1
00:00:01,260 --> 00:00:02,093
Welcome back.
2
00:00:02,093 --> 00:00:04,290
We're gonna switch gear a little bit now
3
00:00:04,290 --> 00:00:07,260
and move away from sort
of mathematical treatment
4
00:00:07,260 --> 00:00:08,820
of this continuous model
5
00:00:08,820 --> 00:00:11,520
and think about how to
implement them computationally.
6
00:00:13,950 --> 00:00:15,600
So as I said in the previous video,
7
00:00:15,600 --> 00:00:18,930
one of the big strengths
of differential equations
8
00:00:18,930 --> 00:00:23,580
is that they give you, without
any additional alteration,
9
00:00:23,580 --> 00:00:25,980
a complete view of the flow diagram
10
00:00:25,980 --> 00:00:27,000
because that's literally
11
00:00:27,000 --> 00:00:29,160
what those differential equations are
12
00:00:29,160 --> 00:00:31,920
is the flow or the speed of the system
13
00:00:31,920 --> 00:00:34,980
at any point in variable space.
14
00:00:34,980 --> 00:00:38,160
So if you tell me the
state, x of a system,
15
00:00:38,160 --> 00:00:42,060
the time derivative of x
of t gives me the direction
16
00:00:42,060 --> 00:00:46,560
of the movement of the
system, as well as its speed,
17
00:00:46,560 --> 00:00:48,513
how fast is it going there?
18
00:00:49,770 --> 00:00:51,600
I have three different
notations right now.
19
00:00:51,600 --> 00:00:53,580
So you're probably most used
20
00:00:53,580 --> 00:00:56,550
to the explicit time derivative.
21
00:00:56,550 --> 00:00:59,070
As I said, I like using Newton's notation.
22
00:00:59,070 --> 00:01:01,645
It's just a little simpler,
23
00:01:01,645 --> 00:01:05,670
this dot representing a time derivative,
24
00:01:05,670 --> 00:01:09,690
and really the way we write
those differential equation
25
00:01:09,690 --> 00:01:12,510
is having x dot or s dot
or i dot, whatever it is,
26
00:01:12,510 --> 00:01:15,454
equal to some function of
the state of the system x,
27
00:01:15,454 --> 00:01:18,573
maybe also explicitly of time, t.
28
00:01:20,160 --> 00:01:23,119
So these differential
equations are always written
29
00:01:23,119 --> 00:01:26,130
as x dot equals f of x of t.
30
00:01:26,130 --> 00:01:27,120
That's our differential,
31
00:01:27,120 --> 00:01:29,163
our system of differential equations.
32
00:01:30,300 --> 00:01:34,230
So in this video, what
I'd like to describe is
33
00:01:34,230 --> 00:01:36,030
what you'll be doing often,
34
00:01:36,030 --> 00:01:38,100
which is not necessarily
mathematical analysis
35
00:01:38,100 --> 00:01:39,450
over these differential equations,
36
00:01:39,450 --> 00:01:41,160
but just tracking them in time.
37
00:01:41,160 --> 00:01:43,200
So you put the system
in some initial state,
38
00:01:43,200 --> 00:01:46,590
you have some initial
conditions, x at t equals zero,
39
00:01:46,590 --> 00:01:49,473
and you wanna follow
this flow of dynamics.
40
00:01:53,970 --> 00:01:56,340
So let's say that our system here
41
00:01:56,340 --> 00:01:57,930
has only one degree of freedom.
42
00:01:57,930 --> 00:02:00,540
This x of t is just one variable,
43
00:02:00,540 --> 00:02:04,800
and we're interested in knowing from x0.
44
00:02:04,800 --> 00:02:07,260
An initial condition of T equals zero,
45
00:02:07,260 --> 00:02:10,173
the state of the system
for all t, for all time.
46
00:02:13,440 --> 00:02:18,440
Now there might be some true
evolution for this system,
47
00:02:19,350 --> 00:02:20,883
which might look like this,
48
00:02:23,400 --> 00:02:26,433
but that would be the true x of t.
49
00:02:29,550 --> 00:02:33,150
The simplest way to follow it numerically
50
00:02:33,150 --> 00:02:35,309
is discretized time,
the same way we're doing
51
00:02:35,309 --> 00:02:37,590
in discretized model.
52
00:02:37,590 --> 00:02:39,300
So at t equal zero,
53
00:02:39,300 --> 00:02:43,650
you can evaluate your
differential equation, x dot of t.
54
00:02:43,650 --> 00:02:47,643
So you evaluate this
function, f of x of t here,
55
00:02:48,750 --> 00:02:52,680
and what it gives you is
the instantaneous derivative
56
00:02:52,680 --> 00:02:53,523
at that point.
57
00:02:55,200 --> 00:02:58,420
So then one thing you can
do once you have that is say
58
00:02:59,580 --> 00:03:04,580
x at time, t, plus h
where h is some time step,
59
00:03:09,480 --> 00:03:13,110
simply gonna be equal to where I am
60
00:03:13,110 --> 00:03:16,800
plus the projection over h of the slope,
61
00:03:16,800 --> 00:03:20,640
f of x of t, my differential equation.
62
00:03:20,640 --> 00:03:22,410
And then that might put me here.
63
00:03:22,410 --> 00:03:25,140
And then I reevaluate my equation,
64
00:03:25,140 --> 00:03:30,003
project myself, reevaluate
project, project, project.
65
00:03:31,590 --> 00:03:32,580
And as you can see,
66
00:03:32,580 --> 00:03:37,580
I don't expect to get exactly
the right distribution.
67
00:03:39,600 --> 00:03:43,380
I mean, I'm doing a little
bit of whatever here,
68
00:03:43,380 --> 00:03:46,294
but the point is I expect to get something
69
00:03:46,294 --> 00:03:50,433
that's close to the true
solution, but that deviates.
70
00:03:52,350 --> 00:03:54,900
So this method, the simplest method
71
00:03:54,900 --> 00:03:56,763
is called Euler's method.
72
00:04:05,550 --> 00:04:09,570
And we're gonna say that its local error
73
00:04:13,260 --> 00:04:18,260
is of order, it's a big
O here, order h squared.
74
00:04:21,210 --> 00:04:25,923
The reason for that is we know
that there's a true x of t,
75
00:04:27,870 --> 00:04:30,693
and then there's the x
of t from Euler's method.
76
00:04:32,760 --> 00:04:37,000
And we can write x of t plus h
77
00:04:38,820 --> 00:04:41,160
by developing around x of t,
78
00:04:41,160 --> 00:04:46,160
the Taylor series of the true solution.
79
00:04:48,150 --> 00:04:53,150
So we know that it's
gonna be x of t plus h
80
00:04:53,700 --> 00:04:57,570
f of x of t plus...
81
00:04:57,570 --> 00:04:59,160
While we're doing that step,
82
00:04:59,160 --> 00:05:02,730
we know that the system
continues evolving,
83
00:05:02,730 --> 00:05:04,140
continues following the dynamics.
84
00:05:04,140 --> 00:05:05,760
So there is some curvature,
85
00:05:05,760 --> 00:05:10,760
curvature f prime of x of t plus h cubed.
86
00:05:11,190 --> 00:05:13,470
If you remember your Taylor series,
87
00:05:13,470 --> 00:05:17,793
three factorials, f prime
prime of x of t, and so on.
88
00:05:19,110 --> 00:05:21,660
And so the reason why we
say that Euler's method
89
00:05:21,660 --> 00:05:25,120
as a local error term in h squared
90
00:05:26,130 --> 00:05:30,720
is because it captures those first term
91
00:05:30,720 --> 00:05:32,343
of the Taylor series,
92
00:05:33,360 --> 00:05:37,893
but ignores the higher order term,
93
00:05:38,970 --> 00:05:41,163
so ignored by Euler.
94
00:05:43,200 --> 00:05:46,800
And so at every time step h that we do,
95
00:05:46,800 --> 00:05:50,520
the biggest error term is
gonna be this one here.
96
00:05:50,520 --> 00:05:53,550
So if the curvature of f
prime is equal to zero,
97
00:05:53,550 --> 00:05:55,512
if we have a linear growth
of a system, let's say,
98
00:05:55,512 --> 00:05:57,960
so f prime is gonna be zero.
99
00:05:57,960 --> 00:05:59,820
Euler's method is gonna do awesome
100
00:05:59,820 --> 00:06:02,343
because this error term is equal to zero,
101
00:06:04,440 --> 00:06:07,080
but for any like non-zero error,
102
00:06:07,080 --> 00:06:08,740
really the key point is that
103
00:06:09,990 --> 00:06:12,930
the biggest term that we have
is proportional to h squared,
104
00:06:12,930 --> 00:06:13,950
and that means two things.
105
00:06:13,950 --> 00:06:17,070
That means just as we move
from discrete to continuous.
106
00:06:17,070 --> 00:06:21,210
If h goes to zero, if we
have a very good computer
107
00:06:21,210 --> 00:06:23,310
and we're very patient
and we say I'm gonna do,
108
00:06:23,310 --> 00:06:25,714
you know, infinitely many step
109
00:06:25,714 --> 00:06:29,100
of tiny, tiny, tiny, tiny durations,
110
00:06:29,100 --> 00:06:30,720
then Euler's method is gonna be fine,
111
00:06:30,720 --> 00:06:33,720
but otherwise, you know,
it's important to remember
112
00:06:33,720 --> 00:06:36,300
that you're going to have
those local error terms.
113
00:06:36,300 --> 00:06:41,300
And often, the difference
here, let me put it in blue,
114
00:06:41,340 --> 00:06:43,833
so the difference here is the local error.
115
00:06:46,560 --> 00:06:51,450
And often as in my cartoon,
this error sort of grows
116
00:06:51,450 --> 00:06:53,460
with time steps, right?
117
00:06:53,460 --> 00:06:54,484
It might shrink sometimes.
118
00:06:54,484 --> 00:06:56,040
You might be lucky,
119
00:06:56,040 --> 00:06:58,263
but overall you expect it to grow.
120
00:07:00,540 --> 00:07:04,860
And so, we're often gonna
say that the global error
121
00:07:07,560 --> 00:07:12,560
of our numerical integrator is the sum.
122
00:07:19,380 --> 00:07:21,363
Let me rewrite that more clearly here.
123
00:07:23,370 --> 00:07:28,370
Global error is gonna be the sum
124
00:07:29,220 --> 00:07:34,220
over all time steps of the local error.
125
00:07:36,660 --> 00:07:37,611
And we had the fact...
126
00:07:37,611 --> 00:07:39,985
So we had this idea that the local error
127
00:07:39,985 --> 00:07:44,313
was an h square for Euler.
128
00:07:48,870 --> 00:07:50,940
Well, this is gonna be equal
129
00:07:50,940 --> 00:07:55,940
to the duration of our
numerical integration.
130
00:07:57,840 --> 00:08:00,100
So we're integrating from t equals zero
131
00:08:01,500 --> 00:08:04,353
to, let's say T, capital T.
132
00:08:06,450 --> 00:08:08,142
So how many steps are we gonna have to do?
133
00:08:08,142 --> 00:08:11,340
How many local error are we solving here,
134
00:08:11,340 --> 00:08:12,173
are we summing up?
135
00:08:12,173 --> 00:08:15,873
So it's capital T divided by h.
136
00:08:17,970 --> 00:08:20,280
That's the number of
steps that we have to do,
137
00:08:20,280 --> 00:08:24,033
times local error, which
is of order h squared.
138
00:08:25,230 --> 00:08:27,453
So this is gonna be an h.
139
00:08:29,430 --> 00:08:31,680
And what I'm trying to say here is that
140
00:08:31,680 --> 00:08:33,030
there's this key thing is,
141
00:08:33,030 --> 00:08:37,380
which is that Euler's method
is capturing up to order h
142
00:08:37,380 --> 00:08:41,280
the Taylor series of the true solution,
143
00:08:41,280 --> 00:08:43,351
and it's so that its local error
144
00:08:43,351 --> 00:08:45,688
and what's being ignored is h square,
145
00:08:45,688 --> 00:08:47,340
but the global error,
146
00:08:47,340 --> 00:08:49,470
because you're doing a
lot of different steps,
147
00:08:49,470 --> 00:08:52,020
the global error is
still gonna be of order h
148
00:08:52,020 --> 00:08:54,540
even though you're capturing the thing.
149
00:08:54,540 --> 00:08:57,213
You're capturing the term of order h.
150
00:08:59,677 --> 00:09:02,726
Really, I wanna highlight this distinction
151
00:09:02,726 --> 00:09:04,606
between local error and global error
152
00:09:04,606 --> 00:09:06,806
because it is kind of
surprising to have something
153
00:09:06,806 --> 00:09:09,690
that captures the derivative.
154
00:09:09,690 --> 00:09:12,360
And yet as you integrate
a trajectory in time,
155
00:09:12,360 --> 00:09:15,960
you do expect its error to scale
156
00:09:15,960 --> 00:09:18,390
with your step size linearly.
157
00:09:18,390 --> 00:09:21,270
So basically if you wanna
do a numerical integration
158
00:09:21,270 --> 00:09:24,420
that's 10 times closer
to the true solution,
159
00:09:24,420 --> 00:09:27,390
if I wanted my red that's
here to be 10 times closer,
160
00:09:27,390 --> 00:09:29,790
I would need to do 10 times more steps.
161
00:09:29,790 --> 00:09:33,930
But we can have integrator that do better.
162
00:09:33,930 --> 00:09:36,949
So, I apologize here.
163
00:09:36,949 --> 00:09:39,033
I expected a blank sheet.
164
00:09:46,620 --> 00:09:49,770
So we wanna integrate
again from t equals zero
165
00:09:49,770 --> 00:09:51,273
to t equals T,
166
00:09:53,543 --> 00:09:55,293
whoops, x of t.
167
00:09:56,940 --> 00:10:01,940
We have something
different, true x of t here.
168
00:10:03,117 --> 00:10:04,833
And now we want to do better.
169
00:10:05,850 --> 00:10:08,700
Well, what's the key part that
Euler's method is missing?
170
00:10:11,640 --> 00:10:14,160
It's the fact that as you do one step,
171
00:10:14,160 --> 00:10:16,950
your derivative is already changing
172
00:10:16,950 --> 00:10:18,060
during that step, right?
173
00:10:18,060 --> 00:10:20,313
There might be a second derivative.
174
00:10:21,420 --> 00:10:24,930
And so it's a mistake
to to project it too far
175
00:10:24,930 --> 00:10:26,193
into the future.
176
00:10:27,960 --> 00:10:30,810
So there is another
method, which I've learned
177
00:10:30,810 --> 00:10:33,693
under the name of improved Euler.
178
00:10:34,890 --> 00:10:38,493
Improved Euler's method.
179
00:10:42,900 --> 00:10:46,171
If you Google around, in English,
180
00:10:46,171 --> 00:10:48,870
I believe it's often called Heun's method.
181
00:10:48,870 --> 00:10:49,703
Whoops.
182
00:10:50,940 --> 00:10:52,023
Too sensitive here.
183
00:10:54,180 --> 00:10:57,033
Heun's method.
184
00:10:59,790 --> 00:11:01,770
And this, I'll tell you right away
185
00:11:01,770 --> 00:11:05,793
is a second order method in global error.
186
00:11:07,230 --> 00:11:10,590
So this method is gonna
have a global error
187
00:11:14,700 --> 00:11:18,420
that scales of the order of h squared,
188
00:11:18,420 --> 00:11:21,300
meaning that, you know, if
you wanted Euler's method
189
00:11:21,300 --> 00:11:23,250
to be 100 times more accurate,
190
00:11:23,250 --> 00:11:25,290
you needed 100 times more time step.
191
00:11:25,290 --> 00:11:28,690
Here, you would only need
10 times more time steps
192
00:11:30,330 --> 00:11:32,853
because the global error
scales better with h.
193
00:11:34,080 --> 00:11:38,460
And so that means that the local error
194
00:11:38,460 --> 00:11:43,460
for the same reason as
before has to be an h cubed
195
00:11:45,120 --> 00:11:48,780
because the number of
steps we're making grows
196
00:11:48,780 --> 00:11:51,600
as one over h as we
decrease our time step.
197
00:11:51,600 --> 00:11:56,600
So the local error error
and the global error,
198
00:11:56,640 --> 00:11:59,103
there's one difference in order, okay.
199
00:12:00,840 --> 00:12:03,120
Well, what is this method?
200
00:12:03,120 --> 00:12:04,920
Let me just fix my typo here.
201
00:12:04,920 --> 00:12:05,753
Heun's method.
202
00:12:09,720 --> 00:12:13,653
The method itself looks super simple.
203
00:12:15,210 --> 00:12:17,790
What we're gonna say,
and I'm going to write it
204
00:12:17,790 --> 00:12:22,790
as xh for Heun at time t plus h,
205
00:12:24,150 --> 00:12:25,810
it's gonna be equal to x of t
206
00:12:27,930 --> 00:12:31,960
plus h f of x of t
207
00:12:33,840 --> 00:12:38,640
plus f of xe time t plus h.
208
00:12:38,640 --> 00:12:41,310
So I'm switching up my
notation here a little more.
209
00:12:41,310 --> 00:12:43,860
So let me explain what I mean,
210
00:12:43,860 --> 00:12:46,500
so that the prediction from Heun's method
211
00:12:46,500 --> 00:12:51,390
is gonna be a projection from
x of t over a time step h
212
00:12:51,390 --> 00:12:56,220
of the average between
the measured derivative.
213
00:12:56,220 --> 00:13:00,600
So this term here, this f of x of t
214
00:13:00,600 --> 00:13:02,150
is the initial derivative here.
215
00:13:03,112 --> 00:13:07,023
And now what I have, this
f of xe, let me go back,
216
00:13:09,600 --> 00:13:13,800
is gonna be the prediction
from Euler's method.
217
00:13:13,800 --> 00:13:16,840
So basically what you do here is you say
218
00:13:22,380 --> 00:13:24,813
I look at my initial derivative,
219
00:13:26,250 --> 00:13:28,893
I project myself over time h.
220
00:13:34,901 --> 00:13:38,553
Okay, project myself over time h,
221
00:13:40,148 --> 00:13:44,550
so I get an Euler prediction of here.
222
00:13:44,550 --> 00:13:45,870
Well, now at this point,
223
00:13:45,870 --> 00:13:49,200
my derivative here is much faster,
224
00:13:49,200 --> 00:13:52,680
much steeper than it
was at my initial point.
225
00:13:52,680 --> 00:13:54,670
So I'm gonna take the average
226
00:13:56,640 --> 00:14:00,300
of both, my first point
and my second point.
227
00:14:00,300 --> 00:14:03,150
So my average might look
something like this,
228
00:14:03,150 --> 00:14:05,430
and now this is what I'm gonna project,
229
00:14:05,430 --> 00:14:06,730
and I'm gonna end up here.
230
00:14:08,910 --> 00:14:11,430
So basically every step of Heun's method
231
00:14:11,430 --> 00:14:13,290
does one step with Euler's method,
232
00:14:13,290 --> 00:14:15,663
looks at the derivative and average.
233
00:14:16,747 --> 00:14:19,817
The derivative that Euler's would have
234
00:14:19,817 --> 00:14:22,560
for its second step with the initial step
235
00:14:22,560 --> 00:14:24,843
and then does a true
first step under Heun.
236
00:14:26,556 --> 00:14:28,650
Let me give another example here.
237
00:14:28,650 --> 00:14:32,190
So with Euler, I evaluate
my derivative here.
238
00:14:32,190 --> 00:14:36,210
I project into the future up to 2h,
239
00:14:36,210 --> 00:14:38,283
which would be, you know, somewhere here.
240
00:14:39,780 --> 00:14:42,870
Euler is gonna end up way off base here.
241
00:14:42,870 --> 00:14:45,243
It's a steep derivative,
might end up here.
242
00:14:48,330 --> 00:14:51,633
Now would evaluate the derivative
here, which is negative.
243
00:14:52,560 --> 00:14:54,360
So Heun would average both,
244
00:14:54,360 --> 00:14:56,730
which would be pretty
much flat in this case
245
00:14:56,730 --> 00:14:57,960
and end up here.
246
00:14:57,960 --> 00:15:00,573
Doing much better than
the blue dot, right?
247
00:15:02,490 --> 00:15:06,300
So Euler is doing this and this,
248
00:15:06,300 --> 00:15:08,343
and Heun is doing this and this.
249
00:15:09,420 --> 00:15:13,740
And the key part is that the local error
250
00:15:13,740 --> 00:15:16,800
and the global error are both
one order of magnitude better.
251
00:15:16,800 --> 00:15:19,350
So if you want to be
100 times more accurate,
252
00:15:19,350 --> 00:15:22,023
you only have to do 10 times more steps.
253
00:15:24,840 --> 00:15:28,560
Now every single evaluation
takes two evaluation
254
00:15:28,560 --> 00:15:29,790
of your derivatives.
255
00:15:29,790 --> 00:15:32,850
So every step does take
twice as many operation,
256
00:15:32,850 --> 00:15:36,000
twice as many evaluation
of this f function
257
00:15:36,000 --> 00:15:38,130
for your derivatives.
258
00:15:38,130 --> 00:15:41,940
But really you're still winning,
you know, a factor of five
259
00:15:41,940 --> 00:15:43,710
if you wanna be 100 times more accurate.
260
00:15:43,710 --> 00:15:48,240
So you can have your time
steps be 10 times smaller.
261
00:15:48,240 --> 00:15:49,800
You do twice as many operations.
262
00:15:49,800 --> 00:15:54,240
So that's 20 operations for
your 100 fold precision,
263
00:15:54,240 --> 00:15:55,260
which is still much better
264
00:15:55,260 --> 00:15:57,663
than what you would do with Euler.
265
00:15:58,845 --> 00:16:00,603
Now as you'll see in the homework,
266
00:16:05,160 --> 00:16:09,090
Heun's method, which is
this definition here,
267
00:16:09,090 --> 00:16:13,740
is actually not defined as
a simple arithmetic average
268
00:16:13,740 --> 00:16:18,740
between f of x of t and
f of xe of t plus h.
269
00:16:21,450 --> 00:16:24,090
The basic idea looks like
you're just taking an average,
270
00:16:24,090 --> 00:16:25,320
but that's not what it is.
271
00:16:25,320 --> 00:16:29,670
The goal here, and this is
a hint for the homework,
272
00:16:29,670 --> 00:16:34,670
is that you can ask, well, what is f of xe
273
00:16:34,740 --> 00:16:36,453
at time t plus h?
274
00:16:37,410 --> 00:16:39,963
You can develop that
with the Taylor series.
275
00:16:49,230 --> 00:16:54,000
And then you can show that this
one half is actually chosen
276
00:16:54,000 --> 00:16:58,800
such that this xh approximation
277
00:16:58,800 --> 00:17:03,150
at time t plus h captures one more term,
278
00:17:03,150 --> 00:17:05,010
so captures the h square term
279
00:17:05,010 --> 00:17:07,650
in the original Taylor
series for the true solution.
280
00:17:07,650 --> 00:17:10,722
This is something that like sounds weird.
281
00:17:10,722 --> 00:17:13,560
Like your only evaluating
first order derivative.
282
00:17:13,560 --> 00:17:16,260
How could you like go and get information
283
00:17:16,260 --> 00:17:18,210
about the second order derivative?
284
00:17:18,210 --> 00:17:21,510
But if you do work it out
by using Taylor series
285
00:17:21,510 --> 00:17:26,360
of both x of t, which we did right here,
286
00:17:30,196 --> 00:17:32,746
so this is our Taylor series
for the true solution,
287
00:17:33,990 --> 00:17:37,710
and then you go do your
Taylor series for this term,
288
00:17:37,710 --> 00:17:41,100
for the derivative itself
under Euler's method,
289
00:17:41,100 --> 00:17:42,810
then you can see that you're able
290
00:17:42,810 --> 00:17:45,530
to combine multiple
evaluation of your derivatives
291
00:17:45,530 --> 00:17:47,890
in a smart way to capture more and more
292
00:17:49,260 --> 00:17:51,453
or capture higher order terms.
293
00:17:53,790 --> 00:17:55,557
In a future video, so next week,
294
00:17:55,557 --> 00:17:59,700
I'll go into deeper detail
in the in the Runge-Kutta
295
00:17:59,700 --> 00:18:01,117
or Dormand-Prince method,
296
00:18:01,117 --> 00:18:04,110
which captures up to order four terms,
297
00:18:04,110 --> 00:18:06,870
and there it's just gonna
be a lot of arithmetics,
298
00:18:06,870 --> 00:18:10,080
but we're gonna be able
to see that similarly
299
00:18:10,080 --> 00:18:14,042
by doing four evaluations or derivative,
300
00:18:14,042 --> 00:18:16,350
we can get like a higher order term,
301
00:18:16,350 --> 00:18:18,510
and there it's gonna look
nothing like an average.
302
00:18:18,510 --> 00:18:21,060
So the fact that you have
a simple average here,
303
00:18:21,060 --> 00:18:26,060
this divided by two for both evaluation
304
00:18:27,060 --> 00:18:29,223
of the derivative is sort of coincidence,
305
00:18:30,480 --> 00:18:31,980
but it's a nice explanation.
306
00:18:31,980 --> 00:18:34,590
And as you can see, like in any cartoon,
307
00:18:34,590 --> 00:18:36,360
like the one in the top left here,
308
00:18:36,360 --> 00:18:38,700
it's clear that it's gonna do better.
309
00:18:38,700 --> 00:18:43,700
So, apologies, by the
way, if my picture here
310
00:18:44,250 --> 00:18:48,843
was, in the way for some
of the words up there.
311
00:18:52,110 --> 00:18:56,160
But, you know, this is a topic
312
00:18:56,160 --> 00:18:58,375
that we're gonna revisit
time and time again
313
00:18:58,375 --> 00:19:02,103
because the key part is
that even if you ask,
314
00:19:03,690 --> 00:19:05,945
you know, a scientific library,
315
00:19:05,945 --> 00:19:10,470
please integrate my differential
equations through time,
316
00:19:10,470 --> 00:19:13,563
it's super important to know what order,
317
00:19:14,940 --> 00:19:16,260
how much error are you getting,
318
00:19:16,260 --> 00:19:18,330
what's the order of the
method that it's using
319
00:19:18,330 --> 00:19:20,970
by default or that
you're asking it to solve
320
00:19:20,970 --> 00:19:23,670
because what comes out
of discrete computer
321
00:19:23,670 --> 00:19:26,040
and discrete integration
is not the true x of t,
322
00:19:26,040 --> 00:19:27,660
but an approximation of it.
323
00:19:27,660 --> 00:19:29,520
And as I showed in the cartoon,
324
00:19:29,520 --> 00:19:32,223
those approximations can be way off.
325
00:19:33,510 --> 00:19:36,029
So it it's gonna be important
to distinguish error
326
00:19:36,029 --> 00:19:38,015
due to our model specification
327
00:19:38,015 --> 00:19:42,723
and error due to how we
solve the model numerically.
328
00:19:46,230 --> 00:19:49,068
And one way to test how well we're doing
329
00:19:49,068 --> 00:19:51,780
is gonna be to compare our predictions
330
00:19:51,780 --> 00:19:53,580
from the mathematical models
331
00:19:53,580 --> 00:19:55,200
with simulations of the results.
332
00:19:55,200 --> 00:19:58,380
So in the next video,
we're gonna see how to use
333
00:19:58,380 --> 00:20:00,090
some of the tools we quickly saw
334
00:20:00,090 --> 00:20:01,800
to simulate discrete models,
335
00:20:01,800 --> 00:20:05,160
but this time in continuous time.
336
00:20:05,160 --> 00:20:06,560
So I'll see in the next one.