Harmonic Numbers: The Series That Grows Forever, Slowly

The nth harmonic number, written Hₙ, is the sum of the reciprocals of the whole numbers from 1 to n — and the sum itself is the formula. No algebraic shortcut turns n into Hₙ the way n(n+1)/2 turns n into a triangular number; a closed form does not exist. What does exist is the definition on the left below and, on the right, an approximation so accurate it might as well be one:

Hₙ = 1 + 1/2 + 1/3 + … + 1/n ≈ ln n + γ + 1/(2n) − 1/(12n²)

That is the complete answer to "harmonic number formula." The exact value comes only from adding the terms. The estimate combines the natural logarithm with the Euler–Mascheroni constant γ = 0.5772156649…, whose digits NIST's Digital Library of Mathematical Functions prints at equation 5.2.3, plus two small correction terms. How small? At n = 10 the estimate already agrees with the true value to within a millionth — worked digit by digit below. The rest of this guide covers what those two expressions hide: the exact small values, the full H₁₀ computation, why the infinite series has no sum at all, and the odd fact that the factorial suffers from — and cures — the very same problem.

The first six values, as exact fractions

Adding fractions with different denominators means putting them over a common one, so exact harmonic numbers are ratios over the least common multiple of 1 through n:

Notice the pace: six terms in, the running total has not reached 2.5, and each new term is smaller than the last. Notice also that the fractions are not collapsing toward any pattern — 3/2, 11/6, 25/12, 137/60, 49/20 share no visible structure. The nearest thing mathematics has to a closed form is the identity Hₙ = γ + ψ₀(n+1), where ψ₀ is the digamma function (NIST DLMF equation 5.4.14). It is exact, but it merely renames the difficulty: the digamma function needs numerical evaluation of its own.

H₁₀ by hand, over 2520

Every harmonic-number computation is the same three moves: find the common denominator, convert, add. For n = 10 the least common multiple of 1 through 10 is 2520, and dividing 2520 by each of 1, 2, 3, …, 10 gives the numerators 2520, 1260, 840, 630, 504, 420, 360, 315, 280 and 252.

Adding them in order: 2520 + 1260 = 3780; + 840 = 4620; + 630 = 5250; + 504 = 5754; + 420 = 6174; + 360 = 6534; + 315 = 6849; + 280 = 7129; + 252 = 7381.

So H₁₀ = 7381/2520 = 2.9289682540. The harmonic number calculator returns this same value — it adds every term exactly for any n up to 25,000, switches to the asymptotic expansion beyond that point — by which stage the expansion's first dropped term is around 10⁻²⁰, far beneath the ten decimal places displayed — and always states which method produced the number you are reading.

Testing the estimate at n = 10

γ is defined as the limit of Hₙ − ln n as n grows: the fixed gap that opens up between the harmonic number and the plain logarithm. The definition itself tells you ln n alone will underestimate Hₙ; adding γ closes most of the distance, and the correction terms handle the rest. Assembling the estimate for n = 10, one operation per row:

Exact minus estimate: 2.9289682540 − 2.9289674246 = 0.0000008294 — the promised millionth, after only ten terms. Better still, the discrepancy is explained rather than mysterious. The next term of the full expansion, the first one this estimate drops, is +1/(120n⁴) = 1/1,200,000 ≈ 0.0000008333, which accounts for almost the entire gap. Wolfram MathWorld lists the continuation — +1/(120n⁴), then −1/(252n⁶) — for anyone who needs the error smaller still.

"Harmonic series sum" is a trick question

The infinite series 1 + 1/2 + 1/3 + … has no finite sum. Its partial sums grow past every bound, so the series diverges; the standard classroom proof is the integral test, set out in OpenStax's Calculus Volume 2, §5.3. But an older argument needs nothing beyond comparing fractions. Group the terms in blocks that end at powers of two:

• 1/3 + 1/4 = 7/12, which beats 1/2;
• 1/5 + 1/6 + 1/7 + 1/8 beats 4 × 1/8 = 1/2;
• every following block of doubled length beats 1/2 the same way.

Each doubling of n therefore adds more than 1/2, so after k doublings the sum exceeds 1 + k/2, and no ceiling can hold it. The same argument exposes how slow the growth is: forcing the sum past 10 with this bound takes 1 + k/2 ≥ 10, meaning k = 18 doublings, or 2¹⁸ = 262,144 terms. The true crossing comes earlier — Hₙ first exceeds 10 at n = 12,367 — but "earlier" is relative. Inverting Hₙ ≈ ln n + γ shows that reaching a target sum S takes about e^(S−γ) terms, exponentially many, and passing 20 takes roughly 2.7 × 10⁸ of them.

Where the slow growth earns its keep

Harmonic numbers appear wherever a cost accumulates at rate 1/k across k steps — quicksort's average-case comparison count and the analysis of self-balancing structures both reduce to them. The cleanest example is the coupon collector. Suppose a cereal brand hides one of 10 toys, chosen uniformly, in each box. While you still lack k of the 10, the chance a new box helps is k/10, so you expect 10/k boxes before it does. The total expected count is 10/10 + 10/9 + … + 10/1 = 10 × H₁₀ = 10 × 2.928968254 = 29.2897 — call it 29.3 boxes to complete the set, nearly three times the naive guess of ten, and every box of the excess is the harmonic number's doing.

The name collision: harmonic mean

A harmonic number is not a harmonic mean, though searches conflate the two relentlessly. The harmonic mean is an average: hand it a dataset and it returns n divided by the sum of the reciprocals of your values, always landing between your smallest and largest entries. A harmonic number takes no data at all — just the single index n — and grows without bound. Exactly one identity bridges them: the harmonic mean of the integers 1, 2, …, n equals n ÷ Hₙ. For the first ten integers that is 10 ÷ 2.928968254 = 25200/7381 ≈ 3.414 — a perfectly ordinary average, falling out of the same fraction worked above.

The factorial has the same disease, and the same cure

Take logarithms of a factorial and something familiar appears: ln(n!) = ln 2 + ln 3 + … + ln n. That is again a sum over the first n integers with no elementary closed form, and the remedy is again the Euler–Maclaurin machinery — which for factorials produces Stirling's approximation, given in Abramowitz & Stegun §6.1 as ln(n!) ≈ n ln n − n + ½ ln(2πn) + 1/(12n) − 1/(360n³). The factorial calculator runs the same honest split as the harmonic tool: exact computation where it is cheap (the log-sum term by term up to n = 100, exact integer products for n! itself up to 170) and Stirling's expansion where it is not, with the n = 100 crossover documented right beside the ln(n!) output. For n = 10 it reports 10! = 3,628,800 with ln(10!) ≈ 15.1044 — and setting that beside H₁₀ = 2.9289682540 is this whole post in one comparison: reciprocals sum to almost nothing while logarithms pile up fast, yet both sums surrender to the same expansion trick.

There is something quietly satisfying about a quantity this stubborn. Hₙ refuses every closed form, yet yields completely to patient fractions over a common denominator — every digit above came from arithmetic anyone can redo, and that is the standard the rest of Quanta is held to as well. Publishing exact arithmetic is a promise that it stays exact, so the contact page doubles as an erratum desk for this post: recomputations that disagree are reports, not complaints.

Sources

• Wolfram MathWorld — Harmonic Number
• NIST DLMF §5.2 — Euler's constant γ
• NIST DLMF §5.4 — the psi (digamma) function and ψ(n+1) = Hₙ − γ
• OpenStax Calculus Vol. 2, §5.3 — the integral test and why the harmonic series diverges