{ "cells": [ { "attachments": {}, "cell_type": "markdown", "metadata": {}, "source": [ "# 变分量子线性求解器\n", "*Copyright (c) 2022 Institute for Quantum Computing, Baidu Inc. All Rights Reserved.*\n", "## 背景介绍\n", "\n", "线性方程组是数学中一个基本但非常有用的工具。 一个例子是，在经济学中，可以使用线性方程对经济进行建模。 此外，它还为非线性的大型系统提供了简单的估计。 因此求解线性方程组是一项重要的任务。\n", "\n", "变分量子线性求解器（Variational quantum linear solver, VQLS）是一种求解线性方程组的变分量子算法，采用了经典-量子混合的方案，可以在近期的含噪中等规模量子计算机上运行。具体来说，对于一个矩阵 $A$ 和一个向量 $\\boldsymbol{b}$，我们的目标是找到一个向量 $\\boldsymbol{x}$ 使得 $A \\boldsymbol{x} = \\boldsymbol{b}$. 使用 VQLS 算法可以得到一个与 $\\boldsymbol{x}$ 成比例的量子态，即一个归一化的向量 $|x\\rangle = \\frac{\\boldsymbol{x}}{\\lVert \\boldsymbol{x} \\rVert_2}$。" ] }, { "attachments": {}, "cell_type": "markdown", "metadata": {}, "source": [ "## 模型原理\n", "\n", "量子场景的线性方程求解问题和通常的设定略有不同，因为量子计算需要将酉算子应用到量子态上。对于输入的矩阵 $A$，我们需要将其分解成酉算子的线性组合 $A = \\sum_n c_n A_n$，其中每个 $A_n$ 都是酉算子，可以在量子线路上运行。对于输入的向量 $\\boldsymbol{b}$，我们需要假设它是一个能够被某个酉算子 $U$ 制备的量子态 $|b\\rangle$，即 $U|0\\rangle = |b\\rangle$。我们可以用下面这张图来概括 VQLS 算法的整体架构:\n", "\n", "![VQLS](vqls.png)\n", "\n", "可以看到，VQLS 算法是一种混合优化算法，可以分为经典和量子两部分，需要在量子计算机上准备参数化量子电路 $V(\\alpha)$ 并计算损失函数 $C(\\alpha)$，然后在经典计算机上对参数 $\\alpha$ 进行优化从而最小化损失函数，直到损失低于某个阈值，最后输出目标量子态 $|x\\rangle$。其中参数化电路 $V(\\alpha)$ 可以生成一个量子态 $|\\psi(\\alpha)\\rangle$，电路 $F(A)$ 可以计算 $A|\\psi(\\alpha)\\rangle$ 与 $|b\\rangle$ 的近似程度，即损失函数 $C(\\alpha)$。当量子态 $A|\\psi(\\alpha)\\rangle$ 与 $|b\\rangle$ 足够接近时，这就意味着量子态 $|\\psi(\\alpha)\\rangle$ 与目标态 $|x\\rangle$ 足够接近，我们可以输出量子态 $|\\psi(\\alpha)\\rangle$ 作为目标态 $|x\\rangle$ 的近似。" ] }, { "attachments": {}, "cell_type": "markdown", "metadata": {}, "source": [ "## 量桨实现\n", "\n", "我们使用量桨中的 `Circuit` 类结合飞桨优化器来实现 VQLS 算法，其中量子部分中参数化量子电路 $V(\\alpha)$ 为 `Circuit` 中内置的 `complex_entangled_layer` 模板，损失函数计算电路 $F(A)$ 由 Hadamard Test 或 Hadamard-Overlap Test 组成，主要使用了量桨中的 `oracle` 量子门来实现控制 $A_n$ 门，在经典优化部分中我们使用 Adam 优化器来最小化损失函数。\n", "\n", "用户可以使用 toml 文件指定算法的输入，矩阵 $A$ 和向量 $\\boldsymbol{b}$，分别以 `.npy` 文件形式存储。用户可以使用以下代码，通过改变$n$的值随机生成一个 $n\\times n$ 的矩阵 $A$ 以及向量 $\\boldsymbol{b}$。" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "这是一个随机生成的A:\n", "[[4.1702199e+00+7.203245j 1.1437482e-03+3.0233257j\n", " 1.4675589e+00+0.9233859j 1.8626021e+00+3.4556072j\n", " 3.9676747e+00+5.3881674j ]\n", " [2.0445225e+00+8.781175j 2.7387592e-01+6.704675j\n", " 4.1730480e+00+5.5868983j 1.4038694e+00+1.9810148j\n", " 8.0074453e+00+9.682616j ]\n", " [8.7638912e+00+8.946067j 8.5044211e-01+0.39054784j\n", " 1.6983042e+00+8.781425j 9.8346835e-01+4.2110763j\n", " 9.5788956e+00+5.3316526j ]\n", " [6.8650093e+00+8.346256j 1.8288277e-01+7.5014434j\n", " 9.8886108e+00+7.4816566j 2.8044400e+00+7.892793j\n", " 1.0322601e+00+4.4789352j ]\n", " [2.8777535e+00+1.3002857j 1.9366957e-01+6.7883554j\n", " 2.1162813e+00+2.6554666j 4.9157314e+00+0.5336254j\n", " 5.7411761e+00+1.4672858j ]]\n", "这是一个随机生成的b:\n", "[4.191945 +6.852195j 3.1342418+6.9232264j 6.9187713+3.1551564j\n", " 9.085955 +2.9361415j 5.8930554+6.9975834j]\n" ] } ], "source": [ "n = 5\n", "\n", "import numpy as np\n", "\n", "\n", "np.random.seed(1)\n", "A = np.zeros([n, n], dtype=\"complex64\")\n", "b = np.zeros(n, dtype=\"complex64\")\n", "for i in range(n):\n", " for j in range(n):\n", " x = np.random.rand() * 10\n", " y = np.random.rand() * 10\n", " A[i][j] = complex(x, y)\n", " x = np.random.rand() * 10\n", " y = np.random.rand() * 10\n", " b[i] = complex(x, y)\n", "np.save(\"./A.npy\", A)\n", "np.save(\"./b.npy\", b)\n", "print(\"这是一个随机生成的A:\")\n", "print(A)\n", "print(\"这是一个随机生成的b:\")\n", "print(b)\n" ] }, { "attachments": {}, "cell_type": "markdown", "metadata": {}, "source": [ "用户可以在 toml 文件中指定 VQLS 算法的参数 `depth`，`iterations`，`LR` 以及 `gamma`，分别对应参数化量子电路 $V(\\alpha)$ 的层数，优化器的迭代次数，优化器的学习率，和损失函数的阈值。在命令行输入 `python vqls.py --config config.toml` 即可完成线性方程组求解。这里我们给出一个在线演示的例子，首先定义配置文件的内容如下，用户可以自行更改 `test_toml` 中的参数：" ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [], "source": [ "test_toml = r\"\"\"\n", "# 存储矩阵A的.npy文件的路径。\n", "A_dir = './A.npy'\n", "# 存储向量b的.npy文件的路径。\n", "b_dir = './b.npy'\n", "# 参数化量子电路的层数。\n", "depth = 4\n", "# 优化器迭代次数。\n", "iterations = 200\n", "# 优化器的学习率。\n", "LR = 0.1\n", "# 损失函数的阈值。默认为0。\n", "gamma = 0\n", "\"\"\"" ] }, { "attachments": {}, "cell_type": "markdown", "metadata": {}, "source": [ "运行 VQLS 算法如下：" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "c:\\Users\\yuzhan01\\Miniconda3\\envs\\pq_model\\lib\\site-packages\\paddle\\tensor\\creation.py:125: DeprecationWarning: `np.object` is a deprecated alias for the builtin `object`. To silence this warning, use `object` by itself. Doing this will not modify any behavior and is safe. \n", "Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations\n", " if data.dtype == np.object:\n", " 88%|████████▊ | 176/200 [02:04<00:16, 1.42it/s]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "Threshold value gamma reached, ending optimization\n", "这是求解Ax=b的x: [ 1.3475237 -0.7860472j 0.22970617-0.88826376j -0.35111237-0.31225887j\n", " 0.07606918+1.2138402j -0.729564 +0.48393282j]\n", "实际b的值: [4.191945 +6.852195j 3.1342418+6.9232264j 6.9187713+3.1551564j\n", " 9.085955 +2.9361415j 5.8930554+6.9975834j]\n", "算法得到的Ax的值: [4.185339 +6.8523855j 3.1297188+6.923625j 6.924285 +3.1467872j\n", " 9.092921 +2.932943j 5.8879805+6.999589j ]\n", "相对误差: 0.0008446976\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\n" ] } ], "source": [ "import argparse\n", "import os\n", "import warnings\n", "\n", "warnings.filterwarnings(\"ignore\")\n", "os.environ[\"PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION\"] = \"python\"\n", "\n", "import toml\n", "import numpy as np\n", "import paddle\n", "from paddle_quantum.data_analysis.vqls import compute\n", "\n", "paddle.seed(0)\n", "\n", "if __name__ == \"__main__\":\n", " config = toml.loads(test_toml)\n", " A_dir = config.pop(\"A_dir\")\n", " A = np.load(A_dir)\n", " b_dir = config.pop(\"b_dir\")\n", " b = np.load(b_dir)\n", " result = compute(A, b, **config)\n", "\n", " print(\"求解 Ax=b 的x:\", result)\n", " print(\"实际 b 的值:\", b)\n", " print(\"算法得到的 Ax 的值:\", np.matmul(A, result))\n", " relative_error = np.linalg.norm(b - np.matmul(A, result)) / np.linalg.norm(b)\n", " print(\"相对误差: \", relative_error)\n", " np.save(\"./answer.npy\", result)\n" ] }, { "attachments": {}, "cell_type": "markdown", "metadata": {}, "source": [ "## 引用信息\n", "\n", "```\n", "@misc{bravo-prieto2020variational,\n", " title = {Variational {{Quantum Linear Solver}}},\n", " author = {{Bravo-Prieto}, Carlos and LaRose, Ryan and Cerezo, M. and Subasi, Yigit and Cincio, Lukasz and Coles, Patrick J.},\n", " year = {2020},\n", " month = jun,\n", " number = {arXiv:1909.05820},\n", " eprint = {1909.05820},\n", " eprinttype = {arxiv},\n", " doi = {10.48550/arXiv.1909.05820}\n", "}\n", "```\n", "\n", "## 参考文献\n", "\n", "[1] “Variational Quantum Linear Solver: A Hybrid Algorithm for Linear Systems.” Carlos Bravo-Prieto, Ryan LaRose, Marco Cerezo, Yigit Subasi, Lukasz Cincio, Patrick J. Coles. arXiv:1909.05820, 2019." ] } ], "metadata": { "kernelspec": { "display_name": "pq-dev", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.8.15 (default, Nov 10 2022, 13:17:42) \n[Clang 14.0.6 ]" }, "orig_nbformat": 4, "vscode": { "interpreter": { "hash": "5fea01cac43c34394d065c23bb8c1e536fdb97a765a18633fd0c4eb359001810" } } }, "nbformat": 4, "nbformat_minor": 2 }