Merge pull request #66 from qutech/feature/small_improvements

Small improvements

filter_functions/basis.py

@@ -568,7 +568,7 @@ def _full_from_partial(elems: Sequence, traceless: bool, labels: Sequence[str])
     else:
         ggm = Basis.ggm(elems.d)
 
-    coeffs = expand(elems, ggm, tidyup=True)
+    coeffs = expand(elems, ggm, hermitian=elems.isherm, tidyup=True)
     # Throw out coefficient vectors that are all zero (should only happen for
     # the identity)
     coeffs = coeffs[(coeffs != 0).any(axis=-1)]
@@ -637,7 +637,7 @@ def normalize(b: Basis) -> Basis:
 
 
 def expand(M: Union[ndarray, Basis], basis: Union[ndarray, Basis],
-           normalized: bool = True, tidyup: bool = False) -> ndarray:
+           normalized: bool = True, hermitian: bool = False, tidyup: bool = False) -> ndarray:
     r"""
     Expand the array *M* in the basis given by *basis*.
 
@@ -650,6 +650,9 @@ def expand(M: Union[ndarray, Basis], basis: Union[ndarray, Basis],
         The basis of shape (m, d, d) in which to expand.
     normalized: bool {True}
         Wether the basis is normalized.
+    hermitian: bool (default: False)
+        If M is hermitian along its last two axes, the result will be
+        real.
     tidyup: bool {False}
         Whether to set values below the floating point eps to zero.
 
@@ -673,15 +676,18 @@ def expand(M: Union[ndarray, Basis], basis: Union[ndarray, Basis],
                     {\mathrm{tr}\big(C_j^\dagger C_j\big)}.
 
     """
-    coefficients = np.tensordot(M, basis, axes=[(-2, -1), (-1, -2)])
 
+    def cast(arr): return arr.real if hermitian and basis.isherm else arr
+
+    coefficients = cast(np.tensordot(M, basis, axes=[(-2, -1), (-1, -2)]))
     if not normalized:
-        coefficients /= np.einsum('bij,bji->b', basis, basis).real
+        coefficients /= cast(np.einsum('bij,bji->b', basis, basis))
 
     return util.remove_float_errors(coefficients) if tidyup else coefficients
 
 
-def ggm_expand(M: Union[ndarray, Basis], traceless: bool = False) -> ndarray:
+def ggm_expand(M: Union[ndarray, Basis], traceless: bool = False,
+               hermitian: bool = False) -> ndarray:
     r"""
     Expand the matrix *M* in a Generalized Gell-Mann basis [Bert08]_.
     This function makes use of the explicit construction prescription of
@@ -699,11 +705,14 @@ def ggm_expand(M: Union[ndarray, Basis], traceless: bool = False) -> ndarray:
         expansion. If it is known beforehand that M is traceless, the
         corresponding coefficient is zero and thus doesn't need to be
         computed.
+    hermitian: bool (default: False)
+        If M is hermitian along its last two axes, the result will be
+        real.
 
     Returns
     -------
     coefficients: ndarray
-        The coefficient array with shape (d**2,) or (..., d**2)
+        The real coefficient array with shape (d**2,) or (..., d**2)
 
     References
     ----------
@@ -715,6 +724,8 @@ def ggm_expand(M: Union[ndarray, Basis], traceless: bool = False) -> ndarray:
     if M.shape[-1] != M.shape[-2]:
         raise ValueError('M should be square in its last two axes')
 
+    def cast(arr): return arr.real if hermitian else arr
+
     # Squeeze out an extra dimension to be shape agnostic
     square = M.ndim < 3
     if square:
@@ -740,18 +751,19 @@ def ggm_expand(M: Union[ndarray, Basis], traceless: bool = False) -> ndarray:
     diag_idx_shifted = tuple([...] + [tuple(i for i in range(1, d))]*2)
 
     # Compute the coefficients
-    coeffs = np.zeros((*M.shape[:-2], d**2), dtype=complex)
+    coeffs = np.zeros((*M.shape[:-2], d**2), dtype=float if hermitian else complex)
     if not traceless:
         # First element is proportional to the trace of M
-        coeffs[..., 0] = np.einsum('...jj', M)/np.sqrt(d)
+        coeffs[..., 0] = cast(np.einsum('...jj', M))/np.sqrt(d)
 
     # Elements proportional to the symmetric GGMs
-    coeffs[..., sym_rng] = (M[triu_idx] + M[tril_idx])/np.sqrt(2)
+    coeffs[..., sym_rng] = cast(M[triu_idx] + M[tril_idx])/np.sqrt(2)
     # Elements proportional to the antisymmetric GGMs
-    coeffs[..., sym_rng + n_sym] = 1j*(M[triu_idx] - M[tril_idx])/np.sqrt(2)
+    coeffs[..., sym_rng + n_sym] = cast(1j*(M[triu_idx] - M[tril_idx]))/np.sqrt(2)
     # Elements proportional to the diagonal GGMs
-    coeffs[..., diag_rng + 2*n_sym] = M[diag_idx].cumsum(axis=-1) - diag_rng*M[diag_idx_shifted]
-    coeffs[..., diag_rng + 2*n_sym] /= np.sqrt(diag_rng*(diag_rng + 1))
+    coeffs[..., diag_rng + 2*n_sym] = cast(M[diag_idx].cumsum(axis=-1)
+                                           - diag_rng*M[diag_idx_shifted])
+    coeffs[..., diag_rng + 2*n_sym] /= cast(np.sqrt(diag_rng*(diag_rng + 1)))
 
     return coeffs.squeeze() if square else coeffs
 

filter_functions/gradient.py

@@ -174,11 +174,12 @@ def _liouville_derivative(dt: Coefficients, propagators: ndarray, basis: Basis,
 
     # Calculate the whole propagator derivative
     propagators_deriv = np.zeros((c_opers_transformed.shape[1], n-1, n, d, d), dtype=complex)
+    U_deriv_transformed = np.zeros((c_opers_transformed.shape[1], n-1, d, d), dtype=complex)
     for g in range(n-1):
-        propagators_deriv[:, g, :g+1] = (propagators[g+1]
-                                         @ propagators[1:g+2].conj().swapaxes(-1, -2)
-                                         @ U_deriv[:, :g+1]
-                                         @ propagators[:g+1])
+        U_deriv_transformed[:, g] = (propagators[g+1].conj().swapaxes(-1, -2)
+                                     @ U_deriv[:, g]
+                                     @ propagators[g])
+        propagators_deriv[:, g, :g+1] = propagators[g+1] @ U_deriv_transformed[:, :g+1]
 
     # Equivalent but usually slower contraction: 'htsba,jbc,tcd,kda->thsjk'
     # Can just take 2*Re(·) when calculating x + x*
@@ -466,8 +467,8 @@ def calculate_derivative_of_control_matrix_from_scratch(
 
     See Also
     --------
-    :func:`_liouville_derivative`
-    :func:`_control_matrix_at_timestep_derivative`
+    _liouville_derivative
+    _control_matrix_at_timestep_derivative
     """
     # Distinction between control and drift operators and only
     # calculate the derivatives in control direction

filter_functions/numeric.py

@@ -444,8 +444,8 @@ def calculate_noise_operators_from_atomic(
 
     See Also
     --------
-    :func:`calculate_noise_operators_from_scratch`: Compute the operators from scratch.
-    :func:`calculate_control_matrix_from_atomic`: Same calculation in Liouville space.
+    calculate_noise_operators_from_scratch: Compute the operators from scratch.
+    calculate_control_matrix_from_atomic: Same calculation in Liouville space.
     """
     n = len(noise_operators_atomic)
     # Allocate memory
@@ -572,8 +572,8 @@ def calculate_noise_operators_from_scratch(
 
     See Also
     --------
-    :func:`calculate_noise_operators_from_atomic`: Compute the operators from atomic segments.
-    :func:`calculate_control_matrix_from_scratch`: Same calculation in Liouville space.
+    calculate_noise_operators_from_atomic: Compute the operators from atomic segments.
+    calculate_control_matrix_from_scratch: Same calculation in Liouville space.
     """
     if t is None:
         t = np.concatenate(([0], np.asarray(dt).cumsum()))
@@ -624,7 +624,7 @@ def calculate_noise_operators_from_scratch(
     return noise_operators
 
 
-@util.parse_optional_parameters({'which': ('total', 'correlations')})
+@util.parse_optional_parameters(which=('total', 'correlations'))
 def calculate_control_matrix_from_atomic(
         phases: ndarray,
         control_matrix_atomic: ndarray,
@@ -859,7 +859,7 @@ def calculate_control_matrix_from_scratch(
 
 def calculate_control_matrix_periodic(phases: ndarray, control_matrix: ndarray,
                                       total_propagator_liouville: ndarray,
-                                      repeats: int) -> ndarray:
+                                      repeats: int, check_invertible: bool = True) -> ndarray:
     r"""
     Calculate the control matrix of a periodic pulse given the phase
     factors, control matrix and transfer matrix of the total propagator,
@@ -877,6 +877,11 @@ def calculate_control_matrix_periodic(phases: ndarray, control_matrix: ndarray,
         pulse.
     repeats: int
         The number of repetitions.
+    check_invertible: bool, optional
+        Check for all frequencies if the inverse :math:`\mathbb{I} -
+        e^{i\omega T} \mathcal{Q}^{(1)}` exists by calculating the
+        determinant. If it does not exist, the sum is evaluated
+        explicitly for those points. The default is True.
 
     Returns
     -------
@@ -909,28 +914,23 @@ def calculate_control_matrix_periodic(phases: ndarray, control_matrix: ndarray,
     # evaluate it explicitly.
     eye = np.eye(total_propagator_liouville.shape[0])
     T = np.multiply.outer(phases, total_propagator_liouville)
-
-    # Mask the invertible frequencies. The chosen atol is somewhat empiric.
     M = eye - T
-    M_inv = nla.inv(M)
-    good_inverse = np.isclose(M_inv @ M, eye, atol=1e-10, rtol=0).all((1, 2))
+    if check_invertible:
+        invertible = ~np.isclose(nla.det(M), 0)
+    else:
+        invertible = np.array(True)
 
     S = np.empty((*phases.shape, *total_propagator_liouville.shape), dtype=complex)
-    # Evaluate the sum for invertible frequencies
-    S[good_inverse] = M_inv[good_inverse] @ (eye - nla.matrix_power(T[good_inverse], repeats))
-
-    # Evaluate the sum for non-invertible frequencies
-    # HINT: Using numba, this could be a factor of ten or so faster. But since
-    # usually very few omega-values have a bad inverse, the compilation
-    # overhead is not compensated.
-    if (~good_inverse).any():
-        S[~good_inverse] = eye + sum(accumulate(repeat(T[~good_inverse], repeats-1), np.matmul))
+    # Solve LSE instead of computing inverse, faster + numerically more stable
+    S[invertible] = nla.solve(M[invertible], eye - nla.matrix_power(T[invertible], repeats))
+    if (~invertible).any():
+        S[~invertible] = eye + sum(accumulate(repeat(T[~invertible], repeats-1), np.matmul))
 
     control_matrix_tot = (control_matrix.transpose(2, 0, 1) @ S).transpose(1, 2, 0)
     return control_matrix_tot
 
 
-@util.parse_optional_parameters({'which': ('total', 'correlations')})
+@util.parse_optional_parameters(which=('total', 'correlations'))
 def calculate_cumulant_function(
         pulse: 'PulseSequence',
         spectrum: Optional[ndarray] = None,
@@ -1096,23 +1096,25 @@ def calculate_cumulant_function(
     if d == 2 and pulse.basis.btype in ('Pauli', 'GGM'):
         # Single qubit case. Can use simplified expression
         cumulant_function = np.zeros(decay_amplitudes.shape, decay_amplitudes.dtype)
-        diag_mask = np.eye(N, dtype=bool)
+        diag_mask = np.zeros((N, N), dtype=bool)
+        diag_mask[1:, 1:] = ~np.eye(N-1, dtype=bool)
 
         # Offdiagonal terms
-        cumulant_function[..., ~diag_mask] = decay_amplitudes[..., ~diag_mask]
+        cumulant_function[..., diag_mask] = decay_amplitudes[..., diag_mask]
 
         # Diagonal terms K_ii given by sum over diagonal of Gamma excluding
         # Gamma_ii. Since the Pauli basis is traceless, K_00 is zero, therefore
         # start at K_11.
-        diag_idx = deque((True, False, True, True))
+        diag_deque = deque((False, True, True))
         for i in range(1, N):
+            diag_idx = [False] + list(diag_deque)
             cumulant_function[..., i, i] = -decay_amplitudes[..., diag_idx, diag_idx].sum(axis=-1)
             # shift the item not summed over by one
-            diag_idx.rotate()
+            diag_deque.rotate()
 
         if second_order:
-            cumulant_function -= frequency_shifts
-            cumulant_function += frequency_shifts.swapaxes(-1, -2)
+            cumulant_function[..., 1:, 1:] -= frequency_shifts[..., 1:, 1:]
+            cumulant_function[..., 1:, 1:] += frequency_shifts[..., 1:, 1:].swapaxes(-1, -2)
     else:
         # Multi qubit case. Use general expression. Drop imaginary part since
         # result is guaranteed to be real (if we didn't do anything wrong)
@@ -1134,7 +1136,7 @@ def calculate_cumulant_function(
     return cumulant_function.real
 
 
-@util.parse_optional_parameters({'which': ('total', 'correlations')})
+@util.parse_optional_parameters(which=('total', 'correlations'))
 def calculate_decay_amplitudes(
         pulse: 'PulseSequence',
         spectrum: ndarray,
@@ -1353,7 +1355,7 @@ def calculate_frequency_shifts(
     return frequency_shifts
 
 
-@util.parse_which_FF_parameter
+@util.parse_optional_parameters(which=('fidelity', 'generalized'))
 def calculate_filter_function(control_matrix: ndarray, which: str = 'fidelity') -> ndarray:
     r"""Compute the filter function from the control matrix.
 
@@ -1524,8 +1526,10 @@ def calculate_second_order_filter_function(
         intermediates = dict()
 
     # Work around possibly populated intermediates dict with missing keys
-    n_opers_transformed = intermediates.get('n_opers_transformed',
-                                            _transform_hamiltonian(eigvecs, n_opers, n_coeffs))
+    n_opers_transformed = intermediates.get('n_opers_transformed')
+    if n_opers_transformed is None:
+        n_opers_transformed = _transform_hamiltonian(eigvecs, n_opers, n_coeffs)
+
     try:
         basis_transformed_cache = intermediates['basis_transformed']
         ctrlmat_step_cache = intermediates['control_matrix_step']
@@ -1585,7 +1589,7 @@ def calculate_second_order_filter_function(
     return result
 
 
-@util.parse_which_FF_parameter
+@util.parse_optional_parameters(which=('fidelity', 'generalized'))
 def calculate_pulse_correlation_filter_function(control_matrix: ndarray,
                                                 which: str = 'fidelity') -> ndarray:
     r"""Compute pulse correlation filter function from control matrix.
@@ -1786,7 +1790,7 @@ def error_transfer_matrix(
     For non-Gaussian noise the expression above is perturbative and
     includes noise up to order :math:`\xi^2` and hence
     :math:`\tilde{\mathcal{U}} = \mathbb{1} + \mathcal{K}(\tau) +
-    \mathcal{O}(\xi^2)`
+    \mathcal{O}(\xi^4)`
     (although it is evaluated as a matrix exponential in any case).
 
     Given the above expression of the error transfer matrix, the
@@ -1829,7 +1833,7 @@ def error_transfer_matrix(
     return error_transfer_matrix
 
 
-@util.parse_optional_parameters({'which': ('total', 'correlations')})
+@util.parse_optional_parameters(which=('total', 'correlations'))
 def infidelity(
         pulse: 'PulseSequence',
         spectrum: Union[Coefficients, Callable],
@@ -2028,12 +2032,12 @@ def infidelity(
         else:
             raise ValueError("spacing should be either 'linear' or 'log'.")
 
-        delta_n = (n_max - n_min)//n_points
+        delta_n = (n_max - n_min)//(n_points - 1)
         n_samples = np.arange(n_min, n_max + delta_n, delta_n)
 
         convergence_infids = np.empty((len(n_samples), len(idx)))
         for i, n in enumerate(n_samples):
-            freqs = xspace(omega_IR, omega_UV, n//2)
+            freqs = xspace(omega_IR, omega_UV, n)
             convergence_infids[i] = infidelity(pulse, spectrum(freqs), freqs,
                                                n_oper_identifiers=n_oper_identifiers,
                                                which='total', show_progressbar=show_progressbar,