Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
We prove the
claim.
The change z=λw, followed by division of P by λ4,
multiplies the renormalized length by λ. It therefore suffices
to treat
∣a∣≤1,∣b∣≤1,max(∣a∣,∣b∣)=1.(1)
The real constant c is unrestricted. All constants below are independent
of it.
The projection inequality
For α∈R/πZ and s∈R let
N(α,s)=#{t∈R:ReP(eiα(t+is))=0}.
Except for finitely many directions and a null set of tangent lines,
this is a number between zero and four. If A is any measurable set of
such lines, the projection inequality is
J(P)≤−21∫A(4−N(α,s))dsdα.(2)
Here J(P) exists: centering gives eight ends with positions
rei(π/8+jπ/4)+O(r−1) and tangent derivatives differing
from those of rays by O(r−2). Each end therefore has a finite
length excess over its radial coordinate. The compact portion of the
real algebraic curve has finite length.
For clarity, (2) is the normalization of the projection argument in
the proof of Fryntov and Nazarov's Lemma 1, pp. 5--6 of the preprint
identified on the
source card.
Project the plane centrally onto the lower hemisphere of a sphere of
radius R tangent to it at zero. Spherical Crofton gives projected
total length 4πR−DR, with
DR=21∫0π∫R(4−N(α,s))(1+s2/R2)−3/2dsdα.
For all sufficiently large r, every circle of radius ρ≥r
meets H at least eight times, by the leading term Rez4.
The radial coordinate on the hemisphere therefore bounds the length of
the projected part outside Dr from below by
4πR−8Rarctan(r/R). Its part inside has length at most
8Rarctan(r/R)−DR. Letting R→∞, restricting the
nonnegative deficit integral to A, and then letting r→∞
gives (2). Equivalently the unoriented Euclidean line measure is
21dsdα, which fixes the factor in (2).
A quadratic coefficient bounded away from zero
Suppose ∣a∣≥1/100. There is a direction
α0∈{−π/12,π/12,π/2−π/12,π/2+π/12}
for which
cos4α0=1/2 and
Re(ae2iα0)≥∣a∣/2.
Indeed, writing a=x+iy, the first two choices yield
3x/2±y/2, and the other two change their signs. The largest
absolute value of the first two is at least
3x2/4+y2/4≥∣a∣/2.
For ∣α−α0∣≤1/100, the elementary Lipschitz estimates for
the exponential give, with A4=cos4α and
B=Re(ae2iα),
A4≥21−1004>52,B≥2∣a∣−1002∣a∣>52∣a∣.(3)
In the real polynomial
q(t)=ReP(eiα(t+is)) the leading, cubic and
quadratic coefficients are respectively
A4,−4ssin4α,B−6s2A4.
After division by A4 and removal of the cubic term, the quadratic
coefficient is exactly
A4B−A426s2.(4)
For ∣s∣≤∣a∣/10 it is positive, since
A4B>4∣a∣/25>6∣a∣/100 by (3). A monic real quartic with four real
roots and zero cubic coefficient has quadratic coefficient
−21∑jxj2≤0. Thus q has a nonreal conjugate pair,
and N(α,s)≤2 throughout this rectangle, apart from irrelevant
null sets.
The rectangle has angular width 1/50 and offset width
∣a∣/5. Substitution in (2) gives
J(P)≤−250∣a∣≤−25001.(5)
Angles are taken modulo π; crossing an endpoint of a fundamental
interval does not change the rectangle's line measure.
A substantial linear coefficient
It remains to treat ∣a∣<1/100, so (1) gives ∣b∣=1.
Choose α0∈{0,π/2} such that
∣Re(beiα0)∣≥1/2>7/10.
For ∣α−α0∣≤1/100 and ∣s∣≤1/100, write
These inequalities follow directly by inserting (6); the two error
terms in the second line sum to less than 1/1000.
If the depressed quartic had four real roots xj, Newton's identities
would give
This contradicts (7). Hence again N(α,s)≤2. The angular and
offset widths are both 1/50, so (2) gives J(P)≤−1/2500.
Together with (5), this proves the normalized estimate. Restoring the
scale λ proves the claim.