Counting Irreducible polynomials: Van Der Waerden, Mahler measure, Chela

We want to understand statistics of “random” integer polynomials. How do we choose the polynomials?

There are several ways to order polynomials — by degree, discriminants, heights of the coefficients etc etc.

Here we fix the degree and focus on the ordering by the coefficient heights.

If we take f(x) =x^{d}+a_{d-1} x^{d-1}+\cdots+a_{1} x+a_{0} , where |a_i| \le N are chosen uniformly, how do the statistics behave?


1. For instance, what fraction of the polynomials are irreducible?
2. If the coefficients are chosen uniformly from 0, 1 or say some bounded set, how likely is it to get a reducible polynomial?
3. What is distribution of Galois groups for a random polynomial? In general how do the invariants of the number fields distribute?

Van Der Waerden:

The main ideas is the relation between the height and a measure of complexity called Mahler measure.

\displaystyle M(f)=|a| \prod_{\left|\alpha_{i}\right| \geq 1}\left|\alpha_{i}\right|=|a| \prod_{i=1}^{d} \max \left\{1,\left|\alpha_{i}\right|\right\} where

\displaystyle f(z)=a\left(z-\alpha_{1}\right)\left(z-\alpha_{2}\right) \cdots\left(z-\alpha_{d}\right) .

We can see that {M(f)} as an integral on the unit circle –“geometric mean” of {f(e^{i \theta})}.

\displaystyle M(f)=\exp \left(\frac{1}{2 \pi} \int_{0}^{2 \pi} \ln \left(\left|f\left(e^{i \theta}\right)\right|\right) d \theta\right) = \lim_{\tau \rightarrow 0} \left(\frac{1}{2 \pi} \int_{0}^{2 \pi}\left|f\left(e^{i \theta}\right)\right|^{\tau} d \theta\right)^{1 / \tau}

It’s easy to see that the multiplicativity property

\displaystyle M(fg) = M(f) M(g)

Different measures of complexity: a) Height

\displaystyle f=a_{0}+a_{1} x+a_{2} x^{2}+\cdots+a_{d} x^{d}, the height {H(f)} is defined to be the largest magnitude of the coefficients:

\displaystyle H(f)=\max _{i}\left|a_{i}\right|

b) Mahler measure

c) Length of {f} is defined to be

{\displaystyle L(f) = \sum_{i=0}^{d}\left|a_{i}\right|}

All of these measures are comparable up to factors that only depend on the degree. That is we have

\displaystyle {d \choose {\lfloor d / 2\rfloor} }^{-1} H(f) \leq M(f) \leq \sqrt{d+1}H(f)

\displaystyle L(f) \leq 2^{d} M(f) \leq 2^{d} L(f)

\displaystyle H(f) \leq L(f) \leq(d+1) H(f)


Main ideas of Van Der Waerden :

Van Der Waerden now counts reducible polynomials by counting all possible products {f(x)g(x)} such that {P(x) =f(x)g(x)} is of height bounded by {N}.

Using the Mahler measure we have

\displaystyle M(f)M(g) =M(P) \le c_d N. Thus for each values of {m} and {n} with {mn \le c_d N}, we can count all the polynomials with {M(f)=m} and {M(g)=n}.

For each value of {m}, and fixing degree of {f} to be {k}, we get the bound

\displaystyle 2 k(2 m+1)^{k-1}(2 n+1)^{d-k} \ll_{k, d} m^{k-1} N^{d-k} \ll_{d} m^{k-1} c^{d-k} \ll_{d} N^{k-1}\left(\frac{N}{m}\right)^{d-k}

and hence summing over {m} get

\displaystyle \ll_{d} 2\sum_{m=1}^{N} m^{k-1}\left(\frac{N}{m}\right)^{d-k}=2 N^{d-k} \sum_{m=1}^{N} \frac{1}{m^{d-2 k+1}}

So we see that the number of reducible polynomial with a factor of degree {k \le d/2 } is at most

\displaystyle 2 N^{d-k} \sum_{m=1}^{N} \frac{1}{m^{d-2 k+1}} .

This quantity is

{\ll_d N^{d-k}} or {\ll_{d} N^{d-k} \log N}

according to {k < d/2} and {k=d/2} respectively.

Summing over the degrees {1\le k \le d/2}, we get that the number of reducible polynomials is {O_{d}(N^{d-1}).}.

This already show that the fraction of reducible polynomials of degree {d} goes to {0} as {N \rightarrow \infty}. I

n fact, we see that the probability of the polynomial being reducible is {O(\frac{1}{N})} for {d >2} . This is the best bound we can expect because polynomial with vanishing constant coefficients already contribute {\frac{1}{2N+1}.}

Note that for {d=2}, the contribution for {k=1} gives {N\log N} and hence the probability is slightly larger {\frac{\log N}{N}}. In fact counting {(x+a)(x+b)} such that the product has height atmost {N}, we get that {|ab| \le N} and hence the reducible polynomials contribute {\displaystyle \sum_{a=1}^{\sqrt N} \frac{N}{a} \asymp N\log N}


Chela’s improved results: Chela gets more precise estimates on the reducible polynomials. In fact, the number of reducible polynomials of degree {d>2} is given by

\displaystyle 2^{d}\left(\zeta(d-1)-\frac{1}{2}+\frac{\alpha_{d}}{2^{d-1}}\right)N^{d-1} \left( 1+ o(1) \right)

where {\alpha_d} is the volume of the region

\displaystyle \left|x_{1}\right| \leq 1, i=1, \ldots, d-1,\left|\sum_{i=1}^{d-1} x_{i}\right| \leq 1

For {d=2}, the number of reducible quadratic polynomials is given by

\displaystyle 2 N \ln (N)+(2 \gamma +o(1))N

Main ideas: Reduce the counting to polynomials which have a linear factor because rest of them contribute {O(N^{d-2}}. Now count polynomials with a fixed linear factor {X+a}. There will be overcounting of polynomials with multiple linear factors, but such polynomials have a degree {2} factor and hence contribute only {N^{d-2}}.

For instance {a=0} counts polynomials with no constant terms– {\sim (2N)^{d-1}}

The cases {|a|=1, |a| =N} and {1<|a|<N} have to be dealt separately.

If {1<|a|<N}, writing conditions for {(x+a)g(x)} to have at most atmost {N}, you get a bound {~\frac{2N}{|a|}^{d-1}}

and thus summing over {a}, we get

\displaystyle 2^dN^{d-1} \sum_{a=2}^{N-1} \frac{1}{|a|^{d-1}} =2^d N^{d-1} \left (\zeta(d-1) -1 + o(1)\right)

{|a|=N}, it’s easy to see that the contribution is small.

Finally {|a|=1} corresponds to counting with the lattice points inside

\displaystyle \{ (a_0, a_1, \cdots, a_{d-1}) : a_{d-1}+\cdots+a_{0} =0 , |a_i| \le N \}

which by rescaling by {N} can be seen to be approximately the volume of the region

\displaystyle \left|x_{1}\right| \leq 1, i=1, \ldots, d-1,\left|\sum_{i=1}^{d-1} x_{i}\right| \leq 1.

Putting all these together we get the count

\displaystyle 2^{d}\left(\zeta(d-1)-\frac{1}{2}+\frac{\alpha_{d}}{2^{d-1}}\right)N^{d-1} \left( 1+ o(1) \right).


(Added Later) Rivin’s idea: To capture reducibility in terms of algebraic relations on the coefficients and then counting points on the corresponding variety.

For instance if a degree {k} polynomial {g(x)} is a factor of {P(x) = (x-\alpha_1)(x-\alpha_2) \cdot (x-\alpha_d)}, then the constant term {c_g} of {g(x)} divides constant term of {P(x)} and is a product of a subset of {k} roots {\alpha_i}. Hence we have the equation

\displaystyle \prod_{1 \leq i_{1} \leq \cdots \leq i_{k} \leq d}\left(\alpha_{i_{1}} \alpha_{i_{2}} \ldots \alpha_{i_{k}}-c_g\right)=0For each choice of the constant term {c_g} and {c_P}, the above gives a polynomial relations between the coefficient of {P}. Apply Schwarz-Zippel type of ideas one bound the number of solutions to the polynomial relations. (To apply Schwarz-Zippel you might have to embed {[-N, N]} inside a finite field {F_p} for {p \asymp N.}) We get a bound {N^{d-2}} for each constant term {c_g|c_P}, and summing over {c_P} we get {N^{d-1}\log N}. (Note that this bound is weaker, but this idea allows us to count other statistics of the polynomials like probability of having a given Galois group)


Remarks:

1. What about other models of ordering the polynomials?
2. What happens if we change how we pick the coefficients? In the case of where each coefficient is picked uniformly from a small set like say {0} and {1}. 3. What about non-uniform distributions on {[-N, N]}? 4. What if we fix some of the coefficients?

Resources:

Van der Waerden, B. L. Die Seltenheit der reduziblen Gleichungen und der Gleichungen mit Affekt. Monatsh. Math. Phys., 43(1):133?147, 1936
Mahler, K. An Application of Jensen?s Formula to Polynomials. Mathematika, vol. 7, no. 2, 1960, pp. 98?100., doi:10.1112/S0025579300001637.
Chela, R. Reducible polynomials. J. Lond. Math. Soc. 38 (1963), 183?188.
Rivin, I. Galois Groups of Generic Polynomials.

Posted in $.

Leave a comment