Merge pull request #116 from Langhaarzombie/feature/ntbk_jax

QuTiPv5 Paper Notebook: JAX

Author: nwlambert

tutorials-v5/miscellaneous/v5_paper-jax.md

@@ -0,0 +1,335 @@
+---
+jupyter:
+  jupytext:
+    text_representation:
+      extension: .md
+      format_name: markdown
+      format_version: '1.3'
+      jupytext_version: 1.13.8
+  kernelspec:
+    display_name: Python 3 (ipykernel)
+    language: python
+    name: python3
+---
+
+# QuTiPv5 Paper Example: QuTiP-JAX with mesolve and auto-differnetiation
+
+Authors: Maximilian Meyer-Mölleringhof ([email protected]), Rochisha Agarwal ([email protected]), Neill Lambert ([email protected])
+
+For many years now, GPUs have been a fundamental tool for accelerating numerical tasks.
+Today, many libraries enable off-the-shelf methods to leverage GPUs' potential to speed up costly calculations.
+QuTiP’s flexible data layer can directly be used with many such libraries and thereby drastically reduce computation time.
+Despite a big variety of frameworks, in connection to QuTiP, development has centered on the QuTiP-JAX integration [\[1\]](#References) due to JAX's robust auto-differentiation features and widespread adoption in machine learning.
+
+In these examples we illustrate how JAX naturally integrates into QuTiP v5 [\[2\]](#References) using the QuTiP-JAX package.
+As a simple first example, we look at a one-dimensional spin chain and how we might employ `mesolve()` and JAX to solve the related master equation.
+In the second part, we focus on the auto-differentiation capabilities.
+For this we first consider the counting statistics of an open quantum system connected to an environment.
+Lastly, we look at a driven qubit system and computing the gradient of its state population in respect to the drive's frequency using `mcsolve()`.
+
+## Introduction
+
+In addition to the standard QuTiP package, in order to use QuTiP-JAX, the JAX package needs to be installed.
+This package also comes with `jax.numpy` which mirrors all of `numpy`s functionality for seamless integration with JAX.
+
+```python
+import jax.numpy as jnp
+import matplotlib.pyplot as plt
+import qutip_jax as qj
+from diffrax import PIDController, Tsit5
+from jax import default_device, devices, grad, jacfwd, jacrev, jit
+from qutip import (CoreOptions, about, basis, destroy, lindblad_dissipator,
+                   liouvillian, mcsolve, mesolve, projection, qeye, settings,
+                   sigmam, sigmax, sigmay, sigmaz, spost, spre, sprepost,
+                   steadystate, tensor)
+
+%matplotlib inline
+```
+
+An immediate effect of importing `qutip_jax` is the availability of the data layer formats `jax` and `jax_dia`.
+They allow for dense (`jax`) and custom sparse (`jax_dia`) formats.
+
+```python
+print(qeye(3, dtype="jax").dtype.__name__)
+print(qeye(3, dtype="jaxdia").dtype.__name__)
+```
+
+In order to use the JAX data layer also within the master equation solver, there are two settings we can choose.
+First, is simply adding `method: diffrax` to the options parameter of the solver.
+Second, is to use the `qutip_jax.set_as_default()` method.
+It automatically switches all data data types to JAX compatible versions and sets the default solver method to `diffrax`.
+
+```python
+qj.set_as_default()
+```
+
+To revert this setting, we can set the function parameter `revert = True`.
+
+```python
+# qj.set_as_default(revert = True)
+```
+
+## Using the GPU with JAX and Diffrax - 1D Ising Spin Chain
+
+Before diving into the example, it is worth noting here that GPU acceleration depends heavily on the type of problem.
+GPUs are good at parallelizing many small matrix-vector operations, such as integrating small systems across multiple parameters or simulating quantum circuits with repeated small matrix operations.
+For a single ODE involving large matrices, the advantages are less straightforward since ODE solvers are inherently sequential.
+However, as it is illustrated in the QuTiP v5 paper [\[2\]](#References), there is a cross-over point at which using JAX becomes beneficial.
+
+### 1D Ising Spin Chain
+
+To illustrate the usage of QuTiP-JAX, we look at the one-dimensional spin chain with the Hamiltonian
+
+$H = \sum_{i=1}^N g_0 \sigma_z^{(n)} - \sum_{n=1}^{N-1} J_0 \sigma_x^{(n)} \sigma_x^{(n+1)}$.
+
+We hereby consider $N$ spins that share an energy splitting of $g_0$ and have a coupling strength $J_0$.
+The end of the chain connects to an environment described by a Lindbladian dissipator which we model with the collapse operator $\sigma_x^{(N-1)}$ and coupling rate $\gamma$.
+
+As part of the [QuTiPv5 paper](#References), we see an extensive study on the computation time depending on the dimensionality $N$.
+In this example we cannot replicate the performance of a supercomputer of course, so we rather focus on the correct implementation to solve the Lindblad equation for this system using JAX and `mesolve()`.
+
+```python
+# system parameters
+N = 4  # number of spins
+g0 = 1  # energy splitting
+J0 = 1.4  # coupling strength
+gamma = 0.1  # dissipation rate
+
+# simulation parameters
+tlist = jnp.linspace(0, 5, 100)
+opt = {
+    "normalize_output": False,
+    "store_states": True,
+    "method": "diffrax",
+    "stepsize_controller": PIDController(
+        rtol=settings.core["rtol"], atol=settings.core["atol"]
+    ),
+    "solver": Tsit5(),
+}
+```
+
+```python
+with CoreOptions(default_dtype="jaxdia"):
+    # Operators for individual qubits
+    sx_list, sy_list, sz_list = [], [], []
+    for i in range(N):
+        op_list = [qeye(2)] * N
+        op_list[i] = sigmax()
+        sx_list.append(tensor(op_list))
+        op_list[i] = sigmay()
+        sy_list.append(tensor(op_list))
+        op_list[i] = sigmaz()
+        sz_list.append(tensor(op_list))
+
+    # Hamiltonian - Energy splitting terms
+    H = 0.0
+    for i in range(N):
+        H += g0 * sz_list[i]
+
+    # Interaction terms
+    for n in range(N - 1):
+        H += -J0 * sx_list[n] * sx_list[n + 1]
+
+    # Collapse operator acting locally on single spin
+    c_ops = [gamma * sx_list[N - 1]]
+
+    # Initial state
+    state_list = [basis(2, 1)] * (N - 1)
+    state_list.append(basis(2, 0))
+    psi0 = tensor(state_list)
+
+    result = mesolve(H, psi0, tlist, c_ops, e_ops=sz_list, options=opt)
+```
+
+```python
+for i, s in enumerate(result.expect):
+    plt.plot(tlist, s, label=rf"$n = {i+1}$")
+plt.xlabel("Time")
+plt.ylabel(r"$\langle \sigma^{(n)}_z \rangle$")
+plt.legend()
+plt.show()
+```
+
+## Auto-Differentiation
+
+We have seen in the previous example how the new JAX data-layer in QuTiP works.
+On top of that, JAX adds the features of auto-differentiation.
+To compute derivatives, it is often numerical approximations (e.g., finite difference method) that need to be employed.
+Especially for higher order derivatives, these methods can turn into costly and inaccurate calculations.
+
+Auto-differentiation, on the other hand, exploits the chain rule to compute such derivatives.
+The idea is that any numerical function can be expressed by elementary analytical functions and operations.
+Consequently, using the chain rule, the derivatives of almost any higher-level function become accessible.
+
+Although there are many applications for this technique, in this chapter we want to focus on two examples where auto-differentiation becomes relevant.
+
+### Statistics of Excitations between Quantum System and Environment
+
+We consider an open quantum system that is in contact with an evironment via a single jump operator.
+Additionally, we have a measurement device that tracks the flow of excitations between the system and the environment.
+The probability distribution that describes the number of such exchanged excitations $n$ in a certain time $t$ is called the full counting statistics and denoted by $P_n(t)$.
+This statistics is a defining property that allows to derive many experimental observables like shot noise or current.
+
+For the example here, we can calculate this statistics by using a modified version of the density operator and Lindblad master equation.
+We introduce the *tilted* density operator $G(z,t) = \sum_n e^{zn} \rho^n (t)$ with $\rho^n(t)$ being the density operator of the system conditioned on $n$ exchanges by time $t$, so $\text{Tr}[\rho^n(t)] = P_n(t)$.
+The master equation for this operator, including the jump operator $C$, is then given as
+
+$\dot{G}(z,t) = -\dfrac{i}{\hbar} [H(t), G(z,t)] + \dfrac{1}{2} [2 e^z C \rho(t)C^\dagger - \rho C^\dagger C - C^\dagger C \rho(t)]$.
+
+We see that for $z = 0$, this master equation becomes the regular Lindblad equation and $G(0,t) = \rho(t)$.
+However, it also allows us to describe the counting statistics through its derivatives
+
+$\langle n^m \rangle (t) = \sum_n n^m \text{Tr} [\rho^n (t)] = \dfrac{d^m}{dz^m} \text{Tr} [G(z,t)]|_{z=0}$.
+
+These derivatives are precisely where the auto-differention by JAX finds its application for us.
+
+```python
+# system parameters
+ed = 1
+GammaL = 1
+GammaR = 1
+
+# simulation parameters
+options = {
+    "method": "diffrax",
+    "normalize_output": False,
+    "stepsize_controller": PIDController(rtol=1e-7, atol=1e-7),
+    "solver": Tsit5(scan_kind="bounded"),
+    "progress_bar": False,
+}
+```
+
+When working with JAX you can choose the type of device / processor to be used.
+In our case, we will resort to the CPU since this is a simple Jupyter Notebook.
+However, when running this on your machine, you can opt for using your GPU by simpy changing the argument below.
+
+```python
+with default_device(devices("cpu")[0]):
+    with CoreOptions(default_dtype="jaxdia"):
+        d = destroy(2)
+        H = ed * d.dag() * d
+        c_op_L = jnp.sqrt(GammaL) * d.dag()
+        c_op_R = jnp.sqrt(GammaR) * d
+
+        L0 = (
+            liouvillian(H)
+            + lindblad_dissipator(c_op_L)
+            - 0.5 * spre(c_op_R.dag() * c_op_R)
+            - 0.5 * spost(c_op_R.dag() * c_op_R)
+        )
+        L1 = sprepost(c_op_R, c_op_R.dag())
+
+        rho0 = steadystate(L0 + L1)
+
+        def rhoz(t, z):
+            L = L0 + jnp.exp(z) * L1  # jump term
+            tlist = jnp.linspace(0, t, 50)
+            result = mesolve(L, rho0, tlist, options=options)
+            return result.final_state.tr()
+
+        # first derivative
+        drhozdz = jacrev(rhoz, argnums=1)
+        # second derivative
+        d2rhozdz = jacfwd(drhozdz, argnums=1)
+```
+
+```python
+tf = 100
+Itest = GammaL * GammaR / (GammaL + GammaR)
+shottest = Itest * (1 - 2 * GammaL * GammaR / (GammaL + GammaR) ** 2)
+ncurr = drhozdz(tf, 0.0) / tf
+nshot = (d2rhozdz(tf, 0.0) - drhozdz(tf, 0.0) ** 2) / tf
+
+print("===== RESULTS =====")
+print("Analytical current", Itest)
+print("Numerical current", ncurr)
+print("Analytical shot noise (2nd cumulant)", shottest)
+print("Numerical shot noise (2nd cumulant)", nshot)
+```
+
+### Driven One Qubit System & Frequency Optimization
+
+As a second example for auto differentiation, we consider the driven Rabi model, which is given by the time-dependent Hamiltonian
+
+$H(t) = \dfrac{\hbar \omega_0}{2} \sigma_z + \dfrac{\hbar \Omega}{2} \cos (\omega t) \sigma_x$
+
+with the energy splitting $\omega_0$, $\Omega$ as the Rabi frequency, the drive frequency $\omega$ and $\sigma_{x/z}$ are Pauli matrices.
+When we add dissipation to the system, the dynamics is given by the Lindblad master equation, which introduces collapse operator $C = \sqrt{\gamma} \sigma_-$ to describe energy relaxation.
+
+For this example, we are interested in the population of the excited state of the qubit
+
+$P_e(t) = \bra{e} \rho(t) \ket{e}$
+
+and its gradient with respect to the frequency $\omega$.
+
+We want to optimize this quantity by adjusting the drive frequency $\omega$.
+To achieve this, we compute the gradient of $P_e(t)$ in respect to $\omega$ by using JAX's auto-differentiation tools and QuTiP's `mcsolve()`.
+
+```python
+# system parameters
+gamma = 0.1  # dissipation rate
+```
+
+```python
+# time dependent drive
+@jit
+def driving_coeff(t, omega):
+    return jnp.cos(omega * t)
+
+
+# system Hamiltonian
+def setup_system():
+    H_0 = sigmaz()
+    H_1 = sigmax()
+    H = [H_0, [H_1, driving_coeff]]
+    return H
+```
+
+```python
+# simulation parameters
+psi0 = basis(2, 0)
+tlist = jnp.linspace(0.0, 10.0, 100)
+c_ops = [jnp.sqrt(gamma) * sigmam()]
+e_ops = [projection(2, 1, 1)]
+```
+
+```python
+# Objective function: returns final exc. state population
+def f(omega):
+    H = setup_system()
+    arg = {"omega": omega}
+    result = mcsolve(H, psi0, tlist, c_ops, e_ops=e_ops, ntraj=100, args=arg)
+    return result.expect[0][-1]
+```
+
+```python
+# Gradient of the excited state population with respect to omega
+grad_f = grad(f)(2.0)
+```
+
+```python
+print(grad_f)
+```
+
+## References
+
+
+
+
+[1] [QuTiP-JAX](https://github.com/qutip/qutip-jax)
+
+[2] [QuTiP 5: The Quantum Toolbox in Python](https://arxiv.org/abs/2412.04705)
+
+
+## About
+
+```python
+about()
+```
+
+## Testing
+
+```python
+assert jnp.isclose(Itest, ncurr, rtol=1e-5), "Current calc. deviates"
+assert jnp.isclose(shottest, nshot, rtol=1e-1), "Shot noise calc. deviates."
+```