Convex Least Squares Programming (CLSP) is a two-step estimator for solving underdetermined, ill-posed, or structurally constrained least-squares problems. It combines pseudoinverse-based estimation with convex-programming correction (Lasso, Ridge, Elastic Net) to ensure numerical stability, constraint enforcement, and interpretability. The package also provides numerical stability analysis and CLSP-specific diagnostics, including partial R², normalized RMSE (NRMSE), Monte Carlo t-tests for mean NRMSE, and condition-number-based confidence bands. All calculations use numpy.float64 precision.

pip install pyclsp

import   numpy as     np
from     clsp  import CLSP

# CMLS (RP), based on known stationary points for y = D @ x + e, x to be estimated

seed     = 123456789
rng      = np.random.default_rng(seed)

# sample (dataset)
k        = 500                                               # number of observations in D
p        = 6                                                 # number of regressors
c        = 1                                                 # sum of coefficients
D        = np.empty((k, p))
D[:, 0 ] = 1.0                                               # constant
D[:, 1:] = rng.normal(size=(k, p - 1))                       # D.,j ~ N(0,1), 2 <= j <= p
b_true   = rng.normal(size=p)                                # b_true ~ N(0,1)
b_true   = (b_true / b_true.sum()) * c
e        = rng.normal(size=(k, 1))                           # e ~ N(0,1)
y        = (D @ b_true).reshape(-1, 1) + e                   # y_t = D @ b_true + e_t

# model
b        = np.vstack([
               np.asarray([c]),                              # c (the sum of coefficients)
               np.zeros((k - 2, 1)),                         # zeros
               np.zeros((k - 1, 1)),                         # zeros
               y                                             # values of y_t
           ])
C        = np.vstack([
               np.ones((1, p)),                              # a row of ones
               np.diff(D, n=2, axis=0),                      # the 2nd differences
               np.diff(D, n=1, axis=0)                       # the 1st differences
           ])
S        = np.block([
               [np.zeros((  1, k-2))],                       # a zero vector
               [np.diag(np.sign(np.diff(y.ravel(), n=2)))],  # a diagonal sign matrix
               [np.zeros((k-1, k-2))],                       # a zero matrix
           ])
model    = CLSP().solve(
               problem="cmls", b=b, C=C, S=S, M=D,
               r=1,                                          # a solution without refinement
               alpha=1.0                                     # a unique MNBLUE estimator
           )

# results
print("true beta (x_M):")
print(np.round(np.asarray(b_true).flatten(), 4))

print("beta hat (x_M hat):")
print(np.round(model.x.flatten(), 4))

print("\nNumerical stability:")
print("  kappaC :", round(model.kappaC, 4))
print("  kappaB :", round(model.kappaB, 4))
print("  kappaA :", round(model.kappaA, 4))

print("\nGoodness-of-fit:")
print("  R2_partial           :", round(model.r2_partial,    6))
print("  NRMSE_partial        :", round(model.nrmse_partial, 6))
print("  Diagnostic band (min):", np.round(np.min(model.x_lower), 4))
print("  Diagnostic band (max):", np.round(np.max(model.x_upper), 4))
print("  Monte Carlo t-test:")
for kw, val in model.ttest(sample_size=30,                   # NRMSE_partial sample
                           seed=seed, distribution="normal", # seed and distribution
                           partial=True).items():
    print(f"    {kw}: {float(val):.6f}")

For comprehensive information on the estimator’s capabilities, advanced configuration options, and implementation details, please refer to the docstrings provided in each of the individual .py source files. These docstrings contain complete descriptions of available methods, their parameters, expected input formats, and output structures.

self.__init__()

Stores the solution, goodness-of-fit statistics, and ancillary parameters.

The class has three core methods: solve(), corr(), and ttest().

Selected attributes: self.A : np.ndarray design matrix A = [C | S; M | Q], where Q is either a zero matrix or S_residual.

self.b : np.ndarray vector of the right-hand side.

self.zhat : np.ndarray vector of the first-step estimate.

self.r : int number of refinement iterations performed in the first step.

self.z : np.ndarray vector of the final solution. If the second step is disabled, it equals self.zhat.

self.x : np.ndarray m x p matrix or vector containing the variable component of z.

self.y : np.ndarray vector containing the slack component of z.

self.kappaC : float spectral κ() for C_canon.

self.kappaB : float spectral κ() for B = C_canon^+A.

self.kappaA : float spectral κ() for A.

self.rmsa : float total root mean square alignment (RMSA).

self.r2_partial : float R^2 for the M block in A.

self.nrmse : float mean square error calculated from A and normalized by standard deviation (NRMSE).

self.nrmse_partial : float mean square error calculated from the M block in A and normalized by standard deviation (NRMSE).

self.z_lower : np.ndarray lower bound of the diagnostic interval (confidence band) based on κ(A).

self.z_upper : np.ndarray upper bound of the diagnostic interval (confidence band) based on κ(A).

self.x_lower : np.ndarray lower bound of the diagnostic interval (confidence band) based on κ(A).

self.x_upper : np.ndarray upper bound of the diagnostic interval (confidence band) based on κ(A).

self.y_lower : np.ndarray lower bound of the diagnostic interval (confidence band) based on κ(A).

self.y_upper : np.ndarray upper bound of the diagnostic interval (confidence band) based on κ(A).

self.solve(problem, C, S, M, b, m, p, i, j, zero_diagonal, r, Z, tolerance, iteration_limit, final, alpha)

Solves the Convex Least Squares Programming (CLSP) problem.

This method performs a two-step estimation: (1) a pseudoinverse-based solution using either the Moore–Penrose or Bott–Duffin inverse, optionally iterated for refinement; (2) a convex-programming correction using Lasso, Ridge, or Elastic Net regularization (if enabled).

Parameters: problem : str, optional Structural template for matrix construction. One of: - 'ap' or 'tm' : allocation (tabular) matrix problem (AP). - 'cmls' or 'rp' : constrained-model least squares (regression) problem. - anything else: general CLSP problem (user-defined C and/or M).

C, S, M : np.ndarray or None Blocks of the design matrix A = [C | S; M | Q]. If C and/or M are provided, the matrix A is constructed accordingly (please note that for AP, C is constructed automatically and known values are specified in M).

b : np.ndarray or None Right-hand side vector. Must have as many rows as A (please note that for AP, it should start with row sums). Required.

m, p : int or None Dimensions of X ∈ ℝ^{m×p}, relevant for AP.

i, j : int, default = 1 Grouping sizes for row and column sum constraints in AP.

zero_diagonal : bool, default = False If True, enforces structural zero diagonals.

r : int, default = 1 Number of refinement iterations for the pseudoinverse-based estimator.

Z : np.ndarray or None A symmetric idempotent matrix (projector) defining the subspace for Bott–Duffin pseudoinversion. If None, the identity matrix is used, reducing the Bott–Duffin inverse to the Moore–Penrose case.

tolerance : float, default = square root of machine epsilon Convergence tolerance for NRMSE change between refinement iterations.

iteration_limit : int, default = 50 Maximum number of iterations allowed in the refinement loop.

final : bool, default = True If True, a convex programming problem is solved to refine zhat. The resulting solution z minimizes a weighted L1/L2 norm around zhat subject to Az = b.

alpha : float, list[float] or None, default = None Regularization parameter (weight) in the final convex program: - α = 0: Lasso (L1 norm) - α = 1: Tikhonov Regularization/Ridge (L2 norm) - 0 < α < 1: Elastic Net If a scalar float is provided, that value is used after clipping to [0, 1]. If a list/iterable of floats is provided, each candidate is evaluated via a full solve, and the α with the smallest NRMSE is selected. If None, α is chosen, based on an error rule: α = min(1.0, NRMSE_{α = 0} / (NRMSE_{α = 0} + NRMSE_{α = 1} + tolerance))

*args, **kwargs : optional CVXPY arguments passed to the CVXPY solver.

Returns: self

self.corr(reset, threshold)

Computes the structural correlogram of the CLSP constraint part.

This method performs a row-deletion sensitivity analysis on the canonical constraint matrix [C | S], denoted as C_canon, and evaluates the marginal effect of each constraint row on numerical stability, angular alignment, and estimator sensitivity.

For each row i in C_canon, it computes: - The Root Mean Square Alignment (RMSA_i) with all other rows j ≠ i. - The change in condition numbers κ(C), κ(B), and κ(A) when row i is deleted. - The effect on estimation quality: changes in nrmse, zhat, z, and x when row i is deleted.

Additionally, it computes the total rmsa statistic across all rows, summarizing the overall angular alignment of C_canon.

Parameters: reset : bool, default = False If True, forces recomputation of all diagnostic values (the results are preserved for eventual reproduction after the method is called).

threshold : float, default = 0 If positive, limits the output to constraints with RMSA_i ≥ threshold.

Returns: dict of list A dictionary containing per-row diagnostic values: { "constraint" : [1, 2, ..., k], # 1-based indices "rmsa_i" : list of RMSA_i values, "rmsa_dkappaC" : list of Δκ(C) after deleting row i, "rmsa_dkappaB" : list of Δκ(B) after deleting row i, "rmsa_dkappaA" : list of Δκ(A) after deleting row i, "rmsa_dnrmse" : list of Δnrmse after deleting row i, "rmsa_dzhat" : list of Δzhat after deleting row i, "rmsa_dz" : list of Δz after deleting row i, "rmsa_dx" : list of Δx after deleting row i, }

self.ttest(reset, sample_size, seed, distribution)

Performs a Monte Carlo-based one- or two-sided t-test on the NRMSE statistic.

This function simulates right-hand side vectors b using a user-defined or default distribution and recomputes the estimator for every new b. It tests whether the observed NRMSE significantly deviates from the null distribution (under H₀) of simulated NRMSE values. The quality of the test depends on the size of the simulated sample.

Parameters: reset : bool, default = False If True, forces recomputation of the NRMSE null distribution (under H₀) (the results are preserved for eventual reproduction after the method is called).

sample_size : int, default = 50 Size of the Monte Carlo simulated sample under H₀.

seed : int or None, optional Optional random seed to override the default.

distribution : str or None, default = ’normal’ Distribution for generating simulated b vectors. One of (standard): 'normal', 'uniform', or 'laplace'.

partial : bool, default = False If True, runs the t-test on the partial NRMSE: during simulation, the C-block entries are preserved and the M-block entries are simulated.

Returns: dict Dictionary with test results and null distribution statistics: { 'p_one_left' : P(nrmse ≤ null mean), 'p_one_right' : P(nrmse ≥ null mean), 'p_two_sided' : 2-sided t-test p-value, 'nrmse' : observed value, 'mean_null' : mean of the null distribution (under H₀), 'std_null' : standard deviation of the null distribution (under H₀) }

To be added.

MIT License — see the LICENSE file.