convergio 0.2.1

Automatic convergence detection for iterative numerical methods. Stop wasting compute.

Convergio

Automatic convergence detection for iterative numerical methods.

Stop wasting compute. One function call tells you if your solver has converged, is stalling, is oscillating, or is diverging.

"Your solver finished 3000 iterations ago."

Install

pip install convergio

Only dependency: NumPy.

Quick Start

Analyze a completed run

from convergio import detect

result = detect(my_residual_history)

print(result.state)        # "converged" | "stalling" | "oscillating" | "diverging" | ...
print(result.quality)      # 0.0 — 1.0
print(result.converged_at) # step where convergence was detected

Detect stalling (new in v0.1.3)

A solver can look "converged" (low variance) but actually be stuck above the target residual. Convergio distinguishes between true convergence and stalling:

result = detect(residuals, residual_target=1e-8)

if result.state == "stalling":
    print("Solver plateaued — try different parameters")
elif result.state == "converged":
    print(f"Solved at step {result.converged_at}")

Monitor a running simulation

from convergio import watch

mon = watch(var_threshold=1e-6, residual_target=1e-8)

for step in range(10000):
    residual = solver.step()
    stop, info = mon.step(residual)
    if stop:
        print(f"Done at step {step} — saved {10000 - step} iterations")
        break

What It Detects

State	Meaning	Recommendation
converged	Signal has stabilized below target	Stop
stalling	Signal is flat but above target	Adjust parameters
oscillating	Signal oscillates periodically without settling	Continue
noisy	High zero-crossing rate but no periodicity (looks like noise)	Continue
bistable	Signal jumps between 2+ distinct states	Investigate
diverging	Variance is genuinely growing	Restart
numerical_failure	Signal contains NaN/Inf — run collapsed	Restart
warming_up	Not enough data yet	Continue

Benchmark Results

Tested on synthetic signals across all convergence states (v0.1.4):

Detection Accuracy

Signal Type	N=200	N=500	N=1000
Converging (exp. decay)	converged	converged	converged
Stalling (flat, above target)	stalling	stalling	stalling
Oscillating (sine + noise)	—	oscillating	oscillating
Diverging (random walk)	—	diverging	diverging

Overall: 87% correct across all signal types and lengths. Hardest case: bistable signals (multi-modal jumping), which can resemble oscillation.

Compute Savings with watch()

Scenario	Budget	Stopped at	Saved
Fast convergence	2,000	1,100	45%
Medium convergence	5,000	4,200	16%
Diverging (early abort)	5,000	350	93%

watch() detects convergence and divergence. Diverging runs are killed early instead of burning through the full iteration budget.

Stalling signals are flagged but not auto-stopped — you decide what to do.

Speed

Signal Length	Time	Throughput
1,000	5 ms	200K pts/s
10,000	52 ms	190K pts/s
100,000	517 ms	193K pts/s
1,000,000	5.1 s	197K pts/s

detect() runs at ~190,000 points/second. For a typical FEM solver with 5,000 iterations, analysis takes ~25 ms — negligible compared to solve time.

API Reference

detect(signal, window=0, var_threshold=1e-6, osc_threshold=0.005, residual_target=0.0)

Analyze a complete time series.

Parameter	Type	Default	Description
signal	array-like	—	1D array of scalar values
window	int	0 (auto)	Analysis window size
var_threshold	float	1e-6	Variance below this = stable
osc_threshold	float	0.005	Min oscillation frequency (zero-crossings / 2N) to flag oscillation
residual_target	float	0.0	If > 0: stable signal above this = stalling

Returns ConvergenceResult.

watch(var_threshold=1e-6, check_every=50, min_steps=100, window=0, residual_target=0.0, max_history=10000)

Create a live monitor. Returns Watch object. Call .step(value) each iteration. Memory is bounded by max_history.

Parameter	Type	Default	Description
var_threshold	float	1e-6	Variance below this = stable
check_every	int	50	Check convergence every N steps
min_steps	int	100	Minimum steps before first check
window	int	0 (auto)	Analysis window size
residual_target	float	0.0	If > 0: stable signal above this = stalling
max_history	int	10000	Keep at most this many recent values (bounds memory)

ConvergenceResult

Field	Type	Description
state	str	converged, stalling, oscillating, noisy, bistable, diverging, numerical_failure, warming_up
quality	float	0.0 — 1.0 convergence quality
converged_at	int or None	Step where convergence detected
final_variance	float	Variance of last window
oscillation_freq	float	Detected oscillation frequency
n_modes	int	Number of distinct stable states
recommendation	str	stop, continue, restart, adjust_params
saved_steps	int	Steps saved by early stopping
total_steps	int	Total steps in analyzed signal
note	str	Optional hint (e.g. missing residual_target, numerical failure); empty if none

What's New

v0.2.1

• Documentation: README now reflects the v0.2 states, parameters and note field.

v0.2.0 — audit-driven robustness pass

• NaN/Inf no longer crash — new state numerical_failure instead of a ValueError.
• Stationary noise is separated from periodic oscillation — new state noisy (autocorrelation check); noise is no longer mislabelled diverging.
• osc_threshold is now actually used (it was inert); default rescaled to 0.005 (the documented zero-crossing frequency), behaviour-preserving.
• count_modes over-counting fixed — a finite unimodal sample no longer produces spurious extra modes (valley-prominence merge).
• rolling_variance now really uses cumulative sums (O(n)); previously a per-window Python loop despite the comment.
• watch() bounds memory via max_history instead of re-analysing an unbounded history every check.
• New note field — a stabilized signal with no residual_target that plateaus without dropping from its start now flags the missing target (state stays converged for backward compatibility).

v0.1.5

• All 6 convergence states: converged, stalling, oscillating, diverging, bistable, warming_up
• Improved trend-check robustness for near-convergent signals
• Battle-tested in CAP Stability Engine v0.8 (15/16 SuiteSparse matrices solved)

v0.1.4

• Trend-check: smooth decline is no longer misclassified as stalling

v0.1.3

• residual_target parameter: distinguishes true convergence from stalling
• stalling state: flat signal above target is now correctly identified
• watch() reports stalling without auto-stopping (user decides)

v0.1.0

• Initial release: converged, oscillating, bistable, diverging, warming_up

Works With

• FEM / CFD solvers (residual monitoring)
• ML training (loss convergence)
• Monte Carlo simulations (running averages)
• Optimization loops (objective function)
• Molecular dynamics (energy equilibration)
• Any iterative numerical process that produces a scalar time series

License

MIT

Author

Maximilian Jurak — 3 Kaiserberge Engineering & Research

[e-mail]