scipy 1.15.3__cp312-cp312-macosx_14_0_arm64.whl → 1.16.0rc2__cp312-cp312-macosx_14_0_arm64.whl

scipy/signal/_filter_design.py

@@ -4,18 +4,23 @@ import operator
 import warnings
 
 import numpy as np
-from numpy import (atleast_1d, poly, polyval, roots, real, asarray,
-                   resize, pi, absolute, sqrt, tan, log10,
+from numpy import (atleast_1d, asarray,
+                   pi, absolute, sqrt, tan, log10,
                    arcsinh, sin, exp, cosh, arccosh, ceil, conjugate,
-                   zeros, sinh, append, concatenate, prod, ones, full, array,
-                   mintypecode)
+                   sinh, concatenate, prod, array)
 from numpy.polynomial.polynomial import polyval as npp_polyval
-from numpy.polynomial.polynomial import polyvalfromroots
 
 from scipy import special, optimize, fft as sp_fft
 from scipy.special import comb
 from scipy._lib._util import float_factorial
 from scipy.signal._arraytools import _validate_fs
+from scipy.signal import _polyutils as _pu
+
+import scipy._lib.array_api_extra as xpx
+from scipy._lib._array_api import (
+    array_namespace, xp_promote, xp_size, xp_default_dtype, is_jax, xp_float_to_complex,
+)
+from scipy._lib.array_api_compat import numpy as np_compat
 
 
 __all__ = ['findfreqs', 'freqs', 'freqz', 'tf2zpk', 'zpk2tf', 'normalize',
@@ -55,6 +60,31 @@ def _is_int_type(x):
         return True
 
 
+# https://github.com/numpy/numpy/blob/v2.2.0/numpy/_core/function_base.py#L195-L302
+def _logspace(start, stop, num=50, endpoint=True, base=10.0, dtype=None, *, xp):
+    if not isinstance(base, float | int) and xp.asarray(base).ndim > 0:
+        # If base is non-scalar, broadcast it with the others, since it
+        # may influence how axis is interpreted.
+        start, stop, base = map(xp.asarray, (start, stop, base))
+        ndmax = xp.broadcast_arrays(start, stop, base).ndim
+        start, stop, base = (
+            xpx.atleast_nd(a, ndim=ndmax)
+            for a in (start, stop, base)
+        )
+        base = xp.expand_dims(base)
+    try:
+        result_dt = xp.result_type(start, stop, base)
+    except ValueError:
+        # all of start, stop and base are python scalars
+        result_dt = xp_default_dtype(xp)
+    y = xp.linspace(start, stop, num=num, endpoint=endpoint, dtype=result_dt)
+
+    yp = xp.pow(base, y)
+    if dtype is None:
+        return yp
+    return xp.astype(yp, dtype, copy=False)
+
+
 def findfreqs(num, den, N, kind='ba'):
     """
     Find array of frequencies for computing the response of an analog filter.
@@ -90,27 +120,42 @@ def findfreqs(num, den, N, kind='ba'):
              3.16227766e-01,   1.00000000e+00,   3.16227766e+00,
              1.00000000e+01,   3.16227766e+01,   1.00000000e+02])
     """
+    xp = array_namespace(num, den)
+    num, den = map(xp.asarray, (num, den))
+
     if kind == 'ba':
-        ep = atleast_1d(roots(den)) + 0j
-        tz = atleast_1d(roots(num)) + 0j
+        ep = xpx.atleast_nd(_pu.polyroots(den, xp=xp), ndim=1, xp=xp)
+        tz = xpx.atleast_nd(_pu.polyroots(num, xp=xp), ndim=1, xp=xp)
     elif kind == 'zp':
-        ep = atleast_1d(den) + 0j
-        tz = atleast_1d(num) + 0j
+        ep = xpx.atleast_nd(den, ndim=1, xp=xp)
+        tz = xpx.atleast_nd(num, ndim=1, xp=xp)
     else:
         raise ValueError("input must be one of {'ba', 'zp'}")
 
-    if len(ep) == 0:
-        ep = atleast_1d(-1000) + 0j
+    ep = xp_float_to_complex(ep, xp=xp)
+    tz = xp_float_to_complex(tz, xp=xp)
 
-    ez = np.r_[ep[ep.imag >= 0], tz[(np.abs(tz) < 1e5) & (tz.imag >= 0)]]
+    if ep.shape[0] == 0:
+        ep = xp.asarray([-1000], dtype=ep.dtype)
 
-    integ = np.abs(ez) < 1e-10
-    hfreq = np.round(np.log10(np.max(3 * np.abs(ez.real + integ) +
-                                     1.5 * ez.imag)) + 0.5)
-    lfreq = np.round(np.log10(0.1 * np.min(np.abs((ez + integ).real) +
-                                           2 * ez.imag)) - 0.5)
+    ez = xp.concat((
+        ep[xp.imag(ep) >= 0],
+        tz[(xp.abs(tz) < 1e5) & (xp.imag(tz) >= 0)]
+    ))
 
-    w = np.logspace(lfreq, hfreq, N)
+    integ = xp.astype(xp.abs(ez) < 1e-10, ez.dtype) # XXX True->1, False->0
+    hfreq = xp.round(
+        xp.log10(xp.max(3*xp.abs(xp.real(ez) + integ) + 1.5*xp.imag(ez))) + 0.5
+    )
+
+    # the fudge factor is for backwards compatibility: round(-1.5) can be -1 or -2
+    # depending on the the floating-point jitter in -1.5
+    fudge = 1e-14 if is_jax(xp) else 0
+    lfreq = xp.round(
+        xp.log10(0.1*xp.min(xp.abs(xp.real(ez + integ)) + 2*xp.imag(ez))) - 0.5 - fudge
+    )
+
+    w = _logspace(lfreq, hfreq, N, xp=xp)
     return w
 
 
@@ -175,16 +220,18 @@ def freqs(b, a, worN=200, plot=None):
     >>> plt.show()
 
     """
+    xp = array_namespace(b, a, worN)
+
     if worN is None:
         # For backwards compatibility
         w = findfreqs(b, a, 200)
     elif _is_int_type(worN):
         w = findfreqs(b, a, worN)
     else:
-        w = atleast_1d(worN)
+        w = xpx.atleast_nd(xp.asarray(worN), ndim=1, xp=xp)
 
     s = 1j * w
-    h = polyval(b, s) / polyval(a, s)
+    h = _pu.polyval(b, s, xp=xp) / _pu.polyval(a, s, xp=xp)
     if plot is not None:
         plot(w, h)
 
@@ -251,8 +298,13 @@ def freqs_zpk(z, p, k, worN=200):
     >>> plt.show()
 
     """
-    k = np.asarray(k)
-    if k.size > 1:
+    xp = array_namespace(z, p)
+    z, p = map(xp.asarray, (z, p))
+
+    # NB: k is documented to be a scalar; for backwards compat we keep allowing it
+    # to be a size-1 array, but it does not influence the namespace calculation.
+    k = xp.asarray(k, dtype=xp_default_dtype(xp))
+    if xp_size(k) > 1:
         raise ValueError('k must be a single scalar gain')
 
     if worN is None:
@@ -263,10 +315,10 @@ def freqs_zpk(z, p, k, worN=200):
     else:
         w = worN
 
-    w = atleast_1d(w)
+    w = xpx.atleast_nd(xp.asarray(w), ndim=1, xp=xp)
     s = 1j * w
-    num = polyvalfromroots(s, z)
-    den = polyvalfromroots(s, p)
+    num = _pu.npp_polyvalfromroots(s, z, xp=xp)
+    den = _pu.npp_polyvalfromroots(s, p, xp=xp)
     h = k * num/den
     return w, h
 
@@ -430,8 +482,15 @@ def freqz(b, a=1, worN=512, whole=False, plot=None, fs=2*pi,
     (2, 1024)
 
     """
-    b = atleast_1d(b)
-    a = atleast_1d(a)
+    xp = array_namespace(b, a)
+
+    b, a = map(xp.asarray, (b, a))
+    if xp.isdtype(a.dtype, 'integral'):
+        a = xp.astype(a, xp_default_dtype(xp))
+    res_dtype = xp.result_type(b, a)
+
+    b = xpx.atleast_nd(b, ndim=1, xp=xp)
+    a = xpx.atleast_nd(a, ndim=1, xp=xp)
 
     fs = _validate_fs(fs, allow_none=False)
 
@@ -449,40 +508,51 @@ def freqz(b, a=1, worN=512, whole=False, plot=None, fs=2*pi,
         lastpoint = 2 * pi if whole else pi
         # if include_nyquist is true and whole is false, w should
         # include end point
-        w = np.linspace(0, lastpoint, N,
-                        endpoint=include_nyquist and not whole)
+        w = xp.linspace(0, lastpoint, N,
+                        endpoint=include_nyquist and not whole, dtype=res_dtype)
         n_fft = N if whole else 2 * (N - 1) if include_nyquist else 2 * N
-        if (a.size == 1 and (b.ndim == 1 or (b.shape[-1] == 1))
+        if (xp_size(a) == 1 and (b.ndim == 1 or (b.shape[-1] == 1))
                 and n_fft >= b.shape[0]
                 and n_fft > 0):  # TODO: review threshold acc. to benchmark?
-            if np.isrealobj(b) and np.isrealobj(a):
+ 
+            if (xp.isdtype(b.dtype, "real floating") and
+                xp.isdtype(a.dtype, "real floating")
+            ):
                 fft_func = sp_fft.rfft
             else:
                 fft_func = sp_fft.fft
-            h = fft_func(b, n=n_fft, axis=0)[:N]
+
+            h = fft_func(b, n=n_fft, axis=0)
+            h = h[:min(N, h.shape[0]), ...]
             h /= a
+
             if fft_func is sp_fft.rfft and whole:
                 # exclude DC and maybe Nyquist (no need to use axis_reverse
                 # here because we can build reversal with the truncation)
-                stop = -1 if n_fft % 2 == 1 else -2
-                h_flip = slice(stop, 0, -1)
-                h = np.concatenate((h, h[h_flip].conj()))
+                stop = None if n_fft % 2 == 1 else -1
+                h_flipped = xp.flip(h[1:stop, ...], axis=0)
+                h = xp.concat((h, xp.conj(h_flipped)))
             if b.ndim > 1:
                 # Last axis of h has length 1, so drop it.
                 h = h[..., 0]
                 # Move the first axis of h to the end.
-                h = np.moveaxis(h, 0, -1)
+                h = xp.moveaxis(h, 0, -1)
     else:
-        w = atleast_1d(worN)
+        if isinstance(worN, complex):
+            # backwards compat
+            worN = worN.real
+        w = xpx.atleast_nd(xp.asarray(worN, dtype=res_dtype), ndim=1, xp=xp)
+        if xp.isdtype(w.dtype, 'integral'):
+            w = xp.astype(w, xp_default_dtype(xp))
         del worN
-        w = 2*pi*w/fs
+        w = 2 * pi * w / fs
 
     if h is None:  # still need to compute using freqs w
-        zm1 = exp(-1j * w)
-        h = (npp_polyval(zm1, b, tensor=False) /
-             npp_polyval(zm1, a, tensor=False))
+        zm1 = xp.exp(-1j * w)
+        h = (_pu.npp_polyval(zm1, b, tensor=False, xp=xp) /
+             _pu.npp_polyval(zm1, a, tensor=False, xp=xp))
 
-    w = w*(fs/(2*pi))
+    w = w * (fs / (2 * pi))
 
     if plot is not None:
         plot(w, h)
@@ -573,7 +643,13 @@ def freqz_zpk(z, p, k, worN=512, whole=False, fs=2*pi):
     >>> plt.show()
 
     """
-    z, p = map(atleast_1d, (z, p))
+    xp = array_namespace(z, p)
+    z, p = map(xp.asarray, (z, p))
+
+    z = xpx.atleast_nd(z, ndim=1, xp=xp)
+    p = xpx.atleast_nd(p, ndim=1, xp=xp)
+    res_dtype = xp.result_type(z, p)
+    res_dtype = xp.float64 if res_dtype in (xp.float64, xp.complex128) else xp.float32
 
     fs = _validate_fs(fs, allow_none=False)
 
@@ -584,15 +660,19 @@ def freqz_zpk(z, p, k, worN=512, whole=False, fs=2*pi):
 
     if worN is None:
         # For backwards compatibility
-        w = np.linspace(0, lastpoint, 512, endpoint=False)
+        w = xp.linspace(0, lastpoint, 512, endpoint=False, dtype=res_dtype)
     elif _is_int_type(worN):
-        w = np.linspace(0, lastpoint, worN, endpoint=False)
+        w = xp.linspace(0, lastpoint, worN, endpoint=False, dtype=res_dtype)
     else:
-        w = atleast_1d(worN)
-        w = 2*pi*w/fs
+        w = xp.asarray(worN)
+        if xp.isdtype(w.dtype, 'integral'):
+            w = xp.astype(w, xp_default_dtype(xp))
+        w = xpx.atleast_nd(w, ndim=1, xp=xp)
+        w = 2 * pi * w / fs
 
-    zm1 = exp(1j * w)
-    h = k * polyvalfromroots(zm1, z) / polyvalfromroots(zm1, p)
+    zm1 = xp.exp(1j * w)
+    func = _pu.npp_polyvalfromroots
+    h = xp.asarray(k, dtype=res_dtype) * func(zm1, z, xp=xp) / func(zm1, p, xp=xp)
 
     w = w*(fs/(2*pi))
 
@@ -725,15 +805,19 @@ def group_delay(system, w=512, whole=False, fs=2*pi):
     return w, gd
 
 
-def _validate_sos(sos):
+def _validate_sos(sos, xp=None):
     """Helper to validate a SOS input"""
-    sos = np.atleast_2d(sos)
+    if xp is None:
+        xp = np    # backcompat, cf sosfilt, sosfiltfilt
+
+    sos = xp.asarray(sos)
+    sos = xpx.atleast_nd(sos, ndim=2, xp=xp)
     if sos.ndim != 2:
         raise ValueError('sos array must be 2D')
     n_sections, m = sos.shape
     if m != 6:
         raise ValueError('sos array must be shape (n_sections, 6)')
-    if not (sos[:, 3] == 1).all():
+    if not xp.all(sos[:, 3] == 1):
         raise ValueError('sos[:, 3] should be all ones')
     return sos, n_sections
 
@@ -787,12 +871,14 @@ def freqz_sos(sos, worN=512, whole=False, fs=2*pi):
 
     See Also
     --------
-    freqz, sosfilt
+    freqz, sosfilt, sosfreqz
 
     Notes
     -----
-    .. versionadded:: 0.19.0
-
+    This function used to be called ``sosfreqz`` in older versions (≥ 0.19.0)
+    
+    .. versionadded:: 1.15.0
+        
     Examples
     --------
     Design a 15th-order bandpass filter in SOS format.
@@ -850,13 +936,16 @@ def freqz_sos(sos, worN=512, whole=False, fs=2*pi):
     >>> plt.show()
 
     """
+    xp = array_namespace(sos)
+
     fs = _validate_fs(fs, allow_none=False)
 
-    sos, n_sections = _validate_sos(sos)
+    sos, n_sections = _validate_sos(sos, xp)
     if n_sections == 0:
         raise ValueError('Cannot compute frequencies with no sections')
     h = 1.
-    for row in sos:
+    for j in range(sos.shape[0]):
+        row = sos[j, :]
         w, rowh = freqz(row[:3], row[3:], worN=worN, whole=whole, fs=fs)
         h *= rowh
     return w, h
@@ -864,11 +953,14 @@ def freqz_sos(sos, worN=512, whole=False, fs=2*pi):
 
 def sosfreqz(*args, **kwargs):
     """
-    Compute the frequency response of a digital filter in SOS format.
-
-    .. warning:: This function is an alias, provided for backward
-                 compatibility. New code should use the function
-                 :func:`scipy.signal.freqz_sos`.
+    Compute the frequency response of a digital filter in SOS format (legacy).
+
+   .. legacy:: function
+        
+        This function is an alias, provided for backward compatibility. 
+        New code should use the function :func:`scipy.signal.freqz_sos`.
+        This function became obsolete from version 1.15.0.
+        
     """
     return freqz_sos(*args, **kwargs)
 
@@ -1122,13 +1214,15 @@ def tf2zpk(b, a):
         array([ -2.5+2.59807621j ,  -2.5-2.59807621j]),
         3.0)
     """
+    xp = array_namespace(b, a)
     b, a = normalize(b, a)
-    b = (b + 0.0) / a[0]
-    a = (a + 0.0) / a[0]
+
+    a, b = xp_promote(a, b, xp=xp, force_floating=True)
+
     k = b[0]
-    b /= b[0]
-    z = roots(b)
-    p = roots(a)
+    b = b / b[0]
+    z = _pu.polyroots(b, xp=xp)
+    p = _pu.polyroots(a, xp=xp)
     return z, p, k
 
 
@@ -1170,38 +1264,26 @@ def zpk2tf(z, p, k):
     >>> zpk2tf(z, p, k)
     (   array([  5., -40.,  60.]), array([ 1., -9.,  8.]))
     """
-    z = atleast_1d(z)
-    k = atleast_1d(k)
-    if len(z.shape) > 1:
-        temp = poly(z[0])
-        b = np.empty((z.shape[0], z.shape[1] + 1), temp.dtype.char)
-        if len(k) == 1:
+    xp = array_namespace(z, p)
+    z, p, k = map(xp.asarray, (z, p, k))
+
+    z = xpx.atleast_nd(z, ndim=1, xp=xp)
+    k = xpx.atleast_nd(k, ndim=1, xp=xp)
+    if xp.isdtype(k.dtype, 'integral'):
+        k = xp.astype(k, xp_default_dtype(xp))
+
+    if z.ndim > 1:
+        temp = _pu.poly(z[0], xp=xp)
+        b = xp.empty((z.shape[0], z.shape[1] + 1), dtype=temp.dtype)
+        if k.shape[0] == 1:
             k = [k[0]] * z.shape[0]
         for i in range(z.shape[0]):
-            b[i] = k[i] * poly(z[i])
+            b[i] = k[i] * _pu.poly(z[i], xp=xp)
     else:
-        b = k * poly(z)
-    a = atleast_1d(poly(p))
-
-    # Use real output if possible. Copied from np.poly, since
-    # we can't depend on a specific version of numpy.
-    if issubclass(b.dtype.type, np.complexfloating):
-        # if complex roots are all complex conjugates, the roots are real.
-        roots = np.asarray(z, complex)
-        pos_roots = np.compress(roots.imag > 0, roots)
-        neg_roots = np.conjugate(np.compress(roots.imag < 0, roots))
-        if len(pos_roots) == len(neg_roots):
-            if np.all(np.sort_complex(neg_roots) == np.sort_complex(pos_roots)):
-                b = b.real.copy()
-
-    if issubclass(a.dtype.type, np.complexfloating):
-        # if complex roots are all complex conjugates, the roots are real.
-        roots = np.asarray(p, complex)
-        pos_roots = np.compress(roots.imag > 0, roots)
-        neg_roots = np.conjugate(np.compress(roots.imag < 0, roots))
-        if len(pos_roots) == len(neg_roots):
-            if np.all(np.sort_complex(neg_roots) == np.sort_complex(pos_roots)):
-                a = a.real.copy()
+        b = k * _pu.poly(z, xp=xp)
+
+    a = _pu.poly(p, xp=xp)
+    a = xpx.atleast_nd(xp.asarray(a), ndim=1, xp=xp)
 
     return b, a
 
@@ -1297,17 +1379,20 @@ def sos2tf(sos):
     (   array([0.91256522, 0.91256522, 0.        ]),
         array([1.        , 0.82513043, 0.        ]))
     """
-    sos = np.asarray(sos)
+    xp = array_namespace(sos)
+    sos = xp.asarray(sos)
+
     result_type = sos.dtype
-    if result_type.kind in 'bui':
-        result_type = np.float64
+    if xp.isdtype(result_type, 'integral'):
+        result_type = xp_default_dtype(xp)
+
+    b = xp.asarray([1], dtype=result_type)
+    a = xp.asarray([1], dtype=result_type)
 
-    b = np.array([1], dtype=result_type)
-    a = np.array([1], dtype=result_type)
     n_sections = sos.shape[0]
     for section in range(n_sections):
-        b = np.polymul(b, sos[section, :3])
-        a = np.polymul(a, sos[section, 3:])
+        b = _pu.polymul(b, sos[section, :3], xp=xp)
+        a = _pu.polymul(a, sos[section, 3:], xp=xp)
     return b, a
 
 
@@ -1338,15 +1423,17 @@ def sos2zpk(sos):
 
     .. versionadded:: 0.16.0
     """
-    sos = np.asarray(sos)
+    xp = array_namespace(sos)
+    sos = xp.asarray(sos)
+
     n_sections = sos.shape[0]
-    z = np.zeros(n_sections*2, np.complex128)
-    p = np.zeros(n_sections*2, np.complex128)
+    z = xp.zeros(n_sections*2, dtype=xp.complex128)
+    p = xp.zeros(n_sections*2, dtype=xp.complex128)
     k = 1.
     for section in range(n_sections):
         zpk = tf2zpk(sos[section, :3], sos[section, 3:])
-        z[2*section:2*section+len(zpk[0])] = zpk[0]
-        p[2*section:2*section+len(zpk[1])] = zpk[1]
+        z = xpx.at(z, slice(2*section, 2*section + zpk[0].shape[0])).set(zpk[0])
+        p = xpx.at(p, slice(2*section, 2*section + zpk[1].shape[0])).set(zpk[1])
         k *= zpk[2]
     return z, p, k
 
@@ -1561,6 +1648,12 @@ def zpk2sos(z, p, k, pairing=None, *, analog=False):
     # 4. Further optimizations of the section ordering / pole-zero pairing.
     # See the wiki for other potential issues.
 
+    xp = array_namespace(z, p)
+
+    # convert to numpy, convert back on exit   XXX
+    z, p = map(np.asarray, (z, p))
+    k = np.asarray(k)
+
     if pairing is None:
         pairing = 'minimal' if analog else 'nearest'
 
@@ -1574,9 +1667,9 @@ def zpk2sos(z, p, k, pairing=None, *, analog=False):
 
     if len(z) == len(p) == 0:
         if not analog:
-            return np.array([[k, 0., 0., 1., 0., 0.]])
+            return xp.asarray(np.asarray([[k, 0., 0., 1., 0., 0.]]))
         else:
-            return np.array([[0., 0., k, 0., 0., 1.]])
+            return xp.asarray(np.asarray([[0., 0., k, 0., 0., 1.]]))
 
     if pairing != 'minimal':
         # ensure we have the same number of poles and zeros, and make copies
@@ -1687,10 +1780,10 @@ def zpk2sos(z, p, k, pairing=None, *, analog=False):
 
     # put gain in first sos
     sos[0][:3] *= k
-    return sos
+    return xp.asarray(sos)
 
 
-def _align_nums(nums):
+def _align_nums(nums, xp):
     """Aligns the shapes of multiple numerators.
 
     Given an array of numerator coefficient arrays [[a_1, a_2,...,
@@ -1715,19 +1808,19 @@ def _align_nums(nums):
         # The statement can throw a ValueError if one
         # of the numerators is a single digit and another
         # is array-like e.g. if nums = [5, [1, 2, 3]]
-        nums = asarray(nums)
+        nums = xp.asarray(nums)
 
-        if not np.issubdtype(nums.dtype, np.number):
+        if not xp.isdtype(nums.dtype, "numeric"):
             raise ValueError("dtype of numerator is non-numeric")
 
         return nums
 
     except ValueError:
-        nums = [np.atleast_1d(num) for num in nums]
-        max_width = max(num.size for num in nums)
+        nums = [xpx.atleast_nd(xp.asarray(num), ndim=1) for num in nums]
+        max_width = max(xp_size(num) for num in nums)
 
         # pre-allocate
-        aligned_nums = np.zeros((len(nums), max_width))
+        aligned_nums = xp.zeros((len(nums), max_width))
 
         # Create numerators with padded zeros
         for index, num in enumerate(nums):
@@ -1792,29 +1885,37 @@ def normalize(b, a):
     Badly conditioned filter coefficients (numerator): the results may be meaningless
 
     """
-    num, den = b, a
+    try:
+        xp = array_namespace(b, a)
+    except TypeError:
+        # object arrays, test_ltisys.py::TestSS2TF::test_simo_round_trip
+        xp = np_compat
+
+    den = xp.asarray(a)
+    den = xpx.atleast_nd(den, ndim=1, xp=xp)
 
-    den = np.atleast_1d(den)
-    num = np.atleast_2d(_align_nums(num))
+    num = xp.asarray(b)
+    num = xpx.atleast_nd(_align_nums(num, xp), ndim=2, xp=xp)
 
     if den.ndim != 1:
         raise ValueError("Denominator polynomial must be rank-1 array.")
     if num.ndim > 2:
         raise ValueError("Numerator polynomial must be rank-1 or"
                          " rank-2 array.")
-    if np.all(den == 0):
+    if xp.all(den == 0):
         raise ValueError("Denominator must have at least on nonzero element.")
 
     # Trim leading zeros in denominator, leave at least one.
-    den = np.trim_zeros(den, 'f')
+    den = _pu._trim_zeros(den, 'f')
 
     # Normalize transfer function
     num, den = num / den[0], den / den[0]
 
     # Count numerator columns that are all zero
     leading_zeros = 0
-    for col in num.T:
-        if np.allclose(col, 0, atol=1e-14):
+    for j in range(num.shape[-1]):
+        col = num[:, j]
+        if xp.all(xp.abs(col) <= 1e-14):
             leading_zeros += 1
         else:
             break
@@ -1892,15 +1993,20 @@ def lp2lp(b, a, wo=1.0):
     >>> plt.legend()
 
     """
-    a, b = map(atleast_1d, (a, b))
+    xp = array_namespace(a, b)
+    a, b = map(xp.asarray, (a, b))
+    a, b = xp_promote(a, b, force_floating=True, xp=xp)
+    a = xpx.atleast_nd(a, ndim=1, xp=xp)
+    b = xpx.atleast_nd(b, ndim=1, xp=xp)
+
     try:
         wo = float(wo)
     except TypeError:
         wo = float(wo[0])
-    d = len(a)
-    n = len(b)
+    d = a.shape[0]
+    n = b.shape[0]
     M = max((d, n))
-    pwo = pow(wo, np.arange(M - 1, -1, -1))
+    pwo = wo ** xp.arange(M - 1, -1, -1, dtype=xp.float64)
     start1 = max((n - d, 0))
     start2 = max((d - n, 0))
     b = b * pwo[start1] / pwo[start2:]
@@ -1908,6 +2014,28 @@ def lp2lp(b, a, wo=1.0):
     return normalize(b, a)
 
 
+def _resize(a, new_shape, xp):
+    # https://github.com/numpy/numpy/blob/v2.2.4/numpy/_core/fromnumeric.py#L1535
+    a = xp.reshape(a, (-1,))
+
+    new_size = 1
+    for dim_length in new_shape:
+        new_size *= dim_length
+        if dim_length < 0:
+            raise ValueError(
+                'all elements of `new_shape` must be non-negative'
+            )
+
+    if xp_size(a) == 0 or new_size == 0:
+        # First case must zero fill. The second would have repeats == 0.
+        return xp.zeros_like(a, shape=new_shape)
+
+    repeats = -(-new_size // xp_size(a))  # ceil division
+    a = xp.concat((a,) * repeats)[:new_size]
+
+    return xp.reshape(a, new_shape)
+
+
 def lp2hp(b, a, wo=1.0):
     r"""
     Transform a lowpass filter prototype to a highpass filter.
@@ -1966,27 +2094,33 @@ def lp2hp(b, a, wo=1.0):
     >>> plt.legend()
 
     """
-    a, b = map(atleast_1d, (a, b))
+    xp = array_namespace(a, b)
+
+    a, b = map(xp.asarray, (a, b))
+    a, b = xp_promote(a, b, force_floating=True, xp=xp)
+    a = xpx.atleast_nd(a, ndim=1, xp=xp)
+    b = xpx.atleast_nd(b, ndim=1, xp=xp)
+
     try:
         wo = float(wo)
     except TypeError:
         wo = float(wo[0])
-    d = len(a)
-    n = len(b)
+    d = a.shape[0]
+    n = b.shape[0]
     if wo != 1:
-        pwo = pow(wo, np.arange(max((d, n))))
+        pwo = wo ** xp.arange(max((d, n)), dtype=b.dtype)
     else:
-        pwo = np.ones(max((d, n)), b.dtype.char)
+        pwo = xp.ones(max((d, n)), dtype=b.dtype)
     if d >= n:
-        outa = a[::-1] * pwo
-        outb = resize(b, (d,))
+        outa = xp.flip(a) * pwo
+        outb = _resize(b, (d,), xp=xp)
         outb[n:] = 0.0
-        outb[:n] = b[::-1] * pwo[:n]
+        outb[:n] = xp.flip(b) * pwo[:n]
     else:
-        outb = b[::-1] * pwo
-        outa = resize(a, (n,))
+        outb = xp.flip(b) * pwo
+        outa = _resize(a, (n,), xp=xp)
         outa[d:] = 0.0
-        outa[:d] = a[::-1] * pwo[:d]
+        outa[:d] = xp.flip(a) * pwo[:d]
 
     return normalize(outb, outa)
 
@@ -2051,16 +2185,20 @@ def lp2bp(b, a, wo=1.0, bw=1.0):
     >>> plt.ylabel('Amplitude [dB]')
     >>> plt.legend()
     """
+    xp = array_namespace(a, b)
 
-    a, b = map(atleast_1d, (a, b))
-    D = len(a) - 1
-    N = len(b) - 1
-    artype = mintypecode((a, b))
+    a, b = map(xp.asarray, (a, b))
+    a, b = xp_promote(a, b, force_floating=True, xp=xp)
+    a = xpx.atleast_nd(a, ndim=1, xp=xp)
+    b = xpx.atleast_nd(b, ndim=1, xp=xp)
+
+    D = a.shape[0] - 1
+    N = b.shape[0] - 1
     ma = max([N, D])
     Np = N + ma
     Dp = D + ma
-    bprime = np.empty(Np + 1, artype)
-    aprime = np.empty(Dp + 1, artype)
+    bprime = xp.empty(Np + 1, dtype=b.dtype)
+    aprime = xp.empty(Dp + 1, dtype=a.dtype)
     wosq = wo * wo
     for j in range(Np + 1):
         val = 0.0
@@ -2139,15 +2277,20 @@ def lp2bs(b, a, wo=1.0, bw=1.0):
     >>> plt.ylabel('Amplitude [dB]')
     >>> plt.legend()
     """
-    a, b = map(atleast_1d, (a, b))
-    D = len(a) - 1
-    N = len(b) - 1
-    artype = mintypecode((a, b))
+    xp = array_namespace(a, b)
+
+    a, b = map(xp.asarray, (a, b))
+    a, b = xp_promote(a, b, force_floating=True, xp=xp)
+    a = xpx.atleast_nd(a, ndim=1, xp=xp)
+    b = xpx.atleast_nd(b, ndim=1, xp=xp)
+
+    D = a.shape[0] - 1
+    N = b.shape[0] - 1
     M = max([N, D])
     Np = M + M
     Dp = M + M
-    bprime = np.empty(Np + 1, artype)
-    aprime = np.empty(Dp + 1, artype)
+    bprime = xp.empty(Np + 1, dtype=b.dtype)
+    aprime = xp.empty(Dp + 1, dtype=a.dtype)
     wosq = wo * wo
     for j in range(Np + 1):
         val = 0.0
@@ -2170,88 +2313,141 @@ def lp2bs(b, a, wo=1.0, bw=1.0):
 
 
 def bilinear(b, a, fs=1.0):
-    r"""
-    Return a digital IIR filter from an analog one using a bilinear transform.
-
-    Transform a set of poles and zeros from the analog s-plane to the digital
-    z-plane using Tustin's method, which substitutes ``2*fs*(z-1) / (z+1)`` for
-    ``s``, maintaining the shape of the frequency response.
+    r"""Calculate a digital IIR filter from an analog transfer function by utilizing
+    the bilinear transform.
 
     Parameters
     ----------
     b : array_like
-        Numerator of the analog filter transfer function.
+        Coefficients of the numerator polynomial of the analog transfer function in
+        form of a complex- or real-valued 1d array.
     a : array_like
-        Denominator of the analog filter transfer function.
+        Coefficients of the denominator polynomial of the analog transfer function in
+        form of a complex- or real-valued 1d array.
     fs : float
-        Sample rate, as ordinary frequency (e.g., hertz). No prewarping is
+        Sample rate, as ordinary frequency (e.g., hertz). No pre-warping is
         done in this function.
 
     Returns
     -------
-    b : ndarray
-        Numerator of the transformed digital filter transfer function.
-    a : ndarray
-        Denominator of the transformed digital filter transfer function.
+    beta : ndarray
+        Coefficients of the numerator polynomial of the digital transfer function in
+        form of a complex- or real-valued 1d array.
+    alpha : ndarray
+        Coefficients of the denominator polynomial of the digital transfer function in
+        form of a complex- or real-valued 1d array.
+
+    Notes
+    -----
+    The parameters :math:`b = [b_0, \ldots, b_Q]` and :math:`a = [a_0, \ldots, a_P]`
+    are 1d arrays of length :math:`Q+1` and :math:`P+1`. They define the analog
+    transfer function
+
+    .. math::
+
+        H_a(s) = \frac{b_0 s^Q + b_1 s^{Q-1} + \cdots + b_Q}{
+                       a_0 s^P + a_1 s^{P-1} + \cdots + a_P}\ .
+
+    The bilinear transform [1]_ is applied by substituting
+
+    .. math::
+
+        s = \kappa \frac{z-1}{z+1}\ , \qquad \kappa := 2 f_s\ ,
+
+    into :math:`H_a(s)`, with :math:`f_s`  being the sampling rate.
+    This results in the digital transfer function in the :math:`z`-domain
+
+    .. math::
+
+        H_d(z) = \frac{b_0 \left(\kappa \frac{z-1}{z+1}\right)^Q +
+                       b_1 \left(\kappa \frac{z-1}{z+1}\right)^{Q-1} +
+                       \cdots + b_Q}{
+                       a_0 \left(\kappa \frac{z-1}{z+1}\right)^P +
+                       a_1 \left(\kappa \frac{z-1}{z+1}\right)^{P-1} +
+                       \cdots + a_P}\ .
+
+    This expression can be simplified by multiplying numerator and denominator by
+    :math:`(z+1)^N`, with :math:`N=\max(P, Q)`. This allows :math:`H_d(z)` to be
+    reformulated as
+
+    .. math::
+
+       & & \frac{b_0 \big(\kappa (z-1)\big)^Q     (z+1)^{N-Q} +
+                b_1 \big(\kappa (z-1)\big)^{Q-1} (z+1)^{N-Q+1} +
+                \cdots + b_Q(z+1)^N}{
+                a_0 \big(\kappa (z-1)\big)^P     (z+1)^{N-P} +
+                a_1 \big(\kappa (z-1)\big)^{P-1} (z+1)^{N-P+1} +
+                \cdots + a_P(z+1)^N}\\
+        &=:& \frac{\beta_0 + \beta_1  z^{-1} + \cdots + \beta_N  z^{-N}}{
+                 \alpha_0 + \alpha_1 z^{-1} + \cdots + \alpha_N z^{-N}}\ .
+
+
+    This is the equation implemented to perform the bilinear transform. Note that for
+    large :math:`f_s`, :math:`\kappa^Q` or :math:`\kappa^P` can cause a numeric
+    overflow for sufficiently large :math:`P` or :math:`Q`.
+
+    References
+    ----------
+    .. [1] "Bilinear Transform", Wikipedia,
+           https://en.wikipedia.org/wiki/Bilinear_transform
 
     See Also
     --------
-    lp2lp, lp2hp, lp2bp, lp2bs
-    bilinear_zpk
+    lp2lp, lp2hp, lp2bp, lp2bs, bilinear_zpk
 
     Examples
     --------
+    The following example shows the frequency response of an analog bandpass filter and
+    the corresponding digital filter derived by utilitzing the bilinear transform:
+
     >>> from scipy import signal
     >>> import matplotlib.pyplot as plt
     >>> import numpy as np
+    ...
+    >>> fs = 100  # sampling frequency
+    >>> om_c = 2 * np.pi * np.array([7, 13])  # corner frequencies
+    >>> bb_s, aa_s = signal.butter(4, om_c, btype='bandpass', analog=True, output='ba')
+    >>> bb_z, aa_z = signal.bilinear(bb_s, aa_s, fs)
+    ...
+    >>> w_z, H_z = signal.freqz(bb_z, aa_z)  # frequency response of digitial filter
+    >>> w_s, H_s = signal.freqs(bb_s, aa_s, worN=w_z*fs)  # analog filter response
+    ...
+    >>> f_z, f_s = w_z * fs / (2*np.pi), w_s / (2*np.pi)
+    >>> Hz_dB, Hs_dB = (20*np.log10(np.abs(H_).clip(1e-10)) for H_ in (H_z, H_s))
+    >>> fg0, ax0 = plt.subplots()
+    >>> ax0.set_title("Frequency Response of 4-th order Bandpass Filter")
+    >>> ax0.set(xlabel='Frequency $f$ in Hertz', ylabel='Magnitude in dB',
+    ...         xlim=[f_z[1], fs/2], ylim=[-200, 2])
+    >>> ax0.semilogx(f_z, Hz_dB, alpha=.5, label=r'$|H_z(e^{j 2 \pi f})|$')
+    >>> ax0.semilogx(f_s, Hs_dB, alpha=.5, label=r'$|H_s(j 2 \pi f)|$')
+    >>> ax0.legend()
+    >>> ax0.grid(which='both', axis='x')
+    >>> ax0.grid(which='major', axis='y')
+    >>> plt.show()
 
-    >>> fs = 100
-    >>> bf = 2 * np.pi * np.array([7, 13])
-    >>> filts = signal.lti(*signal.butter(4, bf, btype='bandpass',
-    ...                                   analog=True))
-    >>> filtz = signal.lti(*signal.bilinear(filts.num, filts.den, fs))
-    >>> wz, hz = signal.freqz(filtz.num, filtz.den)
-    >>> ws, hs = signal.freqs(filts.num, filts.den, worN=fs*wz)
-
-    >>> plt.semilogx(wz*fs/(2*np.pi), 20*np.log10(np.abs(hz).clip(1e-15)),
-    ...              label=r'$|H_z(e^{j \omega})|$')
-    >>> plt.semilogx(wz*fs/(2*np.pi), 20*np.log10(np.abs(hs).clip(1e-15)),
-    ...              label=r'$|H(j \omega)|$')
-    >>> plt.legend()
-    >>> plt.xlabel('Frequency [Hz]')
-    >>> plt.ylabel('Amplitude [dB]')
-    >>> plt.grid(True)
+    The difference in the higher frequencies shown in the plot is caused by an effect
+    called "frequency warping". [1]_ describes a method called "pre-warping" to
+    reduce those deviations.
     """
+    b, a = np.atleast_1d(b), np.atleast_1d(a)  # convert scalars, if needed
+    if not a.ndim == 1:
+        raise ValueError(f"Parameter a is not a 1d array since {a.shape=}")
+    if not b.ndim == 1:
+        raise ValueError(f"Parameter b is not a 1d array since {b.shape=}")
+    b, a = np.trim_zeros(b, 'f'), np.trim_zeros(a, 'f')  # remove leading zeros
     fs = _validate_fs(fs, allow_none=False)
-    a, b = map(atleast_1d, (a, b))
-    D = len(a) - 1
-    N = len(b) - 1
-    artype = float
-    M = max([N, D])
-    Np = M
-    Dp = M
-    bprime = np.empty(Np + 1, artype)
-    aprime = np.empty(Dp + 1, artype)
-    for j in range(Np + 1):
-        val = 0.0
-        for i in range(N + 1):
-            for k in range(i + 1):
-                for l in range(M - i + 1):
-                    if k + l == j:
-                        val += (comb(i, k) * comb(M - i, l) * b[N - i] *
-                                pow(2 * fs, i) * (-1) ** k)
-        bprime[j] = real(val)
-    for j in range(Dp + 1):
-        val = 0.0
-        for i in range(D + 1):
-            for k in range(i + 1):
-                for l in range(M - i + 1):
-                    if k + l == j:
-                        val += (comb(i, k) * comb(M - i, l) * a[D - i] *
-                                pow(2 * fs, i) * (-1) ** k)
-        aprime[j] = real(val)
 
-    return normalize(bprime, aprime)
+    # Splitting the factor fs*2 between numerator and denominator reduces the chance of
+    # numeric overflow for large fs and large N:
+    fac = np.sqrt(fs*2)
+    zp1 = np.polynomial.Polynomial((+1, 1)) / fac  # Polynomial (z + 1) / fac
+    zm1 = np.polynomial.Polynomial((-1, 1)) * fac  # Polynomial (z - 1) * fac
+    # Note that NumPy's Polynomial coefficient order is backward compared to a and b.
+
+    N = max(len(a), len(b)) - 1
+    numerator   = sum(b_ * zp1**(N-q) * zm1**q for q, b_ in enumerate(b[::-1]))
+    denominator = sum(a_ * zp1**(N-p) * zm1**p for p, a_ in enumerate(a[::-1]))
+    return normalize(numerator.coef[::-1], denominator.coef[::-1])
 
 
 def _validate_gpass_gstop(gpass, gstop):
@@ -2685,7 +2881,7 @@ def _relative_degree(z, p):
     """
     Return relative degree of transfer function from zeros and poles
     """
-    degree = len(p) - len(z)
+    degree = p.shape[0] - z.shape[0]
     if degree < 0:
         raise ValueError("Improper transfer function. "
                          "Must have at least as many poles as zeros.")
@@ -2755,8 +2951,11 @@ def bilinear_zpk(z, p, k, fs):
     >>> plt.ylabel('Amplitude [dB]')
     >>> plt.grid(True)
     """
-    z = atleast_1d(z)
-    p = atleast_1d(p)
+    xp = array_namespace(z, p)
+
+    z, p = map(xp.asarray, (z, p))
+    z = xpx.atleast_nd(z, ndim=1, xp=xp)
+    p = xpx.atleast_nd(p, ndim=1, xp=xp)
 
     fs = _validate_fs(fs, allow_none=False)
 
@@ -2769,10 +2968,10 @@ def bilinear_zpk(z, p, k, fs):
     p_z = (fs2 + p) / (fs2 - p)
 
     # Any zeros that were at infinity get moved to the Nyquist frequency
-    z_z = append(z_z, -ones(degree))
+    z_z = xp.concat((z_z, -xp.ones(degree)))
 
     # Compensate for gain change
-    k_z = k * real(prod(fs2 - z) / prod(fs2 - p))
+    k_z = k * xp.real(xp.prod(fs2 - z) / xp.prod(fs2 - p))
 
     return z_z, p_z, k_z
 
@@ -2832,8 +3031,12 @@ def lp2lp_zpk(z, p, k, wo=1.0):
     >>> lp2lp_zpk(z, p, k, wo)
     (   array([2.8, 0.8]), array([2. , 5.2]), 0.8)
     """
-    z = atleast_1d(z)
-    p = atleast_1d(p)
+    xp = array_namespace(z, p)
+
+    z, p = map(xp.asarray, (z, p))
+    z = xpx.atleast_nd(z, ndim=1, xp=xp)
+    p = xpx.atleast_nd(p, ndim=1, xp=xp)
+
     wo = float(wo)  # Avoid int wraparound
 
     degree = _relative_degree(z, p)
@@ -2909,8 +3112,13 @@ def lp2hp_zpk(z, p, k, wo=1.0):
         array([-0.6 , -0.15]),
         8.5)
     """
-    z = atleast_1d(z)
-    p = atleast_1d(p)
+    xp = array_namespace(z, p)
+
+    z, p = map(xp.asarray, (z, p))
+    # XXX: no xp_promote here since that breaks TestButter
+    z = xpx.atleast_nd(z, ndim=1, xp=xp)
+    p = xpx.atleast_nd(p, ndim=1, xp=xp)
+
     wo = float(wo)
 
     degree = _relative_degree(z, p)
@@ -2921,10 +3129,10 @@ def lp2hp_zpk(z, p, k, wo=1.0):
     p_hp = wo / p
 
     # If lowpass had zeros at infinity, inverting moves them to origin.
-    z_hp = append(z_hp, zeros(degree))
+    z_hp = xp.concat((z_hp, xp.zeros(degree)))
 
     # Cancel out gain change caused by inversion
-    k_hp = k * real(prod(-z) / prod(-p))
+    k_hp = k * xp.real(xp.prod(-z) / xp.prod(-p))
 
     return z_hp, p_hp, k_hp
 
@@ -2995,8 +3203,13 @@ def lp2bp_zpk(z, p, k, wo=1.0, bw=1.0):
         array([1.04996339e+02+0.j, -1.60167736e-03+0.j,  3.66108003e-03+0.j,
                -2.39998398e+02+0.j]), 0.8)
     """
-    z = atleast_1d(z)
-    p = atleast_1d(p)
+    xp = array_namespace(z, p)
+
+    z, p = map(xp.asarray, (z, p))
+    z, p = xp_promote(z, p, force_floating=True, xp=xp)
+    z = xpx.atleast_nd(z, ndim=1, xp=xp)
+    p = xpx.atleast_nd(p, ndim=1, xp=xp)
+
     wo = float(wo)
     bw = float(bw)
 
@@ -3007,17 +3220,17 @@ def lp2bp_zpk(z, p, k, wo=1.0, bw=1.0):
     p_lp = p * bw/2
 
     # Square root needs to produce complex result, not NaN
-    z_lp = z_lp.astype(complex)
-    p_lp = p_lp.astype(complex)
+    z_lp = xp.astype(z_lp, xp.complex128)
+    p_lp = xp.astype(p_lp, xp.complex128)
 
     # Duplicate poles and zeros and shift from baseband to +wo and -wo
-    z_bp = concatenate((z_lp + sqrt(z_lp**2 - wo**2),
-                        z_lp - sqrt(z_lp**2 - wo**2)))
-    p_bp = concatenate((p_lp + sqrt(p_lp**2 - wo**2),
-                        p_lp - sqrt(p_lp**2 - wo**2)))
+    z_bp = xp.concat((z_lp + xp.sqrt(z_lp**2 - wo**2),
+                      z_lp - xp.sqrt(z_lp**2 - wo**2)))
+    p_bp = xp.concat((p_lp + xp.sqrt(p_lp**2 - wo**2),
+                      p_lp - xp.sqrt(p_lp**2 - wo**2)))
 
     # Move degree zeros to origin, leaving degree zeros at infinity for BPF
-    z_bp = append(z_bp, zeros(degree))
+    z_bp = xp.concat((z_bp, xp.zeros(degree)))
 
     # Cancel out gain change from frequency scaling
     k_bp = k * bw**degree
@@ -3090,8 +3303,13 @@ def lp2bs_zpk(z, p, k, wo=1.0, bw=1.0):
         array([14.2681928 +0.j, -0.02506281+0.j,  0.01752149+0.j, -9.97493719+0.j]),
         -12.857142857142858)
     """
-    z = atleast_1d(z)
-    p = atleast_1d(p)
+    xp = array_namespace(z, p)
+
+    z, p = map(xp.asarray, (z, p))
+    z, p = xp_promote(z, p, force_floating=True, xp=xp)
+    z = xpx.atleast_nd(z, ndim=1, xp=xp)
+    p = xpx.atleast_nd(p, ndim=1, xp=xp)
+
     wo = float(wo)
     bw = float(bw)
 
@@ -3102,21 +3320,21 @@ def lp2bs_zpk(z, p, k, wo=1.0, bw=1.0):
     p_hp = (bw/2) / p
 
     # Square root needs to produce complex result, not NaN
-    z_hp = z_hp.astype(complex)
-    p_hp = p_hp.astype(complex)
+    z_hp = xp.astype(z_hp, xp.complex128)
+    p_hp = xp.astype(p_hp, xp.complex128)
 
     # Duplicate poles and zeros and shift from baseband to +wo and -wo
-    z_bs = concatenate((z_hp + sqrt(z_hp**2 - wo**2),
-                        z_hp - sqrt(z_hp**2 - wo**2)))
-    p_bs = concatenate((p_hp + sqrt(p_hp**2 - wo**2),
-                        p_hp - sqrt(p_hp**2 - wo**2)))
+    z_bs = xp.concat((z_hp + xp.sqrt(z_hp**2 - wo**2),
+                      z_hp - xp.sqrt(z_hp**2 - wo**2)))
+    p_bs = xp.concat((p_hp + xp.sqrt(p_hp**2 - wo**2),
+                      p_hp - xp.sqrt(p_hp**2 - wo**2)))
 
     # Move any zeros that were at infinity to the center of the stopband
-    z_bs = append(z_bs, full(degree, +1j*wo))
-    z_bs = append(z_bs, full(degree, -1j*wo))
+    z_bs = xp.concat((z_bs, xp.full(degree, +1j*wo)))
+    z_bs = xp.concat((z_bs, xp.full(degree, -1j*wo)))
 
     # Cancel out gain change caused by inversion
-    k_bs = k * real(prod(-z) / prod(-p))
+    k_bs = k * xp.real(xp.prod(-z) / xp.prod(-p))
 
     return z_bs, p_bs, k_bs