Section 3.8 · PDF pages 21–24
The slope-to-gap estimate
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
3.8 The slope-to-gap estimate
We spell out the scalar slope hypotheses used to bound the gap. This is the interface at which the geometric construction enters the analytic argument. These inequalities must be established for the constructed family; they are not consequences of classical properties of Dirichlet L-functions.
Fix a translate, a finite nonempty set W with positive weights d_{u}, a finite node set of size n_{T} , and m, f ∈ℕ, f > 0. Let a_{t} ≥0, w_{j,ρ} > 0, b, k, v > 0, a, b^{′}, B_{0}, N_{c}, L, L_{a}, X^{′′}≥0. The remaining scalar data are
T_{u} > 0, h_{u}, E_{u}, µ_{N,u}, µ^{max}_{N,u}, µ_{M,u}, µ^{max}_{M,u}, µ_{C}, Fo, Ga, ASK, IO, λ_{g}, ℓ_{C}, ℋ.
Put S_{A} =∑_{t}a_{t}, and for functions on W write ⟨X⟩_{d} = ∑ d_{u}X_{u}/ ∑ d_{u}. Define
s = ⟨µ_{N} −h⟩_{d}, u_{F} = max(µ^{max}_{N,u}−h_{u}) −s.
Selectable PDF notation • transcription unchecked
u
The constants c_{Div}, c_{E}, c_{H}, c_{Count}, c_{Slope}, c_{Sl1}, c_{Loss} and e_{Div}, e_{Loss}, c_{HH} are nonnegative. Define
log Ω= γ + log(π/v) −ℓ_{C},
Selectable PDF notation • transcription unchecked
∑
S_{Div} = m (a_{t} log a_{t} −a_{t} −a_{t} log Ω),
Selectable PDF notation • transcription unchecked
t
E_{Div} = c_{Div}n_{T} aL_{a} + c_{Div}m(B_{0} + N_{c})x^{eDivϵ},
L_{P} = c_{Loss}n_{T} mbx^{eLossϵ},
∑ 1∑Λ_{5} = log(w_{j,ρ}v/π) + mλ_{g} −c_{Count}(L_{a} + m^{2}L_{a}/k),
Selectable PDF notation • transcription unchecked
j<m∑∑^{f}ρ<f( ∑ )Main = a_{t}(log a_{t} −1 −γ + ℓ_{C} −^{1} log w_{j,ρ}),
Selectable PDF notation • transcription unchecked
j<m t^{f}ρ<fL_{u} = c_{Sl1}IO + c_{Sl1}n_{T} m(a/k + b^{′}) log T_{u} + c_{Sl1}X^{′′}+ 1.
The convention 0 log 0 = 0 handles any zero node size. The supplied scalar hypotheses are
µ_{C} + h_{u} − Fo ≤S_{Div} −S_{A} log T_{u} + E_{Div},
Selectable PDF notation • transcription unchecked
|E_{u} −mS_{A}λ_{g}| ≤c_{E}mB_{0},|(h_{u} + S_{A} log T_{u}) −(h_{u}′ + S_{A} log T_{u}′)| ≤c_{H}(n_{T} aL_{a} + mB_{0}),
Λ_{5} ≤ log T_{u} ≤c_{Count}mL, | log T_{u} − log T_{u}′| ≤c_{Count} log k,
Selectable PDF notation • transcription unchecked
−c_{Slope}n_{T} aL_{a}(1 + m/k) ≤µ_{M,u}, µ^{max}_{M,u}≤c_{Slope}n_{T}aL_{a},^{(27)}µ^{max}_{N,u}≤µ_{M,u}^{max} + Ga + Fo −E_{u} + c_{Sl1}mB_{0},
µ_{M,u} + Ga + Fo −E_{u} −L_{u} ≤µ_{N,u} ≤µ^{max}_{N,u},
ASK − Ga ≤L_{P} , IO ≤L_{P} .
They hold for every u, u^{′}∈W. There are also numbers 0 ≤δ ≤δ <^{̄} 1, 0 ≤η ≤η̄, 0 ≤e ≤ē ≤1/4, and 0 < ℋ ≤x^{cHH}b, for which the supplied slope inequality is
_{C} 4u_{F} log(1 + ℋ/u_{F} ) + (η + 4^{√}e)ℋ s −µ ≤ log |W| + . (28)
Selectable PDF notation • transcription unchecked
1 −δ
The function z log(1 + ℋ/z) is given its continuous value 0 at z = 0. In the original slope hypothesis, (28) is conditional on s −µ^{min}_{C} ≤ℋ and ℋ > 0; both inequalities are required from the family construction. This formulation has displayed their used consequence rather than silently treating that slope inequality as an established general theorem.
Set
V = 2mc_{E}B_{0} + c_{H}(n_{T} aL_{a} + mB_{0}) + S_{A}c_{Count} log k,
Selectable PDF notation • transcription unchecked
K_{3} = c_{Slope}n_{T} aL_{a}(2 + m/k) + V + c_{Sl1}mB_{0},
K_{0} = S_{A}c_{Count}(L_{a} + m^{2}L_{a}/k) + E_{Div} + mc_{E}B_{0} + c_{Slope}n_{T} aL_{a}(1 + m/k),
M_{L} = c_{Sl1}L_{P} + c_{Sl1}n_{T} m(a/k + b^{′})c_{Count}mL + c_{Sl1}X^{′′}+ 1,
U_{B} = K_{3} + M_{L}.Lemma 3.8 (Pointwise slope bound). Under the preceding hypotheses,
ASK − Main ≤ Err := K_{0} + log |W| + L_{P} + M_{L}
_{+ }4 log(1 + x^{cHH})(U_{B} + b) + (̄η + 4^{√}ē)x^{cHH}b
Selectable PDF notation • transcription unchecked
1 −δ^{̄.}Proof. Define
Selectable PDF notation • transcription unchecked
K_{5} = c_{Slope}n_{T} aL_{a}(1 + m/k) + m(S_{A}λ_{g} + c_{E}B_{0}) + S_{Div} −S_{A}Λ_{5} + E_{Div}.The lower quotient bound, lower bound for µ_{M}, upper bound for E_{u}, and divisor inequality in (27) imply
µ_{N,u} −h_{u} ≥ Ga + µ_{C} −K_{5} −L_{u}.
Here the count lower bound gives S_{A} log T_{u} ≥S_{A}Λ_{5} because S_{A} ≥0. Take weighted means to obtain
s ≥ Ga + µ_{C} −K_{5} −⟨L⟩_{d}. (29)
The comparison bounds for E, h and log T give
Selectable PDF notation • transcription unchecked
(−E_{u} −h_{u}) −(−E_{u}′ −h_{u}′) ≤V.In detail the E difference costs 2mc_{E}B_{0}. The h comparison costs c_{H}(n_{T} aL_{a} + mB_{0}) plus S_{A}(log T_{u} − log T_{u}′), and the count comparison bounds the latter by S_{A}c_{Count} log k. Combining the two quotient bounds and the two bounds for the M slopes therefore gives
Selectable PDF notation • transcription unchecked
µ^{max}_{N,u}−h_{u} −K_{3} −L_{u}′ ≤µ_{N,u}′ −h_{u}′.Average over u^{′}, then maximize over u, to obtain u_{F} ≤K_{3} + ⟨L⟩_{d}. Also u_{F} ≥0, since µ_{N} ≤µ^{max}_{N} and a weighted mean is at most the maximum of its entries.
For 0 ≤z ≤U, 0 < ℋ ≤Xb, X, b > 0,
z log(1 + ℋ/z) ≤ log(1 + X)(U + b).
For z = 0 use the specified value. For z > 0, apply log q ≤q −1 to q = (1 + ℋ/z)/(1 + X), multiply by z, and discard the negative term −zX/(1 + X). The remaining term ℋ/(1 + X) is at most bX/(1 + X) ≤ b log(1 + X). The last inequality follows from the same logarithmic inequality applied to 1/(1 + X). Use this estimate with z = u_{F} , U = K_{3} + ⟨L⟩_{d}, and X = x^{cHH} in (28). Replacing δ, η, e, ℋ by their allowed upper bounds is legitimate since all multiplying quantities are nonnegative.
Expanding the logarithms in Λ_{5} gives the exact identity
S_{Div} −S_{A}Λ_{5} + mS_{A}λ_{g} = Main + S_{A}c_{Count}(L_{a} + m^{2}L_{a}/k),
so K_{5} = K_{0} + Main. Combine this identity with (29) to bound Ga − Main by the claimed expression with ⟨L⟩_{d} in place of M_{L}. The upper count bound and IO ≤ L_{P} give ⟨L⟩_{d} ≤ M_{L}. All its coefficients in the expression are nonnegative, so this replacement increases the bound. Finally ASK ≤ Ga+L_{P} , which proves the result.
Lemma 3.9 (The uniform error budget). In addition to the preceding assumptions, suppose
Selectable PDF notation • transcription unchecked
x ≥1, 0 < ϵ ≤1, b ≥1, xb ≤a, n_{T} ≥1,m ≤x^{1/2+ϵ/2}, xm ≤k, 1 ≤k ≤ min(a, x^{eK}), e_{K} ≥0,
Selectable PDF notation • transcription unchecked
0 ≤L, L_{a} ≤2x^{ϵ}, kb^{′}≤a, X^{′′}≤a,_{√}0 ≤a_{t} ≤2a, δ^{̄} ≤1/2,log |W| ≤(k −3) log 2, (̄η + 4 ē)x^{cHH}b ≤Kx^{ϵ}n_{T} a.
Assume each of the seven constants c_{∗} lies in [0, K], and
log x ≤x^{ϵ}, 200(1 + K)^{2}(1 + c_{HH})(1 + e_{K}) ≤x^{ϵ}.
Then, for E_{T} = n_{T} a(1 + m^{2}/k) + m(B_{0} + N_{c}),
Err ≤x^{(eDiv+eLoss+3)ϵ}E_{T} .
Proof. Write y = x^{ϵ}, y_{D} = x^{eDivϵ}, y_{L} = x^{eLossϵ}, h = 1 + c_{HH}, g = 1 + e_{K}. All five quantities are at least 1. From ϵ ≤1 we have m ≤x, hence mb ≤xb ≤a; also m ≤k, b ≤a, a ≥1, and S_{A} ≤2n_{T} a. The cardinality hypothesis gives log |W| ≤k ≤a. Furthermore
0 ≤ log k ≤gy, 0 ≤ log(1 + x^{cHH}) ≤hy, 0 ≤(1 −δ^{̄})^{−1}≤2.
For the second logarithm use 1 + x^{cHH} ≤2x^{cHH} and log 2 ≤1.
Put Y = y_{D}y_{L}y, P = Y E_{T} , r = m^{2}/k. The nonnegative terms in E_{T} show
Selectable PDF notation • transcription unchecked
n_{T} a(1 + r) ≤E_{T} , mB_{0} ≤E_{T} , 1 ≤a ≤E_{T} ≤P.Since L, L_{a} ≤2Y , m/k ≤1, and b^{′}≤a/k, the individual terms of K_{0} have bounds
Selectable PDF notation • transcription unchecked
S_{A}c_{Count}(L_{a} + m^{2}L_{a}/k) ≤4KP,
E_{Div} ≤3KP, mc_{E}B_{0} ≤KP,
c_{Slope}n_{T} aL_{a}(1 + m/k) ≤4KP.For the first use S_{A}(1 + r) ≤ 2E_{T} ; for E_{Div} bound its two defining summands by 2KP and KP. Also L_{P} ≤KP, because mb ≤a and y_{L} ≤Y . The middle summand of M_{L} is at most 4K^{2}P: indeed
Selectable PDF notation • transcription unchecked
n_{T} m^{2}(a/k + b^{′}) ≤2n_{T} ar ≤2E_{T} .The other summands are at most K^{2}P, KP, P. Thus M_{L} ≤(5K^{2}+ K + 1)P.
For U_{B} + b, its slope term is at most 6KP, the c_{H} term at most 3KP, its two mc_{E}B_{0} terms at most 2KP, the c_{Sl1} term at most KP, and the logarithmic-count term at most 2KgP. Together with the bound for M_{L} and b ≤P, this gives
Selectable PDF notation • transcription unchecked
U_{B} + b ≤uP, u = 5K^{2}+ 13K + 2Kg + 2.The remaining noise term is at most KP by hypothesis. Consequently the full error is at most
Selectable PDF notation • transcription unchecked
( (5K^{2}+ 14K + 2)P + 2) (4hy uP + KP).Since y ≥1, this is at most (5K^{2}+ 16K + 2) + 8hu yP. For h, g ≥1,
Selectable PDF notation • transcription unchecked
(5K^{2}+ 16K + 2) + 8hu ≤200(1 + K)^{2}hg ≤y.To verify the polynomial inequality explicitly, bound its first summand by (5K^{2}+ 16K + 2)hg, and bound u ≤(5K^{2}+ 15K + 2)g. The resulting coefficient is 45K^{2}+ 136K + 18 ≤200(1 + K)^{2}; the difference is 155K^{2}+ 264K + 182 ≥0. The error is therefore at most
Selectable PDF notation • transcription unchecked
y^{2}P = y_{D}y_{L}y^{3}E_{T} = x^{(eDiv+eLoss+3)ϵ}E_{T} .Proposition 3.10 (Gap supplied by slope data). Suppose the slope data just described exist at every translate u_{0} in a finite set U_{0}, with constants independent of u_{0}. Suppose there is a map tr : U_{0} →H, whose image contains A_{0}, and bijections of the nodes with T and of the f weight indices with Φ, identifying a_{t}, w_{j,ρ} with the corresponding arithmetic node sizes and weights. Require ℓ_{C} = ℜℓ(χ_{C}) and require that the slope datum’s count ASK equal the sum of the arithmetic ASK(a_{t}, t, w_{tr(u0),j}). Finally suppose
Selectable PDF notation • transcription unchecked
B_{0} ≤2nmbx^{ϵ}, N_{c} ≤|T_{2}|kbx^{ϵ}/(n^{2}x^{−ϵ}).Then
gap(u) ≤x^{(eDiv+eLoss+3)ϵ}ℰ (u ∈A_{0}).
Proof. Under the stated bijections, the formula for Main is precisely the sum of the real parts of the arithmetic main terms. Thus gap(tr(u_{0})) = ASK(u_{0}) − Main(u_{0}). Apply Lemmas 3.8 and 3.9. The two final bounds replace E_{T} by the nonnegative bracket ℰ in Proposition 3.7. Every u ∈A_{0} has a preimage under tr, giving the assertion at all required translates.