3-State Asymmetrically Weighted Oscillator Optimizer

• The paper introduces a multi-phase CMOS oscillator network based on a 3-state Potts model to encode SAT constraints with asymmetric coupling for clause satisfaction.
• It employs continuous-time oscillator dynamics and a global Lyapunov function to minimize a Potts-Hamiltonian energy landscape, ensuring convergence to low-energy solutions.
• The design integrates clause, weighted sum, and energy AND networks to reduce hardware complexity while supporting diverse NP problems like SAT, TSP, and integer programming.

Searching arXiv for the cited paper to ground the article in the source record. The 3-State Asymmetrically Weighted Oscillator Optimizer is a multi-phase CMOS oscillator network proposed as an analog solver for SAT-like and other NP formulations in "Oscillator Formulations of Many NP Problems" (Cai et al., 16 Sep 2025). It is formulated within a 3-state Potts-style oscillator model in which each oscillator phase encodes one of three discrete states, $\sigma_i \in \{F, T, B\}$, standing for False, True, Blue. The design is "3-state" because each oscillator is constrained to one of three discrete phase states rather than two, and "asymmetrically weighted" because it uses different coupling strengths for different directions of influence, denoted $W_l \neq W_r$, in a clause-centered feedback mechanism (Cai et al., 16 Sep 2025). Within the paper, the optimizer is presented as part of a broader program of expressing many NP problems through first-order and multi-phase Potts Hamiltonians and then minimizing the resulting energy by oscillator dynamics.

1. State representation and architectural principle

The optimizer uses three discrete phase states corresponding to the phase angles

$\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$

Equivalently, the three states are represented as 2D unit vectors at the vertices of an equilateral triangle: $\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$ This triangular embedding gives the model a Potts-style state geometry rather than a binary Ising geometry (Cai et al., 16 Sep 2025).

The paper states that the third state, Blue, is not just a third logic value; it is introduced to handle dynamics/control of updates and to reduce the number of connections required for clause-level feedback. In the architecture summarized in the paper, the optimizer uses four components: a base network, a clause network, a Weighted Sum network, and an Energy AND network. The 3-state dynamics constitute the primitive mechanism through which these blocks implement SAT and other NP encodings (Cai et al., 16 Sep 2025).

The architectural asymmetry is centered on a clause auxiliary node $r$. The clause network introduces this node so that literals $l_1,l_2,l_3$ influence $r$, and $r$ then feeds back to the literals with a different weight. The paper explicitly states that $W_l < 0$ for the influence of literals on $r$, while $W_l \neq W_r$0 for the feedback from $W_l \neq W_r$1 to literals. This asymmetry is used to encode a clause-satisfaction heuristic: if a clause is already satisfied, $W_l \neq W_r$2 becomes Blue, which exerts no preference; if the clause is unsatisfied, $W_l \neq W_r$3 becomes False, which pushes the literals toward True (Cai et al., 16 Sep 2025).

2. Oscillator dynamics and Potts-Hamiltonian formulation

The starting point is an oscillator/Kuramoto description. For oscillator $W_l \neq W_r$4, the voltage is written as

$W_l \neq W_r$5

and the phase evolution is modeled by

$W_l \neq W_r$6

The associated Lyapunov function is

$W_l \neq W_r$7

which decreases over time. The paper states that this resembles the Potts Hamiltonian

$W_l \neq W_r$8

(Cai et al., 16 Sep 2025).

For the 3-state formulation, the interaction between two states is given by the dot product of their vector embeddings: $W_l \neq W_r$9 This yields the Potts-like interaction rule in which same states have positive alignment and different states are penalized (Cai et al., 16 Sep 2025).

The paper further defines a local field

$\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$0

together with the update rule

$\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$1

The network evolves until a stable state or local minimum is reached. In the optimizer, this rule is the local energy-minimizing state update for the 3-state Potts Hopfield dynamics (Cai et al., 16 Sep 2025).

3. Clause-network mechanism and asymmetric weighting

For a clause with literals $\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$2, the auxiliary node $\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$3 is set according to the clause’s satisfaction state: if all $\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$4, then $\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$5; otherwise $\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$6. The paper gives a system of inequalities that enforces this behavior: $\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$7 which simplify to

$\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$8

(Cai et al., 16 Sep 2025).

The simplified conditions state the key directional asymmetry directly: $\left\{0,\tfrac{2}{3}\pi,\tfrac{4}{3}\pi\right\}.$9 and $\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$0. Within the clause logic, this means the literals collectively determine whether the clause feedback node acts neutrally or correctionally. When a clause is satisfied, the node becomes Blue, and its effect is neutral. When a clause is unsatisfied, the node becomes False, and the positive feedback weight pushes literals toward True. The paper characterizes this as direction-dependent coupling designed for clause satisfaction dynamics rather than merely unequal edge weights in a static graph (Cai et al., 16 Sep 2025).

The three fixed sources connected to $\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$1 are merged into a single Biased Source: $\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$2

$\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$3

This modification reduces the number of connections in the SAT construction while preserving the clause-network biasing function (Cai et al., 16 Sep 2025).

A plausible implication is that the Blue state serves both as a representational state and as a control state in the clause-feedback circuit. The paper explicitly links this state to reduced connection count and update control, which distinguishes the design from binary oscillator encodings.

4. Network blocks for objectives and constraints

The paper’s optimizer combines clause satisfaction with direct encoding of linear and quadratic objectives. The Weighted Sum network encodes linear objectives of the form

$\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$4

Each $\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$5 is coupled to a fixed True-phase oscillator with coupling $\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$6, inducing the Potts Hamiltonian

$\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$7

The linear map

$\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$8

preserves the optimizer’s solution (Cai et al., 16 Sep 2025).

The Energy AND network is introduced to encode quadratic terms $\vec{s}_F = \begin{pmatrix}1\0\end{pmatrix},\quad \vec{s}_T = \begin{pmatrix}-1/2\ \sqrt{3}/2\end{pmatrix},\quad \vec{s}_B = \begin{pmatrix}-1/2\ -\sqrt{3}/2\end{pmatrix}.$9. For binary variables $r$0, the paper uses oscillator states such that

$r$1

Thus the quadratic objective

$r$2

is encoded up to a linear transform (Cai et al., 16 Sep 2025).

These two blocks allow a decomposition that recurs throughout the paper’s problem reductions: linear penalties are mapped to the Weighted Sum network, while pairwise interactions are mapped to the Energy AND network. Hard constraints may also be translated into CNF clauses and handled by the clause network. The paper repeatedly describes the encoding pattern as: encode combinatorial choices as binary variables $r$3, translate hard constraints into CNF clauses or penalty terms, implement linear penalties with Weighted Sum, implement pairwise terms with Energy AND, and then let oscillator phase dynamics minimize the corresponding Potts energy (Cai et al., 16 Sep 2025).

This suggests that the optimizer is intended not only as a SAT solver but as a unifying analog substrate for mixed logical and algebraic combinatorial formulations.

5. Problem mappings and representational scope

The paper states that the 3-state oscillator optimizer is used as a hardware-like analog solver for a broad range of NP formulations, including SAT / 3-SAT / mixed k-SAT, 0-1 integer programming and optimization, Hamiltonian path/circle, Traveling salesman problem, Clique, Node cover, Set cover, Chromatic number, Feedback node set, Feedback arc set, Clique cover, Exact cover, Hitting set, 3-dimensional matching, Number partitioning, Knapsack, Graph partitioning / min cut, Independent set, Max-cut, Maze solving, Shortest path, and more (Cai et al., 16 Sep 2025).

Several explicit mappings are given. For 0-1 Integer Programming, with

$r$4

the paper uses the penalty form

$r$5

After expansion and using $r$6, this becomes

$r$7

The linear terms are assigned to Weighted Sum, and the quadratic terms to Energy AND (Cai et al., 16 Sep 2025).

For the Traveling salesman problem, the paper states that SAT-like Hamiltonian-cycle constraints are used to ensure valid tours, after which weighted edge costs are minimized via Energy AND: $r$8 For Set cover, the optimization form is

$r$9

Here the objective uses Weighted Sum, while the coverage constraints become CNF disjunctions (Cai et al., 16 Sep 2025).

The paper also emphasizes direct support for mixed k-SAT, avoiding extra padding variables. This is presented as a structural advantage over formulations that first convert all clauses to 3-SAT (Cai et al., 16 Sep 2025).

6. Algorithmic behavior, scaling, and demonstrations

The optimizer is not presented as a discrete combinatorial algorithm with explicit branching. Instead, the paper describes it as a continuous-time analog relaxation process: initialize oscillators, apply external fixed-phase sources (True, False, Blue / biased source), let couplings drive phase changes, and repeatedly update until a stable state is reached. The local update rule remains

$l_1,l_2,l_3$0

(Cai et al., 16 Sep 2025).

The stability claim is based on the existence of a global Lyapunov function $l_1,l_2,l_3$1, which is minimized over time by the oscillator network. The paper states that this suggests monotonic descent to a local minimum, assuming the dynamics obey the modeled coupling rules. Practical stability conditions highlighted in the paper include using large bias weights to constrain nodes to $l_1,l_2,l_3$2 when Blue should neutralize them, using $l_1,l_2,l_3$3 in mixed objective settings so constraints are satisfied before optimizing the secondary objective, and selecting weights so unsatisfied clauses exert stronger corrective influence than irrelevant satisfied clauses (Cai et al., 16 Sep 2025).

For SAT, the network size scales linearly in variables $l_1,l_2,l_3$4 and clauses $l_1,l_2,l_3$5: $l_1,l_2,l_3$6 or, with the Biased Source,

$l_1,l_2,l_3$7

The paper gives the following comparison.

Design	# Nodes	# Connections
Oscillator Optimizer	$l_1,l_2,l_3$8	$l_1,l_2,l_3$9
w/ Biased Source	$r$0	$r$1

It also explicitly notes that for 110-variable, 1400-clause SAT, the number of connections drops from 8733 to 5933 when using the Biased Source (Cai et al., 16 Sep 2025).

The reported demonstrations are illustrative rather than benchmark-heavy. The paper provides examples and truth tables for 3-state Potts Hopfield updates, clause satisfaction dynamics, maze solving, logic gates, and mixed k-SAT handling, but it does not report large-scale experimental accuracy, runtime, or hardware benchmark results for this specific optimizer in the way a full systems paper would (Cai et al., 16 Sep 2025).

7. Assumptions, limitations, and interpretive context

The paper includes several assumptions. It states that many reductions assume the input instance has a solution or is otherwise "well-formed" for the mapping. For some formulations, parameters are assumed large enough to avoid trivial cases. For graph partitioning, $r$2 is assumed even. Some reductions to SAT rely on nontrivial clause blowup or auxiliary variables (Cai et al., 16 Sep 2025).

Several limitations are stated directly. The hardware has not yet been designed, so transistor/component scaling is not established. The method is an analog optimizer and therefore naturally searches for low-energy states, not guaranteed global optima. Some problems are described as inefficiently handled; Sequencing and Steiner Tree are mentioned as requiring tree-like structures that make the oscillator approach inefficient. XOR/XNOR are not directly implementable using the presented oscillator logic-gate scheme. The paper is also characterized as mainly a unification/formulation work with limited detailed experimental validation (Cai et al., 16 Sep 2025).

Within that scope, the optimizer’s significance lies in the way it combines a Clause network for satisfiability constraints, a Weighted Sum network for linear objectives, an Energy AND network for quadratic terms, and a Base network that constrains variable/complement pairs and ties them to fixed sources. The network phase dynamics descend an energy landscape whose minima correspond to solutions of the encoded NP problem. This suggests a general oscillator-based route from combinatorial formulation to Potts energy minimization, with the 3-state and asymmetrically weighted design serving specifically to implement clause satisfaction dynamics and mixed objective handling in a structurally compact form (Cai et al., 16 Sep 2025).