Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
We prove the
stated obstruction.
All lengths are Euclidean arc lengths; crossing points are counted once,
which has no effect on length integrals.
The family and its limiting line
For y>0 put
Fy(z)=z4+(2y2−23)z2−(1+4y2)z.
Its derivative is
4(z−1)((z+21)2+y2), so its three simple critical points are
1 and −21±iy. Their values are
vy=−23−2y2,uy±4iy3,uy=163+25y2−y4.
Thus dy=vy−uy=y4−29y2−1627.
Its positive zero is
Y=49+233, and dy>0 for y>Y.
The circle Γy through these values has real center and radius
Cy=vy−Ry,Ry=2dydy2+16y6.
Indeed ∣vy−Cy∣=Ry and
(uy−Cy)2+16y6=(Ry−dy)2+16y6=Ry2.
Consequently
py(z)=Ry−1Fy(Ry1/4z)−Cy/Ry is centered and monic and has
three distinct critical values on the unit circle. Its critical points
are distinct too.
As y↓Y, the radius tends to infinity and
−Cy/Ry=1−vy/Ry→1. The other two nonleading coefficients tend to
zero. Hence py→z4+1. On every bounded set Γy converges
smoothly to the line
L=−c+iR,c=6+33,
and
FY(z)=z4+(3+33)z2−(10+63)z has all three critical values
on L.
The limiting integral is positive
For any centered monic quartic F, set
GF(σ)=F(z)=σ∑∣F′(z)∣1
away from its critical values. Write F=FY and define
Both differences are absolutely integrable. Near a critical value the
inverse branches have integrable square-root singularities, since the
critical points are simple. At infinity the estimate proved below gives
GF(σ)=∣σ∣−3/4+O(∣σ∣−7/4), and translating the
reference by the fixed number c preserves integrability.
Let H={z:ReF(z)=−c}. Centering and monicity give
eight ends of H, with asymptotic directions
ej=exp(i(π/8+jπ/4)),0≤j<8.
For sufficiently large r each end meets ∣z∣=r once, at
rej+O(r−1). This follows by applying the implicit function theorem
to Re(r−4(F(reiθ)+c))=0 near the eight
simple zeros of cos4θ; differentiating also makes the end radial
coordinate eventually monotone. The same expansion shows
J=r→∞lim(∣H∩{∣z∣≤r}∣−8r).(1)
For completeness, cutting the integral defining J at ∣v∣≤r4
gives the length of the part of H with ∣F(z)+c∣≤r4, minus 8r.
Along each end its cutting point differs from that on ∣z∣=r by
O(r−1), because F has no cubic term. Thus the two length cuts
differ by O(r−1).
The only vertices of H are its three critical points
w0=1, w+=−21+iY, w−=−21−iY, each of valence four.
There is no cycle: a cycle would contain a Jordan curve on which the
harmonic function ReF+c vanishes, forcing that function
to vanish on the enclosed domain by the maximum principle. Cut the eight
ends at a sufficiently large circle and regard the cut points as vertices
of valence one. The resulting finite forest has eleven vertices and ten
edges, by the degree sum 3⋅4+8=20. It is therefore a tree. In
particular it has two finite edges connecting the three critical points.
Their total length is at least
249+Y2>5,(2)
since the two distances from w0 to w± are the two shortest
pairwise distances and Y>2.
For a unit vector e let
h(e)=maxw∈{w0,w+,w−}Re(weˉ).
An end beginning at a critical point has length, up to its intersection
with ∣z∣=r, at least r−h(ej)+o(1). In the upper half-plane the
maximizing point is w0 in direction π/8, and w+ in directions
3π/8,5π/8,7π/8. To check the first two assertions, use
Y<9/4<23cot(π/8) and
Y>2>23cot(3π/8); the last two are immediate. Conjugation
handles the lower half-plane. Writing q=cos(π/8) and
s=sin(π/8), we obtain
j=0∑7h(ej)=3q+2Y(2q+s)<3+29(2+21)=457.(3)
Here 2<Y<9/4, q<1 and s<1/2. The bound on Y follows directly
from its displayed square: Y2>4 and Y2<81/16.
Combining (1)--(3) gives J>−37/4.
The difference between the two reference integrals is exactly
I−J=c1/4K,K=∫−∞∞(∣v∣−3/4−(1+v2)−3/8)dv>6.
Indeed the integrand is positive everywhere and its integral on
[−1,1] is greater than 2∫01(v−3/4−1)dv=6.
Also c>11>(9/5)4. Therefore
I>−437+554=2031.(4)
Passing to the nearby circles
Inverse-branch integration and the change of scale give exactly
The reference is correct because
∣σ∣−3/4 is the inverse-derivative density of z4, and
{z:z4∈Γy} has length
Ry1/4ℓ(z4−Cy/Ry).
We justify convergence of the right side of (5) to I, including its
singularities and expanding tail. Restrict y to a small fixed compact
interval about Y. Its critical points stay simple and separated and
their second derivatives stay bounded away from zero. The local inverse
function theorem in a square-root coordinate therefore gives uniform
constants A,η0>0 such that, near each critical value fy,
GFy(σ)≤A∣σ−fy∣−1/2+A(0<∣σ−fy∣<η0).
Since Γy passes through fy and its radius tends to infinity,
integrating this estimate over ∣σ−fy∣<η gives
O(η1/2), uniformly for y>Y sufficiently close to Y.
One can see this directly from the chord formula
∣σ−fy∣=2Ry∣sin(t/(2Ry))∣ in arclength t from fy.
The reference has no singularity there: the distance from 0 to
Γy is ∣vy∣→c>0.
Away from these windows, smooth convergence of the circles and the
implicit function theorem for simple roots give convergence on each
bounded set. To control infinity uniformly, take ρ4=σ and
ω4=1. For Fy(z)=z4+ayz2+byz, its four inverse branches
have the uniform expansions
These are convergent inverse series near infinity with a uniform radius
for the bounded coefficient family; differentiation is legitimate there.
On summing the moduli, the linear terms of orders ρ−2 and
ρ−3 cancel because
∑ω4=1ω−2=∑ω4=1ω−3=0.
Consequently
GFy(σ)−∣σ∣−3/4≤B∣σ∣−7/4
beyond a fixed radius, with B independent of y. The length of the
part of any circle in a disc of radius r is at most 2πr (and that
of a line is at most 2r). Summing over the annuli
2jT<∣σ∣≤2j+1T bounds the integral of
∣σ∣−7/4 on the tail by
4πT−3/4∑j≥02−3j/4, independently of the circle.
Thus the tails tend uniformly to zero.
First remove the critical windows and the tail, take y↓Y
on the remaining compact arcs, and then let the windows shrink and the
tail radius increase. This proves convergence in (5) to I. Inequality
(4) proves the claimed limit and hence the strict length comparison for
all y>Y sufficiently close to Y.