BsplineQuantReg

Citation

If you use this package in your research, please cite:

```bibtex @Article{Abbes2026, author = {Alexandre Abbes}, title = {Quantile regression with cubic polynomial splines under shape constraints with applications}, year = {2026}, doi = {10.5281/zenodo.17427913} }

Beta Version Notice

This package is currently in beta (version 0.1.0-beta).

What to expect:

✅ Core functionality (quantile regression, monotonicity, convexity) is stable
⚠️ API may change based on user feedback
📝 Documentation is being refined
🐛 Please report issues on GitHub

R Packages

Package	Description	Constraint Type	Spline Degree
ConstrainedQuantileSplines (this package)	Quantile regression with Karlin-Studden constraints	Monotonicity, Convexity	Cubic
quantreg	Classical quantile regression	None (linear programming)	Linear
cobs	Constrained B-splines	Monotonicity, Convexity	Linear, Quadratic

Comparison with cobs

The cobs package (Constrained B-Splines) is the closest to this package, but with key differences:

Feature	ConstrainedQuantileSplines	cobs
Spline degree	Cubic (degree 3)	Linear, Quadratic
Constraint method	Karlin-Studden SOCP	Traditional constraints
Convexity	✅ Yes	✅ Yes
Monotonicity	✅ Yes	✅ Yes
Quantile regression	✅ Yes	✅ Yes
Partial constraints	✅ Yes (per interval)	Limited
Polynomial coefficient export	✅ Yes	❌ No

When to use this package vs cobs

Use cobs : For linear or quadratic splines, simpler constraints
Use this package : For cubic splines testing , Karlin-Studden exact constraints, polynomial coefficient extraction

⚠️ Performance Notice

This R package is currently intended for demonstration, prototyping, and educational purposes only.

Due to the current implementation (pure R with CVXR), the package is significantly slower than its Python counterpart. Cubic B-spline quantile regression with constraints involves solving SOCP problems, and the R implementation does not yet leverage optimized linear algebra libraries. Python version

Current benchmarking (median regression, n=1000, kn=20, cubic splines):

⚠️ The R version is currently 50-100x slower than Python (estimation)

Future Improvements

We plan to improve performance in future releases by: - Linking with faster optimization libraries (OSQP, Gurobi) - Implementing more efficient SOCP solvers - Optimizing the B-spline basis computation

The Python version remains the recommended choice for production use.

Installation for beta testing:

Install from GitHub

pak::pak(“alexandreabbes/BsplineQuantReg”)

mirror server hosted at Truenetwork, Russian Federation.