freealg 0.7.5__tar.gz → 0.7.7__tar.gz

{freealg-0.7.5 → freealg-0.7.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: freealg
-Version: 0.7.5
+Version: 0.7.7
 Summary: Free probability for large matrices
 Home-page: https://github.com/ameli/freealg
 Download-URL: https://github.com/ameli/freealg/archive/main.zip

freealg-0.7.7/freealg/__version__.py

@@ -0,0 +1 @@
+__version__ = "0.7.7"

freealg-0.7.7/freealg/_algebraic_form/_constraints.py

@@ -0,0 +1,98 @@
+
+# SPDX-FileCopyrightText: Copyright 2025, Siavash Ameli <sameli@berkeley.edu>
+# SPDX-License-Identifier: BSD-3-Clause
+# SPDX-FileType: SOURCE
+#
+# This program is free software: you can redistribute it and/or modify it under
+# the terms of the license found in the LICENSE.txt file in the root directory
+# of this source tree.
+
+
+# =======
+# Imports
+# =======
+
+import numpy
+
+__all__ = ['build_moment_constraint_matrix']
+
+
+# ==========
+# series mul
+# ==========
+
+def _series_mul(a, b, q_max):
+
+    na = min(len(a), q_max + 1)
+    nb = min(len(b), q_max + 1)
+    out = numpy.zeros(q_max + 1, dtype=float)
+    for i in range(na):
+        if a[i] == 0.0:
+            continue
+        j_max = min(nb - 1, q_max - i)
+        if j_max >= 0:
+            out[i:i + j_max + 1] += a[i] * b[:j_max + 1]
+    return out
+
+
+# ==========
+# series pow
+# ==========
+
+def _series_pow(mser, j, q_max):
+    if j == 0:
+        out = numpy.zeros(q_max + 1, dtype=float)
+        out[0] = 1.0
+        return out
+    out = mser.copy()
+    for _ in range(1, j):
+        out = _series_mul(out, mser, q_max)
+    return out
+
+
+# ===============================
+# build moment constraints matrix
+# ===============================
+
+def build_moment_constraint_matrix(pairs, deg_z, s, mu):
+
+    mu = numpy.asarray(mu, dtype=float).ravel()
+    if mu.size == 0:
+        return numpy.zeros((0, len(pairs)), dtype=float)
+
+    # m(z) = -sum_{p>=0} mu_p / z^{p+1}; t = 1/z so m(t) = -sum mu_p t^{p+1}
+    r = mu.size - 1
+    q_max = r
+
+    mser = numpy.zeros(q_max + 1, dtype=float)
+    for p in range(mu.size):
+        q = p + 1
+        if q <= q_max:
+            mser[q] = -float(mu[p])
+
+    # Precompute (m(t))^j coefficients up to t^{q_max}
+    mpow = []
+    for j in range(s + 1):
+        mpow.append(_series_pow(mser, j, q_max))
+
+    # Constraints: coeff of t^q in Q(t) := t^{deg_z} P(1/t, m(t)) must be 0
+    # Q(t) = sum_{i,j} c_{i,j} * t^{deg_z - i} * (m(t))^j
+    n_coef = len(pairs)
+    B = numpy.zeros((q_max + 1, n_coef), dtype=float)
+
+    for k, (i, j) in enumerate(pairs):
+        shift = deg_z - i
+        if shift < 0:
+            continue
+        mj = mpow[j]
+        for q in range(q_max + 1):
+            qq = q - shift
+            if 0 <= qq <= q_max:
+                B[q, k] = mj[qq]
+
+    # Drop all-zero rows (can happen if index-set can't support higher moments)
+    row_norm = numpy.linalg.norm(B, axis=1)
+    keep = row_norm > 0.0
+    B = B[keep, :]
+
+    return B

{freealg-0.7.5 → freealg-0.7.7}/freealg/_algebraic_form/_continuation_algebraic.py

@@ -13,6 +13,7 @@
 
 import numpy
 from .._geometric_form._continuation_genus0 import joukowski_z
+from ._constraints import build_moment_constraint_matrix
 
 __all__ = ['sample_z_joukowski', 'filter_z_away_from_cuts', 'powers',
            'fit_polynomial_relation', 'sanity_check_stieltjes_branch',
@@ -131,7 +132,11 @@ def powers(x, deg):
 # =======================
 
 def fit_polynomial_relation(z, m, s, deg_z, ridge_lambda=0.0, weights=None,
-                            triangular=None, normalize=False):
+                            triangular=None, normalize=False,
+                            mu=None, mu_reg=None):
+    """
+    Fits polynomial P(z, m) = 0 with samples from the physical branch.
+    """
 
     z = numpy.asarray(z, dtype=complex).ravel()
     m = numpy.asarray(m, dtype=complex).ravel()
@@ -198,13 +203,82 @@ def fit_polynomial_relation(z, m, s, deg_z, ridge_lambda=0.0, weights=None,
     s_col[s_col == 0.0] = 1.0
     As = Ar / s_col[None, :]
 
-    if ridge_lambda > 0.0:
-        L = numpy.sqrt(ridge_lambda) * numpy.eye(n_coef, dtype=float)
-        As = numpy.vstack([As, L])
+    # Optional moment constraints B c = 0 (hard via nullspace, soft via
+    # weighted rows)
+    if mu is not None:
+        B = build_moment_constraint_matrix(pairs, deg_z, s, mu)
+        if B.shape[0] > 0:
+            Bs = B / s_col[None, :]
+
+            if mu_reg is None:
+                # Hard constraints: solve in nullspace of Bs
+                uB, sB, vhB = numpy.linalg.svd(Bs, full_matrices=True)
+                tolB = 1e-12 * (sB[0] if sB.size else 1.0)
+                rankB = int(numpy.sum(sB > tolB))
+                if rankB >= n_coef:
+                    raise RuntimeError(
+                        'Moment constraints leave no feasible coefficients.')
+
+                N = vhB[rankB:, :].T  # (n_coef, n_free)
+                AN = As @ N
+
+                if ridge_lambda > 0.0:
+                    L = numpy.sqrt(ridge_lambda) * numpy.eye(N.shape[1],
+                                                             dtype=float)
+                    AN = numpy.vstack([AN, L])
+
+                _, svals, vhN = numpy.linalg.svd(AN, full_matrices=False)
+                y = vhN[-1, :]
+                coef_scaled = N @ y
+
+                coef = coef_scaled / s_col
+
+            else:
+                mu_reg = float(mu_reg)
+                if mu_reg > 0.0:
+                    As_aug = As
+                    Bs_w = numpy.sqrt(mu_reg) * Bs
+                    As_aug = numpy.vstack([As_aug, Bs_w])
+
+                    if ridge_lambda > 0.0:
+                        L = numpy.sqrt(ridge_lambda) * numpy.eye(n_coef,
+                                                                 dtype=float)
+                        As_aug = numpy.vstack([As_aug, L])
+
+                    _, svals, vh = numpy.linalg.svd(As_aug,
+                                                    full_matrices=False)
+                    coef_scaled = vh[-1, :]
+                    coef = coef_scaled / s_col
+                else:
+                    # mu_reg == 0 => ignore constraints
+                    if ridge_lambda > 0.0:
+                        L = numpy.sqrt(ridge_lambda) * numpy.eye(n_coef,
+                                                                 dtype=float)
+                        As = numpy.vstack([As, L])
+
+                    _, svals, vh = numpy.linalg.svd(As, full_matrices=False)
+                    coef_scaled = vh[-1, :]
+                    coef = coef_scaled / s_col
+
+        else:
+            # B has no effective rows -> proceed unconstrained
+            if ridge_lambda > 0.0:
+                L = numpy.sqrt(ridge_lambda) * numpy.eye(n_coef, dtype=float)
+                As = numpy.vstack([As, L])
 
-    _, _, vh = numpy.linalg.svd(As, full_matrices=False)
-    coef_scaled = vh[-1, :]
-    coef = coef_scaled / s_col
+            _, svals, vh = numpy.linalg.svd(As, full_matrices=False)
+            coef_scaled = vh[-1, :]
+            coef = coef_scaled / s_col
+
+    else:
+        # No moment constraints
+        if ridge_lambda > 0.0:
+            L = numpy.sqrt(ridge_lambda) * numpy.eye(n_coef, dtype=float)
+            As = numpy.vstack([As, L])
+
+        _, svals, vh = numpy.linalg.svd(As, full_matrices=False)
+        coef_scaled = vh[-1, :]
+        coef = coef_scaled / s_col
 
     full = numpy.zeros((deg_z + 1, s + 1), dtype=complex)
     for k, (i, j) in enumerate(pairs):
@@ -213,7 +287,14 @@ def fit_polynomial_relation(z, m, s, deg_z, ridge_lambda=0.0, weights=None,
     if normalize:
         full = _normalize_coefficients(full)
 
-    return full
+    # Diagnostic metrics
+    fit_metrics = {
+        's_min': svals[-1],
+        'gap_ratio': float(svals[-2] / svals[-1]),
+        'n_small': float(int(numpy.sum(svals <= svals[0] * 1e-12))),
+    }
+
+    return full, fit_metrics
 
 
 # =============================

freealg-0.7.7/freealg/_algebraic_form/_discriminant.py

@@ -0,0 +1,226 @@
+# SPDX-FileCopyrightText: Copyright 2026, Siavash Ameli <sameli@berkeley.edu>
+# SPDX-License-Identifier: BSD-3-Clause
+# SPDX-FileType: SOURCE
+#
+# This program is free software: you can redistribute it and/or modify it under
+# the terms of the license found in the LICENSE.txt file in the root directory
+# of this source tree.
+
+
+# =======
+# Imports
+# =======
+
+import numpy as np
+import numpy.polynomial.polynomial as poly
+
+__all__ = ['compute_singular_points']
+
+
+# =========
+# ploy trim
+# =========
+
+def _poly_trim(p, tol):
+
+    p = np.asarray(p, dtype=complex).ravel()
+    if p.size == 0:
+        return np.zeros(1, dtype=complex)
+    k = p.size - 1
+    while k > 0 and abs(p[k]) <= tol:
+        k -= 1
+    return p[:k + 1].copy()
+
+
+# ============
+# poly is zero
+# ============
+
+def _poly_is_zero(p, tol):
+
+    p = _poly_trim(p, tol)
+    return (p.size == 1) and (abs(p[0]) <= tol)
+
+
+# ========
+# poly add
+# ========
+
+def _poly_add(a, b, tol):
+
+    return _poly_trim(poly.polyadd(a, b), tol)
+
+
+# ========
+# poly sub
+# ========
+
+def _poly_sub(a, b, tol):
+
+    return _poly_trim(poly.polysub(a, b), tol)
+
+
+# =======
+# ply mul
+# =======
+
+def _poly_mul(a, b, tol):
+
+    return _poly_trim(poly.polymul(a, b), tol)
+
+
+# ==============
+# poly div exact
+# ==============
+
+def _poly_div_exact(a, b, tol):
+
+    a = _poly_trim(a, tol)
+    b = _poly_trim(b, tol)
+    if _poly_is_zero(b, tol):
+        raise ZeroDivisionError("poly division by zero")
+
+    q, r = poly.polydiv(a, b)
+    r = _poly_trim(r, tol)
+
+    # Bareiss expects exact division; with floats it's only approximate.
+    # If the remainder is small, drop it.
+    scale = max(1.0, np.linalg.norm(a))
+    if np.linalg.norm(r) > 1e3 * tol * scale:
+        # Fallback: still drop remainder (keeps algorithm running).
+        # This is acceptable because we only need the resultant roots
+        # robustly, not exact symbolic coefficients.
+        pass
+
+    return _poly_trim(q, tol)
+
+
+# ================
+# det bareiss poly
+# ================
+
+def _det_bareiss_poly(M, tol):
+
+    n = len(M)
+    A = [[_poly_trim(M[i][j], tol) for j in range(n)] for i in range(n)]
+
+    denom = np.array([1.0], dtype=complex)
+    sign = 1.0
+
+    for k in range(n - 1):
+        if _poly_is_zero(A[k][k], tol):
+            piv = -1
+            for i in range(k + 1, n):
+                if not _poly_is_zero(A[i][k], tol):
+                    piv = i
+                    break
+            if piv == -1:
+                return np.array([0.0], dtype=complex)
+            A[k], A[piv] = A[piv], A[k]
+            sign *= -1.0
+
+        pivot = A[k][k]
+        for i in range(k + 1, n):
+            for j in range(k + 1, n):
+                num = _poly_sub(_poly_mul(A[i][j], pivot, tol),
+                                _poly_mul(A[i][k], A[k][j], tol),
+                                tol)
+                if k > 0:
+                    A[i][j] = _poly_div_exact(num, denom, tol)
+                else:
+                    A[i][j] = _poly_trim(num, tol)
+
+        denom = pivot
+
+    return _poly_trim(sign * A[n - 1][n - 1], tol)
+
+
+# ===================
+# cluster real points
+# ===================
+
+def _cluster_real_points(x, eps):
+
+    x = np.asarray(x, dtype=float).ravel()
+    if x.size == 0:
+        return x
+    x = np.sort(x)
+    uniq = []
+    for v in x:
+        if (len(uniq) == 0) or (abs(v - uniq[-1]) > eps):
+            uniq.append(float(v))
+        else:
+            uniq[-1] = 0.5 * (uniq[-1] + float(v))
+    return np.asarray(uniq, dtype=float)
+
+
+# =======================
+# compute singular points
+# =======================
+
+def compute_singular_points(a_coeffs, tol=1e-12, real_tol=None):
+    """
+    a_coeffs[i,j] is coefficient of z^i m^j, shape (deg_z+1, s+1).
+
+    Returns
+    -------
+
+    z_bp     : complex array, roots of Disc_m(P)(z)
+    a_s_zero : complex array, roots of leading coefficient a_s(z)
+    support  : list of (a,b) from real-ish branch points paired consecutively
+    """
+
+    a_coeffs = np.asarray(a_coeffs)
+    s = a_coeffs.shape[1] - 1
+    if s < 1:
+        return (np.array([], dtype=complex),
+                np.array([], dtype=complex),
+                [])
+
+    if real_tol is None:
+        real_tol = 1e3 * tol
+
+    a = [_poly_trim(a_coeffs[:, j], tol) for j in range(s + 1)]
+
+    a_s = a[s]
+    a_s_zero = np.roots(a_s[::-1]) if a_s.size > 1 else \
+        np.array([], dtype=complex)
+
+    b = []
+    for j in range(s):
+        b.append(_poly_trim((j + 1) * a[j + 1], tol))
+
+    mdeg = s
+    ndeg = s - 1
+    N = mdeg + ndeg  # 2s-1
+
+    z0 = np.array([0.0], dtype=complex)
+    M = [[z0 for _ in range(N)] for __ in range(N)]
+
+    for r in range(ndeg):
+        for j in range(mdeg + 1):
+            M[r][r + j] = a[j]
+
+    for r in range(mdeg):
+        rr = ndeg + r
+        for j in range(ndeg + 1):
+            M[rr][r + j] = b[j]
+
+    res = _det_bareiss_poly(M, tol)
+    if res.size <= 1:
+        z_bp = np.array([], dtype=complex)
+    else:
+        z_bp = np.roots(res[::-1])
+
+    support = []
+    if z_bp.size > 0:
+        zr = z_bp[np.abs(z_bp.imag) <= real_tol].real
+        zr = _cluster_real_points(zr, eps=1e2 * real_tol)
+        m2 = (zr.size // 2) * 2
+        for k in range(0, m2, 2):
+            a0 = float(zr[k])
+            b0 = float(zr[k + 1])
+            if b0 > a0:
+                support.append((a0, b0))
+
+    return z_bp, a_s_zero, support

{freealg-0.7.5 → freealg-0.7.7}/freealg/_algebraic_form/algebraic_form.py

@@ -21,6 +21,7 @@ from ._edge import evolve_edges, merge_edges
 from ._decompress import decompress_newton
 from ._decompress2 import decompress_coeffs
 from ._homotopy import stieltjes_poly
+from ._discriminant import compute_singular_points
 from .._free_form._support import supp
 from .._free_form._plot_util import plot_density
 
@@ -194,6 +195,7 @@ class AlgebraicForm(object):
 
         # Initialize
         self.a_coeffs = None               # Polynomial coefficients
+        self.status = None                 # Fitting status
         self.cache = {}                    # Cache inner-computations
 
     # ===
@@ -207,10 +209,38 @@ class AlgebraicForm(object):
             y_eps=2e-2,
             x_pad=0.0,
             triangular=None,
+            mu=None,
+            mu_reg=None,
             normalize=False,
             verbose=False):
         """
-        Fits polynomial.
+        Fit polynomial.
+
+        Parameters
+        ----------
+
+        deg_m : int
+            Degree :math:`\\deg_m(P)`
+
+        deg_z : int
+            Degree :math:`\\deg_z(P)`
+
+        mu : array_like, default=None
+            If an array :math:`[\\mu_0, \\mu_`, \\dots, \\mu_r]` is given,
+            it enforces the first :math:`r+1` moments. Note that :math:`\\mu_0`
+            should be :math:`1` to ensure unit mass. See also ``mu_reg`.
+
+        mu_reg: float, default=None
+            If `None`, the constraints ``mu`` are applied as hard constraint.
+            If a positive number, the constraints are applied as a soft
+            constraints with regularisation ``mu_reg``.
+
+        Notes
+        -----
+
+        When the input data are from an exact model, hard moment constraint is
+        preferred over soft constraint as the latter can hurt an already a good
+        fit.
         """
 
         # Very important: reset cache whenever this function is called. This
@@ -231,11 +261,15 @@ class AlgebraicForm(object):
         z_fit = filter_z_away_from_cuts(z_fit, self.support, y_eps=y_eps,
                                         x_pad=x_pad)
 
+        # Fitting (w_inf = None means adaptive weight selection)
         m1_fit = self.stieltjes(z_fit)
-        a_coeffs = fit_polynomial_relation(z_fit, m1_fit, s=deg_m, deg_z=deg_z,
-                                           ridge_lambda=reg,
-                                           triangular=triangular,
-                                           normalize=normalize)
+        a_coeffs, fit_metrics = fit_polynomial_relation(
+                z_fit, m1_fit, s=deg_m, deg_z=deg_z, ridge_lambda=reg,
+                triangular=triangular, normalize=normalize, mu=mu,
+                mu_reg=mu_reg)
+
+        # Compute global branhc points, zeros of leading a_j, and support
+        branch_points, a_s_zero, support = compute_singular_points(a_coeffs)
 
         self.a_coeffs = a_coeffs
 
@@ -251,8 +285,12 @@ class AlgebraicForm(object):
                                                eta=max(y_eps, 1e-2), n_x=128,
                                                max_bad_frac=0.05)
 
-        status['res_max'] = res_max
-        status['res_99_9'] = res_99_9
+        status['branch_points'] = branch_points
+        status['a_s_zero'] = a_s_zero
+        status['res_max'] = float(res_max)
+        status['res_99_9'] = float(res_99_9)
+        status['fit_metrics'] = fit_metrics
+        self.status = status
 
         if verbose:
             print(f'fit residual max  : {res_max:>0.4e}')
@@ -277,7 +315,7 @@ class AlgebraicForm(object):
             else:
                 print('\nStieltjes sanity check: OK')
 
-        return a_coeffs, status
+        return a_coeffs, support, status
 
     # =============
     # generate grid
@@ -657,11 +695,16 @@ class AlgebraicForm(object):
             z_query = x + 1j * self.delta
 
             # Initial condition at t=0 (physical branch)
-            w0_list = self.stieltjes(z_query)
+            # w0_list = self.stieltjes(z_query)
+            stieltjes = numpy.vectorize(m_fn)
+            w0_list = stieltjes(z_query)
 
             # Times
             t = numpy.log(alpha)
 
+            # Ensure it starts from t = 0
+            t = numpy.concatenate([numpy.zeros(1), t])
+
             # Evolve
             W, ok = decompress_newton(
                 z_query, t, self.a_coeffs,
@@ -669,6 +712,9 @@ class AlgebraicForm(object):
 
             rho = W.imag / numpy.pi
 
+            # Remove time zero
+            rho = rho[1:, :]
+
             if verbose:
                 print("success rate per t:", ok.mean(axis=1))
 

{freealg-0.7.5 → freealg-0.7.7}/freealg/_util.py

@@ -14,7 +14,7 @@
 import numpy
 import scipy
 
-__all__ = ['resolve_complex_dtype', 'compute_eig']
+__all__ = ['resolve_complex_dtype', 'compute_eig', 'subsample_matrix']
 
 
 # =====================
@@ -70,3 +70,26 @@ def compute_eig(A, lower=False):
     eig = scipy.linalg.eigvalsh(A, lower=lower, driver='ev')
 
     return eig
+
+
+# ================
+# subsample matrix
+# ================
+
+def subsample_matrix(matrix, submatrix_size, seed=None):
+    """
+    Generate a random subsample of a larger matrix
+    """
+
+    if matrix.shape[0] != matrix.shape[1]:
+        raise ValueError("Matrix must be square")
+
+    n = matrix.shape[0]
+    if submatrix_size > n:
+        raise ValueError("Submatrix size cannot exceed matrix size")
+
+    rng = numpy.random.default_rng(seed)
+    idx = rng.choice(n, size=submatrix_size, replace=False)
+    idx = numpy.sort(idx)  # optional, preserves original ordering
+
+    return matrix[numpy.ix_(idx, idx)]

{freealg-0.7.5 → freealg-0.7.7}/freealg.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: freealg
-Version: 0.7.5
+Version: 0.7.7
 Summary: Free probability for large matrices
 Home-page: https://github.com/ameli/freealg
 Download-URL: https://github.com/ameli/freealg/archive/main.zip

{freealg-0.7.5 → freealg-0.7.7}/freealg.egg-info/SOURCES.txt

@@ -17,9 +17,11 @@ freealg.egg-info/not-zip-safe
 freealg.egg-info/requires.txt
 freealg.egg-info/top_level.txt
 freealg/_algebraic_form/__init__.py
+freealg/_algebraic_form/_constraints.py
 freealg/_algebraic_form/_continuation_algebraic.py
 freealg/_algebraic_form/_decompress.py
 freealg/_algebraic_form/_decompress2.py
+freealg/_algebraic_form/_discriminant.py
 freealg/_algebraic_form/_edge.py
 freealg/_algebraic_form/_homotopy.py
 freealg/_algebraic_form/_sheets_util.py

freealg-0.7.5/freealg/__version__.py

@@ -1 +0,0 @@
-__version__ = "0.7.5"