The Arithmetic Manifold

Executive Summary (read this first)#

One sentence: The Arithmetic Manifold is the theory that explains why our solver is faster — and it connects hard combinatorial problems to the deepest patterns in number theory.

Who should read this: A technical reader (CTO, staff engineer, research collaborator) who wants to understand why the engine works, not just that it works. If you’re evaluating Navokoj for a business problem, the homepage is the better starting point.

What you’ll get from this page: The five core mechanisms that make the solver work, explained with their mathematical justification, but with a plain-English gloss before each technical block.

The Manifesto#

The Arithmetic Manifold is a research program that treats constraint satisfaction, optimization, and physical simulation as different projections of the same underlying geometric structure.

The central thesis:

The reason NP-hard problems are hard is the same reason primes are irregular: both arise from the competition between geometric structure (spectral gap, Laplacian) and arithmetic structure (prime distribution, modular symmetry). The Riemann Hypothesis governs whether this competition stabilizes or destabilizes.

The Five Proof Points#

1. Why primes?#

In plain terms: Every constraint in your problem gets a unique “fingerprint” derived from a prime number. Two constraints never produce the same fingerprint, so they never compete for the same gradient signal. This is what keeps the solver from getting stuck.

Each constraint in a problem gets a unique prime weight:

$W(p_c) = \frac{1}{1 + \ln(p_c)}$

Why primes?

Fundamental Theorem of Arithmetic: Primes are multiplicatively independent — no prime can be expressed as a product of other primes
Smaller primes exert stronger force; larger primes refine with logarithmic suppression
This guarantees no two constraints create identical gradient signatures

Ablation studies show: 4x speedup and 75% reduction in topological complexity when using prime weights vs. uniform weights.

2. Why multiplicative instead of additive?#

In plain terms: Traditional solvers combine constraints by adding them up (L = A + B), which creates gradient conflicts — one constraint pulling against another. We combine them by multiplying (L = A × B), which preserves the gradient signal and lets the solver find a valid solution without thrashing.

Traditional optimization:

Loss = DataLoss + λ × PhysicsLoss   (additive — causes gradient conflicts)

Arithmetic Manifold approach:

Loss = DataLoss × ConstraintFactor(PhysicsLoss)   (multiplicative — preserves gradient flow)

The constraint factor combines:

Euler Gate: G(v) = ∏(1 - p^{-τv}) — attenuates loss when constraints are satisfied
Exponential Barrier: B(v) = exp(γv) — amplifies gradients on constraint violations

This creates a “superconducting” optimization landscape where gradients flow without resistance when physics constraints are met.

3. Why do hard problems have a “phase transition”?#

In plain terms: Hard problems don’t get hard gradually — they crack suddenly at a specific density or size. Most solvers don’t see this coming and fail. We monitor the energy landscape and detect the crack before it happens, then navigate around it.

Hard problems “crack” at phase transitions. When the landscape fractures (thermodynamic fracture), standard gradient descent fails.

BAHA detects fractures via:

\rho(\beta) = \left| \frac{d}{d\beta} \log Z(\beta) \right|

When ρ exceeds a threshold, the system enters a saddle-node bifurcation governed by the Lambert W function:

$\ln K^* = -C \cdot W\left(-\frac{1}{C}\right)$

This ln K / ln ln K scaling is a fingerprint of the prime weight function — no other weighting produces this specific scaling law.

4. What is the “partition function” and why does it appear everywhere?#

In plain terms: The partition function is a single quantity from statistical mechanics that summarizes the entire energy landscape of a problem. It turns out to be the right lens for every project we ship — SAT, scheduling, neural networks, even physical simulation. One tool, many domains.

Every project begins with the partition function:

$Z(\beta) = \sum_{s \in \mathcal{S}} e^{-\beta E(s)}$

where $\mathcal{S}$ is the state space, $E(s)$ is the energy of state s, and β is inverse temperature.

Project	Role of Z(β)
BAHA	Monitors ρ =
Navokoj	Defines energy E = -w·log P(clause satisfied) for geometric flow
Multiplicative PINN	Constraint factor C(v) ~ e^{γv} is a local partition function
Casimir SAT	Langevin dynamics samples from P(s) ∝ e^{-βE(s)}
Spectral-Multiplicative	Heat kernel trace Tr(e^{-tL}) ≈ Z(β) on graph Laplacian

5. Why does the Riemann Hypothesis show up in a solver?#

In plain terms: The Riemann Hypothesis — the deepest unsolved problem in number theory — turns out to be exactly the stability condition our solver needs. If the prime distribution is well-behaved (the RH case), our solver stays stable as problems scale. If primes misbehave (a counterexample to RH), our solver would eventually destabilize too. We don’t solve the Riemann Hypothesis. We embed it as the boundary we engineer against.

As clause count K grows, two forces compete:

Geometric Weakening: The spectral gap closes at rate λ₂(G_K) ~ K^{-γ}
Prime Fluctuation Decay: Relative prime variance vanishes at rate Φ(K) ~ K^{σ-1}

where σ is the supremum of real parts of ζ(s) zeros.

Stability requires: 1 - σ > γ

The Asymptotic Lock Conjecture: On critical mesh-like geometries (γ → 1/2), preserving stability strictly requires σ → 1/2 (the Riemann Hypothesis).

“NitroSAT does not prove the Riemann Hypothesis. But it embeds RH as a phase boundary: the solver functions as a physical instrument whose asymptotic stability threshold coincides exactly with the critical line Re(s) = 1/2.” — thermodynamic-number-line/index.html

The Grand Synthesis#

The Arithmetic Manifold is not a metaphor — it is a load-bearing mathematical structure:

Manifold Concept	Computational Implementation
Operators as Neurons	Energy gradient operators per constraint
Geometric Curvature	Energy landscape shaped by constraint violations
Dynamical Flow	Adiabatic cooling schedule (β ramp)
Functional Compression	Prime-weighted operator basis
Self-Modification	Continuous state evolution without backtracking
Uncertainty as Geometry	Soft probabilities before collapse
Identity Kernel	Prime weighting ensures operator uniqueness

What This Means Practically#

The Arithmetic Manifold provides:

A common language — The partition function, Euler products, and Lambert W function appear across all projects
Proven speedups — Prime weighting alone delivers 4x speedup; multiplicative enforcement eliminates gradient conflicts entirely
Theoretical grounding — The connection to the Riemann Hypothesis isn’t decorative — it’s a stability condition with measurable consequences
A path to scale — As problems grow, the RH connection tells us when our algorithms will remain stable and when they’ll collapse

What this is NOT#

To help you place this work in context, here is what the Arithmetic Manifold is not:

Not a general SAT competitor. We are not trying to replace Z3, Kissat, or CaDiCaL on generic Boolean SAT. Our wins are on weighted, structured MaxSAT.
Not a ZK rollup or proving system. Zero-knowledge cryptography is a future wedge (2027+) that builds on top of our solver. It is not the product today.
Not a neural network. We borrow ideas from statistical mechanics and number theory. The Multiplicative PINN project uses neural networks, but the core solver does not.
Not a proof of the Riemann Hypothesis. We embed RH as a stability condition. If RH is ever disproven, our solver’s stability guarantee weakens — but we do not claim to have settled the conjecture.
Not a replacement for exact solvers. When you need a certified optimality certificate, use MaxHS, Pacose, or RC2. Our mini engine gets close but is not 100% certified.