WorksMathematics
A Computational Verification of the Markov Uniqueness Conjecture for
Abstract
We describe a finite computational verification of Markov Uniqueness Conjecture for normalized positive integer solutions of with . The archived run reports triples with distinct maximal coordinates.The method combines a disjoint tree decomposition, certified fixed-point range bounds, a deterministic 48-bit candidate screen,and replay modulo . We give the mathematical justification and a reproducibility procedure covering package checks,retained-data auditing, and a full rerun. The conclusion depends on correct execution of the exhaustive screen and replay; the retained audit does not regenerate the deleted screening buckets.
Introduction
The positive integer solutions of the Markov equation
have a classical connection with Diophantine approximation and indefinite binary quadratic forms. Both the spellings Markov and Markoff occur in the literature. We use the former in the text and preserve original spellings in references. A solution is normalized when $1 \le x \le y \le z$; permutations are not counted as different triples.
The uniqueness conjecture states that a positive integer is the largest coordinate of at most one normalized Markov triple. It does not assert that a Markov number occurs in only one triple in any coordinate. Baragar [2] gives this formulation, attributes the conjecture to Frobenius, and reports a verification for maximal coordinates below . Button [3] proves uniqueness when the largest coordinate is prime. These results provide context; the present computation does not depend on primality testing or the classification of special families.
Zagier [1] studies the counting function for normalized triples and proves an asymptotic law with leading constant multiplying . His paper also contains numerical data up to . Counting triples and establishing uniqueness of their maxima are different tasks. We make no priority claim and do not present these references as an exhaustive survey of previous computational bounds.
This article documents the calculation archived on 26 October 2026 [4]. Its purpose is to state the finite result precisely, expose the assumptions connecting the implementation to that result, and provide reviewable evidence. New checks of the retained data are reported separately from the original execution.
Write . For , define
The tested statement is injectivity of , . The bound is inclusive; is its exponent.
Reported computational finding. For , the archived implementation reports
and injectivity of . The final record has matching visited and expected counts and zero boundary ambiguities, missing candidate groups, and 128-bit residue collisions. Subject to the correctness conditions below, the number in (2) is consequently also the number of distinct maximal Markov numbers in the tested range.
Section gives the enumeration, range test, and collision exclusion that make the computation possible. Section describes the retained evidence and the steps for checking it or repeating the calculation.
Method
Tree enumeration and partition
The Vieta involution replaces a coordinate by . After the exceptional triples and , the normalized tree starts at and has children
For fixed , the polynomial has roots and . Since outside the exceptional cases, . Mutating the maximum therefore gives the unique descending neighbor; mutating either smaller coordinate gives an ascending neighbor. Equal coordinates occur only in the exceptional triples: if , then , and writing gives , hence and ; the case is excluded by . Descent decreases a positive integer and terminates in the exceptional prefix. Reversing it proves that (3) enumerates every normalized triple exactly once.
Exact integer enumeration up to gives a prefix of 1456 nodes and 1455 frontier roots, each the first node on its path with maximum exceeding . Thus
where each subtree includes its root. Increasing maxima and unique descent give disjointness and coverage. Exact reconstruction of the stored frontier checks the 1456 prefix records, the 1455 roots, their ordering, and their 256-bit coordinate encodings. In particular, every root has maximum below .
For replay identification, is assigned , with and . Subtracting the smaller coordinate from the larger recovers the path uniquely. The two exceptional prefix nodes use and . These labels identify nodes; they are not the size bounds used next.
Certified range decisions
Large exact integers are avoided during the main traversal by storing modular coordinates and bounds for
Because is increasing, is equivalent to .
Lemma 2.1 (Additive defect). For a normalized Markov triple with ,
Proof. Put , , , and . The larger-root relation gives
Also , so . Together with , this gives ; and . The mean value theorem bounds the defect by .
Each frontier seed stores integer endpoints enclosing and . Their containment is checked by an integer-only certificate, independent of the high-precision calculation that proposed them. At scale , exact integer square roots give outward bounds for the radicals in (5). After reducing a logarithm to , set and use
Outward rounding and exact cross-multiplication then certify the endpoints. For the denominator, use . Indeed, with and , this difference is , bounded by . All 2910 endpoint checks pass in the recorded certificate.
Let . The certified Q255 endpoints are converted to Q63 bounds by
and likewise for . Bounds for each new maximum are obtained by addition: and . This addition encloses the size function; the modular coordinates continue to follow the Vieta recurrence.
To justify the lower bound, put . Every maximum constructed in a frontier subtree exceeds , so its additive defect is below . The subtraction of one in (6) gives each smaller seed coordinate a lower margin of at least . If both current coordinates have this margin, then
The margin survives induction. Upper bounds follow from . Consequently a child is accepted if , excluded if , and otherwise marked ambiguous. Excluded children may be pruned because all their descendants have larger maxima. Zero ambiguities therefore gives an exhaustive range decision.
Roots are inserted before their children are tested, so their preconditions are checked explicitly: positive lower bounds, exact in-range root values, and root upper sums at most $q. For the supplied seeds, the smallest lower endpoint is and the largest root upper sum is . At an included parent, positivity and imply and ; hence child upper sums fit in an unsigned 64-bit word. The implementation checks carries, parent invariants, and path-label overflow. A positive overflow in a lower sum certifies exclusion.
Candidate screening and modular replay
The first traversal associates to a maximum the deterministic key
The mixing function successively applies 0xbf58476d1ce4e5b9, 0x94d049bb133111eb, and , with products modulo . The upper eight key bits select one of 256 buckets; the remaining 40 bits are stored in five bytes.
The subtrees are balanced into groups. Every unordered group pair, including the diagonals, is processed, giving 91 jobs. A job includes the prefix and either one group or two distinct groups. Two nodes in the same group meet in its diagonal job; two nodes in different groups meet in their joint job; the prefix meets every group. Thus every repeated key appears as a duplicate in at least one job. Each node is inserted only once within a job.
The bucket scanner stores the full 40-bit payload as key + 1 in an open-addressed table, with zero reserved for an empty slot. Exact payload comparison and linear probing handle table-address collisions; a load at most 0.60 leaves an empty slot. The repeated keys from all jobs are sorted and deduplicated. With and , the total processed records are
which includes repeated coverage across jobs.
A second traversal computes (3) modulo and records every occurrence of a listed key. Each 28-byte record contains a four-byte candidate index, a 16-byte residue, and two four-byte path labels. The prefix and each frontier subtree have separate shards. Polynomial evaluation commutes with reduction modulo , so exact seed residues determine the residues of all descendants. The two-limb implementation propagates the multiplication carry and subtraction borrow; equality compares both 64-bit limbs.
The final merge requires at least two occurrences per candidate key and checks for equal residues. It processes at most four million keys per block, giving 26 blocks. A residue collision requires exact investigation before any uniqueness conclusion.
Theorem 2.2 (Collision exclusion). Suppose the traversal visits every element of exactly once, the screen lists every repeated key, and replay records every occurrence with its correct residue modulo . If no two occurrences of any listed key have equal residues, then the maximal-coordinate map is injective.
Proof. Equal maxima at distinct nodes give equal keys, so both occurrences are replayed under a listed key. Their residues are equal, contradicting the hypothesis.
This implication is deterministic: unequal residues certify unequal integers, regardless of the size of . No statistical assumption about the fingerprint distribution is required.
Recorded result
The final computation record in the supplementary materials reports the quantities in Table 1. Its visited and expected counts agree, and its three failure counters are zero. The complete screen and replay, with the correctness conditions above, yield the finite result stated in the introduction.
| Quantity | Value |
| Visited nodes and expected nodes | 239 536 376 556 |
| Prefix nodes / frontier subtrees | 1 456 / 1 455 |
| Groups / pair jobs | 13 / 91 |
| Candidate keys | 101 871 768 |
| Candidate occurrences | 203 772 186 |
| Replay shards / merge blocks | 1 456 / 26 |
| Boundary ambiguities | 0 |
| Missing candidate groups | 0 |
| 128-bit residue collisions | 0 |
Table 1. Recorded result for .
Reproducibility
Supplementary materials and environment
The supplementary materials accompanying this paper contain the computational evidence and reproducibility instructions [4]. They were updated using Codex during manuscript preparation. Consult these materials for the directory layout, detailed commands, integrity checks, and test procedures.
The retained production evidence consists of the source programs, frozen seeds and inputs, 91 candidate lists, merged candidates, 1456 replay shards, logs, manifests, and tests. The review scripts and audit records used during manuscript preparation are described below.
Production C# code is compiled through PowerShell 7 with a .NET runtime supporting UInt128; Python handles certificates, metadata, grouping, and candidate merging. The supplied audit uses Python and NumPy. The recorded review environment is Python 3.12.14, NumPy 2.3.5, and .NET 10.0.11; these are not established as the original run’s exact versions. The full production workflow uses a Windows process-held lock. Use new working and scratch directories, keep the archived evidence unchanged, and run the original Python checks with assertions enabled.
Verification and rerun procedure
The following sequence separates checks on retained evidence from a new execution. The supplementary instructions give the commands and working-directory settings for each step.
Check package integrity. Follow the integrity-checking procedure in the supplementary materials. Check file lengths and hashes, frozen-input bindings, and the target result fields. Hash agreement binds the files to the manifest; it does not establish arithmetic correctness.
Audit retained candidates and residues. Run the retained-data audit described in the supplementary materials in a new working directory, allowing several GB of temporary space. It reconstructs the exact frontier, checks all pair memberships and the exact union of their saved candidate lists, and validates every shard’s structure, bindings, count, and digest. It recomputes candidate fingerprints and multiplicities for all 203 772 186 occurrences. Independently of the production merge, it partitions residues by their top six bits and sorts each partition to detect equal 128-bit values.
The supplied audit record reports a pass on all 1456 shards, zero repeated residues, and 4368 reconstructed sample paths matching their stored residues. Candidate multiplicities are 101 843 122 keys occurring twice, 28 642 occurring three times, and four occurring four times. Full residue reconstruction is sampled; all saved residues are included in the equality test.
Recompute certificates and small tests. Recompute the integer seed certificate and run the original exact regression in a separate copy. At , it compares 9670 exactly enumerated triples with an eight-bit screen that forces collisions. Also run the supplementary tests for full-width arithmetic, boundary decisions, and failure handling. The supplied records report successful checks of all 16428 small-instance child decisions, a full-width test across 256 buckets, extremal modular arithmetic, classifier boundaries, and malformed-input rejection. These tests exercise the code; they do not replace the target computation.
Recheck or regenerate replay. Follow the supplementary replay instructions to prepare a fresh destination. Rechecking saved shards validates and merges the retained records. A fresh traversal instead revisits all target nodes modulo ; this step still uses the saved 48-bit candidate list.
Repeat the full calculation. Follow the supplementary instructions for a full rerun in a separate directory: regenerate and certify seeds, count the subtrees, construct the 13-group plan, run all 91 pair jobs, merge candidates, and execute fresh replay and its 26-block merge. The largest pair has approximately 184.27 GB of five-byte payload; the supplied guidance recommends at least 250 GB of available disk. The original target replay used 16 workers. The archived multi-day execution included interruptions and resumed jobs, so its partial timers are not a reliable total-runtime estimate.
A successful rerun at must report the expected 239 536 376 556 visited nodes, zero boundary ambiguities, zero missing candidate groups, and zero 128-bit residue collisions, with matching input and source bindings and a successful final status. Agreement between counting and replay is a consistency check because they share range logic.
The original temporary screening buckets were deleted and are absent from the retained evidence. Their digests, pair candidate lists, and replay shards remain. Accordingly, the retained-data audit verifies the saved candidate union and occurrence records; a fresh full calculation is required to regenerate and recheck the original exhaustive screen. The review records document the retained-data audit and small tests; a fresh target enumeration was not performed during manuscript preparation. The result concerns the stated finite range and does not resolve the unrestricted conjecture.
Declaration of AI use
OpenAI’s Codex was used extensively throughout this project, including the development of the computational method; generation and revision of the enumeration, screening, replay, and auditing code; execution and inspection of computational runs; formulation and checking of mathematical arguments; and drafting and editing of this manuscript. Codex was also used to update and assemble the supplementary reproducibility materials during manuscript preparation. Both the original calculation and the separately implemented retained-data audit were conducted within this Codex-assisted workflow. The latter provides an additional implementation check, rather than independent external replication. The authors take responsibility for the mathematical claims, software, computational evidence, and final text.
References
- [1]D. Zagier, On the number of Markoff numbers below a given bound, Mathematics of Computation 39 (1982), no. 160, 709–723. doi:10.2307/2007348.DOI
- [2]A. Baragar, On the unicity conjecture for Markoff numbers, Canadian Mathematical Bulletin 39 (1996), no. 1, 3–9. doi:10.4153/CMB-1996-001-x.DOI
- [3]J. O. Button, The uniqueness of the prime Markoff numbers, Journal of the London Mathematical Society (2) 58 (1998), no. 1, 9–17. doi:10.1112/S0024610798006292.DOI
- [4]Markov finite-range computation at 10^5000000 (see https://zenodo.org/records/23670834), computational evidence and manuscript-stage reproducibility materials. Supplementary materials record as this paper; verification procedure in Section 3.