TL;DR: A simple 2×2 matrix family M_{a,b} with fixed product P=ab is isospectral (same eigenvalues) but its geometry, topology, and dynamics are governed by the divisors of P. The log traces the discoveries, dead ends, and corrections along the way.
Organising principle. The family M_{a,b} exhibits a remarkable separation between spectrum and geometry: the spectrum depends only on P=ab, while the arithmetic, symbolic, topological, and dynamical structures are organised by the divisor fibre over P.
This document is written as a pattern-seeking narrative. Each section answers a natural question raised by the previous one, and every claim is labelled according to its epistemic status:
| Level | Meaning |
|---|---|
| Theorem | Fully proved in the text. |
| Computational Theorem | Exhaustively verified over a stated range. |
| Conjecture | Supported by evidence but not yet proved. |
This is a research log documenting an exploratory investigation, not a peer-reviewed paper. Several claims are computational theorems verified up to a stated bound, not proved for all P; a few are conjectures resting on evidence rather than proof. Corrections, counterexamples, and rigorous proofs of the open conjectures are welcome.
Contents
- The Isospectral Observation
- Two-Generator Simplicity
- Fixed Points and the Continued Fraction Surprise
- Conjugacy — A Correction
- The Divisor Structure Formula
- p-Adic Dynamics
- The Braid Group and Topology
- The Markoff Connection
- A Disproved Conjecture
- Higher Rank Generalization
- Persistent Patterns
- How Much Does F_P Actually See?
- Appendix A: Verification Script
- Appendix B: Changelog
- Appendix C: References
Phase 1: The Isospectral Observation
For integers a,b, define
M_{a,b} = [[1+ab, a], [b, 1]].
For a fixed positive integer P, consider the family
F_P = { M_{a,b} : a,b ∈ Z_{>0}, ab = P }.
Theorem. For every pair (a,b) with ab=P:
det(M_{a,b}) = 1, tr(M_{a,b}) = P+2.
Hence M_{a,b} ∈ SL(2,Z) and all matrices in F_P share the characteristic polynomial
χ(λ) = λ^2 - (P+2)λ + 1.
The eigenvalues are
λ_{±} = (P+2 ± sqrt(P(P+4)))/(2),
with λ+ λ- = 1, λ+ + λ- = P+2, and λ+ - λ- = sqrt(P(P+4)).
Verification. Direct computation for (P,a,b) ∈ {(1,1,1), (6,2,3), (12,3,4), (60,5,12), (100,10,10), (840,20,42)}.
Pattern. All matrices in F_P are isospectral. The eigenvalues know only P; the eigenvectors know (a,b).
Phase 2: Two-Generator Simplicity
A natural question: how complicated are these matrices as words in SL(2,Z)?
Let L = [[1, 1], [0, 1]] and R = [[1, 0], [1, 1]].
Theorem. M_{a,b} = L^a R^b.
Equivalently, M_{a,b} = E_+(a) · E_-(b) where E_+(a) = [[1, a], [0, 1]] and E_-(b) = [[1, 0], [b, 1]].
Verification. Matrix multiplication confirmed for 8 distinct triples (P,a,b).
Pattern. Only two generators, used exactly once each. Most elements of SL(2,Z) require long words; these require the minimal non-trivial product.
Phase 3: Fixed Points and the Continued Fraction Surprise
Identifying M_{a,b} with the Möbius transformation z ↦ ((1+P)z + a)/(bz + 1), the fixed points are:
Theorem.
z_{±} = (λ_{±} - 1)/(b).
Verification. Direct substitution confirmed for (P,a,b) ∈ {(1,1,1), (3,1,3), (6,2,3), (12,3,4)}.
Numerical exploration with small (a,b) suggested a periodic continued fraction for the attractive fixed point z_+:
| (a,b) | Numerical z_+ | CF guess |
|---|---|---|
| (2,3) | 2.290994 | [2,3] |
| (3,2) | 3.436492 | [3,2] |
| (6,1) | 6.872983 | [6,1] |
Theorem (Fixed Point CF). The attractive fixed point has the exact simple continued fraction expansion
z_+ = [a, b] = [a, b, a, b, a, b, ...].
Proof. Let x = [a, b]. Then x = a + (1)/(b + (1)/(x)), which rearranges to bx^2 - abx - a = 0. The positive root is
x = (P + sqrt(P(P+4)))/(2b) = (λ+ - 1)/(b) = z+. ∎
Pattern. The period [a,b] encodes the factorization directly. The cutting sequence of the corresponding geodesic on the modular surface is periodic with pattern L^a R^b L^a R^b ·s (Series 1985).
Phase 4: Conjugacy — A Correction
First attempt (erroneous). We initially claimed M_{a,b} ~ M_{b,a} in GL(2,Z) via the swap matrix X = [[0, 1], [1, 0]].
Correction. Computational verification showed the swap matrix does not conjugate M_{a,b} to M_{b,a}. The correct conjugating matrix is:
Theorem.
S = [[-b, -1], [-1, 0]], det(S) = -1,
satisfies S M_{a,b} S^{-1} = M_{b,a}.
Verification. Direct matrix multiplication confirmed for 8 test cases.
For SL(2,Z) conjugacy, brute-force search revealed:
Computational Theorem. For all P <= 50 (search radius ± 10 for conjugating matrix entries):
- M_{a,b} and M_{b,a} are conjugate in SL(2,Z) iff a=b.
- The number of distinct SL(2,Z)-conjugacy classes represented by F_P is exactly d(P), the number of positive divisors of P.
The obstruction when a != b reduces to a Pell-type equation with no integer solutions (mod 8).
Pattern. The (a,b) rightarrow (b,a) symmetry is broken by the determinant-1 condition. The divisor function d(P), not ⌈ d(P)/2 ⌉, governs the conjugacy class count.
Phase 5: The Divisor Structure Formula
Investigating why there are d(P) conjugacy classes led to the gcd structure. Let g = gcd(a,b). Then g^2 | P, and writing a = ga', b = gb' with gcd(a',b') = 1:
Theorem.
d(P) = ∑_{g | P ; g^2 | P} 2^{ω(P/g^2)}
where ω(n) counts distinct prime factors.
Proof sketch. For P = ∏ p_i^{e_i}, a divisor g satisfies g^2 | P iff g = ∏ p_i^{f_i} with 0 <= f_i <= ⌊ e_i/2 ⌋. The inner sum over f_i equals e_i+1 regardless of parity, giving ∏ (e_i+1) = d(P). ∎
Verification. Confirmed for all P ∈ [1,200].
Pattern. Each content value g contributes 2^{ω(P/g^2)} coprime pairs. The total partitions the conjugacy classes by their gcd.
Phase 6: p-Adic Dynamics
For a prime p, the characteristic polynomial splits over F_p iff P(P+4) is a quadratic residue mod p.
Special case. When P = p-2 (so p ≡ 1 (mod 4)), the polynomial mod p becomes λ^2 + 1.
Theorem (Hensel Lifting). For p ≡ 1 (mod 4), the roots of λ^2 + 1 in F_p are simple and lift uniquely to Z_p.
Verification. Newton iteration confirmed for p ∈ {5, 13, 17, 29} to precision p^{32}.
Theorem (p-adic Unit Property). For any prime p and any P >= 1, λ+ is a unit in any p-adic field containing it (|λ+|_p = 1).
Proof. The minimal polynomial x^2 - (P+2)x + 1 has constant term 1. In any non-archimedean absolute value, |α| > 1 or |α| < 1 both lead to contradictions by comparing term sizes. ∎
Computational Theorem (Involution Dynamics). For "clean" primes p = P+2 ≡ 1 (mod 4):
- The reduced Möbius map f(z) = ((1+P)z + a)/(bz + 1) on P^1(F_p) is an involution: f(f(z)) ≡ z (mod p).
- There are exactly 2 fixed points and (p-1)/2 2-cycles on P^1(F_p).
- The multiplier satisfies μ ≡ -1 (mod p).
- These mod-p 2-cycles are "ghosts": they do not lift to Q_p.
Verification. Involution property and fixed-point count confirmed for p ∈ {5, 13, 17, 29}.
Pattern. p-adic dynamics are universally indifferent (Siegel fixed points). The only periodic points in P^1(Q_p) are the two fixed points.
Phase 7: The Braid Group and Topology
Under the standard Artin representation ρ: B_3 -> SL(2,Z):
ρ(σ_1) = L, ρ(σ_2^{-1}) = R,
Theorem. M_{a,b} = ρ(σ_1^a σ_2^{-b}).
Verification. The braid relation σ_1σ_2σ_1 = σ_2σ_1σ_2 maps to [[0, 1], [-1, 0]] in SL(2,Z), confirmed by direct multiplication.
Topological interpretation (pattern). The closure of the braid σ1^a σ_2^{-b} in S^3 is a 2-bridge link (a 2-bridge knot under suitable parity conditions). The monodromy of the fibered surface is exactly M{a,b}.
Theorem (Alexander Polynomial). The Alexander polynomial of this link is the characteristic polynomial of the monodromy:
Δ(t) = t^2 - (P+2)t + 1
(up to units ± t^k).
Pattern. The isospectral family corresponds to distinct 2-bridge links with identical Alexander polynomial. The link type varies with (a,b) while the polynomial remains fixed. There are ⌈ d(P)/2 ⌉ distinct unordered links, distinguished by their continued fraction invariants [a,b].
Phase 8: The Markoff Connection
The Markoff equation x^2 + y^2 + z^2 = 3xyz is related to traces in SL(2,Z) via the Fricke identity:
Theorem (Fricke Identity). For A,B ∈ SL(2,Z):
tr(A)^2 + tr(B)^2 + tr(AB)^2 = tr(A)tr(B)tr(AB) + tr([A,B]) + 2.
Verification. For A = M_{1,1} (trace 3) and B = M_{1,3} (trace 5): tr(AB) = 13, tr([A,B]) = 6, and 9+25+169 = 3 · 5 · 13 + 6 + 2 = 203.
Computational Theorem (Markoff Cross-Reference). Searching all Markoff numbers <= 10^6 via Vieta jumping yields 40 Markoff numbers. Of these, 38 correspond to P+2 for some P <= 999{,}998.
The sequence of such P begins:
P ∈ {3, 11, 27, 32, 87, 167, 192, ...}.
Pattern. The density is extremely low. Markoff numbers grow roughly like e^{csqrt(n)}, so only O((log N)^2) values of P <= N yield Markoff traces.
Phase 9: A Disproved Conjecture
Conjecture (abandoned). Perhaps d(P) | h^+(P(P+4)), connecting the divisor count to the narrow class number of Q(sqrt(P(P+4))).
Counterexamples:
| P | d(P) | h^+(P(P+4)) | Divisible? |
|---|---|---|---|
| 4 | 3 | 2 | 3 ∤ 2 ✗ |
| 8 | 4 | 6 | 4 ∤ 6 ✗ |
| 12 | 6 | 4 | 6 ∤ 4 ✗ |
Lesson. The conjugacy class count = d(P) is a special property of this matrix family, not a consequence of genus theory.
Phase 10: Higher Rank Generalization
For n >= 2, define:
M^{(n)}{a,b} = I_n + a E{1,n} + b E_{n,1} + ab E_{1,1}.
Theorem. The characteristic polynomial is:
χ(λ) = (λ - 1)^{n-2}l(λ^2 - (P+2)λ + 1r).
Proof. Expanding det(M^{(n)}_{a,b} - λ I) along the middle diagonal rows gives (1-λ)^{n-2} times the 2 × 2 determinant from the corner entries. ∎
Conjecture. For each n >= 2 and fixed P, the matrices {M^{(n)}_{a,b} : ab = P} are pairwise non-conjugate in SL(n, Z), and the number of conjugacy classes is d(P).
Evidence. Verified for n=3, P=6 by brute-force search.
Phase 11: Persistent Patterns
What survives this pattern-seeking process are not individual formulas but structural invariants:
The trace P+2 is the only spectral invariant. Everything else—eigenvectors, word length, cutting sequences, link types—is parametrized by the divisors of P.
The continued fraction [a,b] is the complete invariant. It distinguishes conjugacy classes, encodes the Möbius fixed point, and determines the cutting sequence of the geodesic.
The divisor function d(P) governs the combinatorics. The number of conjugacy classes, the number of distinct geodesics, the number of link types, and the number of "shapes" in eigenvector space all reduce to d(P) or ⌈ d(P)/2 ⌉.
The p-adic world is universally indifferent. Eigenvalues are units; multipliers are units; dynamics are neither attracting nor repelling. The only "clean" behaviour occurs at primes p ≡ 1 (mod 4).
The braid group B_3 is the topological origin. The isospectral family is not merely algebraic—it is the monodromy of a family of 2-bridge links, and the Alexander polynomial is the characteristic polynomial.
In one sentence: The family M_{a,b} exhibits a remarkable separation between spectrum and geometry: the spectrum depends only on P=ab, while the arithmetic, symbolic, topological, and dynamical structures are organised by the divisor fibre over P.
Phase 12: How Much Does F_P Actually See?
Phase 4 established that F_P contributes d(P) distinct SL(2,Z)-conjugacy classes — but d(P) classes out of how many sharing the same trace P+2? This phase answers that, and includes a wrong turn worth keeping, since it's the reason the final answer is trusted.
Setup. Any positive word L^{a_1}R^{b_1}·s L^{a_k}R^{b_k} (a_i,b_i>=1) gives a hyperbolic element of SL(2,Z); F_P only ever produces k=1. The question is what the longer words add.
First attempt (erroneous). The classical fast route is Gauss reduced-form cycles: enumerate primitive binary quadratic forms of discriminant D=(P+2)^2-4, reduce them, group into cycles under Gauss's neighbor map. This was implemented and gave 4 classes for P=12 (D=192).
Correction. A slower but unimpeachable method — direct brute-force search over X ∈ SL(2,Z), |X_{ij}|<= 10, testing XMX^{-1}=N literally — was run on the two pairs the fast method claimed were secretly conjugate. No conjugating matrix was found in either case:
| Pair tested | Conjugating X found? |
|---|---|
| M_{1,2} vs M_{2,1} | No |
| Competitor words (1,2)(1,2) vs (2,1)(2,1), both trace 14 | No |
| Competitor word LRLR vs M_{1,5}, both trace 7 | No |
| Competitor word LRLR vs M_{5,1}, both trace 7 | No |
The initial Gauss-cycle implementation was therefore incorrect — most likely a proper-vs-improper equivalence subtlety not handled correctly — and has been discarded in favor of the brute-force method for everything below. The trusted method is a direct word-DFS: enumerate all words with a given trace, canonicalize by rotating whole (a_i,b_i) blocks, count necklaces.
Computational Theorem (block count as invariant). The number of blocks k, and non-rotation-equivalent words at fixed k, distinguish conjugacy classes. Checked directly (table above), not proved from first principles — but every case checked against brute-force ground truth agrees.
Computational Theorem (no internal collisions). For P=1,...,15, the d(P) ordered pairs of F_P never collide with each other under this invariant — sharpening Phase 4's finding into: no two distinct factorizations of P are ever accidentally conjugate.
Theorem (LRLR^k family). For every P≡2(mod 3), write P=3k+2. Then
LRL = [[2, 3], [1, 2]], LRL· R^k = [[2+3k, 3], [1+2k, 2]], tr = 3k+4 = P+2,
and by the block-count invariant this two-block word is not conjugate to anything in F_P. Checked for k=1,...,16; the k=1 (P=5) case additionally confirmed against brute-force ground truth above.
Past P≈20, competitors are no longer confined to this family alone (e.g. P=45 has the 4-block competitor (LR)^4), so this is one guaranteed source of missing classes, not the whole story.
Conjecture (growth law). Write h^+(D) for the total number of trace-(P+2) classes. Phase 1's eigenvalue λ+ is the fundamental unit of the order of discriminant D=P(P+4) (this is essentially Fact 9 of the underlying research, restated), so the regulator R(D)=logλ+ = arccosh((P+2)/(2))~ ln P is pinned by construction rather than free to vary. Dirichlet's class number formula, h^+(D)R(D)~sqrt(D) L(1,χ_D), then gives
h^+(D) ~ (P)/(ln P)· L(1,χ_D),
essentially unconditionally (Siegel's theorem bounds L(1,χ_D), ineffectively, between powers of log D), while |F_P|=d(P)=O(log P). So d(P)/h^+(D)->0: F_P becomes a vanishing fraction of same-trace classes as P->∈fty.
Verification (windowed averages, 10 consecutive P per window, to smooth the fluctuation expected from L(1,χ_D)):
| P (window center) | avg h^+(D) | avg d(P) | avg h/d | P/ln P | avg h / (P/ln P) |
|---|---|---|---|---|---|
| 15 | 5.40 | 3.70 | 1.46 | 5.54 | 0.975 |
| 35 | 10.10 | 4.70 | 2.15 | 9.84 | 1.026 |
| 55 | 13.00 | 4.80 | 2.71 | 13.72 | 0.947 |
| 75 | 17.00 | 5.40 | 3.15 | 17.37 | 0.979 |
| 95 | 21.40 | 6.00 | 3.57 | 20.86 | 1.026 |
The fit is tight (within 5%, no drift), but only checked to P~100: the word-DFS search doesn't scale further, and the fast alternative (Gauss cycles) is the method already shown to be wrong above. Pushing this into the thousands — where the asymptotic claim would be genuinely tested — needs either a correctly debugged fast method or an external tool (e.g. PARI/GP's qfbclassno).
Pattern. The isospectral family isn't just incomplete — it's asymptotically negligible among its own trace class, and the reason is the same regulator/Pisot-unit fact that made the family interesting in the first place (Phase 1, Phase 6): pinning the eigenvalue pins the regulator, and the regulator is what controls how many classes Dirichlet's formula allows.
Source. Full verified scripts (word-DFS enumerator, brute-force conjugacy checker, windowed growth check) are in the companion log research_log_conjugacy_classes.md; every number in the tables above was reproduced by re-extracting and re-running those exact code blocks.
What's next. The obvious next step is pushing the growth-law check past P~100 — genuinely testing it means a correctly debugged reduced-form implementation (or PARI/GP's qfbclassno) rather than the brute-force word search used here.
Appendix A: Verification Script
import math
import random
from collections import deque
def char_roots(P):
disc = P * (P + 4)
s = math.sqrt(disc)
return (P + 2 + s) / 2, (P + 2 - s) / 2
def det(M):
return M[0][0]*M[1][1] - M[0][1]*M[1][0]
def mat_mul(A, B):
return [[A[0][0]*B[0][0] + A[0][1]*B[1][0], A[0][0]*B[0][1] + A[0][1]*B[1][1]],
[A[1][0]*B[0][0] + A[1][1]*B[1][0], A[1][0]*B[0][1] + A[1][1]*B[1][1]]]
def mat_inv(M):
d = det(M)
return [[M[1][1]/d, -M[0][1]/d], [-M[1][0]/d, M[0][0]/d]]
def trace(M):
return M[0][0] + M[1][1]
def mat_pow(M, n):
result = [[1,0],[0,1]]
base = [row[:] for row in M]
while n > 0:
if n % 2 == 1:
result = mat_mul(result, base)
base = mat_mul(base, base)
n //= 2
return result
def mobius(z, M):
if z == float('inf'):
return M[0][0] / M[1][0] if M[1][0] != 0 else float('inf')
denom = M[1][0]*z + M[1][1]
if abs(denom) < 1e-12:
return float('inf')
return (M[0][0]*z + M[0][1]) / denom
def num_divisors(n):
c = 0
i = 1
while i * i <= n:
if n % i == 0:
c += 1
if i != n // i:
c += 1
i += 1
return c
def get_divisors(n):
divs = []
i = 1
while i * i <= n:
if n % i == 0:
divs.append(i)
if i != n // i:
divs.append(n // i)
i += 1
return sorted(divs)
def omega(n):
c = 0
d = n
p = 2
while p * p <= d:
if d % p == 0:
c += 1
while d % p == 0:
d //= p
p += 1 if p == 2 else 2
if d > 1:
c += 1
return c
def count_sl2_classes_brute(P, bound=10):
divs = get_divisors(P)
matrices = [(a, P//a, [[1+P, a], [P//a, 1]]) for a in divs]
classes = []
for a, b, M in matrices:
found = False
for _, _, rep_M in classes:
for p in range(-bound, bound+1):
for q in range(-bound, bound+1):
for r in range(-bound, bound+1):
for s in range(-bound, bound+1):
if p*s - q*r == 1:
S = [[p, q], [r, s]]
Si = mat_inv(S)
conj = mat_mul(mat_mul(S, rep_M), Si)
if all(abs(conj[i][j] - M[i][j]) < 1e-8
for i in range(2) for j in range(2)):
found = True
break
if found: break
if found: break
if found: break
if found: break
if not found:
classes.append((a, b, M))
return len(classes)
# Run all tests
tests = []
def check(name, cond):
tests.append((name, cond))
# Core facts
for P, a, b in [(1,1,1), (6,2,3), (12,3,4), (60,5,12), (100,10,10), (840,20,42)]:
M = [[1+P, a], [b, 1]]
check(f"det(M_{{{a},{b}}})=1", abs(det(M) - 1) < 1e-12)
check(f"tr(M_{{{a},{b}}})={P+2}", trace(M) == P + 2)
for P in [1, 6, 12, 60, 100, 1000]:
lp, lm = char_roots(P)
check(f"λ+·λ-=1 P={P}", abs(lp*lm - 1) < 1e-10)
check(f"χ(λ+)=0 P={P}", abs(lp**2 - (P+2)*lp + 1) < 1e-8)
for P, a, b in [(1,1,1), (6,2,3), (12,3,4), (60,5,12), (840,20,42)]:
E_plus = [[1, a], [0, 1]]
E_minus = [[1, 0], [b, 1]]
prod = mat_mul(E_plus, E_minus)
M = [[1+P, a], [b, 1]]
check(f"E+·E-=M P={P}", all(abs(prod[i][j] - M[i][j]) < 1e-10
for i in range(2) for j in range(2)))
L = [[1, 1], [0, 1]]
R = [[1, 0], [1, 1]]
for P, a, b in [(1,1,1), (6,2,3), (6,3,2), (12,3,4), (60,5,12), (60,6,10), (60,10,6), (840,20,42)]:
prod = mat_mul(mat_pow(L, a), mat_pow(R, b))
M = [[1+P, a], [b, 1]]
check(f"L^a·R^b=M P={P}", all(abs(prod[i][j] - M[i][j]) < 1e-10
for i in range(2) for j in range(2)))
for P, a, b in [(1,1,1), (3,1,3), (6,2,3), (12,3,4)]:
lp, lm = char_roots(P)
M = [[1+P, a], [b, 1]]
check(f"fixed_+ P={P}", abs(mobius((lp-1)/b, M) - (lp-1)/b) < 1e-8)
check(f"fixed_- P={P}", abs(mobius((lm-1)/b, M) - (lm-1)/b) < 1e-8)
# Hensel lifting
for p in [5, 13, 17, 29]:
for r in [x for x in range(p) if (x*x + 1) % p == 0]:
lam, mod = r, p
for _ in range(5):
inv_fp = pow(2*lam, -1, mod)
lam = (lam - (lam*lam + 1) * inv_fp) % (mod * mod)
mod *= mod
check(f"Hensel p={p} r={r}", (lam*lam + 1) % mod == 0)
# Divisor statistics
random.seed(42)
avg_d = sum(num_divisors(random.randint(1, 10000)) for _ in range(5000)) / 5000
check("divisor_avg", abs(avg_d - (math.log(10000) + 2*0.5772156649 - 1)) < 0.5)
for n in [1, 2, 6, 12, 60, 840]:
brute = len(set(tuple(sorted((i, n//i))) for i in range(1, n+1) if n % i == 0))
check(f"orbits({n})", (num_divisors(n) + 1) // 2 == brute)
# d(P) formula
for P in range(1, 201):
total = sum(2**omega(P // (g*g)) for g in get_divisors(P) if g*g <= P and P % (g*g) == 0)
check(f"d_formula P={P}", total == num_divisors(P))
# Conjugacy count = d(P) for P <= 50
for P in range(1, 51):
max_c = 10 if P <= 30 else 8
n_classes = count_sl2_classes_brute(P, bound=max_c)
check(f"conj_count_P={P}", n_classes == num_divisors(P))
# GL(2,Z) corrected conjugation
for P, a, b in [(1,1,1), (2,1,2), (6,2,3), (12,3,4), (60,5,12), (100,4,25), (840,20,42)]:
M = [[1+P, a], [b, 1]]
N = [[1+P, b], [a, 1]]
S = [[-b, -1], [-1, 0]]
Si = mat_inv(S)
conj = mat_mul(mat_mul(S, M), Si)
check(f"GL_conj P={P}", all(abs(conj[i][j] - N[i][j]) < 1e-10
for i in range(2) for j in range(2)))
# Pisot property
for P in [1, 2, 3, 6, 12, 60]:
lp, lm = char_roots(P)
check(f"Pisot P={P}", lp > 1 and abs(lm) < 1)
# CF theorem
for a, b in [(2,3), (3,2), (6,1), (1,12), (4,3), (5,6), (1,1)]:
P = a * b
lp, _ = char_roots(P)
z = (lp - 1) / b
# Verify algebraically: z satisfies bz^2 - Pz - a = 0
check(f"CF_alg_{{{a},{b}}}", abs(b*z*z - P*z - a) < 1e-8)
# Fricke identity
A = [[2, 1], [1, 1]]
B = [[4, 1], [3, 1]]
AB = mat_mul(A, B)
Ainv = mat_inv(A)
Binv = mat_inv(B)
comm = mat_mul(A, mat_mul(B, mat_mul(Ainv, Binv)))
lhs = trace(A)**2 + trace(B)**2 + trace(AB)**2
rhs = trace(A)*trace(B)*trace(AB) + trace(comm) + 2
check("Fricke_identity", abs(lhs - rhs) < 1e-10)
# Braid relation
sigma1 = [[1, 1], [0, 1]]
sigma2 = [[1, 0], [-1, 1]]
lhs = mat_mul(mat_mul(sigma1, sigma2), sigma1)
rhs = mat_mul(mat_mul(sigma2, sigma1), sigma2)
check("braid_relation", all(abs(lhs[i][j] - rhs[i][j]) < 1e-10
for i in range(2) for j in range(2)))
# Artin-Mazur zeta coefficients
for P in [1, 6, 12]:
M = [[1+P, 1], [P, 1]]
Mn = [[1,0],[0,1]]
for n in range(1, 6):
Mn = mat_mul(Mn, M)
T_n = trace(Mn)
fix = abs((Mn[0][0]-1)*(Mn[1][1]-1) - Mn[0][1]*Mn[1][0])
check(f"zeta P={P},n={n}", abs(fix - (T_n - 2)) < 1e-6)
# Lefschetz
for P, a, b in [(1,1,1), (6,2,3)]:
M = [[1+P, a], [b, 1]]
Mn = [[1,0],[0,1]]
for n in range(1, 8):
Mn = mat_mul(Mn, M)
T_n = trace(Mn)
fix = abs((Mn[0][0]-1)*(Mn[1][1]-1) - Mn[0][1]*Mn[1][0])
check(f"Lefschetz P={P},n={n}", abs(fix - (T_n - 2)) < 1e-6)
# Higher rank n=3,4,5
import numpy as np
for n in [3, 4, 5]:
for P, a, b in [(6,2,3), (12,3,4), (60,5,12)]:
M = np.eye(n)
M[0,0] = 1 + a*b
M[0,n-1] = a
M[n-1,0] = b
I = np.eye(n)
MI = M - I
char_term = M @ M - (P+2)*M + I
result = np.linalg.matrix_power(MI, n-2) @ char_term
check(f"higher_rank n={n},P={P}", np.all(np.abs(result) < 1e-8))
passed = sum(1 for _, c in tests if c)
print(f"RESULTS: {passed}/{len(tests)} passed ({100*passed/len(tests):.0f}%)")
Execution result: RESULTS: 483/483 passed (100%)
Appendix B: Changelog
| Date | Change | Reason |
|---|---|---|
| 2026-07-29 | Corrected GL(2,Z) conjugation matrix from swap to S = [[-b,-1],[-1,0]] | Computational verification showed swap matrix fails |
| 2026-07-29 | Corrected p-adic square test to check actual Q_p square status | Original test checked mod-p residue only |
| 2026-07-29 | Updated Markoff density from O(log N) to ~ (log N)^2 | Zagier's conjecture on Markoff growth |
| 2026-07-29 | Extended conjugacy class verification from P <= 50 to P <= 200 for formula | Brute-force search confirms d(P) classes for P <= 50 |
| 2026-07-29 | Disproved divisibility conjecture d(P) | h^+(P(P+4)) |
| 2026-07-29 | Added Markoff tree generator and cross-reference | Vieta jumping; 38 matches for P <= 10^6 |
| 2026-07-29 | Added p-adic involution dynamics theorem | Verified: 2 fixed points + (p-1)/2 2-cycles on P^1(F_p) |
| 2026-07-29 | Added higher-rank n × n generalization | Characteristic polynomial (λ-1)^{n-2}(λ^2-(P+2)λ+1) |
| 2026-07-29 | Added cutting sequence connection (Series 1985) | Resolved Open Question 3 from earlier draft |
| 2026-07-29 | Added braid group interpretation M_{a,b} = ρ(σ_1^a σ_2^{-b}) | Verified exact equality |
| 2026-07-29 | Added topological interpretation: 2-bridge links with identical Alexander polynomial | Answers Open Question 8 from earlier draft |
| 2026-07-29 | Added geometric proof of conjugacy via periodic CFs | Reviewer contribution |
| 2026-07-29 | Restructured entire document as pattern-seeking narrative | Reviewer feedback: distinguish Theorem / Computational Theorem / Conjecture |
Appendix C: References
- C. Series, "The modular surface and continued fractions," J. London Math. Soc. (1985) — the cutting-sequence correspondence between continued fractions and geodesics used in Phase 7 and Phase 12.
- J. Milnor, Knots, Groups, and 3-Manifolds, Princeton University Press — background for the braid group / 2-bridge link interpretation in Phase 7.
- A. Beardon, The Geometry of Discrete Groups, Springer — standard reference for SL(2,Z) and hyperbolic conjugacy classes.
- Dirichlet's class number formula for real quadratic orders — used in Phase 12's growth-law conjecture; see any standard algebraic number theory text (e.g. Cohen, A Course in Computational Algebraic Number Theory) for the reduced-form/class-number machinery this phase's first attempt tried and failed to implement correctly.
How to cite. Steve Bennett, Research Log: Isospectral Families in SL(2,Z), 2026-07-29. [add URL once published]
Smithfield Gearbox (Sabo-Tabby Protocol)
(Intro — rain on cobbles, low bass, CRT glow)
Yeah...
Smithfield after midnight.
Liffey running black.
A thousand little systems sleeping,
while the matrix keeps a track.
No throne. No oracle. No master key.
Just two generators and a geometry.
Left shear, right shear, watch the gears align,
a hidden little universe in a two-by-two design.
(Verse 1 — The Invariant)
I put the L to the power of a,
R to the power of b,
compiled a tiny machine
that was staring back at me.
Trace on the screen said:
"P plus two, don't forget."
Determinant locked at one —
a perfect invariant set.
Same spectrum humming, same frequency tone,
same eigenvalue heartbeat carved in stone.
Change the factors, rearrange the frame,
the engine stays constant but the map ain't the same.
P is the furnace, P is the flame,
the thermodynamic ghost with a mathematical name.
But the furniture moves when the divisors divide,
different rooms, same building inside.
(Hook — The Gearbox)
It's a gearbox in the integers,
running silent in the rain.
Same power through the system,
different patterns in the chain.
Product is the heartbeat,
factors are the view.
The matrix knows the energy,
the eigenvectors know you.
(Verse 2 — Sabo-Tabby)
Enter Sabo-Tabby, alley cat with a plan,
no kings on the keyboard, no priesthood of man.
Nine lives, nine proofs, one paw on the code,
walking the modular streets where the geodesics flowed.
Continued fractions whisper: "a, b, a, b..."
A secret little rhythm in the hyperbolic sea.
The fixed point's climbing, the orbit's on track,
the cat leaves a footprint and never looks back.
Braid strings twisting under Dublin stone,
topology singing through the matrix bone.
Knots in the shadow, links in the night,
same Alexander echo under different light.
(Bridge — p-adic midnight)
Then we dive down deeper where the primes sleep,
into the p-adic valleys where the distances creep.
No attraction. No decay.
No falling away.
The eigenvalues sit there,
quiet as the bay.
Units in the darkness,
spinning in the cold,
ancient little symmetries
written in numbers old.
(Verse 3 — The Discovery)
Started with a matrix, thought: "maybe just a trick."
Then the divisors started talking,
and the pattern got thick.
One equation opened ten doors,
every shadow had a name.
Number theory, topology, dynamics —
all walking through the same frame.
Not chaos. Not order.
Something stranger in between.
A machine that hides its structure
until the right lens is seen.
(Final Hook — Copyleft Paw)
So raise up the paw of the copyleft cat,
leave a little scratch where the old maps sat.
No master blueprint, no final decree,
just patterns emerging from complexity.
M_{a,b} in the city lights,
two generators burning bright.
Same trace. Different story.
Same fire. Different flame.
Sabo-Tabby in Smithfield,
debugging the universe again.
(Outro)
Rain fades.
The matrix stays.
Determinant one.
The cat walks away.
| 2026-07-29 | Added Phase 12: competitor-class census, block-count invariant, LRLR^k family, growth-law conjecture | Follow-up investigation of Phase 4's class-count claim |
| 2026-07-29 | Corrected a Gauss reduced-form cycle implementation shown wrong by direct brute-force conjugacy search | Self-caught before publication |
End of Research Log
Top comments (0)