Zipf's Law

Open any novel and count words. Rank them by how often they appear. The most common word — typically the in English — might appear 60,000 times. The second-most-common? About 30,000. The third? About 20,000. The pattern is almost too clean: the nth most common word appears roughly 1/n times as often as the most common one.

This is Zipf's law, named after linguist George Kingsley Zipf, who popularized it in his 1949 book Human Behavior and the Principle of Least Effort. Zipf did not discover it — the mathematician Felix Auerbach had noticed similar patterns in city populations decades earlier — but Zipf saw it everywhere and made it famous.

The law is empirical. Nobody handed down a theorem. It just shows up: in English, French, Mandarin, and ancient Latin; in city populations and earthquake magnitudes; in the incomes of the richest individuals and the access frequencies of web pages. The rank-frequency curve bends in the same way across wildly different systems — and no single explanation has won consensus.

Type or paste any text below and click Analyze. The demo ranks every word by frequency, then plots rank on the horizontal axis and count on the vertical axis — both on a log scale. If Zipf's law holds, the points will line up along a straight line with slope −1.

<div class="controls">
  <textarea id="input" rows="5" placeholder="{{placeholder}}">To be or not to be that is the question whether tis nobler in the mind to suffer the slings and arrows of outrageous fortune or to take arms against a sea of troubles and by opposing end them to die to sleep no more and by a sleep to say we end the heartache and the thousand natural shocks that flesh is heir to tis a consummation devoutly to be wished to die to sleep to sleep perchance to dream</textarea>
  <button id="analyze" type="button">{{btn_analyze}}</button>
  <button id="clear" type="button" class="ghost">{{btn_clear}}</button>
</div>
<div id="stats" class="stats"></div>
<canvas id="chart" width="480" height="300"></canvas>
<div id="table-wrap">
  <table id="freq-table">
    <thead><tr><th>{{th_rank}}</th><th>{{th_word}}</th><th>{{th_count}}</th><th>{{th_expected}}</th></tr></thead>
    <tbody id="tbody"></tbody>
  </table>
</div>

* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; margin: 0; color: #222; }
.controls { display: flex; flex-direction: column; gap: .5rem; margin-bottom: .7rem; }
textarea { width: 100%; border: 1px solid #cdd9e3; border-radius: 8px; padding: .5rem .7rem;
           font: 14px system-ui, sans-serif; resize: vertical; color: #222; background: #f7fafc; }
button { font: 600 14px system-ui, sans-serif; padding: .45rem .9rem; border: 1px solid #1d3557;
         background: #1d3557; color: #fff; border-radius: 8px; cursor: pointer; align-self: flex-start; }
button.ghost { background: #fff; color: #1d3557; }
.stats { font-size: .88rem; color: #555; margin-bottom: .5rem; min-height: 1.2em; }
canvas { display: block; width: 100%; max-width: 480px; border: 1px solid #dde4eb;
         border-radius: 8px; background: #f7fafc; margin-bottom: .7rem; }
#table-wrap { max-height: 160px; overflow-y: auto; border: 1px solid #dde4eb; border-radius: 8px; }
table { width: 100%; border-collapse: collapse; font-size: .82rem; }
thead th { background: #e8eef3; padding: .3rem .6rem; text-align: left; position: sticky; top: 0; }
tbody tr:nth-child(even) { background: #f7fafc; }
td { padding: .25rem .6rem; }
td.word { font-family: ui-monospace, monospace; color: #1d3557; }
td.count { color: #0a7d33; font-weight: 600; }
td.exp { color: #888; }

// Code not found

Try short texts (a few sentences) and long ones (paste a whole paragraph). Notice how the straight line emerges more clearly the more text you add. The slope shown tells you how close reality is to the ideal −1. Natural language tends to land between −0.8 and −1.2.

Zipf's law is not proven from first principles. It is an open empirical regularity — one of the most reproducible patterns in science, yet one whose mechanism remains disputed. Here are the leading explanations and why none is definitive:

• Least effort (Zipf's own theory): speakers minimize effort by reusing a small set of very common words; listeners prefer variety. The 1/rank distribution is a compromise. Critics note this is vague and hard to test rigorously.
• Random typing (Miller 1957): George Miller showed that a monkey randomly hitting keys on a typewriter, with a spacebar, produces letter sequences whose "word" frequencies follow a power law. This suggests Zipf's law might be a mathematical artifact of any process that segments a stream into chunks — not a deep property of language at all.
• Preferential attachment: in models like Simon (1955), each new word token is copied from an existing token proportional to how often that token already appears. Rich get richer. This produces power laws mechanically, but the exponent is sensitive to parameters.
• Maximum entropy: given a constraint on the average information per word, the distribution that maximizes entropy is a power law. This is elegant but assumes the constraint without explaining it.
• Status in complexity theory: Zipf's law is not NP-hard, not undecidable, not a Millennium Prize problem. It is an open question in statistical physics and linguistics — we can describe it precisely, verify it empirically, and simulate it, but we cannot yet derive it from a universally accepted mechanism.

The deepest puzzle is universality. Why does the same 1/rank curve appear in city sizes, protein expression levels, and GitHub repository stars? Each domain has a completely different generative process. A truly satisfying explanation must work for all of them — and none of the theories above does.

The 1/rank curve is not just pretty — it has real engineering and scientific consequences:

• Text compression: because a few words account for most tokens, short codes for common words (like in Huffman coding, explored in our compression article) yield enormous savings. Zipf's law quantifies exactly how skewed the distribution is.
• Natural language processing: language models, spell-checkers, and search engines all maintain word frequency tables. Knowing the distribution follows a power law helps allocate vocabulary size, set smoothing parameters, and estimate out-of-vocabulary rates.
• Information retrieval: the inverse-document-frequency component of TF-IDF is a direct consequence of Zipf's law — common words carry less information.
• Economics and inequality: city population distributions (Gabaix 1999) and individual wealth distributions both fit Zipf-like power laws. Understanding the exponent helps model how inequality evolves.
• Biology: messenger RNA expression levels in cells follow a Zipf-like distribution. Rare transcripts vastly outnumber common ones, which affects sequencing depth requirements.
• Anomaly detection: a text or dataset that violates Zipf's law is suspicious. Benford's law (a cousin of Zipf's) is used in forensic accounting to detect fabricated numbers.

At its core, Zipf's law is the mathematical signature of heavy-tailed distributions — the statistical backbone behind everything from PageRank to the structure of the internet.

Zipf's law is deceptively simple: rank anything by frequency and a straight line appears on a log-log plot. It takes minutes to verify on any large text. Yet after nearly a century of study, we still do not have a single agreed explanation for why it holds across such wildly different systems.

That gap — between a pattern so reliable you can bet on it and a mechanism too elusive to pin down — is what makes Zipf's law one of the most tantalizing open questions in quantitative science. It is not a Millennium Prize problem, and no complexity class captures it. It lives in the borderland between empirical regularity and mathematical mystery.

The next time you write a document, remember: the words you reach for most often are obeying a law that neither you nor anyone else has fully explained.