Gradient_EN.ipynb 36.8 KB
Newer Older
Q
Quleaf 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861
{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "e8062c41",
   "metadata": {},
   "source": [
    "# Calculating Gradient Using Quantum Circuit\n",
    "\n",
    "<em> Copyright (c) 2021 Institute for Quantum Computing, Baidu Inc. All Rights Reserved. </em>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d8947f99",
   "metadata": {},
   "source": [
    "## Overview\n",
    "\n",
    "When using variational quantum algorithms like [Variational Quantum Eigensolver (VQE)](../quantum_simulation/VQE_EN.ipynb) and [Quantum Approximate Optimization Algorithm (QAOA)](../combinatorial_optimization/QAOA_EN.ipynb), which involves varying parameters of quantum circuit to minimize an objective function of interest, it brings out an important question - what are the ways to calculate the gradient of a parameterized quantum circuit? Since the objective function is being evaluated using a quantum circuit, it's necessary to evaluate its gradient using quantum algorithms as well. Compared with computing gradient classically, this is more challenging for sure. Below we provide three different methods to accomplish this task on a quantum computer. And here we demonstrate the code and simulate their effect of running on a quantum computer using Paddle Quantum."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "61395a57",
   "metadata": {},
   "source": [
    "## Introduction\n",
    "\n",
    "Suppose the objective function is the typical parameterized cost function used in VQA: $O(\\theta) = \\left\\langle00\\right| U^{\\dagger}(\\theta)HU(\\theta) \\left|00\\right\\rangle$ where H is a Hamiltonian, $U(\\theta)$ represents the parameterized quantum circuit and $\\theta = [\\theta_1, \\theta_2, \\dots, \\theta_n]$ is a list of trainable parameters in the circuit, then our goal is to find\n",
    "\n",
    "$$\n",
    "\\nabla O(\\theta) = \\begin{bmatrix} \\frac{\\partial O}{\\partial \\theta_1} \\\\ \\frac{\\partial O}{\\partial \\theta_2}\\\\ \\vdots\\\\ \\frac{\\partial O}{\\partial \\theta_n} \\end{bmatrix}.\n",
    "\\tag{1}\n",
    "$$\n",
    "\n",
    "First, let's import all the required packages."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "5e438617",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import paddle\n",
    "from paddle_quantum.circuit import UAnsatz\n",
    "from paddle_quantum.utils import pauli_str_to_matrix, Hamiltonian\n",
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ec29c111",
   "metadata": {},
   "source": [
    "Then, let's construct our $U(\\theta)$ and Hamiltonian $H$ of the objective function $O(\\theta) = \\left\\langle00\\right| U^{\\dagger}(\\theta)HU(\\theta) \\left|00\\right\\rangle$.\n",
    "\n",
    "We will demonstrate our example on a 2-qubit quantum circuit, constructing $U(\\theta)$ with randomly generated theta of size four. We choose our Hamiltonian H to be $H = Z \\otimes Z$."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "00c29cb2",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Hamiltonian H: \n",
      " [[ 1.+0.j  0.+0.j  0.+0.j  0.+0.j]\n",
      " [ 0.+0.j -1.+0.j  0.+0.j  0.+0.j]\n",
      " [ 0.+0.j  0.+0.j -1.+0.j  0.+0.j]\n",
      " [ 0.+0.j  0.+0.j  0.+0.j  1.+0.j]]\n",
      "\n",
      "U(theta):\n",
      "--Ry(5.690)----*----x----Ry(3.107)--\n",
      "               |    |               \n",
      "--Ry(2.521)----x----*----Ry(0.437)--\n",
      "                                    \n"
     ]
    }
   ],
   "source": [
    "# Define Hamiltonian H\n",
    "pauli_str = [[1.0, 'Z0,Z1']]\n",
    "H = Hamiltonian(pauli_str)\n",
    "\n",
    "# Randomly generate parameters in range from 0 to 2*PI\n",
    "theta_np = np.random.uniform(0, 2 * np.pi, 4)\n",
    "# Warning: Note that when defining theta, if we mark stop_gradient=False, then this parameter is trainable; \n",
    "# otherwise, the parameter will be treated as constant, its gradient will not be calculated and it will not be updated in training process.\n",
    "theta_tensor = paddle.to_tensor(theta_np, 'float64', stop_gradient=False)\n",
    "\n",
    "def U_theta(theta):\n",
    "    cir = UAnsatz(2)\n",
    "    cir.ry(theta[0], 0)\n",
    "    cir.ry(theta[1], 1)\n",
    "    cir.cnot([0, 1])\n",
    "    cir.cnot([1, 0])\n",
    "    cir.ry(theta[2], 0)\n",
    "    cir.ry(theta[3], 1)\n",
    "    cir.run_state_vector()\n",
    "    return cir\n",
    "\n",
    "print('Hamiltonian H: \\n', H.construct_h_matrix())\n",
    "print('\\nU(theta):')\n",
    "print(U_theta(theta_tensor))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fc5028f6",
   "metadata": {},
   "source": [
    "## Finite Difference Method\n",
    "\n",
    "The Finite difference method is one of the most traditional and common numerical methods to estimate the gradient of a function. The main idea is that the error of the derivative of a function $f(x)$ tends to zero as $h$ tends to zero: \n",
    "\n",
    "$$\n",
    "f'(x)= \\lim_{h \\to 0}\\frac{f(x+h) - f(x)}{h}.\n",
    "\\tag{2}\n",
    "$$\n",
    "\n",
    "By choosing a sufficiently small $h$, we can get a good approximation of the derivative.\n",
    "\n",
    "For example, by using one type of finite difference, the central finite difference method, the objective function's gradient will be\n",
    "\n",
    "$$\n",
    "\\nabla O(\\theta) \\approx \\frac{O(\\theta+\\delta) - O(\\theta-\\delta)}{2\\delta} = \\frac{\\left\\langle00\\right| U^{\\dagger}(\\theta + \\delta)HU(\\theta + \\delta) \\left|00\\right\\rangle - \\left\\langle00\\right| U^{\\dagger}(\\theta - \\delta)HU(\\theta-\\delta) \\left|00\\right\\rangle)}{2\\delta}.\n",
    "\\tag{3}\n",
    "$$\n",
    "\n",
    "When implementing this, we can simply loop through the parameter list, shift the specific parameter in the original circuit and evaluate the objective function over and over again. There's no need to build extra circuits or using extra qubits.\n",
    "\n",
    "Using Paddle Quantum's built-in method, we can build the circuit of $U(\\theta)$, and compute the finite-difference gradient easily by passing in the corresponding Hamiltonian H and the delta. Note: The built-in method currently does not support assigning an input state or running on a noisy circuit."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "c836f71e",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "--Ry(5.690)----*----x----Ry(3.107)--\n",
      "               |    |               \n",
      "--Ry(2.521)----x----*----Ry(0.437)--\n",
      "                                    \n",
      "Gradient of this objective function is:  [-0.79156126  0.12584221 -0.26541629  0.78068313]\n"
     ]
    }
   ],
   "source": [
    "# We reuse the predefined Hamiltonian H and parameters of the circuit. \n",
    "# Again, be sure to mark stop_gradient=False when defining tensor for parameters of the circuit\n",
    "\n",
    "# Constructing circuit U(theta)\n",
    "cir = U_theta(theta_tensor)\n",
    "print(cir)\n",
    "\n",
    "# Calculating finite-difference gradient\n",
    "gradients = cir.finite_difference_gradient(H, delta=0.01)\n",
    "print(\"Gradient of this objective function is: \", gradients.numpy())"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8055ed76",
   "metadata": {},
   "source": [
    "## Parameter-shift Method\n",
    "Again, we use this objective function $O(\\theta) = \\left\\langle0\\right| U^{\\dagger}(\\theta)HU(\\theta) \\left|0\\right\\rangle$ as our example. If $U(\\theta)$ can be written as $e^{-ia\\theta G}$ where $G$ has two unique eigenvalues $\\lambda_1$ and $\\lambda_2$, we can apply parameter-shift method to find its gradient [1]:\n",
    "\n",
    "$$\n",
    "\\nabla O(\\theta) = r \\left[O(\\theta+\\frac{\\pi}{4r}) - O(\\theta-\\frac{\\pi}{4r})\\right],\n",
    "\\tag{4}\n",
    "$$ \n",
    "\n",
    "where shift constant $r = \\frac{a}{2} (\\lambda_2 - \\lambda_1)$. Note that we're getting a theoretically exact gradient instead of an estimation like finite difference gradient. Moreover, this method doesn't require the construction of new circuits or adding ancilla qubits. Evaluations could be done just by changing the parameters inside the circuit.\n",
    "\n",
    "The fundamental rotation gates provided by Paddle Quantum are $R_x(\\theta), R_y(\\theta), R_z(\\theta)$ gates, which can be written as $e^{-i\\frac{1}{2}\\theta X}, e^{-i\\frac{1}{2}\\theta Y}, e^{-i\\frac{1}{2}\\theta Z}$ respectively. Since the eigenvalues of $X$, $Y$, and $Z$ gates are unique, which are -1 and 1, it's not hard to see that $r = \\frac{1}{2}$, and the gradient of those gates is just\n",
    "\n",
    "$$\n",
    "\\frac{1}{2}[ O(\\theta + \\frac{\\pi}{2}) - O(\\theta - \\frac{\\pi}{2})].\n",
    "\\tag{5}\n",
    "$$\n",
    "\n",
    "We will demonstrate this formula's derivation using a single rotation gate $R_x$."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "81a1defe",
   "metadata": {},
   "source": [
    "### Derivation\n",
    "\n",
    "In this section, we will go through all the steps and arrive at the resulting formula for calculating $R_x$ gate's derivative. Using $R_x(\\theta)$ as $U(\\theta)$, we have\n",
    "\n",
    "$$\n",
    "O(\\theta) = \\left\\langle0\\right| R_x^{\\dagger}(\\theta)HR_x(\\theta) \\left|0\\right\\rangle.\n",
    "\\tag{6}\n",
    "$$\n",
    "\n",
    "Given that $R_x(\\theta) = e^{-i\\frac{1}{2}\\theta X}$, where $X$ is the Pauli-X matrix, we know $\\frac{\\partial}{\\partial \\theta}  R_x(\\theta) =-i\\frac{1}{2}Xe^{-i\\frac{\\theta}{2}X}=-i\\frac{1}{2}XR_x(\\theta)$.  Using the product rule, its derivative can be written as\n",
    "\n",
    "$$\n",
    "O'(\\theta) = \\left\\langle0\\right| [\\frac{i}{2}X] R_x^{\\dagger}(\\theta)HR_x(\\theta)\\left|0\\right\\rangle + \\left\\langle0\\right| R_x^{\\dagger}(\\theta)H [-\\frac{i}{2}X] R_x(\\theta)\\left|0\\right\\rangle.\n",
    "\\tag{7}\n",
    "$$\n",
    "\n",
    "We take out the shift constant $r$ in front for the purpose of later rearranging the equation using some trick, where $r$ for $R_x$ gate is $\\frac{1}{2}$. We have\n",
    "\n",
    "$$\n",
    "O'(\\theta) = r \\left\\langle0\\right| [\\frac{i}{2r}X] R_x^{\\dagger}(\\theta)HR_x(\\theta)\\left|0\\right\\rangle + \\left\\langle0\\right| R_x^{\\dagger}(\\theta)H [-\\frac{i}{2r}X] R_x(\\theta)\\left|0\\right\\rangle.\n",
    "\\tag{8}\n",
    "$$\n",
    "\n",
    "Since we recognize that for any operators $U$, $V$ and $Q$, and for an arbitrary state $|\\psi\\rangle$, \n",
    "\n",
    "$$\n",
    "\\langle\\psi|U^\\dagger QV|\\psi\\rangle + \\langle\\psi|V^\\dagger QU|\\psi\\rangle = \\frac{1}{2} \\big(\\langle\\psi|(U+V)^\\dagger Q(U+V)|\\psi\\rangle - \\langle\\psi|(U-V)^\\dagger Q(U-V)|\\psi\\rangle \\big),\n",
    "\\tag{9}\n",
    "$$\n",
    "\n",
    "we get \n",
    "\n",
    "$$\n",
    "O'(\\theta) = \\frac{r}{2} \\big( \\left\\langle0\\right|R_x^{\\dagger}(\\theta) [I + \\frac{i}{2r}X]H[I - \\frac{i}{2r}X]R_x(\\theta)\\left|0\\right\\rangle - \\left\\langle0\\right| R_x^{\\dagger}(\\theta) [I - \\frac{i}{2r}X] H [I+\\frac{i}{2r}X] R_x(\\theta)\\left|0\\right\\rangle \\big).\n",
    "\\tag{10}\n",
    "$$\n",
    "\n",
    "Using Euler's identity and knowing that $G$ has two unique eigenvalues, we can rewrite $U(\\theta)$ as $e^{-ia\\theta G} = I\\cos(r\\theta) - i\\frac{a}{r}G\\sin(r\\theta)$ [1]. Thus, we have $R_x(\\theta) = I\\cos(r\\theta) - i\\frac{1}{2r}X\\sin(r\\theta)$, we notice that \n",
    "\n",
    "$$\n",
    "R_x(\\frac{\\pi}{4r}) = I\\cos(\\frac{\\pi}{4}) - i\\frac{1}{2r}X\\sin(\\frac{\\pi}{4}) = \\frac{1}{\\sqrt2}(I-\\frac{i}{2r}X).\n",
    "\\tag{11}\n",
    "$$\n",
    "\n",
    "We can use the same method to get \n",
    "\n",
    "$$\n",
    "R_x(-\\frac{\\pi}{4r}) = \\frac{1}{\\sqrt2}(I+\\frac{i}{2r}X).\n",
    "\\tag{12}\n",
    "$$\n",
    "\n",
    "Thus, the equation can be simplified to\n",
    "\n",
    "$$\n",
    "O'(\\theta) = r\\big[ \\left\\langle0\\right|R_x^{\\dagger}(\\theta+\\frac{\\pi}{4r})HR_x(\\theta+\\frac{\\pi}{4r})\\left|0\\right\\rangle - \\left\\langle0\\right| R_x^{\\dagger}(\\theta-\\frac{\\pi}{4r}) H R_x(\\theta-\\frac{\\pi}{4r})\\left|0\\right\\rangle \\big],\n",
    "\\tag{13}\n",
    "$$\n",
    "\n",
    "and get the final formula,\n",
    "\n",
    "$$\n",
    "O'(\\theta) = r\\big[O(\\theta+\\frac{\\pi}{4r}) - O(\\theta-\\frac{\\pi}{4r}))\\big] = \\frac{1}{2}\\big[ O(\\theta + \\frac{\\pi}{2}) - O(\\theta - \\frac{\\pi}{2})\\big].\n",
    "\\tag{14}\n",
    "$$"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c460666d",
   "metadata": {},
   "source": [
    "### Paddle Quantum implementation\n",
    "Here we demonstrate how to use Paddle Quantum's built-in parameter shift method to calculate the gradient. Note: The built-in method currently does not support assigning an input state or running on a noisy circuit."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "18d78ca9",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "--Ry(5.690)----*----x----Ry(3.107)--\n",
      "               |    |               \n",
      "--Ry(2.521)----x----*----Ry(0.437)--\n",
      "                                    \n",
      "Gradient of this objective function is:  [-0.79156457  0.12584274 -0.2654174   0.7806864 ]\n"
     ]
    }
   ],
   "source": [
    "# We reuse the predefined Hamiltonian H and parameters of the circuit. \n",
    "# Again, be sure to mark stop_gradient=False when defining tensor for parameters of the circuit.\n",
    "\n",
    "# Constructing circuit U(theta)\n",
    "cir = U_theta(theta_tensor)\n",
    "print(cir)\n",
    "\n",
    "gradients = cir.param_shift_gradient(H)\n",
    "print(\"Gradient of this objective function is: \", gradients.numpy())"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2490a251",
   "metadata": {},
   "source": [
    "## Linear Combination of Unitary Gradients\n",
    "\n",
    "Building a parameterized circuit $U(\\theta)$ using Paddle Quantum requires many parameterized one-qubit and two-qubit gates like $R_x$ and $CR_x$. So we can rewrite $U(\\theta)$ as $U_{1}(\\theta_1)U_{2}(\\theta_2)\\cdots U_{m}(\\theta_m)$, where $U_i(\\theta_i)$ is one of the one-qubit and two-qubit gates and $m$ is the total number of parameterized gates in this circuit $U(\\theta)$. To get the gradient of an individual parameter, we consider this equation $\\frac{\\partial U(\\theta)}{\\partial \\theta_i}=U_{1}(\\theta_1)U_{2}(\\theta_2)\\cdots\\frac{\\partial U_i{(\\theta_i)}}{\\partial \\theta_i}\\cdots U_{m}(\\theta_m)$. We notice that as long as we know $\\frac{\\partial U_i{(\\theta_i)}}{\\partial \\theta_i}$ for all parameterized gates, we can get the gradients for all parameters easily [2]."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8c98fb33",
   "metadata": {},
   "source": [
    "### Single qubit gate gradient\n",
    "\n",
    "Let's consider single qubit gates first. We also take $R_x(\\theta)$ as an example. In the previous sections, we've already shown that $\\frac{\\partial R_x(\\theta)}{\\partial \\theta}=-i\\frac{1}{2}XR_x(\\theta)$, which can be easily constructed using a circuit. Let's try to implement it using Paddle Quantum."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "97435cf4",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original circuit: \n",
      "--Rx(1.047)--\n",
      "             \n",
      "The circuit for gradient of Rx: \n",
      "------------x----Rx(1.047)--\n",
      "            |               \n",
      "--H---SDG---*--------H------\n",
      "                            \n"
     ]
    }
   ],
   "source": [
    "# Construct the circuit with a single one-qubit gate Rx\n",
    "theta = paddle.to_tensor(np.pi / 3, 'float64')\n",
    "cir = UAnsatz(1)\n",
    "cir.rx(theta, 0)\n",
    "print('Original circuit: ')\n",
    "print(cir)\n",
    "\n",
    "print('The circuit for gradient of Rx: ')\n",
    "# The first parameter here is the index of the gate, the second parameter is the name of the gate\n",
    "print(cir.pauli_rotation_gate_partial(0, 'rx'))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e3d30aa3",
   "metadata": {},
   "source": [
    "It's a lot more complicated to do the same for a $u3(\\theta, \\phi, \\lambda)$ gate. But don't worry, we provide built-in methods for generating the circuits needed for calculating gradients of all parameterized single qubit gates in Paddle Quantum, i.e., $R_x$, $R_y$, $R_z$, and $u3$. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "6a7e5f3c",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original circuit: \n",
      "--U--\n",
      "     \n",
      "Circuits for gradient of u3: \n",
      "------------z----U--\n",
      "            |       \n",
      "--H---SDG---*----H--\n",
      "                    \n",
      "--Rz(4.706)---------y----Ry(3.292)----Rz(1.748)--\n",
      "                    |                            \n",
      "------H-------SDG---*--------H-------------------\n",
      "                                                 \n",
      "--Rz(4.706)----Ry(3.292)----z----Rz(1.748)--\n",
      "                            |               \n",
      "------H-----------SDG-------*--------H------\n",
      "                                            \n"
     ]
    }
   ],
   "source": [
    "cir = UAnsatz(1)\n",
    "theta = paddle.uniform([3], min=0.0, max=2*np.pi, dtype='float64')\n",
    "cir.u3(theta[0], theta[1], theta[2], 0)\n",
    "print('Original circuit: ')\n",
    "print(cir)\n",
    "\n",
    "# Since the u3 gate has three parameters, we need a total of three circuits. Each corresponds to one parameter.\n",
    "print('Circuits for gradient of u3: ')\n",
    "# The first parameter here is the index of the gate, the second parameter is the index of parameter\n",
    "print(cir.u3_partial(0, 0))\n",
    "print(cir.u3_partial(0, 1))\n",
    "print(cir.u3_partial(0, 2))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "95d11a2b",
   "metadata": {},
   "source": [
    "### Two-qubit gate gradient\n",
    "\n",
    "Paddle Quantum provides many two-qubit parameterized gates as well. They can be categorized into two types: one is control rotation gates like $CR_x$, the other is two-qubit rotation gates like $R_{xx}$. Circuits for gradients of two-qubit rotation gates are easy to construct. Let's take $R_{xx}$ as an example. Following the idea of one-qubit rotation gates, we first write it as $R_{xx}(\\theta)=e^{-i\\frac{\\theta}{2}X\\otimes X}$, then get the equation $\\frac{\\partial R_{xx}(\\theta)}{\\partial \\theta}=-i\\frac{1}{2}X\\otimes Xe^{-i\\frac{\\theta}{2}X\\otimes X}$, which can be converted into a circuit easily."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3fe81c6c",
   "metadata": {},
   "source": [
    "We need to be careful when calculating the gradients for control rotation gates. Usually, we will need two circuits for a control rotation gate with one parameter. For example, let's consider $CR_x(\\theta)$.\n",
    "\n",
    "$CR_x(\\theta)$ can be written as $\\left|0\\right>\\left<0\\right|\\otimes I + \\left|1\\right>\\left<1\\right|\\otimes R_x(\\theta)$, so its gradient is: \n",
    "\n",
    "$$\n",
    "\\frac{\\partial CR_x(\\theta)}{\\partial \\theta}=\\left|1\\right>\\left<1\\right|\\otimes \\frac{\\partial R_x(\\theta)}{\\partial \\theta}=-\\frac{i}{2}\\left|1\\right>\\left<1\\right|\\otimes Xe^{-i\\frac{\\theta}{2}X}.\n",
    "\\tag{15}\n",
    "$$\n",
    "\n",
    "However, this equation cannot be represented directly using one circuit. We need to use a tiny 'trick' here. Instead of using this formula directly, we decompose it into two terms \n",
    "\n",
    "$$\n",
    "\\frac{\\partial CR_x(\\theta)}{\\partial \\theta}=-\\frac{i}{4}(\\left|0\\right>\\left<0\\right|\\otimes I + \\left|1\\right>\\left<1\\right|\\otimes R_x(\\theta))I\\otimes X + \\frac{i}{4}(\\left|0\\right>\\left<0\\right|\\otimes I + \\left|1\\right>\\left<1\\right|\\otimes R_x(\\theta))Z\\otimes X.\n",
    "\\tag{16}\n",
    "$$ \n",
    "\n",
    "You can easily verify that this formula is equivalent to the former one. By doing so, we can use two circuits to compute the gradients for $CR_x$."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "23a999c4",
   "metadata": {},
   "source": [
    "As always, we provide built-in methods for calculating the gradients of all two-qubit parameterized gates in Paddle Quantum, i.e., $R_{xx}$, $R_{yy}$, $R_{zz}$, $CR_x$, $CR_y$, $CR_z$, $CU$. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "59733ca2",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original circuit: \n",
      "--*----Rzz(3.76)--------*------\n",
      "  |        |            |      \n",
      "--U----Rzz(3.76)----Ry(2.246)--\n",
      "                               \n",
      "The circuit for gradient of rzz: \n",
      "--*---------z---------Rzz(3.76)--------*------\n",
      "  |         |             |            |      \n",
      "--U---------|----z----Rzz(3.76)----Ry(2.246)--\n",
      "            |    |                            \n",
      "--H---SDG---*----*--------H-------------------\n",
      "                                              \n",
      "The circuit for gradient of cry: \n",
      "--*----Rzz(3.76)----y--------*------\n",
      "  |        |        |        |      \n",
      "--U----Rzz(3.76)----|----Ry(2.246)--\n",
      "                    |               \n",
      "--H-------SDG-------*--------H------\n",
      "                                    \n",
      "--*----Rzz(3.76)---------y--------*------\n",
      "  |        |             |        |      \n",
      "--U----Rzz(3.76)----z----|----Ry(2.246)--\n",
      "                    |    |               \n",
      "--H--------S--------*----*--------H------\n",
      "                                         \n"
     ]
    }
   ],
   "source": [
    "theta = paddle.uniform([5], min=0.0, max=2*np.pi, dtype='float64')\n",
    "cir = UAnsatz(2)\n",
    "cir.cu(theta[0], theta[1], theta[2], [0, 1])\n",
    "cir.rzz(theta[3], [0, 1])\n",
    "cir.cry(theta[4], [0, 1])\n",
    "print('Original circuit: ')\n",
    "print(cir)\n",
    "\n",
    "# The first parameter here is the index of the gate, the second parameter is the index of parameter\n",
    "# Since we have three parameters for cu gate, and we need two circuits for each parameter, we have a total of 6 circuits.\n",
    "# Circuits for gradients of cu:\n",
    "cu3_00 = cir.cu3_partial(0, 0)[0]\n",
    "cu3_01 = cir.cu3_partial(0, 0)[1]\n",
    "cu3_10 = cir.cu3_partial(0, 1)[0]\n",
    "cu3_11 = cir.cu3_partial(0, 1)[1]\n",
    "cu3_20 = cir.cu3_partial(0, 2)[0]\n",
    "cu3_21 = cir.cu3_partial(0, 2)[1]\n",
    "\n",
    "# The first parameter here is the index of the gate, the second parameter is the name of the gate\n",
    "print('The circuit for gradient of rzz: ')\n",
    "print(cir.pauli_rotation_gate_partial(1, 'RZZ_gate'))\n",
    "\n",
    "# The first parameter here is the index of the gate, the second parameter is the name of the gate\n",
    "print('The circuit for gradient of cry: ')\n",
    "print(cir.control_rotation_gate_partial(2, 'cry')[0])\n",
    "print(cir.control_rotation_gate_partial(2, 'cry')[1])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9d264a86",
   "metadata": {},
   "source": [
    "Now that we have all individual circuits prepared for calculating gradients, the next step is to get the exact value of the gradients. How to do so? We need to plug these circuits into our objective function, then the results are our desired gradients. For gates like $CR_x$, we will take the mean of the two circuits to be the gradient, as indicated in the analytical formula. We also provide a built-in method (Note: The built-in method currently does not support assigning an input state or running on a noisy circuit):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "02e8ea72",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "--U----*----x----Ry(0.667)--\n",
      "       |    |               \n",
      "--U----x----*----Ry(4.807)--\n",
      "                            \n",
      "Gradient of this objective function is:  [ 0.         -0.76471634  0.00700639  0.         -0.29062181 -0.01701887\n",
      " -0.07729092  0.7766131 ]\n"
     ]
    }
   ],
   "source": [
    "# Randomly generate parameters for our circuit\n",
    "theta = paddle.uniform(shape=[8], dtype='float64', min=0.0, max=np.pi * 2)\n",
    "theta.stop_gradient = False\n",
    "\n",
    "# Construct circuit of U(theta)\n",
    "cir = UAnsatz(2)\n",
    "cir.complex_entangled_layer(theta[:6], 1)\n",
    "cir.ry(theta=theta[6], which_qubit=0)\n",
    "cir.ry(theta=theta[7], which_qubit=1)\n",
    "cir.run_state_vector()\n",
    "print(cir)\n",
    "\n",
    "# Calculate gradient using our built-in method\n",
    "# We pass in our Hamiltonian H used in the objective function\n",
    "gradient = cir.linear_combinations_gradient(H, shots=0)\n",
    "print(\"Gradient of this objective function is: \", gradient.numpy())"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d7fcdbdf",
   "metadata": {},
   "source": [
    "## Application: Simulating VQE with Paddle Quantum\n",
    "\n",
    "Variational Quantum Eigensolver (VQE) [3] is designed to find the ground state energy of a given molecular Hamiltonian using variational quantum circuits. Interested readers can find more details from the previous tutorial [VQE](../quantum_simulation/VQE_EN.ipynb).\n",
    "\n",
    "We will demonstrate how to use VQE to find the ground state energy for the Hamiltonian of hydrogen molecule $H_2$. In the process, we will use the methods introduced above to calculate the gradient."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8e9d6e7b",
   "metadata": {},
   "source": [
    "### Using Paddle's Optimizer\n",
    "\n",
    "First, we will use Paddle's optimizer Adam to run our example. We can choose using either finite difference method or parameter-shift method to calculate gradient."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "cce057c2",
   "metadata": {},
   "outputs": [],
   "source": [
    "from paddle_quantum.VQE.chemistrysub import H2_generator\n",
    "from paddle_quantum.expecval import ExpecVal\n",
    "\n",
    "# Set up our Hamiltonian H\n",
    "pauli_str, N = H2_generator()\n",
    "H = Hamiltonian(pauli_str)\n",
    "\n",
    "# Hyper-parameters\n",
    "ITR = 80  # Set the number of optimization iterations\n",
    "LR = 0.4   # Set the learning rate\n",
    "D = 2      # Set the depth of the repetitive calculation module in QNN\n",
    "\n",
    "def U_theta(theta, Hamiltonian, N, D):\n",
    "    \"\"\"\n",
    "    Quantum Neural Network\n",
    "    \"\"\"\n",
    "    # Initialize the quantum neural network according to the number of qubits N\n",
    "    cir = UAnsatz(N)\n",
    "\n",
    "    # Built-in {R_y + CNOT} circuit template\n",
    "    theta = paddle.reshape(theta, [D+1, N, 1])\n",
    "    cir.real_entangled_layer(theta[:D], D)\n",
    "\n",
    "    # Lay R_y gates in the last row\n",
    "    for i in range(N):\n",
    "        cir.ry(theta=theta[D][i][0], which_qubit=i)\n",
    "\n",
    "    # The quantum neural network acts on the default initial state |0...0>\n",
    "    cir.run_state_vector()\n",
    "    \n",
    "    return cir"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c0cdee79",
   "metadata": {},
   "source": [
    "Here in our forward propagation mechanism, we use the updated parameters to calculate gradient using parameter-shift rule, and calculate the expectation value. You can change the method to 'finite_diff' if you'd like to try it out using finite difference method. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "0d172081",
   "metadata": {},
   "outputs": [],
   "source": [
    "class StateNet(paddle.nn.Layer):\n",
    "\n",
    "    def __init__(self, cir):\n",
    "        super(StateNet, self).__init__()\n",
    "        \n",
    "        self.cir = cir\n",
    "        params = cir.get_param()\n",
    "        \n",
    "        # Assign the theta parameter list to be the trainable parameter list of the circuit\n",
    "        self.theta = self.create_parameter(shape=[len(params)], \n",
    "                                           default_initializer=paddle.nn.initializer.Assign(params),\n",
    "                                           dtype='float32', is_bias=False)\n",
    "        \n",
    "    # Define loss function and forward propagation mechanism\n",
    "    def forward(self):\n",
    "        # Calculate the loss function/expectation value using Parameter-shift rule to calculate gradient\n",
    "        loss = ExpecVal.apply(self.cir, self.theta.cast('float64'), 'param_shift', H, shots=0)\n",
    "        \n",
    "        return loss, self.cir"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "0b8235c9",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "iter: 20 loss: -1.1114\n",
      "iter: 20 Ground state energy: -1.1114 Ha\n",
      "iter: 40 loss: -1.1316\n",
      "iter: 40 Ground state energy: -1.1316 Ha\n",
      "iter: 60 loss: -1.1357\n",
      "iter: 60 Ground state energy: -1.1357 Ha\n",
      "iter: 80 loss: -1.1361\n",
      "iter: 80 Ground state energy: -1.1361 Ha\n",
      "\n",
      "The trained circuit:\n",
      "--Ry(6.282)----*--------------x----Ry(6.289)----*--------------x----Ry(3.148)--\n",
      "               |              |                 |              |               \n",
      "--Ry(3.138)----x----*---------|----Ry(0.207)----x----*---------|----Ry(3.142)--\n",
      "                    |         |                      |         |               \n",
      "--Ry(6.278)---------x----*----|----Ry(0.001)---------x----*----|----Ry(3.143)--\n",
      "                         |    |                           |    |               \n",
      "--Ry(0.001)--------------x----*----Ry(3.156)--------------x----*----Ry(3.154)--\n",
      "                                                                               \n",
      "\n",
      "Ground state energy obtained:  [-1.13609609] Ha\n",
      "Actual ground state energy:  -1.13618 Ha\n"
     ]
    }
   ],
   "source": [
    "# Initialize the theta parameter list and fill the initial value with a uniform distribution of [0, 2*pi]\n",
    "theta = paddle.to_tensor(np.random.uniform(0.0, 2*np.pi, (D+1) * N), stop_gradient=False)\n",
    "\n",
    "# Initialize the circuit\n",
    "cir = U_theta(theta, H, N, D)\n",
    "\n",
    "# Determine the parameter dimension of the network\n",
    "net = StateNet(cir)\n",
    "\n",
    "# Generally speaking, we use Adam optimizer to obtain relatively good convergence,\n",
    "# You can change it to SGD or RMS prop.\n",
    "opt = paddle.optimizer.Adam(learning_rate=LR, parameters=net.parameters())\n",
    "\n",
    "# Record optimization results\n",
    "summary_iter, summary_loss = [], []\n",
    "\n",
    "# Optimization loop\n",
    "for itr in range(1, ITR + 1):\n",
    "\n",
    "    # Forward propagation to calculate loss function\n",
    "    loss, cir = net()\n",
    "\n",
    "    # Use back propagation to minimize the loss function\n",
    "    loss.backward()\n",
    "    opt.minimize(loss)\n",
    "    opt.clear_grad()\n",
    "\n",
    "    # Record optimization results\n",
    "    summary_loss.append(loss.numpy())\n",
    "    summary_iter.append(itr)\n",
    "\n",
    "    # Print result\n",
    "    if itr % 20 == 0:\n",
    "        print(\"iter:\", itr, \"loss:\", \"%.4f\" % loss.numpy())\n",
    "        print(\"iter:\", itr, \"Ground state energy:\", \"%.4f Ha\" \n",
    "                                            % loss.numpy())\n",
    "    if itr == ITR:\n",
    "        print(\"\\nThe trained circuit:\")\n",
    "        print(cir)\n",
    "\n",
    "print('\\nGround state energy obtained: ', summary_loss[-1], \"Ha\")\n",
    "print('Actual ground state energy: ', -1.13618, \"Ha\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2e672bed",
   "metadata": {},
   "source": [
    "We can see that the ground state energy we obtained is close to the theoretical value."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8d58ebab",
   "metadata": {},
   "source": [
    "### Using SciPy's Optimizer\n",
    "\n",
    "We will also demonstrate how to use SciPy's optimizer to run VQE easily with Paddle Quantum. For this example, we will use Conjugate Gradient (CG) optimizer along with linear combination method to find the ground state energy of our Hamiltonian. \n",
    "\n",
    "Other SciPy methods we support include Newton-CG, Powell, and SLSQP."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "7f7ef671",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "loss:  [-0.91176578]\n",
      "loss:  [-1.03555093]\n",
      "loss:  [-1.11965221]\n",
      "loss:  [-1.13435502]\n",
      "loss:  [-1.13577104]\n",
      "loss:  [-1.13615947]\n",
      "loss:  [-1.13618601]\n",
      "loss:  [-1.13618942]\n",
      "loss:  [-1.13618945]\n",
      "loss:  [-1.13618945]\n",
      "loss:  [-1.13618945]\n",
      "Optimization terminated successfully.\n",
      "Actual ground state energy:  -1.13618 Ha\n"
     ]
    }
   ],
   "source": [
    "from paddle_quantum.optimizer import ConjugateGradient\n",
    "\n",
    "# Initialize the circuit\n",
    "cir = U_theta(theta, H, N, D)\n",
    "\n",
    "optimizer = ConjugateGradient(cir, H, shots=0, grad_func_name='linear_comb')\n",
    "optimizer.minimize(iterations=80)\n",
    "print('Actual ground state energy: ', -1.13618, \"Ha\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1ad4927e",
   "metadata": {},
   "source": [
    "## Conclusion\n",
    "\n",
    "As you can see, finite-difference and parameter-shift methods have similar forms - both of them require two function evaluations per parameter. The benefits of these methods are that the gradients can be calculated without knowing much about the circuit or the objective function. We can treat them as a black box and get the gradient just by feeding in different parameters. Our preferred choice between those two is parameter-shift method because its result is an analytical gradient, while finite difference method can only get an estimation of the gradient. However, parameter-shift only applies to $U(\\theta)$ that can be generated by $G$ that has two distinct eigenvalues:  $U(\\theta) = e^{-ia\\theta G}$ or can be decomposed into a product of gates in this form.\n",
    "\n",
    "Using linear combination of unitary gates to calculate the gradients of a given circuit is probably the most straightforward analytical method. By differentiating unitary gates under their mathematical forms, we can use circuits to represent the resulting formula. The number of circuits required is proportional to the number of parameters in the original circuit as the other two methods. We can even calculate gradients by constructing only one single circuit for simple gates like $R_x$, $R_{xx}$ and etc. However, note that we will use an ancilla qubit in this method. Moreover, you might have noticed that this method takes a long time to run on complex circuits. That's because as the number of qubits increases, the number of circuits used to represent the first order differentiation of a single multi-qubit gate also increases. "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "26a8f5fd",
   "metadata": {},
   "source": [
    "_______\n",
    "\n",
    "## References\n",
    "\n",
    "[1] Crooks, Gavin E. \"Gradients of parameterized quantum gates using the parameter-shift rule and gate decomposition.\" [arXiv preprint arXiv:1905.13311 (2019)](https://arxiv.org/abs/1905.13311).\n",
    "\n",
    "[2] Somma, Rolando, et al. \"Simulating physical phenomena by quantum networks.\" [Physical Review A 65.4 (2002): 042323](https://arxiv.org/abs/quant-ph/0108146).\n",
    "\n",
    "[3] Peruzzo, Alberto, et al. \"A variational eigenvalue solver on a photonic quantum processor.\" [Nature communications 5.1 (2014): 1-7](https://www.nature.com/articles/ncomms5213).\n",
    "\n",
    "[4] Schuld, Maria, et al. \"Evaluating analytic gradients on quantum hardware.\" [Physical Review A 99.3 (2019): 032331](https://arxiv.org/abs/1811.11184)."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.7.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}