scipy 1.16.2__cp313-cp313t-win_arm64.whl

scipy/sparse/linalg/tests/test_norm.py

@@ -0,0 +1,154 @@
+"""Test functions for the sparse.linalg.norm module
+"""
+
+import pytest
+import numpy as np
+from numpy.linalg import norm as npnorm
+from numpy.testing import assert_allclose, assert_equal
+from pytest import raises as assert_raises
+
+import scipy.sparse
+from scipy.sparse.linalg import norm as spnorm
+
+
+# https://github.com/scipy/scipy/issues/16031
+# https://github.com/scipy/scipy/issues/21690
+def test_sparray_norm():
+    row = np.array([0, 0, 1, 1])
+    col = np.array([0, 1, 2, 3])
+    data = np.array([4, 5, 7, 9])
+    test_arr = scipy.sparse.coo_array((data, (row, col)), shape=(2, 4))
+    test_mat = scipy.sparse.coo_matrix((data, (row, col)), shape=(2, 4))
+    for ord in (1, np.inf, None):
+        for ax in [0, 1, None, (0, 1), (1, 0)]:
+            for A in (test_arr, test_mat):
+                expected = npnorm(A.toarray(), ord=ord, axis=ax)
+                actual = spnorm(A, ord=ord, axis=ax)
+                assert hasattr(actual, "dtype")
+                assert_equal(actual, expected)
+    # test 1d array and 1d-like (column) matrix
+    test_arr_1d = scipy.sparse.coo_array((data, (col,)), shape=(4,))
+    test_mat_col = scipy.sparse.coo_matrix((data, (col, [0, 0, 0, 0])), shape=(4, 1))
+    for ord in (1, np.inf, None):
+        for ax in [0, None]:
+            for A in (test_arr_1d, test_mat_col):
+                expected = npnorm(A.toarray(), ord=ord, axis=ax)
+                assert_equal(spnorm(A, ord=ord, axis=ax), expected)
+
+
+class TestNorm:
+    def setup_method(self):
+        a = np.arange(9) - 4
+        b = a.reshape((3, 3))
+        self.b = scipy.sparse.csr_array(b)
+
+    @pytest.mark.thread_unsafe
+    def test_matrix_norm(self):
+
+        # Frobenius norm is the default
+        assert_allclose(spnorm(self.b), 7.745966692414834)
+        assert_allclose(spnorm(self.b, 'fro'), 7.745966692414834)
+
+        assert_allclose(spnorm(self.b, np.inf), 9)
+        assert_allclose(spnorm(self.b, -np.inf), 2)
+        assert_allclose(spnorm(self.b, 1), 7)
+        assert_allclose(spnorm(self.b, -1), 6)
+        # Only floating or complex floating dtype supported by svds.
+        with pytest.warns(UserWarning, match="The problem size"):
+            assert_allclose(spnorm(self.b.astype(np.float64), 2),
+                            7.348469228349534)
+
+        # _multi_svd_norm is not implemented for sparse array
+        assert_raises(NotImplementedError, spnorm, self.b, -2)
+
+    def test_matrix_norm_axis(self):
+        for m, axis in ((self.b, None), (self.b, (0, 1)), (self.b.T, (1, 0))):
+            assert_allclose(spnorm(m, axis=axis), 7.745966692414834)
+            assert_allclose(spnorm(m, 'fro', axis=axis), 7.745966692414834)
+            assert_allclose(spnorm(m, np.inf, axis=axis), 9)
+            assert_allclose(spnorm(m, -np.inf, axis=axis), 2)
+            assert_allclose(spnorm(m, 1, axis=axis), 7)
+            assert_allclose(spnorm(m, -1, axis=axis), 6)
+
+    def test_vector_norm(self):
+        v = [4.5825756949558398, 4.2426406871192848, 4.5825756949558398]
+        for m, a in (self.b, 0), (self.b.T, 1):
+            for axis in a, (a, ), a-2, (a-2, ):
+                assert_allclose(spnorm(m, 1, axis=axis), [7, 6, 7])
+                assert_allclose(spnorm(m, np.inf, axis=axis), [4, 3, 4])
+                assert_allclose(spnorm(m, axis=axis), v)
+                assert_allclose(spnorm(m, ord=2, axis=axis), v)
+                assert_allclose(spnorm(m, ord=None, axis=axis), v)
+
+    def test_norm_exceptions(self):
+        m = self.b
+        assert_raises(TypeError, spnorm, m, None, 1.5)
+        assert_raises(TypeError, spnorm, m, None, [2])
+        assert_raises(ValueError, spnorm, m, None, ())
+        assert_raises(ValueError, spnorm, m, None, (0, 1, 2))
+        assert_raises(ValueError, spnorm, m, None, (0, 0))
+        assert_raises(ValueError, spnorm, m, None, (0, 2))
+        assert_raises(ValueError, spnorm, m, None, (-3, 0))
+        assert_raises(ValueError, spnorm, m, None, 2)
+        assert_raises(ValueError, spnorm, m, None, -3)
+        assert_raises(ValueError, spnorm, m, 'plate_of_shrimp', 0)
+        assert_raises(ValueError, spnorm, m, 'plate_of_shrimp', (0, 1))
+
+
+class TestVsNumpyNorm:
+    _sparse_types = (
+            scipy.sparse.bsr_array,
+            scipy.sparse.coo_array,
+            scipy.sparse.csc_array,
+            scipy.sparse.csr_array,
+            scipy.sparse.dia_array,
+            scipy.sparse.dok_array,
+            scipy.sparse.lil_array,
+            )
+    _test_matrices = (
+            (np.arange(9) - 4).reshape((3, 3)),
+            [
+                [1, 2, 3],
+                [-1, 1, 4]],
+            [
+                [1, 0, 3],
+                [-1, 1, 4j]],
+            )
+
+    def test_sparse_matrix_norms(self):
+        for sparse_type in self._sparse_types:
+            for M in self._test_matrices:
+                S = sparse_type(M)
+                assert_allclose(spnorm(S), npnorm(M))
+                assert_allclose(spnorm(S, 'fro'), npnorm(M, 'fro'))
+                assert_allclose(spnorm(S, np.inf), npnorm(M, np.inf))
+                assert_allclose(spnorm(S, -np.inf), npnorm(M, -np.inf))
+                assert_allclose(spnorm(S, 1), npnorm(M, 1))
+                assert_allclose(spnorm(S, -1), npnorm(M, -1))
+
+    def test_sparse_matrix_norms_with_axis(self):
+        for sparse_type in self._sparse_types:
+            for M in self._test_matrices:
+                S = sparse_type(M)
+                for axis in None, (0, 1), (1, 0):
+                    assert_allclose(spnorm(S, axis=axis), npnorm(M, axis=axis))
+                    for ord in 'fro', np.inf, -np.inf, 1, -1:
+                        assert_allclose(spnorm(S, ord, axis=axis),
+                                        npnorm(M, ord, axis=axis))
+                # Some numpy matrix norms are allergic to negative axes.
+                for axis in (-2, -1), (-1, -2), (1, -2):
+                    assert_allclose(spnorm(S, axis=axis), npnorm(M, axis=axis))
+                    assert_allclose(spnorm(S, 'f', axis=axis),
+                                    npnorm(M, 'f', axis=axis))
+                    assert_allclose(spnorm(S, 'fro', axis=axis),
+                                    npnorm(M, 'fro', axis=axis))
+
+    def test_sparse_vector_norms(self):
+        for sparse_type in self._sparse_types:
+            for M in self._test_matrices:
+                S = sparse_type(M)
+                for axis in (0, 1, -1, -2, (0, ), (1, ), (-1, ), (-2, )):
+                    assert_allclose(spnorm(S, axis=axis), npnorm(M, axis=axis))
+                    for ord in None, 2, np.inf, -np.inf, 1, 0.5, 0.42:
+                        assert_allclose(spnorm(S, ord, axis=axis),
+                                        npnorm(M, ord, axis=axis))

scipy/sparse/linalg/tests/test_onenormest.py

@@ -0,0 +1,252 @@
+"""Test functions for the sparse.linalg._onenormest module
+"""
+
+import numpy as np
+from numpy.testing import assert_allclose, assert_equal, assert_
+import pytest
+import scipy.linalg
+import scipy.sparse.linalg
+from scipy.sparse.linalg._onenormest import _onenormest_core, _algorithm_2_2
+
+
+class MatrixProductOperator(scipy.sparse.linalg.LinearOperator):
+    """
+    This is purely for onenormest testing.
+    """
+
+    def __init__(self, A, B):
+        if A.ndim != 2 or B.ndim != 2:
+            raise ValueError('expected ndarrays representing matrices')
+        if A.shape[1] != B.shape[0]:
+            raise ValueError('incompatible shapes')
+        self.A = A
+        self.B = B
+        self.ndim = 2
+        self.shape = (A.shape[0], B.shape[1])
+
+    def _matvec(self, x):
+        return np.dot(self.A, np.dot(self.B, x))
+
+    def _rmatvec(self, x):
+        return np.dot(np.dot(x, self.A), self.B)
+
+    def _matmat(self, X):
+        return np.dot(self.A, np.dot(self.B, X))
+
+    @property
+    def T(self):
+        return MatrixProductOperator(self.B.T, self.A.T)
+
+
+class TestOnenormest:
+
+    @pytest.mark.xslow
+    def test_onenormest_table_3_t_2(self):
+        # This will take multiple seconds if your computer is slow like mine.
+        # It is stochastic, so the tolerance could be too strict.
+        np.random.seed(1234)
+        t = 2
+        n = 100
+        itmax = 5
+        nsamples = 5000
+        observed = []
+        expected = []
+        nmult_list = []
+        nresample_list = []
+        for i in range(nsamples):
+            A = scipy.linalg.inv(np.random.randn(n, n))
+            est, v, w, nmults, nresamples = _onenormest_core(A, A.T, t, itmax)
+            observed.append(est)
+            expected.append(scipy.linalg.norm(A, 1))
+            nmult_list.append(nmults)
+            nresample_list.append(nresamples)
+        observed = np.array(observed, dtype=float)
+        expected = np.array(expected, dtype=float)
+        relative_errors = np.abs(observed - expected) / expected
+
+        # check the mean underestimation ratio
+        underestimation_ratio = observed / expected
+        assert_(0.99 < np.mean(underestimation_ratio) < 1.0)
+
+        # check the max and mean required column resamples
+        assert_equal(np.max(nresample_list), 2)
+        assert_(0.05 < np.mean(nresample_list) < 0.2)
+
+        # check the proportion of norms computed exactly correctly
+        nexact = np.count_nonzero(relative_errors < 1e-14)
+        proportion_exact = nexact / float(nsamples)
+        assert_(0.9 < proportion_exact < 0.95)
+
+        # check the average number of matrix*vector multiplications
+        assert_(3.5 < np.mean(nmult_list) < 4.5)
+
+    @pytest.mark.xslow
+    def test_onenormest_table_4_t_7(self):
+        # This will take multiple seconds if your computer is slow like mine.
+        # It is stochastic, so the tolerance could be too strict.
+        np.random.seed(1234)
+        t = 7
+        n = 100
+        itmax = 5
+        nsamples = 5000
+        observed = []
+        expected = []
+        nmult_list = []
+        nresample_list = []
+        for i in range(nsamples):
+            A = np.random.randint(-1, 2, size=(n, n))
+            est, v, w, nmults, nresamples = _onenormest_core(A, A.T, t, itmax)
+            observed.append(est)
+            expected.append(scipy.linalg.norm(A, 1))
+            nmult_list.append(nmults)
+            nresample_list.append(nresamples)
+        observed = np.array(observed, dtype=float)
+        expected = np.array(expected, dtype=float)
+        relative_errors = np.abs(observed - expected) / expected
+
+        # check the mean underestimation ratio
+        underestimation_ratio = observed / expected
+        assert_(0.90 < np.mean(underestimation_ratio) < 0.99)
+
+        # check the required column resamples
+        assert_equal(np.max(nresample_list), 0)
+
+        # check the proportion of norms computed exactly correctly
+        nexact = np.count_nonzero(relative_errors < 1e-14)
+        proportion_exact = nexact / float(nsamples)
+        assert_(0.15 < proportion_exact < 0.25)
+
+        # check the average number of matrix*vector multiplications
+        assert_(3.5 < np.mean(nmult_list) < 4.5)
+
+    def test_onenormest_table_5_t_1(self):
+        # "note that there is no randomness and hence only one estimate for t=1"
+        t = 1
+        n = 100
+        itmax = 5
+        alpha = 1 - 1e-6
+        A = -scipy.linalg.inv(np.identity(n) + alpha*np.eye(n, k=1))
+        first_col = np.array([1] + [0]*(n-1))
+        first_row = np.array([(-alpha)**i for i in range(n)])
+        B = -scipy.linalg.toeplitz(first_col, first_row)
+        assert_allclose(A, B)
+        est, v, w, nmults, nresamples = _onenormest_core(B, B.T, t, itmax)
+        exact_value = scipy.linalg.norm(B, 1)
+        underest_ratio = est / exact_value
+        assert_allclose(underest_ratio, 0.05, rtol=1e-4)
+        assert_equal(nmults, 11)
+        assert_equal(nresamples, 0)
+        # check the non-underscored version of onenormest
+        est_plain = scipy.sparse.linalg.onenormest(B, t=t, itmax=itmax)
+        assert_allclose(est, est_plain)
+
+    @pytest.mark.xslow
+    def test_onenormest_table_6_t_2(self):
+        #TODO this test seems to give estimates that match the table,
+        #TODO even though no attempt has been made to deal with
+        #TODO complex numbers in the one-norm estimation.
+        # This will take multiple seconds if your computer is slow like mine.
+        # It is stochastic, so the tolerance could be too strict.
+        np.random.seed(1234)
+        t = 2
+        n = 100
+        itmax = 5
+        nsamples = 5000
+        observed = []
+        expected = []
+        nmult_list = []
+        nresample_list = []
+        for i in range(nsamples):
+            A_inv = np.random.rand(n, n) + 1j * np.random.rand(n, n)
+            A = scipy.linalg.inv(A_inv)
+            est, v, w, nmults, nresamples = _onenormest_core(A, A.T, t, itmax)
+            observed.append(est)
+            expected.append(scipy.linalg.norm(A, 1))
+            nmult_list.append(nmults)
+            nresample_list.append(nresamples)
+        observed = np.array(observed, dtype=float)
+        expected = np.array(expected, dtype=float)
+        relative_errors = np.abs(observed - expected) / expected
+
+        # check the mean underestimation ratio
+        underestimation_ratio = observed / expected
+        underestimation_ratio_mean = np.mean(underestimation_ratio)
+        assert_(0.90 < underestimation_ratio_mean < 0.99)
+
+        # check the required column resamples
+        max_nresamples = np.max(nresample_list)
+        assert_equal(max_nresamples, 0)
+
+        # check the proportion of norms computed exactly correctly
+        nexact = np.count_nonzero(relative_errors < 1e-14)
+        proportion_exact = nexact / float(nsamples)
+        assert_(0.7 < proportion_exact < 0.8)
+
+        # check the average number of matrix*vector multiplications
+        mean_nmult = np.mean(nmult_list)
+        assert_(4 < mean_nmult < 5)
+
+    def _help_product_norm_slow(self, A, B):
+        # for profiling
+        C = np.dot(A, B)
+        return scipy.linalg.norm(C, 1)
+
+    def _help_product_norm_fast(self, A, B):
+        # for profiling
+        t = 2
+        itmax = 5
+        D = MatrixProductOperator(A, B)
+        est, v, w, nmults, nresamples = _onenormest_core(D, D.T, t, itmax)
+        return est
+
+    @pytest.mark.slow
+    def test_onenormest_linear_operator(self):
+        # Define a matrix through its product A B.
+        # Depending on the shapes of A and B,
+        # it could be easy to multiply this product by a small matrix,
+        # but it could be annoying to look at all of
+        # the entries of the product explicitly.
+        np.random.seed(1234)
+        n = 6000
+        k = 3
+        A = np.random.randn(n, k)
+        B = np.random.randn(k, n)
+        fast_estimate = self._help_product_norm_fast(A, B)
+        exact_value = self._help_product_norm_slow(A, B)
+        assert_(fast_estimate <= exact_value <= 3*fast_estimate,
+                f'fast: {fast_estimate:g}\nexact:{exact_value:g}')
+
+    def test_returns(self):
+        np.random.seed(1234)
+        A = scipy.sparse.rand(50, 50, 0.1)
+
+        s0 = scipy.linalg.norm(A.toarray(), 1)
+        s1, v = scipy.sparse.linalg.onenormest(A, compute_v=True)
+        s2, w = scipy.sparse.linalg.onenormest(A, compute_w=True)
+        s3, v2, w2 = scipy.sparse.linalg.onenormest(A, compute_w=True, compute_v=True)
+
+        assert_allclose(s1, s0, rtol=1e-9)
+        assert_allclose(np.linalg.norm(A.dot(v), 1), s0*np.linalg.norm(v, 1), rtol=1e-9)
+        assert_allclose(A.dot(v), w, rtol=1e-9)
+
+
+class TestAlgorithm_2_2:
+
+    @pytest.mark.thread_unsafe
+    def test_randn_inv(self):
+        rng = np.random.RandomState(1234)
+        n = 20
+        nsamples = 100
+        for i in range(nsamples):
+
+            # Choose integer t uniformly between 1 and 3 inclusive.
+            t = rng.randint(1, 4)
+
+            # Choose n uniformly between 10 and 40 inclusive.
+            n = rng.randint(10, 41)
+
+            # Sample the inverse of a matrix with random normal entries.
+            A = scipy.linalg.inv(rng.randn(n, n))
+
+            # Compute the 1-norm bounds.
+            g, ind = _algorithm_2_2(A, A.T, t)

scipy/sparse/linalg/tests/test_propack.py

@@ -0,0 +1,165 @@
+import os
+import pytest
+
+import numpy as np
+from numpy.testing import assert_allclose
+from pytest import raises as assert_raises
+from scipy.sparse.linalg._svdp import _svdp
+from scipy.sparse import csr_array, csc_array
+
+
+# dtype_flavour to tolerance
+TOLS = {
+    np.float32: 1e-4,
+    np.float64: 1e-8,
+    np.complex64: 1e-4,
+    np.complex128: 1e-8,
+}
+
+
+def is_complex_type(dtype):
+    return np.dtype(dtype).kind == "c"
+
+
+_dtypes = []
+for dtype_flavour in TOLS.keys():
+    marks = []
+    if is_complex_type(dtype_flavour):
+        marks = [pytest.mark.slow]
+    _dtypes.append(pytest.param(dtype_flavour, marks=marks,
+                                id=dtype_flavour.__name__))
+_dtypes = tuple(_dtypes)  # type: ignore[assignment]
+
+
+def generate_matrix(constructor, n, m, f,
+                    dtype=float, rseed=0, **kwargs):
+    """Generate a random sparse array"""
+    rng = np.random.RandomState(rseed)
+    if is_complex_type(dtype):
+        M = (- 5 + 10 * rng.rand(n, m)
+             - 5j + 10j * rng.rand(n, m)).astype(dtype)
+    else:
+        M = (-5 + 10 * rng.rand(n, m)).astype(dtype)
+    M[M.real > 10 * f - 5] = 0
+    return constructor(M, **kwargs)
+
+
+def assert_orthogonal(u1, u2, rtol, atol):
+    """Check that the first k rows of u1 and u2 are orthogonal"""
+    A = abs(np.dot(u1.conj().T, u2))
+    assert_allclose(A, np.eye(u1.shape[1], u2.shape[1]), rtol=rtol, atol=atol)
+
+
+def check_svdp(n, m, constructor, dtype, k, irl_mode, which, f=0.8):
+    tol = TOLS[dtype]
+
+    M = generate_matrix(np.asarray, n, m, f, dtype)
+    Msp = constructor(M)
+
+    u1, sigma1, vt1 = np.linalg.svd(M, full_matrices=False)
+    u2, sigma2, vt2, _ = _svdp(Msp, k=k, which=which, irl_mode=irl_mode,
+                               tol=tol, rng=np.random.default_rng(0))
+
+    # check the which
+    if which.upper() == 'SM':
+        u1 = np.roll(u1, k, 1)
+        vt1 = np.roll(vt1, k, 0)
+        sigma1 = np.roll(sigma1, k)
+
+    # check that singular values agree
+    assert_allclose(sigma1[:k], sigma2, rtol=tol, atol=tol)
+
+    # check that singular vectors are orthogonal
+    assert_orthogonal(u1, u2, rtol=tol, atol=tol)
+    assert_orthogonal(vt1.T, vt2.T, rtol=tol, atol=tol)
+
+
+@pytest.mark.parametrize('ctor', (np.array, csr_array, csc_array))
+@pytest.mark.parametrize('dtype', _dtypes)
+@pytest.mark.parametrize('irl', (True, False))
+@pytest.mark.parametrize('which', ('LM', 'SM'))
+def test_svdp(ctor, dtype, irl, which):
+    np.random.seed(0)
+    n, m, k = 10, 20, 3
+    if which == 'SM' and not irl:
+        message = "`which`='SM' requires irl_mode=True"
+        with assert_raises(ValueError, match=message):
+            check_svdp(n, m, ctor, dtype, k, irl, which)
+    else:
+        check_svdp(n, m, ctor, dtype, k, irl, which)
+
+
+@pytest.mark.xslow
+@pytest.mark.parametrize('dtype', _dtypes)
+@pytest.mark.parametrize('irl', (False, True))
+def test_examples(dtype, irl):
+    # Note: atol for complex64 bumped from 1e-4 to 1e-3 due to test failures
+    # with BLIS, Netlib, and MKL+AVX512 - see
+    # https://github.com/conda-forge/scipy-feedstock/pull/198#issuecomment-999180432
+    atol = {
+        np.float32: 1.3e-4,
+        np.float64: 1e-9,
+        np.complex64: 1e-3,
+        np.complex128: 1e-9,
+    }[dtype]
+
+    path_prefix = os.path.dirname(__file__)
+    # Test matrices from `illc1850.coord` and `mhd1280b.cua` distributed with
+    # PROPACK 2.1: http://sun.stanford.edu/~rmunk/PROPACK/
+    relative_path = "propack_test_data.npz"
+    filename = os.path.join(path_prefix, relative_path)
+    with np.load(filename, allow_pickle=True) as data:
+        if is_complex_type(dtype):
+            A = data['A_complex'].item().astype(dtype)
+        else:
+            A = data['A_real'].item().astype(dtype)
+
+    k = 200
+    u, s, vh, _ = _svdp(A, k, irl_mode=irl, rng=np.random.default_rng(0))
+
+    # complex example matrix has many repeated singular values, so check only
+    # beginning non-repeated singular vectors to avoid permutations
+    sv_check = 27 if is_complex_type(dtype) else k
+    u = u[:, :sv_check]
+    vh = vh[:sv_check, :]
+    s = s[:sv_check]
+
+    # Check orthogonality of singular vectors
+    assert_allclose(np.eye(u.shape[1]), u.conj().T @ u, atol=atol)
+    assert_allclose(np.eye(vh.shape[0]), vh @ vh.conj().T, atol=atol)
+
+    # Ensure the norm of the difference between the np.linalg.svd and
+    # PROPACK reconstructed matrices is small
+    u3, s3, vh3 = np.linalg.svd(A.todense())
+    u3 = u3[:, :sv_check]
+    s3 = s3[:sv_check]
+    vh3 = vh3[:sv_check, :]
+    A3 = u3 @ np.diag(s3) @ vh3
+    recon = u @ np.diag(s) @ vh
+    assert_allclose(np.linalg.norm(A3 - recon), 0, atol=atol)
+
+
+@pytest.mark.parametrize('shifts', (None, -10, 0, 1, 10, 70))
+@pytest.mark.parametrize('dtype', _dtypes[:2])
+def test_shifts(shifts, dtype):
+    rng = np.random.default_rng(0)
+    n, k = 70, 10
+    A = rng.random((n, n))
+    if shifts is not None and ((shifts < 0) or (k > min(n-1-shifts, n))):
+        with pytest.raises(ValueError):
+            _svdp(A, k, shifts=shifts, kmax=5*k, irl_mode=True, rng=rng)
+    else:
+        _svdp(A, k, shifts=shifts, kmax=5*k, irl_mode=True, rng=rng)
+
+
+@pytest.mark.slow
+@pytest.mark.xfail()
+def test_shifts_accuracy():
+    rng = np.random.default_rng(0)
+    n, k = 70, 10
+    A = rng.random((n, n)).astype(np.float64)
+    u1, s1, vt1, _ = _svdp(A, k, shifts=None, which='SM', irl_mode=True, rng=rng)
+    u2, s2, vt2, _ = _svdp(A, k, shifts=32, which='SM', irl_mode=True, rng=rng)
+    # shifts <= 32 doesn't agree with shifts > 32
+    # Does agree when which='LM' instead of 'SM'
+    assert_allclose(s1, s2)