the two forests have the same number of free bonds. At a critical point the Hessian pulls back by \(D_a \text{mv}\), so near \(0\)

\(\displaystyle \det \text{Hess}_a \mathcal{T}^H_k (x, \rho (x)) = (\det D_a \text{mv} (x, \rho (x)))^2 \det \text{Hess}_a \mathcal{T}^{H'}_k (x, \rho' (x)),\)

and the right-hand determinant is positive (by the gap). Lean: F05.hessian_ratio, hessian_pos, move.

Lemma 4.16 (Haar Jacobian identity). Near \(0\),

\(\displaystyle \prod_{e \in H^{\text{fr}}} J_{\mathscr{J}} (a_e) = | \det D_a \text{mv} (x, a) | \prod_{e' {\in H'}^{\text{fr}}} J_{\mathscr{J}} (\text{sw} (x, a)_{e'}) .\)

Sketch:

Lean: F05.toll_switch.

Lemma 4.17 (cancellation). If \(B, j, k > 0\), \(A = d^2 B\) and \(j = |d| k\), then \(d \ne 0\), \(A > 0\), and \(\tfrac{1}{2} \log A - \log j = \tfrac{1}{2} \log B - \log k\). With Lemmas 4.13–4.16 and the equality of the loop terms, this gives III.6.5(ii) with constant \(0\). Lean: F05.same_thermal.

4.5 The volume-independent branch and its jets

Lemma 4.18 (Combes–Thomas bound). Let \(S\) be a real symmetric matrix on a finite index set assigned to sites of a four-dimensional torus, with multiplicity at most \(K_0\). Suppose \(a_0, M > 0\), \(C_0 \geqslant 0\), \(|S_{ij} | \leqslant C_0 e^{- a_0 \text{dist} (i, j)}\), and \(S \geqslant MI\). Then

\(\displaystyle | (S^{- 1})_{ij} | \leqslant \frac{2}{M} e^{- r_{\text{CT}} \text{dist} (i, j)}, \qquad r_{\text{CT}} = \min \left( \frac{a_0}{4}, \frac{M}{2 (1 + 4 C_0 K_0 c (a_0 / 2) / a_0)} \right) .\)

Here the site labels are understood in the distances, and

\(\displaystyle c (\alpha) = \frac{16 \cdot 6! \hspace{0.17em} e^{2 \alpha}}{\alpha^6}, \qquad \sum_y e^{- \alpha \text{dist} (x, y)} \leqslant c (\alpha) \quad (\alpha > 0),\)

uniformly over the torus side. Sketch: Conjugate \(S\) by the diagonal weight \(e^{r_{\text{CT}} \text{dist} (\cdot, j)}\). The difference has both absolute row and column sums at most \(M / 2\). Its quadratic form is therefore bounded by \((M / 2) \|v\|_2^2\), so the conjugated matrix retains coercivity \(M / 2\), although it need not be symmetric. Applying this inequality to the tilted \(j\)-th inverse column gives the displayed entry bound.

Lean: tiltedCharge, inverse_sealed.

Lemma 4.19 (uniform implicit function theorem; III.6.7(a)). In the notation of III.6.7(a), set \(D = K_0 c (a_0)\) and

\(\displaystyle \Pi = 1 + \frac{4}{m} K_0 c (r_{\text{CT}}),\)

where \(r_{\text{CT}}\) is the rate of Lemma 4.18 with gap \(M = m / 2\). The radii may be chosen as

\(\displaystyle \begin{aligned} r_{\text{safe}} & = \min \left( \frac{r}{2}, \frac{m}{2 (1 + C_0 D^2)} \right),\\ r_{\text{zone}} & = \min \left( r_{\text{safe}}, \frac{1}{4 \Pi (1 + C_0 D^2)} \right),\\ r_{\text{hor}} & = \min \left( r_{\text{zone}}, \frac{r_{\text{zone}}}{4 \Pi (1 + C_0 D)} \right) . \end{aligned}\)

Sketch: The third-derivative estimate bounds the change of a Hessian entry by \(r_{\text{safe}} C_0 D\) times its spatial decay factor. Summing once more gives absolute row and column sums at most \(r_{\text{safe}} C_0 D^2 \leqslant m / 2\). Thus the vertical Hessian stays \(\geqslant (m / 2) I\), and its inverse has sup-norm operator bound \(\Pi\) throughout the closed safety ball. For \(\|x\|_{\infty} < r_{\text{hor}}\), set \(H_0 = D^2_{\tau \tau} f (0)\) and

\(\displaystyle Y (v) = v - H_0^{- 1} \nabla_{\tau} f (x, v) .\)

On the closed ball of radius \(r_{\text{zone}}\), \(\|DY\|_{\infty \to \infty} \leqslant \Pi r_{\text{zone}} C_0 D^2 \leqslant 1 / 4\), while \(\|Y (0)\|_{\infty} \leqslant \Pi \|x\|_{\infty} C_0 D \leqslant r_{\text{zone}} / 4\). Hence \(Y\) maps this ball into the ball of half the radius and has a unique fixed point. The analytic implicit function theorem at each fixed point, together with uniqueness, makes the resulting branch analytic at every allowed \(x\). Lean: Receipt.adjusted, Receipt.snaps, Receipt.priced_inverse, Receipt.throughout.

Lemma 4.20 (the tested action satisfies III.6.7(a)). Take the auxiliary constants of III.6.7(b). For every clip \(p\), larger \(S\) with order \(J\) and kernel choices \(K\), forest \(H\) and \(W\) with bounds \((p, \mu, C, \Lambda)\), the tested action in sup-norm colour coordinates satisfies:

\(\displaystyle m = \frac{\min (c_p / |\mathcal{S}|, \kappa)}{1 + \beta (L, \ell_L) + 2 L^4} .\)

This is a coercivity estimate for the same vertical Hessian, with no bed or attachment hypothesis. In it:

Lean: Before.receipt, symbolic_gap, zeroGap, Before.full_pack. (proof read for the gap and the assembly; the derivative bounds of the action and the direct shell estimate: statement only)

Lemma 4.21 (derivatives of the inverse and of \(\log \det\); III.6.8). Let \(M (y)\) be a finite matrix function analytic at \(y_0\), with \(M (y_0)\) invertible. Assign its matrix and variable indices to lattice sites, with matrix-index multiplicity at most \(K_0\). Suppose its coefficient derivatives through order \(N\) have the localization bounds of III.6.8 at rate \(a_0 > 0\), and \(M (y_0)^{- 1}\) has exponentially decaying entries at that rate. Then the derivatives of \(M^{- 1}\) through order \(N\) are localized, with rates and constants defined recursively from these bounds and \(K_0\). Sketch: Differentiate

\(\displaystyle D (M^{- 1}) = - M^{- 1} (DM) M^{- 1} .\)

The product rule reduces each new order to lower inverse derivatives and derivatives of \(M\). Exponential convolution bounds control the internal index sums; each multiplication reduces the available rate. For a positive determinant, use \(D \log \det M = \text{tr} (M^{- 1} DM)\) and repeat the same estimates. Under the coercivity hypotheses of III.6.8 this applies to every principal submatrix with every nonnegative shift. Composing with the stationary graph gives the estimates along the branch; the joint action has order \(N + 3\), leaving order \(N + 1\) after taking two Hessian derivatives.

Lean: inverse_lines, establish, along_block, Counter.background. (proof read for the inverse induction and the assembly; the convolution and composition estimates: statement only)

4.6 One level of the iteration

Lemma 4.22 (gap quote and branch selection). Fix a stage index \(j\). Apply III.6.6 with the block factor and floor bound assigned to that stage. This gives post-\(L\) gap constants \(\eta_j\) (Iteration.post). At level \(m\), use \(\eta_{J (m + 1)}\). Every funded \(T\) at level \(m + 1\) then has, for each period and forest, a branch satisfying III.6.9(i) (select).

Lemma 4.23 (the glued datum is funded). The fresh-data hypotheses supply fresh cubic and quartic atoms \(t, v\) and fresh loop atoms \(l\). Let \(R\) be the bundle on the bed \(M_{N, m}\) with:

Then \(R\) is funded at level \(m\). The inherited cubic and quartic atoms are charged with relative factors \(L^{- 2}\) and \(L^{- 4}\), and the resulting charges stay within \(6\) times the level price. Lean: resume. (proof read for the assembly; the inherited-atom estimates: statement only)

Lemma 4.24 (the move). The fresh one-loop data supply \(| \alpha + \kappa_G \log L| \leqslant s_{j, k}\), where \(s_{j, k}\) is the prescribed slope bound for the two stage indices. The schedule gives \(s_{j, k} + 1 \leqslant \text{drift}_{m + 1}\), proving (ii). The matching (iii) is proved in three steps:

The bound (iv) is proved on the reference forest from the fresh determinant hypothesis, and transferred to every forest by the equality of thermal functions (transplant). Lean: extend. (proof read for the assembly; inherited-atom matching: statement only)

4.7 Remaining packages (statements only)

\(\displaystyle F (x) E = EF' (x) \quad \text{near } 0, \qquad E_{i, j} =\boldsymbol{1}_{\pi (i) = j} .\)

Let \(y, z\) be analytic scalar functions with \(y = Mz\) near \(0\). Fix two directions. Suppose the entries of \(F (0)^{- 1}\), the two first variations and the mixed second variation are bounded by \(B_0 e^{- d \hspace{0.17em} \text{dist}}\), \(T_0 e^{- d \hspace{0.17em} \text{dist}}\) and \(Q_0 e^{- d \hspace{0.17em} \text{dist}}\), respectively, where \(d > 0\) and \(B_0, T_0, Q_0 \geqslant 0\). The mixed second derivatives of \(\tfrac{1}{2} \log | \det F| - y\) and \(M (\tfrac{1}{2} \log | \det F' | - z)\) then differ by at most

\(\displaystyle | \beta | \hspace{0.17em} C (K_0, T_0, Q_0, B_0, d) e^{- dN / 8} .\)

For positive determinants these are the thermal determinant terms. The proof expands the second derivative into two traces. Compatibility identifies the small-system trace with sums over covering fibres; the error consists of off-diagonal entries between distinct elements of a fibre, whose separation is at least \(N\).

5. Where the difficulty lies

The main estimate is the volume-independent lower bound on the vertical Hessian. Curvature alone leaves gauge and flat directions; the forest condition and predictor penalty together control every free link. The denominator of the gap contains a term proportional to \(C_{\text{E02}}^2 L^4 |X|^2\), and \(|X| = (2 (J + 7) L + 1)^4\). The gap is chosen after \(L\); no lower bound uniform in \(L\) is asserted.

A spectral gap gives an \(\ell^2\) inverse bound. To obtain a sup-norm radius independent of the volume, one also needs the exponential entry bounds and their summability. The uniform construction uses a direct shell estimate and the positive Wilson coefficient, so it needs no bed or attachment hypothesis. It reuses the linearization and counting estimates of F01, while supplying its own gap and radii.

Forest independence requires both the stationary change-of-coordinates formula for the Hessian and preservation of product Haar measure. Their Jacobians cancel exactly, including the Haar normalization scales. The proof needs only the gap for the target forest. Other unused hypotheses are the coarse floor in F04, the identity base value in F02, and the size clauses of “larger” in F03–F05.

The conclusions concern a branch near the identity and uniqueness in a specified sup-ball. Global constrained minimization is not asserted. The one-loop coefficient and the classical remainders remain inputs to the iteration through the fresh-data hypotheses in III.6.9.

6. Notation

Written proof Lean Meaning
\(S\), \(\kappa\), \(K\), \(J\), \(L\) StepData, KernelChoices step datum; penalty; kernel choices; order; block factor
larger Bid.Larger well posed, \(100 R_S \leqslant f\), \(c > 1\)
\(Q_k\), \(P\), \(\bar{Q}\) prediction, probe, mean predictor; its linearization; linearized block average
\(H\), \(H^{\text{fr}}\) Forest, Free forest gauge; free bonds
\(\mathcal{R}\), \(\iota\) replica, linearCopy coarse field copied onto crossing fine bonds
\(U^H (x, a)\), \(a\) lifting, AxisChart fine field in the axis chart
\(s\), \(x - Ps\) stretch, crop, bills linearized fine field; linearized penalty argument
\(W\), \(p\), \(c_p\) PathData, Clip path datum (row series, loop/cubic/quartic atoms); clip; Wilson coefficient
\(\mathcal{M}_p (W ; \cdot)\), \(\mathcal{M}^{(2)}\) mainTerm, secondTerm symmetrized main term; Wilson-plus-row-square term
\(\mathcal{T}^H_k\) tested, noiseTaylor tested action; penalty term
\(B\), \(M_B\), \(R_B\), \(K_B\) Bed, K bed; metric; roots; \(d^{\top} M_B d\)
\(\langle dA, M_B \hspace{0.17em} dA' \rangle\) gaugeFiscal quadratic form of the bed
floor, attached, sworn Floor, Attached, Sworn ellipticity of the bed; \(\ell_W\) linearizes to \(R_B \hspace{0.17em} d\); symmetry invariance
\(T\), \(\mathcal{W}\), \(\hat{\mathcal{W}}_H\) Towers.Diagonal, Inputs, extracted towers; linear background map; its forest-gauge version
\(\text{rel}_H\) relative, Guessed linear background field in chart coordinates
critical branch \(\rho\) Critical background field
gap \(\eta\) PriceGap lower bound of the vertical Hessian at \(0\)
second jet; value jet Comptroller, HeatG, SameJet III.6.1; III.6.4(iii); equality of \(j\)-th derivatives
\(J_{\mathscr{J}}\) ChartRatio, LocalRatio Haar density in exponential coordinates
\(\Theta^H_k\) thermal, hessian thermal function; Hessian matrix in the colour basis
\(\text{sw}\), \(\text{mv}\) switch, move chart change between forests
localized derivatives; localized matrix Scatter, Sealed, Cable exponential decay of entries and derivatives
hypotheses of III.6.7(a) Receipt analytic, localized, gapped, critical at \(0\)
auxiliary constants chosen before \(L\) Before radii and costs of the uniform layer
block package; jet service BlockPack, Counter III.6.8
fresh-data hypotheses Rates cutoff atoms, fresh classical remainders and one-loop matching
funded datum; move Work item, Move, OnShell III.6.9

"Price", "bill", "work item", "receipt", "toll" and "cost" are numerical bounds and their bookkeeping; they have no economic meaning.

7. What remains unexplained

  1. Fresh classical and one-loop inputs. Proposition III.6.9 is conditional on Rates. The adapter IterationRecipe.stocked supplies the cutoff atoms and retains the classical and quantum matching assumptions from Heating; it does not discharge those assumptions. Their discharge in the main construction remains unverified here.

  2. Small-field domains. The comparison between the componentwise radii in III.6.7–III.6.8 and the small-field thresholds used downstream remains unverified. The relevant branch and derivative interfaces are linked in III.6.7–III.6.8.

  3. Iteration over all levels. The induction beyond one step is in Iteration.sweep and Luna.uniformZero (III.12). Their proofs are outside this verification.

  4. Supporting proofs not checked in full. The remaining unread supporting proofs are the metric/path-limit estimates identified after Lemma 4.3; the direct shell and action-derivative estimates used in Lemma 4.20; the convolution and graph-composition estimates used in Lemma 4.21; and the inherited-atom matching and bounds in Lemmas 4.23–4.24. F06–F09 and the primary-integral, atom-regularity, rotation and Ward packages are represented by their statements only. The period-comparison trace argument was checked; its geometric realization for the actual Hessians was not.

  5. Attached packages. The primary-integral package feeds the repeatable step in III.10; the rotation and Ward material belongs with VI.1. The application of the period comparison in Fresco.subscribed remains unverified here. The size figures in the header are the inherited dependency-closure counts; they have not been recounted.

← III.5 Cluster expansion of the polymer gas (notes)Part III·Unit outlineIII.7 Matching the Wilson energy to fourth order (notes) →