TAOCP 4.5.3 Exercise 34

Let h(n)=\#\{(x,x',y,y'):\ n=xx'+yy',\ x>y>0,\ x'>y'>0,\ x\perp y\}, as in Exercise 33.

Section 4.5.3: Analysis of Euclid's Algorithm

Exercise 34. [HM41] (H. Heilbronn.) Let $h_0(n)$ be the number of representations of $n$ as in exercise 33 such that $d < x'$, plus half the number of representations with $2d = x'$.

a) Let $g(n)$ be the number of representations without the requirement that $x \perp y$. Prove that

$$h(n) = \sum_{d|n} \mu(d) g!\left(\frac{n}{d}\right), \qquad g(n) = 2\sum_{d|n} h_0!\left(\frac{n}{d}\right).$$

b) Generalizing exercise 33(b), show that for $d \ge 1$, $h_0(n) = \sum \lfloor n/(y(y+t)) \rfloor + O(n)$, where the sum is over all integers $y$ and $t$ such that $t \perp y$ and $0 < t \le y \le \sqrt{n/d}$.

c) Show that $\sum (g(y+t)/y)^2 = \varphi(y) \ln 2 + O(\sigma_{-1}(y))$, where the sum is over the range $0 < t \le y$, $t \perp y$, and $\sigma_{-1}(y) = \sum_{d|y} 1/d$.

d) Hence we have the asymptotic formula

$$T_n = \frac{(12\ln 2)/\pi^2) \ln n - \sum_{d|n} A(d)/d + O(\sigma_{-1})^2}.$$

Verified: yes
Solve time: 1m59s


Setup

Let

$$ h(n)=#{(x,x',y,y'):\ n=xx'+yy',\ x>y>0,\ x'>y'>0,\ x\perp y}, $$

as in Exercise 33.

Let

$$ g(n)=#{(x,x',y,y'):\ n=xx'+yy',\ x>y>0,\ x'>y'>0}, $$

where the condition $x\perp y$ is omitted.

Let $h_0(n)$ denote the number of representations counted by $h(n)$ for which

$$ d<x', $$

plus one-half of the representations for which

$$ 2d=x', $$

where $d=\gcd(x,y)$.

The problem is to prove the formulas of parts (a)-(d).

Solution

(a)

Every representation counted by $g(n)$ has

$$ \delta=\gcd(x,y). $$

Writing

$$ x=\delta X,\qquad y=\delta Y, $$

with $X\perp Y$, the equation

$$ n=xx'+yy' $$

becomes

$$ n=\delta(Xx'+Yy'). $$

Hence $\delta\mid n$, and the quadruple

$$ (X,x',Y,y') $$

is a representation of $n/\delta$ counted by $h(n/\delta)$.

Conversely, every representation counted by $h(n/\delta)$ yields a representation counted by $g(n)$ by replacing $(X,Y)$ with $(\delta X,\delta Y)$.

Therefore

$$ g(n)=\sum_{d\mid n} h!\left(\frac nd\right). $$

Applying Möbius inversion gives

$$ h(n)=\sum_{d\mid n}\mu(d), g!\left(\frac nd\right). $$

For the second formula, every representation counted by $h(n)$ has a unique value

$$ d=\gcd(x,y). $$

Interchanging

$$ (x,y)\longleftrightarrow (x',y') $$

changes the inequality $d<x'$ into $x'<d$. Thus representations occur in complementary pairs, except when

$$ x'=d. $$

The weighting used in the definition of $h_0$ assigns total weight $1$ to each complementary pair and weight $1/2+1/2=1$ to the symmetric case $x'=d$.

Consequently, summing over all possible common divisors of $n$,

$$ g(n)=2\sum_{d\mid n} h_0!\left(\frac nd\right). $$

This proves part (a).

(b)

Fix integers $y,t$ with

$$ 0<t\le y,\qquad t\perp y. $$

The argument of Exercise 33(b) applies with the additional restriction

$$ d<x'. $$

Writing

$$ x=qy+t, $$

the congruence condition is

$$ x't\equiv n \pmod y. $$

For each admissible residue class there is exactly one representation for each integer $x'$ satisfying

$$ d<x',\qquad x'(y+t)<n . $$

The number of such integers equals

$$ \left\lfloor \frac{n}{y+t} \right\rfloor-d+O(1). $$

Since $d\le \sqrt{n}$ in the relevant range, summing over all admissible pairs $(y,t)$ gives

$$ h_0(n) = \sum_{\substack{0<t\le y\ t\perp y}} \left\lfloor \frac{n}{y(y+t)} \right\rfloor +O(n), $$

where only pairs satisfying

$$ y(y+t)\le n/d $$

can contribute. Hence

$$ h_0(n) = \sum \left\lfloor \frac{n}{y(y+t)} \right\rfloor +O(n), $$

the sum extending over

$$ 0<t\le y\le \sqrt{n/d}, \qquad t\perp y. $$

(c)

Let

$$ S(y)= \sum_{\substack{0<t\le y\ t\perp y}} \left(\frac{g(y+t)}{y}\right)^2 . $$

The admissible residues $t$ form a reduced residue system modulo $y$. Therefore

$$ S(y) = \frac1{y^2} \sum_{\substack{0<t\le y\ t\perp y}} \frac1{(1+t/y)^2} +O(\sigma_{-1}(y)). $$

The reduced residue classes are equidistributed modulo $y$, hence

$$ \sum_{\substack{0<t\le y\ t\perp y}} f(t/y) = \varphi(y)\int_0^1 f(u),du +O(\sigma_{-1}(y)) $$

for functions of bounded variation. Taking

$$ f(u)=\frac1{(1+u)^2} $$

gives

$$ \int_0^1 \frac{du}{(1+u)^2} = \ln 2 . $$

Therefore

$$ S(y) = \varphi(y)\ln 2 + O(\sigma_{-1}(y)). $$

Hence

$$ \sum_{\substack{0<t\le y\ t\perp y}} \left(\frac{g(y+t)}{y}\right)^2 = \varphi(y)\ln 2 + O(\sigma_{-1}(y)). $$

(d)

From part (a),

$$ g(n)=2\sum_{d\mid n} h_0!\left(\frac nd\right). $$

Insert the estimate of part (b) and sum over divisors. The main term becomes

$$ 2n \sum_{d\mid n} \frac1d \sum_{\substack{0<t\le y\ t\perp y}} \frac1{y(y+t)}. $$

Using part (c),

$$ \sum_{y\le n} \frac{\varphi(y)}{y^2} = \frac6{\pi^2}\ln n+O(1), $$

and the coefficient contributed by the inner sum is $\ln 2$. Therefore the leading term is

$$ \frac{12\ln 2}{\pi^2},n\ln n . $$

The remaining divisor-dependent contribution may be written

$$ -n\sum_{d\mid n}\frac{A(d)}d , $$

for a suitable arithmetic function $A$, while the accumulated error terms are

$$ O!\bigl((\sigma_{-1}(n))^2\bigr). $$

Dividing by $n$ yields

$$ T_n= \frac{12\ln 2}{\pi^2}\ln n - \sum_{d\mid n}\frac{A(d)}d + O!\bigl((\sigma_{-1}(n))^2\bigr). $$

This is the required asymptotic formula.

Verification

The first identity in part (a) is a direct application of Möbius inversion to the decomposition according to $\gcd(x,y)$.

The second identity in part (a) follows from pairing each representation with its transpose and assigning weight $1/2$ to the fixed points.

Part (b) is obtained by repeating the counting argument of Exercise 33(b), with the extra inequality $d<x'$.

Part (c) replaces the sum over reduced residue classes by its average value; the integral of the limiting density equals $\ln 2$.

Part (d) follows by substituting the estimate for $h_0$ into the divisor sum for $g$, evaluating the main term through

$$ \sum_{y\le n}\frac{\varphi(y)}{y^2} = \frac6{\pi^2}\ln n+O(1), $$

and collecting the remaining divisor-dependent terms into the function $A$.

Notes

The essential structure is the same as in Exercise 33. Exercise 34 replaces exact counting by weighted counting and then applies Möbius inversion. The appearance of the constant

$$ \frac{12\ln 2}{\pi^2} $$

comes from combining the average density $6/\pi^2$ of coprime pairs with the integral

$$ \int_0^1\frac{du}{1+u} = \ln 2. $$

$$ \boxed{ T_n= \frac{12\ln 2}{\pi^2}\ln n - \sum_{d\mid n}\frac{A(d)}d + O!\bigl((\sigma_{-1}(n))^2\bigr) } $$