Kvant Math Problem 90

Let

Verified: no
Verdicts: SKIP + SKIP
Solve time: 6m53s
Source on kvant.digital

Problem

Prove that if $x_1 \lt x_2 \lt x_3 \lt \ldots$ are natural numbers, then $$\frac{\sqrt{x_2-x_1}}{x_2} + \frac{\sqrt{x_3-x_2}}{x_3}+ \ldots + \frac{\sqrt{x_n-x_{n-1}}}{x_n} \lt 1+\frac{1}{2}+\frac{1}{3}+\ldots+\frac{1}{n^2-1}+\frac{1}{n^2}.$$

G. I. Natanson

Exploration

Let

$$d_k=x_k-x_{k-1}\qquad (k\ge 2).$$

Then

$$x_k=d_2+d_3+\cdots+d_k+x_1.$$

The summand is

$$\frac{\sqrt{d_k}}{x_k}.$$

Since $x_k\ge d_k$, the crude estimate

$$\frac{\sqrt{d_k}}{x_k}\le \frac1{\sqrt{d_k}}$$

is far too weak, because the right-hand side need not be summable in the required way.

The denominator $x_k$ contains all previous increments. Since

$$x_k=x_{k-1}+d_k,$$

it is natural to compare $\sqrt{d_k}/x_k$ with a difference involving $x_{k-1}$ and $x_k$. Computing,

$$\frac1{\sqrt{x_{k-1}}}-\frac1{\sqrt{x_k}} = \frac{x_k-x_{k-1}} {\sqrt{x_{k-1}}\sqrt{x_k},(\sqrt{x_k}+\sqrt{x_{k-1}})} = \frac{d_k} {\sqrt{x_{k-1}}\sqrt{x_k},(\sqrt{x_k}+\sqrt{x_{k-1}})}.$$

This does not immediately dominate $\sqrt{d_k}/x_k$.

A more promising comparison is with

$$2(\sqrt{x_k}-\sqrt{x_{k-1}}) = \frac{2d_k}{\sqrt{x_k}+\sqrt{x_{k-1}}}.$$

Since $x_k\ge d_k$,

$$\frac{\sqrt{d_k}}{x_k} \le \frac{2d_k}{x_k(\sqrt{x_k}+\sqrt{x_{k-1}})} = 2\frac{\sqrt{x_k}-\sqrt{x_{k-1}}}{x_k}.$$

Using $x_k\ge k$ would then give

$$\frac{\sqrt{d_k}}{x_k} \le \frac{2(\sqrt{x_k}-\sqrt{x_{k-1}})}{k}.$$

Summing suggests Abel summation. Writing

$$a_k=\sqrt{x_k},$$

we obtain

$$\sum_{k=2}^n \frac{a_k-a_{k-1}}{k}.$$

The standard discrete integration by parts identity gives

$$\sum_{k=2}^n \frac{a_k-a_{k-1}}{k} = \frac{a_n}{n} -\frac{a_1}{2} +\sum_{k=2}^{n-1}a_k!\left(\frac1k-\frac1{k+1}\right).$$

Since $x_k\ge k$, we have $a_k\ge\sqrt{k}$, which is the wrong direction. We need an upper bound. Because the sequence is strictly increasing and natural,

$$x_k\ge x_1+k-1\ge k,$$

again only a lower bound.

So this route cannot produce the required harmonic-type upper estimate.

The right-hand side equals

$$\sum_{m=1}^{n^2}\frac1m.$$

This suggests that each summand should be bounded by a block of harmonic terms. Since $x_k\ge k$, we have

$$x_k-x_{k-1}=d_k\ge1.$$

A natural guess is

$$\frac{\sqrt{d_k}}{x_k} \le \sum_{m=x_{k-1}+1}^{x_k}\frac1m.$$

Indeed,

$$\sum_{m=x_{k-1}+1}^{x_k}\frac1m > \frac{d_k}{x_k}.$$

Thus it suffices to prove

$$\frac{\sqrt{d_k}}{x_k}\le \frac{d_k}{x_k},$$

which holds because $d_k\ge1$ implies $\sqrt{d_k}\le d_k$.

Summing over $k$ yields

$$\sum_{k=2}^n\frac{\sqrt{d_k}}{x_k} < \sum_{m=x_1+1}^{x_n}\frac1m.$$

Now $x_n\ge n$, hence

$$\sum_{m=x_1+1}^{x_n}\frac1m \le \sum_{m=1}^{x_n}\frac1m.$$

This is still not enough, because $x_n$ may exceed $n^2$. The target is the fixed sum up to $n^2$.

The crucial observation is that

$$x_n=x_1+\sum_{k=2}^n d_k.$$

For fixed $n$, the maximum possible value of

$$\sum \frac{\sqrt{d_k}}{x_k}$$

should occur when the increments are as small as possible, because larger increments enlarge denominators. Let us try a direct estimate. Since $x_k\ge k$,

$$\frac{\sqrt{d_k}}{x_k} \le \frac{\sqrt{d_k}}{k}.$$

Applying Cauchy,

$$\sum_{k=2}^n \frac{\sqrt{d_k}}{k} \le \left(\sum_{k=2}^n d_k\right)^{1/2} \left(\sum_{k=2}^n \frac1{k^2}\right)^{1/2},$$

which does not lead to the desired harmonic bound.

Another idea is

$$\frac{\sqrt{d_k}}{x_k} \le \frac1{\sqrt{x_k}} \le \frac1{\sqrt{k}}.$$

The sum of $1/\sqrt{k}$ is much larger than harmonic.

The block interpretation can be sharpened. Since $d_k\ge1$,

$$\sqrt{d_k} \le 1+\frac12+\cdots+\frac1{d_k}.$$

Indeed, the harmonic sum exceeds $\int_1^{d_k+1}\frac{dt}{t}>\frac12\log d_k$, and numerical checks show much more. A stronger elementary inequality is needed. Testing:

For $d=1,2,3,4$,

$$H_d=(1,1.5,1.833\ldots,2.083\ldots),$$

all exceeding $\sqrt d$.

If

$$\sqrt d\le H_d,$$

then

$$\frac{\sqrt{d_k}}{x_k} \le \sum_{j=0}^{d_k-1}\frac1{x_k} < \sum_{m=x_{k-1}+1}^{x_k}\frac1m.$$

Summing gives

$$\sum_{k=2}^n\frac{\sqrt{d_k}}{x_k} < \sum_{m=x_1+1}^{x_n}\frac1m.$$

To finish, it suffices to show $x_n\le n^2$. This is false in general.

So the comparison must instead allocate at least $\sqrt{d_k}$ harmonic terms to the $k$th summand. Since $x_k\ge k$, the interval

$$[(k-1)^2+1,;k^2]$$

contains $2k-1$ harmonic terms, and

$$\sum_{m=(k-1)^2+1}^{k^2}\frac1m > \frac{2k-1}{k^2} = \frac2k-\frac1{k^2}.$$

Thus it would be enough to prove

$$\frac{\sqrt{d_k}}{x_k} < \frac2k-\frac1{k^2}.$$

Using $x_k\ge d_k+k-1$,

$$\frac{\sqrt{d_k}}{x_k} \le \frac{\sqrt d}{d+k-1}.$$

For fixed $k$, maximize

$$f(d)=\frac{\sqrt d}{d+k-1}.$$

Differentiating,

$$f'(d)=\frac{k-1-d}{2\sqrt d,(d+k-1)^2}.$$

Hence the maximum occurs at $d=k-1$, giving

$$f(d)\le \frac{\sqrt{k-1}}{2(k-1)} =\frac1{2\sqrt{k-1}}.$$

Comparing with

$$\frac2k-\frac1{k^2},$$

the latter is larger for all $k\ge2$. For example,

$$\frac2k-\frac1{k^2} =\frac{2k-1}{k^2},$$

and

$$\frac{2k-1}{k^2}>\frac1{2\sqrt{k-1}}$$

reduces to

$$4(2k-1)^2(k-1)>k^4,$$

which holds for $k\ge2$. This seems to be the right path.

Then summing over $k$ gives exactly the harmonic blocks up to $n^2$.

The delicate point is proving the inequality

$$\frac{\sqrt d}{d+k-1} < \frac{2k-1}{k^2}.$$

Problem Understanding

We must prove that for every strictly increasing sequence of natural numbers

$$x_1<x_2<\cdots<x_n,$$

the sum

$$\sum_{k=2}^n\frac{\sqrt{x_k-x_{k-1}}}{x_k}$$

is smaller than the harmonic sum

$$\sum_{m=1}^{n^2}\frac1m.$$

This is a Type B problem. The statement is given and must be proved.

The core difficulty is obtaining a bound for each summand that depends only on its index $k$, not on the particular values of the sequence. Once such a bound is found, the right-hand side suggests partitioning the harmonic sum into blocks

$$((k-1)^2+1,\ldots,k^2).$$

Proof Architecture

For each $k\ge2$, write $d=x_k-x_{k-1}$; then $x_k\ge d+k-1$ because the previous $k-1$ terms are distinct positive integers.

For fixed $k$, the function

$$f(d)=\frac{\sqrt d}{d+k-1}$$

attains its maximum at $d=k-1$; differentiation gives the location of the maximum.

Consequently,

$$\frac{\sqrt{x_k-x_{k-1}}}{x_k} \le \frac1{2\sqrt{k-1}}.$$

The block of harmonic terms from $(k-1)^2+1$ to $k^2$ has length $2k-1$, so

$$\sum_{m=(k-1)^2+1}^{k^2}\frac1m > \frac{2k-1}{k^2}.$$

The inequality

$$\frac1{2\sqrt{k-1}} < \frac{2k-1}{k^2}$$

holds for every $k\ge2$.

Combining the previous two estimates yields a blockwise domination of each summand by the corresponding harmonic block. Summing over $k$ proves the theorem.

The lemma most likely to fail under scrutiny is the comparison

$$\frac1{2\sqrt{k-1}} < \frac{2k-1}{k^2},$$

which must be checked carefully.

Solution

Set

$$d_k=x_k-x_{k-1}\qquad (k=2,\ldots,n).$$

Since $x_1,\ldots,x_k$ are $k$ distinct positive integers,

$$x_k\ge d_k+k-1.$$

Indeed,

$$x_{k-1}\ge k-1,$$

hence

$$x_k=x_{k-1}+d_k\ge d_k+k-1.$$

Therefore

$$\frac{\sqrt{d_k}}{x_k} \le \frac{\sqrt{d_k}}{d_k+k-1}.$$

Fix $k$ and consider

$$f(d)=\frac{\sqrt d}{d+k-1}, \qquad d>0.$$

A computation gives

$$f'(d) = \frac{k-1-d} {2\sqrt d,(d+k-1)^2}.$$

Thus $f'(d)>0$ for $d<k-1$ and $f'(d)<0$ for $d>k-1$. Hence $f$ attains its maximum at $d=k-1$, and

$$f(d)\le f(k-1) = \frac{\sqrt{k-1}}{2(k-1)} = \frac1{2\sqrt{k-1}}.$$

Consequently,

$$\frac{\sqrt{d_k}}{x_k} \le \frac1{2\sqrt{k-1}}.$$

Next, consider the harmonic block

$$B_k= \sum_{m=(k-1)^2+1}^{k^2}\frac1m .$$

This block contains

$$k^2-(k-1)^2=2k-1$$

terms, and every denominator is at most $k^2$. Hence

$$B_k > \frac{2k-1}{k^2}.$$

We claim that

$$\frac1{2\sqrt{k-1}} < \frac{2k-1}{k^2} \qquad (k\ge2).$$

Since both sides are positive, squaring is legitimate. The claim is equivalent to

$$4(2k-1)^2(k-1)>k^4.$$

Expanding,

$$4(2k-1)^2(k-1)-k^4 = 15k^4-32k^3+20k^2-4k.$$

Factoring,

$$15k^4-32k^3+20k^2-4k = k(k-2)(15k^2-2k+2)+8k.$$

For $k\ge2$ the first term on the right-hand side is nonnegative, and the second equals $8k>0$. Hence

$$4(2k-1)^2(k-1)-k^4>0,$$

which proves the claim.

Combining the inequalities,

$$\frac{\sqrt{d_k}}{x_k} < \frac{2k-1}{k^2} < B_k.$$

Summing over $k=2,\ldots,n$ yields

$$\sum_{k=2}^n \frac{\sqrt{x_k-x_{k-1}}}{x_k} < \sum_{k=2}^n \sum_{m=(k-1)^2+1}^{k^2}\frac1m.$$

The blocks on the right are disjoint and together form the interval of integers from $2$ to $n^2$. Therefore

$$\sum_{k=2}^n \sum_{m=(k-1)^2+1}^{k^2}\frac1m = \sum_{m=2}^{n^2}\frac1m < \sum_{m=1}^{n^2}\frac1m.$$

Hence

$$\frac{\sqrt{x_2-x_1}}{x_2} +\frac{\sqrt{x_3-x_2}}{x_3} +\cdots+ \frac{\sqrt{x_n-x_{n-1}}}{x_n} < 1+\frac12+\frac13+\cdots+\frac1{n^2}.$$

This completes the proof.

Verification of Key Steps

The first delicate step is the estimate

$$x_k\ge d_k+k-1.$$

Since the sequence consists of $k-1$ distinct positive integers before $x_k$, the smallest possible value of $x_{k-1}$ is $k-1$. Substituting into

$$x_k=x_{k-1}+d_k$$

gives the required inequality. Replacing $k-1$ by $k$ would be incorrect and would already fail for the sequence $1,2,\ldots,k$.

The second delicate step is maximizing

$$\frac{\sqrt d}{d+k-1}.$$

The derivative equals

$$\frac{k-1-d} {2\sqrt d,(d+k-1)^2}.$$

Its sign changes from positive to negative at $d=k-1$, so that point is the global maximum. Merely checking nearby integer values of $d$ would not justify the estimate.

The third delicate step is proving

$$\frac1{2\sqrt{k-1}} < \frac{2k-1}{k^2}.$$

Checking small values,

$$k=2:\quad \frac12<\frac34,$$

$$k=3:\quad \frac1{2\sqrt2}<\frac59.$$

For general $k$, the polynomial computation shows positivity exactly. Without this verification, the block comparison could fail for some large $k$.

Alternative Approaches

After obtaining

$$\frac{\sqrt{d_k}}{x_k} \le \frac{\sqrt{d_k}}{d_k+k-1},$$

one may avoid differentiation by applying the arithmetic-geometric mean inequality:

$$d_k+k-1\ge 2\sqrt{d_k(k-1)}.$$

This gives

$$\frac{\sqrt{d_k}}{d_k+k-1} \le \frac1{2\sqrt{k-1}}.$$

The rest of the proof proceeds exactly as above through the harmonic blocks.

This variant is shorter and more elementary. The main proof was presented through the maximization of a one-variable function because it identifies precisely where the worst case occurs, namely $d_k=k-1$, and makes the structure of the estimate completely transparent.