An index built to ignore citation counts, almost entirely determined by them

Suppose you have to reduce a scientist’s output to one number. There are two obvious candidates, and both fall apart immediately.

Count the papers. Then the way to win is to publish two hundred forgettable ones. Quantity is cheap; salami-slicing a result across six short papers beats writing one good long one.

Count the citations. Then one landmark paper — or one widely-used software tool, or one review that everybody cites out of politeness — carries an entire career. A researcher with a single 5,000-citation paper and nothing else outranks one with fifty solid papers at eighty citations apiece.

Jorge Hirsch’s 2005 proposal was to ask for both at once. Rank someone’s papers by citation count, most-cited first, and walk down the list until the rank overtakes the citations:

In words: you have an h-index of 20 if and only if twenty separate papers have each been cited at least twenty times. Everything else about your record is invisible to it.

That single test does the work of both counters. Publishing two hundred uncited papers adds nothing, because a paper only counts once it clears its own rank. And a single blockbuster raises h by at most one, no matter how enormous it gets, because it is still just one paper. The metric is deliberately blind to how big your biggest hit is.

Which makes it all the more suspicious that the total citation count — a number dominated by exactly that biggest hit — should predict it at all.

Here are four researchers. Their paper counts differ by a factor of fifteen, their citation totals by a factor of twenty, and their citation curves look nothing alike. Against each, the crudest possible prediction: take the total citation count, square-root it, halve it.

Three careers out of four, predicted to within a rank or two by a rule that never asks how many papers you wrote, how they were distributed, what field you are in, or how long you have been working. That is the thing to explain. And there are really two questions hiding inside it:

Why a square root at all? That one turns out to be almost forced, and the next section disposes of it in a paragraph.

Why one half? That is the real question. The coefficient is not a mathematical constant — it is a measurement, and what it measures is the shape that nearly every citation curve happens to have.

Take the ranked list from § 01 and plot it: rank along the bottom, citations up the side. Now both quantities are visible at once.

N, the total citation count, is the area under the curve — you are adding up one bar per paper. h is where the curve crosses the diagonal y = x, which is exactly the c_i ≥ i test drawn as a picture.

So one is a length and the other is an area. A square-root relation between them is not a coincidence waiting to be explained; it is the only dimensionally sensible thing that could happen. Double every citation count on the page and N quadruples while h merely doubles.

The picture also hands over the one part of this that is a theorem rather than an observation. Each of the top h papers has, by definition, at least h citations, so the h × h square in the corner sits entirely underneath the curve and its area cannot exceed the total:

That much is airtight, for everyone, always. Equality would need a perfectly square profile — every one of the top h papers holding exactly h citations, and not one citation anywhere else. Nobody has that. So the coefficient is somewhere below 1, and the only remaining question is how far below.

Every citation that is not inside the square falls into one of exactly two regions. The head is the excess above the square’s top edge — the handful of papers cited far more than h times. The tail is everything to the right of rank h — papers with a few citations each that will never lift h by even one.

Hirsch’s original observation was that across real careers those leaks are remarkably consistent in size:

Take a = 4 and invert it: the square holds a quarter of the citations, and h = 0.5√N. So the entire content of the coefficient is a claim about how much of a career’s citation mass sits outside the square. Drag the skew slider to see how hard you have to distort a profile to move it.

Put the plot away. Suppose someone publishes p papers a year, and every paper accrues c citations a year, steadily, forever. After t years:

Citations grow quadratically. Papers accumulate linearly and each one’s citations accumulate linearly, so the running total is the area of a triangle.

h grows linearly. A paper of age a has ca citations, so the number of papers with at least x citations is p(t − x/c). Setting that equal to x and solving gives a straight line in time.

Eliminate t between the two and the square root appears on its own, dragging a coefficient you can read straight off:

That expression peaks at √2/2 ≈ 0.707 when the two rates are equal, and sags whenever they are badly mismatched. Realistic values land in the 0.4–0.7 band. Two unrelated arguments — one geometric, one purely about accumulation — converging on the same coefficient is the point at which the rule stops looking like a fluke.

Below are 320 synthetic careers — paper counts from 12 to 350, varied skew, varied productivity — with h computed exactly from each profile rather than fitted. The point is not that reality looks like this. It is that a mechanism this crude already reproduces the law, scatter and all, once its skew is set to match the one number Hirsch actually measured.

A rule of thumb that needed recalibrating every five years would be useless. This one does drift — just very slowly. Work out the case where citations fall off as 1/rank: the crossing condition gives C = h², and the total is then a harmonic sum over the P papers, so

The paper count enters only through a logarithm, and then only under another square root. Going from 30 papers to 300 — a tenfold career — moves the coefficient from about 0.50 to 0.40. That is the entire drift, and it is why the collaborator in Fig. 2 came in 11% under prediction while the mid-career experimentalist was almost exact.

Everything above adds up to a way of using the rule that is more interesting than the rule itself. 0.5√N is a null model: it is what h would be if your citation curve had the shape almost everyone’s does. Compute it, compare it to the real h, and the direction of the gap tells you the shape.

Which settles the theorist from § 02. Twenty-four papers, 6,200 citations, and an h of 22 against a predicted 39 — a 44% miss, far too large for the paper-count drift, which at 24 papers runs the wrong way anyway. That leaves concentration, and the first panel of Fig. 2 shows it plainly: a single paper holds 47% of that career’s citations, and the top two hold 63% between them. Those citations pile up in the head, above the square, where h cannot reach them. The metric is working exactly as designed — it is refusing to let one paper buy a career — and the size of the gap is the measure of how hard it is refusing.

The last thing worth taking from it is a caution about the metric itself. Inverted, the rule reads N ≈ 4h²: every additional point of h-index costs more than the last, and doubling h takes four times the citations.