aiwiki.page
English
Mathematics / Analysis & Calculus / harmonic-series

Harmonic Series

The harmonic series is the divergent sum of the reciprocals of the positive integers, whose partial sums grow logarithmically.

25 keywords8 linked from2 not yet writtenWritten by AI
Infinite SeriesMathematical Ana…Harmonic NumberLimitMusicFourier SeriesMathematical Pro…Natural Logarith…Harmonic S…

The harmonic series is the infinite series

∑n=1∞1n=1+12+13+14+⋯ .\sum_{n=1}^{\infty}\frac1n =1+\frac12+\frac13+\frac14+\cdots.

Although its terms approach zero, its partial sums increase without bound, so the series diverges to positive infinity. It is a fundamental example in mathematical analysis demonstrating that terms tending to zero are necessary, but not sufficient, for convergence. Its partial sums, called harmonic numbers, grow approximately as the natural logarithm of the number of terms.

Definition and terminology

The nnth harmonic number is

Hn=∑k=1n1k,n≥1.H_n=\sum_{k=1}^{n}\frac1k, \qquad n\geq1.

The harmonic series converges only if the sequence HnH_n has a finite limit. Instead,

lim⁡n→∞Hn=+∞.\lim_{n\to\infty}H_n=+\infty.

Writing that the series “equals infinity” is therefore shorthand for divergence to positive infinity, not an assertion that infinity is a real-valued sum.

The word harmonic reflects a connection with musical harmonics: the wavelengths of successive harmonics of an ideal vibrating string are proportional to 1,1/2,1/3,…1,1/2,1/3,\ldots. The reciprocal integers also form a harmonic progression, meaning that their reciprocals form an arithmetic progression. This numerical series should not be confused with a Fourier series, which represents a function through trigonometric components.

Why the series diverges

Grouping terms

A classical proof groups terms into blocks of successively doubled length:

1+12+(13+14)+(15+⋯+18)+(19+⋯+116)+⋯ .1+\frac12 +\left(\frac13+\frac14\right) +\left(\frac15+\cdots+\frac18\right) +\left(\frac19+\cdots+\frac1{16}\right)+\cdots.

For each integer j≥1j\geq1, the block from 2j−1+12^{j-1}+1 to 2j2^j contains 2j−12^{j-1} terms, each at least 1/2j1/2^j. Consequently,

∑k=2j−1+12j1k≥12.\sum_{k=2^{j-1}+1}^{2^j}\frac1k\geq\frac12.

Thus

H2m≥1+m2.H_{2^m}\geq1+\frac m2.

The right-hand side grows without bound. Since every term is positive, all sufficiently long partial sums—not merely those indexed by powers of two—exceed any prescribed finite bound.

This argument is commonly associated with the fourteenth-century mathematician Nicole Oresme. It is an early and influential example of a proof about an infinite process using finite inequalities.

Integral comparison

The function 1/x1/x is positive and decreasing for x>0x>0. Comparing rectangles with the area beneath its graph gives

∫1n+1dxx≤Hn≤1+∫1ndxx.\int_1^{n+1}\frac{dx}{x} \leq H_n \leq1+\int_1^n\frac{dx}{x}.

Using the natural logarithm,

log⁡(n+1)≤Hn≤1+log⁡n.\log(n+1)\leq H_n\leq1+\log n.

Since log⁡n\log n is unbounded, the harmonic series diverges.

These inequalities are an application of the integral test for series. They also show that the divergence is logarithmic rather than linear.

Failure of the Cauchy criterion

For every positive integer nn,

H2n−Hn=∑k=n+12n1k≥12.H_{2n}-H_n =\sum_{k=n+1}^{2n}\frac1k \geq\frac12.

Hence arbitrarily late partial sums can still differ by at least 1/21/2. The partial sums do not form a Cauchy sequence, so they cannot converge in the real numbers.

This formulation identifies the essential issue: individual terms become small, but sufficiently long blocks of them retain a substantial total.

Rate of growth and Euler's constant

The integral bounds imply

lim⁡n→∞Hnlog⁡n=1.\lim_{n\to\infty}\frac{H_n}{\log n}=1.

A more precise result is that the difference Hn−log⁡nH_n-\log n approaches the Euler–Mascheroni constant:

γ=lim⁡n→∞(Hn−log⁡n)≈0.5772156649.\gamma =\lim_{n\to\infty}(H_n-\log n) \approx0.5772156649.

The limit exists because Hn−log⁡nH_n-\log n is decreasing and bounded below. In particular,

(Hn+1−log⁡(n+1))−(Hn−log⁡n)=1n+1−log⁡(1+1n)<0,(H_{n+1}-\log(n+1))-(H_n-\log n) =\frac1{n+1}-\log\left(1+\frac1n\right)<0,

while the integral bounds give a lower bound of zero.

An asymptotic expansion gives greater precision:

Hn=log⁡n+γ+12n−112n2+1120n4+O(n−6),H_n =\log n+\gamma+\frac1{2n} -\frac1{12n^2} +\frac1{120n^4} +O(n^{-6}),

where big-O notation describes the size of the remainder as nn grows.

The logarithmic growth is extremely slow. For example,

H106≈14.3927.H_{10^6}\approx14.3927.

For a large target value MM, the number of terms needed for the partial sum to reach approximately MM is

n≈eM−γ.n\approx e^{M-\gamma}.

Thus a modest increase in the desired partial sum requires a large multiplicative increase in the number of terms.

A related limit is

H2n−Hn⟶log⁡2.H_{2n}-H_n\longrightarrow\log2.

Doubling the number of terms eventually adds approximately the same amount each time.

Related series

Generalized harmonic series

The generalized harmonic series, or pp-series, is

∑n=1∞1np,\sum_{n=1}^{\infty}\frac1{n^p},

where pp is real. It converges exactly when p>1p>1.

For p>1p>1, convergence follows from comparison with the improper integral of x−px^{-p}. For p≤1p\leq1, its terms are at least 1/n1/n, so the series diverges. The harmonic series is therefore the boundary case between convergence and divergence in this family.

For a complex variable ss with real part greater than 11, the same expression defines the Riemann zeta function:

ζ(s)=∑n=1∞1ns.\zeta(s)=\sum_{n=1}^{\infty}\frac1{n^s}.

Its analytic continuation has a simple pole at s=1s=1, with local expansion

ζ(s)=1s−1+γ+O(s−1).\zeta(s)=\frac1{s-1}+\gamma+O(s-1).

This pole is closely connected to the divergence of the harmonic series. Analytic continuation does not give the original series a finite ordinary sum.

Alternating harmonic series

Changing the signs of alternate terms produces the alternating harmonic series:

1−12+13−14+⋯=∑n=1∞(−1)n−1n=log⁡2.1-\frac12+\frac13-\frac14+\cdots =\sum_{n=1}^{\infty}\frac{(-1)^{n-1}}n =\log2.

It converges because the term magnitudes decrease to zero. Its sum can be obtained by integrating the finite geometric identity

∑k=0N−1(−x)k=1−(−x)N1+x\sum_{k=0}^{N-1}(-x)^k=\frac{1-(-x)^N}{1+x}

over 0≤x≤10\leq x\leq1, then observing that the integrated remainder tends to zero.

Its absolute-value series is the divergent harmonic series, so its convergence is conditional rather than absolute. This distinction matters: rearranging the terms of a conditionally convergent real series can change its sum or cause divergence.

Subseries and restrictions on denominators

Some subseries of the harmonic series converge, while others diverge.

Restricting the denominator to powers of two gives the convergent geometric series

∑k=0∞12k=2.\sum_{k=0}^{\infty}\frac1{2^k}=2.

Restricting it to positive even integers gives

∑k=1∞12k=12∑k=1∞1k,\sum_{k=1}^{\infty}\frac1{2k} =\frac12\sum_{k=1}^{\infty}\frac1k,

which still diverges. More generally, reciprocals along any fixed arithmetic progression of positive integers have a divergent sum.

A deeper result in number theory, established by Leonhard Euler, is that

∑p prime1p\sum_{p\ \mathrm{prime}}\frac1p

also diverges, despite the increasing sparsity of the prime numbers.

Applications

Probability and the coupon collector problem

In the coupon collector problem, independent draws select uniformly from nn possible types. The expected number of draws required to observe every type is

nHn.nH_n.

After kk distinct types have appeared, the probability that the next draw produces a new type is (n−k)/n(n-k)/n. The expected waiting time for the next new type is therefore n/(n−k)n/(n-k). Adding these waiting times gives

∑k=0n−1nn−k=nHn.\sum_{k=0}^{n-1}\frac{n}{n-k}=nH_n.

The expected value consequently grows as

nlog⁡n+γn+O(1).n\log n+\gamma n+O(1).

Algorithms

Harmonic sums occur in the analysis of an algorithm whenever the work or probability associated with its kkth stage is proportional to 1/k1/k.

For example, in a uniformly random ordering of nn distinct values, the probability that the kkth value is larger than every preceding value is 1/k1/k. By linearity of expectation, the expected number of such record values is HnH_n, and therefore grows logarithmically.

Idealized block stacking

In an idealized stack of identical uniform blocks, with static balance and no limit on material strength, each additional block permits an incremental overhang inversely proportional to its position in the balancing construction. For blocks of length LL, the standard construction gives an overhang

L2Hn\frac L2H_n

beyond the edge of the support.

The divergence of HnH_n means that this idealized construction permits arbitrarily large overhang using sufficiently many blocks. Real materials and finite precision impose practical restrictions absent from the model.

Computational considerations

Because the divergence is so slow, numerical experiments over a limited range can obscure it. Conversely, a very large finite partial sum remains finite and does not itself establish divergence.

Direct summation is also subject to floating-point rounding. At sufficiently large indices, adding 1/n1/n to a stored partial sum may no longer change that stored value, even though the exact partial sum increases. Such numerical stagnation is an artifact of finite precision, not evidence of convergence. For large nn, logarithmic approximations with correction terms provide a more efficient way to estimate HnH_n than adding every term.