arXiv reader

Bézout bounds and rational points on quasianalytic graphs

Armin Rainer

arXiv
2610.21725v1 · PDF · 15 October 2026
Categories
math.CA, math.FA, math.LO, math.NT
Licence
CC BY 4.0 The site shows the text.

Abstract

For functions f : [0,1] \to [0,1] in quasianalytic Denjoy-Carleman classes \mathcal C_M, we establish explicit bounds for the number of intersection points of the graph \Gamma(f) with algebraic curves of degree at most d. The bounds are given in terms of the Bang degree determined by the weight sequence M, the degree d, and the norming constant of \Gamma(f) with respect to the space of real bivariate polynomials of degree at most d. We give instances, where the bounds are at most polynomial in d and, consequently, the graph \Gamma(f) contains few rational points: the number of rational points of height at most H is bounded by a polynomial in \log H.

Items 39

Definition

Definition 2.3 ((\mu^{|N},M_0)-smooth functions)

RLIH

Let \mu^{|\infty} = (\mu_j)_{j\ge 1} be an admissible weight and M_0>0. Let N \in \mathbb{N}_{\ge 1} \cup \{\infty\}. A function f : I \to \mathbb{R}, defined on a compact interval I \subseteq \mathbb{R}, is called (\mu^{|N},M_0)-smooth if f is of class \mathcal{C}^{N} and

\|f^{(j)}\|_{I} \le M_j, \quad 0 \le j < N +1.
Definition

Definition 2.4

EYNW

Let \mu^{|\infty}= (\mu_j)_{j\ge 1} be an admissible weight and M_0,a,b>0. We say that (\mu^{|\infty},M_0,a,b) is admissible, if

\sum_{j=j_0+1}^\infty \frac{1}{\mu_j} > a\, e, \quad \text{ where } j_0 = \lceil \log (\tfrac{M_0}{b})\rceil_{\mathbb{N}}.

In that case, we define the integer

\mathbf{N}= \mathbf{N}(\mu^{|\infty},M_0,a,b):= 2\, \sup\Big\{n \in \mathbb{N} : \sum_{j=j_0+1}^n \frac{1}{\mu_j} < a\, e\Big\}

and say that \mathbf{N} is the degree associated with (\mu^{|\infty},M_0,a,b).

Remark

Remark 2.5

HQRD

(1) By definition, for r,s>0, (r\, \mu^{|\infty},s\, M_0,a,b) is admissible if and only if (\mu^{|\infty},M_0,r\, a, \frac{b}{s}) is admissible, and

\mathbf{N}(r \, \mu^{|\infty},s\, M_0,a,b) = \mathbf{N}(\mu^{|\infty},M_0,r\, a, \tfrac{b}{s}).

(2) Assuming admissibility, the integer \mathbf{N}(\mu^{|\infty},M_0,a,b) is increasing in a and decreasing in b.

Theorem

Theorem 3.1

SD4X

Let \mu^{|\infty} = (\mu_j)_{j\ge 1} be an admissible weight satisfying (?) with constant H\ge 1. Let M_0>0, B_0\ge b_0 > 0, and d \in \mathbb{N}_{\ge 1}. Assume that (\mu^{|\infty},\tfrac{1}{2} e^{4H\overline m}\, d^{2d} B_0,2,b_0) is admissible and let

\mathbf{N} = \mathbf{N}(\mu^{|\infty},\tfrac{1}{2} e^{4H\overline m} d^{2d} B_0,2,b_0).

Let f : [0,1] \to [0,1] be (\mu^{|\mathbf{N}},M_0)-smooth and p \in \mathcal{P}_2^d satisfy \|p\|_{\Gamma(f)} \ge b_0 and \|p\|_{[0,1]^2} \le B_0. Then

\mathbf{z}(p,f)\le \mathbf{N},

where \mathbf{z}(p,f) denotes the total number of zeros of p(x,f(x)), for x \in [0,1], counted with multiplicities.

Proposition

Proposition 3.2

FSIQ

Let I \subseteq \mathbb{R} be a compact interval. Let \mu^{|\infty} = (\mu_j)_{j \ge 1} be an admissible weight and M_0>b_0>0. Assume that (\mu^{|\infty},M_0,|I|,b_0) is admissible and let

\mathbf{N} = \mathbf{N}(\mu^{|\infty},M_0,|I|,b_0).

Let f : I \to \mathbb{R} be a (\mu^{|\mathbf{N}}, M_0)-smooth function satisfying \|f\|_I \ge b_0. Then the total number of zeros of f in I (counted with multiplicities) is finite and bounded by \mathbf{N}.

Proposition

Proposition 3.3

OS21

Let \mu^{|\infty} = (\mu_j)_{j\ge 1} be an admissible weight satisfying (?) with constant H\ge 1. Let M_0>0. If f : [0,1] \to [0,1] is (\mu^{|N},M_0)-smooth and p \in \mathcal{P}_2^d, then h := p \circ \overline f : [0,1] \to \mathbb{R}, where \overline f : [0,1] \to [0,1]^2, \overline f(x) = (x,f(x)), is (2\, \mu^{|N}, \tfrac{1}{2} e^{4H\overline m}\, d^{2d}\, \|p\|_{[0,1]^2})-smooth.

Lemma

Lemma 3.4

76V1

Let \mu^{|\infty} = (\mu_j)_{j\ge 1} be an admissible weight satisfying (?) with constant H\ge 1. Then the sequence b_j :=a_j^j satisfies

b_{j_1}b_{j_2} \cdots b_{j_\ell} \le H^\ell \, b_k

whenever j_1 + \cdots + j_\ell = k and j_i \ge 1 for all 1 \le i \le \ell.

Proof of Lemma 3.4

By (?), we have b_i \le H \, b_{j}^{i/j} if i\le j. Thus, if j_1 + \cdots + j_\ell = k and j_i \ge 1 for all 1 \le i \le \ell, then

b_{j_1}b_{j_2} \cdots b_{j_\ell} \le H\, b_k^{j_1/k}\, H\, b_k^{j_2/k} \cdots H\, b_k^{j_\ell/k} = H^\ell \, b_k

as claimed.

Lemma

Lemma 3.5

2D6Q

For all p \in \mathcal{P}_n^d and j \in \mathbb{N},

\sup_{x \in [0,1]^n}\|d^j p(x)\|_{L^j(\mathbb{R}^n,\mathbb{R})} \le (2nd^2)^j \|p\|_{[0,1]^n}.
Proof of Lemma 3.5

Markov’s inequality implies

\|\partial^\alpha p\|_{[0,1]^n} \le (2d^2)^{|\alpha|}\, \|p\|_{[0,1]^n},

for all \alpha \in \mathbb{N}^n. Then, for v^i= (v^i_1,\ldots,v^i_n) \in \mathbb{R}^n with |v^i|\le 1 and x \in [0,1]^n,

\begin{aligned} |(d^jp)(x)(v^1,\ldots,v^j)| &= \Big|\sum_{i_1=1, \ldots,i_j=1}^n \partial_{i_1}\cdots\partial_{i_j} p(x) \, v^1_{i_1}\cdots v^j_{i_j}\Big| \le (2nd^2)^{j}\, \|p\|_{[0,1]^n}. \end{aligned}

The lemma is proved.

Proposition

Proposition 3.6 (Faà di Bruno’s formula; cf.)

76Y8

Let f : U \to V and g : V \to W be \mathcal{C}^k maps, where U,V,W are open subsets of normed spaces. Then, for x \in U,

\frac{d^k(g \circ f)(x)}{k!} = \operatorname{sym} \,\sum \frac{d^jg(f(x))}{j_1!\cdots j_k!} \Big( \frac{d^1 f(x)}{1!}\Big)^{j_1} \cdots \Big( \frac{d^k f(x)}{k!}\Big)^{j_k},

where j = j_1+ \cdots+j_k, the sum is taken over all j_1,\ldots,j_k such that j_1 + 2j_2 + \cdots + kj_k = k, and \operatorname{sym} denotes symmetrization of a multilinear operator.

Proposition

Proposition 3.7

KV8J

In the setting of ?, the composite h = p \circ \overline f : [0,1] \to \mathbb{R} is (2\, \mu^{|N}, \tfrac{1}{2} (4 d^2 H \hat m)^d\, \|p\|_{[0,1]^2})-smooth, where

\hat m := \max \Big\{ M_0\sqrt{1+\tfrac{1}{M_1^2}},1\Big\}.
Proof of Proposition 3.7

It suffices to modify the estimate (?) in the following way:

\begin{aligned} \frac{|h^{(k)}(x)|}{k!} &\le \|p\|_{[0,1]^2} \, b_k\, \sum_{j \le d} \frac{(4\, d^2 H \hat m)^j}{j_1!\cdots j_k!} \le (4\, d^2 H \hat m)^d\, \|p\|_{[0,1]^2} \, b_k\, \sum \frac{j! }{j_1!\cdots j_k!}. \end{aligned}

Then the proof of ? implies the statement.

Remark

Remark 3.8

R47V

An example, where d is small, is the number of fixed points of f. Indeed, the polynomial p(X,Y) = X-Y satisfies \|p\|_{[0,1]^2} = 1 and \deg p =1. Assume that (\mu^{|\infty},M_0(1+\frac{1}{M_1}),1,b_0) is admissible and let

\mathbf{N} = \mathbf{N}(\mu^{|\infty}, M_0(1+\tfrac{1}{M_1}) ,1,b_0).

Let f : [0,1] \to [0,1] be (\mu^{|\mathbf{N}},M_0)-smooth with \sup_{x\in [0,1]} |x-f(x)|\ge b_0. By a direct computation, the composite function h(x) = p(x,f(x)) = x-f(x), x \in [0,1], is (\mu^{|\mathbf{N}}, M_0(1+\tfrac{1}{M_1}))-smooth. (Alternatively, one can use ? which however gives weaker bounds in this case.) Moreover, \|h\|_{[0,1]}\ge b_0. By ?, the total number of fixed points of f (counted with multiplicities) is finite and bounded by \mathbf{N}.

Definition

Definition 4.1

TH85

Let \mathcal{V} \subseteq C^0([0,1]^n,\mathbb{R}) be a finite dimensional subspace. A compact subset Z \subseteq [0,1]^n is said to be \mathcal{V}-norming if there is a constant C>0 such that

\|v\|_{[0,1]^n} \le C \, \|v\|_Z, \quad v \in \mathcal{V}.

The infimum of all such constants C is denoted by N_{\mathcal{V}}(Z) and called the \mathcal{V}-norming constant of Z. This is an inequality of Remez type on a set of zero measure.

Theorem

Theorem 4.2

OEIQ

Let \mu^{|\infty} = (\mu_j)_{j\ge 1} be an admissible quasianalytic weight satisfying (?) with constant H\ge1. Let M_0>0. If f : [0,1] \to [0,1] is a transcendental (\mu^{|\infty},M_0)-smooth function, then, for each p \in \mathcal{P}_2^d\setminus \{0\},

\mathbf{z}(p,f) \le \eta_{\mu^{|\infty}}\big( 2d \log d + \mathbf{e}(p,f) + 4H \overline m -\log 2 \big).

Consequently,

\mathbf{z}_d(f) \le \eta_{\mu^{|\infty}} \big( 2d \log d + \mathbf{e}_d(f) + 4H \overline m -\log 2 \big), \quad d \ge 1.
Proof of Theorem 4.2

By (the proof of) ?,

\mathbf{z}(p,f) \le \mathbf{N}(\mu^{|\infty},\tfrac{1}{2} e^{4 H \overline m} d^{2d}\,\|p\|_{[0,1]^2},2,\|p\|_{\Gamma(f)}).

Now, by the properties of \mathbf{N} (see ?),

\begin{aligned} \mathbf{N}(\mu^{|\infty},\tfrac{1}{2} e^{4 H \overline m} d^{2d}\,\|p\|_{[0,1]^2},2,\|p\|_{\Gamma(f)}) &= \mathbf{N}\Big(\mu^{|\infty},\tfrac{1}{2} e^{4 H \overline m} d^{2d}\,\frac{\|p\|_{[0,1]^2}}{\|p\|_{\Gamma(f)}},2,1\Big) \\ &= \eta_{\mu^{|\infty}}\big(2d \log d + \mathbf{e}(p,f) + 4H \overline m -\log 2\big) \end{aligned}

implying (4.2). Since \eta_{\mu^{|\infty}} is increasing, (4.2) is immediate from (4.2).

Corollary

Corollary 4.3

5R8Z

Let f : [0,1] \to [0,1] be real analytic and transcendental. There exist C,\rho>0 (depending on f) such that

\mathbf{z}_d(f) \le 2\, e^{2e\rho+1} \Big(2d\log d + \mathbf{e}_d(f) + 4 C\sqrt{1 + \tfrac{1}{C^2\rho^2}} +1 -\log 2\Big), \quad d\ge 1.
Used by Proof, Remark 4.7
Proof of Corollary 4.3

In this case, H=1, M_0 = C, and M_1 = C\rho. The bound follows from (4.2) and (?).

Corollary

Corollary 4.4

7S5T

Let f : [0,1] \to [0,1] be transcendental and the restriction of an entire function. There exists C=C(f)>0 such that

\mathbf{z}_d(f) \le 12\,d\log d + 6\, \mathbf{e}_d(f) + C, \quad d\ge 1.
Proof of Corollary 4.4

Choose \tau>0 such that 2e\tau +1 \le \log 3 and apply (4.3) for \rho = \tau.

Theorem

Theorem 4.5

CIC0

Let f be a non-polynomial entire function on \mathbb{C}^n that is real on \mathbb{R}^n. Assume that the graph \Gamma(f|_{\mathbb{R}^n}) of f|_{\mathbb{R}^n} intersects [0,1]^{n+1} in a subset Z of real dimension n. Then there exists a sequence of positive integers d_j \to \infty and a sequence of positive reals \epsilon_j \to 0 such that

\log N_{\mathcal{P}_{n+1}^{d_j}}(Z) \le d_j^{2+\epsilon_j},\quad \text{ for all } j.

The sequence (d_j) depends only on f and the sequence (\epsilon_j) only on f, n, and the n-dimensional Hausdorff measure of Z.

Corollary

Corollary 4.6

57S0

Let f : [0,1] \to [0,1] be transcendental and the restriction of an entire function that is real on \mathbb{R}. There exists a sequence of positive integers d_j \to \infty, a sequence of positive reals \epsilon_j \to 0, and a constant C>0 such that

\mathbf{z}_{d_j}(f) \le 6\, d_j^{2+ \epsilon_j} + 12\, d_j \log d_j + C, \quad \text{ for all } j.
Proof of Corollary 4.6

Combine ? with ?.

Remark

Remark 4.7

XYZT

For each \zeta \in (0,1), there exists an analytic function f defined in a neighborhood of [0,1] and a sequence of positive integers d_j \to \infty satisfying

\mathbf{z}_{d_j}(f) \ge e^{d_j^\zeta}, \quad \text{ for all } j.

See [Comte:2018aa, Remark 5.3]. In view of (4.3), we conclude that

\mathbf{e}_{d_j}(f) \ge c_1\, e^{d_j^\zeta} - 2 d_j \log d_j - C_2, \quad \text{ for all } j,

for constants c_1 = c_1(f) \in (0,1) and C_2 = C_2(f) >0.

Definition

Definition 4.8 (Denjoy weights)

8SCH

Let s\in [0,\infty) and set k = \lceil s\rceil \in \mathbb{N} and \sigma = k-s \in [0,1). We define

(\mu_s)_j := j L_1(e_kj) L_2(e_k j) \cdots L_{k-1}(e_k j) (L_k(e_k j))^{1-\sigma}, \quad j\ge 1.

Then \mu_s^{|\infty} := ((\mu_s)_j)_{j \ge 1} is an admissible quasianalytic weight for all s \in [0,\infty). (Quasianalyticity follows from the integral test, in view of (Proof).) Since (\mu_s)_j/j is increasing, so is the sequence a_j defined in (?), i.e., (?) holds with H=1.

The sequences \mu_s^{|\infty} are called Denjoy weights. Note that \mu_0^{|\infty} coincides with the analytic weight \mu_{\operatorname{an}}^{|\infty}.

Proposition

Proposition 4.9

RTZY

Let s \ge 0 and a\ge 1. Then \eta_{a\,\mu_s^{|\infty}} : \mathbb{N}_{\ge 1} \to \mathbb{N} is bounded by a polynomial if and only if s \in [0,1].

For s=k \in \mathbb{N} and m \in \mathbb{N}_{\ge 1},

\frac{E_{k+1}(f_k(a,m)-1)}{e_k}-1 \le \tfrac{1}{2}\eta_{a\, \mu_{k}^{|\infty}}(m) \le \frac{E_{k+1}(f_k(a,m)+1)}{e_k}-1,

where f_k(a,m) := 2a e+ L_{k+1}(e_k(m+1)). In particular,

\begin{aligned} e^{2ae-1}(m+1)-1 \le &\tfrac{1}{2}\eta_{a\, \mu_{0}^{|\infty}}(m) \le e^{2ae+1} (m+1) - 1, \\ e^{e^{2ae -1}-1} (m+1)^{e^{2ae -1}}-1 \le &\tfrac{1}{2}\eta_{a\, \mu_{1}^{|\infty}}(m) \le e^{e^{2ae +1}-1} (m+1)^{e^{2ae +1}}-1. \end{aligned}
Proof of Proposition 4.9

Let us first prove (4.9). Set A^k_n := \sum_{j=1}^n \frac{1}{(\mu_k)_j}, for n \ge 1, and A^k_0 := 0. Consider g(t) = L_{k+1}(e_k t), for t \ge 1. Then

\begin{aligned} g'(t) &= \frac{1}{t L_1(e_kt) L_2(e_k t) \cdots L_k(e_k t)} \end{aligned}

is decreasing, i.e., g is concave. Thus

g(j+1) - g(j) \le g'(j) \le g(j) - g(j-1),

where the left inequality holds for j\ge 1 and the right for j \ge 2. Summing over j, we find

g(n+1) - g(1) \le A^k_n \quad \text{ and } \quad A^k_n -A^k_1 \le g(n)-g(1)

which implies

L_{k+1}(e_k(n+1)) \le A^k_n \le L_{k+1}(e_k n) + 1, \quad n \ge 1.

For x > 1 define the positive integer n(x) by

A^k_{n(x)} < x \le A^k_{n(x)+1}.

By (Proof) and (?),

\frac{E_{k+1}(x-1)}{e_k}- 1 \le n(x) < \frac{E_{k+1}(x)}{e_k}- 1.

By (?), we have

\sum_{j=m+1}^{n(x)} \frac{1}{(\mu_k)_j} < x- A^k_{m} \le \sum_{j=m+1}^{n(x)+1} \frac{1}{(\mu_k)_j}

so that, setting x = 2ae +A^k_{m}, we get

\sum_{j=m+1}^{n} \frac{1}{(\mu_k)_j} < 2ae \le \sum_{j=m+1}^{n+1} \frac{1}{(\mu_k)_j}

with n=n(2ae + A^k_{m}), since 2ae+ A^k_{m}> 1. Thus, by (Proof) and (?),

\frac{E_{k+1}(2ae-1 +L_{k+1}(e_k (m+1)))}{e_k}-1 \le n \le \frac{E_{k+1}(2ae+1 +L_{k+1}(e_k (m+1)))}{e_k}-1.

This implies (4.9), since n=\frac{1}{2} \eta_{a\,\mu_k^{|\infty}}(m).

Now (?) and (?) easily follow from (4.9). Since (\mu_{s})_j \le (\mu_{t})_j for all j if s \le t and thus \eta_{a\, \mu_{s}^{|\infty}} \le \eta_{a\, \mu_{t}^{|\infty}}, we conclude that \eta_{a\, \mu_{s}^{|\infty}} is bounded by a polynomial, for all s\in [0,1].

To see that \eta_{a\, \mu_{s}^{|\infty}} cannot be bounded by a polynomial if s>1 it is enough to show this for s \in (1,2) and a=1. Fix such s, let \sigma = 2 -s \in (0,1), and consider g(t) = (\log\log(e^e t))^{\sigma}, for t\ge 1. Since

g'(t) = \frac{\sigma}{t \log(e^e t) (\log\log(e^et))^{1-\sigma}}

and g is concave, we find similarly to (Proof),

(\log \log(e^e (n+1)))^{\sigma} -1 \le \sigma A^s_n \le (\log \log(e^e n))^{\sigma} - 1 +\tfrac{\sigma}{e}, \quad n\ge 1,

where A^s_n = \sum_{j=1}^n \frac{1}{(\mu_s)_j} and noting that A^s_1 = \frac{1}{e}. As above, let n(x), for x>1, be the positive integer satisfying A^s_{n(x)} < x \le A^s_{n(x)+1}. Then, by (?),

e_2(n(x)+1) \ge E_2((\sigma x +1 -\tfrac{\sigma}{e})^{\frac{1}{\sigma}})

and, setting n = n(2e + A^s_m) =\frac{1}{2} \eta_{\mu_{s}^{|\infty}}(m), we find

\begin{aligned} e_2(n+1) &\ge E_2((2\sigma e + \sigma A^s_m +1 -\tfrac{\sigma}{e})^{\frac{1}{\sigma}}) \ge E_2(((2 e -\tfrac{1}{e})\sigma + (L_2(e_2 (m+1)))^{\sigma} )^{\frac{1}{\sigma}}). \end{aligned}

Suppose for contradiction that n can be bounded by a polynomial in m. Then there exist C\ge 1 and N \in \mathbb{N}_{\ge 1} such that, for all m,

E_2(((2 e -\tfrac{1}{e})\sigma + (L_2(e_2 (m+1)))^{\sigma} )^{\frac{1}{\sigma}}) \le C (e_2(m+1))^N,

or equivalently,

(2 e -\tfrac{1}{e})\sigma \le (L_2(C (e_2(m+1))^N))^\sigma - (L_2(e_2 (m+1)))^{\sigma}.

Setting z=e_2(m+1), we have

\begin{aligned} L_2(C (e_2(m+1))^N)& = L_2(C z^N) = \log(\log C + N \log z) \\ &= \log\Big(N\log z \Big(1+\frac{\log C}{N \log z}\Big) \Big) \\ &= \log\log z + \log N + o(1) \end{aligned}

as z \to \infty. By the mean value theorem, the right-hand side of (?) equals

\begin{aligned} (\log\log z + \log N + o(1))^\sigma - (\log\log z)^{\sigma} = \sigma \zeta^{\sigma-1} (\log N + o(1)), \end{aligned}

for some \zeta between \log\log z + \log N + o(1) and \log\log z. Since \sigma \in (0,1), (?) tends to zero as z \to \infty which contradicts (?).

Corollary

Corollary 4.10

DLNJ

Let s \in [0,1]. If f : [0,1] \to [0,1] is a transcendental (\mu_s^{|\infty},M_0)-smooth function, then \mathbf{z}_d(f) is bounded by a polynomial in d and \mathbf{e}_d(f).

Proof of Corollary 4.10

This is a consequence of ? and ?.

Corollary

Corollary 5.1

1H8O

Let s \in [0,1] and let f : [0,1] \to [0,1] be a transcendental function of class \mathcal{C}_{M_s} associated with the Denjoy weight \mu_s^{|\infty} (with (M_s)_0=1) such that \mathbf{e}_d(f) is bounded by a polynomial in d. Then the number of rational points of height at most H on the graph of f is bounded by a polynomial in \log H.

Proof of Corollary 5.1

By [Pila:2006ab, Proposition 2.4], the set \Gamma(f)(\mathbb{Q};H) of rational points of height at most H in \Gamma(f) is contained in the union of at most 6B (\frac{(d+1)(d+2)}{2})^C\, H^{8/(d+3)} algebraic curves of degree d, where d \ge 5 and B,C are constants depending on the class of f. (The Denjoy–Carleman class \mathcal{C}_{M_s} is contained in the class M(B,C) of mild functions used in [Pila:2006ab].)

Let d = \lfloor \log H \rfloor, assuming that H \ge e. Then \Gamma(f)(\mathbb{Q};H) is contained in the union of at most

6B\, 3^C\, e^8 (\log H)^{2C}

algebraic curves of degree d, using that \frac{(d+1)(d+2)}{2} \le 3 (\log H)^2 and H^{8/(d+3)} \le e^8. Thus, the number of points in \Gamma(f)(\mathbb{Q};H) is bounded by

6B\, 3^C\, e^8 (\log H)^{2C} \cdot \mathbf{z}_{\lfloor \log H \rfloor}(f)

which is bounded by a polynomial in \log H, by ?, provided that \mathbf{e}_d(f) is bounded by a polynomial in d. Thus ? is proved.

Definition

Definition 5.2 (The functions f_\xi)

U908

Let \mu^{|\infty} = (\mu_j)_{j \ge 1} be any admissible weight (not necessarily quasianalytic) and M_0=1 such that the associated function

\omega_M(t) := \sup_{k \in \mathbb{N}} \log \frac{t^k}{M_k}, \quad t>0,

satisfies \omega_M(t) = o(t) and \omega_M(2t) = O(\omega_M(t)) as t \to \infty. Let \xi = (\xi_n)_{n \ge 1} be a sequence of independent random variables, each uniformly distributed on [-1,1]. Define

f_\xi(x) := \tfrac{1}{2} + \sum_{n\ge 1} \xi_n c_n \cos(2\pi n x), \quad \text{ where } c_n := \frac{1}{4 n^2 e^{\omega_M(2\pi n)}}.
Remark

Remark 5.3

8E4T

Note that \omega_M(t) = o(t) as t \to \infty means that \mathcal{C}_M contains the real analytic functions as a proper subclass. This condition is satisfied if the associated sequence a_j (see (?)) tends to \infty. Moreover, \omega_M(2t) = O(\omega_M(t)) as t \to \infty is a technical condition which holds provided that \liminf_{j \to \infty} \frac{\mu_{\ell j}}{\mu_j} >1 for some positive integer \ell; see [BMM07, Lemma 12]. Clearly, these conditions are satisfied by all Denjoy weights. Moreover, \log t = o(\omega_M(t)) as t \to \infty by definition.

Proposition

Proposition 5.4

SM4E

The family \{f_\xi\} has the following properties:

  1. f_\xi : [0,1] \to [0,1] belongs to \mathcal{C}_M.

  2. Almost surely, f_\xi is not real analytic. More generally, almost surely f_\xi does not belong to any class \mathcal{C}_{L} with \omega_M(t) = o(\omega_L(t)) as t \to \infty, where L is any weight sequence as in ?.

  3. Almost surely, \mathbf{e}_d(f_\xi) is bounded by a polynomial in d of degree 7 (whose coefficients depend on \xi).

Theorem

Theorem 5.5

JV5U

Let M = M_s a Denjoy weight for s \in (0,1] with (M_s)_0 =1. Then, for the family \{f_\xi\}, the number of rational points of height at most H on the graph of f_\xi is almost surely bounded by a polynomial in \log H.

Proof of Theorem 5.5

This follows from ? and ?. Note that f_\xi is almost surely transcendent, because if there exists a nonzero polynomial p \in \mathbb{R}[X,Y] such that p(x,f_\xi(x)) = 0, for all x \in [0,1], then f_\xi is real analytic. Indeed, the 1-periodic function f_\xi belongs to \mathcal{C}_{M_s}(\mathbb{R}) and, by quasianalyticity, p(x,f_\xi(x))=0, for all x \in \mathbb{R}. By [Siciak70, Theorem 2] or [Malgrange66, Proposition VI.3.11], f_\xi is real analytic.

Source and licence

The text on this page is from arXiv:2610.21725v1 by Armin Rainer, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/2610/2610.21725.jsonl.