Skip to content

Section 3.2 · PDF pages 11–13

The windowed Landau theorem

The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.

3.2 The windowed Landau theorem

For a measurable real function g, write 𝒲(g; θ, δ), where δ > 0, for the following condition: for every η > 0 and every real U_{0}, there exist U ≥U_{0} and a function R such that on J = [U, (1 + δ)U]

Selectable PDF notation • transcription unchecked
R ≥0, g ≤R, Re^{−θu}∈L^{1}(J), ∫_{J}R(u)e^{−θu}du ≤e^{ηU}.

The choices U, R may depend on both η and U_{0}.

Theorem 3.2 (Windowed Landau theorem). Suppose g, Φ, Z have the properties in Proposition 3.1, with the growth exponent 2. If 0 < θ < 1, δ > 0, and 𝒲(g; θ, δ) holds, then ℜρ ≤θ for every ρ ∈Z.

Proof. For σ > 2, differentiation under the integral gives

Selectable PDF notation • transcription unchecked
_{(−1)}k_{Φ}(k)_{(s)} ∫ ∞_{u}k_{e}−su
A_{k}(s) := _{k!} =_{0} g(u)_{k!} du (ℜs > 2). (12)

To justify all derivatives, at a fixed s use b = (ℜs −2)/2 > 0: on the disc of radius b, the integrands for the j-th and (j + 1)-st derivatives are dominated by C_{0}u^{j}e^{−bu} and C_{0}u^{j+1}e^{−bu}. These are integrable. Dominated differentiation and induction on j prove (12).

We first establish concentration of this kernel on a window. For a > 0,

Selectable PDF notation • transcription unchecked
∫ ∞_{u}k_{e}−au_{u}k_{e}−au
_{0k!} du = a^{−k−1},_{k!} ≤a^{−k} (u ≥0).

The first identity follows by integration by parts from the exponential integral, with the polynomial-exponential boundary terms zero. The second follows from (au)^{k}/k! ≤e^{au}. If 0 ≤l < s, splitting u < U or u > V gives ∫

Selectable PDF notation • transcription unchecked
_{u}k_{e}−su_{e}−lV_{e}lU
_{(0,∞)\[U,V ]} k!^{du ≤} (s −l)^{k+1 +}(s + l)^{k+1 .(13)}

Indeed the original exponential is bounded by the sum e^{−lV}e^{−(s−l)u}+ e^{lU}e^{−(s+l)u} on that complement; integrate the nonnegative majorant on the full positive half-line.

Put

Selectable PDF notation • transcription unchecked
l = 2 + 6/δ, M = l(1 + δ), σ = M(2 + l), C_{T} =_{σ −}^{e}_{2 −l}+_{σ −}^{1}_{2 + l}.

Both denominators are positive and σ > 2. For U ≥ 1, choose k = ⌈(1 + δ/2)σU⌉, so k ≥ U. The elementary estimate

Selectable PDF notation • transcription unchecked
_{e}c
e^{x}a^{−k−1}≤_{aσk} if x + k(σ −a)/a ≤c

follows from σ/a ≤e^{(σ−a)/a}. Apply (13) with s = σ−2, V = (1+δ)U. For its first term take a = σ−2−l, x = −MU, c = 1. The identities

Selectable PDF notation • transcription unchecked
(1 + δ/2)(2 + l) = M −1, (1 + δ/2)(l −2) = l + 1

give (σ −a)/a = 1/(M −1), (1+δ/2)σ = M(M −1), and M −1 ≥1. The upper ceiling bound therefore verifies the required inequality. For its second term take a = σ −2 + l, x = lU, c = 0. The lower ceiling bound gives k(l −2) ≥(l + 1)σU, and l(σ −2 + l) ≤(l + 1)σ; hence it too verifies that inequality. Thus the same σ, C_{T} , depending only on δ, satisfy

Selectable PDF notation • transcription unchecked
∫_{u}k_{e}(2−σ)u_{−k}
(0,∞)\[U,(1+δ)U]^{k!du ≤CT σ.(14)}

Let w_{k}(u) = u^{k}e^{−σu}/k!. For a majorant R on a window J, the triangle inequality in (12) gives, uni-formly for all real t,

Selectable PDF notation • transcription unchecked
∫ ∞
|A_{k}(σ + it)| + ℜA_{k}(σ) ≤_{0} (|g(u)| + g(u))w_{k}(u) du.

On J, |g| + g ≤2R and w_{k}(u) ≤(σ −θ)^{−k}e^{−θu}. On its complement, |g| + g ≤2C_{0}e^{2u}. All integrals are absolutely convergent by the exponential bound. Taking U ≥ max(U_{0}, 1) from 𝒲 and then its kernel index k, (14) proves

|A_{k}(σ + it)| + ℜA_{k}(σ) ≤C_{∗}e^{ηU}(σ −θ)^{−k} for every t ∈ℝ, (15)

where C_{∗}= 2 + 2C_{0}C_{T} . The constants σ, C_{∗} are fixed before η, U_{0}; the same U, k work for every t.

We next record the required pole geometry. Every pole has real part at most 2: otherwise the Laplace integral would bound Φ on a neighborhood of that pole, contradicting its nonzero residue. More explicitly, on ℜs ≥ σ^{′} > 2, |Φ(s)| ≤ C_{0}/(σ^{′}−2), whereas the local pole expansion makes (s −ρ)Φ(s) tend to a nonzero value. Furthermore Z ∩B(z, r) is finite whenever r < ℜz. Every point of this compact disc has a neighborhood meeting Z in at most that point; a finite subcover proves the assertion.

Suppose a pole has real part greater than θ. Apply this finiteness assertion to Z ∩B(σ, σ −θ). One of the following alternatives holds:

(a) There is D ≤σ −θ for which B(σ, D) ∩Z = ∅, and a pole ρ_{1} with σ −ℜρ_{1} < D. (b) A real pole β > θ is the only pole in B(σ, r), for some σ −β < r ≤σ −θ. To verify the dichotomy, if that finite set is empty use D = σ−θ and the original pole. Otherwise choose a member of minimum distance d from σ. If some pole satisfies σ −ℜρ < d, take D = d. If no pole does, a member at distance d must be the real point β = σ −d: the real-part distance is at least d, and its squared modulus is that distance squared plus the imaginary part squared. Also d < σ −θ, because the original pole has real part greater than θ. The positive gap to the remaining finitely many distances gives the radius r in (b).

In case (a), move the center slightly in the imaginary direction to make its nearest pole unique. Here is the needed justification. Let z_{0} = σ +iℑρ_{1}, and choose ϵ > 0 with ϵ < D −(σ −ℜρ_{1}) and D +ϵ < σ. The poles in B(z_{0}, D +ϵ) form a finite set. For two distinct members a, b, the equality |σ +it−a| = |σ +it−b| determines at most one real t. If their imaginary parts agree, equality would force their real parts to agree too, since both are less than σ. Choose t within ϵ of ℑρ_{1} avoiding these finitely many values. Then z = σ + it has a uniquely nearest pole ρ, at distance d < D. The gap to the other finite distances supplies d < r ≤D such that B(z, r) contains only ρ. Poles in this ball lie in the preceding finite set by the triangle inequality, so none have been omitted.

If the residue at this pole is c, subtract c/(s−ρ) and fill in its analytic value at ρ. The result is holomorphic on B(z, r). Cauchy’s coefficient estimate at any e with d < e < r gives

Selectable PDF notation • transcription unchecked
|A_{k}(z)| ≥|c|d^{−k−1}−Me^{−k}.

The pole-free disc B(σ, D) gives |A_{k}(σ)| ≤M^{′}e^{−k}. Since r ≤D ≤σ −θ, insert these inequalities into (15) to obtain

Selectable PDF notation • transcription unchecked
|c|/d_{≤} M + M^{′}_{+} C_{∗}e^{ηU}_{ .(16)}
d^{k} e^{k} (σ −θ)^{k}

All discs here lie in ℜs > 0, because their radii are less than σ.

In case (b), put d = σ −β, and choose d < e < r. The same pole subtraction and Cauchy estimate now give

Selectable PDF notation • transcription unchecked
ℜA_{k}(σ) ≥c_{β}d^{−k−1}−Me^{−k}.

The residue c_{β} is strictly positive by hypothesis. Taking t = 0 in (15) and dropping its nonnegative modulus term gives the same form as (16), with positive leading constant c_{β}/d.

Both conclusions are impossible. In fact if c > 0 and 0 < d < e ≤E, put η =_{2}^{1} log(e/d) > 0. For

Selectable PDF notation • transcription unchecked
k ≥U,
(M_{0} + C_{0})e^{ηU}/e^{k}_{≤} |M_{0}| + |C_{0}|_{e−ηU−→0.}
c/d^{k} c

Enlarge U_{0} so this ratio is less than 1. The choices allowed in (15) still give U ≥U_{0} and k ≥U, contradicting the corresponding lower bound. This proves the theorem.