Home·The theorem·The proof in one reading·Architecture·Units·Manuscript·Glossary·FAQ·Work items·Data·Method·Provenance·Changes
← II.2 Marker algebra, polymer combinatorics and Gaussian estimates (notes)Part III·Unit outlineIII.2 Choosing the constants: the order of choice and the RG schedule (notes) →

III.1. The block-spin step and the symmetric gauge fixing — detailed notes

Contents

1. What this unit proves 3

2. What it uses 12

3. Strategy 13

4. Steps 14

4.1 The kernel and its normalization 14

4.2 Covariance 14

4.3 Axial gauges and the branch constants 15

4.4 Synchronization and the \(L^{- 2}\) estimate 16

4.5 Near the trivial configuration 18

4.6 Tree gauge 18

4.7 Formats and energy accounting 19

5. Where the difficulty lies 19

6. Notation 20

7. What remains unexplained 21

The full account of unit III.1: 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: YM/RG/ (Block, Patch, Measurable, Push, Masks, Activities, PathActions, Bounded, History), YM/Branch/, YM/Pressure/, YM/PREC/, YM/PCalc/, YM/Tree/, YM/Exterior/, and the small packages YM/RG17/, YM/Push/CORRECTED_COLUMNS/, YM/Planned/CALIBRATED_MIXED_HISTORY/, YM/UVNext/ANNULUS_COMPARISON/, YM/FullVolume/Train/, YM/SectorTools/ACTUAL_SCALAR_LEDGER/. Size: 4,219 lines (693 declarations) in the main-theorem closure.

Conventions. \(N \geqslant 1\) and \(G \subset U (N)\) is a closed subgroup with probability Haar measure \(dg\). Its Lie algebra \(\mathfrak{g} \subset M_N (\mathbb{C})\) carries the real inner product \(\text{Re tr} (X^{\dagger} Y)\), and \(\Pi_{\mathfrak{g}}\) denotes the orthogonal projection onto \(\mathfrak{g}\). Matrix norms below are Frobenius norms. In path actions, Lie-algebra fields, Hessians and the six works, the physical inner product is \(N^{- 1} \text{Re tr} (X^{\dagger} Y)\); write \(|X| = \|X\| / \sqrt{N}\) for its norm. This is the coordinate normalization of Gauge.inner_frob_color. The curvature field is \(\text{curv} (u) = \Pi_{\mathfrak{g}} (u - u^{- 1}) / 2\), with \(| \text{curv} (u) | \leqslant \text{ch} (u)\). The chord is \(\text{ch} (u) = \|u - 1\| / \sqrt{N}\) and the price is \(\text{pr} (u) = 1 - \text{Re tr} \hspace{0.17em} u / N\). They satisfy \(\text{ch}^2 = 2 \hspace{0.17em} \text{pr}\), both are class functions, \(\text{ch} (u^{- 1}) = \text{ch} (u) \leqslant 2\), \(\text{ch} (uv) \leqslant \text{ch} (u) + \text{ch} (v)\), and \(\|u - v\|^2 = 2 N \hspace{0.17em} \text{pr} (uv^{- 1})\). The Wilson energy is \(\sum_p \text{pr} (U_p)\).

A step goes from the fine torus \(\Lambda_f = (\mathbb{Z}/ f)^4\) to the coarse torus \(\Lambda_c = (\mathbb{Z}/ c)^4\), with \(c \geqslant 1\), \(f = cL\) and \(L \geqslant 3\) odd. The coarse site \(x\) has root \(Lx\) and block \(Lx + B_{(L - 1) / 2}\), where \(B_r = \{- r, \ldots, r\}^4 \subset \mathbb{Z}^4\). Configurations are \(U \in G^{\text{links}}\), and \(\mu_f\) is the product of the probability Haar measures. The gauge action is \((k \cdot U)_{(y, i)} = k_y U_{(y, i)} k_{y + e_i}^{- 1}\). \(T_U (y ; w)\) is the parallel transport along a walk \(w\). \(T_U (y \to z ; o)\) is the transport along the coordinate route that changes the coordinates in the order given by \(o \in S_4\). For a coarse site \(x\), \(U^{(x)}\) is the periodic lift of \(U\) to \(\mathbb{Z}^4\), translated so that \(Lx\) sits at the origin. All patch constructions take place on \(\mathbb{Z}^4\), so a patch may be larger than the torus. For finite \(X \subset \mathbb{Z}^4\), \(E_X (U)\) is the sum of the chords of the links with both ends in \(X\), and \(F_X (U)\) is the sum of the chords of the plaquettes with all four corners in \(X\); \(F_X\) is gauge invariant. \(\mathfrak{B}_4\) is the group of signed permutations of the axes (384 elements). For \(s \in \mathfrak{B}_4\) and a point \(b\), \(s_b U\) is the pullback of \(U\) under \(z \mapsto b + s (z)\); a link reversed by \(s\) carries the inverse variable. Further, \(\kappa_{\text{slope}} = 11 C_A / (24 \pi^2)\), where \(C_A\) is the adjoint Casimir, and \(\text{temp} (x) = x^{- 1 / 2} e^{(\log \log x)^2}\) for \(x > 1\).

Definition (the block kernel). A patch is \(P = (L, J)\) with \(L \geqslant 3\) odd and \(J \geqslant 4\). Its radius is \(R = (J + 3) L\) and its points are those of \(B_R\), with centre \(0\). The kernel constants are \(K = (\kappa, \delta, \varepsilon, a)\), all positive: penalty, frame cutoff, predictor cutoff, axial factor. For a configuration \(W\) on \(\mathbb{Z}^4\) the following objects are defined.

  1. Frame tags and axial frames. A frame tag is \(t = (r, o) \in B_R \times S_4\). The axial frame of \(t\) is \(a^t_y = T_W (r \to y ; o)\).

  2. Minimizing branch. \(W \in M_{\delta}\) if \(\text{ch} \left( a^{(0, o)}_y \hspace{0.17em} T_W (y \to z ; o') \hspace{0.17em} (a^{(0, o)}_z)^{- 1} \right) < \delta\) for all \(o, o' \in S_4\) and \(y, z \in B_R\). In words: in each of the 24 axial gauges from the centre, every route transport in the patch is \(\delta\)-close to \(1\).

  3. Score. For \(h : B_R \to G\), \(S_W (h) = \text{avg}_{y, z \in B_R} \text{avg}_{o \in S_4} \hspace{0.17em} \text{pr} (h_y T_W (y \to z ; o) h_z^{- 1})\).

  4. Eligible tags. \(\mathcal{E} (W) = \{t : E_{B_R} (a^t \cdot W) \leqslant a (2 R + 1) F_{B_R} (W)\}\).

  5. Frame and tag mass. If \(W \in M_{\delta}\), the frame \(h^W\) is a minimizer of \(S_W\) subject to \(h_0 = 1\), and the tag mass is uniform, \(1 / (24| B_R |)\). Otherwise \(h^{W, t}_y = (a^t_0)^{- 1} a^t_y\) (the axial gauge from \(r\), normalized at the centre), with mass \(\boldsymbol{1}_{\mathcal{E} (W)} (t) / |\mathcal{E}(W) |\).

  6. Weight. \(w\) is the law on \(B_R\) of \(\xi_1 + \ldots + \xi_J\), where the \(\xi_k\) are i.i.d. uniform on \(B_{(L - 1) / 2}\). It is the \(J\)-fold convolution of the uniform weight on a centred block; it is \(\mathfrak{B}_4\)-invariant and supported in \(B_{J (L - 1) / 2}\).

  7. Connector. For a coarse link \(b = (x, i)\) and frame tags \((t_x)_{x \in \Lambda_c}\), set \(Y_y = h^{(x)}_y \hspace{0.17em} T_{U^{(x)}} (y ; L \text{ steps along } e_i) \hspace{0.17em} (h^{(x + e_i)}_y)^{- 1}\) for \(y \in B_R\). Here \(h^{(x)}\) is the frame of \(U^{(x)}\) with tag \(t_x\). This is the fine transport from \(Lx + y\) to \(L (x + e_i) + y\), written in the frames of the two patches.

  8. Prediction. Let \(\Phi_Y (u) = \sum_y w_y \hspace{0.17em} \text{pr} (Y_y u^{- 1})\). If \(\min \Phi_Y < \varepsilon\), then \(Q_b\) is a chosen minimizer, equivalently a nearest point of \(G\) to the matrix mean \(\sum_y w_y Y_y\) (an extrinsic, chordal mean); the bond tag then has uniform mass \(1 / |B_R |\). Otherwise \(Q_b = Y_{r_b}\), with the bond tag \(r_b\) drawn with probability \(w_{r_b}\).

  9. Kernel. The tag law is \(\lambda (U, t) = \prod_x m^{\text{fr}} (U^{(x)}, t_x) \prod_b m^{\text{pr}} (Y_b, r_b)\), and

\(\displaystyle \rho_{\beta} (U, V) = \sum_t \lambda (U, t) \prod_b Z_{\beta}^{- 1} e^{- \beta \kappa \hspace{0.17em} \text{pr} (V_b Q_b (U, t)^{- 1})}, \qquad Z_{\beta} = \int_G e^{- \beta \kappa \hspace{0.17em} \text{pr} (g)} \hspace{0.17em} dg.\)

The branch conditions for \((P, K, G)\) are:

Compactness gives the chosen minimizers; (B1)–(B2) make the choices unique on the minimizing branches. Lean: Patch, KernelChoices, usesMin, score, eligible, frameMass, frame, weight, bondCost, predict, predictorMass, Branches, StepData, connector, prediction, likelihood, density.

1. What this unit proves

The block-spin kernel is normalized, Borel and covariant under gauge and lattice symmetries. Its branch cutoffs can be chosen independently of the patch size. The geometric estimates synchronize neighbouring frames and bound the predicted coarse curvature by \(L^{- 2}\) times the fine curvature in a surrounding box. Near the trivial configuration the prediction is analytic, with linearization \(\mathcal{M}+ d_0 \mathcal{F}\). Tree gauge reduces the step integral to the free links. The unit also defines effective densities, RG histories and the requirements used in the later energy and stationary-branch estimates.

Proposition III.1.1 (the block kernel is a Markov kernel). Assume (B1)–(B3). For every \(\beta \in \mathbb{R}\):

Consequently, for every Banach-space-valued \(F \in L^1 (\mu_f)\) the push \((\mathcal{P}_{\beta} F) (V) = \int \rho_{\beta} (U, V) F (U) \hspace{0.17em} d \mu_f (U)\) is defined for every \(V\). Moreover \(\mathcal{P}_{\beta} F \in L^1 (\mu_c)\), \(\mathcal{P}_{\beta}\) is linear, and \(\int \mathcal{P}_{\beta} F \hspace{0.17em} d \mu_c = \int F \hspace{0.17em} d \mu_f\). This is the operation that takes level \(m + 1\) to level \(m\). Status: theorem. Positivity and integrability in \(V\) hold without (B1)–(B3). Lean: IsKernel, isKernel, push, push_integrable, push_integral, push_add, push_smul.

Proposition III.1.2 (exact covariance). Assume (B1)–(B3). For all \(\beta \in \mathbb{R}\), \(U\), \(V\):

Status: theorem. Lean: covary, channel, Calendar.blocks_go, Calendar.BlockControl, Levels.transition_mass.

Proposition III.1.3 (branch constants depending only on \(G\)). Let \(d = d (G)\) be the constant of III.1.5. Let \(r_G > 0\) be such that every \(X \in M_N (\mathbb{C})\) at distance \(< r_G\) from \(G\) has a unique nearest point in \(G\). Put \(\delta_0 = d / \sqrt{N}\) and \(\varepsilon_0 = r_G^2 / (2 N)\). Then for every \(\kappa > 0\), \(\delta \in (0, \delta_0]\), \(\varepsilon \in (0, \varepsilon_0]\) and \(a \geqslant 16\), the constants \((\kappa, \delta, \varepsilon, a)\) satisfy (B1)–(B3) for every patch, that is, for every odd \(L \geqslant 3\) and every \(J \geqslant 4\). Condition (B3) holds in a stronger form: for every \(W\) and every ordering \(o\) there is a root \(r\) with \((r, o)\) eligible. Status: theorem. The existence theorem records the case \((\delta_0, \varepsilon_0, 16)\) for a given penalty \(\kappa\). Lean: Axial.viable, RG.frame_unique, RG.predictor_unique, Axial.eligible.

Proposition III.1.4 (axial gauge controlled by curvature). Let \(l\) be a list of distinct axes, \(r \in \mathbb{N}\), \(o \in B_r\), and \(U\) any configuration on \(\mathbb{Z}^4\). Let \(S = \{y \in B_r : y_j = o_j \text{ for } j \not\in l\}\) be the slab. Then there is a root \(x \in S\) such that the axial gauge \(g_y = T_U (x \to y ; \hspace{0.17em} \text{axes in the order } l)\) satisfies

\(\displaystyle E_S (g \cdot U) \leqslant 4 \hspace{0.17em} |l| \hspace{0.17em} (2 r + 1) \hspace{0.17em} F_S (U) .\)

Status: theorem. Lean: Axial.existence.

Proposition III.1.5 (least-squares synchronization: uniqueness up to a global left multiple). There is \(d = d (G) \in (0, 1 / 4]\) with the following property. Let \(n, \Lambda\) be finite nonempty sets. Let \(W : n \times n \times \Lambda \to G\) satisfy \(W_{xxk} = 1\) and \(W_{yx \sigma (k)} = W_{xyk}^{- 1}\) for some permutation \(\sigma\) of \(\Lambda\), and \(\|W_{xyk} - 1\| \leqslant d\) for all indices. Then any two global minimizers \(h, h'\) of \(\text{cost} (h) = \text{avg}_{x, y, k} \hspace{0.17em} \text{pr} (h_x W_{xyk} h_y^{- 1})\) over \(G^n\) satisfy \(h' = c \hspace{0.17em} h\) for a single \(c \in G\). The constant \(d\) does not depend on \(|n|\) or \(| \Lambda |\). Status: theorem. Lean: Branch.align, with Pairs.

Proposition III.1.6 (synchronization of the realized frame with any gauge). There is \(A = A (G) > 0\) with the following property. Let \(P\) be any patch and \(K\) any kernel constants; no branch condition is needed. Let \(W\) be a configuration on \(\mathbb{Z}^4\), \(t\) a frame tag with \(m^{\text{fr}} (W, t) > 0\), \(h\) the realized frame, and \(H : \mathbb{Z}^4 \to G\) any gauge. Then there is \(c \in G\) with

\(\displaystyle \sum_{y \in B_R} \text{ch} (c^{- 1} h_y H_y^{- 1}) \leqslant A \hspace{0.17em} (2 R + 1) \left( E_{B_R} (H \cdot W) + a \hspace{0.17em} (2 R + 1) \hspace{0.17em} F_{B_R} (W) \right) .\)

Status: theorem. Lean: Pressure.synchronized.

Proposition III.1.7 (one gauge for a neighbourhood of blocks). Fix \(G\), \(K\) and \(j \geqslant 4\), and put \(\ell_j = 2 j + 15\). Then \(D = 16 \hspace{0.17em} \ell_j\) and \(C = A \hspace{0.17em} \ell_j (D + a \hspace{0.17em} \ell_j)\), with \(A\) from III.1.6, have the following property. Let \(S\) be any step with kernel constants \(K\) and \(J = j\) (any \(L\), any \(c\)). Let \(U\) be fine, \(x\) a coarse site, and \(t\) a tag that is usable at \(x\), meaning that \(m^{\text{fr}} (U^{(x + y)}, t_{x + y}) > 0\) for all \(y \in B_2\). Then there are a gauge \(H : \mathbb{Z}^4 \to G\) and pads \(p_y \in G\) (\(y \in B_2\)) with

\(\displaystyle E_{B_{R'}} (H \cdot U^{(x)}) \leqslant D \hspace{0.17em} L \hspace{0.17em} \text{port} (x, U), \qquad \sum_{i \in B_R} \text{ch} \left( p_y^{- 1} h^{(x + y)}_i \hspace{0.17em} H_{i + Ly}^{- 1} \right) \leqslant C \hspace{0.17em} L^2 \hspace{0.17em} \text{port} (x, U) \quad (y \in B_2) .\)

Here \(R' = R + 4 L = (J + 7) L\) and \(\text{port} (x, U) = F_{B_{R'}} (U^{(x)})\) is the fine curvature in the macro-box around \(x\). Every tag with \(\lambda (U, t) > 0\) is usable at every coarse site. Status: theorem. Lean: Pressure.planting, usable, wraps, port, macroR.

Proposition III.1.8 (predicted versus actual coarse curvature). For \(G\), \(K\), \(j\) let \(C_8 = 8 \max (1, \varepsilon^{- 1}) (2 C + D)\), with \(C, D\) from III.1.7. Let \(S\) be any step with constants \(K\) and \(J = j\). Then for all \(U\) and \(V\), every coarse site \(x\), every tag \(t\) usable at \(x\) and every plane \(p\):

\(\displaystyle \text{ch} (Q (U, t)_{(x, p)}) \leqslant \frac{C_8}{L^2} \hspace{0.17em} \text{port} (x, U), \qquad \text{ch} (V_{(x, p)}) \leqslant \text{fuel} (Q, V)_{(x, p)} + \frac{C_8}{L^2} \hspace{0.17em} \text{port} (x, U) .\)

Here \(Q_{(x, p)}\) denotes the plaquette holonomy, and \(\text{fuel} (Q, V)_{(x, p)} = \sum_{e \in \partial (x, p)} \text{ch} (V_e Q_e^{- 1})\) is the sum of the four wires of the plaquette, counted with multiplicity even when edges coincide on a short torus. The macro-box contains \(6 (2 R')^2 (2 R' + 1)^2\) internal plaquettes. Thus \(L^{- 4} \text{port}\) differs from their average chord by a factor bounded above and below in terms of \(J\) only. The bound therefore says that the coarse predicted plaquette is at most \(C (G, K, J) \hspace{0.17em} L^2\) times the average fine curvature: the area scaling of the non-abelian Stokes formula, with no further loss in \(L\). Status: theorem. Lean: Pressure.qdc, fuel, wire.

Proposition III.1.9 (effective densities and RG histories). The following definitions carry four elementary consequences.

(a) Budget points, flags, masks. A budget point is a tuple (integer power \(\geqslant 2\), integer order \(\geqslant 6\), support exponent \(> 0\), cover size \(B \geqslant 1\), threshold \(t \in (0, 1]\)). The flagged sites \(\text{flags}_t (U)\) are the sites touched by a plaquette of chord \(> t\). \(U\) is healthy on \(X\) if \(X \cap \text{flags}_t (U) = \emptyset\), and the masked function is \(\boldsymbol{1}_{\{ \text{healthy on } X\}} F\). A function is regular on \(X\) if it depends only on links in \(X\), is Borel and gauge invariant, and at each configuration healthy on \(X\) is \(C^q\) in exponential left increments, where \(q\) is the order; its ordered left derivatives through order \(q\) must exist. Its sup bounds control derivatives of orders \(0, \ldots, q\) in directions with \(|A_e | \leqslant t (1 + \text{dist} (e^-, o))^{\text{power}}\), where \(e^-\) is the base of the positive link. Theorem: a function that reads the links in \(X\) and the flags in \(X\) depends only on the links in \(\text{inspected} (X)\), the union of \(X\) and the corners of all plaquettes meeting \(X\).

(b) Activities. There are regular activities \(E_i\), with an address (a site \(o_i\) in a support \(X_i\)), and forced activities \(K_i\), with a support \(Y_i\), an inventory \(\emptyset \ne I_i \subset Y_i\) and an extra test. Validity at bounds \(\delta_E, w_K \geqslant 0\) means the following:

\(\displaystyle \sum_{i : o_i = x} e^{\sigma \hspace{0.17em} \text{cov}_B (X_i)} e_i \leqslant \delta_E, \qquad \sum_{i : x \in Y_i} e^{\sigma \hspace{0.17em} \text{cov}_B (Y_i)} k_i \leqslant w_K .\)

Here \(\sigma\) is the support exponent and \(\text{cov}_B (X)\) is the least number of radius-\(B\) balls with connected union covering \(X\); both series are required to converge.

The gas is \(\sum_b \prod_{i \in b} K_i (U)\), summed over finite families \(b\) with pairwise disjoint supports whose inventories cover \(\text{flags}_t (U)\) exactly. Theorem: the gas equals \(1\) when \(\text{flags}_t (U) = \emptyset\).

(c) Leading action. A clip has \(c_{\text{dir}} > 16 \hspace{0.17em} c_{\text{lin}} > 0\) and a relaxation length. Path data consist of row series \(\ell_p (U) = \sum_k r_{p, k} \hspace{0.17em} \text{slot}_k (U)\) (weighted parallel-transported curvatures) and loop, cubic and quartic atoms. Their bounds require exponentially weighted summability, invariance, the normalization \(c_{\text{dir}} \delta_{pq} + \sum_k \hat{s}_k (p) \hat{s}_k (q) = \delta_{pq}\) of the zero-momentum symbols (the sums of row coefficients at each plaquette orientation), and vanishing zeroth and first moments of the loop coefficients. The row term at a plaquette based at \(x\) is kept exactly when \(\text{ch} (U_q) \leqslant t \exp (\text{dist} (x, q^-) / r_{\text{rel}})\) for every plaquette \(q\), where \(r_{\text{rel}} > 0\) is the relaxation length and \(q^-\) its base. The per-plaquette term is \(\text{clip}_t (\text{ch} U_p) +\boldsymbol{1}_{\text{kept}} \tfrac{1}{2} | \ell_p (U) |^2\), where \(\text{clip}_t (s) = c_{\text{dir}} s^2 / 2\) for \(s \leqslant t\) and \(c_{\text{lin}} ts\) for \(s > t\). The atoms are masked by health on their path and plaquette footprints. The leading action \(\mathcal{S}\) is the \(\mathfrak{B}_4\)-average of \(\beta \hspace{0.17em} [\sum_p (\cdot) + \text{cubic} + \text{quartic}] + \text{loop}\). A representation \(\mathcal{R}= (\text{path data, clip, activities}, p, \beta, v)\) evaluates to \(\text{eval}_{\mathcal{R}} = \exp \left( v| \Lambda | -\mathcal{S}- \sum_i \boldsymbol{1}_{\text{healthy}} E_i \right) \cdot \text{gas}\). Theorem: if \(\mathcal{R}\) is lawful (path bounds and valid activities), then \(\text{eval}_{\mathcal{R}}\) is Borel and bounded on that finite torus.

(d) Histories.

The lawful bounds at level \(m\) use the stage path width and row and higher-atom bound, the loop bound \(2 \hspace{0.17em} \text{monitor}_m^{\theta (J_m)}\), and activity bounds \(\delta_E = \hat{\beta}_m^{- 11 / 10}\) and \(w_K = \exp [- \text{monitor}_m^{1 / 5}]\), where \(\text{monitor}_m = \exp [(\log \log \hat{\beta}_m)^2]\).

The remaining fits conditions are: \(\log L_m / \text{scale} (m) \to 0\); \(L_m\) is the factor assigned to order \(J_{m + 1}\), and the assigned factors are nondecreasing; the derivative orders, powers and support exponents are nondecreasing along the levels; \(\theta\) is antitone with \(0 < \theta (J_m) \leqslant J_m^{- 1}\); and

\(\displaystyle \frac{\text{drift}_{m + 1}}{\log L_m} \leqslant J_{m + 1}^{- 2}, \qquad \frac{\log (1 + C_{m + 1} + \sigma_{m + 1}^{- 1} + C^{\text{col}}_{m + 1} + r_{m + 1} + s_{m + 1})}{\log L_m} \leqslant J_{m + 1}^{- 2} .\)

Here \(C, \sigma, C^{\text{col}}, r, s\) are the stage path bound, path width, column bound, relaxation length and support exponent.

Theorem: for a selected history, \(\int \text{eval}_{N, p, m} \hspace{0.17em} d \mu = Z_{N, p} (\beta_{N, p, N})\), the Wilson partition function, for every \(m \leqslant N\).

Status: definitions. The four theorems are proved. The quantitative path and activity bounds needed to construct such histories belong to III.3–III.10. Lean: BudgetPoint, BudgetPoint.Regular, flags, UsesFinal.allReads, Activities, RegularFamily, ForcedFamily, Moves, Valid, admissible, gas, gas_one, Clip, PathData, PathData.Bounds, baseTerm, principal, inChannels, Rep, Rep.eval, Rep.envelope, Calendar, Calendar.Fits, Levels, Levels.budget, History, History.Selected, History.partition_eq.

Proposition III.1.10 (the tube identity). A tube for a step is a tower of transfer kernels together with path data and clips at the fine level (up) and at the coarse level (down). Fix a path-length cutoff, a fine threshold \(t\) and a coarse threshold \(s\). From \(V\) one builds:

Use counting inner products and norms on these finite arrays. Let \(C_f\) and \(C_c\) be the sums of the fine and coarse per-plaquette crowns, and let \(P = \sum_b \tfrac{\kappa}{2} \hspace{0.17em} \text{ch} (V_b Q_b^{- 1})^2\). A crown is the clipped direct term plus half the kept row's squared physical norm, with the row truncated at the path cutoff. The totals of the six local works are

\(\displaystyle \begin{array}{rlll} W_F & = C_f + \tfrac{1}{2} \| \Phi w\|^2 - \langle \Phi w, X_f \rangle, & W_I & = P + \tfrac{1}{2} \| \Phi' w\|^2 - \langle \Phi' w, Z \rangle,\\ W_C & = \langle w, X_n \rangle - \tfrac{1}{2} \|w\|^2 - C_c, & W_L & = \langle w, \Phi^{\dagger} X_f {+ \Phi'}^{\dagger} Z - X_n \rangle,\\ W_N & = \tfrac{1}{2} \|\mathcal{N}w\|^2, & W_D & = \tfrac{1}{2} \langle w, (1 - \Phi^{\dagger} {\Phi - \Phi'}^{\dagger} \Phi' -\mathcal{N}^{\dagger} \mathcal{N}) w \rangle . \end{array}\)

For every path cutoff, thresholds \(t, s \in \mathbb{R}\) and configurations \(U, Q, V\),

\(\displaystyle W_F + W_I + W_C + W_N + W_L + W_D = C_f + P - C_c .\)

The fine and coarse data have two components per plaquette: \(\sqrt{c_{\text{dir}}} \hspace{0.17em} \text{curv} (U_p)\) and the truncated row, using the respective clip and configuration. The fine transfer uses parallel transports of the blockwise copy of \(V\); the other two maps use \(V\) on the coarse torus. The vision caps the direct component by \(\min (1, s / \text{ch} (V_p))\), taking this factor to be \(1\) at zero chord, and masks the row by the kept condition. For the local works, assign each square or inner-product summand to its array coordinate. The two components at each plaquette are summed together: \(W_F\) is indexed by fine plaquettes, \(W_I\) by coarse links, \(W_C, W_L, W_D\) by coarse plaquettes, and \(W_N\) by coarse gain coordinates. The crowns and penalties are charged at the same indices.

Status: theorem (an algebraic identity). Lean: Tube, Tube.locates, workF, workD.

Proposition III.1.11 (energy per seed: the requirement, and its transfer to a history).

The requirement. Write \(\hat{\beta}\) for the coarse reference coupling, \(t = \text{temp} (\hat{\beta} + \kappa_{\text{slope}} \log L)\) for the fine-level threshold and \(\varrho = \lceil (\log \hat{\beta})^2 \rceil\) for the cluster radius.

The transfer (theorem). Let \(B\) be a directed history and suppose the ringing condition holds: for each pair of orders \(j_{\min} \leqslant i \leqslant j \leqslant i + 1\), the inequality holds with some \(\mu > 0\) for every coarse reference coupling above the assigned entry threshold and wire parameter \(F (j, i) > 0\). Here \(j_{\min} = J_0\) is the first admissible order. Then for each such pair there is \(\mu > 0\) with the following property. For all \(N\), all \(m < N\) with \(J_{m + 1} = j\) and \(J_m = i\), all \(p\), and every plot at \((\hat{\beta}_m, F (j, i))\) with \(U\) trivial on the trunks,

\(\displaystyle \mu \hspace{0.17em} t_{m + 1}^2 \hspace{0.17em} |A| \leqslant \text{Work} (A) .\)

This uses \(\hat{\beta}_{m + 1} = \hat{\beta}_m + \kappa_{\text{slope}} \log L_m\), so that \(t = \text{temp} (\hat{\beta}_{m + 1}) = t_{m + 1}\).

Status: the requirement is a definition (a hypothesis of other units). It is proved for every choice of estimator data in III.10 (Next.Fluctuations.grants). Its proof applies Integrated.pressure_grants, which follows from the work items E01–E09 and from the coarse-row bound EBare, itself proved by Fluctuations.bare. The ringing condition is derived from these grants by raising the \(\beta\)-entry thresholds (Bid.pricing_order, pressure_pricing, III.10). The transfer is a theorem of this unit; its constant is uniform in \(N, m, p\) for the fixed pair \((j, i)\). Lean: Stores, Stores.Plot, Stores.assessed, Teller, Teller.Funds, Teller.live, Teller.cash, Teller.Grants, Testing.Stamp, Testing.Ringing, Testing.notation_ticket.

Proposition III.1.12 (kinetic specification of a directed history). Fix levels, a horizon of quadratic kernels, a local exponential-chart density and positive stage constants. Kinetic data assign, for each \(N, m\), a shift \(\alpha_{N, m}\) and path data, and, for each volume and forest \(F\), a map \(\psi\) from coarse Lie-algebra fields to Lie-algebra fields on the free fine links.

To state the stationary requirements, let \(\mathcal{A}_f^{\circ}\) be the unmasked, unclipped fine action with coupling \(1\) and without loop terms, averaged over signed symmetries. Write \(\mathcal{L}_f^{\circ}\) for the corresponding loop action and \(\mathcal{Q}_c^{\circ}\) for the coarse quadratic part. For a coarse field \(u\) and a free-link field \(b\), let \(V = \exp u\) and form \(U_F (u, b)\) by putting \(1\) on trunks and \(\exp (b_e) (\iota V)_e\) on free links, where \(\iota V\) is \(1\) on internal links and copies \(V\) on links crossing blocks. For each tag \(t\), set

\(\displaystyle \mathcal{I}_t (u, b) =\mathcal{A}_f^{\circ} (U_F (u, b)) + \kappa \sum_e \text{pr} (V_e Q_e (U_F (u, b), t)^{- 1}) .\)

The kinetic requirements are:

A directed history consists of paid levels (III.2), a horizon satisfying the decay, positivity and transfer identities of III.3, kinetic data satisfying (i)–(vi), and a recipe with the same path data. It also carries a local Haar-chart formula: on some ball \(B_r \subset \mathfrak{g}\), the exponential is injective with open image, \(J\) is positive and analytic, \(J (0) = 1\), and the pushforward of \(c_J J (X) \hspace{0.17em} dX\) is Haar measure restricted to that image, for some \(c_J > 0\); \(J\) is Borel globally.

A move between a fine representation \(\mathcal{R}_{m + 1}\) and a coarse representation \(\mathcal{R}_m\) requires the exact push identity, the recipe's continuation and retention conditions, and agreement of the coarse path data, clip, norm parameters and lawful bounds with level \(m\). It also requires a forced-activity bound \(\exp [- g_j \hspace{0.17em} \text{monitor}_{m + 1}^2]\), where \(g_j > 0\) is prescribed and \(j = J_{m + 1}\), and

\(\displaystyle | \beta_m - \beta_{m + 1} - \alpha_{N, m} | \leqslant 1.\)

Together with (v), this implies the drift bound of III.1.9(d).

Status: definitions of requirements. For simple nonabelian \(\mathfrak{g}\), the conditional construction theorems Roost.companyK and FundedCompanySelection.company supply directed histories with prescribed paid levels as part of the assembly in III.2. The kinetic requirements themselves do not assert a coupling reset.

Lean: PriceGap, Critical, Guessed, ClassicalMatch, extracted, Horizon.Suffices, Kin, Kin.Planned, FiniteBound, thermal, hessian, LocalRatio, Directed, Directed.Move.

Proposition III.1.13 (tree gauge on blocks). Functions in (c)–(d) take values in a real Banach space. A forest assigns to every non-root fine site \(y\) (one with \(y \ne L \cdot \text{parent} (y)\)) a unit step \(\text{up} (y)\) and a height \(\leqslant 4 L\), such that the height decreases along \(\text{up}\) and the parent block does not change. The trunks are the links used by these steps; the others are free.

Status: theorem. Lean: Tree.Forest, Tree.select, Forest.compress, Forest.names, Forest.pack, Forest.pushing.

Proposition III.1.14 (prediction near the trivial configuration). Assume (B1)–(B3). For a \(\mathfrak{g}\)-valued 1-form \(v\) on \(\Lambda_f\), let \(U_v = (\exp v_e)_e\). Then for every tag \(t\) and coarse link \(b = (x, i)\):

\(\displaystyle \tfrac{d}{ds} h^{(x)}_y (U_{sv}) |_0 = \tau_x (y) - \tau_x (0), \qquad \tfrac{d}{ds} Q_b (U_{sv}, t) |_0 = (\mathcal{M}v + d_0 \mathcal{F}v) (b) .\)

Here \(d_0 \varphi (x, i) = \varphi (x + e_i) - \varphi (x)\) and \(\tau_x (y) = - \text{avg}_{z \in B_R} \text{avg}_o \int_{\text{route} (Lx + y \to Lx + z ; \hspace{0.17em} o)} v\), \((\mathcal{M}v) (x, i) = \sum_y w_y \int_{Lx + y}^{Lx + y + Le_i} v\) (a straight line integral), and \((\mathcal{F}v) (x) = \tau_x (0) - \sum_y w_y \tau_x (y)\).

Status: theorem, with the neighbourhood allowed to depend on the fixed step. Since both tag laws are locally uniform, the kernel itself is analytic in fine and coarse exponential coordinates near \((1, 1)\) for each fixed \(\beta\). Lean: Push.settled, Push.prediction_cont, Push.derivative, Push.prices, Precision.probe, Precision.root_gradient.

2. What it uses

From other units.

Standard mathematics.

Imported constants and assumptions.

3. Strategy

The step follows the block-spin method: fix a frame around each coarse site, transport across coarse links, average the connectors, and add independent Haar-normalized Wilson noise. Two rules define the frames on all configurations. When route holonomies in axial gauges are small, least-squares synchronization gives a unique frame normalized at the centre. Otherwise an axial frame is chosen uniformly among tags satisfying the curvature bound. The predictor likewise uses a unique nearest point to the weighted matrix mean when the minimum price is small, and a connector sampled with weight \(w\) otherwise. Uniqueness and symmetric sampling make these rules covariant. The tag probabilities sum to \(1\), and the noise has total mass \(1\), so the full kernel is normalized.

The branch cutoffs are independent of the number of patch sites. For frame uniqueness, normalize each minimizer by a nearest point to its centroid. Centroid differences then have zero tangential projection, making the feedback comparison a contraction. Predictor uniqueness follows from the tubular neighbourhood of \(G\). Eligible axial frames exist by induction over coordinate slabs: choose a layer of least curvature, fix its gauge, and compare the other layers by plaquette bounds.

For quantitative control, one axial gauge on the macro-box synchronizes the surrounding patch frames. Each frame defect costs \(O (L^2)\) times the curvature sum; the averaging weight contributes \(L^{- 4}\). Fine transports cost another \(L\) times a link sum of order \(L\), giving the same \(L^{- 2}\) coarse-curvature estimate. Near the trivial configuration, the analytic implicit function theorem gives the frame and predictor derivatives. Tree gauge then reduces integration to free links. These geometric and analytic results enter the later energy estimates and stationary-branch analysis.

4. Steps

4.1 The kernel and its normalization

Lemma 4.1 (the tag law is a probability). Under (B3), \(\sum_t \lambda (U, t) = 1\). Sketch: first sum over the bond tags with the frame tags fixed. The connectors are then fixed, and each bond factor sums to \(1\): the mass is uniform on the minimizing branch, and it is \(w\), with \(\sum w = 1\), otherwise. The remaining sum factorizes over coarse sites, and each frame factor sums to \(1\): the mass is uniform on all tags, or on the nonempty eligible set. Lean: StepData.chosen, frameMass_sum, predictorMass_sum.

Lemma 4.2 (Borel measurability). Under (B1)–(B2), for each tag the frames, connectors, predictions and \(\lambda (U, t)\) are Borel functions of \(U\). Sketch: \(M_{\delta}\) and \(\{\min \Phi_Y < \varepsilon\}\) are open. The argmin relations are closed and the targets compact, so a unique minimizer depends continuously on the parameter. Hence the selected frame is continuous on \(M_{\delta}\) and the minimizing predictor is continuous where \(\min \Phi_Y < \varepsilon\). For a fixed tag, fallback frames are continuous in \(U\). Eligibility is a closed condition. Piecewise definitions over Borel sets are Borel. Lean: minArg_cont, selected_cont, best_cont, min_open.

Lemma 4.3 (Proposition III.1.1). Sketch: for each link, right invariance of Haar measure moves \(Q_b\) to \(1\), so \(\int Z_{\beta}^{- 1} e^{- \beta \kappa \hspace{0.17em} \text{pr} (V_b Q_b^{- 1})} dV_b = 1\). Fubini over the coarse links then gives \(\int \rho_{\beta} \hspace{0.17em} dV = \sum_t \lambda = 1\). The noise is continuous on a compact space, hence bounded by some \(A_{\beta}\), and \(\rho_{\beta} \leqslant A_{\beta} \sum_t \lambda = A_{\beta}\). The properties of the push follow from Fubini for the bounded kernel times an integrable \(F\). Lean: noise_mass, density_total, density_bound, push_slice, push_integral.

4.2 Covariance

Lemma 4.4 (gauge covariance of the patch data). Assume (B1)–(B2). For every \(k : \mathbb{Z}^4 \to G\), \(M_{\delta}\), the eligible sets and frame masses are unchanged under \(W \mapsto k \cdot W\), and the frame becomes \(h_y \mapsto k_0 h_y k_y^{- 1}\). For a gauge \(k : \Lambda_f \to G\), the connector of \((x, i)\) becomes \(k_{Lx} \hspace{0.17em} Y_y \hspace{0.17em} k_{L (x + e_i)}^{- 1}\). The prediction transforms in the same way, with unchanged bond mass. Sketch: \(T_{k \cdot W} (y \to z ; o) = k_y T_W (y \to z ; o) k_z^{- 1}\), and ch, pr are class functions. Then \(S_{k \cdot W} (k_0 hk^{- 1}) = S_W (h)\), the normalization \(h_0 = 1\) is preserved, and (B1) identifies the minimizer. The axial costs are conjugated, hence unchanged. \(\Phi_{uYv} (uqv) = \Phi_Y (q)\), and (B2) identifies the minimizer. Lean: frame_gauged, min_gauged, updateGauged, bond_gauge, predicts.

Lemma 4.5 (lattice-symmetry covariance of the patch data). Assume (B1)–(B2). Let \(s \in \mathfrak{B}_4\) act about the centre of a patch. Then:

Sketch: routes and axial gauges are mapped to routes and axial gauges, and the score is invariant; (B1) gives the minimizer. Moreover \(\text{pr} (k^{- 1} u) = \text{pr} (ku^{- 1})\), so \(\Phi_{Y^{- 1}} (u^{- 1}) = \Phi_Y (u)\), and (B2) applies. Lean: frames_channel, axial_channel, min_channel, good_channel, weight_invariant, bond_pull, bond_inv, StepData.positive, StepData.negative. (proof read, including Axial.channels)

Lemma 4.6 (Proposition III.1.2). Sketch: (b) The bijection of tag configurations that relocates frame tags to image sites, maps them by \(s\), and maps bond tags by \(s\) gives \(\lambda (s_{Lb} U, t) = \lambda (U, t')\) and \(Q (s_{Lb} U, t) = s_b Q (U, t')\). The noise is invariant under the simultaneous symmetry, including inversion of both arguments. Reindex the sum over tags. (a) \(\lambda\) is invariant and \(Q \mapsto (k \circ \text{embed}) \cdot Q\); the noise is invariant because \(\text{pr} {(kVk'}^{- 1} {(kQk'}^{- 1})^{- 1}) = \text{pr} (VQ^{- 1})\). (c) is immediate. Lean: mass_channel, prediction_channel, noise_channel, withinChoice, channel, covary.

4.3 Axial gauges and the branch constants

Lemma 4.7 (comparison of parallel layers). Fix \(r, o, l\) as in III.1.4 and an axis \(j \not\in l\). Let \(S'\) be the slab for \(l + [j]\), and suppose all \(j\)-links inside \(S'\) equal \(1\). Let \(S_z\) be the layer of \(S'\) with \(j\)-th coordinate \(z\). Then \(|E_{S_z} - E_{S_u} | \leqslant 4 F_{S'}\) for every \(z, u \in \{- r, \ldots, r\}\). Sketch: with trivial \(j\)-links, the \((i, j)\)-plaquette at \(y\) has holonomy \(U_{(y, i)} U_{(y + e_j, i)}^{- 1}\). The inequality \(| \text{ch} (a) - \text{ch} (b) | \leqslant \text{ch} (ab^{- 1})\) is a discrete Stokes inequality. Summing it over the four axes \(i\) and over \(y\) in a layer bounds \(|E_{S_{z + 1}} - E_{S_z} |\) by \(4 \times\) the plaquette sum at the sites of \(S_z\). Telescoping uses each layer once. Lean: variation, squares, slide_le, window_le.

Lemma 4.8 (Proposition III.1.4). Sketch: induction on \(|l|\); the case \(l = \emptyset\) is trivial. For \(l' = l + [j]\), the slab \(S'\) is the disjoint union of \(2 r + 1\) layers \(S_z\), each a slab for \(l\). Choose the layer \(z_0\) with the least curvature, so that \((2 r + 1) F_{S_{z_0}} \leqslant F_{S'}\). By induction there is a root \(x \in S_{z_0}\) with \(E_{S_{z_0}} \leqslant 4 |l| (2 r + 1) F_{S_{z_0}} \leqslant 4 |l| F_{S'}\) in its axial gauge along \(l\). Extend that gauge by moving along \(j\) last. Then all \(j\)-links in \(S'\) are trivial, and the gauge on \(S_{z_0}\) is unchanged. By Lemma 4.7,

\(\displaystyle E_{S'} = \sum_z E_{S_z} \leqslant (2 r + 1) (4| l| + 4) F_{S'} = 4 |l' | (2 r + 1) F_{S'} .\)

The choice of the cheapest layer is what closes the induction. Lean: existence, total_slab, vertical, faces_pins.

Lemma 4.9 ((B3) for \(a \geqslant 16\)). For every \(W\) and every ordering \(o\) there is a root \(r\) with \(E_{B_R} (a^{(r, o)} \cdot W) \leqslant 16 (2 R + 1) F_{B_R} (W)\). Sketch: apply III.1.4 to the whole box with the four axes in the order \(o\); then \(4 \cdot 4 = 16\). Lean: Axial.eligible, edge_cost.

Lemma 4.10 (structure of the synchronization problem). Let \(W\) be as in III.1.5, with centroid \(m (h) = \text{avg}_x h_x\) and feedback \(F_x (h) = \text{avg}_{y, k} h_y W_{xyk}^{- 1}\), both computed in \(M_N (\mathbb{C})\).

Lean: min_iff, price_diff, min_cov, self_consistent, compare_energy, cluster, closeness, changes, weak_changes.

Lemma 4.11 (Proposition III.1.5). Sketch: let \(\delta_1\) be the radius for which the strict-differentiability bound of \(\pi\) at \(1\) holds with \(\eta = 1 / 4\), and take \(d = \min (\delta_1 / 4, 1 / 4)\), so that \(3 d < \delta_1\). Left-multiply each minimizer by the inverse of a nearest point to its centroid. Then \(1\) is nearest to both centroids, and by 4.10(b) both are still minimizers. Let \(p, q\) be the normalized minimizers, \(u = \max_x \|p_x - q_x \|\) attained at \(x\), and \(B = F_x (p) - F_x (q)\). By 4.10(c) both feedbacks lie within \(3 d < \delta_1\) of \(1\), and self-consistency makes \(p_x, q_x\) nearest points to these feedbacks. Hence \(\|p_x - q_x - \Pi_{\mathfrak{g}} B\| \leqslant \frac{1}{4} \|B\| \leqslant \frac{u}{4}\). Since \(m (p) - 1\) and \(m (q) - 1\) are normal to \(\mathfrak{g}\), \(\Pi_{\mathfrak{g}} (m (p) - m (q)) = 0\), and 4.10(d) gives \(\| \Pi_{\mathfrak{g}} B\| \leqslant ud\). So \(u \leqslant ud + u / 4 \leqslant u / 2\), hence \(u = 0\) and \(p = q\). Undoing the normalization gives \(h' = ch\). Lean: align.

Lemma 4.12 ((B1) and (B2) on patches). (B1) holds if \(\delta \leqslant d / \sqrt{N}\), and (B2) holds if \(\varepsilon \leqslant r_G^2 / (2 N)\), for every patch. Sketch: (B1). Take \(n = B_R\), \(\Lambda = S_4\) and \(W_{yzo} = a_y T (y \to z ; o) a_z^{- 1}\), where \(a\) is the axial frame from the centre with the identity ordering. The involution \(\sigma\) reverses the ordering, since the reversed route is the route with the reversed ordering. On \(M_{\delta}\), \(\|W - 1\| = \sqrt{N} \hspace{0.17em} \text{ch} (W) < d\). Since \(S_W (h) = \text{cost} (ha^{- 1})\) and the cost is left invariant, a constrained minimizer is a global minimizer. By III.1.5 two of them differ by a constant, which equals \(1\) at the centre. (B2). If \(\Phi_Y (u) < r_G^2 / (2 N)\), Jensen's inequality gives \(\| \sum w_y Y_y - u\|^2 \leqslant \sum w_y \|Y_y - u\|^2 = 2 N \Phi_Y (u) < r_G^2\). So the mean has a unique nearest point, which by 4.10(a) is the unique minimizer. Lean: frame_unique, predictor_unique, weighted_unique.

Lemma 4.13 (Proposition III.1.3). Combine 4.9 and 4.12. Lean: viable.

4.4 Synchronization and the \(L^{- 2}\) estimate

Lemma 4.14 (clustering of minimizers for arbitrary data). There is \(C_G > 0\) such that for all finite nonempty \(n, \Lambda\) and all \(W\) satisfying the diagonal and reflection conditions of III.1.5, every global minimizer \(h\) admits \(y \in G\) with \(\text{avg}_x \text{ch} (h_x y^{- 1}) \leqslant C_G \hspace{0.17em} \text{avg}_{x, z, k} \text{ch} (W_{xzk})\). Sketch: write \(\text{inc}_x = \text{avg}_{z, k} \text{ch} (W_{xzk})\) and \(\bar{\mu} = \text{avg}_x \text{inc}_x\). Since \(\text{ch}^2 \leqslant 2 \hspace{0.17em} \text{ch}\), \(\text{cost} (h) \leqslant \text{cost} (1) \leqslant \bar{\mu}\), so \(\text{avg} \|h_x - h_z \|^2 \leqslant 8 N \bar{\mu}\). Let \(\delta_1\) be the radius for which the strict-differentiability bound of \(\pi\) at \(1\) holds with \(\eta = 1\).

Lean: Pressure.cluster, price_square, income_bounds.

Lemma 4.15 (route counting). For every ordering \(o\): \(\sum_{y, z \in B_r} \text{ch} \hspace{0.17em} T_U (y \to z ; o) \leqslant 4 \hspace{0.17em} |B_r | \hspace{0.17em} (2 r + 1) \hspace{0.17em} E_{B_r} (U)\). Sketch: induction over the axes of the route. The last segment, along the axis \(j\), starts at the point \(\zeta (y, z)\) that has the already-processed coordinates of \(z\) and the other coordinates of \(y\). Its chord is at most the sum of the link chords on the whole \(j\)-line through \(\zeta (y, z)\). The map \((y, z) \mapsto \zeta (y, z)\) is \(|B_r |\)-to-one onto \(B_r\), and each \(j\)-link lies on one line containing \(2 r + 1\) points. Lean: route_sum, marginal, slidings, Road.revenue.

Lemma 4.16 (Proposition III.1.6). Sketch: Minimizing branch. Since \(T_{H \cdot W} = H \hspace{0.17em} T_W H^{- 1}\), the family \(y \mapsto h_y H_y^{- 1}\) is a global minimizer of the synchronization cost for the data \(T_{H \cdot W} (y \to z ; o)\). By 4.14 and 4.15 there is \(c\) with \(\sum_y \text{ch} (h_y H_y^{- 1} c^{- 1}) \leqslant 4 C_G (2 R + 1) E_{B_R} (H \cdot W)\). The cutoff \(\delta\) plays no role here. Fallback branch. Here \(h\) is the normalized axial frame of an eligible tag, so \(E_{B_R} (h \cdot W) \leqslant a (2 R + 1) F_{B_R} (W)\). Put \(g = hH^{- 1}\); the pure-gauge configuration \(g \cdot 1\) satisfies \(E (g \cdot 1) \leqslant E (h \cdot W) + E (H \cdot W)\) by the triangle inequality. Apply 4.15 to \(g \cdot 1\), whose route transports are \(g_y g_z^{- 1}\), and choose \(v\) with \(\sum_y \text{ch} (g_y g_v^{- 1})\) at most its average over \(v\). Then \(c = g_v\). In both cases \(A = \max (4 C_G, 4)\), and \(\text{ch} (c^{- 1} x) = \text{ch} (xc^{- 1})\). Lean: synchronized, Road.strips, risks, fallback.

Lemma 4.17 (elementary bounds on the step geometry).

Lean: spread, light, traverse, price_query.

Lemma 4.18 (Proposition III.1.7). Sketch: apply III.1.4 once, on \(B_{R'}\) around \(Lx\) with all four axes. This gives one gauge \(H\) with \(E_{B_{R'}} (H \cdot U^{(x)}) \leqslant 16 (2 R' + 1) \hspace{0.17em} \text{port} \leqslant 16 \hspace{0.17em} \ell_j L \hspace{0.17em} \text{port}\). For \(y \in B_2\), the patch of \(x + y\) is the translate of \(B_R\) by \(Ly\) and lies inside \(B_{R'}\). Apply III.1.6 to it, with the translated gauge \(H (\cdot + Ly)\); this gives \(p_y\). Bound its edge and face sums by those of \(B_{R'}\) using 4.17(a), and use \(2 R + 1 \leqslant \ell_j L\). Every tag with \(\lambda > 0\) is usable, because a vanishing frame mass would make \(\lambda\) vanish. Lean: planting, wraps.

Lemma 4.19 (link comparison). In the setting of III.1.7, for \(y, y + e_i \in B_2\) with pads \(\xi = p_y\) and \(\eta = p_{y + e_i}\):

\(\displaystyle L^2 \hspace{0.17em} \text{ch} (\xi^{- 1} Q_{(x + y, i)} \eta) \leqslant 2 \max (1, \varepsilon^{- 1}) (2 C + D) \hspace{0.17em} \text{port} (x, U) .\)

Sketch: write \(\xi^{- 1} Y_k \eta = f_k \cdot T_{H \cdot U^{(x)}} (k + Ly \to k + Ly + Le_i) \cdot g_k^{- 1}\), where \(f_k, g_k\) are the frame defects of III.1.7. Then \(\text{ch} (\xi^{- 1} Y_k \eta) \leqslant \text{ch} f_k + \text{ch} T + \text{ch} g_k\). Multiply by \(w_k L^4 \leqslant 1\) and sum: the defects contribute \(\leqslant 2 CL^2 \hspace{0.17em} \text{port}\), and by 4.17(c) the transports contribute \(\leqslant L \hspace{0.17em} E_{B_{R'}} \leqslant DL^2 \hspace{0.17em} \text{port}\). Conclude with 4.17(d). Lean: indicated.

Lemma 4.20 (Proposition III.1.8). Sketch: conjugate the plaquette holonomy of \(Q\) at \(x\) by \(p_0\). It factors into four padded links \(p_y^{- 1} Q \hspace{0.17em} p_{y'}\), with \(y, y' \in \{0, e_i, e_k, e_i + e_k \} \subset B_1\), and the intermediate pads cancel. The triangle inequality and 4.19 give \(\text{ch} \leqslant 4 \cdot 2 \max (1, \varepsilon^{- 1}) (2 C + D) L^{- 2} \hspace{0.17em} \text{port}\). For \(V\), \(\text{ch} (V_{(x, p)} Q_{(x, p)}^{- 1}) \leqslant \text{fuel}\) by a telescoping product estimate. Lean: qdc, comp.

4.5 Near the trivial configuration

Lemma 4.21 (Proposition III.1.14). Sketch:

  1. \(U_0\) lies in \(M_{\delta}\), which is open, and at \(v = 0\) the frame is \(h \equiv 1\) by (B1).

  2. By 4.10(b) the frame solves \(h_0 = 1\) and \(h_y = \pi (\text{avg}_{z, o} h_z T_{U_v} (y \to z ; o)^{- 1})\) for \(y \ne 0\). This system is real analytic near \((0, 1)\). A variation \(u\) in the kernel of its partial derivative satisfies \(u_0 = 0\) and \(u_y = \Pi_{\mathfrak{g}} (\text{avg} \hspace{0.17em} u)\) for \(y \ne 0\). If \(n = |B_R |\) and \(A = \Pi_{\mathfrak{g}} (\text{avg} \hspace{0.17em} u)\), then \(A = (n - 1) A / n\), so \(u = 0\). Finite dimensionality makes the derivative invertible. Continuity of the selected minimizer identifies it with the analytic implicit solution.

  3. Differentiate the fixed-point equation, using \(D \pi (1) = \Pi_{\mathfrak{g}}\) and the fact that the derivative of a transport along \(w\) is the line integral of \(v\) along \(w\). This gives \(h'_y = \Pi_{\mathfrak{g}} (\text{avg} \hspace{0.17em} h') + \tau (y)\); with \(h'_0 = 0\) this is \(h'_y = \tau (y) - \tau (0)\).

  4. Near \(0\), \(\min \Phi_Y < \varepsilon\) (it is \(0\) at \(v = 0\)), so \(Q_b = \pi (\sum w_y Y_y)\), which is analytic. Its derivative is \(\sum_y w_y Y'_y = (\mathcal{M}v + d_0 \mathcal{F}v) (b)\).

  5. For an exact form \(d_0 \varphi\), the line integrals telescope; this gives (d).

Lean: hats_zero, stars, scheme, focused, derivative, lips, prices, root_gradient, Exterior.curvature_linear.

4.6 Tree gauge

Lemma 4.22 (Proposition III.1.13). Sketch: (a) Because \(L\) is odd, blocks are centred at their roots and the offset \(q\) lies in \(B_{(L - 1) / 2}\). A step towards \(0\) along an axis with \(q_a \ne 0\) decreases \(|q|_1 \leqslant 4 L\) and stays in the block. (b) \(\text{fold}_y \hspace{0.17em} U_{\text{up} (y)} = \text{fold}_{y + \text{up} (y)}\), by induction on the height. (c) Split the links into trunks and free links; this split preserves the measure. For fixed trunk values, gauging by \(\text{fold}\) (trunks set to \(1\)) sends the trunk values to \(1\). On the free links it acts as \(s_e \mapsto k_{e^-} s_e k_{e^+}^{- 1}\), which preserves product Haar measure, and \(F\) is unchanged because \(k = 1\) at the roots. (d) \(U \mapsto \rho_{\beta} (U, V) F (U)\) is invariant under gauge transformations that are trivial at the roots, by III.1.2(a). Lean: select, compress, names, pack, pushing.

4.7 Formats and energy accounting

Lemma 4.23 (the theorems of III.1.9).

Lean: gas_one, UsesFinal.allReads, Rep.envelope, partition_eq.

Lemma 4.24 (Proposition III.1.10). Sketch: expand the six works. The adjointness \(\langle \Phi q, p \rangle = \langle q, \Phi^{\dagger} p \rangle\) matches the cross terms of \(W_F\), \(W_I\) with \(W_L\), and the squares \(\tfrac{1}{2} \| \Phi w\|^2\), \(\tfrac{1}{2} \| \Phi' w\|^2\), \(\tfrac{1}{2} \|\mathcal{N}w\|^2\), \(\tfrac{1}{2} \|w\|^2\) cancel against \(W_D\), \(W_N\), \(W_C\). What remains is the right-hand side. III.1.11 requires a lower bound for the corresponding local sum near the seeds; the identity alone supplies no such bound. Lean: Tube.locates, arity.

Lemma 4.25 (transfer of the energy per seed, III.1.11). Assume the ringing condition and \(j_{\min} \leqslant i \leqslant j \leqslant i + 1\). Sketch: the stamp for the pair \((j, i)\) gives \(\mu\) and a \(\beta\)-entry threshold. In a directed history:

Lean: notation_ticket.

5. Where the difficulty lies

The synchronization cutoff must be independent of the patch size. The feedback map alone has Lipschitz constant \(1\); normalizing the centroids removes its tangential average and gives the contraction factor \(d + \tfrac{1}{4} \leqslant \tfrac{1}{2}\).

The curvature estimate must hold on both branches. It combines clustering of arbitrary minimizers, route multiplicities and curvature-controlled axial gauges. Losing a power of \(L\) in any of these comparisons would spoil the area scaling in III.1.8.

Normalization and tree-gauge integration use finite sums and Haar invariance. Global regularity is more limited: the kernel is Borel, while analyticity is proved near the trivial configuration. The energy bound and stationary requirements in III.1.11–III.1.12 need the later units.

6. Notation

Written proof Lean Meaning
\(\rho_{\beta} (U, V)\), \(\mathcal{P}_{\beta}\) StepData.density, push block kernel and push
\(P = (L, J)\), \(R = (J + 3) L\) Patch, radius block factor, order, patch radius
\(K = (\kappa, \delta, \varepsilon, a)\) KernelChoices penalty, frame cutoff, predictor cutoff, axial factor
(B1)–(B3) Patch.Branches branch conditions
\(M_{\delta}\) usesMin minimizing branch
\(S_W (h)\) score synchronization cost on a patch
frame \(h\), tag \(t = (r, o)\) frame, FrameTag gauge fixing on a patch; root and axis ordering
\(\mathcal{E} (W)\) eligible eligible axial tags
\(w\) weight \(J\)-fold convolution of the block-uniform weight
\(Y\), \(\Phi_Y\), \(Q\) connector, bondCost, prediction connector, bond cost, prediction
\(\lambda (U, t)\) likelihood tag law
ch, pr Gauge.chord, Gauge.price chord distance to \(1\); Wilson plaquette density
\(E_X\), \(F_X\) Axial.edges, Axial.faces link and plaquette chord sums in \(X\)
\(m (h)\), \(F_x (h)\) Pairs.centroid, Pairs.feedback centroid and feedback of the synchronization problem
\(\pi\) Gauge.focus nearest point of \(G\) (near \(1\))
port, \(R'\) port, macroR fine curvature in the macro-box of radius \((J + 7) L\)
usable usable positive frame masses on the \(5^4\) neighbours
toll, wire, fuel toll, wire, fuel plaquette chord; chord of \(V_b Q_b^{- 1}\); sum of four wires
\(\mathcal{M}+ d_0 \mathcal{F}\), \(\tau\) probe, mean, fixRoot, tau linearized prediction (averaging operator)
forest, trunks, fold Tree.Forest, trunks, foldAt tree gauge in blocks
budget point, flags BudgetPoint, flags norm parameters; large-field sites
\(\mathcal{R}\), eval Rep, Rep.eval effective density: exponential of the vacuum term minus the action and regular activities, times the gas
\(\hat{\beta}_m\), \(t_m\) Calendar.hatBeta, Calendar.threshold reference coupling; small-field threshold \(\text{temp} (\hat{\beta}_m)\)
tube, works Tube, Tube.vision transfer tower with fine and coarse path data; six local works
store, plot, Work Stores, Stores.Plot, assessed seeds; their large-field characterization; local work
estimator data, energy-per-seed requirement Teller, Teller.Grants parameters of one energy estimate; energy-per-seed requirement
\(t = \text{temp} (\hat{\beta} + \kappa_{\text{slope}} \log L)\) fiery, temp, betaF fine-level threshold
\(\varrho\) radiusB cluster radius \(\lceil (\log \hat{\beta})^2 \rceil\)
kinetic data, directed history Kin.Planned, Directed one-loop shift and stationary-branch requirements

7. What remains unexplained

  1. Energy estimates. III.1.10 is an identity for the clipped direct term and truncated rows. The estimates converting it into III.1.11, including bounds on \(W_D\) and the truncation remainders, belong to III.10. No sign is asserted here for \(1 - \Phi^{\dagger} {\Phi - \Phi'}^{\dagger} \Phi' -\mathcal{N}^{\dagger} \mathcal{N}\) at arbitrary \(V\).

  2. Construction of the kinetic data. The stationary conditions and the outputs of the two cited construction theorems are identified in III.1.12. Their full construction from the estimates of III.3, III.6–III.7 and III.9 is outside this unit. The alternative constructor Extra.Iteration.planned has the required output but is outside the main-theorem closure.

  3. Choice of neighbourhood and smoothing order. The curvature proof uses pads at plaquette corners in \(B_1\), whereas the stated synchronization result supplies them throughout \(B_2\). The need for the larger neighbourhood in later estimates is not determined here. Likewise, \(w \leqslant L^{- 4}\) needs only \(J \geqslant 1\); the later use of the stronger patch condition \(J \geqslant 4\) remains to be specified.

  4. Additional files. The small packages listed in the header are not described by the numbered results. Their contributions, and any further results from the exterior calculus or quadratic-kernel files beyond those cited here, remain to be incorporated into the relevant later units.

← II.2 Marker algebra, polymer combinatorics and Gaussian estimates (notes)Part III·Unit outlineIII.2 Choosing the constants: the order of choice and the RG schedule (notes) →