Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Use the notation of the
claim.
The inner product below is linear in its first argument and uses the
measure dθ/(2π) on the unit circle.
The exterior inverse and its boundary norm
There is a univalent exterior inverse Ψ, normalized by
Ψ(w)=w+O(w−1), such that
p(Ψ(w))=w4,∣w∣>1.(1)
One can construct it directly by continuing the root asymptotic to w
near infinity. All finite critical values have modulus one, so no roots
collide for ∣w∣>1. A circuit in the w-plane winds the target w4
four times around all finite critical values; the root monodromy around
them is the four-cycle seen at infinity, so this continuation is
single-valued. The four roots are Ψ(w),Ψ(iw),Ψ(−w),Ψ(−iw).
They are distinct and account for every inverse root. This also proves
univalence: two arguments with the same image must differ by a fourth
root of unity, and the four listed roots never coincide.
The nonvanishing derivative has winding number zero around infinity.
Consequently it has an analytic square root
g(w)=Ψ′(w)=1+k≥2∑gkw−k,∣w∣>1.(2)
At a critical point of multiplicity r in p′, the derivative of an
inverse branch has singular order r/(r+1)≤3/4. Its modulus is
integrable. The finitely many such boundary expansions show that g has
square-integrable boundary values and that the radial boundary norms
converge. The boundary parameterization by Ψ counts each regular
arc of ∣p∣=1 once: w↦w4 runs through the four inverse sheets.
Thus Parseval's identity gives
S=2πℓ(p)=∥g∥2=1+k≥2∑∣gk∣2.(3)
If S=1, every gk vanishes, so Ψ′=1 and its normalization gives
Ψ(w)=w. Then p=z4, whose critical value is zero, contrary to the
hypothesis. Hence S>1.
The modes fixed by polynomial division
Solving (1) at infinity and taking the square root gives
g2=8a,g3=4b,g4=83c−1287a2,g5=−32ab.(4)
For example, the inverse begins
Ψ=w−a/(4w)−b/(4w2)+(a2/32−c/4)w−3+0w−4+O(w−5),
which verifies all four coefficients in (4).
Let V be the polynomial quotient of p2 by p′, so V′=Q, and put
R=p2−p′V, of degree at most two. Differentiating (1) yields
p′(Ψ)Ψ′=4w3, whence
The last term starts at w−2. The coefficients of w4,w3,w2,w,1,w−1
in f are therefore, respectively,
45,0,32a,−16b,A=51221a2−329c,1285ab.(6)
Since Q stays bounded on the compact lemniscate, f has
square-integrable boundary values as well. With arclength transferred by
Ψ, the moment identities are
⟨f,g⟩=Sm,∥f−mg∥2=SVarμ(Q).(7)
The projection bound
Set h=f−mg. It is orthogonal to g. Because g has no positive modes
and no w−1 mode, (6) fixes the squared norm in the four modes
w4,w2,w,w−1 of h to be
D=1625+1024∣a∣2+256∣b∣2+1638425∣ab∣2.(8)
Its constant coefficient is h0=A−m. Write h−k for its coefficient
of w−k. Orthogonality to g, with g0=1 and g1=0, says
The pushforward of normalized arclength by p has density proportional
to ∣ζ−c∣−3/4 on the unit circle. Writing ζ=ceiθ,
the beta-integral recurrence gives
Esin2(θ/2)=1/5, hence
Eζ=3c/5. It follows that m=0 and
Var(Q)=(25/16)(1−9/25)=1.
The Cauchy--Schwarz step above is strict here. In fact
g4j=(3/8)jcj/j! and all other nonconstant modes vanish.
For j≥1, (9) gives
g4jf−4j=c(45j+1j+3/8−43)=cj+1j/2−9/32.(10)
These ratios differ already at j=1,2, and both corresponding
g4j are nonzero. The two tail vectors are not proportional, so
equality in Cauchy--Schwarz is impossible. The stated lower bound is
therefore strictly less than one at (9). Its exact value there is
16S025+1024(S0−1)81,S0=2πℓ(z4−1).
This strict loss concerns the finite projection used in the proof, not
the unproved variance inequality for regular quartics.