qiskit 1.3.0b1__cp39-abi3-macosx_10_9_x86_64.whl

qiskit/synthesis/unitary/aqc/elementary_operations.py

@@ -0,0 +1,108 @@
+# This code is part of Qiskit.
+#
+# (C) Copyright IBM 2021.
+#
+# This code is licensed under the Apache License, Version 2.0. You may
+# obtain a copy of this license in the LICENSE.txt file in the root directory
+# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# Any modifications or derivative works of this code must retain this
+# copyright notice, and modified files need to carry a notice indicating
+# that they have been altered from the originals.
+"""
+These are a number of elementary functions that are required for the AQC routines to work.
+"""
+
+import numpy as np
+
+from qiskit.circuit.library.standard_gates import RXGate, RZGate, RYGate
+
+
+def place_unitary(unitary: np.ndarray, n: int, j: int) -> np.ndarray:
+    """
+    Computes I(j - 1) tensor product U tensor product I(n - j), where U is a unitary matrix
+    of size ``(2, 2)``.
+
+    Args:
+        unitary: a unitary matrix of size ``(2, 2)``.
+        n: num qubits.
+        j: position where to place a unitary.
+
+    Returns:
+        a unitary of n qubits with u in position j.
+    """
+    return np.kron(np.kron(np.eye(2**j), unitary), np.eye(2 ** (n - 1 - j)))
+
+
+def place_cnot(n: int, j: int, k: int) -> np.ndarray:
+    """
+    Places a CNOT from j to k.
+
+    Args:
+        n: number of qubits.
+        j: control qubit.
+        k: target qubit.
+
+    Returns:
+        a unitary of n qubits with CNOT placed at ``j`` and ``k``.
+    """
+    if j < k:
+        unitary = np.kron(
+            np.kron(np.eye(2**j), [[1, 0], [0, 0]]), np.eye(2 ** (n - 1 - j))
+        ) + np.kron(
+            np.kron(
+                np.kron(np.kron(np.eye(2**j), [[0, 0], [0, 1]]), np.eye(2 ** (k - j - 1))),
+                [[0, 1], [1, 0]],
+            ),
+            np.eye(2 ** (n - 1 - k)),
+        )
+    else:
+        unitary = np.kron(
+            np.kron(np.eye(2**j), [[1, 0], [0, 0]]), np.eye(2 ** (n - 1 - j))
+        ) + np.kron(
+            np.kron(
+                np.kron(np.kron(np.eye(2**k), [[0, 1], [1, 0]]), np.eye(2 ** (j - k - 1))),
+                [[0, 0], [0, 1]],
+            ),
+            np.eye(2 ** (n - 1 - j)),
+        )
+    return unitary
+
+
+def rx_matrix(phi: float) -> np.ndarray:
+    """
+    Computes an RX rotation by the angle of ``phi``.
+
+    Args:
+        phi: rotation angle.
+
+    Returns:
+        an RX rotation matrix.
+    """
+    return RXGate(phi).to_matrix()
+
+
+def ry_matrix(phi: float) -> np.ndarray:
+    """
+    Computes an RY rotation by the angle of ``phi``.
+
+    Args:
+        phi: rotation angle.
+
+    Returns:
+        an RY rotation matrix.
+    """
+    return RYGate(phi).to_matrix()
+
+
+def rz_matrix(phi: float) -> np.ndarray:
+    """
+    Computes an RZ rotation by the angle of ``phi``.
+
+    Args:
+        phi: rotation angle.
+
+    Returns:
+        an RZ rotation matrix.
+    """
+    return RZGate(phi).to_matrix()

qiskit/synthesis/unitary/aqc/fast_gradient/__init__.py

@@ -0,0 +1,164 @@
+# This code is part of Qiskit.
+#
+# (C) Copyright IBM 2022.
+#
+# This code is licensed under the Apache License, Version 2.0. You may
+# obtain a copy of this license in the LICENSE.txt file in the root directory
+# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# Any modifications or derivative works of this code must retain this
+# copyright notice, and modified files need to carry a notice indicating
+# that they have been altered from the originals.
+
+r"""
+================================================================================
+Fast implementation of objective function class
+(:mod:`qiskit.synthesis.unitary.aqc.fast_gradient`)
+================================================================================
+
+.. currentmodule:: qiskit.synthesis.unitary.aqc.fast_gradient
+
+Extension to the implementation of Approximate Quantum Compiler as described in the paper [1].
+
+Interface
+=========
+
+The main public class of this module is FastCNOTUnitObjective. It replaces the default objective
+function implementation :class:`.DefaultCNOTUnitObjective` for faster computation.
+The individual classes include the public one (FastCNOTUnitObjective) and few
+internal ones:
+
+.. autosummary::
+    :toctree: ../stubs
+    :template: autosummary/class_no_inherited_members.rst
+
+    FastCNOTUnitObjective
+    LayerBase
+    Layer1Q
+    Layer2Q
+    PMatrix
+
+
+Mathematical Details
+====================
+
+In what follows we briefly outline the main ideas underlying the accelerated implementation
+of objective function class.
+
+* The key ingredient of approximate compiling is the efficient optimization procedure
+  that minimizes :math:`\|V - U\|_{\mathrm{F}}` on a classical computer, where :math:`U`
+  is a given (target) unitary matrix and :math:`V` is a matrix of approximating quantum
+  circuit. Alternatively, we maximize the Hilbert-Schmidt product between :math:`U` and
+  :math:`V` as outlined in the main part of the documentation.
+
+* The circuit :math:`V` can be represented as a sequence of 2-qubit gates (layers)
+  applied one after another. The corresponding matrix takes the form:
+  :math:`V = C_0 C_1 \ldots C_{L-1} F`, where :math:`L` is the length of the sequence
+  (number of layers). If the total number of qubits :math:`n > 2`, every
+  :math:`C_i = C_i(\Theta_i)` is a sparse, :math:`2^n \times 2^n` matrix of 2-qubit gate
+  (CNOT unit block) parameterized by a sub-set of parameters :math:`\Theta_i`
+  (4 parameters per unit block), and :math:`F` is a matrix that comprises the action
+  of all 1-qubit gates in front of approximating circuit. See the paper [1] for details.
+
+* Over the course of optimization we compute the value of objective function and its
+  gradient, which implies computation of :math:`V` and its derivatives
+  :math:`{\partial V}/{\partial \Theta_i}` for all :math:`i`, given the current estimation
+  of all the parameters :math:`\Theta`.
+
+* A naive implementation of the product :math:`V = C_0 C_1 \ldots C_{L-1} F` and its
+  derivatives would include computation and memorization of forward and backward partial
+  products as required by the backtracking algorithm. This is wasteful in terms of
+  performance and resource allocation.
+
+* Minimization of :math:`\|V - U\|_{\mathrm{F}}^2` is equivalent to maximization of
+  :math:`\text{Re}\left(\text{Tr}\left(U^{\dagger} V\right)\right)`. By cyclic permutation
+  of the sequence of matrices under trace operation, we can avoid memorization of intermediate
+  partial products of gate matrices :math:`C_i`. Note, matrix size grows exponentially with
+  the number of qubits, quickly becoming prohibitively large.
+
+* Sparse structure of :math:`C_i` can be exploited to speed up matrix-matrix multiplication.
+  However, using sparse matrices as such does not give performance gain because sparse patterns
+  tend to violate data proximity inside the cache memory of modern CPUs. Instead, we make use
+  of special structure of gate matrices :math:`C_i` coupled with permutation ones. Although
+  permutation is not cache friendly either, its impact is seemingly less severe than that
+  of sparse matrix multiplication (at least in Python implementation).
+
+* On every optimization iteration we, first, compute :math:`V = C_0 C_1 \ldots C_{L-1} F`
+  given the current estimation of all the parameters :math:`\Theta`.
+
+* As for the gradient of objective function, it can be shown (by moving cyclically around
+  an individual matrices under trace operation) that:
+
+.. math::
+    \text{Tr}\left( U^{\dagger} \frac{\partial V}{\partial \Theta_{l,k}} \right) =
+    \langle \text{vec}\left(E_l\right), \text{vec}\left(
+    \frac{\partial C_l}{\partial \Theta_{l,k}}\right) \rangle,
+
+where :math:`\Theta_{l,k}` is a :math:`k`-th parameter of :math:`l`-th CNOT unit block,
+and :math:`E_l=C_{l-1}\left(C_{l-2}\left(\cdots\left(C_0\left(U^{\dagger}V
+C_0^{\dagger}\right)C_1^{\dagger}\right) \cdots\right)C_{l-1}^{\dagger}\right)C_l^{\dagger}`
+is an intermediate matrix.
+
+* For every :math:`l`-th gradient component, we compute the trace using the matrix
+  :math:`E_l`, then this matrix is updated by multiplication on left and on the right
+  by corresponding gate matrices :math:`C_l` and :math:`C_{l+1}^{\dagger}` respectively
+  and proceed to the next gradient component.
+
+* We save computations and resources by not storing intermediate partial products of
+  :math:`C_i`.  Instead, incrementally updated matrix :math:`E_l` keeps all related
+  information. Also, vectorization of involved matrices (see the above formula) allows
+  us to replace matrix-matrix multiplication by "cheaper" vector-vector one under the
+  trace operation.
+
+* The matrices :math:`C_i` are sparse. However, even for relatively small matrices
+  (< 1M elements) sparse-dense multiplication can be very slow. Construction of sparse
+  matrices takes a time as well.  We should update every gate matrix on each iteration
+  of optimization loop.
+
+* In fact, any gate matrix :math:`C_i` can be transformed to what we call a standard
+  form: :math:`C_i = P^T \widetilde{C}_i P`, where :math:`P` is an easily computable
+  permutation matrix and :math:`\widetilde{C}_i` has a block-diagonal layout:
+
+.. math::
+    \widetilde{C}_i = \left(
+    \begin{array}{ccc}
+    G_{4 \times 4} & \ddots & 0 \\
+    \ddots & \ddots & \ddots \\
+    0 & \ddots & G_{4 \times 4}
+    \end{array}
+    \right)
+
+* The 2-qubit gate matrix :math:`G_{4 \times 4}` is repeated along diagonal of the full
+  :math:`2^n \times 2^n` :math:`\widetilde{C}_i`.
+
+* We do not actually create neither matrix :math:`\widetilde{C}_i` nor :math:`P`.
+  In fact, only :math:`G_{4 \times 4}` and a permutation array (of size :math:`2^n`)
+  are kept in memory.
+
+* Consider left-hand side multiplication by some dense, :math:`2^n \times 2^n` matrix :math:`M`:
+
+.. math::
+    C_i M = P^T \widetilde{C}_i P M = P^T \left( \widetilde{C}_i \left( P M \right) \right)
+
+* First, we permute rows of :math:`M`, which is equivalent to the product :math:`P M`, but
+  without expensive multiplication of two :math:`2^n \times 2^n` matrices.
+
+* Second, we compute :math:`\widetilde{C}_i P M` multiplying every block-diagonal sub-matrix
+  :math:`G_{4 \times 4}` by the corresponding rows of :math:`P M`. This is the dense-dense
+  matrix multiplication, which is very well optimized on modern CPUs. Important: the total
+  number of operations is :math:`O(2^{2 n})` in contrast to :math:`O(2^{3 n})` as in general
+  case.
+
+* Third, we permute rows of :math:`\widetilde{C}_i P M` by applying :math:`P^T`.
+
+* Right-hand side multiplication is done in a similar way.
+
+* In summary, we save computational resources by exploiting some properties of 2-qubit gate
+  matrices :math:`C_i` and using hardware optimized multiplication of dense matrices. There
+  is still a room for further improvement, of course.
+
+References:
+
+    [1]: Liam Madden, Andrea Simonetto, Best Approximate Quantum Compiling Problems.
+    `arXiv:2106.05649 <https://arxiv.org/abs/2106.05649>`_
+"""

qiskit/synthesis/unitary/aqc/fast_gradient/fast_grad_utils.py

@@ -0,0 +1,237 @@
+# This code is part of Qiskit.
+#
+# (C) Copyright IBM 2022.
+#
+# This code is licensed under the Apache License, Version 2.0. You may
+# obtain a copy of this license in the LICENSE.txt file in the root directory
+# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# Any modifications or derivative works of this code must retain this
+# copyright notice, and modified files need to carry a notice indicating
+# that they have been altered from the originals.
+
+"""
+Utility functions in the fast gradient implementation.
+"""
+from __future__ import annotations
+from typing import Union
+import numpy as np
+
+
+def is_permutation(x: np.ndarray) -> bool:
+    """
+    Checks if array is really an index permutation.
+
+    Args:
+        1D-array of integers that supposedly represents a permutation.
+
+    Returns:
+        True, if array is really a permutation of indices.
+    """
+    return (
+        isinstance(x, np.ndarray)
+        and x.ndim == 1
+        and x.dtype == np.int64
+        and np.all(np.sort(x) == np.arange(x.size, dtype=np.int64))
+    )
+
+
+def reverse_bits(x: Union[int, np.ndarray], nbits: int, enable: bool) -> Union[int, np.ndarray]:
+    """
+    Reverses the bit order in a number of ``nbits`` length.
+    If ``x`` is an array, then operation is applied to every entry.
+
+    Args:
+        x: either a single integer or an array of integers.
+        nbits: number of meaningful bits in the number x.
+        enable: apply reverse operation, if enabled, otherwise leave unchanged.
+
+    Returns:
+        a number or array of numbers with reversed bits.
+    """
+
+    if not enable:
+        if isinstance(x, int):
+            pass
+        else:
+            x = x.copy()
+        return x
+
+    if isinstance(x, int):
+        res: int | np.ndarray = int(0)
+    else:
+        x = x.copy()
+        res = np.full_like(x, fill_value=0)
+
+    for _ in range(nbits):
+        res <<= 1
+        res |= x & 1
+        x >>= 1
+    return res
+
+
+def swap_bits(num: int, a: int, b: int) -> int:
+    """
+    Swaps the bits at positions 'a' and 'b' in the number 'num'.
+
+    Args:
+        num: an integer number where bits should be swapped.
+        a: index of the first bit to be swapped.
+        b: index of the second bit to be swapped.
+
+    Returns:
+        the number with swapped bits.
+    """
+    x = ((num >> a) ^ (num >> b)) & 1
+    return num ^ ((x << a) | (x << b))
+
+
+def bit_permutation_1q(n: int, k: int) -> np.ndarray:
+    """
+    Constructs index permutation that brings a circuit consisting of a single
+    1-qubit gate to "standard form": ``kron(I(2^n/2), G)``, as we call it. Here n
+    is the number of qubits, ``G`` is a 2x2 gate matrix, ``I(2^n/2)`` is the identity
+    matrix of size ``(2^n/2)x(2^n/2)``, and the full size of the circuit matrix is
+    ``(2^n)x(2^n)``. Circuit matrix in standard form becomes block-diagonal (with
+    sub-matrices ``G`` on the main diagonal). Multiplication of such a matrix and
+    a dense one is much faster than generic dense-dense product. Moreover,
+    we do not need to keep the entire circuit matrix in memory but just 2x2 ``G``
+    one. This saves a lot of memory when the number of qubits is large.
+
+    Args:
+        n: number of qubits.
+        k: index of qubit where single 1-qubit gate is applied.
+
+    Returns:
+        permutation that brings the whole layer to the standard form.
+    """
+    perm = np.arange(2**n, dtype=np.int64)
+    if k != n - 1:
+        for v in range(2**n):
+            perm[v] = swap_bits(v, k, n - 1)
+    return perm
+
+
+def bit_permutation_2q(n: int, j: int, k: int) -> np.ndarray:
+    """
+    Constructs index permutation that brings a circuit consisting of a single
+    2-qubit gate to "standard form": ``kron(I(2^n/4), G)``, as we call it. Here ``n``
+    is the number of qubits, ``G`` is a 4x4 gate matrix, ``I(2^n/4)`` is the identity
+    matrix of size ``(2^n/4)x(2^n/4)``, and the full size of the circuit matrix is
+    ``(2^n)x(2^n)``. Circuit matrix in standard form becomes block-diagonal (with
+    sub-matrices ``G`` on the main diagonal). Multiplication of such a matrix and
+    a dense one is much faster than generic dense-dense product. Moreover,
+    we do not need to keep the entire circuit matrix in memory but just 4x4 ``G``
+    one. This saves a lot of memory when the number of qubits is large.
+
+    Args:
+        n: number of qubits.
+        j: index of control qubit where single 2-qubit gate is applied.
+        k: index of target qubit where single 2-qubit gate is applied.
+
+    Returns:
+        permutation that brings the whole layer to the standard form.
+    """
+    dim = 2**n
+    perm = np.arange(dim, dtype=np.int64)
+    if j < n - 2:
+        if k < n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(swap_bits(v, j, n - 2), k, n - 1)
+        elif k == n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(swap_bits(v, n - 2, n - 1), j, n - 2)
+        else:
+            for v in range(dim):
+                perm[v] = swap_bits(v, j, n - 2)
+    elif j == n - 2:
+        if k < n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(v, k, n - 1)
+        else:
+            pass
+    else:
+        if k < n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(swap_bits(v, n - 2, n - 1), k, n - 1)
+        else:
+            for v in range(dim):
+                perm[v] = swap_bits(v, n - 2, n - 1)
+    return perm
+
+
+def inverse_permutation(perm: np.ndarray) -> np.ndarray:
+    """
+    Returns inverse permutation.
+
+    Args:
+        perm: permutation to be reversed.
+
+    Returns:
+        inverse permutation.
+    """
+    inv = np.zeros_like(perm)
+    inv[perm] = np.arange(perm.size, dtype=np.int64)
+    return inv
+
+
+def make_rx(phi: float, out: np.ndarray) -> np.ndarray:
+    """
+    Makes a 2x2 matrix that corresponds to X-rotation gate.
+    This is a fast implementation that does not allocate the output matrix.
+
+    Args:
+        phi: rotation angle.
+        out: placeholder for the result (2x2, complex-valued matrix).
+
+    Returns:
+        rotation gate, same object as referenced by "out".
+    """
+    a = 0.5 * phi
+    cs, sn = np.cos(a).item(), -1j * np.sin(a).item()
+    out[0, 0] = cs
+    out[0, 1] = sn
+    out[1, 0] = sn
+    out[1, 1] = cs
+    return out
+
+
+def make_ry(phi: float, out: np.ndarray) -> np.ndarray:
+    """
+    Makes a 2x2 matrix that corresponds to Y-rotation gate.
+    This is a fast implementation that does not allocate the output matrix.
+
+    Args:
+        phi: rotation angle.
+        out: placeholder for the result (2x2, complex-valued matrix).
+
+    Returns:
+        rotation gate, same object as referenced by "out".
+    """
+    a = 0.5 * phi
+    cs, sn = np.cos(a).item(), np.sin(a).item()
+    out[0, 0] = cs
+    out[0, 1] = -sn
+    out[1, 0] = sn
+    out[1, 1] = cs
+    return out
+
+
+def make_rz(phi: float, out: np.ndarray) -> np.ndarray:
+    """
+    Makes a 2x2 matrix that corresponds to Z-rotation gate.
+    This is a fast implementation that does not allocate the output matrix.
+
+    Args:
+        phi: rotation angle.
+        out: placeholder for the result (2x2, complex-valued matrix).
+
+    Returns:
+        rotation gate, same object as referenced by "out".
+    """
+    exp = np.exp(0.5j * phi).item()
+    out[0, 0] = 1.0 / exp
+    out[0, 1] = 0
+    out[1, 0] = 0
+    out[1, 1] = exp
+    return out