1. What this unit proves 2
2. What it uses 8
3. Strategy 9
4. Steps 10
4.1 Filters and histories 10
4.2 Explicit metric data 11
4.3 Transfer kernels of one flight 12
4.4 The defect and the amendment 14
4.5 The catalogue 15
4.6 Along a schedule: pre-repair, repair, output 15
4.7 All tori 17
4.8 Free covariance of the curvature (III.3.6) 17
4.9 Calculus lemmas (III.3.7) 18
4.10 Auxiliary results 18
5. Where the difficulty lies 19
6. Notation 20
7. What remains unexplained 21
The full account of unit III.3: 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/Next/FREE_KERNELS/ (5,450 lines), YM/MultiNext/SIGNED_SOURCE_CONTINUATION/ (1,674), YM/Further/FREE_COVARIANCE_KP/ (924), YM/UVNext/PRICED_KERNELS/ (533), YM/Further/CORRECTED_LIFT_COEFFICIENTS/ (248), YM/Push/CUTOFF_PROJECTORS/ (139), YM/ProofPorts/GRID_RENEWED_RESERVES/ (30). Size: 8,998 lines in the main-theorem closure.
Conventions.
Tori and forms. \(\Lambda_n = (\mathbb{Z}/ n\mathbb{Z})^4\), \(n \geqslant 1\). \(\Omega^r (\Lambda_n)\) is the space of real \(r\)-cochains, with the counting inner product. \(d_r : \Omega^r \to \Omega^{r + 1}\) is the exterior derivative, and \(d = d_1\). \(X^{\ast}\) denotes the transpose on real cochains and the Hermitian adjoint of a symbol at real momentum. Displacements use \(|y|_1\); real momenta use the Euclidean norm \(|p|\).
Levels. In mesh \(N\) the levels are \(i = N, \ldots, 0\). Level \(N\) is the bare lattice. Step \(i < N\) maps level \(i + 1\) (fine torus \(\Lambda_f\)) to level \(i\) (coarse torus \(\Lambda_c\)), with \(f = cL_i\) and \(L_i \geqslant 3\) odd. Level \(i\) has a stage index \(J (i) \geqslant 4\), and \(J (i) \leqslant J (i + 1) \leqslant J (i) + 1\).
Kernels. A kernel is a map \(R : \mathbb{Z}^4 \to \text{Hom} (\mathbb{R}^{\beta}, \mathbb{R}^{\alpha})\), where \(\alpha, \beta\) are sets of face types. It acts on every torus by periodisation: \((R_n \omega) (x) = \sum_{y \in \mathbb{Z}^4} R (y) \hspace{0.17em} \omega (x + y)\). Its symbol is \(\hat{R} (p) = \sum_y R (y) e^{ip \cdot y}\).
Cross kernels. A cross kernel of block size \(L\) also carries a fine residue \(r \in (\mathbb{Z}/ L)^4\) in its row index. It maps cochains on \(\Lambda_c\) to cochains on \(\Lambda_{cL}\), for every \(c\). Its displacements are measured on the coarse lattice.
Strip bound. \(R \in \mathcal{S} (\alpha, \Omega)\) means \(\sum_y e^{\alpha |y|_1} |R_{ab} (y) | \leqslant \Omega\) for every entry. This implies that \(\hat{R}\) is analytic and bounded on \(| \text{Im} \hspace{0.17em} p_a | < \alpha\). Conversely, a bounded periodic analytic symbol has such a weighted coefficient bound on each narrower strip.
Aliases and lattice momenta. For a coarse momentum \(p\), the fine aliases are \(q_{\ell} = (p + 2 \pi \ell) / L\), \(\ell \in (\mathbb{Z}/ L)^4\) (centred representatives). \(\tau (p) = (e^{ip_a} - 1)_{a = 1}^4\) is the lattice momentum, and \(\sigma_r (p)\) is the symbol of \(d_r\), so \(\sigma_0 = \tau\).
Complex cubes. Write \(\mathcal{C}_{\theta} = \{p \in \mathbb{C}^4 : | \text{Re} \hspace{0.17em} p_a | < \pi + \theta, | \text{Im} \hspace{0.17em} p_a | < \theta\}\) and \(\mathcal{C}_{\theta}^{\mathbb{R}} =\mathcal{C}_{\theta} \cap \mathbb{R}^4\). Absolute squares in the real-momentum formulas are continued holomorphically as \(g (p) g (- p)\) for real-coefficient filters, and \(| \tau (p) |^2\) as \(\sum_a (e^{ip_a} - 1) (e^{- ip_a} - 1)\).
Constants from the kernel choice. \(\kappa > 0\) is the penalty of the kernel choice \(K\) (the number \(b\) of the main theorem). \(c_i > 0\) is the direct coefficient of the clip of level \(i\), that is, the coefficient of the bare plaquette term in the quadratic part of the action.
The unit computes the Gaussian part of the block-spin step of III.1 in closed form, for every history of steps. It proves three things about it:
exact algebraic identities on every torus, giving an orthogonal splitting of the quadratic form;
exponential decay, uniform in the block sizes, history length and torus, with constants fixed by the kernel choice and stage index, and, after repair, by the sheet's direct clip coefficients;
uniform ellipticity and hypercubic symmetry of the quadratic forms.
The constants are fixed before the block sizes (Proposition III.3.1). Positivity of the rolled kernel requires the lower bound \(L \geqslant b_j\).
The package also contains:
the free covariance of the curvature and its comparison with the continuum (III.3.6, used in III.11);
calculus lemmas used in the nonlinear step (III.3.7, used in III.10 and III.2).
The identities of one step. Fix a step with tori \(\Lambda_f, \Lambda_c\), \(f = cL\). Let \(Q : \Omega^1 (\Lambda_f) \to \Omega^1 (\Lambda_c)\) be the linearised block average of III.1: a weighted average, over a \(J\)-fold smoothing patch, of line integrals of the fine 1-form along \(L\)-link paths.
Metric data on a level is a quintuple \(\mathcal{B}= (D, D^{- 1}, H, H^{- 1}, R)\) of kernels on 2-forms: a diagonal multiplier, its inverse, a metric, its inverse, and a root. With the direct coefficient \(c\) of the level's clip, set
Then \(V^{\ast} V = c \hspace{0.17em} 1 + R^{\ast} R = H\) and \(\langle dA, H \hspace{0.17em} dA \rangle = |V \hspace{0.17em} dA|^2\).
Transfer kernels are \(\mathcal{U}= (W, T, M_f, M_n, O)\) with:
\(W : \Omega^1 (\Lambda_c) \to \Omega^1 (\Lambda_f)\) and \(T : \Omega^2 (\Lambda_c) \to \Omega^2 (\Lambda_f)\) (cross kernels);
\(M_f : (\Omega^2)^{\oplus 2} (\Lambda_c) \to (\Omega^2)^{\oplus 2} (\Lambda_f)\) (cross kernel);
\(M_n : (\Omega^2)^{\oplus 2} (\Lambda_c) \to \Omega^1 (\Lambda_c)\);
\(O : (\Omega^2)^{\oplus 2} (\Lambda_c) \to ((\Omega^2)^{\oplus 2} \oplus \Omega^3) (\Lambda_c)\).
With fine data \((\mathcal{B}, V)\) and coarse data \((\mathcal{B}', V')\), the identities are:
(D1) \(d^{\ast} H \hspace{0.17em} d \hspace{0.17em} W = Q^{\ast} d^{\ast} H' d\)
(D2) \(QW = 1 - \kappa^{- 1} d^{\ast} H' d\)
(D3) \(T^{\ast} d = dQ\)
(D4) \(T \hspace{0.17em} D' d = D \hspace{0.17em} d \hspace{0.17em} W\)
(R1) \(M_f^{\ast} M_f + M_n^{\ast} M_n + O^{\ast} O = 1\)
(R2) \(\sqrt{\kappa} \hspace{0.17em} M_n^{\ast} = V' d\)
(R3) \(M_f V' d = V \hspace{0.17em} d \hspace{0.17em} W\)
(R4) \(M_n V' d = \kappa^{- 1 / 2} d^{\ast} H' d\)
(R5) \(OV' d = 0\)
(R6) \(M_f^{\ast} V \hspace{0.17em} d = \sqrt{\kappa} \hspace{0.17em} M_n^{\ast} Q\)
(Δ) \(H' \left( {H'}^{- 1} - T^{\ast} H^{- 1} T - \kappa^{- 1} d \hspace{0.17em} d^{\ast} \right) H' = d_2^{\ast} P \hspace{0.17em} d_2\), for a 3-form kernel \(P\) on the coarse level.
What the identities mean:
(D1)–(D2). These say that \(A = WB\) is a critical point, hence a minimiser, of \(A \mapsto \langle dA, H \hspace{0.17em} dA \rangle + \kappa |QA - B|^2\), and that the minimum is \(\langle dB, H' dB \rangle\). So \(W\) is Balaban's background-field (minimiser) map, and \(H'\) is the effective metric of the coarse level. The value identity is III.1's action.
(D3)–(D4). (D3) makes \(T^{\ast}\) a 2-form average compatible with \(Q\). (D4) says that \(T\) carries the weighted coarse curvature \(D' dB\) to the weighted fine curvature \(D \hspace{0.17em} dWB\) of the background field.
(R1)–(R6): orthogonal splitting. For \(A \in \Omega^1 (\Lambda_f)\), \(B \in \Omega^1 (\Lambda_c)\) and \(Z = A - WB\), (R3), (R4) and (D2) give
The two summands are orthogonal by (R6), and \(| (M_f, M_n) V' dB| = |V' dB|\) by (R1) and (R5). Hence
(Δ). This is the same splitting for inverse metrics. It says \({H'}^{- 1} = T^{\ast} H^{- 1} T + \kappa^{- 1} dd^{\ast} {+ H'}^{- 1} d_2^{\ast} Pd_2 {H'}^{- 1}\), and the last term sees only \(d_2 {H'}^{- 1} \omega\).
Proposition III.3.1 (catalogue fixed before the block sizes). Let \(K\) be a kernel choice. There are numbers \(\alpha_j > 0\), \(\eta_j > 0\), \(\Omega_j \in \mathbb{R}\), \(\lambda_j \geqslant 0\) (amendments), \(\gamma_j > 0\) (clip bounds) and \(b_j \in \mathbb{N}\) (block thresholds), \(j \in \mathbb{N}\), depending only on \(K\), with the following properties.
(a) Metric data of a history. Let \(j \geqslant 4\), let the clip have direct coefficient \(c \leqslant \gamma_j\), and let \(v\) be a history of stage index \(j\). Then there is metric data \((D, D^{- 1}, H, H^{- 1}, R)\) such that:
\(D\) and \(D^{- 1}\) are the explicit kernels of \(v\) with penalty \(\kappa\), and \(H^{- 1} = D^{- 1} + \lambda_j (d_2 D^{- 1})^{\ast} (d_2 D^{- 1})\) (Lemmas 4.4–4.7);
all five kernels lie in \(\mathcal{S} (\alpha_j, \Omega_j)\);
on every torus: \(DD^{- 1} = HH^{- 1} = 1\), \(c \hspace{0.17em} 1 + R^{\ast} R = H\), \(\hat{H} (0) = 1\), \(Hd = Dd\), and \(H\) and \(D\) are symmetric;
on every torus: \(\eta_j | \omega |^2 \leqslant \langle \omega, H \omega \rangle \leqslant \Omega_j | \omega |^2\), and the same for \(D\);
for every gauge group \(G\), the form \(\langle A, d^{\ast} Hd \hspace{0.17em} A \rangle\) on \(\mathfrak{g}\)-valued 1-forms is invariant under the signed lattice symmetries;
\(D = 1\) if \(v\) is the empty history.
(b) One step. Let \(j, t \geqslant 4\) with \(t \in \{j, j - 1\}\), and let \(P\) be a flight of index \((j, t)\): odd block size \(L \geqslant 3\) and fine history of index \(j\). Then:
the explicit transfer kernels \(W, T\) of \(P\) satisfy \(\nabla_u W \in \mathcal{S} (\alpha_j, \Omega_j L^{- 1 - |u|})\) and \(\nabla_u T \in \mathcal{S} (\alpha_j, \Omega_j L^{- 2 - |u|})\), for every word \(u\) of at most \(j - 2\) fine forward differences;
the rolled 3-form kernel \(\mathcal{P}\) of \(P\), built with fine amendment \(\lambda_j\) and coarse amendment \(\lambda_t\) (Lemma 4.12), lies in \(\mathcal{S} (\alpha_j, \Omega_j)\);
if \(L \geqslant b_j\), then at every real momentum \(\hat{\mathcal{P}} (p)\) is Hermitian and \(\eta_j \leqslant \hat{\mathcal{P}} (p) \leqslant \Omega_j\).
Status: theorem, with no hypothesis beyond the kernel choice. It uses III.6's square-root theorem to construct \(R\). The catalogue does not depend on \(G\). Lean: Catalog, pledges, BottomQuote, HopQuote, Mag, Defined, Floor, Donate, Funds.
Proposition III.3.2 (free kernels along an admissible schedule). Let \(G\) be a gauge group and \(K\) a kernel choice. There is a catalogue \((\alpha, \Omega, \eta, \lambda, \gamma, b)\) as in III.3.1 with the following property. For every sheet \(\mathfrak{S}\) with kernel \(K\) (the pre-\(L\) constants, III.2) there is a second table \((\alpha'_j, \Omega'_j, \eta_j)\) such that the conclusion below holds.
Let \(A\) be an RG trajectory with kernel \(K\) and the stage profile of \(\mathfrak{S}\), satisfying the credit condition:
(C1) \(c_i \leqslant \gamma_{J (i)}\) for every level \(i\);
(C2) \(b_{J (i + 1)} \leqslant L_i\) for every step \(i\).
Then, simultaneously for all \(N \geqslant 0\) and \(0 \leqslant i \leqslant N\), there are two families of metric data and transfer kernels: a pre-repair family \(\textsf{M} = (\mathcal{B}^{N, i}, \mathcal{U}^{N, i})\) and an output family \(\textsf{H} = (\mathcal{B}^{N, i}, \tilde{\mathcal{U}}^{N, i})\). They have the same metric data and the same \(W, T\), and satisfy:
Explicit formulas. \(\mathcal{B}^{N, i}\) is the metric data of III.3.1(a) for the history of level \(i\) in mesh \(N\), with amendment \(\lambda_{J (i)}\). That history is the composite of the stage hops of steps \(N - 1, \ldots, i\), starting from the empty history at level \(N\). \(W^{N, i}, T^{N, i}\) are the explicit transfer kernels of the flight of step \(i\), of index \((J (i + 1), J (i))\).
Pre-repair conditions (\(\textsf{M}\)).
\(\mathcal{B}^{N, i} \in \mathcal{S} (\alpha_{J (i)}, \Omega_{J (i)})\), with lower and upper bounds \(\eta_{J (i)}, \Omega_{J (i)}\) on every torus for the clip of level \(i\), and the symmetry;
\(D^{N, N} = 1\);
\(W, T\) have the derivative bounds of III.3.1(b) at stage \(J (i + 1)\), with \(J (i + 1) - 2\) differences;
(D1)–(D4) hold on every schedule torus pair of step \(i\);
there is a 3-form kernel \(P \in \mathcal{S} (\alpha_{J (i + 1)}, \Omega_{J (i + 1)})\), Hermitian at real momenta, with \(\langle v, \hat{P} (p) v \rangle \geqslant \eta_{J (i + 1)} |v|^2\) for real \(p\) and \(v\) in the range of \(\sigma_2 (p)\), such that (Δ) holds on every schedule torus pair.
Output conditions (\(\textsf{H}\)), with the second table.
the metric bounds of III.3.1(a), with strip constants \(\alpha'_j, \Omega'_j\) and the same lower bounds \(\eta_j\);
\(\tilde{M}_n, \tilde{O} \in \mathcal{S} (\alpha'_{J (i + 1)}, \Omega'_{J (i + 1)})\) and \(\tilde{M}_f \in \mathcal{S} (\alpha'_{J (i + 1)}, \Omega'_{J (i + 1)} L_i^{- 2})\), and the derivative bounds on \(W, T\) with \(\alpha'_j, \Omega'_j\);
(D1)–(D4) and (R1)–(R6) on every schedule torus pair, with the clips of levels \(i + 1, i\);
\(D^{N, N} = 1\), and the symmetry of every \(\mathcal{B}^{N, i}\).
Status: theorem. The inputs F06–F09 (III.6) are discharged inside the proof. For trajectories with the stated kernel and profile, the additional hypothesis is the credit condition. The variant before_paid (assigned to III.4) replaces it by:
two conditions on the sheet: clip coefficients \(\leqslant \gamma_j\), and block-size floors \(\geqslant b_j\), for all stages from the first one on;
the schedule's payment record (III.2), which gives \(L_i \geqslant\) floor.
III.9 builds the sheet so that these hold: its floors are exactly \(b\) and its clip coefficients are \(\leqslant \gamma\) (housing). The variant also lets the second table be computed from a preliminary sheet \(\mathfrak{S}\) and used for any later sheet \(\mathfrak{S}'\), provided \(\mathfrak{S}'\) has the same direct clip coefficients, the same kernel, a drift at least as large, and logarithmic parameter size at least as large.
Lean: retained_support; conditions Present, Craft.Rack, Bid.Retained, Horizon.Suffices, Bid.Sworn, Credit; discharge credit_from_sheet, Willing.
Proposition III.3.3 (identities on every torus). Let a family satisfy the output conditions of III.3.2(3) for some table. For every \(N\), \(m < N\) and positive sides \(c, f\) with \(f = cL_m\), (D1)–(D4) and (R1)–(R6) hold, with the same kernels and the clips of levels \(m + 1, m\). The diagonal identities follow by rigidity (III.10, panes). They also follow from the explicit formulas when the family realises them (universal_diagonal). Given the diagonal identities, the roll identities follow from the output conditions by the same rigidity argument. Status: theorem. Lean: complete_roll, Towers.Diagonal, Towers.Roll, Inputs. Used in III.10.
Proposition III.3.4 (positivity of the rolled 3-form kernel). Fix a kernel choice \(K\) and \(j, t \geqslant 4\). There are \(C_0, B > 0\), depending only on \((K, j, t)\), with the following property. For every coarse amendment \(\lambda \geqslant C_0\) and fine amendment \(\mu \geqslant 0\) there are \(\eta, \sigma, \Omega > 0\) such that, for every flight \(P\) of index \((j, t)\):
\(\mathcal{P}_{\mu, \lambda} (P) \in \mathcal{S} (\sigma, \Omega)\);
if \(\mu B \leqslant L^2\), then at every real \(p\), \(\hat{\mathcal{P}}_{\mu, \lambda} (p)\) is Hermitian and \(\eta \leqslant \hat{\mathcal{P}}_{\mu, \lambda} (p) \leqslant \Omega\).
Status: theorem. Lean: Flight.frontier, Flight.absorb_before.
Proposition III.3.5 (line and box filters). For positive integers \(L, M\), define the uncentred line average below; for odd \(L, M\), also define the centred line and box averages:
(centred line average, uncentred line average, box average). Then:
(a) \(\text{rail}_L (Mz) \hspace{0.17em} \text{rail}_M (z) = \text{rail}_{LM} (z)\) for all positive \(L, M\), and \(\text{plank}_L (Mz) \hspace{0.17em} \text{plank}_M (z) = \text{plank}_{LM} (z)\) for odd \(L, M\). For odd \(M\), \(\text{plank}_M\) is even, \(2 \pi\)-periodic and real on \(\mathbb{R}\).
(b) \(| \text{plank}_M (z) |, | \text{rail}_M (z) | \leqslant e^{M | \text{Im} \hspace{0.17em} z|}\).
(c) There are absolute constants \(c_{\ast}, c_1 > 0\) such that \(| \text{rail}_M (z) | \geqslant c_{\ast}\) for complex \(z\) with \(M |z| \leqslant 3 \pi / 2\), and \(| \text{plank}_M (z) | \geqslant e^{- c_1 M / L}\) for real \(z\), \(M \leqslant L\) and \(L |z| \leqslant 3 \pi / 2\).
(d) There is an absolute \(B\) such that \(| \text{rail}_L ((q_{\ell})_a) |, | \text{plank}_L ((q_{\ell})_a) | \leqslant B / (1 + | \ell_a |)\). This holds for odd \(L\), every \(p\) in the complex cube \(\{| \text{Re} \hspace{0.17em} p_a | < \tfrac{5 \pi}{4}, | \text{Im} \hspace{0.17em} p_a | < \tfrac{\pi}{4} \}\), every alias \(\ell\) and every \(a\).
Status: theorem. Lean: plank, rail, area, plank_bind, rail_bind, plank_height, rail_lower, principal_plank, alias_rail.
Proposition III.3.6 (free covariance of the curvature). Call a real 2-form kernel \(f\) firm with constants \((\alpha, \Omega, \eta)\) if \(f \in \mathcal{S} (\alpha, \Omega)\), \(\hat{f} (0) = 1\), and \(\eta | \omega |^2 \leqslant \langle \omega, f_n \omega \rangle \leqslant \Omega | \omega |^2\) for every real cochain on every torus. Symmetry is not required. Fix \(\alpha, \eta > 0\) and \(\Omega \in \mathbb{R}\), and let \(U = \{p \in \mathbb{R}^4 \setminus \{0\}: |p_a | < 3 \pi / 2\}\). For \(p \in U\) define:
\(\Phi_f (p) = \sigma_1 (p)^{\ast} \hat{f} (p) \sigma_1 (p) + \sigma_0 (p) \sigma_0 (p)^{\ast}\), the symbol of \(d_1^{\ast} fd_1 + d_0 d_0^{\ast}\) on 1-forms;
\(G_f (p) = \sigma_1 (p) \Phi_f (p)^{- 1} \sigma_1 (p)^{\ast}\), the curvature covariance symbol when \(f\) is symmetric and the gauge-fixed potential has precision \(d_1^{\ast} fd_1 + d_0 d_0^{\ast}\);
\(G_{\infty} (p) = s (p) |p|^{- 2} s (p)^{\ast}\), where \(s (p)\) is wedging with \(ip\) on 1-forms: the continuum Feynman-gauge curvature symbol.
(a) There is \(b > 0\), depending only on \(\alpha, \Omega, \eta\), such that \(\text{Re} \hspace{0.17em} \langle x, \Phi_f (p) x \rangle \geqslant b |p|^2 |x|^2\) for every firm \(f\), \(p \in U\) and complex 1-form \(x\).
(b) For every \(n \in \mathbb{N}\) there is \(C_n \geqslant 0\), depending only on \(n, \alpha, \Omega, \eta\), such that, for all firm \(f\) and \(p \in U\), \(\|D^n G_f (p)\| \leqslant C_n |p|^{- n}\) and \(\|D^n (G_f - G_{\infty}) (p)\| \leqslant C_n |p|^{1 - n}\).
(c) If \(f\) is a scalar kernel with \(\sum_x e^{\alpha |x|_1} |f (x) | \leqslant \Omega\), then \(\hat{f}\) is \(C^{\infty}\) on \(\mathbb{R}^4\) and \(\|D^n \hat{f} (p)\| \leqslant A_n \Omega\) for every \(p\), where \(A_0 = 1\) and \(A_n = (n / \alpha)^n\) for \(n \geqslant 1\). Here \(D^n\) is the full \(n\)th Fréchet derivative, with its multilinear operator norm; this notation also applies to (b).
Status: theorems. Part (b) is used in III.11. Lean: tank_full, gem_quote, radial_prices, Firm, fuel, garnet, pearl.
Proposition III.3.7 (calculus lemmas for the nonlinear step). These do not concern free kernels; they are consumed in III.10 and III.2.
(a) Discrete Taylor expansion with scaled bounds. Let \(b, k \in \mathbb{N}\) with \(k \leqslant b\). There is \(C = C (b, k) \geqslant 1\) with the following property. Let \(L \geqslant 1\), \(M \geqslant 0\), and let \(p : \mathbb{Z}^4 \to E\) (\(E\) a real normed vector space) satisfy \(\| \nabla_u p (x)\| \leqslant ML^{- 1 - |u|} (1 + L^{- 1} |x|_1)^b\) for all words \(u\) of forward differences with \(|u| \leqslant k\). Then the Newton pieces \(P_i (x) = \sum_{| \nu | = i} \binom{x}{\nu} \nabla^{\nu} p (0)\), of degree at most \(i\), satisfy \(\|P_i (x)\| \leqslant CML^{- 1 - i} (1 + |x|_1)^b\) for \(i < k\), and the remainder satisfies \(\|p (x) - \sum_{i < k} P_i (x)\| \leqslant CML^{- 1 - k} (1 + |x|_1)^b\).
(b) Weighted Taylor coefficients. Let \(d \in \mathbb{N}\) and let \(f : E \to F\) be \(C^d\) at \(0\) between real normed vector spaces. Let \(p_v : E \to E\) (\(v \in V\) finite) be continuous linear maps with \(\sum_v p_v = 1\), and let \(Y_v \geqslant 1\) be integers. Suppose that for every \(z\) the function \(t \mapsto f (\sum_v t^{Y_v} p_v z)\) has vanishing derivatives of orders \(0, \ldots, d\) at \(0\). Then \(D^m f (0) (x_1, \ldots, x_m) = 0\) whenever there are integers \(k_i \geqslant 1\) with \(\sum_i k_i \leqslant d\) and \(p_v (x_i) = 0\) for every \(v\) with \(Y_v > k_i\).
(c) Moments of local gauge-invariant functions. Fix \(b, d \in \mathbb{N}\) with \(b \leqslant d\), a torus \(\Lambda_n\) with \(n \geqslant 2\), an anchor \(o\), and a finite support \(X\) satisfying \(10 \hspace{0.17em} \text{dist} (x, o) < n\) for every \(x \in X\), where \(\text{dist}\) is the torus \(\ell^{\infty}\) distance. Let \(u : \text{Config} (G, \Lambda_n) \to \mathbb{C}\) be gauge invariant and depend only on links whose two endpoints lie in \(X\), with \(u \circ \exp\) of class \(C^b\) at \(0\). For every radial polynomial connection jet \(J\) through weight \(d\), let \(U_{J, o} (t)\) be its link holonomies at scale \(t\), based at \(o\): a homogeneous connection component of degree \(r - 1\) has weight \(r\), \(2 \leqslant r \leqslant d\), and is multiplied by \(t^r\). Assume \(\frac{d^k}{dt^k} u (U_{J, o} (t)) |_{t = 0} = 0\) for every \(J\) and \(0 \leqslant k \leqslant b\). For any \(m \in \mathbb{N}\), \(p_1, \ldots, p_m \geqslant 1\) with \(\sum_i p_i \leqslant b\), and link fields \(w_i\) whose values on those links are polynomials of degree at most \(p_i - 1\) in the centred lift of the initial site relative to \(o\), one has \(D^m (u \circ \exp) (0) (w_1, \ldots, w_m) = 0\). Values of \(w_i\) on other links are unrestricted.
(d) Radial decomposition. Let \(j \in \mathbb{N}\) and let \(Q_a : \mathbb{R}^4 \to \mathfrak{g}\) (\(a = 1, \ldots, 4\)) be homogeneous polynomials of degree \(j + 1\). Then \(Q_a = K_a + \partial_a r\), where \(r = (j + 2)^{- 1} \sum_a x_a Q_a\) is homogeneous of degree \(j + 2\), and \(K\) is homogeneous of degree \(j + 1\) with \(\sum_a x_a K_a = 0\) (the Fock–Schwinger condition).
(e) Small-field condition along rays. There is \(C = C (G) > 0\) with the following property. Let \(0 \leqslant b \leqslant 1\), \(\tau \geqslant Cb^2\), and let \(v\) be Lie-algebra link variables with \(|v_x | \leqslant b\) on all links within distance \(5\) of \(Y\). If every plaquette touching \(Y\) has holonomy within chord distance \(\tau\) of \(1\) for the configuration \(\exp (v)\), then the same plaquette bounds hold for \(\exp (sv)\), \(0 \leqslant s \leqslant 1\).
Status: theorems. Lean: capital, cancel, converse, wiped, locale, strings.
From other units.
I.1. The lattice, the signed (hypercubic) symmetries, and the gauge group. In the free-kernel construction the gauge group enters only the symmetry statement; III.3.7 also uses its exponential and holonomy maps.
III.1. The block-spin data:
the block ratio, the patch (odd \(L \geqslant 3\), \(J \geqslant 4\)), the kernel choices KernelChoices and the clips Clip;
the linearised average \(Q\) (mean) and the lattice exterior derivative;
the trajectory record Levels: block sizes, stage indices with \(J (i) \leqslant J (i + 1) \leqslant J (i) + 1\), and schedule tori whose sides tend to infinity.
III.1 also defines the output objects:
transfer kernels Towers and their bounds Towers.Deep;
the identities (D1)–(D4), (R1)–(R6), with \(d^{\ast} Hd\) as K;
kernel families Horizon, the output conditions Horizon.Suffices, and price tables MapPrices.
III.2. The sheet Sheet, the block sizes Rolls, and the payment record Paid.
III.6. The four linear-algebra theorems F06, F07, F08, F09. Their statements are in Lemma 4.16. F07 is also used in Lemma 4.8.
III.10.
The interface conditions Craft.Rack, Bleeps, Reaches, Funds, Donate, Replaces, Bid.Retained, Bid.Sworn.
The repair theorem Bid.pinning (Lemma 4.16), and panes for III.3.3.
III.16. Kernels and their action on tori: Row, Cross, installed. The weighted norms Row.strip and Laurent.Width, where the size of a displacement is its \(\ell^1\) norm. The alias-resolved Fourier transform on \(\Lambda_f\), which is unitary up to \(f^4\).
Standard mathematics.
Fourier analysis on \(\mathbb{Z}^4\) and on periodic lattices. Bloch (alias) decomposition of operators that commute with coarse translations. Periodic strip analyticity gives exponential coefficient bounds on narrower strips.
Gaussian block-spin transformations computed in Fourier space: K. Gawędzki, A. Kupiainen, Massless lattice \(\varphi^4_4\) theory, CMP 99 (1985). Balaban's averaging and background-field operators for lattice gauge fields: T. Balaban, CMP 95 and 96 (1984), CMP 98 (1985). Exposition: J. Dimock, The renormalization group according to Balaban I, Rev. Math. Phys. 25 (2013).
Lagrange's identity. The Schur-complement (orthogonal) splitting of a quadratic form.
Exactness of the lattice de Rham (Koszul) complex for exponentially decaying translation-invariant kernels.
Newton forward-difference interpolation. Weighted-homogeneous Taylor expansions. The Fock–Schwinger (radial) gauge.
Imported constants and assumptions.
The kernel choice \(K\); only the penalty \(\kappa\) enters the free kernels.
The analyticity half-width \(\pi / 4\) (zone) and the radius \(3 \pi / 2\) (domainRadius). The proved compactness bounds \(c_- |w| \leqslant |e^{iw} - 1| \leqslant c_+ |w|\) on \(|w| \leqslant 3 \pi / 2\) (meter).
The credit condition, a hypothesis of III.3.2. It is discharged through before_paid, the construction of the sheet in III.9, and the payment record of III.2.
F06–F09 are proved in III.6 and plugged in inside retained_support.
The Gaussian block-spin problem is solved in Fourier space. Iterating the block average and its independent noise gives a diagonal 1-form covariance symbol. Minimising its precision over gauge directions gives the effective form \(d^{\ast} Dd\) by Lagrange's identity. The background-field map \(W\) and curvature transfer \(T\) are then explicit alias sums satisfying (D1)–(D4).
Uniform decay follows from the filters. A stage-\(J\) box filter gives \(\prod_a (1 + | \ell_a |)^{- J}\) alias decay, leaving \(J - 2\) fine differences summable. Along a history, successive dilation ratios decrease at least geometrically by \(1 / 3\), so the extra filter products and accumulated noise stay uniformly bounded. Before estimating inverse symbols, the principal-alias poles are cancelled. The remaining denominators are bounded below at real momenta and hence on a thinner complex cube. These constants depend on the kernel choice and stage, uniformly in history length and block sizes.
The inverse-metric defect factors through \(d_2\) as \({X'}^{\ast} CX'\), where \(X' = d_2 {D'}^{- 1}\) and \(C\) need not be positive. Amending the fine and coarse inverse metrics leaves their quadratic forms on potentials unchanged and replaces \(C\) by \(C + \lambda_t - \lambda_j S^{\ast} S\). The squared 3-form transfer satisfies \(S^{\ast} S = O (L^{- 2})\), including the sum over fine residues. Choose the coarse amendments first, then require \(\lambda_j B / L^2 \leqslant 1\). The resulting positive defect yields \(M_f, M_n, O\) and (R1)–(R6) through III.6 and the repair theorem of III.10. The repair constants also depend on the sheet's direct clip coefficients. Rigidity extends the schedule identities to all tori.
Lemma 4.1 (filter algebra). This is III.3.5(a). Sketch: \(M \hspace{0.17em} \text{rail}_M (z) (e^{iz} - 1) = e^{iMz} - 1\); apply it twice and cancel \(e^{iz} - 1\) when nonzero. If \(e^{iz} = 1\), every uncentred average is \(1\). For the centred filter, the centring phases add up, since \(M \hspace{0.17em} \frac{L - 1}{2} + \frac{M - 1}{2} = \frac{LM - 1}{2}\). Lean: rail_bind, plank_bind.
Lemma 4.2 (size of filters). This is III.3.5(b)–(d). Sketch:
(b) Triangle inequality.
(c) The compactness constants and Lemma 4.1 give \(| \text{rail}_M | \geqslant c_- / c_+\). On the principal alias, averaging \(|e^{isz} - 1| \leqslant c_+ |sz|\) gives \(| \text{plank}_M (z) - 1| \leqslant c_+ \tfrac{3 \pi}{2} \tfrac{M}{L}\), so \(| \text{plank}_M | \geqslant \max (c_{\ast}, 1 - kx) \geqslant e^{- c_1 x}\) with \(x = M / L\). The exponent is linear in \(x\) with no prefactor, which is what makes products along a history converge.
(d) \(L \hspace{0.17em} \text{rail}_L (z) (e^{iz} - 1) = e^{ip_a} - 1\) at \(z = (q_{\ell})_a\), and \(L |z| \gtrsim 1 + | \ell_a |\) when \(\ell_a \ne 0\).
Lean: plank_height, rail_height, rail_lower, plank_lower, principal_plank, alias_rail.
Lemma 4.3 (histories). A history of stage index \(j\) is either empty, or a hop \((L, J)\) with \(j \leqslant J \leqslant j + 1\) and odd \(L \geqslant 3\), applied after a history of index \(J\). Its dilation \(M_v\) is the product of its block sizes. Its cumulative filter is defined by \(s_{\emptyset} = 1\) and \(s_{(L, J) \circ v} (p) = a_L (M_v p)^J s_v (p)\). Equivalently, \(s_v = a_{M_v}^{\hspace{0.17em} j} e_v\), where \(e_v\) is a product of box filters at the hops where the index drops. Then:
\(s_v\) is entire, \(2 \pi \mathbb{Z}^4\)-periodic and invariant under signed permutations;
on \(\mathcal{C}_{\pi / 4}\), \(|s_v (q_{\ell}) | \leqslant C_j \prod_a (1 + | \ell_a |)^{- j}\) at the aliases \(q_{\ell} = (p + 2 \pi \ell) / M_v\);
for \(p \in \mathcal{C}_{\pi / 4}^{\mathbb{R}}\), \(|s_v (p / M_v) | \geqslant c_j > 0\).
Here \(C_j, c_j\) do not depend on \(v\), that is, on the number of hops or the block sizes. Sketch: After \(k\) further hops, a hop of dilation \(M'\) is evaluated at the momentum scaled by \(M' / M_v \leqslant 3^{- k}\). Lemma 4.2 gives factors \(e^{\pm CM' / M_v}\), and these products converge. The factor \(a_{M_v}^j\) gives the alias decay by 4.2(d); the geometric product bounds the extra factors. Lean: Train, Train.smooth, Train.bind, Train.smooth_upper, Train.smooth_lower.
Lemma 4.4 (auxiliary 1-form covariance symbols). Fix \(\kappa > 0\) and a history \(v\) of dilation \(M\). For a direction \(a\), let \(\chi_{v, a} (p) = s_v (p) \hspace{0.17em} \text{rail}_M (p_a)\). For \(p \in \mathcal{C}_{\pi / 4}^{\mathbb{R}} \setminus \{0\}\) and \(q_{\ell} = (p + 2 \pi \ell) / M\), define
The second sum runs over the proper final segments \(v'\) of \(v\) (the last \(k\) hops, \(0 \leqslant k < \sharp v\)), with \(q'_{\ell'} = (p + 2 \pi \ell') / M_{v'}\). For the empty history, \(D_a = | \tau |^{- 2}\). The formula obeys the exact one-step recursion \(D'_a (p) = L^{- 2} \sum_{\ell} |a_L (q_{\ell})^j \text{rail}_L ((q_{\ell})_a) |^2 D_a (q_{\ell}) + \kappa^{- 1}\), which is the covariance recursion for an averaged field with independent noise. These singular symbols are used only away from zero; Lemma 4.6 constructs the regular 2-form symbols. Lean: Train.levelD, D_feedback, Train.levelNoise, feed. (definitions and covariance recursion proof read)
Lemma 4.5 (the diagonal). At the nonzero real momenta of Lemma 4.4, let \(\rho = \sum_a | \tau_a |^2 / D_a\) and \(D_{\mu \nu} = (\rho \hspace{0.17em} D_{\mu} D_{\nu})^{- 1}\); this is a multiplier diagonal in the plaquette orientation. Then, with \(F = \text{diag} (D_a)\),
So \(d^{\ast} Dd\) is the precision \(F^{- 1}\) minimised over gauge directions. For the empty history, \(D = 1\) (the Wilson quadratic form). Sketch: Lagrange's identity \(\sum_{\mu < \nu} w_{\mu} w_{\nu} | \tau_{\mu} x_{\nu} - \tau_{\nu} x_{\mu} |^2 = (\sum_a w_a | \tau_a |^2) (\sum_a w_a |x_a |^2) - | \sum_a w_a \bar{\tau}_a x_a |^2\), with \(w_a = D_a^{- 1}\). The weights and \(\rho\) are positive. Lean: rotor_law, pulse, vault, pink, blue, top_metric.
Lemma 4.6 (filled symbols, uniform analyticity). Fix \(\kappa > 0\). Away from \(p = 0\) in the real cube, the singular formulas become regular combinations. With \(| \tau_M |^2 = \sum_a M^2 |e^{ip_a / M} - 1|^2\):
Here \(\beta_a\) is the principal term \(| \chi_{v, a} (p / M) |^2\) plus \(| \tau_M |^2\) times the non-principal and noise terms, and \(h\) is a rational combination of filters whose denominators stay nonzero on a sufficiently thin cube. The regular formulas give \(\beta_a (0) = h (0) = 1\). For each \(j \geqslant 4\) there are \(\theta, C, c > 0\), depending only on \(\kappa, j\), such that, for all histories of index \(j\), \(\beta_a\) and \(h\) are analytic and bounded by \(C\) on \(\mathcal{C}_{\theta}\), and \(| \beta_a |, |h| \geqslant c\) on \(\mathcal{C}_{\theta}^{\mathbb{R}}\). Hence \(D\) and \(D^{- 1}\) lie in a common \(\mathcal{S} (\alpha, \Omega)\), uniformly over histories of index \(j\). Sketch:
Upper bounds: Lemma 4.3 summed over the aliases.
Lower bound on \(\beta_a\): its principal term (4.3 and 4.2(c)).
Lower bound on \(h\): away from zero use \(| \tau_M |^2 h = \sum_a | \tau_a |^2 / \beta_a\), \(\beta_a \leqslant C\) and \(| \tau |^2 \asymp | \tau_M |^2 \asymp |p|^2\); at zero use \(h (0) = 1\).
Inverses: a lower bound at real momenta plus an analytic bound gives a lower bound on a thinner complex cube, so the inverses stay analytic.
Lean: Train.bulb, Train.bend, Train.pole_split, Train.lift_bend, Train.recover_rho, Train.bulb_firm, Train.bend_firm, heart_banked, Moderate.
Lemma 4.7 (amended metric). Fix \(\kappa > 0\), \(j \geqslant 4\) and \(\lambda \geqslant 0\). For each history of index \(j\), let
Then on every torus: \(HH^{- 1} = 1\), \(Hd = Dd\), \(XH =\mathcal{T}_{\lambda} d_2\), and \(H, \mathcal{T}_{\lambda}\) are symmetric. Moreover \(\hat{H} (0) = 1\), and \(H, H^{- 1}, \mathcal{T}_{\lambda}, \mathcal{T}_{\lambda}^{- 1}\) have strip bounds and two-sided bounds at real momenta, with constants depending only on \(\kappa, j, \lambda\). The amendment changes \(H\) only away from exact 2-forms, and leaves \(d^{\ast} Hd = d^{\ast} Dd\) unchanged. Sketch:
\(XDd = d_2 d = 0\), which gives \(Hd = Dd\).
\((1 + \lambda Xd_2^{\ast}) X = d_2 H^{- 1}\), which gives \(XH =\mathcal{T}_{\lambda} d_2\).
Uniform bounds \(a \hspace{0.17em} 1 \leqslant D^{- 1} \leqslant A \hspace{0.17em} 1\) and \(\|X\| \leqslant B\) give \(a \hspace{0.17em} 1 \leqslant H^{- 1} \leqslant (A + \lambda B^2) \hspace{0.17em} 1\) and hence two-sided bounds for \(H\). The analogous argument applies to \(1 + \lambda d_2 D^{- 1} d_2^{\ast}\).
Lean: pledge, amend, bail, treble, obligations, pledge_warm, insurance, forced_amend.
Lemma 4.8 (metric data of every history; III.3.1(a)). Let \(\kappa > 0\), \(\lambda \geqslant 0\), \(j \geqslant 4\). There are \(\gamma, \sigma, \eta, C > 0\) with the following property. For every clip with \(c \leqslant \gamma\) and every history of index \(j\), the explicit \(D, D^{- 1}, H, H^{- 1}\), together with a positive symmetric square root \(R\) of \(H - c \hspace{0.17em} 1\), satisfy III.3.1(a) with amendment \(\lambda\) and bounds \(\sigma, \eta, C\). These constants depend only on \(\kappa, j, \lambda\). Sketch:
\(\gamma\) is half the real-momentum lower bound of \(H\), so \(H - c \hspace{0.17em} 1\) has two-sided bounds.
III.6's square-root theorem gives \(R \in \mathcal{S}\) with \(R^{\ast} R = H - c \hspace{0.17em} 1\) on every torus.
Torus bounds follow from symbol bounds, since \(\Lambda_n\) is diagonalised by the momenta \(2 \pi \mathbb{Z}^4 / n\).
The symmetry follows from Lemma 4.3.
\(D = 1\) for the empty history by Lemma 4.5.
Lean: bid_bed, Defined, full_cover, low_ceiling, sworn_metric.
A flight of index \((j, t)\) consists of an odd block size \(L \geqslant 3\), a fine history \(v\) of index \(j\), and the coarse history (hop \((L, j)\) after \(v\)) of index \(t \in \{j - 1, j\}\).
Lemma 4.9 (alias symbols and derivative bounds). Fix \(\kappa > 0\) and \(j, t \geqslant 4\). There are \(\theta, \alpha, C > 0\), depending only on \(\kappa, j, t\), such that the following holds for every flight of index \((j, t)\). Its explicit \(W\) and \(T\) are cross kernels; at this stage \(M_f = M_n = O = 0\). Their fibre at coarse momentum \(p\) and fine residue \(r\) is
with explicit per-alias symbols built from the filters and from \(\beta_a, h\) of both histories. At the harmonic point \(w_e (0, 0) = 1\) and \(w_e (0, \ell) = 0\) for \(\ell \ne 0\). Each \(w_e (\cdot, \ell)\) is analytic on \(\mathcal{C}_{\theta}\) and bounded by \(C \prod_a (1 + | \ell_a |)^{- j}\), uniformly over flights. Hence, for words \(u\) with \(|u| \leqslant j - 2\), \(\nabla_u W \in \mathcal{S} (\alpha, CL^{- 1 - |u|})\) and \(\nabla_u T \in \mathcal{S} (\alpha, CL^{- 2 - |u|})\). Sketch: A fine difference in direction \(a\) multiplies the alias term by \(e^{i (q_{\ell})_a} - 1 = O ((1 + | \ell |) / L)\). The sum \(\sum_{\ell} (1 + | \ell |)^{|u|} \prod_a (1 + | \ell_a |)^{- j}\) converges for \(|u| \leqslant j - 2\). In other words, the order \(j\) of the smoothing pays for \(j - 2\) derivatives. Decay is measured on the coarse lattice, so the fine-lattice rate is \(\alpha / L\). Lean: Flight.duo, Flight.pilot_cost, Flight.ahead_cost, Flight.wire_budget, secure_both, Secured.propels.
Lemma 4.10 (exact identities for one flight). Fix a flight of index \((j, t)\) with \(t \geqslant 4\), a kernel choice \(K\), and any pair of tori with \(f = cL\). Then (D1)–(D4) hold for:
\(W\), \(T\) of the flight and \(Q\) of the step;
\(D\), \(D'\) the diagonals of the fine and coarse histories;
\(H\), \(H'\) the amended metrics for any amendments \(\geqslant 0\).
Sketch:
All operators commute with coarse translations. On the Bloch fibre over \(p\), the identities become identities between per-alias symbols, for each alias \(q = q_{\ell}\):
\((\sigma_1^{\ast} D \sigma_1) (q) \hspace{0.17em} W_{\ell} = L^{- 3} h_1 (q)^{\ast} (\sigma_1^{\ast} D' \sigma_1) (p)\), giving (D1);
\(\sigma_1 (q)^{\ast} T_{\ell} = L^{- 3} h_1 (q)^{\ast} \sigma_1 (p)^{\ast}\), giving (D3);
\(T_{\ell} D' (p) \sigma_1 (p) = D (q) \sigma_1 (q) W_{\ell}\), giving (D4);
the alias sum \(\sum_{\ell} Lh_1 (q_{\ell}) W_{\ell} = 1 - \kappa^{- 1} (\sigma_1^{\ast} D' \sigma_1) (p)\), giving (D2).
Here \(W_{\ell} = L^{- 1} w_1 (p, \ell)\), \(T_{\ell} = L^{- 2} w_2 (p, \ell)\), and \(h_r (q) = a_L (q)^j \prod_{a \in \text{face}} \text{rail}_L (q_a)\) is the alias symbol of the \(r\)-form average. The forward block average has alias row \(Lh_1 (q)\), and its adjoint has alias column \(L^{- 3} h_1 (q)^{\ast}\) because the fine-to-coarse volume ratio is \(L^4\). The form averages intertwine the exterior derivatives: \(h_{r + 1} (q) \hspace{0.17em} L \sigma_r (q) = \sigma_r (p) \hspace{0.17em} h_r (q)\), from \(L \hspace{0.17em} \text{rail}_L (q) (e^{iq} - 1) = e^{iLq} - 1\).
The point \(p = 0\) is treated separately: the box filter vanishes at the non-principal aliases of \(0\).
The identities pass from the fibres to the tori through the alias-resolved Fourier transform.
Replacing \(D\) by the amended \(H\) in (D1)–(D2) changes nothing, since \(d^{\ast} Hd = d^{\ast} Dd\).
Lean: Flight.tangent, diagonal, Flight.scaled_tri, hinge_transport, ZeroAlias.lean, pilot, wire, script_main, script_cross, script_dual. (proof read, including the nonzero-alias algebra and harmonic point)
Lemma 4.11 (unamended defect). In the setting of 4.10, let \({\mathcal{V}= D'}^{- 1} - T^{\ast} D^{- 1} T - \kappa^{- 1} dd^{\ast}\) on \(\Omega^2 (\Lambda_c)\). Then \(D' \mathcal{V}D'\) is symmetric, and \(D' \mathcal{V}D' d = 0\) on every torus. Consequently there is an explicit symmetric 3-form kernel \(C\), decaying uniformly over flights of index \((j, t)\), with
Sketch:
By (D4), (D3), (D2): \(T^{\ast} D^{- 1} TD' d = T^{\ast} dW = dQW = d - \kappa^{- 1} dd^{\ast} D' d\). Hence \(\mathcal{V}D' d = 0\).
A symmetric decaying 2-form kernel \(A\) with \(Ad = 0\) on every torus equals \(d_2^{\ast} Bd_2\), for an explicit decaying 3-form kernel \(B\). This is a lattice Poincaré lemma for translation-invariant kernels, proved by an axis-by-axis contracting homotopy of the Koszul complex of \((e^{ip_a} - 1)_a\).
The hypothesis is needed on all tori, not on one: at a single period the harmonic (zero-momentum) forms obstruct the factorisation.
Lean: Flight.gap, Flight.dep_zero, Double.duplex, Flight.fund, Flight.banked_dollars. Quantitative control: duplex_stock.
Lemma 4.12 (amended defect: identity (Δ)). In the setting of Lemma 4.10, take fine amendment \(\mu \geqslant 0\) and coarse amendment \(\lambda \geqslant 0\). Then (Δ) holds on every pair of tori, with \(P =\mathcal{P}_{\mu, \lambda}\) where
Here \(S\) is the 3-form transfer kernel of the flight, with \(S \hspace{0.17em} d_2 {D'}^{- 1} = d_2 D^{- 1} T\), and \(S^{\ast} S\) has strip bounds \(O (L^{- 2})\). Sketch: Write \(X = d_2 D^{- 1}\) and \(X' = d_2 {D'}^{- 1}\). Then
Since \(XT = SX'\), the last term is \({- X'}^{\ast} (\mu S^{\ast} S) X'\); with Lemma 4.11 the whole expression is \({X'}^{\ast} (C + \lambda - \mu S^{\ast} S) X'\). Conjugate by \(H'\) and use \(X' H' =\mathcal{T}_{\lambda} d_2\) (Lemma 4.7). Lean: Flight.rebased, Flight.rollFunds, Flight.render, Flight.intake, Flight.sale, Flight.netted, check_bleep.
Lemma 4.13 (positivity; III.3.4). The constants and bounds of III.3.4 hold uniformly over flights of the fixed index \((j, t)\). Sketch:
The strip bound on \(C\) bounds its entries at real momenta, so \(C \geqslant - C_1\).
\(\mu S^{\ast} S \leqslant \mu B / L^2 \leqslant 1\).
Hence \(C + \lambda - \mu S^{\ast} S \geqslant 1\) for \(\lambda \geqslant C_0 = C_1 + 2\).
\(\mathcal{T}_{\lambda}^{- 1} = 1 + \lambda X' d_2^{\ast}\), with \(X' = d_2 {D'}^{- 1}\), has bounded entries, so \(|x|^2 \leqslant W |\mathcal{T}_{\lambda} x|^2\), and \(\hat{\mathcal{P}} \geqslant W^{- 1}\).
The upper bound and the strip bound follow from products of strip-bounded kernels.
Lean: Flight.absorb_before, Flight.frontier.
Lemma 4.14 (III.3.1). For every kernel choice \(K\), a catalogue satisfying III.3.1 exists. Sketch:
One hop. For each \((j, t)\) there is \(\lambda_0 (j, t)\) with the following property: for every coarse amendment \(\lambda \geqslant \lambda_0\) and fine amendment \(\mu \geqslant 0\), there are rates, constants and an integer \(b > \mu B (j, t)\) for which III.3.1(b) holds. The proof combines Lemma 4.13 with Lemma 4.9 at \(e = 1, 2\).
Assembly. Set \(\lambda_j = \max (\lambda_0 (j, j), \lambda_0 (j + 1, j))\): each stage's amendment acts as the coarse amendment of the hops \((j, j)\) and \((j + 1, j)\), and as the fine amendment of \((j, j)\) and \((j, j - 1)\). Then:
\(\alpha_j\) is the minimum of the rates of the bottom quote (Lemma 4.8 with \(\lambda_j\)) and of the hops \((j, j)\), \((j, j - 1)\);
\(\Omega_j\) is the corresponding maximum, and \(\eta_j\) the corresponding minimum;
\(b_j = \max (b (j, j), b (j, j - 1))\);
weaker constants are always admissible.
The block size enters only through \(b_j\), which guarantees \(\lambda_j B \leqslant L^2\). The stronger choice \(L \geqslant b_j > \lambda_j B (j, t)\) suffices for each admissible \(t\).
Lemma 4.15 (pre-repair family). Fix a catalogue of III.3.1 and a trajectory with its kernel, satisfying the credit condition. Define:
\(\mathcal{B}^{N, i}\): the metric data of Lemma 4.8 for the history of level \(i\) in mesh \(N\), with the clip of level \(i\) (allowed by (C1));
\(W^{N, i}, T^{N, i}\): those of the flight of step \(i\), of index \((J (i + 1), J (i))\) (allowed since \(J (i) \leqslant J (i + 1) \leqslant J (i) + 1\)).
This family satisfies the pre-repair conditions of III.3.2(2):
the bounds, from III.3.1;
(D1)–(D4), from Lemma 4.10;
(Δ), with \(P\) the rolled kernel;
the positivity of \(P\), from III.3.1(b) and (C2).
Lean: shelves, stageHop, location, destination.
Lemma 4.16 (repair; III.10 with III.6's theorems). For every sheet and every table \(X\) there is a table \(X'\) with the following property. Every pre-repair family for \(X\) (for a trajectory with the sheet's profile and kernel) has a repaired family with the same metric data and the same \(W, T\), which satisfies the output conditions for \(X'\). For step \(i\) the construction is:
F06 applied to \(P\) from (Δ): a kernel \(P_1\) with two-sided bounds at real momenta, a strip bound, and \(d_2^{\ast} P_1 d_2 = d_2^{\ast} Pd_2\) on every torus.
F07: a symmetric decaying square root, \(P_1 = P_2^{\ast} P_2\).
F08: kernels with the same \(W, T\) and
with the strip and derivative bounds of III.3.2(3). The constants depend only on \(X\), \(\kappa\) and the sheet's direct clip coefficients.
F09 (algebra) gives (R1)–(R6). Since \({V'}^{\ast} V' = H'\), \(\Pi\) is the orthogonal projection onto the range of \(V'\). Then:
(R1): \(M_f^{\ast} M_f + M_n^{\ast} M_n + O^{\ast} O = V' (T^{\ast} H^{- 1} T + \kappa^{- 1} dd^{\ast} {+ H'}^{- 1} d_2^{\ast} P_1 d_2 {H'}^{- 1}) {V'}^{\ast} + 1 - \Pi = 1\), by (Δ);
(R2) and (R4) follow from \({V'}^{\ast} V' = H'\);
(R3) follows from (D4) and \(H' d = D' d\), \(Hd = Dd\);
(R5) follows from \((1 - \Pi) V' = 0\) and \(d_2 d = 0\);
(R6) follows from (D3).
The second table depends on the sheet only through the direct clip coefficients, which enter \(V, V'\). It can be chosen with \(\alpha'_j \leqslant \alpha_j\), \(\Omega'_j \geqslant \Omega_j\) and unchanged \(\eta_j\). Lean: Bid.pinning, F06, F07, F08, F09, Replaces. (repair proof and F08–F09 proofs read)
Lemma 4.17 (III.3.2 and its variant). The catalogue and repair construction give the families and constants asserted in III.3.2, including its preliminary-sheet variant. Sketch: Take the catalogue (4.14), then the second table from the repair theorem with III.6's theorems (4.16), then the pre-repair family (4.15), then its repair. The explicit formulas and the symmetry carry over because the metric data and \(W, T\) are unchanged. In the variant, the payment record of the schedule turns sheet-level bounds into (C1)–(C2). A preliminary sheet is handled by re-running the construction for a copy of the trajectory with the preliminary profile; this preserves the histories and the direct clip coefficients. Lean: retained_support, retained_presents, before_paid, credit_from_sheet.
Lemma 4.18 (rigidity of decaying kernels). Suppose two finite sums of products of decaying kernels, or of cross kernels of the same block size, define equal matrices on tori \(\Lambda_{c_k}\) (respectively on pairs \((c_k, c_k L)\)) with \(c_k \to \infty\). Then they are equal on every torus (pair). Sketch: Sums and products of decaying kernels are decaying kernels, so each side is a single kernel. A summable kernel is the limit of its periodisations as the period tends to infinity. Dominated convergence justifies this coefficient limit; the schedule sides of each fixed step tend to infinity. Lean: entry_limit, Echo.all, Echo.allAcross, Echo.Sound, Echo.schedule_sides.
Lemma 4.19 (III.3.3). Every output family of III.3.2 satisfies (D1)–(D4) and (R1)–(R6) on every torus pair of the same block ratio. Sketch:
(R1)–(R5) are identities between finite sums of products of kernels, with \(M_f\) and \(W\) as cross kernels. Apply 4.18 on the schedule tori.
(R6) contains \(Q\). Given (D3) and (R2) at \((c, f)\), (R6) is equivalent to \((V \hspace{0.17em} d)^{\ast} M_f = d^{\ast} (T \hspace{0.17em} {V'}^{\ast})\), which is again an identity between kernels.
(D1)–(D4) at arbitrary \((c, f)\) follow by the same argument, because \(Q\) is itself given by a cross kernel (III.10). They also follow directly from Lemma 4.10 for any family realising the explicit formulas.
Lean: complete_roll, orthogonal_iff, panes, universal_diagonal.
Lemma 4.20 (lower bound). The coercivity bound of III.3.6(a) holds uniformly over firm kernels. Sketch:
\(\sigma_1^{\ast} \sigma_1 + \sigma_0 \sigma_0^{\ast} = | \tau |^2 \hspace{0.17em} 1\) on 1-form symbols; this is \(d^{\ast} d + dd^{\ast} = - \Delta\) in Fourier space.
A firm \(f\) satisfies \(\text{Re} \hspace{0.17em} \langle x, \hat{f} (p) x \rangle \geqslant \eta |x|^2\) at every real \(p\): the torus bounds give it at the frequencies \(2 \pi \mathbb{Z}^4 / n\), and these are dense; conclude by continuity.
Hence \(\text{Re} \hspace{0.17em} \langle x, \Phi_f (p) x \rangle \geqslant \min (\eta, 1) | \tau (p) |^2 |x|^2\).
\(| \tau (p) |^2 \geqslant c_0 |p|^2\) for \(|p_a | \leqslant 3 \pi / 2\), since \(e^{ir} \ne 1\) for \(0 < |r| \leqslant 3 \pi / 2\) (compactness).
Lean: blade_id, ion, real_price, tank_full.
Lemma 4.21 (smooth symbols). This is III.3.6(c). Sketch: Differentiate the Fourier series termwise and use \(|y|_1^n \leqslant A_n e^{\alpha |y|_1}\), with \(A_n\) as in III.3.6(c). Lean: radial_prices.
Lemma 4.22 (comparison with the continuum). This is III.3.6(b). Let \(S^k (U)\) be the class of families with \(\|D^n g\| \leqslant C_n |p|^{k - n}\) on \(U\), uniformly over the family. Sums preserve a fixed order, products add orders, and inversion reverses the order when the rescaled matrix has a uniform coercivity bound. Sketch:
\(\tau - ip \in S^2\) (Taylor at \(0\)), and \(\hat{f} - 1 \in S^1\) (since \(\hat{f} (0) = 1\), and by Lemma 4.21).
Hence \(\Phi_f \in S^2\) and \(\Phi_f - |p|^2 \hspace{0.17em} 1 \in S^3\).
By (a) and the resolvent identity, \(\Phi_f^{- 1} \in S^{- 2}\) and \(\Phi_f^{- 1} - |p|^{- 2} \hspace{0.17em} 1 \in S^{- 1}\).
Hence \(G_f \in S^0\) and \(G_f - G_{\infty} \in S^1\).
Lean: gem_quote, Weight, Value. (proof read, including the symbol-class product and inverse estimates)
Lemma 4.23 (nonlinear calculus). The statements of III.3.7 hold under their respective hypotheses. Sketch:
(a) Split at \(|x|_1 = L\).
For \(|x|_1 \leqslant L\): localised Newton remainder estimate, \(\|p (x) - \sum_{i < k} P_i (x)\| \leqslant M' (1 + |x|_1)^k\) with \(M'\) the bound on \(k\)-th differences.
For \(|x|_1 > L\): bound \(p\) and the pieces directly, using \(L^{- 1 - i} (1 + |x|)^i \leqslant L^{- 1 - k} (1 + |x|)^b\) when \(L^{- 1} (1 + |x|) \geqslant 1\).
(b) Substitute \(z \mapsto \sum_v t^{Y_v} p_v z\) in the Taylor polynomial of \(f\) and read off each coefficient of \(t^r\) for \(r \leqslant d\). Rescaling \(z\) by an independent scalar separates the ordinary Taylor degrees, and polarization recovers the multilinear coefficients. All weighted components of weight \(\leqslant d\) of every Taylor coefficient vanish. A component of weight \(> d\) vanishes on the given inputs by the support condition.
(c) Locality restricts the logarithmic link fields to links with both endpoints in \(X\). Gauge invariance identifies \(u\) near the identity with its pullback by the map taking a site gauge and a radial connection jet to gauged logarithmic holonomies. Part (b) gives weighted vanishing for this pullback. The derivative of that map covers every polynomial link field of degree at most \(p - 1\): part (d) supplies the highest degree, line integration changes it only by lower degrees, and induction removes the error. Its higher derivatives preserve the total weight. Faà di Bruno and induction on the number of inputs therefore transfer vanishing to \(u \circ \exp\). Non-wrapping and \(n \geqslant 2\) ensure that the centred lifts of adjacent sites differ by a unit vector.
(d) \(r = (j + 2)^{- 1} \sum_a x_a Q_a\) is homogeneous of degree \(j + 2\), and Euler's identity gives \(\sum_a x_a (Q_a - \partial_a r) = 0\).
(e) On the unit ball, \(\|g (sv) - g (0)\| \leqslant s\|g (v) - g (0) \|+ Q (1 - s) \|v\|^2\) for the matrix \(g\) of the plaquette holonomy of four link variables (second-order Taylor bound). Hence the chord distance at \(sv\) is at most \(s \tau + (1 - s) \tau\).
Lean: see III.3.7; Maps, covers, attenuation, testLinks. (proof read for (a)–(e), including the moment-space definition and transfer argument)
Three auxiliary packages concern local observables and equation spaces:
Corrected observables. An exact identity recovering holonomy-jet probes from the coefficients of corrected observables (Screen.reproduces). Used by the rank-covariance part of III.11, by VII.2 and by VII.4.
Equation projections. Orthogonal projections onto a finite-dimensional equation subspace and its complement (drop, retain). The additional criterion spectral_bank gives spectral separation under construction and comparison hypotheses, exact top-jet tests, and approximation by a fixed triangular model on finitely many probes.
Local jet vanishing. Vanishing of test-jet derivatives below the engineering degree for non-unit basis observables, assuming their local section is readable (Speeds.lean, canon_first). Used in the nonlinear step of III.10.
(statements read; the coefficient identity and jet-vanishing proofs read)
Uniformity. The algebraic identities must be accompanied by bounds uniform over histories of any length and over all block sizes, with the kernel choice and stage index fixed. The mechanism is in Lemmas 4.2–4.3:
geometric dilation along the history;
a lower bound on the principal alias, with an exponent linear in \(M / L\);
\(J\)-fold smoothing, which gives \(\prod_a (1 + | \ell_a |)^{- J}\) alias decay and therefore \(J - 2\) derivatives on \(W, T\).
Positivity is proved only at real momenta. \(| \tau |^{- 2}\) is never estimated on a strip. Every symbol is put in a filled form (Lemma 4.6) whose denominators are bounded below at real momenta, and inverses are taken on a thinner complex cube. The estimates use the Banach algebra of exponentially decaying kernels and its analytic symbols.
Unusual features.
The amendment \(\lambda_j\) (Lemma 4.7) changes the inverse metric. It modifies the metric only off exact 2-forms, so \(d^{\ast} Hd\) is unchanged, and it exists only to make the defect's 3-form factor positive.
\(L\) enters the catalogue only through \(\lambda_j B \leqslant L^2\).
The second table depends on the sheet through its direct clip coefficients.
The factorisation of the defect (Lemma 4.11) needs the identities on all tori, because of the harmonic forms.
For explicit families, universal_diagonal gives the diagonal identities directly. The output conditions alone give them by rigidity.
Units. Decay is in the \(\ell^1\) norm, on the lattice where the kernel acts. For \(W, T, M_f\) that is the coarse lattice, so the fine-lattice rate is \(\alpha / L\), with amplitudes \(L^{- 1}\), \(L^{- 2}\) and an extra \(L^{- 1}\) per fine difference. These are the scalings of Balaban's background field: smooth on scale \(L\).
Statement strength. III.3.2 is an existence statement whose output is a package of bounds and identities. Later units use (R1)–(R6), the floors and the strip bounds. The repaired family retains the pre-repair metric data and \(W, T\), so the defect identity and positivity remain available, although they are not fields of the output conditions.
| Written proof | Lean (glossary) | Meaning |
| \(R\), \(\hat{R}\), \(R_n\); \(\mathcal{S} (\alpha, \Omega)\) | Row, Row.strip | translation-invariant kernel, symbol, action on \(\Lambda_n\); strip bound |
| cross kernel | Cross, installed | coarse-to-fine kernel and its action |
| \(\text{plank}_M, \text{rail}_M, a_L, a_L^J\) | plank, rail, area, massFilter | line, box and \(J\)-fold box filters |
| history \(v\), dilation \(M_v\), filter \(s_v\) | Train, Train.smooth | finite history of block averages |
| flight of index \((j, t)\) | Flight, Flight.land | one step: block size, fine history, coarse history |
| \(D_a\); \(\rho\); \(\beta_a\), \(h\) | Train.levelD, D_feedback; Train.rho; Train.bulb, Train.bend | 1-form covariance; gauge factor; filled symbols |
| \(D\), \(D^{- 1}\) | pink, blue | diagonal 2-form multiplier, inverse |
| \(H^{- 1}\), \(H\); \(\mathcal{T}_{\lambda}\); \(\lambda\) | pledge, amend; treble | amended inverse metric, metric; 3-form factor; amendment |
| metric data \(\mathcal{B}\), \(V\) | Bed, Bed.full, Floor | \((D, D^{- 1}, H, H^{- 1}, R)\); \(V = (\sqrt{c} \hspace{0.17em} 1, R)\) |
| transfer kernels \((W, T, M_f, M_n, O)\) | Towers (A0, T, mf, mn, o), Flight.duo | background-field map, curvature transport, isometry blocks |
| (D1)–(D4); (R1)–(R6); (Δ) | Towers.Diagonal, Inputs; Towers.Roll; Bleeps | identities of one step |
| \(C\), \(S\), \(\mathcal{P}_{\mu, \lambda}\) | Flight.fund, Flight.sale, Flight.rollFunds | unamended 3-form defect, 3-form transport, rolled 3-form kernel |
| derivative bounds | Donate, Towers.Deep | \(\nabla_u\) bounds with \(L^{- 1 - \lvert u \rvert}\), \(L^{- 2 - \lvert u \rvert}\) |
| positive on \(\text{ran} \hspace{0.17em} \sigma_2\); two-sided at real momenta | Reaches, Funds | positivity conditions on 3-form kernels |
| kernel family; pre-repair; output conditions | Horizon; Craft.Rack; Horizon.Suffices | metric data and transfer kernels for all \((N, i)\) |
| explicit formulas; credit condition | Present; Credit | (III.3.2(1)); (C1)–(C2) |
| catalogue; price table | Catalog; MapPrices | constants fixed before \(L\) |
| hypercubic symmetry of \(d^{\ast} Hd\) | Bid.Sworn | invariance under signed permutations |
| firm kernel; \(\Phi_f\); \(G_f\); \(G_{\infty}\) | Firm; fuel; garnet; pearl | curvature covariance, lattice and continuum |
Gaussian interpretation and zero modes. Lemma 4.4 verifies the covariance recursion away from zero momentum. A measure-level derivation requires a stated treatment of gauge directions and harmonic modes; neither the auxiliary singular symbols nor III.1's action alone specifies that Gaussian measure. The identification with an iterated Feynman-gauge covariance remains probable.
Change of Gaussian variables. The quadratic splitting in §1 follows from the identities. Its use as a change of integration variables, including null directions and determinants, belongs to III.4 and III.10 and is not established by this unit alone.
Auxiliary spectral criterion. The finite-test hypotheses of the cited cutoff-projector criterion must be supplied at its point of use. Its statement does not assert that a trajectory satisfies them.
Further proof detail. The analytic coefficient-extraction machinery underlying the strip bounds and the deeper positivity and square-root proofs F06–F07 are inputs here; their full proofs remain outside this account. The symbol-class calculations in Gem.lean and Weight.lean support the covariance comparison, but no position-space continuum limit or contact-term statement is asserted in III.3.6.