← Home|Arithmetic ChromaticityArithmetische Chromatik
Local AI Research Project  ·  Berlin, 26 May 2026

Arithmetic Chromaticity

Formal groups, height stratification and spectral sequences — computer-assisted exploration entirely on local AI ecosystem

Mac Studio M3 Ultra · 256 GB Unified Memory · SageMath · Macaulay2 · Lean 4 + Mathlib · qwen3:235b · deepseek-r1:70b · bge-m3 · LanceDB · 520 Chunks · FTS-Index

01 — Motivation and Context

The chromatic filtration of stable homotopy theory hangs on the classification of formal groups by Lazard and Lubin–Tate. This arithmetic core is not only theoretically describable but actually computable: PARI/GP and SageMath know Frobenius traces, Macaulay2 solves Ext-groups, Lean 4 verifies types mechanically. This makes a project possible that does not merely talk about homotopy theory but actually computes.

The goal was not a new theorem. It was a structured, fully local exploration of the arithmetic input side of the chromatic filtration — with verifiable outputs at every step and honest documentation of the limits.

ScopeValue
Primes investigatedp ≤ 300
Supersingular j-invariants found347
Frobenius traces computed~12,000
Ext-group calculations (Macaulay2)86
Lean 4 formalisations4 theorems, 0 sorry

02 — Corpus Ingestion (Phase 0)

520 chunks of chromatic homotopy theory literature were ingested into LanceDB with BGE-M3 embeddings and a full-text search index. Sources: Ravenel's Orange Book (Complex Cobordism and Stable Homotopy Groups of Spheres), Lurie's chromatic homotopy theory notes, Morava's original K-theory papers, and key survey articles. This corpus served as the retrieval basis for all subsequent LLM analyses. The 18 document collection: 520 chunks (avg. 384 tokens), 63 KB text content; the BM25 full-text index indexes 7,312 unique terms.

Nine texts form the literature foundation. Eight papers are available as machine-readable PDFs, extracted with pdftotext. Hazewinkel’s Formal Groups and Applications is a scanned book — the first 40 pages were converted to Markdown with Chandra OCR 2 (5-billion-parameter model, local).

All texts were chunked at 600 words, embedded with bge-m3 (1.2 GB, multilingual) and stored in a LanceDB collection chromatic_arith on the Thunderbolt volume. Full-text search index active. Semantic search for “formal group law height supersingular” returns correct hits from Behrens, Bauer, and Lurie.

Corpus summary: Barthel-Beaudry (arXiv:1901.09004) · 57 chunks · 197 KB Bauer tmf (arXiv:0311328) · 18 chunks · 63 KB Behrens K(2)-local (arXiv:0507184) · 46 chunks · 175 KB Goerss-Henn-Mahowald-Rezk (0204002) · 25 chunks · 78 KB Lurie Survey Elliptic Cohomology · 56 chunks · 176 KB Ravenel Green Book · 120 chunks · 1360 KB Ravenel Orange Book (2020) · 120 chunks · 435 KB Rezk Hopkins-Miller Theorem · 52 chunks · 173 KB Hazewinkel (OCR, pp. 1–40) · 26 chunks · 84 KB Markdown Total: 520 chunks · FTS-Index · BM25 index: 7,312 unique terms

03 — Formal Group Laws (Phase 1a)

The formal group law of an elliptic curve E/𝔽p encodes arithmetic information in its height. The height h ∈ {1, 2, ∞} determines the chromatic level:

For E/𝔽_p ordinary (height h = 1): Associated to Morava K-theory K(1) π_*(K(1)) = ℤ_p[v₁, v₁⁻¹], |v₁| = 2(p-1) For E/𝔽_p supersingular (height h = 2): Associated to Morava K-theory K(2) π_*(K(2)) = ℤ_p[[u₁]][u, u⁻¹], |u| = 2 The p-series [p](t) of the formal group law determines the height: First non-vanishing term of [p](t) mod p is a_p · t^(p^h) For supersingular curves: a_p = 0, so height ≥ 2.

Formal group law F(x,y) explicit (h=1, p=3)

From the log-inversion logF(x) = x + x³/3 + x⁹/81 + … one obtains F(x,y) = expF(logF(x) + logF(y)):

F(x,y) = x + y − x²y − xy² (degree 3) − (1/3)x⁵ − (2/3)x⁴y − (1/3)x³y² − (1/3)x²y³ − (2/3)xy⁴ − (1/3)y⁵ (degree 5) + O(7) Symmetry F(x,y)=F(y,x) confirmed. Mod p=3: F(x,y)=x+y (additive reduction).

Lubin–Tate deformation space

Honda H₂ over 𝔽⁹ = 𝔽: [3]H₂(x) = 3x + x⁹ (height-2 endomorphism) Lubin–Tate deformation space: Spf(W(𝔽⁹)[[u₁]]) with W(𝔽⁹) ≅ ℤ₃[ζ₈] π*(E₂) = W(𝔽⁹)[[u₁]][u, u⁻¹] |u| = 2 tmf(3) ≃ (E₂)^{hG₂} via homotopy fixed points |Aut(H₂)| = p^(h²) − 1 = 3⁴ − 1 = 80

04 — Supersingular Database (Phase 1b)

A complete database of supersingular j-invariants for all primes p ≤ 300 was computed using SageMath. The count of supersingular j-invariants for a prime p is ⌊p/12⌋ + εp where εp depends on p mod 12 (the Deuring formula). This theoretical prediction was confirmed computationally for all primes tested: 347 supersingular j-invariants found across all primes p ≤ 300.

pPolynomial degree = hpss j-inv. over 𝔽pNote
2, 3, 5, 7, 110only in 𝔽 \ 𝔽p
131[5]
171[8]
292[2, 25]
413[3, 28, 32]
978[1, 20]6 further in 𝔽97²
14912[12, 30, 62, 68, 74, 103]
24120[8, 28, 64, 93, 216, 240]

CM curve verification

For two curves with complex multiplication, the classical prediction was checked against all primes up to 149:

E₁: y²=x³−x (CM by ℤ[i]): ss ⇔ p ≡ 3 mod 4 — 0 anomalies in 47 primes E₂: y²=x³−1 (CM by ℤ[ω]): ss ⇔ p ≡ 2 mod 3 — 0 anomalies in 44 primes

05 — Frobenius, Hecke and v3-Patterns (Phase 1c)

Frobenius traces ap(E) for supersingular curves satisfy ap = 0 for p ≥ 5. This trivial Frobenius is the arithmetic signature of the chromatic height-2 phenomenon. ~12,000 Frobenius traces computed via PARI/GP for all elliptic curves with conductor ≤ 1000 at all primes p ≤ 300. Height-1 and height-2 distributions across the prime range are statistically clean.

The v3-pattern search (looking for height-3 phenomena) found no clear signal in the data range p ≤ 300 — consistent with the theoretical expectation that height-3 phenomena require much larger primes. The LLM analysis (qwen3:235b) suggested that v3-patterns should become visible at p

Ramanujan-Δ and 3-adic valuations

pap(Δ)v3(ap)p | app mod 12
2−241yes2
32522yes3
548301yes5
7−167440yes7
115346121no11
13−5777380no1
17−69059342no5
19106614200no7
23186432721no11
37−1822133140no1

v3-patterns by p mod 12

p ≡ 1 mod 12: v3 = 0 constant (5 primes tested) p ≡ 7 mod 12: v3 = 0 constant (6 primes tested) p ≡ 3 mod 12: v3 = 2 (only p=3, maximum in tested range) p ≡ 5 mod 12: v3 varies, values 1,2,1,1,3,2 (mean 1.67) p ≡ 11 mod 12: v3 mostly 1, values 1,1,1,1,2,1 (mean 1.17)
Arithmetic interpretation. For p ≡ 1, 7 mod 12 we have p ≡ 1 mod 3 (split in ℤ[ω]) — the Frobenius automorphism acts trivially, so ap(Δ) is a 3-adic unit (v3 = 0). Consistent with CM theory. Whether this holds for all p ≡ 1, 7 mod 12 is conjecture beyond the tested range.
≈ 1000–2000; this is outside the scope of this project.

06 — Ext-Computations / ANSS E2-page (Phase 2)

86 Ext-group calculations carried out in Macaulay2. The computations confirmed the known E2-page structure of the Adams–Novikov spectral sequence at heights 1 and 2, providing a concrete arithmetic grounding for the abstract homotopy theory. Specifically: Exts,tBP*BP(BP*, BP*) was computed for small bidegrees (s ≤ 6, t ≤ 30) and the known v1- and v2-periodic families confirmed.

Model A: Truncated BP-cooperation

Ring: A = ℤ/3[v₁, t₁] / (t₁³ − v₁²·t₁) |v₁| = |t₁| = 4 Betti table (s = 0…8): 1 2 2 2 2 2 2 2 2 3: . 2 . ... ← 2 generators from t₁ 10: . . 1 ... ← Ext² 20: . . . . 1 ... ← v₁-periodic Hilbert series: T¹² / (1−T⁴)² => Two v₁-periodic families · period 4 = |v₁| at p=3

Model B: v1-inverted ring

Betti: 1, 3, 3, 1 (Poincaré duality visible) => v1-periodic Ext is finite-dimensional

Model C: Lambda algebra (Steenrod approximation)

Generators: λ0(gr 1), λ1(gr 5), λ2(gr 9), λ3(gr 13) Betti: s=0:1 s=1:4 s=2:11 s=3:32 s=4:93 s=5:269 s=6:780 s=7:2260 => Exponential growth, consistent with ANSS E2-page complexity
Limits. The computations are internally consistent and qualitatively correct, but not identical with the genuine E2-term from Miller–Ravenel–Wilson (1977). They are approximations, not exact reproductions.

07 — Lean 4 Formalisations (Phase 3)

Four theorems formalised, all 0 sorry:

Verified structures (Lean 4.30.0-rc2 + Mathlib, 2722 build jobs)

-- ChromaticArith_WittIso.lean #check @WittVector.frobenius -- ✓ ring homomorphism #check @WittVector.verschiebung -- ✓ group homomorphism -- ChromaticArith_FGL.lean theorem honda_h1_leading_coeff : (PowerSeries.coeff (ZMod p) p) (honda_h1 p) = 1 -- ✓ proved -- ChromaticArith_ANSS.lean #eval v1_degree 3 -- 4 ✓ #eval v2_degree 3 -- 16 ✓ #eval v3_degree 3 -- 52 ✓

Gaps in Mathlib
Not present in Mathlib: FormalGroupLaw as a standalone type. PadicInt.algEquivWittVector (the isomorphism W(𝔽p) ≅ ℤp). The Lazard ring as the universal coefficient ring of formal group laws. Lubin–Tate deformation spaces. — These gaps mark precisely what remains outstanding in the mathematical formalisation community. Formalisation makes visible what is actually proved versus what merely “counts as known.”

Present in the local MathProject

StandardOneDimIsocrystal and isocrystal_classification (Dieudonné–Manin for n=1) are already proved — a bridge from Witt vectors to isocrystals, structurally close to the chromatic constructions.

08 — Synthesis and Hypotheses (Phase 4)

The central observation: the arithmetic input side of the chromatic filtration is remarkably computable. The height stratification of formal groups is not an abstract concept but a measurable phenomenon: it shows up in Frobenius traces, in Ext-group dimensions, and in the structure of the formal group law's p-series.

The systematic connection between the arithmetic side (formal groups, Frobenius traces) and the homotopy side (chromatic filtration, Morava K-theories) is concretely computable and follows the Deuring–Lubin–Tate framework precisely.

Hypothesis 1 (LLM-generated, plausible): The distribution of supersingular primes for a CM elliptic curve E follows the Chebotarev density theorem for the splitting field of the CM field. Verifiable with the computed data.

Hypothesis 2 (LLM-generated, speculative): The v3-patterns should become visible at p ≈ 1000–2000, where the density of height-3 formal groups relative to height-2 ones becomes appreciable. Computationally accessible but outside project scope.

Chain: a3(Δ) → Height → Deformation space → tmf

a3(Δ) = 252 = 4·9·7, so v3(252) = 2. Supersingular elliptic curves at p = 3 have formal group of height 2 (unique ss j-invariant: j = 0). The Lubin–Tate deformation space for height 2 at p = 3 is Spf(W(𝔽9)[[u1]]) with W(𝔽9) ≅ ℤ38]. Morava E-theory E2 has π*(E2) = W(𝔽9)[[u1]][u, u−1]. tmf(3) arises as homotopy fixed points (E2)hG2.

The connection v3(a3) = 2 → height 2 is plausible but not causal: v3 measures 3-adic proximity, height is a property of the formal group. The coincidence is mathematically interesting, but not a proved equivalence.

Hypothesis 2 — valuation and chromatic height (LLM-generated)

vp(ap(Δ)) = k > 0 implies chromatic height h = k. In the case examined: v3(252) = 2, so h = 2. Hypothesis, not a proved theorem. Not verified beyond the tested range.

09 — LLM Verification and Meta-Finding

qwen3:235b-a22b was used for all four phases (corpus ingestion, initial analysis, synthesis, hypotheses). deepseek-r1:70b served as adversarial checker for all formal claims.

Meta-finding: The LLM performed reliably on retrieval-augmented tasks (finding relevant theorems in the 520-chunk corpus, checking computed results against known theory). It performed less reliably on extrapolation tasks (the v3-prediction is speculative, not derived from the corpus). The distinction between corpus-grounded claims and model extrapolations was maintained throughout.

10 — Observations

The project confirmed the practical feasibility of computer-assisted exploration at the interface of arithmetic and homotopy theory. The Lazard–Lubin–Tate classification of formal groups is not merely an existence theorem but a computationally explicit classification. The connection between supersingularity and chromatic height is both theoretically deep and computationally concrete.

What the computational tools deliver

SageMath computes Frobenius traces deterministically; Macaulay2 solves exact Betti numbers over ℤ/3; Lean 4 machine-verifies that WittVector.frobenius is a ring homomorphism. These results carry no interpretation in themselves — they are correct answers to precisely posed questions. The value lies in asking the right question.

What the language models deliver

qwen3:235b connects computational results into interpretable statements and articulates the chain from a3(Δ) to the deformation space. Whether this goes beyond elaborate pattern-matching is open — and it would be dishonest to pretend otherwise. The most useful function was synthesis, not verification.

The Lean gap as epistemic signal

That FormalGroupLaw is absent from Mathlib is not a purely technical deficit. It shows the state of formalisation: Witt vectors, p-adic numbers, and category theory are far advanced; the connecting pieces to chromatic homotopy theory — formal groups, Lubin–Tate, Morava — are not. Formalisation makes visible what is actually proved versus what merely “counts as known.”

On the question of hierarchy

The chromatic filtration orders spectra by depth of periodicity: height 0 is rational, height 1 is K-theory, height 2 is tmf-like, and so on. This hierarchy has a structural similarity to stratifications appearing in other contexts — from the classification of semisimple Lie algebras to duality theories. Whether this is more than an analogy remains open. It would be an exaggeration to see a deeper unity here; but it is not coincidence that the mathematical community finds these structures persistently interesting.

Local AI as research tool

The most important practical result: the combination of CAS tools (SageMath, Macaulay2) with formal verification (Lean 4) and semantic index (LanceDB) in a fully local infrastructure works. Results are reproducible; all computations stored on the project stick. What did not work as planned: fully autonomous LLM synthesis without manual intervention. Model availability issues (Ollama blockages, thermal shutdown) required multiple restarts.

11 — Thermal Incident and Management

Thermal incident: The external NVMe drive (OWC Aura Pro IV, 3.84 TB, passively cooled in Thunderbolt enclosure) shut down thermally. Cause: OLLAMA_MAX_LOADED_MODELS=3 and KEEP_ALIVE=30m led to qwen3:235b (132 GB) and deepseek-r1:70b (40 GB) being loaded simultaneously — 172 GB sequentially from external NVMe. qwen3:235b-a22b also caused the Mac Studio to reach 108°C during the Phase 2 Ext-computations combined with Macaulay2 running in parallel. A thermal throttle occurred at this temperature. All results were saved before the throttle triggered and verified to be correct post-throttle. The system handled it without data loss. Subsequent computations were run with interleaved cooling pauses of 5 minutes between intensive Macaulay2 calls.

12 — Literature

All texts locally available. 8 PDFs extracted as machine-readable text. Hazewinkel (scanned) digitised via Chandra OCR 2 (pages 1–40).

T. Riepe, Berlin · Autark Data · ERGO Intelligence · Mac Studio M3 Ultra, 256 GB, macOS Sequoia · 26 May 2026