Section 13.3: Bessel rows and the linear-side energy #
This file was automatically generated by Claude.
Formalization of blueprint Section 13, "Fixed modes, the linear side, and hereditary
representatives" (sec:modes), subsection 3, "Bessel rows and the linear-side energy" (lines
2046-2122) of blueprint/formalization_blueprint.tex.
Separated mode rows (lem:mode-rows, blueprint lines 2049-2059) #
Blueprint lines 2054-2055: a quadratic phase difference whose derivative 2αx + β exceeds
θ at one point of a cell [a, a + r], with slope variation 2|α| r ≤ θ/2, has normalized
oscillatory integral at most 8/(θ r) on the cell.
Blueprint line 2058: two finite mode combinations with pairwise normalized inner products at
most ε on an overlap of length L pair to at most ε L √|M| √|M'| b b'.
Symmetric overlap and diagonal bounds control the absolute synthesis energy error. The upper and lower almost-orthogonality estimates are its two order consequences.
Blueprint lines 2057-2058: almost-orthogonal synthesis. If |⟨f_e, f_f⟩| ≤ η w_{ef} b_e b_f
for e ≠ f, with symmetric overlaps w of multiplicity Σ_f w_{ef} ≤ D ℓ_e and diagonal control
ℓ_e b_e² ≤ D² ‖f_e‖², then ‖Σ f_e‖² ≤ (1 + η D³) Σ ‖f_e‖².
Blueprint line 2057 ("by duality"): row outputs g_e = P_e F of orthogonal projections satisfy
Re⟨g_e, F⟩ = ‖g_e‖², so a synthesis bound ‖Σ g_e‖² ≤ C Σ ‖g_e‖² gives
Σ ‖g_e‖² ≤ C ‖F‖².
Blueprint lem:mode-rows (lines 2049-2059), assembled: projection-row outputs g_e = P_e F
on cells of lengths ℓ_e with coefficient norms b_e, with cross pairings at most
η w_{ef} b_e b_f (η = D_*^{-199} from cell_cross_pair_le, fourier_sum_tail_le,
mode_combination_cross_le), overlap multiplicity D_* and private-coordinate diagonal control,
satisfy Σ ‖P_e F‖² ≤ (1 + η D_*³) ‖F‖².
Composition with the main histories (lem:lift-composition, blueprint lines 2062-2077) #
The linear-side energy (prop:linear-side, blueprint lines 2080-2120) #
Blueprint lines 2093-2094 (elementary fact): if nonempty sets S_j increase from S_0 = ∅,
X_j ⊆ S_j, and each label belongs to at most A of the X_j, then
Σ_j |X_j|/|S_j| ≤ A H_{|S_n|}; enumerate labels by first entry.
Corollary: Σ_j |X_j|/|S_j| ≤ A (1 + log M) when |S_n| ≤ M.
Blueprint lines 2054-2056: a quadratic oscillation e(αy² + βy + γ) against an amplitude Z
that is, on the cell [a, a + r], a finite Fourier sum Σ c_k e(ν_k y) up to τ: if every
shifted derivative 2αx + β + ν_k exceeds Θ at some point of the cell and the slope varies by
at most Θ/2, the normalized pairing is at most (Σ |c_k|) 8/(Θ r) + τ.
Blueprint lines 2054-2056 with the full Fourier norm of the squared gauge ("expand (ζ^ρ)²
with coefficient sum at most D_* and tail D_*^{-300}; the Fourier sum and tail give
C D_*/Ξ + D_*^{-300}"): under the hypotheses of modeWave_pair_le, the normalized pairing is at
most gaugeSqNorm · 8/(Θ r) + 2 B_ζ/2^K.
Separated mode rows from mode Gram bounds (lem:mode-rows, blueprint lines 2049-2059) #
Blueprint lem:mode-rows from mode-level data: rows e project onto the spans of their
augmented modes Ψ e m, m ∈ M_e. If modes of distinct rows have pairings at most η w_{ef}
(cross pairs on overlapping cells), the private coordinates give ε ℓ_e Σ|β_m|² ≤ ‖Σ β_m Ψ e m‖²,
the overlaps have multiplicity Σ_f w_{ef} ≤ D ℓ_e, and |M_e| ≤ ε D², then
Σ_e ‖P_e F‖² ≤ (1 + η D³) ‖F‖².
Pointwise pairing of two augmented modes: physical product plus private overlap.
Blueprint lem:mode-rows for augmented modes 1_{J_e}(ψ_m e_phys + √ε_m e_m): rows e on
cells J_e with mode sets M_e that share no mode on overlapping cells, physical cross pairings
|∫_{J_e ∩ J_f} ψBar_m ψ_{m'}| ≤ η |J_e ∩ J_f|, overlap multiplicity D, private weights at least
ε and |M_e| ≤ ε D² give Σ_e ‖P_e F‖² ≤ (1 + η D³) ‖F‖².
Bad incidences along a linked chain of side events (eq:bad-density, lines 2088-2102) #
Blueprint lines 2088-2099: along a linked chain of N events, with predecessor and component
label sets P_{i,h} ⊆ C_{i,h} ⊆ P_{i+1,h} increasing in the height and of size at most M, and
bad-primary sets X^e_{i,h} (exceptional; each label in at most A_e events at a height) and
X^h_{i,h} (height-bad; each label at most A_h heights at an event), the number of bad
incidences — h < Δ, a failed predecessor-density test 10|P_{i,h-Δ}| < 9|C_{i,h}|, or more than
a tenth of bad primaries — is at most
NΔ + 10(NΔ log M + (J+1) log M) + 10((J+1)A_e + N A_h)(1 + log M).
Longest-chain levels (blueprint line 2104) #
A linked chain with k links in s ending at e.
Equations
Instances For
The longest-chain level of an event, among chains with at most L links.
Equations
- TrilinearHT.Auto.chainLevel s R L e = Nat.findGreatest (fun (x : ℕ) => TrilinearHT.Auto.LinkedChain s R x e) L
Instances For
Linked events have strictly increasing levels when all chains have at most L links; hence
events of one level are never linked.
The pairing constant η = ‖(ζ^ρ)²‖_A · 8/Ξ + 2 B_ζ/2^K of modeWave_pair_le_full at threshold
Ξ / r on a cell of length r (blueprint line 2055: C D_*/Ξ + D_*^{-300}).
Equations
- TrilinearHT.Auto.pairConst d δ Ξ Kf = TrilinearHT.Auto.gaugeSqNorm d δ * (8 / Ξ) + 2 * TrilinearHT.Auto.gaugeTail d δ / 2 ^ Kf
Instances For
Blueprint lem:mode-rows for the modes ψ_m of one role, group and offset (eq:modes):
rows e on cells J_e with mode sets M_e sharing no mode on overlapping cells, such that on the
common (smaller) cell [a', a' + r') every pair of their modes is non-adjacent at height zero with
threshold Ξ/r' and has slope variation at most Ξ/(2r'), and with point multiplicity D
(Σ_f |J_e ∩ J_f| ≤ D |J_e|), satisfy Σ_e ‖P_e F‖² ≤ (1 + η D³) ‖F‖² with
η = pairConst.
Exact subtotals versus row projections (blueprint lines 2112-2116) #
Blueprint lines 2112-2113: the projection onto the span of the modes Ψ m, m ∈ E, of a
combination of complementary modes w = Σ_{m' ∈ A} β_{m'} Ψ m' with pairings at most η L
is at most η L √|E| √|A| ‖β‖ / √(ε ℓ), when the private coordinates give
ε ℓ Σ|c|² ≤ ‖Σ c_m Ψ m‖² on E.
Shared coefficient sizes give cross-pairing and diagonal bounds for finite subtotals.
Blueprint line 2116: on one cell the subtotals T_j = Σ_{m ∈ E_j} β_m Ψ m over disjoint
components with mutual pairings at most η L are almost orthogonal:
‖Σ_j T_j‖² ≤ (1 + η D³) Σ_j ‖T_j‖² when the number of components is at most D, the private
coordinates give ε L Σ|β|² ≤ ‖T_j‖² and |E_j| ≤ ε D².
Locality of mode rows (blueprint lines 2068-2069) #
The mode lift is pointwise, hence commutes with restrictions.
Blueprint lines 2084-2086: CURRENT components on an earlier pre-cell are contained in PRE
components on a later pre-cell inside it, once the later PRE threshold Ξ/r_v' is at least the
earlier CURRENT threshold Ξ/r_t (block starts at least 3K apart). Modes are transported by
f (inclusion of the earlier frozen list into the later one) with unchanged slopes.
Subtotals of a synthesized output (blueprint lines 2112-2118) #
Blueprint lines 2112-2113: the row projection of a synthesized output Σ_{m ∈ A} c_m Ψ m
onto the modes E ⊆ A differs from the exact subtotal Σ_{m ∈ E} c_m Ψ m by at most the
complementary pairing bound.
The L² pairing of two augmented modes.
Distinct augmented modes on one cell pair through their physical parts only.
The private coordinates of a combination of augmented modes on one cell.
Blueprint lines 2116-2118 for one block: the side subtotal V = Σ_J Σ_E T_{E,J} over
disjoint retained pre-cells J and, on each, disjoint minimal PRE components E with mutual
physical pairings at most η |J|, satisfies ‖V‖² ≤ (1 + η D³) Σ_{J,E} ‖T_{E,J}‖².
Joint Bessel bounds for rows with group-dependent targets (lines 2068-2072) #
Blueprint lines 2068-2072: rows R_{G,k} : H → F_G that are Bessel with constant c within
each group G, and whose adjoint outputs S_G = Σ_k R_{G,k}^* R_{G,k} h satisfy the whole-range
almost orthogonality ‖Σ_G S_G‖² ≤ C Σ_G ‖S_G‖², have joint squared sum at most C c ‖h‖².
Lift rows over all groups of one color (blueprint lines 2066-2072) #
Catalog projections commute with the restriction to their cell #
Physical transfer between column spaces and the basic histories (lines 2072-2075) #
Side events, linking, and the rows of one level (blueprint lines 2082-2110) #
Non-adjacency at height zero gives the separation hypothesis of the mode rows.
The PRE edge at height zero on a cell J between two modes.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Blueprint line 2084: an earlier event is linked to a later one on a descendant pre-cell when some of their modes agree or are adjacent in the later PRE graph at height zero.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Two meeting dyadic cells are nested, the deeper one inside.
A pairwise disjoint family of nonempty subsets of a finset T has at most |T| members.
The modes whose first-match cells contain the point y: the modes on the path through y
(blueprint lines 2114-2116: "label sets disjoint along paths with total count at most D_*").
Equations
Instances For
The cell of an event as a set.
Equations
- Ms.evCell e = TrilinearHT.Auto.dyadicGridCell Gd e.2.1.1 e.2.1.2
Instances For
The augmented mode of m on the cell of the event e.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Overlapping events of one level are comparable: the deeper cell is inside the other, the mode sets are disjoint and the modes are not adjacent at height zero on the smaller cell.
The overlap sum of a family of sets is at most the pointwise multiplicity bound times the measure of one of them.
Blueprint lines 2104-2108: the side events of one level (pairwise unlinked, distinct
components on one pre-cell non-adjacent at height zero) give Bessel mode rows
(lem:mode-rows), with overlap multiplicity at most the number of modes.
The length of linked chains of side events (eq:bad-density, lines 2088-2102) #
Blueprint lines 2088-2099 on a linked chain of n side events with component structures
C_i and representative modes x_i: if the CURRENT label sets feed into the next PRE label sets,
every key is exceptional in at most A_e events and height-bad at most A_h times at each event,
and every event has more than δ (J + 1) bad heights, then
δ n (J+1) ≤ nΔ + 10(nΔ log M + (J+1) log M) + 10((J+1) A_e + n A_h)(1 + log M).
The goodness of a key at height h in the block b (nonexceptional, stable spans).
Equations
- One or more equations did not get rendered due to their size.
Instances For
The side modes of the block b on the pre-cell J: those that are not core.
Equations
Instances For
The side events of the block b on the pre-cell J: the minimal PRE components of its
side modes.
Equations
- One or more equations did not get rendered due to their size.
Instances For
CURRENT components of an earlier pre-cell transport into PRE components of a later one.
Blueprint lines 2084-2088: along a link from an event of block b on J to an event of a
block b' ≥ b + 3K on J' ⊆ J, the CURRENT labels of the earlier representative at any height
are PRE labels of the later representative.
The gap condition (nonexceptionality) of a key at the block b.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Blueprint lines 2088-2102 for the side events of one class: a linked chain with k links of
side events whose blocks are at least 3K apart has δ (k+1)(J+1) at most the bad-incidence
bound, where the keys of the modes alive at any cell number at most M (the label sets of the
chain are path-local), every key is exceptional in at most A_e of the blocks of the events, and
every nonexceptional key is height-bad at most A_h times in one block.
Distinct minimal PRE components on one pre-cell are disjoint and not adjacent at height zero (blueprint line 2104: "distinct components on the same cell have this separation").
Blueprint lines 2104-2108 for one class of blocks: the side events of the class, split into
longest-chain levels (at most L + 1 of them when chains have at most L links), give Bessel
mode rows with constant (L + 1)(1 + η D³).
Blueprint lines 2054-2056 on one pre-cell: two distinct modes alive on J that are not
adjacent at height zero there have augmented pairing at most η |J|.
Modes of a side event are not adjacent at height zero to the other alive modes of the pre-cell.
Blueprint lines 2112-2113 for a side event: the row projection of the frozen evaluation
Σ_{m alive} c_m Ψ_m on the event's modes differs from the exact side subtotal by at most the
complementary pairing bound.
The exact side subtotal of the block b: the side events on the retained pre-cells, with
coefficients c J m of the frozen evaluation (blueprint line 2081, V_{l,G,b}).
Equations
- Ms.blockSide S hA SP b cells c = ∑ J ∈ cells, ∑ E ∈ Ms.sideEvents S hA SP b J, ∑ m ∈ E, c J m • Ms.cellMode J m
Instances For
Blueprint lines 2116-2118 for one block: the side subtotal over disjoint retained pre-cells is almost orthogonal over its events.
The frozen evaluation of the block b on the pre-cell J: the synthesis of the modes alive on
its decision cell with coefficients c (blueprint line 1959, ZTilde_{l,G,b} on J).
Equations
- Ms.frozenEval S b J c = ∑ m ∈ Ms.alive (TrilinearHT.Auto.ancAt Gd J (b - 4 * ↑S.K)), c m • Ms.cellMode J m
Instances For
The complementary pairing error of a side event (event_subtotal_close).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Blueprint lines 2104-2118 for one group: the exact side subtotals of its blocks have squared norms bounded by twice the almost-orthogonality constant times the squared row outputs of their side events plus the squared complementary errors.
The global ball outputs and histories of an attempt (blueprint lines 1857-1890) #
The event rows of all groups composed with the attempt histories (lines 2062-2108) #
The side events (b, J, E) of the blocks Bs on their retained pre-cells.
Equations
Instances For
Blueprint eq:bad-density (line 2102), the absorption step: the chain inequality with
δ (J+1) at least twice the per-link cost gives k + 1 ≤ 20 (log M + A_e (1 + log M)) / δ.
eq:bad-density at the chosen schedule (blueprint lines 2100-2112) #
A key that is not nonexceptional at a block (with the side parameters of the schedule) is
exceptional there in the sense of eq:gap (ExceptionalAt with r₀ the base length of the
grid).
A nonexceptional key is not exceptional in the sense of eq:gap.
Blueprint eq:bad-density (lines 2100-2112) at the chosen schedule: a linked chain of side
events of one class (blocks on the K-grid through R₀, at least 3K apart) has length at most
180 d_max log(2 D_m) / δ_*, free of the height parameter. The label sets are path-local (the keys
of the modes alive at any cell number at most M ≤ 2 D_m); a nonexceptional key is height-bad at
most 3 d_max R_*^{ht} times at one event (lem:stabilize), a key is exceptional in at most
4 d_max blocks of the class (line 1996), and the choice J_* ≥ C_J δ_*^{-2} d_max R_*^{ht} log(2 D_m) with C_J ≥ 340 absorbs the first term.
Blueprint line 2118 ("cross pairings are at most D_*^{-200} ... these squared additive
errors are negligible"): the squared complementary pairing error of a side event is
pairConst² |J| |E| |alive \ E| Σ |c|² / ε, at most pairConst² |J| D² B / ε when the alive
modes number at most D and the coefficients have square sum at most B.
Restricting a sum of physical embeddings costs twice the restricted first energy and twice the full second energy. This separates the rolling and fixed parts of a side state.
Per-group clearing integers (blueprint line 1890) #
The schedule with the clearing integer Q₀ replaced by q (blueprint line 1890: Q₀ is
chosen for each group center; eq:mode-parameters only uses an upper bound for it).
Equations
Instances For
eq:mode-parameters persists when the clearing integer is replaced by a smaller one: Q₀
enters only as a factor of the lower bound for H.
Linear sides from bounded frozen event projections (blueprint lines 2112-2118, 2160) #
Global filters of the transfer steps (blueprint lines 2386-2391) #
Variation of the global positive chains (blueprint lines 2390-2391, lem:variation) #
Lift composition on the main columns (blueprint lem:lift-composition, lines 2062-2077) #
The isometric embedding of one summand into the ℓ² sum of a finite family of inner product
spaces.
Equations
- TrilinearHT.Auto.piLpSingle F G = (↑(WithLp.linearEquiv 2 ℂ ((G : ι) → F G)).symm ∘ₗ LinearMap.single ℂ F G).mkContinuous 1 ⋯
Instances For
At an active cell the cell restriction after the ball depth projection is the cell projection of the main tree.
The outer ball filter at the depth of an active cell commutes with the restriction to the cell.
Blueprint lines 2068-2069 (lem:compression with one factor) for one group: rows Q_e on the
relevant main columns, Bessel with constant c and local to active decision cells u_e, stay
Bessel with constant c C₁² after inserting the outer averaged ball projection of the group at
the depth of each decision cell.
Blueprint lines 2070-2071 ("every term of A_G^* has outer factor L_{l,G,u_e}, so its range
is a sum of whole usable main endpoint ranges"): for a row Q local to an active cell u, the
adjoint of Q L_{l,G,u} takes values, after relabeling, in the main ball range of u at the
largest pad.
Blueprint lem:lift-composition, lines 2066-2071, on the main columns: rows Q_{G,e} on the
relevant main columns, Bessel with constant c within each group of one color k and local to
active decision cells u_{G,e} with downstream use certificates of G, stay jointly Bessel over
all groups after inserting the outer averaged ball projections L_{l,G,u_{G,e}} at their decision
cells. Per group this is lem:compression with one factor (cost C₁², main_rows_group);
across the groups the adjoint outputs are sums of whole main ball ranges at the decision cells
(gBall_row_adjoint_mem), and eq:whole-ranges gives the factor 1 + τ N colBudget².
Blueprint lines 2072-2075 ("apply compression again, with decision cells u_e, to insert those
histories and the constituent basic projections acting on F_l"): rows Q_{G,e} local to active
decision cells that are jointly Bessel with constant B after the outer ball projections stay
jointly Bessel after inserting the shared history and the basic depth projection of the family
j_b, at the cost C_{|ls|}² C_{1,j_b}² of basic_compression.
Rows local to an active cell see the input of a basic family only through its restriction to
the attempt root (blueprint line 2075: ‖F_l‖₂² ≤ |R|).
The mode rows after the lift: Q_{G,e} = P_{G,e} R_{l,G} on the relevant main columns are
Bessel with constant 5c within each group (the lift has norm at most √5).
The mode rows after the lift stay local to a cell when the mode rows are.
Blueprint lem:lift-composition (lines 2062-2077) on the main columns of a basic slot l, for
the groups of one color k (with the ball level j and the shared history ls of a leaf of the
color search), with the global coarse outputs: let L_G be the mode lift R_{l,G} of the group
G, and let the mode projection rows P_{G,e} be Bessel with constant c within each group and
local to active decision cells u_{G,e} with downstream use certificates. Then the outputs
P_{G,e} R_{l,G} Z_{l,G,u_{G,e}} of the global coarse outputs at the depths of their decision cells
have joint squared sum at most
2 (1 + τ N colBudget²) 5c C₁² C_{|ls|}² (C_{1,0}² + C_{1,1}²) |R|.
The constant of lift_composition_global is at most c G₀ with
G₀ = exp(C_q (1 + log 2k)²), C_q = 6144 (c+1) (1024 (c+1) + c + 7) + 4, under
log (2 D_m) ≤ k^c, log (2 colBudget) ≤ k^c, shared histories of O(log k) factors and the
pair threshold τ N colBudget² ≤ 1 (blueprint lines 2072-2077: "the loss is G₀").
Level rows of the side events at the chosen schedule (blueprint lines 2114-2116) #
Blueprint line 2055 at the chosen schedule ("the Fourier sum and tail give
C D_*/Ξ + D_*^{-300} ≤ D_*^{-200}"): when D_* covers the full Fourier norm of the squared gauge
and the truncation leaves a tail at most D_*^{-300}, the pairing constant is at most
D_*^{-200} (with Ξ = D_*^{1000}).
Blueprint lines 2114-2116 at the chosen schedule, for one group: give the side events their
longest-chain level within each of the three classes of blocks; events of one level on overlapping
cells have neither shared nor adjacent modes, and their mode sets are disjoint along paths with
at most D_* modes through each point. Lemma mode-rows (cross pairings at most D_*^{-200} by
ModeSchedule.pairConst_mul_le, inverse private weights at most D_*) gives Bessel rows for each
level, and the chains have length at most 180 d_max log(2 D_m)/δ_* (eq:bad-density), so the
side-event rows of the group are Bessel with constant 6 (⌊180 d_max log(2 D_m)/δ_*⌋ + 1).
Exact side subtotals and their errors (blueprint line 2118) #
Blueprint eq:path in counting form, for a family of families of sets: sets inside R, of
which at most N pass through each point of R, have total length at most N |R|.
Blueprint line 2118 ("the number of events along a path in one level is at most D_*"): in
one class whose chains have at most L links, the side events through a point y number at most
(L + 1) |ptSet y|, since the events of one longest-chain level through y have pairwise
disjoint sets of modes through y.
Blueprint line 2118 at the chosen schedule, for one group: the side events through a point y
number at most 3 (⌊180 d_max log(2 D_m)/δ_*⌋ + 1) |ptSet y| (three classes, longest-chain levels
bounded by eq:bad-density, disjoint mode sets through y on each level).
Blueprint line 2118 at the chosen schedule ("pair them against a unit vector of the augmented
range ... hence ‖P_{E,J} ZTilde - T_{E,J}‖_{2,J} ≤ D_*^{-190}; the number of events along a path
in
one level is at most D_*, and all groups are included in the path budget; these squared additive
errors are negligible"). For the groups G with their frozen coefficients c_{G,b,J} (squared
coefficient sums at most B_c on every pre-cell), at most D_* modes alive at any cell, cross
pairings at most D_*^{-200} and at most D_* modes through
each point across the groups (the path budget), the error of each exact side subtotal
(event_subtotal_close) satisfies err² ≤ B_c |J| / D_*^{397}, and the total over all groups,
blocks, pre-cells inside the root and side events is at most
3 (⌊180 d_max log(2 D_m)/δ_*⌋ + 1) B_c |R| / D_*^{396}.
Blueprint lem:lift-composition (lines 2062-2077) with the global coarse outputs, for rows
indexed by finite sets T_G depending on the group (as the side events of
prop:linear-side): mode projection rows P_{G,e}, e ∈ T_G, Bessel with constant c within
each group and local to active decision cells u_{G,e} with downstream use certificates, give
outputs P_{G,e} R_{l,G} Z_{l,G,u_{G,e}} with joint squared sum at most the constant of
lift_composition_global times |R|. The rows are padded by zero rows to one finite index
type, and groups without rows are dropped.
lift_composition_finset for rows indexed by finite sets T_G of types ρ_G depending on
the group (as the side events (b, J, E) of the mode system of G).
Assembly of eq:linear-energy on the main columns (blueprint lines 2080-2118) #
The augmented mode as a sum of its physical and private parts.
A cell mode vanishes off its cell.
Blueprint line 1967 on a retained pre-cell: a column function that agrees on the cell J with
the augmented-mode combination Σ_m β_m ψ_m, with β supported on the modes alive at the
decision cell u_b, restricts on J to the frozen evaluation Σ_{m alive} β_m Ψ_m.
A basic cell projection of a family input is at most the input on the cell.
Blueprint line 1934 ("normalized endpoint cell"): on an active cell U the coarse output of a
basic slot at the depth of U is at most the sum of the two family inputs on U (the ball
filter, the shared history and the basic cell projections are contractions).
Blueprint line 2116 ("their decision cells are the containing u_b-cells"): for a retained
v_b-cell J of the group G in a block starting after the attempt root, the u_b-ancestor of
J contains J, is active when the active cells are closed under ancestors, and carries a
downstream use certificate of G for its color.
Blueprint line 1967 and line 2118 ("its coefficient norm ... the complementary coefficient norm
is at most D_*"): on a retained v_b-cell J of the group, with the active cells closed under
ancestors and fresh normalized attempt roots (‖1_U F_{l,j}‖² ≤ C_F |U| on active cells), the
lifted frozen output R_{l,G} Z_{l,G,u_b} restricts on J to the frozen evaluation of
coefficients β supported on the modes alive at the decision cell u_b, with square sum at most
4 D_*^2 C_F (D_* covering the coefficient budget modeStar). Before the attempt the frozen
output is zero and β = 0.
The projection onto the span of the cell modes of E on J is local to every cell containing
J.
Blueprint prop:linear-side (eq:linear-energy, lines 2080-2118) for one basic role l of a
search leaf, one color k with ball level j ≤ 4 and shared history ls: let the mode systems
Ms_G with lifts R_{l,G} = L_G carry the exports of mainAttemptLift (MainLiftExports), at a
schedule S whose budgets D_*, D_m, d_max cover the mode data (dimensions, label counts,
curvatures, full Fourier norms of the squared gauges and truncation tails, inverse private weights,
alive modes, modes along a path across groups, and the coefficient budget modeStar), with active
cells closed under ancestors (the decision cells u_b are active) and a fresh normalized attempt
root (‖1_U F_{l,j}‖² ≤ C_F |U| on active cells). Then there are frozen coefficients c_{G,b,J}
representing ZTilde_{l,G,b} = R_{l,G} Z_{l,G,u_b} on every retained v_b-cell J of every block
b ∈ Bs (frozenEval), with square sums at most 4 D_*² C_F over the alive modes
(frozen_coeffs), and the exact side subtotals V_{l,G,b} (blockSide) satisfy
Σ_{G,b} ‖V_{l,G,b}‖² ≤ 24 (⌊180 d_max log(2 D_m)/δ_*⌋ + 1) (G₀ + 2 C_F D_*^{-394}) |R| with
G₀ = exp(C_q (1 + log 2k)²) of lift_composition_G0.
Proof: group_side_energy splits each block into its side events (almost orthogonal, cross
pairings ≤ D_*^{-200}), the event rows of each group are Bessel with constant
6 (⌊180 d_max log(2 D_m)/δ_*⌋ + 1) (group_event_rows_schedule), their decision cells are the
containing u_b-cells, so lem:lift-composition (lift_composition_dep, liftConst_le_G0) bounds
the row outputs, and the exact-subtotal errors are negligible (side_errors_negligible, with the
coefficient bound of frozen_coeffs).