Phase-Synchronization Control Theory

Abstract

This document presents the theoretical basis for using Kuramoto phase dynamics as a universal control abstraction — a domain-agnostic framework where any system with coupled cycles is controlled by manipulating synchronization topology rather than tracking setpoints. We demonstrate three concrete instantiations: tokamak fusion plasma, qubit register coherence, and multi-scale plasma MHD hierarchy, and argue that the approach constitutes a structural advance over classical control methods (PID, MPC, LQR) by operating on the geometric invariant (phase coherence) rather than on domain-specific state variables.

---

1. Core Claim

Synchronization is a universal control variable.

Every physical system with coupled cyclical degrees of freedom — rotating machinery, oscillating plasmas, precessing qubits, rhythmic biological processes, queued computational workloads — admits a description in terms of phase angles θ_i(t) evolving on the N-torus T^N. The Kuramoto order parameter R = |⟨e^{iθ}⟩| then measures the degree of coherence of those phases.

The key insight: instead of designing a bespoke controller for each domain's state variables (temperatures, pressures, currents, gate fidelities), we:

1. Map domain observables → phase angles via principled extraction (Hilbert transform, event frequency, state-ring encoding, or analytic formulas).
2. Couple phases through a Knm matrix whose topology encodes the desired interaction structure.
3. Measure coherence R per layer and partition layers into good (synchrony to promote) and bad (synchrony to suppress).
4. Act on four universal knobs — coupling K, phase lag α, entrainment ζ, reference phase Ψ — which the domain binding translates back into physical actuators.

This four-step pipeline is the SCPN Phase Orchestrator (SPO).

---

2. Mathematical Foundation

2.1 The Unified Phase Dynamics Equation (UPDE)

dθ_i/dt = ω_i + Σ_j K_ij sin(θ_j − θ_i − α_ij) + ζ sin(Ψ − θ_i)

Symbol	Meaning	Controllable?
ω_i	Natural frequency of oscillator i	No (intrinsic)
K_ij	Coupling strength between i and j	Yes (knob K)
α_ij	Phase lag between i and j	Yes (knob α)
ζ	External entrainment drive	Yes (knob ζ)
Ψ	Reference phase	Yes (knob Ψ)

This is the Sakaguchi–Kuramoto equation [sakaguchi1986] with an external drive term. The coupling matrix K_ij, lag matrix α_ij, drive ζ, and reference Ψ are the four degrees of freedom available to the controller.

2.2 Order Parameter as Health Metric

For a group G of oscillators:

R_G e^{iψ_G} = (1/|G|) Σ_{i∈G} e^{iθ_i}

R_G ∈ [0, 1] measures coherence. R = 1 means perfect phase-locking; R = 0 means uniform scatter. This is a topological invariant — it depends only on the relative phase distribution, not on the absolute values or the physical meaning of the oscillators.

2.3 Dual Objective Partition

The binding spec partitions layers into two sets:

• good_layers: R_good should be maximised (healthy coordination).
• bad_layers: R_bad should be suppressed (pathological lock-in).

This captures a pattern universal across domains:

Domain	good (maximise R)	bad (suppress R)
Tokamak plasma	Transport barrier, equilibrium	ELM cascade, tearing modes
Cloud queues	Throughput coordination	Retry storm synchrony
Manufacturing	Line yield coherence	Sensor drift correlation
Neuroscience	Functional connectivity	Seizure-like hypersynchrony
Quantum computing	Logical qubit coherence	Decoherence crosstalk

No classical control formulation natively represents this dual objective without domain-specific cost function engineering.

2.4 Regime State Machine

The supervisor maps (R_good, R_bad, boundary violations) to four regimes:

NOMINAL ←→ DEGRADED ←→ CRITICAL ←→ RECOVERY

• NOMINAL: R_good ≥ 0.6, no hard violations.
• DEGRADED: R_good ∈ [0.3, 0.6), or R_bad rising.
• CRITICAL: R_good < 0.3, or any hard boundary violation.
• RECOVERY: exiting CRITICAL, restoring coupling.

Transitions have hysteresis (0.05) and cooldown (10 steps) to prevent oscillation. This is a finite-state supervisory controller [ramadge1987] realised on top of Kuramoto observables.

---

3. Fusion Plasma Control

3.1 The Control Problem

A tokamak confines plasma (T > 10^8 K) magnetically. The plasma exhibits coupled oscillatory phenomena across timescales spanning six orders of magnitude:

Phenomenon	Timescale	Character
Micro-turbulence (ITG/TEM)	~μs	Broadband fluctuation
Zonal flows	~ms	Shear oscillation
MHD tearing modes	~ms	Rotating magnetic islands
Sawtooth crashes	~10–100 ms	Periodic core relaxation
ELM cycles	~1–10 ms	Edge pressure collapse
Transport barrier	~100 ms	Steady-state H-mode pedestal
Current profile (q)	~1 s	Slow diffusive evolution
Global equilibrium	~1–10 s	Grad-Shafranov force balance

Existing tokamak control uses separate PID loops for each phenomenon: a shape controller for equilibrium, a density controller for Greenwald limit, NTM controllers for tearing modes, ELM pace-making, etc. These loops are designed independently, with ad-hoc priority rules when they conflict.

3.2 SPO Formulation

The plasma_control domainpack maps these phenomena to 8 Kuramoto layers (16 oscillators, 2 per layer). The observable-to-phase mapping:

Observable	Phase formula	Physical rationale
q_profile	2π(q − q_min)/(q_max − q_min)	Safety factor MHD proximity
β_N	2πβ_N/β_limit	Troyon stability margin
τ_E	2πτ_E/τ_ref	Confinement quality fraction
Sawtooth count	count·π mod 2π	π-phase kick per crash event
ELM count	count·π mod 2π	π-phase kick per ELM event
MHD amplitude	2π·amplitude/threshold	Mode activity normalised

Each formula maps the observable's physics-relevant range onto [0, 2π). Event-driven phenomena (sawteeth, ELMs) produce discrete phase kicks rather than continuous evolution — this naturally represents their crash-then-recover dynamics.

Physics boundaries enforce hard limits:

• q_min ≥ 1.0 (Kruskal–Shafranov MHD stability)
• β_N ≤ 2.8 (Troyon no-wall ideal limit)
• Greenwald fraction ≤ 1.2 (density limit)

These map to SPO hard boundaries → immediate CRITICAL regime if violated.

Coupling topology: the Knm matrix uses exponential distance decay (K_ij = K_base · exp(−α|i−j|)) between layers, encoding the physical reality that adjacent-timescale phenomena couple more strongly.

Dual objective: layers 4–6 (transport barrier, current profile, equilibrium) are good — their coherence indicates stable H-mode. Layers 0, 2, 3 (micro-turbulence, tearing, sawtooth/ELM) are bad — their synchronisation indicates confinement-degrading MHD activity.

Policy rules implement two key responses: 1. suppress_elm_storm: when R_bad on the sawtooth/ELM layer exceeds 0.7, inject phase lag α on the turbulence layer to decouple the cascade. 2. restore_transport_barrier: when R_good on the transport barrier drops below 0.3, boost global coupling K.

3.3 What This Replaces

Conventional tokamak control stacks:

Layer	Conventional	SPO equivalent
Shape control	MIMO PID on gaps/currents	Equilibrium layer R + ζ drive
Density	PID on line-integrated density	Greenwald boundary (hard)
NTM suppression	ECCD aimed at island O-point	MHD layer R_bad + α decouple
ELM mitigation	Pellet/RMP pacing	ELM layer R_bad + policy rule
q-profile	Feedforward + slow PID	Current profile layer R_good
Disruption avoidance	Expert-system exception handler	CRITICAL regime + Lyapunov boundary

The SPO formulation replaces ~6 independent SISO/MIMO PID loops and a disruption handler with a single coherence-aware supervisory controller that sees the relationships between phenomena, not just individual setpoints.

3.4 The Kronecker Expansion

The PlasmaControlBridge.import_knm_spec method performs a critical step: it takes an (L×L) layer-level coupling matrix (describing how fast turbulence couples to slow equilibrium) and expands it to an (N×N) oscillator-level matrix via Kronecker product:

K_osc = K_layer ⊗ 1_{n×n}

where n is the number of oscillators per layer. This preserves the inter-layer coupling topology while giving each oscillator within a layer uniform intra-block coupling (with zero self-coupling on the diagonal).

---

4. Quantum Coherence Control

4.1 The Control Problem

A register of N qubits evolving under an XY Hamiltonian:

H = Σ_{i<j} J_ij (X_i X_j + Y_i Y_j) + Σ_i h_i Z_i

can be interpreted as N coupled oscillators in the XY plane. The Bloch-sphere azimuthal angles φ_i play the role of Kuramoto phases, and the coupling constants J_ij play the role of K_ij.

4.2 SPO Formulation

The quantum_simulation domainpack maps this:

• qubit_register layer (4 oscillators): raw XY-plane phases φ_i of individual qubits, extracted from the statevector.
• logical_coherence layer (4 oscillators): logical qubit phases after error correction, treated as informational events (entanglement success/failure timestamps → frequency → phase).

The order parameter R_qubit measures how well the physical qubits maintain phase coherence (≈ entanglement fidelity for the XY Hamiltonian). R_logical measures logical error rate stability.

Boundary: fidelity ≥ 0.5 (hard) — below this, the quantum state is no better than random.

The QuantumControlBridge provides: - import_artifact / export_artifact: pure dict conversion, no quantum library needed. - build_quantum_circuit: constructs a Trotterised circuit from the Knm coupling matrix (requires scpn-quantum-control). - extract_phases_from_statevector: extracts per-qubit Bloch-sphere phases from a Qiskit Statevector.

4.3 What This Adds

Standard quantum optimal control (GRAPE, Krotov, DRAG) optimises pulse waveforms to achieve a target unitary. SPO offers a complementary view: instead of optimising for a specific gate, it monitors ongoing coherence as the system evolves, and adjusts coupling topology to maintain synchrony. This is relevant for:

• Analog quantum simulation (maintaining collective coherence in Hamiltonian evolution)
• Quantum error correction monitoring (tracking logical qubit phase stability across correction cycles)
• Variational quantum circuits (adjusting entangler structure based on coherence feedback)

---

5. The Fusion Core Bridge: Equilibrium-to-Phase Mapping

5.1 The Grad-Shafranov Gap

The Grad-Shafranov equation governs tokamak equilibrium:

Δ*ψ = −μ_0 R² dp/dψ − F dF/dψ

It produces continuous field solutions (ψ, q, pressure, current), not oscillatory dynamics. There is no natural "phase" in an equilibrium solver output.

The FusionCoreBridge resolves this gap by defining six analytic phase-mapping formulas that convert equilibrium observables into [0, 2π) phases (§3.2 table). Each formula maps the observable's physical range onto the circle:

• Linear normalisation for continuous observables (q, β_N, τ_E, MHD amplitude): the ratio of current value to its physical limit sets the phase. A phase near 2π means the system is near its limit.
• Event-driven kicks for discrete phenomena (sawtooth crashes, ELMs): each event advances the phase by π, producing a half-cycle jump. Pairs of rapid events produce near-2π advances, which wrap back near zero — naturally encoding the periodic crash-recovery cycle.

This mapping is invertible (phases_to_feedback recovers the mean phase and coherence), allowing closed-loop operation: equilibrium solver → phases → UPDE evolution → feedback → equilibrium solver.

---

6. Novelty Analysis

6.1 What Exists in the Literature

Approach	Scope	Phase-based?	Multi-scale?	Domain-agnostic?
Kuramoto model [kuramoto1975]	Descriptive	Yes	No	No (physics only)
PID control	Control	No	No	Partially
MPC [rawlings2017]	Control	No	Partially	No
LQR/LQG	Control	No	No	Partially
DIII-D PCS [ferron2006]	Tokamak control	No	Partially	No
GRAPE [khaneja2005]	Quantum control	No	No	No
Synchrophasor (PMU)	Power grid	Yes	No	No
SPO	Control	Yes	Yes	Yes

The Kuramoto model has been used to describe synchronisation in power grids [dorfler2014], neural networks [breakspear2010], and chemical oscillators. But it has never been used as a control abstraction — a framework where:

1. Arbitrary domain observables are systematically mapped into phases.
2. The Kuramoto coupling matrix K_ij becomes a control variable.
3. The order parameter R becomes a health metric.
4. A supervisory policy operates on (R_good, R_bad) to drive the system between regimes.

This four-part combination — observable→phase mapping + coupling as actuator + R as metric + dual-objective supervision — is, to the authors' knowledge, not present in any published control framework.

6.2 Specific Novelties

1. Dual R_good/R_bad objective: no standard cost function formulation natively distinguishes "synchrony to promote" from "synchrony to suppress" as structurally independent objectives. MPC/LQR penalise deviation from a setpoint; they do not partition the state space into "good coherence" and "bad coherence."

2. Observable-to-phase mapping as a formal bridge: the analytic formulas (§3.2, §5.1) constitute a systematic procedure for embedding non-oscillatory observables (equilibrium fields, event counts) into the Kuramoto state space. This has not been formalised as a general control methodology.

3. Kronecker coupling expansion (§3.4): the technique of specifying inter-layer coupling at the hierarchy level and expanding to oscillator level via Kronecker product is a practical contribution that preserves multi-scale topology while keeping the configuration compact.

4. Regime-based supervisory control on Kuramoto observables: integrating a finite-state supervisor (NOMINAL/DEGRADED/CRITICAL/ RECOVERY) with Kuramoto R thresholds and physics-derived hard boundaries (q > 1.0, β_N < 2.8) in a single framework.

5. Domain-agnostic binding spec: the same engine, integrator, supervisor, and policy rules control tokamak plasma, qubit registers, cloud queue systems, manufacturing lines, and biological oscillators — with only a YAML configuration change.

---

7. Why Phase-Synchronization Control Supersedes Classical Methods

7.1 PID: Single-Variable, No Structure

PID tracks a single variable against a setpoint. It has no concept of:

• Multi-scale coupling between phenomena
• Phase relationships (lead/lag) between oscillators
• Coherence as a collective property
• Dual objectives (promote some sync, suppress other sync)

A tokamak running 6 PID loops has 6 independent controllers that can (and do) fight each other. SPO sees the 6 phenomena as one coupled dynamical system and acts on the coupling structure.

7.2 MPC: Domain-Specific, Computationally Expensive

Model Predictive Control solves an optimisation problem at each timestep using a plant model. It has three key limitations:

• Requires a domain-specific model (cannot reuse across domains)
• Scales poorly with state dimension (quadratic or worse in N)
• Solves for actuator trajectories, not for coupling topology
• Cannot express "suppress this coherence" without manual cost function engineering

SPO's per-step cost is O(N²) for the Kuramoto coupling sum — dominated by the sin(θ_j − θ_i) computation — and O(N) for the order parameter. No optimisation problem is solved online.

7.3 LQR: Linear, Assumes Quadratic Cost

Linear-Quadratic Regulator requires linearisation around an operating point. Phase dynamics are inherently nonlinear (sin(·) coupling). Linearising destroys the wrap-around topology of the N-torus and cannot represent phase-locking transitions.

7.4 Expert Systems: Brittle, Non-Transferable

Current tokamak disruption avoidance uses expert-rule exception handlers (e.g., DIII-D's PCS exception handling system). These rules are:

• Written by domain experts for specific machines
• Not transferable between tokamaks (let alone to other domains)
• Triggered by absolute thresholds, not relational coherence
• Unable to detect emergent pathological synchronisation

SPO's policy rules operate on relational observables (R_bad on a specific layer exceeding a threshold), which transfers across machines with the same phenomenological hierarchy.

7.5 The Structural Advantage

The fundamental advantage is dimensional: SPO operates on one scalar per layer (R), plus one regime state, plus four knobs. This is a constant-dimension control space regardless of how many raw state variables the domain has. A tokamak with 10⁴ measurement channels collapses to 8 R-values + 4 boundaries + 4 knobs. The controller complexity does not scale with plant complexity.

---

8. Practical Implications

8.1 For Fusion Energy

• A single SPO instance replaces the 6+ independent PID loops in a tokamak plasma control system.
• Physics boundaries (Kruskal–Shafranov, Troyon, Greenwald) are first-class hard constraints, not afterthought exception handlers.
• The plasma_control policy rules encode ELM suppression and transport barrier recovery as declarative YAML — transferable between machines (ITER, SPARC, DEMO) by adjusting the binding spec.
• The Lyapunov verdict import allows integration with existing disruption prediction systems as a soft boundary signal.

8.2 For Quantum Computing

• Ongoing coherence monitoring during analog quantum simulation, complementing gate-level optimal control.
• Entanglement topology (the Knm matrix) as a control variable, adjusted based on real-time R feedback.
• The fidelity hard boundary (≥ 0.5) triggers regime escalation before the quantum state becomes irrecoverable.

8.3 For General Process Control

• Cloud/queue systems: retry storm detection via R_bad, automatic decoupling via α lag injection (queuewaves domainpack).
• Manufacturing: sensor drift correlation detection, yield coherence monitoring (manufacturing_spc domainpack).
• Biology: multi-scale oscillator monitoring from cellular to systemic (bio_stub domainpack).

All use the same engine with different YAML binding specs.

---

9. Hardware Pipeline: From Sensors to Actuators

This section describes the complete physical data path — how raw sensor signals enter SPO, traverse the phase-synchronization loop, and emerge as actuator commands that drive real hardware.

9.1 General Pipeline Architecture

┌──────────────┐    ┌────────────┐    ┌──────────┐    ┌────────────┐
│ Sensors      │───►│ Phase      │───►│ UPDE     │───►│ Supervisor │
│ (domain HW)  │    │ Extraction │    │ Engine   │    │ + Policy   │
└──────────────┘    └────────────┘    └──────────┘    └─────┬──────┘
                                                            │
              ┌────────────────────────────────────────────┘
              │
              ▼
┌──────────────┐    ┌────────────┐    ┌──────────────────┐
│ Action       │───►│ Actuation  │───►│ Domain Execution │
│ Projector    │    │ Mapper     │    │ (physical HW)    │
└──────────────┘    └────────────┘    └──────────────────┘

Timing: The binding spec declares two periods: - sample_period_s (e.g. 1 ms) — sensor acquisition and phase update rate. - control_period_s (e.g. 10 ms) — supervisor evaluation and actuator command rate. Multiple sample cycles accumulate before one control decision.

Each pipeline stage has a fixed computational cost:

Stage	Operation	Cost
Phase extraction	Per-oscillator mapping	O(N)
UPDE step	Coupling sum	O(N²)
Order parameter	Per-layer mean	O(N)
Supervisor decide	Regime FSM + policy rules	O(R) where R = rule count
Action projector	Rate clamp per action	O(A) where A = action count
Actuation mapper	Scope resolution	O(A·M) where M = actuator count

For the plasma_control domainpack (N=16, R=2, A≤3, M=3), the full pipeline runs in <100 μs on a single core — well within the 1 ms sample budget.

9.2 Tokamak Fusion: Sensor Ingestion

A tokamak produces diagnostic signals across all 8 SPO layers. Each diagnostic maps to a specific oscillator family.

9.2.1 Measurement Systems and Phase Extraction

Diagnostic	Physical quantity	SPO layer	Phase formula
Mirnov coils (poloidal array, ~32 probes)	dB_θ/dt (magnetic fluctuations)	mhd_tearing (2)	Hilbert transform of dominant mode → instantaneous phase
Far-infrared interferometer (multi-chord)	Line-integrated n_e	transport_barrier (4)	2π · n_e / n_Greenwald (density fraction)
Electron Cyclotron Emission (ECE radiometer)	T_e(R) radial profile	micro_turbulence (0)	Hilbert transform of T_e fluctuation envelope
Thomson scattering (multi-pulse laser)	T_e(r), n_e(r) profiles	global_equilibrium (6)	2π · β_N / β_limit from kinetic profiles
Magnetic flux loops (full poloidal set)	ψ(R,Z) equilibrium reconstruction	global_equilibrium (6)	2π · (q − q_min) / (q_max − q_min) from EFIT
Soft X-ray array (diode array)	Core radiation	sawtooth_elm (3)	count · π mod 2π (crash detector)
Divertor Langmuir probes	Ion flux, T_e at target	plasma_wall (7)	2π · heat_flux / heat_flux_limit
Bolometer array	Radiated power P_rad	current_profile (5)	2π · P_rad / P_input (radiation fraction)
Filterscopes / Dα monitors	Dα emission at edge	sawtooth_elm (3)	count · π mod 2π (ELM detector)
Motional Stark Effect (MSE)	Pitch angle → q(r) profile	current_profile (5)	2π · q_axis / q_95 (profile peakedness)

The PlasmaControlBridge receives these as a dict from the plasma control system (scpn-control) or from a direct diagnostic interface:

tick_result = {
    "phases": mirnov_phases + interferometer_phases + ...,  # len=16
    "regime": "NOMINAL",
    "stability": lyapunov_score,
    "layer_sizes": [2, 2, 2, 2, 2, 2, 2, 2],
}
state = bridge.import_snapshot(tick_result)

9.2.2 Phase Extraction Methods

Three extraction methods apply depending on signal character:

Hilbert transform (continuous oscillatory signals): For Mirnov coils and ECE fluctuations, the analytic signal z(t) = x(t) + iH[x(t)] yields instantaneous phase φ(t) = arg(z(t)). Applied to the dominant MHD mode (identified by toroidal mode number analysis), this extracts the rotating island phase.

Normalised ratio (equilibrium quantities): For β_N, q-profile, τ_E, density, and radiation — quantities that do not oscillate but vary slowly — the phase is a linear map of the observable's position within its physical range: θ = 2π · x/x_limit. As x approaches its stability limit, θ → 2π signals proximity to the boundary.

Event counter (discrete crash events): For sawteeth and ELMs, each detected event (identified by a threshold crossing in SXR or Dα) increments a counter. Phase = count · π mod 2π. Two rapid events produce a near-2π advance that wraps to ~0, naturally encoding the crash-recovery cycle.

9.2.3 Diagnostic Timing Budget

Diagnostic	Acquisition rate	Latency to SPO
Mirnov coils	100 kHz–1 MHz	<10 μs (analog + ADC)
Interferometer	10–100 kHz	<100 μs
ECE radiometer	1–10 kHz	<1 ms
Thomson scattering	10–100 Hz	10–100 ms
Magnetic flux loops	10 kHz	<100 μs
SXR array	100 kHz	<10 μs
Langmuir probes	10–100 kHz	<100 μs

Fast diagnostics (Mirnov, SXR) update every sample cycle (1 ms). Slow diagnostics (Thomson) hold their last value between updates. The UPDE engine uses whatever phases are current, treating slow channels as piecewise-constant — consistent with the real physics where equilibrium quantities evolve on ~1 s timescales.

9.3 Tokamak Fusion: Control Pipeline

9.3.1 Supervisor Decision

Every control_period_s (10 ms), the supervisor evaluates:

1. R_good for layers [4, 5, 6] (transport_barrier, current_profile, equilibrium). High R_good → stable H-mode.
2. R_bad for layers [0, 2, 3] (micro_turbulence, mhd_tearing, sawtooth_elm). High R_bad → pathological MHD activity.
3. Boundary violations: q_min < 1.0 or β_N > 2.8 or greenwald > 1.2 → immediate CRITICAL transition.

The RegimeManager updates the state machine:

NOMINAL (R_good ≥ 0.6, no hard violations)
   ↓ R_good drops below 0.55
DEGRADED (R_good ∈ [0.3, 0.6), or R_bad rising)
   ↓ R_good < 0.25 or hard violation
CRITICAL (emergency response)
   ↓ recovery detected
RECOVERY (controlled ramp-back)
   ↓ R_good ≥ 0.65
NOMINAL

Hysteresis (0.05 offset between down/up thresholds) prevents chattering at transitions. Cooldown (10 steps) prevents rapid regime oscillation.

9.3.2 Policy Rule Firing

Two policy rules in plasma_control/policy.yaml:

suppress_elm_storm: fires in NOMINAL or DEGRADED when R_bad on sawtooth_elm layer exceeds 0.7. Emits:

ControlAction(knob="alpha", scope="layer_0", value=0.4, ttl_s=5.0)

This injects phase lag on the micro_turbulence layer, decoupling the turbulence↔ELM cascade. The physical effect: turbulence-driven edge pressure build-up is disrupted, preventing the next ELM trigger.

restore_transport_barrier: fires in DEGRADED or RECOVERY when R_good on transport_barrier layer drops below 0.3. Emits:

ControlAction(knob="K", scope="global", value=0.2, ttl_s=10.0)

This increases global coupling, strengthening the inter-layer synchronization. The physical effect: tighter coordination between equilibrium, transport, and boundary layers, restoring the H-mode pedestal.

9.3.3 Action Projection and Rate Limiting

Each ControlAction passes through the ActionProjector before reaching hardware. For the plasma_control binding:

Knob	Value bounds	Rate limit per step
K	[0.0, 5.0]	0.1
alpha	[0.0, π/2]	0.05
zeta	[0.0, 2.0]	0.2

Rate limiting prevents actuator damage. If the policy demands K=0.2 but the previous value was K=0.05, the projector caps the step at +0.1, yielding K=0.15. The next cycle brings K=0.2. This produces smooth actuator ramps, not discontinuous jumps.

9.3.4 Knob-to-Actuator Mapping

The ActuationMapper resolves abstract knob commands into domain-specific actuator commands. The binding spec defines three actuators:

actuators:
  - name: coupling_global   # knob: K, scope: global
  - name: lag_turbulence     # knob: alpha, scope: layer_0
  - name: damping_global     # knob: zeta, scope: global

The output is a command dict:

{"actuator": "coupling_global", "knob": "K", "scope": "global",
 "value": 0.15, "ttl_s": 10.0}

9.3.5 Actuator-to-Hardware Translation

The final translation from SPO actuator commands to physical tokamak hardware is performed by the domain execution layer — the PlasmaControlBridge.export_control_actions() method or the downstream plasma control system.

SPO actuator	Physical hardware	Translation
coupling_global (K)	Poloidal field (PF) coil currents	K ∝ I_PF shaping current. Higher K → tighter magnetic configuration → stronger inter-layer coupling. The PF coil power supplies (thyristor or IGBT converters, ~10 kA, ~1 kV) receive current setpoints from the shape controller.
coupling_global (K)	Neutral Beam Injection (NBI) power	K ∝ P_NBI. Higher coupling → higher beam power → more torque → stronger rotation → better MHD coupling across flux surfaces. NBI sources (~1 MW per beamline, deuterium/hydrogen, 40–100 keV) respond in ~50 ms.
lag_turbulence (α)	Resonant Magnetic Perturbation (RMP) coil phasing	α maps to the toroidal phase of the n=3 RMP field. The RMP coils (typically 3×6 in-vessel saddle coils) receive phase-shifted sinusoidal currents. Adjusting the phase angle between upper and lower coil rows controls the edge magnetic perturbation spectrum, directly modulating ELM stability.
lag_turbulence (α)	ECRH/ECCD deposition angle	α maps to the poloidal steering angle of the EC launcher mirror. Electron Cyclotron Resonance Heating (ECRH, 170 GHz gyrotrons, ~1 MW per launcher) deposits power at a specific flux surface. Shifting the mirror angle changes the current drive location, modulating the local q-profile shear that controls tearing mode coupling.
damping_global (ζ)	Pellet injection rate	ζ ∝ pellet frequency. Higher entrainment drive → more frequent pellet injection → stronger density pacing. Pellet injectors (frozen D₂ pellets, 1–4 mm diameter, 200–1000 m/s) fire at programmable rates up to ~100 Hz. ELM pacing via pellets is established practice (JET, AUG).
damping_global (ζ)	Gas puff valves	ζ ∝ gas flow rate. Piezoelectric gas valves (D₂, N₂, Ne, Ar) with ~1 ms response time control edge density and impurity seeding. Higher ζ drives the edge density toward a reference state.

9.3.6 Closed-Loop Timing

The complete loop for one control cycle:

t=0.000 ms  Mirnov/SXR/interferometer acquire raw signals
t=0.010 ms  ADC digitisation complete
t=0.100 ms  Phase extraction (Hilbert / ratio / counter)
t=0.200 ms  PlasmaControlBridge.import_snapshot() → UPDEState
t=0.300 ms  UPDE RK4 step (16 oscillators, O(256) operations)
t=0.400 ms  Order parameter R_good, R_bad computed
t=0.500 ms  RegimeManager.evaluate() → regime transition check
t=0.600 ms  SupervisorPolicy.decide() → list[ControlAction]
t=0.700 ms  ActionProjector.project() → rate-limited actions
t=0.800 ms  ActuationMapper.map_actions() → actuator commands
t=0.900 ms  PlasmaControlBridge.export_control_actions() → HW dict
t=1.000 ms  Commands dispatched to PF/NBI/RMP/ECRH/pellet systems

Total pipeline latency: ~1 ms. This is within the 10 ms control period of the plasma_control binding spec (10× margin), and well within the characteristic MHD timescale (~1–10 ms).

For comparison, existing tokamak control systems (DIII-D PCS, KSTAR EPICS) operate at 1–10 ms cycle times. SPO matches this performance while providing multi-scale coherence awareness that single-loop controllers lack.

9.4 Tokamak Fusion: Full Data Flow Diagram

 TOKAMAK VESSEL
 ┌─────────────────────────────────────────────────────────────┐
 │  Diagnostics (sensors)              Actuators (effectors)   │
 │  ┌──────────────────┐               ┌────────────────────┐ │
 │  │ Mirnov coils     │               │ PF coils (6 pairs) │ │
 │  │ ECE radiometer   │               │ CS solenoid        │ │
 │  │ Interferometer   │               │ NBI (~8 sources)   │ │
 │  │ Thomson scatter  │               │ ECRH (~4 gyrotrons)│ │
 │  │ SXR array        │               │ RMP coils (3×6)    │ │
 │  │ Bolometers       │               │ Pellet injector    │ │
 │  │ Langmuir probes  │               │ Gas puff valves    │ │
 │  │ MSE diagnostic   │               │ Feedback coils     │ │
 │  │ Dα monitors      │               │                    │ │
 │  └───────┬──────────┘               └────────▲───────────┘ │
 └──────────┼───────────────────────────────────┼─────────────┘
            │ raw signals                       │ I, P, angle, rate
            ▼                                   │
 ┌──────────────────────┐            ┌──────────┴──────────────┐
 │ Phase Extraction     │            │ Domain Execution         │
 │                      │            │                          │
 │ Hilbert(Mirnov) →θ₂  │            │ K→I_PF, P_NBI           │
 │ ratio(n_e/n_GW) →θ₄  │            │ α→RMP_phase, ECRH_angle │
 │ ratio(β_N/β_lim)→θ₁  │            │ ζ→pellet_rate, gas_flow │
 │ count(SXR crash)→θ₃  │            │ Ψ→reference setpoint    │
 │ ratio(q/q_95)   →θ₅  │            │                          │
 │ ratio(P_rad)    →θ₅  │            └──────────▲──────────────┘
 │ ratio(T_e/T_sep)→θ₇  │                       │
 │ Hilbert(ECE)    →θ₀  │                       │ actuator commands
 └───────┬──────────────┘            ┌───────────┴──────────────┐
         │ θ₀..θ₁₅ (16 phases)      │ ActuationMapper          │
         ▼                           │ scope resolution         │
 ┌──────────────────────┐            │ value clamping           │
 │ UPDE Engine (RK4)    │            └──────────▲──────────────┘
 │                      │                       │
 │ dθᵢ/dt = ωᵢ         │            ┌───────────┴──────────────┐
 │   + ΣⱼKᵢⱼsin(θⱼ−θᵢ) │            │ ActionProjector          │
 │   + ζ sin(Ψ−θᵢ)     │            │ rate limits              │
 │                      │            │ value bounds             │
 │ → θ₀..θ₁₅ (updated) │            └──────────▲──────────────┘
 └───────┬──────────────┘                       │
         │                           ┌───────────┴──────────────┐
         ▼                           │ SupervisorPolicy.decide() │
 ┌──────────────────────┐            │                          │
 │ Order Parameters     │───────────►│ R_good[4,5,6] ≥ 0.6?    │
 │                      │            │ R_bad[0,2,3] < 0.7?      │
 │ R_good (layers 4-6)  │            │ q_min ≥ 1.0?             │
 │ R_bad  (layers 0,2,3)│            │ β_N ≤ 2.8?               │
 │ per-layer Rᵢ, ψᵢ    │            │ greenwald ≤ 1.2?         │
 └──────────────────────┘            └──────────────────────────┘

9.5 Quantum Hardware Pipeline

Superconducting qubit systems (IBM Heron, Google Sycamore) use a parallel pipeline with different physical signals.

9.5.1 Sensor Ingestion

Diagnostic	Physical quantity	SPO layer
Readout resonator (dispersive, ~7 GHz)	Qubit state (IQ plane)	qubit_register (0)
Randomized benchmarking (gate sequences)	Process fidelity	logical_coherence (1)
T₁/T₂ characterization (decay sequences)	Coherence times	logical_coherence (1)

Phase extraction: the readout resonator returns an IQ-plane point for each qubit. The azimuthal angle in the IQ plane is the Bloch-sphere phase φᵢ — no Hilbert transform needed. This is a direct measurement of the oscillator phase.

9.5.2 Actuator Commands

SPO actuator	Physical hardware	Translation
K (coupling)	Tunable coupler flux bias	K ∝ coupler flux Φ_c. Flux-tunable couplers (DC SQUID loop between qubits) set the effective J_ij coupling. A DAC channel (~16 bit, ~1 GHz update rate) drives the coupler's flux line.
α (phase lag)	Microwave drive phase	α maps to the phase offset of the qubit drive pulse. The arbitrary waveform generator (AWG, ~5 GS/s) shifts the IQ modulation phase of the 4–6 GHz drive tone.
ζ (entrainment)	Microwave drive amplitude	ζ ∝ Rabi frequency Ω_R. Higher drive amplitude → stronger pull toward the reference phase Ψ. The AWG scales the drive envelope.
Ψ (reference)	Drive pulse rotation axis	Ψ sets the rotation axis angle on the Bloch sphere equator. Standard single-qubit gates (X, Y, Z, arbitrary) are rotations at programmable phase angles — Ψ selects which axis to drive toward.

9.5.3 Timing

Superconducting qubit operations run at ~GHz clock rates with ~100 ns gate times. The SPO control cycle must fit within the coherence time T₂ (~100 μs for current hardware):

t=0      Readout pulse → IQ data (500 ns)
t=1 μs   Phase extraction (direct IQ angle)
t=2 μs   UPDE step (8 oscillators, O(64) operations)
t=3 μs   Supervisor evaluation
t=4 μs   Coupler/drive command dispatch
t=5 μs   Actuator settling (~100 ns for flux couplers)

Total latency: ~5 μs. This fits within a single T₂ window, allowing real-time coherence-based feedback during quantum computation — not just post-hoc analysis.

9.6 General Process Control Pipeline

For non-physics domains (manufacturing, cloud queues, biological systems), the pipeline follows the same architecture with different sensor and actuator vocabularies.

9.6.1 Manufacturing (SPC)

SPO component	Physical realization
Sensors	Inline metrology (coordinate measuring machines, optical gauges, force sensors)
Phase extraction	Normalised ratio: θ = 2π · (x − LSL) / (USL − LSL) where LSL/USL are spec limits
Actuators (K)	PLC setpoint coupling between stations (e.g., upstream tool compensation based on downstream SPC)
Actuators (α)	Process delay compensation (conveyor speed, batch hold time)
Actuators (ζ)	Reference part injection rate (golden sample pacing)
Cycle time	Seconds to minutes (matches process tempo)

9.6.2 Cloud / Queue Systems

SPO component	Physical realization
Sensors	Queue depth metrics, request latency percentiles, error rate counters
Phase extraction	Normalised ratio (queue_depth / queue_capacity) or event frequency (requests/sec → rad/s)
Actuators (K)	Service mesh routing weights (Istio, Envoy)
Actuators (α)	Retry backoff configuration (exponential base, jitter)
Actuators (ζ)	Rate limiter setpoint (requests/sec ceiling)
Cycle time	Milliseconds to seconds

9.6.3 Biological Oscillators

SPO component	Physical realization
Sensors	EEG electrodes, ECG leads, respiratory belt, skin conductance
Phase extraction	Hilbert transform of band-passed signal (e.g., alpha 8–13 Hz, theta 4–8 Hz)
Actuators (K)	Binaural beat frequency (entrainment coupling strength)
Actuators (α)	Phase offset between left/right audio channels
Actuators (ζ)	Audio amplitude (drive strength)
Actuators (Ψ)	Target brainwave phase (circadian reference, meditation target)
Cycle time	~100 ms (matching EEG epoch length)

9.7 Safety and Fail-Safe Design

The pipeline incorporates multiple safety layers:

1. Boundary violations (hard) trigger immediate CRITICAL regime, which commands conservative actuator values (reduce K, increase ζ toward a known stable state).

2. Rate limiting in the ActionProjector prevents actuator damage. A sudden coupling increase request is smoothed over multiple control cycles.

3. TTL (time-to-live) on every ControlAction ensures commands expire. If the supervisor stops issuing commands (e.g., software crash), actuators revert to their default state within the TTL window.

4. Physics invariant checks in the bridge adapters (PlasmaControlBridge.check_physics_invariants, FusionCoreBridge.check_stability) run independently of SPO and can veto commands that would violate absolute safety limits.

5. Watchdog: the sample/control period mismatch (10:1 in plasma_control) means the UPDE runs 10 phase updates between each control decision. If the UPDE detects divergence (NaN, Inf, or R < 0.01), it halts before the next control cycle.

For nuclear-grade tokamak deployment, the SPO control layer would sit above the existing machine-protection interlock system (which operates on hard-wired analog signals and cannot be overridden by software). SPO optimises within the interlock-safe envelope; it does not replace the interlock.

---

10. Scope of Competence: What SPO Is and Is Not

This section draws explicit boundaries around the framework's claims. Overstating scope would be scientifically dishonest; understating it would obscure the genuine contribution.

10.1 What SPO Is

SPO is a supervisory control layer that monitors multi-scale coherence and issues corrective commands to domain actuators.

It is a control framework, not a physics solver. It does not compute equilibria, propagate wavefunctions, solve transport equations, or simulate fluid dynamics. It consumes the outputs of domain-specific solvers and acts on them through the phase-synchronization abstraction.

The analogy: SPO is to a tokamak what a conductor is to an orchestra. The conductor does not play any instrument (does not solve any physics equation). The conductor listens to the ensemble (monitors phase coherence across layers), detects when sections drift out of sync (R_bad rising), and signals corrections (adjusts coupling K, phase lag α, drive ζ). The musicians (domain solvers + physical actuators) produce the actual sound (maintain the actual equilibrium).

10.2 What SPO Is Not

SPO is NOT	Why this matters
An equilibrium solver	It cannot compute ψ(R,Z). It depends on EFIT, FreeGS, or scpn-fusion-core for force balance.
A turbulence simulator	It cannot predict ITG/TEM growth rates. It monitors turbulence signatures via phase extraction.
A disruption predictor	It detects coherence degradation, which may correlate with disruption precursors. It does not model disruption physics (thermal quench, current quench, halo currents).
A quantum error correction code	It monitors coherence of logical qubits. It does not implement stabiliser measurements or syndrome decoding.
A replacement for machine-protection interlocks	Hard-wired safety systems (vessel protection, magnet quench detection) operate on analog signals below SPO's software layer. SPO cannot override them and must not attempt to.

10.3 What SPO Expects from the Domain

For SPO to function, the domain must provide:

1. Oscillatory or quasi-oscillatory observables. The phase extraction step requires signals that admit meaningful phase angles. Continuous oscillations (MHD modes, qubit precession) map cleanly. Slowly varying quantities (q-profile, β_N) map via normalised ratios — a weaker but still informative encoding. Signals with no temporal structure (static geometry, material properties) cannot be phase-encoded and are outside SPO's scope.

2. A known coupling hierarchy. The Knm matrix encodes which phenomena couple to which. This must be specified by domain knowledge (e.g., "micro-turbulence drives zonal flows which modulate transport"). SPO does not discover coupling topology from data — it exploits a topology provided in the binding spec.

3. Controllable actuators that influence coupling. The four knobs (K, α, ζ, Ψ) must map to physical actuators that can modulate the coupling, phase lag, or drive of the oscillatory phenomena. If the domain has no actuators (pure observation), SPO reduces to a monitoring-only tool (still useful for coherence detection, but not for control).

4. Meaningful good/bad partition. The dual objective requires domain expertise to identify which synchrony is healthy and which is pathological. This is not always obvious — in some systems, all synchrony might be desirable (or undesirable). If no dual partition exists, SPO degenerates to a standard single-objective coherence maximiser.

10.4 Concrete Expectations per Domain

10.4.1 Tokamak Fusion Plasma

SPO's expected contribution: improved disruption avoidance through earlier detection of multi-scale coherence degradation.

SPO does not solve the Grad-Shafranov equation. It does not compute MHD stability limits. It monitors whether the plasma's oscillatory phenomena (turbulence, MHD modes, sawteeth, ELMs, transport barriers) are maintaining a healthy coherence pattern — and detects correlated degradation across timescales before individual diagnostic thresholds are crossed.

The specific value proposition:

• Cascade detection: when R_bad rises simultaneously on layers [0, 2, 3] (turbulence, tearing, sawtooth/ELM), it signals a coupled MHD cascade forming. Individual diagnostics may still be within limits, but the correlation is the precursor. Existing PID loops cannot see this because they monitor channels independently.

• Coordinated actuation: instead of 6 independent PID loops potentially fighting each other (shape controller demanding more current while density controller demands less gas), SPO sees the coupled system and issues coordinated commands through a single coherence-aware policy.

• Transferability: the same binding spec structure (8 layers, 16 oscillators, good/bad partition, physics boundaries) applies to any tokamak. Porting from DIII-D to ITER requires changing the Knm calibration and boundary thresholds, not rewriting the controller.

What SPO does NOT expect to do for fusion: - Replace the real-time equilibrium reconstruction (EFIT runs anyway) - Replace machine-protection interlocks (hard-wired, below SPO) - Guarantee disruption avoidance (it improves detection, not elimination) - Work without calibration (the Knm matrix and phase extraction require tokamak-specific tuning against experimental data)

Confidence level: MEDIUM. The theoretical basis (Kuramoto coherence as a multi-scale health metric) is sound. The practical value depends on: (a) phase extraction fidelity under real diagnostic noise, (b) Knm calibration against real inter-phenomenon coupling, (c) actuator response linearity. None of these have been validated experimentally.

10.4.2 Quantum Coherence

SPO's expected contribution: real-time coherence monitoring and coupling topology adjustment during analog quantum simulation.

SPO does not implement quantum gates, error correction, or compilation. It monitors the Bloch-sphere phases of physical qubits (directly measurable from readout resonators) and the logical coherence of error-corrected qubits (derived from syndrome statistics).

The specific value proposition:

• Continuous fidelity tracking: R_qubit dropping below threshold triggers regime escalation during the computation, not after post-hoc analysis.

• Coupling topology feedback: the Knm matrix corresponds to flux-tunable coupler settings. SPO adjusts couplings to maintain coherence — complementing gate-level optimal control (GRAPE/Krotov) which optimises pulse shapes for specific unitaries.

• Cross-layer monitoring: tracking both physical qubit phases and logical qubit stability in a single framework, with the fidelity boundary triggering corrective action when the physical layer degrades.

Confidence level: MODERATE-HIGH for monitoring, LOWER for active control. Phase extraction from IQ readout is direct and high-fidelity. But the Kuramoto sinusoidal coupling is an approximation of the actual XY Hamiltonian dynamics — it captures the first harmonic but misses higher-order terms. Active coupling adjustment via flux couplers is realistic but untested in a coherence-feedback loop.

10.4.3 General Process Control

SPO's expected contribution: detection of pathological synchronisation (retry storms, sensor drift correlation, cascade failures) in systems with coupled cyclical processes.

• Cloud queues: retry-storm monitoring (R_bad on the retry layer) is a candidate fit because retries can synchronise request timing. SPO can measure that hypothesis and propose an α/backoff adjustment; lead time, false-alarm rate, and intervention benefit remain unvalidated.

• Manufacturing SPC: phase-based sensor-drift correlation is a candidate complement to single-channel SPC. Its incremental detection value and phase extraction from slowly drifting metrology signals require comparative validation.

• Biological oscillators: EEG band coherence monitoring is well-established (functional connectivity analysis uses similar metrics). SPO adds the dual good/bad partition and entrainment control via audio. Confidence: MODERATE-HIGH for monitoring, MODERATE for entrainment effectiveness.

10.5 What Would Constitute Validation

The framework remains at TRL 3–4 (analytical/experimental proof of concept in laboratory environment). Advancing claims requires:

Validation target	Required evidence	Current status
Phase extraction fidelity	Apply Hilbert/ratio/counter extraction to real diagnostic data from a tokamak shot database (DIII-D, EAST, JET). Measure phase quality (SNR, continuity, bandwidth).	Not done. Synthetic-only.
Cascade detection lead time	Compare R_bad onset against disruption timestamps in a shot database. Quantify how many ms earlier the coherence signal appears vs. single-channel thresholds.	Not done.
Knm calibration	Fit Knm matrix to cross-correlation structure of real multi-diagnostic data. Compare exponential-decay default against data-driven matrix.	Not done.
Actuator coordination benefit	Run SPO in advisory mode alongside existing PCS on a real tokamak. Compare shot-to-shot performance (confinement time, ELM frequency, disruption rate).	Not done. Requires facility access.
Quantum coherence feedback	Run SPO feedback loop on IBM Heron or Google Sycamore. Compare fidelity trajectory with and without SPO-driven coupler adjustments.	Not done. scpn-quantum-control has hardware results for open-loop characterisation only.
Cross-domain transfer	Apply the same SPO engine to two unrelated domains (e.g., tokamak + cloud queue) with only binding spec changes. Demonstrate that both benefit.	Demonstrated in simulation (9 domainpacks). Not validated on real systems.

10.6 The Honest Claim

SPO offers a structural advance in control architecture — the ability to monitor multi-scale coherence and act on coupling topology rather than individual setpoints. This is genuinely new (§6).

SPO does not offer a guarantee that this structural advance translates to practical performance gains in any specific domain. That translation depends on calibration, phase extraction fidelity, and actuator response — all domain-specific, all requiring experimental validation.

The framework is most likely to succeed in domains where: - Pathological synchronisation is a known failure mode (ELMs, retry storms, seizure-like cascades) - Multi-scale coupling is strong and poorly handled by independent single-loop controllers - The good/bad coherence partition is physically clear

The framework is least likely to succeed in domains where: - Observables lack oscillatory or quasi-periodic character - Coupling topology is unknown or highly nonlinear beyond first-harmonic - Actuators cannot modulate coupling (observation-only systems) - The required control bandwidth exceeds the UPDE computation time

This section exists to prevent the framework from being oversold. A control abstraction with no experimental validation is a hypothesis, not a solution. The hypothesis is well-motivated (§2–§7), the implementation is complete (§9), and the validation agenda is defined (§10.5). What remains is the experiment.

---

11. Limitations and Open Questions

1. Phase extraction fidelity: the Hilbert transform, event frequency, and ring-encoding methods have finite bandwidth and quality limits. Not all domain signals have a clean oscillatory component. The quality gating system (§ oscillators_PIS.md) mitigates this but does not eliminate it.

2. Coupling matrix design: the exponential-decay Knm is a reasonable default but may not match the true coupling topology of a specific system. Domain calibration is needed.

3. Observable-to-phase mapping invertibility: the analytic formulas (§3.2) are forward-only surjections, not bijections. The phases_to_feedback method recovers aggregate statistics (R, mean phase) but cannot reconstruct the original observables.

4. Computational limits: dense coupling has O(N²) compute and storage costs. Feasible oscillator count and update rate depend on the selected monitors, backend, host, and required latency envelope; no portable real-time limit is established here.

5. Validation: the framework is demonstrated on synthetic simulations. Validation against real tokamak data (DIII-D, EAST, JET) and real quantum hardware (IBM Eagle/Heron) remains future work.

---

12. References

• [kuramoto1975] Y. Kuramoto (1975). Self-entrainment of a population of coupled non-linear oscillators. Lecture Notes in Physics 39, 420–422.
• [sakaguchi1986] H. Sakaguchi & Y. Kuramoto (1986). A soluble active rotater model showing phase transitions via mutual entertainment. Prog. Theor. Phys. 76, 576–581.
• [acebron2005] J. A. Acebrón et al. (2005). The Kuramoto model: a simple paradigm for synchronization phenomena. Rev. Mod. Phys. 77, 137–185.
• [dorfler2014] F. Dörfler & F. Bullo (2014). Synchronization in complex networks of phase oscillators: a survey. Automatica 50, 1539–1564.
• [breakspear2010] M. Breakspear, S. Heitmann & A. Daffertshofer (2010). Generative models of cortical oscillations. Front. Hum. Neurosci. 4, 190.
• [rawlings2017] J. B. Rawlings, D. Q. Mayne & M. M. Diehl (2017). Model Predictive Control: Theory, Computation, and Design. 2nd ed., Nob Hill Publishing.
• [ferron2006] J. R. Ferron et al. (2006). Real time equilibrium reconstruction for tokamak discharge control. Nucl. Fusion 38, 1055.
• [khaneja2005] N. Khaneja et al. (2005). Optimal control of coupled spin dynamics: design of NMR pulse sequences by gradient ascent. J. Magn. Reson. 172, 296–305.
• [ramadge1987] P. J. Ramadge & W. M. Wonham (1987). Supervisory control of a class of discrete event processes. SIAM J. Control Optim. 25(1), 206–230.
• [dormand1980] J. R. Dormand & P. J. Prince (1980). A family of embedded Runge-Kutta formulae. J. Comput. Appl. Math. 6(1), 19–26.
• [troyon1984] F. Troyon et al. (1984). MHD-limits to plasma confinement. Plasma Phys. Control. Fusion 26, 209.
• [kruskal1958] M. D. Kruskal & M. Schwarzschild (1954). Some instabilities of a completely ionized plasma. Proc. R. Soc. Lond. A 223, 348–360.
• [greenwald2002] M. Greenwald (2002). Density limits in toroidal plasmas. Plasma Phys. Control. Fusion 44, R27.
• [snipes2010] J. A. Snipes et al. (2010). Actuator and diagnostic requirements of the ITER plasma control system. Fusion Eng. Des. 85, 461–465.
• [humphreys2015] D. A. Humphreys et al. (2015). Novel aspects of plasma control in ITER. Phys. Plasmas 22, 021806.
• [evans2006] T. E. Evans et al. (2006). Edge stability and transport control with resonant magnetic perturbations in collisionless tokamak plasmas. Nature Phys. 2, 419–423.
• [lang2004] P. T. Lang et al. (2004). ELM pace making and mitigation by pellet injection in ASDEX Upgrade. Nucl. Fusion 44, 665.
• [krantz2019] P. Krantz et al. (2019). A quantum engineer's guide to superconducting qubits. Appl. Phys. Rev. 6, 021318.