6. Ordinary Differential Equations#
A differential equation is an equation that includes a derivative \(f\frac{df(x)}{dx}\) of a function \(f(x)\).
The derivative of a function \(f(x)\) means its slope, mathematically defined as the limit: $\( \frac{df(x)}{dx} = \lim_{\Delta x \rightarrow 0} \frac{f(x+\Delta x) - f(x)}{\Delta x} \)$
If the independent variable \(x\) is single, such as time, it is called an ordinary differential equation (ODE).
If there are multiple independent variables, such as space and time, the equation includes partial derivatives and called a partial differential equation (PDE).
Here we consider ODEs of the form $\( \frac{dy}{dt} = f(y, t) \)\( which describes the temporal dynamics of a varibale \)y\( over time \)t$. It is also called a continuous-time dynamical system.
Finding the variable as an explicit function of time \(y(t)\) is called solving or integrating the ODE. When it is done numerically, it is aslo called simulating.
References:
Stephen Wiggins: Introduction to Applied Nonlinear Dynamical Systems and Chaos, 2nd ed., Springer (2003).
Scipy Lecture Notes: Section 6.7 Numerical Integration
Here are some basic properties of the deriative \(f'(x)=\frac{df(x)}{dx}\)
The derivative of exponential is itself: \( \frac{d e^x}{dx} = e^x \)
Derivatives of sine and cosine: \(\sin'(x)=\cos(x)\), \(\cos'(x)=-\sin(x)\)
Linearity: \( \frac{d af(x)+bg(x)}{dx} = af'(x)+bg'(x) \)
Chain rule: \( \frac{d f(g(x))}{dx} = f'(g(x))g'(x) \)
Analytic Solutions#
Solving a differential equation is an inverse problem of differentiation, for which analytic solution may not be available.
The simplest case where analytic solutions are available is linear differential equations $\( \frac{dy}{dt} = A y \)\( where \)y\( is a real variable or a real vector, and \)A$ is a constant coefficient or a matrix.
Linear ODEs#
In general, a differential equation can have multiple solutions. For example, for a scalar linear ODE $\( \frac{dy}{dt} = a y, \)\( the solution is given by \)\( y(t) = C e^{at}, \)\( where \)C$ can be any real value.
Verify that by differentiating both sides of the equaiton above.
When the value of \(y\) at a certain time is specified, the solution becomes unique.
For example, by specifying an initial condition
$\( y(0)=3, \)\(
from \)e^{a0}=e^0=1\(, we have \)C=3\( and a particular solution \)\( y(t)=3e^{at}. \)$
Another example is a second-order linear ODE $\( \frac{d^2y}{dt^2} = -a^2 y. \)$
The second-order derivative \(\frac{d^2y}{dt^2}\) of a function \(y(t)\) is the devative of the derivative $\( \frac{d^2y}{dt^2} = \frac{d \frac{dy}{dt}}{dt} = \frac{d}{dt} \frac{dy}{dt} \)$ representing the curvature.
In this case, the solution is given by $\( y(t) = C_1 \sin at + C_2 \cos at \)\( where \)C_1\( and \)C_2\( are determined by spedifying \)y\( and \)dy/dt$ at certain time.
Also verify the above.
Analytically solvable ODEs#
Other cases where analytic solutions are well known are:
Linear time-varying coefficient: $\( \frac{dy}{dt} = a(t)y \)$
\(at\) replaced by time integral $\( y(t) = C e^{\int a(t)dt} \)$
Linear time-varying input: $\(\frac{dy}{dt} = a(t)y + b(t)\)$
using the above solution \(y_0(t)\) for \(b(t)=0\) $\( y(t) = Cy_0(t) + y_0(t) \int \frac{b(t)}{y_0(t)}dt \)$
Separation of variables: $\(\frac{dy}{dt} = a(y)b(t)\)$
two related integrals $\(\int \frac{1}{a(y)}dy = \int b(t)dt + C \)$
Other cases that can be reduced to above by change of variables, etc.
You can use dsolve()
function of SymPy
to find some analitic solutions. See Scipy Tutorial, Chapter 16 if you are interested.
Euler Method#
The most basic way of sovling an ODE numerically is Euler Method.
For an ODE
$\( \frac{dy}{dt} = f(y,t) \)\(
with an initial condition \)y(t_0)=y_0\(, the solution is iteratively approximated by
\)\( y(t+\Delta t) = y(t) + f(y,t) \Delta t \)\(
with a small time step \)\Delta t$.
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
def euler(f, y0, dt, n, *args):
"""f: righthand side of ODE dy/dt=f(y,t)
y0: initial condition y(0)=y0
dt: time step
n: iteratons
args: parameter for f(y,t,*args)"""
d = np.array([y0]).size ## state dimension
y = np.zeros((n+1, d))
y[0] = y0
t = 0
for k in range(n):
y[k+1] = y[k] + f(y[k], t, *args)*dt
t = t + dt
return y
Let us test this with a first-order linear ODE.
def first(y, t, a):
"""first-order linear ODE dy/dt = a*y"""
return a*y
y = euler(first, 1, 0.1, 100, -1)
plt.plot(y, '.-');
![_images/6a3b18d1d12f33fd7722d90041e570c8a286bd95c3f321c8116a3a475361224c.png](_images/6a3b18d1d12f33fd7722d90041e570c8a286bd95c3f321c8116a3a475361224c.png)
A second-order ODE $\( \frac{d^2y}{dt^2} = a_2 \frac{dy}{dt} + a_1 y + a_0 \)\( can be converted into a first-order ODE of a 2-dimensional vector \){\bf y} = \pmatrix{y_1 \ y_2} =\pmatrix{y \ \frac{dy}{dt}}\( as \)\( \frac{dy_1}{dt} = y_2 \)\( \)\( \frac{dy_2}{dt} = a_2 y_2 + a_1 y_1 + a_0 \)\( or in a vector-matrix form \)\( \frac{d}{dt}{\bf y} = A{\bf y} + {\bf b}\)\( where \)\( A = \pmatrix{0 & 1 \\ a_1 & a_2} \mbox{ and } {\bf b} = \pmatrix{0 \\ a_0}\)$
def second(y, t, a):
"""second-order linear ODE """
y1, y2 = y
return np.array([y2, a[2]*y2 + a[1]*y1 + a[0]])
y = euler(second, [0, 1], 0.01, 2000, [0, -1, 0])
plt.plot(y);
![_images/95efd4035f272243682b3842a7f486e7bc453b2ba9ecda26e100ca955a1bb496.png](_images/95efd4035f272243682b3842a7f486e7bc453b2ba9ecda26e100ca955a1bb496.png)
y = euler(second, [0, 1], 0.01, 2000, [0, -1, -0.5])
plt.plot(y);
![_images/90e4be2342f3456c0f3c544487a47a82ec686ac1819525ed637afab505505cca.png](_images/90e4be2342f3456c0f3c544487a47a82ec686ac1819525ed637afab505505cca.png)
# phase plot
plt.plot(y[:,0], y[:,1]);
![_images/cfa85e8729c45d30f735595027ab3847cf924dc3b8d5057339d6463481e974f8.png](_images/cfa85e8729c45d30f735595027ab3847cf924dc3b8d5057339d6463481e974f8.png)
Let us see how the time step affects the accuracy of the solution.
steps = [0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5]
tend = 2
a = -5
for dt in steps:
y = euler(first, 1, dt, int(tend/dt), a)
plt.plot(np.linspace(0, tend, len(y)), y)
![_images/86c73b4d3319b4b5c434e33b1565d39b56181c751071e601ec5b16a8ccbe0b92.png](_images/86c73b4d3319b4b5c434e33b1565d39b56181c751071e601ec5b16a8ccbe0b92.png)
Scipy’s Integrate package#
To avoid numerical instability and to improve accuracy and efficiency, there are advanced methods for ODE solutions.
Backward Euler method: solve $\( y(t+\Delta t) = y(t) + f(y(t+\Delta t)) \Delta t\)$
Mixture of forward and backward (Crank-Nicolson): $\( y(t+\Delta t) = y(t) + \frac{1}{2}\{f(y(t))+f(y(t+\Delta t))\} \Delta t\)$
Runge-Kutta method: minimize higher-order erros by Taylor expansion $\( y(t+\Delta t) = y(t) + f(y(t))\Delta t + \frac{1}{2} \frac{d}{dt}f(y(t))\Delta t^2 + ...\)$
Adative time step: adjust \(\Delta t\) depending on the scale of \(\frac{df(y(t))}{dt}\).
The implementation and choice of these methods require a good expertise, but fortunately scipy
includes integrate
package which has been well tested and optimized.
odeint()
implements automatic method switching and time step adaptation.
ode()
is a class interface for multiple methods.
from scipy.integrate import odeint
help(odeint)
Help on function odeint in module scipy.integrate._odepack_py:
odeint(func, y0, t, args=(), Dfun=None, col_deriv=0, full_output=0, ml=None, mu=None, rtol=None, atol=None, tcrit=None, h0=0.0, hmax=0.0, hmin=0.0, ixpr=0, mxstep=0, mxhnil=0, mxordn=12, mxords=5, printmessg=0, tfirst=False)
Integrate a system of ordinary differential equations.
.. note:: For new code, use `scipy.integrate.solve_ivp` to solve a
differential equation.
Solve a system of ordinary differential equations using lsoda from the
FORTRAN library odepack.
Solves the initial value problem for stiff or non-stiff systems
of first order ode-s::
dy/dt = func(y, t, ...) [or func(t, y, ...)]
where y can be a vector.
.. note:: By default, the required order of the first two arguments of
`func` are in the opposite order of the arguments in the system
definition function used by the `scipy.integrate.ode` class and
the function `scipy.integrate.solve_ivp`. To use a function with
the signature ``func(t, y, ...)``, the argument `tfirst` must be
set to ``True``.
Parameters
----------
func : callable(y, t, ...) or callable(t, y, ...)
Computes the derivative of y at t.
If the signature is ``callable(t, y, ...)``, then the argument
`tfirst` must be set ``True``.
y0 : array
Initial condition on y (can be a vector).
t : array
A sequence of time points for which to solve for y. The initial
value point should be the first element of this sequence.
This sequence must be monotonically increasing or monotonically
decreasing; repeated values are allowed.
args : tuple, optional
Extra arguments to pass to function.
Dfun : callable(y, t, ...) or callable(t, y, ...)
Gradient (Jacobian) of `func`.
If the signature is ``callable(t, y, ...)``, then the argument
`tfirst` must be set ``True``.
col_deriv : bool, optional
True if `Dfun` defines derivatives down columns (faster),
otherwise `Dfun` should define derivatives across rows.
full_output : bool, optional
True if to return a dictionary of optional outputs as the second output
printmessg : bool, optional
Whether to print the convergence message
tfirst : bool, optional
If True, the first two arguments of `func` (and `Dfun`, if given)
must ``t, y`` instead of the default ``y, t``.
.. versionadded:: 1.1.0
Returns
-------
y : array, shape (len(t), len(y0))
Array containing the value of y for each desired time in t,
with the initial value `y0` in the first row.
infodict : dict, only returned if full_output == True
Dictionary containing additional output information
======= ============================================================
key meaning
======= ============================================================
'hu' vector of step sizes successfully used for each time step
'tcur' vector with the value of t reached for each time step
(will always be at least as large as the input times)
'tolsf' vector of tolerance scale factors, greater than 1.0,
computed when a request for too much accuracy was detected
'tsw' value of t at the time of the last method switch
(given for each time step)
'nst' cumulative number of time steps
'nfe' cumulative number of function evaluations for each time step
'nje' cumulative number of jacobian evaluations for each time step
'nqu' a vector of method orders for each successful step
'imxer' index of the component of largest magnitude in the
weighted local error vector (e / ewt) on an error return, -1
otherwise
'lenrw' the length of the double work array required
'leniw' the length of integer work array required
'mused' a vector of method indicators for each successful time step:
1: adams (nonstiff), 2: bdf (stiff)
======= ============================================================
Other Parameters
----------------
ml, mu : int, optional
If either of these are not None or non-negative, then the
Jacobian is assumed to be banded. These give the number of
lower and upper non-zero diagonals in this banded matrix.
For the banded case, `Dfun` should return a matrix whose
rows contain the non-zero bands (starting with the lowest diagonal).
Thus, the return matrix `jac` from `Dfun` should have shape
``(ml + mu + 1, len(y0))`` when ``ml >=0`` or ``mu >=0``.
The data in `jac` must be stored such that ``jac[i - j + mu, j]``
holds the derivative of the ``i``\ th equation with respect to the
``j``\ th state variable. If `col_deriv` is True, the transpose of
this `jac` must be returned.
rtol, atol : float, optional
The input parameters `rtol` and `atol` determine the error
control performed by the solver. The solver will control the
vector, e, of estimated local errors in y, according to an
inequality of the form ``max-norm of (e / ewt) <= 1``,
where ewt is a vector of positive error weights computed as
``ewt = rtol * abs(y) + atol``.
rtol and atol can be either vectors the same length as y or scalars.
Defaults to 1.49012e-8.
tcrit : ndarray, optional
Vector of critical points (e.g., singularities) where integration
care should be taken.
h0 : float, (0: solver-determined), optional
The step size to be attempted on the first step.
hmax : float, (0: solver-determined), optional
The maximum absolute step size allowed.
hmin : float, (0: solver-determined), optional
The minimum absolute step size allowed.
ixpr : bool, optional
Whether to generate extra printing at method switches.
mxstep : int, (0: solver-determined), optional
Maximum number of (internally defined) steps allowed for each
integration point in t.
mxhnil : int, (0: solver-determined), optional
Maximum number of messages printed.
mxordn : int, (0: solver-determined), optional
Maximum order to be allowed for the non-stiff (Adams) method.
mxords : int, (0: solver-determined), optional
Maximum order to be allowed for the stiff (BDF) method.
See Also
--------
solve_ivp : solve an initial value problem for a system of ODEs
ode : a more object-oriented integrator based on VODE
quad : for finding the area under a curve
Examples
--------
The second order differential equation for the angle `theta` of a
pendulum acted on by gravity with friction can be written::
theta''(t) + b*theta'(t) + c*sin(theta(t)) = 0
where `b` and `c` are positive constants, and a prime (') denotes a
derivative. To solve this equation with `odeint`, we must first convert
it to a system of first order equations. By defining the angular
velocity ``omega(t) = theta'(t)``, we obtain the system::
theta'(t) = omega(t)
omega'(t) = -b*omega(t) - c*sin(theta(t))
Let `y` be the vector [`theta`, `omega`]. We implement this system
in Python as:
>>> import numpy as np
>>> def pend(y, t, b, c):
... theta, omega = y
... dydt = [omega, -b*omega - c*np.sin(theta)]
... return dydt
...
We assume the constants are `b` = 0.25 and `c` = 5.0:
>>> b = 0.25
>>> c = 5.0
For initial conditions, we assume the pendulum is nearly vertical
with `theta(0)` = `pi` - 0.1, and is initially at rest, so
`omega(0)` = 0. Then the vector of initial conditions is
>>> y0 = [np.pi - 0.1, 0.0]
We will generate a solution at 101 evenly spaced samples in the interval
0 <= `t` <= 10. So our array of times is:
>>> t = np.linspace(0, 10, 101)
Call `odeint` to generate the solution. To pass the parameters
`b` and `c` to `pend`, we give them to `odeint` using the `args`
argument.
>>> from scipy.integrate import odeint
>>> sol = odeint(pend, y0, t, args=(b, c))
The solution is an array with shape (101, 2). The first column
is `theta(t)`, and the second is `omega(t)`. The following code
plots both components.
>>> import matplotlib.pyplot as plt
>>> plt.plot(t, sol[:, 0], 'b', label='theta(t)')
>>> plt.plot(t, sol[:, 1], 'g', label='omega(t)')
>>> plt.legend(loc='best')
>>> plt.xlabel('t')
>>> plt.grid()
>>> plt.show()
Here is an example of first-order linear ODE.
t = np.arange(0, 10, 0.5) # time points
y = odeint(first, 1, t, args=(1,))
plt.plot(t, y, '.-');
![_images/80aaa443241c5f324e27cde16fda1bb8eb2fe211d3baed322aebb979d328ded5.png](_images/80aaa443241c5f324e27cde16fda1bb8eb2fe211d3baed322aebb979d328ded5.png)
Here’s another example of second-order linear ODE.
t = np.arange(0, 10, 0.2) # time points
y = odeint(second, [1, 1], t, args=([0, -1, -1],))
plt.plot(t, y, '.-');
![_images/940be91be182302ddf59cff09349ecd2cba7522d63a16489f9d6b8b16265aaf5.png](_images/940be91be182302ddf59cff09349ecd2cba7522d63a16489f9d6b8b16265aaf5.png)
odeint()
internally uses adaptive time steps, and returns values of y
for time points specified in t
by interpolation.
# If you are interested in the internal time steps used...
t = np.arange(0, 10, 0.5) # time points
y, info = odeint(first, 1, t, args=(-1,), full_output=1)
plt.plot(t, y, '.-')
# the crosses show the time points actually used
plt.plot(info['tcur'], np.zeros_like(info['tcur']), '+');
![_images/3f7b2ac96f3ff98a419b2a03d9068915ed27650bf67595ae750462c05431de0f.png](_images/3f7b2ac96f3ff98a419b2a03d9068915ed27650bf67595ae750462c05431de0f.png)
Fixed Point and Stability#
A point \(y\) that satisfy \(\frac{d}{dt}{\bf y}=f({\bf y})={\bf 0}\) is called a fixed point.
A fixed point is characterized by its stability.
def expos(a, b):
"""Exponentials exp(a*t), exp(b*t)"""
u = np.array(np.exp(a*t))
v = np.array(np.exp(b*t))
x = np.array([u, -u, -u, u]).T
y = np.array([v, v, -v, -v]).T
return x, y
def spirals(a, b):
"""Spirals: exp(a*t)*cos(b*t), exp(a*t)*sin(b*t) """
u = np.array(np.exp(a*t)*np.cos(b*t))
v = np.array(np.exp(a*t)*np.sin(b*t))
x = np.array([u, v, -u, -v]).T
y = np.array([v, -u, -v, u]).T
return x, y
def arrowcurves(x, y, s=0.1):
"""curves with an arrowhead
x, y: time courses in columns
s: arrowhead size"""
plt.plot(x, y)
n = x.shape[1] # columns
for i in range(n):
plt.arrow(x[-2,i], y[-2,i], x[-1,i]-x[-2,i], y[-1,i]-y[-2,i],
head_width=s, head_length=s)
plt.axis('equal')
t = np.linspace(0, 1) # time points
Stable
Attractor
Neutrally stable
plt.figure(figsize=(12,4))
plt.subplot(1,3,1)
plt.title('Attractor')
plt.plot(0, 0, 'o')
x, y = expos(-2, -3)
arrowcurves(x, y)
plt.subplot(1,3,2)
plt.title('Attractor (spiral)')
plt.plot(0, 0, 'o')
x, y = spirals(-1, 3)
arrowcurves(x,y)
plt.subplot(1,3,3)
plt.title('Neutrally Stable')
plt.plot(0, 0, 'o')
x, y = spirals(0, 2)
arrowcurves(x,y)
![_images/c7023ebae9f6748762a130f07355279fcd6a6e95b698ee55f33524b3379b779e.png](_images/c7023ebae9f6748762a130f07355279fcd6a6e95b698ee55f33524b3379b779e.png)
Unstable
repellor
Saddle
plt.figure(figsize=(12,4))
plt.subplot(1,3,1)
plt.title('Repellor')
plt.plot(0, 0, 'o')
x, y = expos(2, 2.5)
arrowcurves(x, y, s=1)
arrowcurves(x, y)
plt.subplot(1,3,2)
plt.title('Repellor (spiral)')
plt.plot(0, 0, 'o')
x, y = spirals(1, 3)
arrowcurves(x, y, s=0.3)
plt.subplot(1,3,3)
plt.title('Saddle')
plt.plot(0, 0, 'o')
x, y = expos(-3, 1.1)
arrowcurves(x, y, s=0.3)
![_images/83874ce8b1d1adea3249c9a4271127a444e9f251d4507b5f9cae31049572d968.png](_images/83874ce8b1d1adea3249c9a4271127a444e9f251d4507b5f9cae31049572d968.png)
Linear dynamical system#
For a linear dynamical system $\( \frac{d}{dt}{\bf y} = A {\bf y} \)\( where \){\bf y}\( is an \)n\( dimensional vector and \)A\( is an \)n\times n\( matrix, the origin \){\bf y}={\bf 0}\( is always a fixed point. Its stability is determined by the eigenvalues of \)A$.
If the real part of all the eigenvalues are negative or zero, the system is stable.
If any of the real part of the eigenvalues is positive, the system is unstable.
If there are complex eigenvalues, the solution is oscillatory.
def linear(y, t, A):
"""Linear dynamcal system dy/dt = Ay
y: n-dimensional state vector
t: time (not used, for compatibility with odeint())
A: n*n matrix"""
# y is an array (row vector), A is a matrix
return A@y
def lode_plot(A, y0=[1,0]):
"""Plot the trajectory and eiven values of linear ode"""
ev,_ = np.linalg.eig(A)
print('A =', A, '\nev =', ev)
t=np.arange(0, 10, 0.1)
y = odeint(linear, y0, t, args=(A,))
plt.plot(y[0,0], y[0,1], 'o') # starting point
plt.plot(y[:,0], y[:,1], '.-') # trajectory
plt.plot(y[-1,0], y[-1,1], '*') # end point
plt.axis('equal')
Try different settings of A.
# spiral in
lode_plot([[-1, 1], [-1, 0]])
A = [[-1, 1], [-1, 0]]
ev = [-0.5+0.8660254j -0.5-0.8660254j]
![_images/5f6df66bf0528d692b9c2da1037386591fd61b266fc5679504e36b40ef8e2392.png](_images/5f6df66bf0528d692b9c2da1037386591fd61b266fc5679504e36b40ef8e2392.png)
# spiral out
lode_plot([[1, 1], [-1, 0]])
A = [[1, 1], [-1, 0]]
ev = [0.5+0.8660254j 0.5-0.8660254j]
![_images/8e84f8de472fdc50e8ac5bcb626d9078370ce75070870f6b07a102712d99b801.png](_images/8e84f8de472fdc50e8ac5bcb626d9078370ce75070870f6b07a102712d99b801.png)
# saddle point
lode_plot([[-1, 0], [0, 1]], [1,0.0001])
A = [[-1, 0], [0, 1]]
ev = [-1. 1.]
![_images/0880babc3a92378773f71668c3030ed0334541717ea4181bc01f22b26fa2afd2.png](_images/0880babc3a92378773f71668c3030ed0334541717ea4181bc01f22b26fa2afd2.png)
Nonlinear ODEs#
While the dynamics of a linear ODE can show only convergence, divergence, or neutrally stable oscillations, nonlinear ODEs can show limit-cycle oscillation and chaos.
Van der Pol oscillator#
This is a classic equation describing an oscillator circuit with a vacuume tube: $\( \frac{d^2y}{dt^2} - \mu(1-y^2)\frac{dy}{dt} + y = 0 \)$
def vdp(y, t, mu):
"""Van der Pol equation"""
y1, y2 = y
return np.array([y2, mu*(1 - y1**2)*y2 - y1])
t = np.arange(0, 50, 0.1)
y = odeint(vdp, [0.1, 0], t, args=(1,))
plt.plot(t, y);
![_images/6e1319be88de29d23e44a8a489748f5c73c00a67277ac9d9f3dcb29bb69190f4.png](_images/6e1319be88de29d23e44a8a489748f5c73c00a67277ac9d9f3dcb29bb69190f4.png)
# phase plot
plt.plot(y[:,0], y[:,1]);
![_images/c7f1074c9dc7bf9028cf524bf86821e80cf278f2d570b7b2a2bdd14ed4ac214c.png](_images/c7f1074c9dc7bf9028cf524bf86821e80cf278f2d570b7b2a2bdd14ed4ac214c.png)
# from outside
y = odeint(vdp, [3, 0], t, args=(1,))
plt.plot(y[:,0], y[:,1]);
![_images/d414106c5587c0013ee9ab518e764d84d87c25902f2d909c513a2013688a982f.png](_images/d414106c5587c0013ee9ab518e764d84d87c25902f2d909c513a2013688a982f.png)
Periodic orbit and Limit cycle#
If a trajectry comes back to itself \(y(t+T) = y(t)\) after some period \(T\), it is called a periodic orbit.
If trajectories around it converges to a periodic orbit, it is called a limit cycle.
Poincaré-Bendixon theorem: In a continuous 2D dynamical system, if a solution stay within a closed set with no fixed point, it converges to a periodic orbit.
It implies that there is no chaos in a continuous 2D dynamic system.
Lorenz attractor#
Edward Lorenz derived a simplified equation describing the convection of atmosphere and found that it shows non-periodic oscillation. $\( \frac{dx}{dt} = p(y - x) \)\( \)\( \frac{dy}{dt} = -xz + rx - y \)\( \)\( \frac{dz}{dt} = xy - bz \)$
def lorenz(xyz, t, p=10., r=28., b=8./3):
"""Lorenz equation"""
x, y, z = xyz
dxdt = p*(y - x)
dydt = -x*z + r*x - y
dzdt = x*y - b*z
return np.array([dxdt, dydt, dzdt])
y0 = np.array([0.1, 0, 0])
t = np.arange(0, 50, 0.01)
y = odeint(lorenz, y0, t, args=(10., 30., 8./3))
plt.plot(t, y);
![_images/326cb73e65187b9846e1c6d99752286db87593551f3c9b5d358b5b727a50a0cc.png](_images/326cb73e65187b9846e1c6d99752286db87593551f3c9b5d358b5b727a50a0cc.png)
%matplotlib notebook
# Plot in 3D
fig = plt.figure(figsize=(8,8))
ax = fig.add_subplot(projection='3d')
ax.plot(y[:,0], y[:,1], y[:,2], lw=0.5);