You are on page 1of 5

A draft on the repartition of algebraic numbers

A. BALI

ABSTRACT. A lot of unsolved number theoretic open problems are about the
irrationality or the transcendence of a number. This paper will share a few
results which are useful to know about the repartition of algebraic numbers.
We will specifically look at quadratic roots, as quadratics are easier to study
than most other polynomials, and because they also contain ℂ \ ℝ roots.

Let Pn = {P | P(x) = ax2 + bx + c = 0, (a, b, c) ∈ ⟦-n, n⟧3, a ≠ 0}, where we note ⟦A, B⟧ as [A, B] ∩
ℤ. We can plot all the ℂ \ ℝ roots of all P ∈ Pn in the complex plane for a few integer values of n.

P3 ℂ \ ℝ roots P4 ℂ \ ℝ roots

P5 ℂ \ ℝ roots P6 ℂ \ ℝ roots

We can see that they form a very particular, symmetrical shape. As we iterate, we might even find
out it is a self-similar fractal.
A simple algorithm we could use to plot these would be :
Int n;
For a = -n, a ≤ n, a++ :
For b = -n, b ≤ n, b++ :
For c = -n, c ≤ n, c++ :
If a ≠ 0:
Plot { x | ax²+bx+c = 0 };
End If;
End For;
End For;
End For;
Let's define first an integer t and add a t++; instruction right after Plot. In order to render
quicker, we thought we could edit the program this way to get approximately the same result with
fewer points :

Int s; Int t; Int n;


For a = -n, a ≤ n, a++ :
For b = -n, b ≤ n, b++ :
For c = -n, c ≤ n, c++ :
If t ≡ 0 (mod s) and a ≠ 0:
Plot { x | ax²+bx+c };
End If;
t++;
End For;
End For;
End For;

However, the result was not actually what we expected it to be :

P15 ℂ \ ℝ roots P15 ℂ \ ℝ roots


t = 1; t = 10;

This sort of cut is actually at Re(z) = 1/2. We can hence ask how many polynomials have -b/a < 1
and b² + 4ac < 0. An algorithm sorted out a few results which, assuming the percentage decreases in
a sort of x↦A/x+B fashion, gives us, by interpolating percentages for n = 60 and n = 100, the values
of A = 1 194 049/5 974 850 ≈ 19.98 % and B = 109 235 351/358 491 540 ≈ 30.47 %.
We can hence deduce the limit is around 30.47 %. We will, however, try to find good lower and
upper bounds for this limit, assuming it even exists. We can draw a table with a ∈ ⟦-n, -1⟧ as
columns, b ∈ ⟦-n, n⟧ as rows, and each cell containing all ordered triples ( a, b, c) ∈ ⟦-n, n⟧3. We
represent (a, b, c) as a black rectangle iff -b/a < 1 and b² + 4ac < 0, not represented otherwise. We
can do this by using the following algorithm (A).

Int n;
New Table;
For a = -n, a < 0, a++:
New Row;
For b = -n, b < n, b++:
New Cell;
For c = -n, c < n, c++:
If b² + 4ac < 0 and -b / a < 1:
Add "█\n" in current cell;
Else:
Add " \n" in current cell;
End If;
End For;
End For;
End For;

We can know check the result of the algorithm for different values of n :

n = 3;
The incrementing number of blank
cells per row corresponds to -b > a.
However, half of each cell contains
a blank area, which corresponds to n = 4;
c > 0. Indeed, we can show that for
negative values of a, then b² – 4ac > 0. Proof :
Assume b² < 4ac (E). b is an integer, so b² > 0. If a < 0 and c > 0, then
ac < 0, and 4ac < 0 as well. However, this means b² < 0.
n = 5;
Hence, we have at least one upper bound, which is all the permanent blank parts of the cells over
the number of cells in the table, the whole substracted to one. 1 + 2 + 3 + ... + n – 1 = (n – 1) n / 2
forceful blank cells due to the -b < a condition constitute the right part of the table. There are
precisely 2n + 1 triples in each cell, and there are precisely (2n + 1)² n cells, which therefore gives
us at least ((2n + 1) (n – 1) n) / (2 (2n + 1)² n) elements of ⟦-n, n⟧3 which don't fit the criteria. We
also know that if c > 0 and a < 0, then there is no solution to b² < 4ac & -b < a, then (n + 1) (2n + 1)
triples in each cell don't fit the equation. Since there are n rows, this gives us another area of
unfitting triples, which makes up (2n + 1) (n + 1) n / ((2n + 1)² n) of the table. However, the area
intersects the triples of the strips. Hence, we need to reduce the strips' triples to (a, b, c) ∈ ⟦-n, n⟧3
such that -b < a & c < 0, which gives us n triples. Hence, these reduced strips would actually
constitute ((n – 1) n) / (2 (2n + 1)²) of all triples in the table. We can now add these two unfitting
triples percentages as a single equation, for which the limit, once substracted from one, would give
us an upper bound for the percentage of quadratics such that b² < 4ac and -b < a.

1 - lim n → ∞ ((n + 1) n / ((2n + 1) n) + ((n – 1) n) / (2 (2n + 1)²)


= 1 - lim n → ∞ ((2n + 1) / (4n + 2) + (2n – 1) / (16n + 8))
= 1/2 - 1/8
= 3/8

We hence know that the percentage of (a, b, c) ∈ ⟦-n, n⟧3 triples such that -b < a & b² < 4ac is
below three halves. However, we also need a lower bound. We first off observed a tendency for
having only three last fitting triples in the last row, and a sort of almost halved hyperbola look in
there as well. There seems to be a decreasing amount of triples per row, so we may say that
multiplying by n the number of triples in the last row would be a good lower bound. Assuming this,
and looking at the very particular almost halved hyperbolic fashion of the triples repartition in the
last row, we assume we could do the following conjecture :

n ∫ ∛0( n / 2 ) (n – x3) dx < Card { (a, b, c) ∈ ⟦-n, n⟧3 | -b < a & b² < 4ac } (E´)

We use the third degree in order to use l'Hôpital's rule later. Of course, we can simplify this definite
integral as ∫∛0( n / 2 ) (n – x3) dx = 3n5/2/(8∛2). Since 5/2 is not a integer power, we should rather use
the fact that n5/2 > n² for all n > 1 in order to apply l'Hôpital's rule and getting a reasonable lower
bound for this supposed percentage limit :

lim n → ∞ (3n²/(16 ∛2 n² + 8 ∛2 n)))


= lim n → ∞ (6n/(32 ∛2 n + 8 ∛2)))
= 3/(16 ∛2)

Therefore, if (E´) is actually true, then we'd have a lower bound of 3/ (16 ∛2) for the percentage of
the ordered triples in ⟦-n, n⟧3. Hence, if conjecture (E´) is true, we can also deduce that :

3/(16 ∛2) < lim n → ∞ Card { (a, b, c) ∈ ⟦-n, n⟧3 | -b < a & b² < 4ac } < 3/8

which is indeed a wide estimate (5/8 – 3/(16 ∛2) ≈ 47.62 %), but sufficient to show that if the limit
does exists, it is surely between those two constants. At the very least, 3/8 is a safe upper bound,
which suffices to show that quadratic roots don't have the same repartition for real parts below or
above a half. We can also conclude that there is a non-uniform repartition of quadratic roots in the
complex plane with real parts below or above some constant c ∈ ℝ. In the rest of the paper, we will
try to find good bounds for ℓz := lim n → ∞ Card { (a, b, c) ∈ ⟦-n, n⟧3 | -b < 2az & b² < 4ac }.
We can show that, if (E´) indeed holds, then for all n > 1/2, ℓn > 3/(16 ∛2) holds, because there are
always less fitting triples in the last row than in any previous row, which we can prove only
focusing at the -b < 2az condition, and the last row always has at least the same amount of fitting
triples as when z = 1/2. A good upper bound would be the one based upon b² < 4ac, which, even
better, is independant of z, and this is the lower half of each cell of the table, which constitutes
precisely (n + 1)/(2n + 1) of the whole table. Since (n + 1)´ = 1 and (2n + 1)´ = 2, all of these results
would guarantee us, once (E´) is proven, that the repartition of quadratic polynomials' roots with a
real part less than any real constant c is between 3/(16 ∛2) and 1/2 of the whole complex plane, and
hence gives us the probability Pz,c for a random complex quadratic root z, with Im(z) ≠ 0, to have a
real part less than or equal to c > 1/2, which is :

• If Re(z) < c, then Pz,c = 1 – ℓc > 1/2 (and P = ℓc < 3/(16 ∛2) as well iff (E´) is true).
• If Re(z) > c, then Pz,c = ℓc < 1/2 (and P = ℓc > 1 – 3/(16 ∛2) as well iff (E´) is true).

However, if we, instead, look at ₰ z := lim n → ∞ Card { (a, b, c) ∈ ⟦-n, n⟧3 | -b < 2az & b² > 4ac },
then we have a lower bound for ₰ z for z > 1/2 which is (n + 1) / (2n + 1) – (n² – n) / (4n² + 4n + 2),
which tends to 1/4 as n tends to infinity. However, there is another area for z > 1/2 at the left part of
the table, which seemingly can be lowered by (n – 1) n / 2. Hence, we could conjecture that the
minimal area is 1/4 + (n² – n) / (2(2n + 1)²), which tends to 3/8. We, however, think that ₰ A > ₰B iff
A > B. Assuming ₰ is a continuously increasing function, we found out there could exists z0 such
that ₰ z₀ = 1/2 and 1/2 < z0 < 1.

We will, in the future, hopefully provide much more research in order to prove the few conjectures
we have made in this little draft, and maybe even improve our results to eventually constitute a
single actual paper which will accumulate all the results we'll need.

You might also like