timeseries-formula-finder 0.1.0

Symbolic Form Discovery - Find interpretable mathematical formulas in data

PPF - Symbolic Form Discovery

Discover interpretable mathematical formulas from data, then deploy them anywhere.

PPF (Promising Partial Form) is a law-aware symbolic regression library that finds compact mathematical expressions describing your data. Unlike black-box neural networks, PPF produces human-readable formulas grounded in physical reality that can be deployed to edge devices in under 100 bytes.

Core Philosophy

A Promising Partial Form is a mathematical expression that explains a meaningful portion of a signal with high explanatory power and low complexity. Complex signals are modeled as layered sums:

y(t) ≈ f₁(t) + f₂(t) + ... + fₖ(t) + ε(t)

where each fᵢ(t) is a discovered form (oscillation, decay, trend, etc.) and ε(t) is noise-like residual.

PPF repeatedly discovers → subtracts → analyzes residuals to reveal the layered mathematical structure hidden in data. This approach applies to any 1D time-series or signal analysis task - from real-time sensor monitoring to scientific data exploration to serving as an interpretable front-end to downstream AI systems.

Why PPF?

Traditional ML	PPF Approach
Train neural network	Discover formula
Deploy 10-100KB model	Deploy 50-byte expression
1000s of MACs/inference	10 FLOPs/evaluation
Black box	Interpretable
Requires TensorFlow Lite	Runs on any microcontroller

Key insight: Many real-world signals (sensor data, biological rhythms, physical systems) follow simple mathematical forms. PPF attempts to discover those forms automatically.

Information-theoretic view: Where neural networks compress data into opaque weight matrices, PPF compresses data into explicit equations. This enables a layered architecture: Raw Signal → PPF (laws) → ML (meaning) → Decisions, where PPF converts high-entropy measurements into low-entropy symbolic representations.

Quick Start

pip install numpy scipy

import numpy as np
from ppf import SymbolicRegressor, export_python, export_c

# Generate example data
t = np.linspace(0, 10, 200)
y = 2.5 * np.exp(-0.3 * t) * np.sin(4.0 * t + 0.5) + np.random.randn(200) * 0.1

# Discover the underlying formula
regressor = SymbolicRegressor(generations=30)
result = regressor.discover(t, y, verbose=True)

print(f"Discovered: {result.best_tradeoff.expression_string}")
print(f"R-squared:  {result.best_tradeoff.r_squared:.4f}")
# Output: Discovered: 2.498*exp(-0.301*x)*sin(4.002*x + 0.497)
# Output: R-squared:  0.9847

# Export to standalone Python (no PPF dependency)
python_code = export_python(result.best_tradeoff.expression, fn_name="predict")
exec(python_code)
print(predict(1.5))  # Works without importing ppf

# Export to C for embedded deployment
c_code = export_c(result.best_tradeoff.expression, use_float=True)
# Compile with: gcc -std=c99 -O2 -lm model.c

Features

Multi-Domain Discovery

PPF includes domain-specific "vocabularies" that guide the search:

from ppf import SymbolicRegressor, DiscoveryMode

regressor = SymbolicRegressor()

# Let PPF auto-detect the best domain
result = regressor.discover(x, y, mode=DiscoveryMode.AUTO)

# Or specify a domain if you know it
result = regressor.discover(x, y, mode=DiscoveryMode.OSCILLATOR)   # Vibrations, waves
result = regressor.discover(x, y, mode=DiscoveryMode.CIRCUIT)      # RC charging, decay
result = regressor.discover(x, y, mode=DiscoveryMode.GROWTH)       # Sigmoids, saturation
result = regressor.discover(x, y, mode=DiscoveryMode.RATIONAL)     # Ratios, feedback
result = regressor.discover(x, y, mode=DiscoveryMode.UNIVERSAL)    # Power laws, Gaussians

Macro Templates

PPF includes "macros" - pre-composed functional forms that capture common physics:

Macro	Formula	Use Case
DAMPED_SIN	a·exp(-k·t)·sin(ω·t + φ)	Vibration decay
RC_CHARGE	a·(1 - exp(-k·t)) + c	Capacitor charging
GAUSSIAN	a·exp(-((x-μ)/σ)²)	Peaks, distributions
SIGMOID	a / (1 + exp(-k·(x-x₀)))	Saturation curves
POWER_LAW	a·x^b + c	Scaling phenomena
HILL	a·x^n / (k^n + x^n)	Enzyme kinetics

These let PPF find complex forms (multiplicative compositions across function families) that pure GP struggles with.

Export to Edge Devices

Discovered formulas export to production-ready code:

from ppf import export_python, export_c, export_json

# Standalone Python evaluator
code = export_python(expr, fn_name="predict_temp", safe=True)
# Includes safety wrappers for div-by-zero, log(0), exp overflow

# C99 for microcontrollers
code = export_c(expr, use_float=True, safe=True)
# Compiles on ESP32, STM32, Arduino, etc.

# JSON bundle for storage/transmission
bundle = export_json(result, variables=["t"])
# Send via MQTT, store in database, audit trail

Feature Extraction for Downstream ML

Use discovered forms as features for classification/regression:

from ppf import extract_features, feature_vector

# Extract interpretable features
features = extract_features(result)
print(features["dominant_family"])  # "oscillation"
print(features["damping_k"])        # 0.301
print(features["omega"])            # 4.002

# Convert to ML-ready vector
vec, names = feature_vector(features, schema="ppf.features.v1.full")
# Feed to sklearn, XGBoost, etc.

Installation

# Clone the repository
git clone https://github.com/pcoz/timeseries-formula-finder.git
cd ppf

# Install dependencies
pip install numpy scipy

# Optional: for hybrid decomposition (EMD/SSA)
pip install PyEMD

# Run tests
pytest tests/ -v

Documentation

Document	Description
USER_GUIDE.md	Complete usage guide with examples
docs/PPF_Paper.md	Core PPF concepts and architecture
docs/PPF_Information_Theory_Paper.md	Information-theoretic foundations
COMPARISON.md	How PPF compares to PySR, gplearn, Eureqa, AI Feynman
USE_CASES.md	Edge AI, ECG analysis, predictive maintenance
TESTING.md	Test suite details and datasets used
docs/PPF_EXPORT_LAYER_TSD.md	Export layer technical specification
docs/IOT_SENSOR_ANALYSIS.md	Real IoT sensor case study
docs/ECG_ANALYSIS.md	ECG waveform analysis case study
DOCUMENTATION.md	Comprehensive API documentation

Use Cases

Edge AI / IoT Sensors

Deploy temperature prediction models to ESP32 in 50 bytes instead of TensorFlow Lite:

# Discover daily temperature cycle from sensor data
result = regressor.discover(hours, temp, mode=DiscoveryMode.OSCILLATOR)
# Found: T(t) = 36 + 5·cos(2π·t/24 - 2)  [R² = 0.58]

# Export to C for microcontroller
c_code = export_c(result.best_tradeoff.expression, use_float=True)
# 3 parameters, 10 FLOPs, runs at 100kHz on ESP32

Biomedical Signal Analysis

Extract interpretable features from ECG waveforms:

# Analyze T-wave morphology
result = regressor.discover(t_wave_time, t_wave_amplitude, mode=DiscoveryMode.UNIVERSAL)
# Found: Damped cosine (R² = 0.96) - indicates repolarization dynamics

# Use parameters as cardiac health features
features = extract_features(result)
# damping_k, omega → feed to arrhythmia classifier

Predictive Maintenance

Classify machine health from vibration signatures:

# Healthy machine: clean sinusoid
# Bearing fault: damped oscillations (impact response)
# Imbalance: dominant 1x RPM

# The discovered FORM is the diagnosis
result = regressor.discover(t, vibration, mode=DiscoveryMode.OSCILLATOR)
if "exp(-" in result.best_tradeoff.expression_string:
    print("Bearing fault detected - damped oscillation signature")

Architecture

ppf/
├── symbolic_types.py     # Expression trees, macros, primitives
├── symbolic.py           # GP engine, symbolic regressor
├── symbolic_utils.py     # Printing, simplification
├── detector.py           # Fixed-form fitting (legacy)
├── residual_layer.py     # Multi-layer decomposition
├── hierarchical.py       # Windowed parameter evolution
├── hybrid.py             # EMD/SSA + PPF interpretation
├── export/
│   ├── python_export.py  # export_python()
│   ├── c_export.py       # export_c()
│   ├── json_export.py    # export_json()
│   └── load.py           # load_json()
└── features/
    ├── extract.py        # extract_features()
    └── vectorize.py      # feature_vector()

Benchmarks

Performance on standard symbolic regression benchmarks:

Benchmark	PPF R²	Complexity	Notes
Kepler's 3rd Law	0.9999	3	T² = a³
Damped Oscillator	0.9847	8	Macro template
Logistic Growth	0.9923	5	Sigmoid macro
Nguyen-1 (x³+x²+x)	0.9998	7	Polynomial mode

See BENCHMARK_COMPARISON.md for detailed results.

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

Areas of interest:

• Additional macro templates for new domains
• Performance optimization of GP engine
• New export targets (Rust, WASM, ONNX)
• Integration with AutoML frameworks

License

MIT License - see LICENSE for details.

Copyright (c) 2026 Edward Chalk (fleetingswallow.com)

Citation

If you use PPF in research, please cite:

@software{timeseries-formula-finder,
  author = {Chalk, Edward},
  title = {Timeseries Formula Finder: Discover Mathematical Forms in Time-Series Data},
  year = {2026},
  url = {https://github.com/pcoz/timeseries-formula-finder}
}

Contact

• Website: fleetingswallow.com
• Email: [e-mail]