Zipf at the word, the phrase, and in chemistry
What’s claimed and how well supported
- Labels on this page
- REAL ×1 SPECULATIVE ×3
- Strongest label used
- REAL
- Main caveat (quoted)
- Grant’s correspondence is HIS and ◆ SPECULATIVE.
Counted automatically from the labels printed on this page. Counts are a reading aid, not a grade of the page.
Terms on this page (1), plain-language definitions
- Valency (ling.)
- Arity: how many actants a predicate takes; formal vs semantic valency can diverge.
Definitions quoted from the Chem for linguists primer glossary.
Date: 2026-10-03 PT
Status: Corpus note. Same-domain counts below are REAL inside the field that measured them. Grant’s correspondence is HIS and SPECULATIVE.
Audience: syntax/semantics, with the chemical rank plots kept as chemistry.
Does not: announce a joint field, revise Bridge A, or treat a linguistic remark as a chemical measurement.
Not Bridge A. Bridge A is concentration windows ↔ valencyArity: how many actants a predicate takes; formal vs semantic valency can diverge. / saturation. This note is ranked frequency. The two strands are not the same claim.
Abstract. Sourced. Ranked word frequency is the classic object: the r-th word has frequency falling about as 1/r (exponent near −1 on a log–log rank plot), not a law fixed at exactly −1. A large English-text fit uses a different exponent, near 2, on the frequency spectrum, and that value moves with the text. Fixed-length n-grams are shallower than −1; a combined word-plus-n-gram list, and mixed-length non-overlapping phrases, are the constructions that have been reported as close to −1 over a long rank range. Opened chemistry sources do not show a Zipf law over molecules as species. They show power laws for molecular features, for reaction-network degree, and a Zipf-like rank plot for BPE tokens of SMILES/SELFIES. One opened abstract says text-fragment frequencies and molecular-fragment frequencies have the same structure, and does not name Zipf. Element abundances, database compound-popularity counts, citation counts, and mass spectra were not opened, so they are not claimed here.
Grant’s hypothesis (HIS, SPECULATIVE, not a finding). Zipf power laws in language sit at the phrase level, and in chemistry at the molecule level, suggesting phrase ↔ molecule and word ↔ atom. He coined “phrasisized” for a lexicalized phrase-level unit. None of the opened papers uses that word. The phrase-level linguistic result is partly in the sources below. The molecule-level chemical Zipf law, and the word ↔ atom pairing, are not findings of those sources.
A prior check of his main Grok Bot chat (mail, calendar, and Krazy To-Go only) found no comments from him on Zipf; this note does not claim those Grok-app comments were recovered. Separately, as supplied for this note and not re-read from the app: his only message in the Grok app thread “Zipf’s Law Holds for Phrases” (10 Jun) is “https://www.nature.com/articles/srep12209 please help me understand this”. The app’s reply, not Grant, said phrases behave like “meaningful molecules” of language. That line is the app’s gloss. It is not his wording, and it is not a chemical result.
Two later remarks of his are recorded in §2 beside the Williams et al. 2015 sentence. Neither is a finding. One is an open question about phrasal verbs. The other is an open analogy (quarks, and maybe anyons). This page does not test either to a conclusion.
1. Classic word law, and the caveats
What is ranked is the word type, by how often it occurs in a corpus. Rank r = 1 is the most frequent word. Classic Zipf is that ranked word frequency, about 1/r, not exactly −1.
Opened:
-
Ha, Sicilia-Garcia, Ming, and Smith, “Extension of Zipf’s Law to Words and Phrases” (COLING 2002), https://aclanthology.org/C02-1117.pdf. They state Zipf (1949) as f = k/r, a log–log slope of −1. They also record that the slope was found to vary slightly from −1, and they give Mandelbrot’s form f = k/(r+α)^β. In that notation they say legal texts have smaller β (≈ 0.9). On the Brown corpus (1 million words) the unigram curve drops below the slope-−1 line for rank greater than about 5000. They say the single-word law is valid only for high-frequency words.
-
Piantadosi, “Zipf’s word frequency law in natural language: A critical review and future directions” (Psychonomic Bulletin & Review, 2014), https://sites.socsci.uci.edu/~lpearl/courses/readings/Piantadosi2014InPress_ZipfsLaw.pdf (DOI 10.3758/s13423-014-0585-6). He states the classic claim, citing Zipf 1936 and 1949, as f(r) ∝ 1/r^α for α ≈ 1, with r the frequency rank of a word. Mandelbrot’s shift is f(r) ∝ 1/(r+β)^α for α ≈ 1 and β ≈ 2.7. He treats this as an approximation (“near-Zipfian”). Fitting rank and frequency on the same corpus correlates the errors and makes the curve look simpler than it is. Residuals after the Zipf–Mandelbrot fit are structured, including for the most frequent words, so no simple equation is strictly right. He sets Heap’s law aside and does not review it. He notes that it is apparently unclear whether n-grams follow Zipf’s law, citing Egghe against Ha and coauthors. Egghe’s own PDF did not load (server 503), so Egghe is not used as an opened source.
-
Moreno-Sánchez, Font-Clos, and Corral, “Large-Scale Analysis of Zipf’s Law in English Texts” (PLOS ONE, 2016), https://fontclos.github.io/files/Moreno-Sanchez_2016_Large_Scale_Analysis_Zipf.pdf (DOI 10.1371/journal.pone.0147073). More than 30,000 Project Gutenberg English texts. They state the classic form as frequency n decreasing approximately as 1/r, and more generally n ∝ r^(−α) with α close to one. They also fit the distribution of frequencies. In that second version the exponent β satisfies β = 1 + 1/α, so values of β lie close to 2, and Zipf’s own formulation implies β = 2. For texts accepted at p > 0.05, estimated mean β is 2.03 (standard deviation 0.15) for one of their three forms and 2.02 (0.17) for another. β varies from text to text, mostly between about 1.6 and 3, and the fitted β decreases as text length grows. A pure power law in the complementary cumulative distribution of frequencies, one free parameter, fits more than 40% of the texts at the 0.05 level across the whole frequency domain. So “exactly −1” on the rank plot, and “exactly 2” on the frequency spectrum, are reference values, not what every text returns.
Heaps and finite size, opened separately:
-
Lü, Zhang, and Zhou, “Zipf’s Law Leads to Heaps’ Law” (PLOS ONE, 2010), https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0014139. Heaps’ law (growth of the number of distinct words) can be derived if Zipf’s law holds. The usual link between the two exponents is an asymptotic result for infinite systems. In a finite system the Heaps exponent depends on system size.
-
Lü, Zhang, and Zhou, “Deviation of Zipf’s and Heaps’ Laws in Human Languages with Limited Dictionary Sizes” (Scientific Reports, 2013), https://www.nature.com/articles/srep01082. They write Zipf as Z(r) ~ r^(−α) and Heaps as N_t ~ t^λ with λ ≤ 1. They say both are observed in the Indo-European family and do not hold in the same way for Chinese, Japanese, and Korean, which they treat as character systems with a limited dictionary. Character frequency follows a power law with exponent close to one, at which the corresponding Zipf exponent diverges, and on a Zipf plot the character frequency decays exponentially. Distinct characters grow in three stages (linear, then logarithmic, then saturation).
2. Phrases, n-grams, and constructions
Opened and measured:
-
Ha et al. 2002 (URL above). Separate rank lists for word n-grams, n = 2 to 5, on Wall Street Journal English (about 19, 16, and 6 million tokens) and on the TREC Mandarin corpus (19,546,872 tokens). Straight-line slopes they tabulate, as the magnitude of the negative slope, are about 0.65–0.67 (English bigrams; Mandarin 0.75), 0.46–0.51 (English trigrams; Mandarin 0.59), and lower still for 4-grams and 5-grams (English 5-grams 0.34–0.42; Mandarin 0.48). Fixed-length n-grams are shallower than −1. Unigram and bigram curves cross near rank 3000 in the English sets. When unigrams and n-grams are put in one frequency-ranked list, both languages give lines with slopes close to −1 for all ranks above 100. They say the same combined pattern appeared in preliminary checks of Irish, Latin, and Vietnamese. These are overlapping fixed-n counts, not a syntactic notion of a construction.
-
Williams, Lessard, Desu, Clark, Bagrow, Danforth, and Dodds, “Zipf’s law holds for phrases, not words” (Scientific Reports 5, 12209, 2015), https://www.nature.com/articles/srep12209 (DOI 10.1038/srep12209; article srep12209). They say word Zipf holds over no more than three to four orders of magnitude before a break. They do not use overlapping n-grams as the main object, because overlap obscures word frequencies. They cut clauses at punctuation and then cut word–word boundaries at random (pure random partitioning, every partition equally likely). What is ranked is the expected frequency of mixed-length phrases. Those phrases are mixed-length and non-overlapping. They are not phrasal verbs. On English Wikipedia (2010), the New York Times (1987–2007), Twitter (2009), and music lyrics (1960–2007), those phrase frequencies scale around exponent θ = 1 for 7–9 orders of rank. Mixed-length phrases in this paper sit near −1 over a long rank range. For about 4000 Gutenberg works, phrase exponents are typically θ ≤ 1, while word exponents are mostly θ > 1. Figure values for one novel were not in the text layer of the page that loaded, so they are not stated here. They also partition letters (graphemes). Letter scaling is not robust. They cite Ha et al. 2002 as having seen a similar long range by combining n-gram lists, and they call that approach unprincipled because it over-counts words.
Grant’s open question (his voice; not a finding). He does not know if this is true. It is only something to explore. It is not a result, and it is not a hypothesis this page tests to a conclusion.
Maybe the power law applies to phrasal verbs (a verb plus its arguments, basically a function) rather than to phrases in general.
Williams et al. 2015 (Scientific Reports 5, 12209, srep12209) define the phrases they rank by random partitions: mixed-length non-overlapping phrases, every partition equally likely. That object is not a phrasal verb, and it is not a verb plus its arguments. No phrasal-verb exponent is stated on this page.
Molecule wording in that paper, and Grant’s open analogy
The paper does use the analogy. Its sentence is that morphemes are the meaningful “atoms” of language, and that the meaningful “molecules” of language are a mixture of individual words and phrases. That is a linguistic remark, not a chemical measurement. The paper does not measure chemical molecules, formulas, or reactions. The app’s sentence that phrases behave like “meaningful molecules” is the app’s, not a sentence of the paper, and it is narrower than the paper: the paper’s “molecules” are words and phrases together, and its “atoms” are morphemes, not words. That remark is not the word ↔ atom / phrase ↔ molecule cut. Grant’s word ↔ atom pairing is not what this paper says.
Beside that morpheme-as-atom line, Grant’s own remark (HIS, SPECULATIVE, not a finding): it is not wrong to say morphemes are like quarks, because they can only be produced by the breakdown of a larger unit such as a word or an atom. He added that maybe anyons, and that he does not know physics particles perfectly. Quarks and anyons are both his open analogy. This page does not choose one over the other. It does not assert confinement or any particle-physics result.
3. Chemistry: what was opened, and what it is not
Distinguish the object. None of the following is a rank-frequency law over chemical species, molecular formulas, or “a molecule” as the analogue of a phrase. Chemistry sources opened do not show Zipf over molecules.
-
Features inside molecules, not molecules. Benz, Swamidass, and Baldi, “Discovery of power-laws in chemical space” (J. Chem. Inf. Model., 2008, DOI 10.1021/ci700353m). PubMed abstract only, via NCBI efetch (https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=18522387&retmode=text&rettype=abstract). The full article did not open. They report log–log rank/frequency linearity for rigid segments, ring systems, molecular paths, circular substructures, and sizes of molecular similarity clusters, with characteristic exponents in the 1.5–3 range, plus Heaps-like growth of the number of unique features. That is substructure and cluster size, not a count of each molecule.
-
Reaction-network degree, not a Zipf plot of formulas. Fialkowski, Bishop, Chubukov, Campbell, and Grzybowski, “Architecture and Evolution of Organic Chemistry” (Angewandte Chemie, 2005, DOI 10.1002/anie.200502272), PDF at http://fennetic.net/irc/grzybowski/Grzybowski_2005.pdf. Molecules are nodes. In- and out-degree distributions obey p(k) = k^(−γ). From 1850 to 2004 the exponents approach about 2.7 (in) and 2.1 (out). They also report preferential attachment and a power-law drop of price with degree for the top 300 economically relevant compounds. This is how many reactions touch a molecule, not how often a formula occurs.
-
Tokens of molecular strings, not molecules. Liu, Fang, Li, Wang, Qi, and Hu, “Zipf-like Statistical Regularities in Molecular Sequence Representations for Chemical Language Models” (J. Phys. Chem. Lett., 2026, DOI 10.1021/acs.jpclett.6c02089), landing page https://pubs.acs.org/doi/abs/10.1021/acs.jpclett.6c02089 (published date on the page: 2026-08-17). Abstract on that page, not the full text: rank–frequency of byte-pair-encoding tokens of SMILES and SELFIES across large molecular databases. Slopes are described as moderately steeper than the canonical −1. No numeric slope is given in the abstract, so none is stated here. Closeness of those slopes to −1 correlated with BERT property-prediction performance (Pearson r = 0.91, p < 0.001; partial r = 0.86 after vocabulary size). The parallel they draw is to natural-language token statistics. The units are substrings of line notations, not molecules and not phrases of chemistry.
Not opened, and therefore not reported: element abundances; molecule or formula counts in a database (an MDPI 2021 popularity paper returned 403); citation counts of compounds; mass spectra. A 2018 RSC reaction-network HTML page stopped at a login wall and is not used.
4. Explicit map from linguistic Zipf onto chemistry
No opened full text maps linguistic Zipf onto molecules. There is no joint field in the sources opened here.
The closest opened abstract is Cadeddu, Wylie, Jurczak, Wampler-Doty, and Grzybowski, “Organic chemistry as a language and the implications of chemical linguistics for structural and retrosynthetic analyses” (Angew. Chem. Int. Ed., 2014, DOI 10.1002/anie.201403708), PubMed abstract only (pmid 25044611). It says English and organic chemistry have the same structure in the frequency of text fragments and of molecular fragments, and that high-information bonds mark symmetry or repeat units and, in asymmetric molecules, possible retrosynthetic cuts. The abstract does not say Zipf, does not give an exponent, and does not identify phrases with molecules or words with atoms. The Wiley full text did not load (bot check). An earlier project gloss that called this paper “Zipf-like” is not what that abstract says. See Prior bridges.
Liu et al. 2026, abstract only, is the opened source that uses the name Zipf for a chemical object, and that object is a BPE vocabulary of molecular strings, as in section 3.
What was not used
Egghe 1999 PDF: HTTP 503. MDPI Molecules 2021, 26, 2394: HTTP 403. Europe PMC and the PubMed HTML shell blocked or required JavaScript; the two abstracts above are from NCBI efetch. RSC 2018 article HTML: login wall. No element-abundance paper and no mass-spectrum paper was opened. Exponents that appeared only in search snippets, and figure readings that were not in an opened text, are omitted. No phrasal-verb exponent was taken from any source. “Phrasisized” is Grant’s coinage, not a term in the opened papers. Quarks and anyons are his labels, not measurements.
2026-10-03 PT, with Grant’s later open question and open analogy recorded beside Williams et al. 2015 (srep12209). No joint field. Not Bridge A.
Follow-up pointer, 2026-10-06
HIS open question above (Zipf on verb + arguments) is taken up in Zipf at the frame (site page: Zipf at the frame): English verb-argument constructions show Zipfian, frame-selective verb fillers (Römer, O'Donnell & Ellis 2015), and the CAS Registry shows a power law over molecular frameworks (Lipkus et al. 2008). That note also corrects §4 in part: Lipkus et al. 2008 explicitly carry Simon's word-frequency model over to frameworks (framework as word type). It is still not a phrase ↔ molecule claim and not a Zipf law over molecules as species.
Corpus note: zipf-language-chemistry-v0.md. Rank counts stay in the field that measured them. Grant’s phrase↔molecule line, phrasal-verb question, and quark/anyon analogy are his. Not Bridge A. Not a joint field.