1
00:00:00,000 --> 00:00:01,620
You know, usually when we think about measuring

2
00:00:01,620 --> 00:00:04,919
something, there is a deep comfort in the physical

3
00:00:04,919 --> 00:00:07,280
reality of it. Oh, absolutely. It's grounded.

4
00:00:07,599 --> 00:00:10,279
Right. Like you take a piece of wood, you lay

5
00:00:10,279 --> 00:00:13,099
a tape measure across it, and it is exactly four

6
00:00:13,099 --> 00:00:16,019
feet long. It's solid. It's undeniable. There's

7
00:00:16,019 --> 00:00:18,199
really no ambiguity there. Yeah, you measure

8
00:00:18,199 --> 00:00:20,620
a liquid in a beaker or you weigh a stone on

9
00:00:20,620 --> 00:00:23,699
a scale. We basically navigate our entire lives

10
00:00:23,699 --> 00:00:25,660
assuming the universe is made of things that

11
00:00:25,660 --> 00:00:28,739
possess a measurable objective size. Exactly.

12
00:00:29,230 --> 00:00:31,269
And it's a foundational assumption for how we

13
00:00:31,269 --> 00:00:34,250
interact with the world. We have this intuitive,

14
00:00:34,409 --> 00:00:37,950
almost primal belief that if a space or an object

15
00:00:37,950 --> 00:00:41,630
exists, it inherently has a volume or a length

16
00:00:41,630 --> 00:00:44,630
or an area. It's binary. Either it's there to

17
00:00:44,630 --> 00:00:46,770
be measured or it's not. Right. I mean, that's

18
00:00:46,770 --> 00:00:50,070
just common sense. But when you step into the

19
00:00:50,070 --> 00:00:53,390
advanced mathematics of randomness, that tape

20
00:00:53,390 --> 00:00:56,229
measure just, well, it snaps. Yeah, the peaker

21
00:00:56,229 --> 00:00:58,909
shatters. Exactly. You suddenly find yourself

22
00:00:58,909 --> 00:01:01,850
in a landscape where the very concept of length

23
00:01:01,850 --> 00:01:05,150
or size fundamentally breaks down. And that is

24
00:01:05,150 --> 00:01:08,049
exactly what we are getting into today. Welcome

25
00:01:08,049 --> 00:01:10,590
to the Deep Dive. I'm your host. And sitting

26
00:01:10,590 --> 00:01:12,750
across from me is our resident mathematics expert.

27
00:01:13,269 --> 00:01:17,129
Today, we are taking a deep dive into how to

28
00:01:17,129 --> 00:01:20,129
put that shattered universe back together. We're

29
00:01:20,129 --> 00:01:22,969
meticulously decoding a truly seminal text, an

30
00:01:22,969 --> 00:01:25,670
advanced course in probability and stochastic

31
00:01:25,670 --> 00:01:28,950
processes published in 2023 by Dirk P. Kroos

32
00:01:28,950 --> 00:01:32,069
and Zdravko I. Botev. And, you know, what makes

33
00:01:32,069 --> 00:01:34,430
this text so crucial is that it acts as a bridge.

34
00:01:34,609 --> 00:01:37,709
There is this massive, frankly, intimidating

35
00:01:37,709 --> 00:01:41,030
chasm between elementary probability. Like the

36
00:01:41,030 --> 00:01:43,129
coin flips and the dice rolls, right? Yeah, exactly.

37
00:01:43,290 --> 00:01:45,790
The dice rolls, the bell curves we learn early

38
00:01:45,790 --> 00:01:48,920
on in school. There's a... chasm between that

39
00:01:48,920 --> 00:01:52,700
and advanced research level probability, which

40
00:01:52,700 --> 00:01:55,180
runs on an entirely different engine called measure

41
00:01:55,180 --> 00:01:58,560
theory. Right. And Kross and Botov, they wanted

42
00:01:58,560 --> 00:02:01,319
to build a rigorous framework that actually connects

43
00:02:01,319 --> 00:02:05,140
those two worlds without getting bogged down

44
00:02:05,140 --> 00:02:07,280
in endless abstraction. They even anchored their

45
00:02:07,280 --> 00:02:09,319
approach with a quote from Richard Feynman right

46
00:02:09,319 --> 00:02:11,879
at the beginning. Oh, I love that quote. It's,

47
00:02:11,879 --> 00:02:14,780
if you cannot explain something in simple terms,

48
00:02:14,919 --> 00:02:18,159
you don't really understand it. Yes. Which, I

49
00:02:18,159 --> 00:02:20,280
mean, that sets an incredibly high bar for a

50
00:02:20,280 --> 00:02:22,580
math textbook. It really does. And before we

51
00:02:22,580 --> 00:02:24,680
get into the actual mechanics of measure theory,

52
00:02:24,879 --> 00:02:28,000
we have to talk about how they enforce that clarity.

53
00:02:28,539 --> 00:02:31,199
Because the authors are, well, for lack of a

54
00:02:31,199 --> 00:02:33,139
better word, they are obsessed with notation.

55
00:02:33,479 --> 00:02:36,280
Oh, heavily obsessed. But for good reason. They

56
00:02:36,280 --> 00:02:38,639
view notation as the heart of mathematics. Because

57
00:02:38,639 --> 00:02:41,659
if you think about it, a poorly designed notation

58
00:02:41,659 --> 00:02:44,659
system forces your brain to spend all its computing

59
00:02:44,659 --> 00:02:47,699
power deciphering the symbols. Rather than understanding

60
00:02:47,699 --> 00:02:50,199
the relationships between the ideas. Exactly.

61
00:02:50,620 --> 00:02:53,680
So they establish a strict, almost architectural

62
00:02:53,680 --> 00:02:56,020
house style. If you're reading this text, you

63
00:02:56,020 --> 00:02:58,159
never have to guess what an object is. Right.

64
00:02:58,159 --> 00:03:00,840
Like composite objects, column vectors and matrices

65
00:03:00,840 --> 00:03:03,960
are always in boldface. Yep. Random variables

66
00:03:03,960 --> 00:03:07,310
are strictly uppercase Roman letters. Sets are

67
00:03:07,310 --> 00:03:09,770
uppercase Roman too, but the collections of those

68
00:03:09,770 --> 00:03:12,150
sets, the really advanced structures we'll get

69
00:03:12,150 --> 00:03:15,830
into later, they're in this sweeping calligraphic

70
00:03:15,830 --> 00:03:18,919
font. It effectively acts as a color coding system

71
00:03:18,919 --> 00:03:21,759
for pure logic. You can just glance at an equation

72
00:03:21,759 --> 00:03:24,000
and immediately know the physical geometry of

73
00:03:24,000 --> 00:03:27,580
the ideas. It's beautiful. It is. And we're going

74
00:03:27,580 --> 00:03:29,300
to need that structure because the first thing

75
00:03:29,300 --> 00:03:31,979
this text does is completely dismantle our intuition

76
00:03:31,979 --> 00:03:34,539
about the physical world. So let's start right

77
00:03:34,539 --> 00:03:36,680
at the foundation. Why do we even need something

78
00:03:36,680 --> 00:03:40,060
called measure theory to study probability? Well,

79
00:03:40,099 --> 00:03:42,199
we need it because probability at its core is

80
00:03:42,199 --> 00:03:44,460
just... the act of measuring the likelihood of

81
00:03:44,460 --> 00:03:46,719
events within a space of all possible outcomes.

82
00:03:47,060 --> 00:03:49,580
Okay, makes sense. If the outcomes are finite,

83
00:03:49,800 --> 00:03:51,979
like the six sides of a die, it's super easy.

84
00:03:52,099 --> 00:03:54,520
You just count them. But the moment you introduce

85
00:03:54,520 --> 00:03:58,199
continuous space, or infinity, the act of measuring

86
00:03:58,199 --> 00:04:02,259
becomes shockingly paradoxical. Yeah. The text

87
00:04:02,259 --> 00:04:03,979
hits us with this right out of the gate in the

88
00:04:03,979 --> 00:04:06,139
very first chapter. They present a scenario that

89
00:04:06,139 --> 00:04:09,180
genuinely broke my brain for a minute. The Cantor

90
00:04:09,180 --> 00:04:12,580
set. Oh, the Cantor set. It's a classic. Let's

91
00:04:12,580 --> 00:04:14,219
walk through this because it proves how completely

92
00:04:14,219 --> 00:04:17,439
our intuition fails us. So imagine you have a

93
00:04:17,439 --> 00:04:19,680
solid line segment, let's call it a piece of

94
00:04:19,680 --> 00:04:22,699
wood, that is exactly one meter long. Okay, a

95
00:04:22,699 --> 00:04:25,800
standard measurable object. Right. Now you take

96
00:04:25,800 --> 00:04:28,300
a saw and you carve away the exact middle third

97
00:04:28,300 --> 00:04:30,660
of that wood. You remove the segment from one

98
00:04:30,660 --> 00:04:33,519
third to two thirds. So now you're left with

99
00:04:33,519 --> 00:04:36,139
two smaller pieces of wood, each one third of

100
00:04:36,139 --> 00:04:38,579
a meter long. Makes perfect physical sense so

101
00:04:38,579 --> 00:04:41,860
far. And then the recursive process begins. You

102
00:04:41,860 --> 00:04:44,680
take those two remaining pieces and you carve

103
00:04:44,680 --> 00:04:47,639
out the exact middle third of each of them. So

104
00:04:47,639 --> 00:04:50,899
now you have four even smaller pieces. Okay.

105
00:04:50,939 --> 00:04:52,899
Let me do the math on this out loud to keep it

106
00:04:52,899 --> 00:04:55,889
straight. First, I removed a piece that was one

107
00:04:55,889 --> 00:04:58,769
third of a meter. Right. Next, I removed two

108
00:04:58,769 --> 00:05:01,250
pieces that were each one ninth of a meter. So

109
00:05:01,250 --> 00:05:04,189
that's two ninths. Yep. If I do it again to the

110
00:05:04,189 --> 00:05:06,930
four remaining pieces, I'm removing four pieces

111
00:05:06,930 --> 00:05:09,600
of one twenty -seventh of a meter. So I'm removing

112
00:05:09,600 --> 00:05:14,189
4 27ths. The series of wood I'm removing is 1

113
00:05:14,189 --> 00:05:19,370
3rd plus 2 9ths plus 4 27ths and so on added

114
00:05:19,370 --> 00:05:22,029
for an item. Right, which forms a standard geometric

115
00:05:22,029 --> 00:05:24,709
series. Exactly, a classic geometric series.

116
00:05:25,290 --> 00:05:28,569
So the initial term, let's call it A, is 1 3rd.

117
00:05:28,870 --> 00:05:31,509
The ratio R we were multiplying by each time

118
00:05:31,509 --> 00:05:35,189
is 2 3rds. The formula for the sum of an infinite

119
00:05:35,189 --> 00:05:39,370
geometric series is A divided by 1 minus R. Correct.

120
00:05:39,709 --> 00:05:43,110
So one third divided by, well, one minus two

121
00:05:43,110 --> 00:05:46,050
thirds is one third. And one third divided by

122
00:05:46,050 --> 00:05:49,290
one third equals exactly one. Exactly one. Which

123
00:05:49,290 --> 00:05:52,350
means if you repeat this process to infinity,

124
00:05:52,449 --> 00:05:54,810
the total length of the wood you have carved

125
00:05:54,810 --> 00:05:57,810
away and thrown in the trash is exactly one meter.

126
00:05:58,290 --> 00:06:02,149
But wait, I started with one meter and I threw

127
00:06:02,149 --> 00:06:04,829
away one meter. So logically, there should be

128
00:06:04,829 --> 00:06:06,730
nothing left. The length of whatever remains

129
00:06:06,730 --> 00:06:10,730
has to be absolutely zero. It is zero. But here's

130
00:06:10,730 --> 00:06:13,370
the paradox that forced mathematicians to invent

131
00:06:13,370 --> 00:06:15,949
measure theory in the first place. What you are

132
00:06:15,949 --> 00:06:18,649
left with is an invisible dust of points, the

133
00:06:18,649 --> 00:06:21,370
Cantor dust. And even though its total length

134
00:06:21,370 --> 00:06:24,230
is mathematically proven to be zero, the number

135
00:06:24,230 --> 00:06:27,069
of points remaining in that dust is uncountable.

136
00:06:27,230 --> 00:06:30,110
That's the crazy part. It is. In fact, there

137
00:06:30,110 --> 00:06:32,490
are exactly as many points left in that zero

138
00:06:32,490 --> 00:06:35,069
-length dust as there are real numbers in the

139
00:06:35,069 --> 00:06:37,310
entire universe. This is the part that is so

140
00:06:37,310 --> 00:06:39,550
incredibly hard to visualize. How can something

141
00:06:39,550 --> 00:06:43,009
have a physical length of zero but still contain

142
00:06:43,009 --> 00:06:45,990
an infinite, uncountable number of physical points?

143
00:06:46,389 --> 00:06:49,529
You have to forcibly separate the concept of

144
00:06:49,529 --> 00:06:51,970
cardinality, which is just the count of items

145
00:06:51,970 --> 00:06:55,129
in a set, from the concept of measure or length.

146
00:06:55,290 --> 00:06:57,670
Just because you have an infinitely dense collection

147
00:06:57,670 --> 00:06:59,889
of points doesn't mean they span a measurable

148
00:06:59,889 --> 00:07:03,290
distance. Huh. They are so fragmented, so completely

149
00:07:03,290 --> 00:07:05,250
disconnected from one another that they hold

150
00:07:05,250 --> 00:07:08,209
no continuous space. It basically proves that

151
00:07:08,209 --> 00:07:11,029
assembling an infinite number of points doesn't

152
00:07:11,029 --> 00:07:13,750
automatically build a line segment. Okay, so

153
00:07:13,750 --> 00:07:16,829
the Cantor set shows us that a massively infinite

154
00:07:16,829 --> 00:07:20,269
set can have a length of zero. That's weird,

155
00:07:20,350 --> 00:07:23,329
but I can kind of accept it as a quirk of infinite

156
00:07:23,329 --> 00:07:26,050
fragmentation. But the text doesn't stop there.

157
00:07:26,230 --> 00:07:30,439
It introduces something far worse. Oh, you mean

158
00:07:30,439 --> 00:07:33,860
example 1 .3? Yes. A set that literally possesses

159
00:07:33,860 --> 00:07:36,959
no length at all. It completely rejects the tape

160
00:07:36,959 --> 00:07:40,319
measure. Ah, the Vitale set. The text frames

161
00:07:40,319 --> 00:07:43,120
it brilliantly using the axiom of choice. And

162
00:07:43,120 --> 00:07:45,740
this is where the mathematics of reality truly

163
00:07:45,740 --> 00:07:48,180
breaks down if we aren't incredibly careful.

164
00:07:48,839 --> 00:07:50,839
Let's try to map this out carefully for everyone

165
00:07:50,839 --> 00:07:54,949
listening. Imagine a circle. We know its circumference

166
00:07:54,949 --> 00:07:57,670
is a fixed length, right? 2 pi. Standard geometry,

167
00:07:57,889 --> 00:08:00,430
yeah. Now, we're going to group every single

168
00:08:00,430 --> 00:08:02,550
point on that circle into infinite families,

169
00:08:02,790 --> 00:08:06,050
or equivalence classes. The rule for being in

170
00:08:06,050 --> 00:08:07,649
the same family is that you can reach another

171
00:08:07,649 --> 00:08:10,269
point by rotating along the circle by a purely

172
00:08:10,269 --> 00:08:13,470
rational amount. Right. So, to be mathematically

173
00:08:13,470 --> 00:08:16,389
precise, two points are in the same equivalence

174
00:08:16,389 --> 00:08:18,410
class if the difference between their angles

175
00:08:18,410 --> 00:08:21,910
is a rational multiple of pi. Got it. And because

176
00:08:21,910 --> 00:08:24,550
there are countably infinitely many rational

177
00:08:24,550 --> 00:08:27,459
numbers, Every single point on the circle belongs

178
00:08:27,459 --> 00:08:30,639
to a family that contains countably, infinitely

179
00:08:30,639 --> 00:08:33,659
many siblings, just spread out all over the circumference.

180
00:08:34,039 --> 00:08:36,799
Okay, so the circle is completely partitioned

181
00:08:36,799 --> 00:08:40,360
into these infinite rational families. Now we

182
00:08:40,360 --> 00:08:42,940
use the axiom of choice, which is a fundamental

183
00:08:42,940 --> 00:08:45,440
rule in set theory that says no matter how many

184
00:08:45,440 --> 00:08:48,059
buckets you have, you can reach in and pick exactly

185
00:08:48,059 --> 00:08:50,460
one item from each bucket to form a new set.

186
00:08:50,840 --> 00:08:54,149
So we form a congress. We pick exactly one representative

187
00:08:54,149 --> 00:08:56,950
point from every single infinite family on that

188
00:08:56,950 --> 00:09:00,309
circle. We take this Congress, this incredibly

189
00:09:00,309 --> 00:09:02,990
strange scattered subset of points, and we ask

190
00:09:02,990 --> 00:09:05,409
the most basic question possible. What is its

191
00:09:05,409 --> 00:09:09,269
length? Seems simple enough. Right. So to find

192
00:09:09,269 --> 00:09:11,809
out, we take our Congress and we rotate it by

193
00:09:11,809 --> 00:09:14,210
a rational amount. Because of how we defined

194
00:09:14,210 --> 00:09:17,350
our families, rotating this Congress shifts every

195
00:09:17,350 --> 00:09:19,450
single representative to a new sibling in its

196
00:09:19,450 --> 00:09:21,649
family. We stamp a copy of the Congress right

197
00:09:21,649 --> 00:09:24,590
there. Okay. Then we rotate to another rational

198
00:09:24,590 --> 00:09:27,570
angle, stamp another copy. And because there

199
00:09:27,570 --> 00:09:29,970
are countably infinitely many rational angles,

200
00:09:30,070 --> 00:09:32,750
we stamp countably infinitely many copies of

201
00:09:32,750 --> 00:09:34,870
our Congress. And because each point in a family

202
00:09:34,870 --> 00:09:36,929
corresponds to exactly one rational rotation

203
00:09:36,929 --> 00:09:39,830
from our representative, these infinitely many

204
00:09:39,830 --> 00:09:42,129
copies of the Congress will perfectly cover the

205
00:09:42,129 --> 00:09:45,610
entire circle. Yep. No overlapping points and

206
00:09:45,610 --> 00:09:48,750
no gaps. We have perfectly reassembled the circle.

207
00:09:48,870 --> 00:09:52,309
And now the trap springs. If we perfectly recreated

208
00:09:52,309 --> 00:09:54,850
the circle, the sum of the lengths of all of

209
00:09:54,850 --> 00:09:56,769
our infinite copies must equal the length of

210
00:09:56,769 --> 00:10:00,389
the circle, which is 2 pi. Obviously. And since

211
00:10:00,389 --> 00:10:02,809
every copy is just a rigid rotation of the original

212
00:10:02,809 --> 00:10:06,250
Congress, they must all have the exact same length.

213
00:10:06,610 --> 00:10:08,009
Well, it's called the length of the Congress

214
00:10:08,009 --> 00:10:12,269
L. We are adding L to itself countably infinitely

215
00:10:12,269 --> 00:10:17,460
many times, and it must equal 2 pi. So, what

216
00:10:17,809 --> 00:10:21,129
on earth is L? Well... Let's test the options.

217
00:10:21,370 --> 00:10:25,490
Okay. If the length of the congress is 0, then

218
00:10:25,490 --> 00:10:28,389
0 plus 0 plus 0 infinitely many times is still

219
00:10:28,389 --> 00:10:32,070
0, but it has to equal 2 pi. So the length cannot

220
00:10:32,070 --> 00:10:36,649
be 0. Exactly. Okay. So what's the alternative?

221
00:10:37,090 --> 00:10:39,950
The alternative is that the length is some positive

222
00:10:39,950 --> 00:10:43,169
number, no matter how microscopically tiny. Let's

223
00:10:43,169 --> 00:10:50,389
say L is 0 .0000001. But... If I add any fixed

224
00:10:50,389 --> 00:10:53,370
positive number to itself infinitely many times,

225
00:10:53,470 --> 00:10:56,980
it immediately explodes to infinity. Right. But

226
00:10:56,980 --> 00:11:00,539
it has to equal exactly 2 pi. So the length cannot

227
00:11:00,539 --> 00:11:03,779
be greater than zero either. It cannot be zero.

228
00:11:04,080 --> 00:11:07,440
It cannot be greater than zero. The concept of

229
00:11:07,440 --> 00:11:10,940
length simply ceases to exist for this set. It

230
00:11:10,940 --> 00:11:14,679
is entirely non -measurable. If a physical tape

231
00:11:14,679 --> 00:11:16,259
measure encountered this congress of points,

232
00:11:16,399 --> 00:11:19,870
the universe would just crash. It's an unresolvable

233
00:11:19,870 --> 00:11:22,750
error. Which brings us to the entire point of

234
00:11:22,750 --> 00:11:25,309
measure theory. These paradoxes are not just

235
00:11:25,309 --> 00:11:27,730
mathematical parlor tricks, you know. They are

236
00:11:27,730 --> 00:11:29,909
warning signs. Warning is about what exactly?

237
00:11:30,169 --> 00:11:33,029
They prove that in a massive continuous space

238
00:11:33,029 --> 00:11:35,649
like the real number line, you cannot safely

239
00:11:35,649 --> 00:11:38,730
assign a measure, whether that's length or area

240
00:11:38,730 --> 00:11:41,649
or probability, to every conceivable subset.

241
00:11:42,299 --> 00:11:44,500
If you insist that every random collection of

242
00:11:44,500 --> 00:11:47,220
points must have a probability, you will eventually

243
00:11:47,220 --> 00:11:49,799
calculate that the probability of an event happening

244
00:11:49,799 --> 00:11:53,639
is simultaneously zero and infinity. So how do

245
00:11:53,639 --> 00:11:56,820
Croce and Botev solve this? How do we fix the

246
00:11:56,820 --> 00:11:59,799
broken tape measure? We don't fix it. We build

247
00:11:59,799 --> 00:12:02,080
a fence around the problem. We introduce the

248
00:12:02,080 --> 00:12:05,399
concept of a sigma algebra. I thought of a sigma

249
00:12:05,399 --> 00:12:07,720
algebra as a kind of mathematical quarantine

250
00:12:07,720 --> 00:12:11,090
zone. Like, we look at the terrifying, glitchy

251
00:12:11,090 --> 00:12:12,889
universe of all possible subsets and we say,

252
00:12:12,929 --> 00:12:15,669
no, we are only going to deal with a specific,

253
00:12:15,789 --> 00:12:18,950
safe collection of subsets. And we will only

254
00:12:18,950 --> 00:12:21,250
assign probabilities to the sets inside this

255
00:12:21,250 --> 00:12:23,409
safe zone. That's a very accurate way to visualize

256
00:12:23,409 --> 00:12:26,450
it. A sigma algebra is just a collection of sets

257
00:12:26,450 --> 00:12:28,990
that agrees to obey three fundamental rules of

258
00:12:28,990 --> 00:12:31,990
logic. Okay, what are they? First, it must contain

259
00:12:31,990 --> 00:12:35,110
the entire space itself. Second, it must be closed

260
00:12:35,110 --> 00:12:37,769
under complements, meaning if the event it rains

261
00:12:37,769 --> 00:12:41,070
is in our safe zone, the event it does not rain

262
00:12:41,070 --> 00:12:43,490
must also be in the safe zone. You can't measure

263
00:12:43,490 --> 00:12:45,490
an event without being able to measure its opposite.

264
00:12:45,789 --> 00:12:48,470
OK, that makes intuitive sense. What's the third

265
00:12:48,470 --> 00:12:51,330
rule? It must be closed under countable unions.

266
00:12:51,669 --> 00:12:54,289
If you have a countable sequence of safe measurable

267
00:12:54,289 --> 00:12:56,769
events, event one, event two, event three, all

268
00:12:56,769 --> 00:12:59,149
the way out to infinity, then the scenario where

269
00:12:59,149 --> 00:13:02,509
at least one of these events happens. must also

270
00:13:02,509 --> 00:13:05,710
be a safe, measurable event. Right. As long as

271
00:13:05,710 --> 00:13:07,710
a collection of sets follows these three rules,

272
00:13:07,870 --> 00:13:11,029
it forms a sigma algebra. The monstrous sets,

273
00:13:11,110 --> 00:13:13,490
like the Vitale set without length, simply fail

274
00:13:13,490 --> 00:13:15,509
these rules and are securely locked outside the

275
00:13:15,509 --> 00:13:18,070
quarantine zone. So we essentially agree to be

276
00:13:18,070 --> 00:13:21,110
slightly blind to the deepest chaotic fractions

277
00:13:21,110 --> 00:13:23,850
of the mathematical universe, just so we can

278
00:13:23,850 --> 00:13:26,110
build a coherent theory of probability. Exactly.

279
00:13:26,830 --> 00:13:29,110
And once we establish this safe measurable zone,

280
00:13:29,350 --> 00:13:31,929
we can finally redefine what probability actually

281
00:13:31,929 --> 00:13:35,049
is. Yeah, we can move beyond the fragmented ways

282
00:13:35,049 --> 00:13:37,889
we are taught probability in school. The text

283
00:13:37,889 --> 00:13:40,570
takes this measure theoretic foundation and uses

284
00:13:40,570 --> 00:13:43,110
it to completely unify discrete and continuous

285
00:13:43,110 --> 00:13:45,629
probability. Right, because usually you treat

286
00:13:45,629 --> 00:13:47,889
rolling a die and measuring a person's height

287
00:13:47,889 --> 00:13:50,429
as two entirely different types of math. For

288
00:13:50,429 --> 00:13:52,690
the die, you use basic addition. The probability

289
00:13:52,690 --> 00:13:56,149
of rolling a 1 or a 2 is 1 sixth plus 1 sixth.

290
00:13:56,309 --> 00:13:59,149
Simple discrete math. But for height, you have

291
00:13:59,149 --> 00:14:02,090
to use calculus. You integrate a continuous probability

292
00:14:02,090 --> 00:14:05,350
density function over a range. They feel like

293
00:14:05,350 --> 00:14:07,769
completely different mechanics. But under measure

294
00:14:07,769 --> 00:14:11,289
theory, they are identical. A probability space

295
00:14:11,289 --> 00:14:14,330
is defined simply as a triplet. A sample space

296
00:14:14,330 --> 00:14:17,149
of all outcomes, a sigma algebra of measurable

297
00:14:17,149 --> 00:14:20,830
events. And a probability measure. And the measure

298
00:14:20,830 --> 00:14:23,909
is the key there. Yes. This measure is just a

299
00:14:23,909 --> 00:14:26,090
function that assigns a number between 0 and

300
00:14:26,090 --> 00:14:29,730
1 to every event inside the sigma algebra. It

301
00:14:29,730 --> 00:14:31,889
doesn't care if the underlying space is discrete

302
00:14:31,889 --> 00:14:34,929
integers or a continuous curve. The measure theory

303
00:14:34,929 --> 00:14:37,029
handles the integration seamlessly across both.

304
00:14:37,090 --> 00:14:40,169
It's one unified engine. And once we have that

305
00:14:40,169 --> 00:14:42,409
unified engine, the tech starts putting it into

306
00:14:42,409 --> 00:14:45,269
motion. Because static probability spaces don't

307
00:14:45,269 --> 00:14:48,049
capture how the world actually works. Reality

308
00:14:48,049 --> 00:14:50,990
unfolds over time. The stock market changes daily.

309
00:14:51,110 --> 00:14:54,549
Weather patterns shift. So they introduce stochastic

310
00:14:54,549 --> 00:14:57,909
processes. Right. Chapter two. Definition 2 .16

311
00:14:57,909 --> 00:15:00,629
states that a stochastic process is a collection

312
00:15:00,629 --> 00:15:03,009
of random variables indexed by a set, usually

313
00:15:03,009 --> 00:15:05,809
representing time. If a single random variable

314
00:15:05,809 --> 00:15:08,509
is a photograph of uncertainty, a stochastic

315
00:15:08,509 --> 00:15:11,149
process is the motion picture. Oh, that's a great

316
00:15:11,149 --> 00:15:13,759
way to put it. The foundational example they

317
00:15:13,759 --> 00:15:17,360
start with is the Bernoulli process, just a sequence

318
00:15:17,360 --> 00:15:20,000
of independent coin flips, a random variable

319
00:15:20,000 --> 00:15:25,080
x1, x2, x3, each taking the value 1 with probability

320
00:15:25,080 --> 00:15:29,539
p and 0 with probability 1 minus p. It seems

321
00:15:29,539 --> 00:15:32,690
incredibly basic. It is basic. But the authors

322
00:15:32,690 --> 00:15:35,769
emphasize that it is the DNA from which massive

323
00:15:35,769 --> 00:15:38,850
complexity is built. But before they build that

324
00:15:38,850 --> 00:15:41,250
complexity, they take a bit of a detour into

325
00:15:41,250 --> 00:15:43,750
the geometry of these random variables. They

326
00:15:43,750 --> 00:15:47,289
explore variance and covariance. This was a brilliant

327
00:15:47,289 --> 00:15:50,289
section. I never thought of probability geometrically

328
00:15:50,289 --> 00:15:52,629
before. I always thought of covariance as just,

329
00:15:52,669 --> 00:15:55,070
you know, a statistical formula you plug into

330
00:15:55,070 --> 00:15:57,710
Excel to see if two stocks move together. Right,

331
00:15:57,789 --> 00:16:00,250
but the text reframes random variables with a

332
00:16:00,250 --> 00:16:02,850
mean of zero as literal vectors in a mathematical

333
00:16:02,850 --> 00:16:05,909
space, specifically a Hilbert space. Let me make

334
00:16:05,909 --> 00:16:08,350
sure I'm visualizing this correctly. If the random

335
00:16:08,350 --> 00:16:11,789
variables are vectors, the origin point, zero,

336
00:16:11,950 --> 00:16:14,750
is their mean. Precisely. If we center them so

337
00:16:14,750 --> 00:16:17,230
their expectation is zero, the variance of a

338
00:16:17,230 --> 00:16:19,210
random variable is no longer just a measure of

339
00:16:19,210 --> 00:16:22,009
statistical spread. It is exactly equivalent

340
00:16:22,009 --> 00:16:25,169
to the squared length of that vector. Okay, so

341
00:16:25,169 --> 00:16:28,750
a high -variance stock is a very long vector

342
00:16:28,750 --> 00:16:31,850
stretching far away from the origin. Yes. And

343
00:16:31,850 --> 00:16:34,590
the covariance between two random variables becomes

344
00:16:34,590 --> 00:16:37,470
the inner product, the dot product between their

345
00:16:37,470 --> 00:16:39,590
two vectors. It literally measures the angle

346
00:16:39,590 --> 00:16:43,350
between them. Wow. Yeah. If two random variables

347
00:16:43,350 --> 00:16:45,590
are completely uncorrelated, their vectors are

348
00:16:45,590 --> 00:16:48,070
perfectly perpendicular. They form a 90 -degree

349
00:16:48,070 --> 00:16:51,429
angle. Their inner product is exactly zero. That's

350
00:16:51,429 --> 00:16:55,399
incredible. And this geometric perspective naturally

351
00:16:55,399 --> 00:16:57,399
spits out one of the most famous inequalities

352
00:16:57,399 --> 00:16:59,899
in math, right? The Cauchy -Schwarz inequality.

353
00:17:00,399 --> 00:17:03,899
It does. Because in pure geometry, the dot product

354
00:17:03,899 --> 00:17:06,099
of two vectors can never be larger than the product

355
00:17:06,099 --> 00:17:08,519
of their individual lengths. It's impossible

356
00:17:08,519 --> 00:17:10,759
for the angle to make the projection longer than

357
00:17:10,759 --> 00:17:13,680
the vectors themselves. Right. That's just basic

358
00:17:13,680 --> 00:17:16,549
geometry. But translated back into probability

359
00:17:16,549 --> 00:17:19,410
via this text, it proves that the covariance

360
00:17:19,410 --> 00:17:21,650
of two random variables can never exceed the

361
00:17:21,650 --> 00:17:24,650
product of their standard deviations. It grounds

362
00:17:24,650 --> 00:17:28,650
an abstract statistical rule in undeniable physical

363
00:17:28,650 --> 00:17:31,369
geometry. You can literally see why the math

364
00:17:31,369 --> 00:17:34,630
works. And this visualization really helps when

365
00:17:34,630 --> 00:17:37,549
we move to the next major conceptual leap, how

366
00:17:37,549 --> 00:17:40,730
information itself grows over time. The text

367
00:17:40,730 --> 00:17:43,930
introduces the concept of filtrations. I spent

368
00:17:43,930 --> 00:17:45,990
a lot of time on this because filtrations become

369
00:17:45,990 --> 00:17:48,269
the absolute backbone for everything that happens

370
00:17:48,269 --> 00:17:51,589
later in the text with Martin Gales and Ito Calculus.

371
00:17:52,200 --> 00:17:54,859
Let me try an analogy on you for this one. I

372
00:17:54,859 --> 00:17:57,380
think of a filtration as watching a complex mystery

373
00:17:57,380 --> 00:18:00,720
movie. Oh, I like that. Go ahead. When you start

374
00:18:00,720 --> 00:18:03,559
the movie at minute zero, your sigma algebra

375
00:18:03,559 --> 00:18:06,420
of knowledge, the collection of events you can

376
00:18:06,420 --> 00:18:08,740
perfectly measure and know the outcome of, is

377
00:18:08,740 --> 00:18:11,319
essentially empty. You know absolutely nothing.

378
00:18:11,700 --> 00:18:15,660
A blank slate. Right. At minute 10, your filtration

379
00:18:15,660 --> 00:18:18,619
has expanded slightly. You know who the victim

380
00:18:18,619 --> 00:18:21,880
is. Maybe you saw a blurry shadow running away.

381
00:18:22,240 --> 00:18:25,220
You can definitively assign a probability of

382
00:18:25,220 --> 00:18:28,619
1 or 0 to certain basic events. You've locked

383
00:18:28,619 --> 00:18:31,480
in some reality. Right. But the identity of the

384
00:18:31,480 --> 00:18:34,900
killer is still a highly uncertain random variable.

385
00:18:35,359 --> 00:18:38,640
As the movie progresses to minute 90, your filtration

386
00:18:38,640 --> 00:18:41,970
has grown massively. contains almost all the

387
00:18:41,970 --> 00:18:45,009
clues, the motives, the alibis. Your sigma algebra

388
00:18:45,009 --> 00:18:47,430
now encompasses a huge number of events that

389
00:18:47,430 --> 00:18:50,609
are fixed and known. That maps perfectly to the

390
00:18:50,609 --> 00:18:53,329
rigorous definition. A filtration is a sequence

391
00:18:53,329 --> 00:18:57,390
of sigma algebras, let's call them F1, F2, F3,

392
00:18:57,509 --> 00:19:00,029
where each one is completely contained within

393
00:19:00,029 --> 00:19:03,730
the next. F1 is a subset of F2, which is a subset

394
00:19:03,730 --> 00:19:07,299
of F3. So it only grows. Exactly. It mathematically

395
00:19:07,299 --> 00:19:09,480
models the accumulation of knowledge without

396
00:19:09,480 --> 00:19:13,680
memory loss. At time t, the filtration f subscript

397
00:19:13,680 --> 00:19:16,559
t contains every single event whose outcome can

398
00:19:16,559 --> 00:19:19,220
be absolutely determined by observing the stochastic

399
00:19:19,220 --> 00:19:22,099
process up to that exact moment. The universe

400
00:19:22,099 --> 00:19:24,740
of things we are certain about is constantly

401
00:19:24,740 --> 00:19:27,579
expanding. And with this framework of time and

402
00:19:27,579 --> 00:19:29,519
information, the authors introduce what we might

403
00:19:29,519 --> 00:19:33,160
call the heavy artillery of the text. The fundamental

404
00:19:33,160 --> 00:19:36,119
stochastic processes. They start with Gaussian

405
00:19:36,119 --> 00:19:39,140
processes. Which is the continuous generalization

406
00:19:39,140 --> 00:19:41,859
of the normal distribution, right? Yes. If you

407
00:19:41,859 --> 00:19:44,700
sample a Gaussian process at any finite set of

408
00:19:44,700 --> 00:19:49,160
times, say T1, T5, and T20, the joint distribution

409
00:19:49,160 --> 00:19:51,920
of the process at those exact moments will always

410
00:19:51,920 --> 00:19:54,740
form a multivariate normal distribution. So it's

411
00:19:54,740 --> 00:19:57,539
very structured. Highly structured. They are

412
00:19:57,539 --> 00:20:00,569
entirely defined by just two things. a mean function,

413
00:20:00,809 --> 00:20:03,349
which tracks the average path over time, and

414
00:20:03,349 --> 00:20:05,869
a covariance function, which dictates how a bump

415
00:20:05,869 --> 00:20:08,849
at time t1 correlates with a bump at time t5.

416
00:20:09,289 --> 00:20:11,490
They are the perfect model for systems where

417
00:20:11,490 --> 00:20:13,990
randomness is the result of millions of tiny

418
00:20:13,990 --> 00:20:16,650
independent shocks adding together. But not everything

419
00:20:16,650 --> 00:20:20,430
in the universe is a tiny continuous shock. Sometimes

420
00:20:20,430 --> 00:20:22,869
things happen abruptly. A phone call arrives

421
00:20:22,869 --> 00:20:26,250
at a switchboard. An alpha particle hits a Geiger

422
00:20:26,250 --> 00:20:29,759
counter. A stock market flash crashes. That brings

423
00:20:29,759 --> 00:20:32,339
us to Poisson random measures and processes.

424
00:20:32,759 --> 00:20:36,119
Right. This is the mathematics of counting unpredictable

425
00:20:36,119 --> 00:20:39,279
arrivals. The Poisson process measures discrete

426
00:20:39,279 --> 00:20:42,440
events happening in continuous time. And the

427
00:20:42,440 --> 00:20:44,839
text provides an incredibly elegant derivation

428
00:20:44,839 --> 00:20:46,400
for this called the Bernoulli approximation.

429
00:20:47,039 --> 00:20:49,299
It takes that basic coin flip we talked about

430
00:20:49,299 --> 00:20:52,220
earlier and stretches it out to infinity to create

431
00:20:52,220 --> 00:20:55,000
a Poisson process. Let's walk through the math

432
00:20:55,000 --> 00:20:57,099
of how this actually happens because it's a beautiful

433
00:20:57,099 --> 00:20:59,759
proof. It really demonstrates the power of limits.

434
00:21:00,200 --> 00:21:02,220
Let's say we want to model the number of phone

435
00:21:02,220 --> 00:21:04,220
calls arriving at a call center over one hour.

436
00:21:04,319 --> 00:21:07,319
We know on average we get lambda calls per hour.

437
00:21:07,480 --> 00:21:09,920
But they arrive randomly. Right. They arrive

438
00:21:09,920 --> 00:21:12,920
randomly. So we take that one hour and we slice

439
00:21:12,920 --> 00:21:16,599
it into n tiny microscopic intervals, milliseconds

440
00:21:16,599 --> 00:21:19,980
or even nanoseconds. We slice it so thinly that

441
00:21:19,980 --> 00:21:22,140
the probability of getting two phone calls in

442
00:21:22,140 --> 00:21:25,200
a single interval is virtually zero. So each

443
00:21:25,200 --> 00:21:28,259
tiny interval becomes a Bernoulli trial. a microscopic

444
00:21:28,259 --> 00:21:31,539
coin flip. It's either a 1, meaning a call arrives,

445
00:21:31,759 --> 00:21:34,619
or a 0, meaning no call. And the probability

446
00:21:34,619 --> 00:21:37,079
of success in that tiny interval is our average

447
00:21:37,079 --> 00:21:40,279
rate, lambda, divided by n intervals, so the

448
00:21:40,279 --> 00:21:43,440
probability is lambda over n. Now we ask a simple

449
00:21:43,440 --> 00:21:45,839
question. What is the probability that we get

450
00:21:45,839 --> 00:21:48,759
exactly zero phone calls in the entire hour?

451
00:21:49,259 --> 00:21:51,579
Well, that means we must get a failure on the

452
00:21:51,579 --> 00:21:53,779
first interval, and a failure on the second,

453
00:21:53,960 --> 00:21:56,579
and a failure on the third, all the way to n.

454
00:21:57,079 --> 00:21:59,599
And since they are independent, the probability

455
00:21:59,599 --> 00:22:02,799
of failure on one interval is 1 minus lambda

456
00:22:02,799 --> 00:22:06,200
over n. We just multiply them together n times.

457
00:22:06,910 --> 00:22:09,950
So the probability of zero calls in an hour is

458
00:22:09,950 --> 00:22:13,210
1 minus lambda over n raised to the power of

459
00:22:13,210 --> 00:22:16,289
n. And here is where the calculus ticks over.

460
00:22:16,750 --> 00:22:20,390
What happens as n approaches infinity? As we

461
00:22:20,390 --> 00:22:23,819
slice time into infinitely thin slivers. The

462
00:22:23,819 --> 00:22:26,940
limit as n goes to infinity of 1 minus lambda

463
00:22:26,940 --> 00:22:30,420
over n all raised to the n is the exact mathematical

464
00:22:30,420 --> 00:22:33,660
definition of the exponential function e to the

465
00:22:33,660 --> 00:22:36,960
negative lambda. Which is exactly the probability

466
00:22:36,960 --> 00:22:40,279
mass function for a Poisson random variable evaluated

467
00:22:40,279 --> 00:22:43,559
at zero. By chopping time into infinite coin

468
00:22:43,559 --> 00:22:47,440
flips, the discrete binomial math perfectly morphs

469
00:22:47,440 --> 00:22:49,460
into the continuous exponential waiting times

470
00:22:49,460 --> 00:22:52,740
of a Poisson process. It's just... Breathtaking

471
00:22:52,740 --> 00:22:55,940
how the math snaps together. It really is. And

472
00:22:55,940 --> 00:22:57,859
the authors don't just leave it as a theoretical

473
00:22:57,859 --> 00:23:01,079
beauty either. They include algorithm 2 .87 to

474
00:23:01,079 --> 00:23:03,500
show you how to physically simulate this continuous

475
00:23:03,500 --> 00:23:06,220
randomness on a computer, which, you know, can

476
00:23:06,220 --> 00:23:08,460
only generate discrete random numbers. Right,

477
00:23:08,500 --> 00:23:10,539
because a computer uses a uniform distribution

478
00:23:10,539 --> 00:23:13,740
to pick random numbers between 0 and 1. It's

479
00:23:13,740 --> 00:23:17,779
flat. Every decimal is equally likely. How do

480
00:23:17,779 --> 00:23:21,559
you get from flat... uniform randomness to the

481
00:23:21,559 --> 00:23:24,900
organic, unpredictable, exponential heartbeat

482
00:23:24,900 --> 00:23:29,200
of a Poisson process? You use the inverse transform

483
00:23:29,200 --> 00:23:32,579
method. You take your uniform random number,

484
00:23:32,660 --> 00:23:35,339
let's call it u, and you run it through a specific

485
00:23:35,339 --> 00:23:38,359
function. You take the natural logarithm of u

486
00:23:38,359 --> 00:23:41,339
and divide it by negative lambda. Oh, wow. Yeah,

487
00:23:41,400 --> 00:23:43,400
because of the deep connection between uniform

488
00:23:43,400 --> 00:23:46,640
and exponential distributions, this simple mathematical

489
00:23:46,640 --> 00:23:50,599
alchemy, taking the log, warps the flat randomness

490
00:23:50,599 --> 00:23:54,059
into the exact skewed waiting times between Poisson

491
00:23:54,059 --> 00:23:57,400
arrivals. It's like pouring liquid metal into

492
00:23:57,400 --> 00:24:00,539
a mold, and out pops a perfectly formed gear.

493
00:24:01,579 --> 00:24:04,819
And this connects to the third major category

494
00:24:04,819 --> 00:24:09,049
they cover, Levy processes. Levy processes are

495
00:24:09,049 --> 00:24:11,890
the grand unifiers of continuous time probability.

496
00:24:12,490 --> 00:24:15,549
A levy process is basically any process that

497
00:24:15,549 --> 00:24:18,509
has independent and stationary increments. Let's

498
00:24:18,509 --> 00:24:21,529
break that down. Stationary increments mean the

499
00:24:21,529 --> 00:24:24,500
rules don't change over time. The probability

500
00:24:24,500 --> 00:24:26,960
of the process jumping up by five points between

501
00:24:26,960 --> 00:24:30,680
minute 10 and minute 11 is exactly the same as

502
00:24:30,680 --> 00:24:32,940
it jumping up five points between minute 100

503
00:24:32,940 --> 00:24:36,569
and minute 101. And independent increments mean

504
00:24:36,569 --> 00:24:39,190
the amnesia we discussed earlier. What happens

505
00:24:39,190 --> 00:24:41,789
between minute 10 and 11 gives you absolutely

506
00:24:41,789 --> 00:24:44,029
zero information about what will happen between

507
00:24:44,029 --> 00:24:46,869
minute 11 and 12. And the text points out something

508
00:24:46,869 --> 00:24:49,650
that feels almost like a cheat code for probability.

509
00:24:50,630 --> 00:24:53,589
Every single Levy process in existence, whether

510
00:24:53,589 --> 00:24:55,890
it's the smooth wandering of Brownian motion

511
00:24:55,890 --> 00:24:59,250
or the jagged jumping arrival of a Poisson process,

512
00:24:59,589 --> 00:25:02,269
can be completely characterized by just three

513
00:25:02,269 --> 00:25:06,039
parameters. the characteristic triplet. Drift,

514
00:25:06,039 --> 00:25:09,039
variance, and the jump measure. Drift is the

515
00:25:09,039 --> 00:25:11,339
deterministic trend, you know, the slow slide

516
00:25:11,339 --> 00:25:14,720
in one direction. Variance dictates the continuous

517
00:25:14,720 --> 00:25:18,039
Gaussian jitter around that trend. And the jump

518
00:25:18,039 --> 00:25:20,940
measure acts as a blueprint for the sudden discontinuous

519
00:25:20,940 --> 00:25:23,900
spikes or crashes. So it's just three dials.

520
00:25:23,900 --> 00:25:26,160
Yep, three dials. If you lock in those three

521
00:25:26,160 --> 00:25:28,539
things, you have entirely defined the universe

522
00:25:28,539 --> 00:25:31,450
of that process. They highlight the gamma process

523
00:25:31,450 --> 00:25:33,990
as a fascinating example of a Levy process that

524
00:25:33,990 --> 00:25:37,230
has no continuous variance at all. It moves entirely

525
00:25:37,230 --> 00:25:40,289
by an infinite number of tiny positive jumps.

526
00:25:40,910 --> 00:25:44,690
So we've built a foundation. We have measure

527
00:25:44,690 --> 00:25:47,430
theory. We have defined probability spaces. We

528
00:25:47,430 --> 00:25:50,049
have filtrations expanding our knowledge. And

529
00:25:50,049 --> 00:25:52,789
we have these massive stochastic processes running

530
00:25:52,789 --> 00:25:56,819
over time. This naturally forces us to ask, where

531
00:25:56,819 --> 00:25:59,710
are they actually going? If you let time run

532
00:25:59,710 --> 00:26:02,390
to infinity, what happens to the randomness?

533
00:26:02,589 --> 00:26:05,609
Which brings us to chapter three of the text,

534
00:26:05,849 --> 00:26:10,470
convergence. Now, if we were just looking at

535
00:26:10,470 --> 00:26:13,869
standard calculus, convergence is easy. A sequence

536
00:26:13,869 --> 00:26:17,210
of numbers like one, one half, one quarter, one

537
00:26:17,210 --> 00:26:19,230
eighth. I shouldn't say dot, dot, dot. I mean,

538
00:26:19,269 --> 00:26:21,910
it goes on forever. And it either gets closer

539
00:26:21,910 --> 00:26:23,869
and closer to a specific target number, in this

540
00:26:23,869 --> 00:26:27,589
case zero, or it doesn't. It's a clean, deterministic

541
00:26:27,589 --> 00:26:30,809
path. Very clean. But when your sequence is made

542
00:26:30,809 --> 00:26:34,480
of random variables, what does it mean for randomness

543
00:26:34,480 --> 00:26:37,849
to converge? It's not a single point moving.

544
00:26:38,269 --> 00:26:41,009
It's a shifting, breathing cloud of probability.

545
00:26:41,369 --> 00:26:44,710
This is where probability theory becomes extremely

546
00:26:44,710 --> 00:26:48,630
demanding because the cloud can settle down in

547
00:26:48,630 --> 00:26:51,089
different ways. The authors methodically unpack

548
00:26:51,089 --> 00:26:54,549
four distinct modes of convergence. And it is

549
00:26:54,549 --> 00:26:56,630
vital to understand the difference between them

550
00:26:56,630 --> 00:26:58,769
because confusing them will lead to catastrophic

551
00:26:58,769 --> 00:27:02,789
errors in forecasting. The four modes are almost

552
00:27:03,009 --> 00:27:05,890
sure convergence, convergence in probability,

553
00:27:06,289 --> 00:27:09,529
convergence in distribution, and convergence

554
00:27:09,529 --> 00:27:13,140
in LP norm. I'll be trying to find a good physical

555
00:27:13,140 --> 00:27:16,559
analogy for this that isn't just reciting formulas.

556
00:27:17,789 --> 00:27:20,170
Think about predicting the path of a massive

557
00:27:20,170 --> 00:27:22,210
hurricane moving across the ocean toward the

558
00:27:22,210 --> 00:27:25,390
coast. Okay, a highly chaotic stochastic system.

559
00:27:26,190 --> 00:27:27,970
Let's see how the modes apply to the hurricane.

560
00:27:28,269 --> 00:27:30,450
Okay, let's start with the strongest one, almost

561
00:27:30,450 --> 00:27:33,930
sure convergence. In our analogy, this means

562
00:27:33,930 --> 00:27:36,130
that if you look at the infinite universe of

563
00:27:36,130 --> 00:27:38,509
all possible physical paths this hurricane could

564
00:27:38,509 --> 00:27:41,309
take, every twist, every turn, every stalling

565
00:27:41,309 --> 00:27:44,210
out, the set of paths where it ultimately hits

566
00:27:44,210 --> 00:27:47,630
our specific coastal city has a probability measure

567
00:27:47,630 --> 00:27:51,890
of exactly one. Wow, okay. It might wobble violently

568
00:27:51,890 --> 00:27:54,430
on Tuesday, it might hook south on Wednesday,

569
00:27:54,569 --> 00:27:57,490
but we are absolutely mathematically certain

570
00:27:57,490 --> 00:28:00,789
that the exact specific storm will eventually

571
00:28:00,789 --> 00:28:03,349
arrive at our coordinates. The sample paths themselves

572
00:28:03,349 --> 00:28:07,349
stabilize. The key there is the sample paths

573
00:28:07,349 --> 00:28:10,730
stabilize. You are making a claim about the holistic

574
00:28:10,730 --> 00:28:13,130
realization of the randomness over the entire

575
00:28:13,130 --> 00:28:16,089
stretch of time. Now, what about convergence

576
00:28:16,089 --> 00:28:19,839
in probability? This is a weaker mode. It doesn't

577
00:28:19,839 --> 00:28:22,519
guarantee the entire path. It says that if you

578
00:28:22,519 --> 00:28:24,680
take a snapshot of the hurricane at any given

579
00:28:24,680 --> 00:28:27,440
hour, the statistical chance that the hurricane

580
00:28:27,440 --> 00:28:29,680
is far away from our predicted trajectory is

581
00:28:29,680 --> 00:28:32,059
shrinking towards zero as time goes on. Okay,

582
00:28:32,119 --> 00:28:34,839
so it could still jump. Right. It might occasionally

583
00:28:34,839 --> 00:28:36,960
dart outside the cone of uncertainty for a split

584
00:28:36,960 --> 00:28:39,759
second, meaning the continuous path isn't perfectly

585
00:28:39,759 --> 00:28:42,900
locked in stone. But the odds of finding it way

586
00:28:42,900 --> 00:28:45,819
off course at any specific moment become mathematically

587
00:28:45,819 --> 00:28:49,140
negligible. Exactly. The formal definition is

588
00:28:49,140 --> 00:28:51,160
that the probability of the absolute difference

589
00:28:51,160 --> 00:28:54,319
between the sequence x sub n and the limit x,

590
00:28:54,480 --> 00:28:57,119
being greater than any arbitrary tiny margin,

591
00:28:57,359 --> 00:29:01,019
epsilon, goes to zero as n approaches infinity.

592
00:29:01,220 --> 00:29:03,619
You are basically measuring the shrinking volume

593
00:29:03,619 --> 00:29:06,829
of the error. What about convergence in distribution?

594
00:29:07,539 --> 00:29:10,680
This is the weakest mode, but heavily used in

595
00:29:10,680 --> 00:29:14,440
statistics. In convergence in distribution, our

596
00:29:14,440 --> 00:29:16,460
specific hurricane might not be coming anywhere

597
00:29:16,460 --> 00:29:20,009
near our city at all. But as we observe more

598
00:29:20,009 --> 00:29:22,730
and more hurricanes over the decades, the overall

599
00:29:22,730 --> 00:29:26,309
pattern of where hurricanes land begins to perfectly

600
00:29:26,309 --> 00:29:29,190
match a theoretical bell curve of risk we drew

601
00:29:29,190 --> 00:29:31,829
on a map. Oh, I see. We aren't predicting this

602
00:29:31,829 --> 00:29:34,349
specific storm anymore. We are predicting the

603
00:29:34,349 --> 00:29:37,799
shape of the danger. The cumulative distribution

604
00:29:37,799 --> 00:29:40,339
functions align, even if the underlying random

605
00:29:40,339 --> 00:29:43,160
variables are completely disconnected. That is

606
00:29:43,160 --> 00:29:45,420
the essence of it. It's why we use it for the

607
00:29:45,420 --> 00:29:47,549
central limit theorem. We only care about the

608
00:29:47,549 --> 00:29:50,710
shape of the aggregate data. And finally, what

609
00:29:50,710 --> 00:29:54,990
about convergence in L sub p norm? This one is

610
00:29:54,990 --> 00:29:57,470
about the magnitude of the error. It means that

611
00:29:57,470 --> 00:30:00,109
the expected value, the mathematical average,

612
00:30:00,349 --> 00:30:02,609
of the distance between the hurricane's location

613
00:30:02,609 --> 00:30:06,710
and our prediction, raised to some power p, is

614
00:30:06,710 --> 00:30:09,890
shrinking to zero. It's heavily focused on penalizing

615
00:30:09,890 --> 00:30:14,140
large deviations. But here's where I struggled

616
00:30:14,140 --> 00:30:17,779
when reading the text. Why is convergence in

617
00:30:17,779 --> 00:30:20,940
probability not enough? Why do we need L sub

618
00:30:20,940 --> 00:30:23,599
p norm? If the chance of an error is shrinking

619
00:30:23,599 --> 00:30:26,240
to zero, isn't the average error also shrinking

620
00:30:26,240 --> 00:30:28,839
to zero? This is the trap. And this is where

621
00:30:28,839 --> 00:30:30,839
the text introduces one of the most technical

622
00:30:30,839 --> 00:30:34,200
but vital concepts in the entire book. Uniform

623
00:30:34,200 --> 00:30:36,480
integrability. Let me give you a mathematical

624
00:30:36,480 --> 00:30:38,920
nightmare scenario to show why convergence in

625
00:30:38,920 --> 00:30:41,980
probability fails to guarantee the average shrinks.

626
00:30:42,259 --> 00:30:44,920
Oh, I'm ready for the nightmare scenario. Let's

627
00:30:44,920 --> 00:30:47,940
create a random sequence xn. Most of the time,

628
00:30:47,960 --> 00:30:51,660
xn is exactly 0. In fact, the probability that

629
00:30:51,660 --> 00:30:56,559
it is 0 is 1 minus 1 over n. So for x100, there

630
00:30:56,559 --> 00:31:00,720
is a 99 % chance it is 0. For x1000, there is

631
00:31:00,720 --> 00:31:04,819
a 99 .9 % chance it is 0. Okay, so it is definitely

632
00:31:04,819 --> 00:31:08,000
converging in probability to 0. The chance of

633
00:31:08,000 --> 00:31:09,960
it being anything else is vanishing rapidly.

634
00:31:10,589 --> 00:31:13,730
But when it is not 0, which happens with a tiny

635
00:31:13,730 --> 00:31:16,710
probability of 1 over n, it suddenly spikes to

636
00:31:16,710 --> 00:31:20,990
the value of n. So x1 ,000 has a 0 .1 % chance

637
00:31:20,990 --> 00:31:24,390
of being exactly 1 ,000. Let me calculate the

638
00:31:24,390 --> 00:31:28,130
expected value on that. The expectation of xn

639
00:31:28,130 --> 00:31:30,809
is the value 0 times its probability plus the

640
00:31:30,809 --> 00:31:33,250
value n times its probability, which is 1 over

641
00:31:33,250 --> 00:31:39,240
n. So 0 plus n times 1 over n. That equals 1.

642
00:31:40,039 --> 00:31:42,019
The expected value is always 1, no matter how

643
00:31:42,019 --> 00:31:45,779
far out we go to infinity. Precisely. The sequence

644
00:31:45,779 --> 00:31:48,819
converges in probability to 0, but its mathematical

645
00:31:48,819 --> 00:31:52,759
expectation remains permanently stuck at 1. It

646
00:31:52,759 --> 00:31:55,160
never converges to the expectation of the limit.

647
00:31:55,359 --> 00:31:58,700
The tail of the probability distribution is getting

648
00:31:58,700 --> 00:32:01,220
thinner and thinner, but it's getting heavier

649
00:32:01,220 --> 00:32:04,700
and heavier in magnitude. It escapes to infinity,

650
00:32:04,960 --> 00:32:07,740
completely ruining your average. It's the ultimate

651
00:32:07,740 --> 00:32:10,849
black swan mathematical glitch. A high impact,

652
00:32:11,049 --> 00:32:13,809
low probability event that entirely skews the

653
00:32:13,809 --> 00:32:16,450
underlying reality. And uniform integrability

654
00:32:16,450 --> 00:32:18,730
is the mathematical clamp that prevents this.

655
00:32:18,849 --> 00:32:21,349
It is a condition that guarantees these extreme

656
00:32:21,349 --> 00:32:24,089
outliers, these heavy tails, are suppressed enough

657
00:32:24,089 --> 00:32:26,150
that they don't corrupt the expectation as we

658
00:32:26,150 --> 00:32:28,990
take the limit to infinity. The text proves that

659
00:32:28,990 --> 00:32:31,990
if a sequence converges in probability and it

660
00:32:31,990 --> 00:32:34,710
possesses uniform integrability, then it is guaranteed

661
00:32:34,710 --> 00:32:38,289
to converge in L1 norm, meaning its expectations

662
00:32:38,289 --> 00:32:41,259
safely converge. It repairs the bridge between

663
00:32:41,259 --> 00:32:44,140
probability and averages. And the authors wield

664
00:32:44,140 --> 00:32:46,619
these convergence proofs to establish the twin

665
00:32:46,619 --> 00:32:49,519
pillars of probability theory. The law of large

666
00:32:49,519 --> 00:32:52,559
numbers, which uses almost sure convergence to

667
00:32:52,559 --> 00:32:55,119
prove that if you flip a fair coin infinitely,

668
00:32:55,200 --> 00:32:58,599
the average will stabilize at 50%. Yep. And the

669
00:32:58,599 --> 00:33:00,640
central limit theorem, which uses convergence

670
00:33:00,640 --> 00:33:03,180
in distribution to prove that the sum of countless

671
00:33:03,180 --> 00:33:06,079
independent random variables, no matter... how

672
00:33:06,079 --> 00:33:08,839
bizarre their original distributions, will inevitably

673
00:33:08,839 --> 00:33:11,319
be crushed into the shape of a perfect Gaussian

674
00:33:11,319 --> 00:33:14,420
bell curve. The universe just has a gravitational

675
00:33:14,420 --> 00:33:17,680
pull toward the bell curve. It does. But everything

676
00:33:17,680 --> 00:33:20,619
we've talked about so far is a bit, well, passive.

677
00:33:20,759 --> 00:33:23,700
We set up a space, we let a process run to infinity,

678
00:33:23,759 --> 00:33:26,539
and we just watch to see what shape it takes.

679
00:33:26,680 --> 00:33:29,279
But what happens when we actively interact with

680
00:33:29,279 --> 00:33:32,059
the process? What happens when we learn new information

681
00:33:32,059 --> 00:33:35,299
halfway through the experiment? Which... moves

682
00:33:35,299 --> 00:33:37,940
us into chapter four, conditioning and markup

683
00:33:37,940 --> 00:33:41,079
processes. We have to update our worldview. Now,

684
00:33:41,099 --> 00:33:43,660
in high school stats, conditioning is easy. What's

685
00:33:43,660 --> 00:33:45,420
the probability of drawing a king from a deck

686
00:33:45,420 --> 00:33:49,099
of cards? Four out of 52. What if I secretly

687
00:33:49,099 --> 00:33:51,160
peek and tell you the card is definitely a face

688
00:33:51,160 --> 00:33:54,579
card? You just shrink the sample space. The probability

689
00:33:54,579 --> 00:33:57,400
updates to four out of 12. But in the infinite,

690
00:33:57,519 --> 00:34:00,599
continuous spaces of advanced probability, you

691
00:34:00,599 --> 00:34:03,480
can't just count cards like that. You can't divide

692
00:34:03,480 --> 00:34:05,839
by the probability of an event if that probability

693
00:34:05,839 --> 00:34:09,019
is zero. And in a continuous space, the probability

694
00:34:09,019 --> 00:34:12,519
of hitting any exact specific point is technically

695
00:34:12,519 --> 00:34:16,260
zero. So the authors redefine conditional expectation

696
00:34:16,260 --> 00:34:19,440
not as a simple ratio, but as a geometric projection.

697
00:34:19,860 --> 00:34:22,659
Wait, a projection? Yes. Remember the Hilbert

698
00:34:22,659 --> 00:34:24,860
space of random variables we discussed earlier?

699
00:34:25,519 --> 00:34:28,780
Imagine you have a random variable y whose outcome

700
00:34:28,780 --> 00:34:30,940
you want to guess. It's a vector pointing off

701
00:34:30,940 --> 00:34:34,090
into space. but you only possess a limited filtration,

702
00:34:34,409 --> 00:34:37,829
a smaller sigma algebra of information. This

703
00:34:37,829 --> 00:34:40,409
filtration forms a subspace, like a flat plane

704
00:34:40,409 --> 00:34:43,230
beneath y. Okay, I'm picturing it. The conditional

705
00:34:43,230 --> 00:34:46,389
expectation of y, given your filtration, is the

706
00:34:46,389 --> 00:34:49,449
orthogonal projection of the vector y straight

707
00:34:49,449 --> 00:34:52,210
down onto your plane of knowledge. It is mathematically

708
00:34:52,210 --> 00:34:55,809
the closest point in your known universe to the

709
00:34:55,809 --> 00:34:58,630
unknown reality. It minimizes the mean squared

710
00:34:58,630 --> 00:35:01,889
error. That is profound. It's not just updating

711
00:35:01,889 --> 00:35:04,250
a fraction, it's finding the best possible shadow

712
00:35:04,250 --> 00:35:06,510
of the truth on the wall of our current knowledge.

713
00:35:06,710 --> 00:35:09,769
And this idea of updating knowledge leads directly

714
00:35:09,769 --> 00:35:11,889
to one of the most important concepts in physics

715
00:35:11,889 --> 00:35:14,429
and finance, which is the Markov property. Right.

716
00:35:14,590 --> 00:35:17,250
The Markov property is the mathematical definition

717
00:35:17,250 --> 00:35:20,210
of living strictly in the present moment. The

718
00:35:20,210 --> 00:35:23,210
text states that a process is Markov if, conditional

719
00:35:23,210 --> 00:35:26,150
on its past history, its future increments depend

720
00:35:26,150 --> 00:35:29,130
only on its present state. I think the best way

721
00:35:29,130 --> 00:35:32,170
to visualize this? is a game of chess. If I walk

722
00:35:32,170 --> 00:35:34,010
up to a chessboard in the middle of a tournament

723
00:35:34,010 --> 00:35:36,469
and I want to calculate the probability of white

724
00:35:36,469 --> 00:35:39,710
winning or the best possible next move, what

725
00:35:39,710 --> 00:35:42,170
information do I actually need? You just need

726
00:35:42,170 --> 00:35:44,110
to know exactly where every piece is currently

727
00:35:44,110 --> 00:35:48,309
located. Right. The present state. I do not need

728
00:35:48,309 --> 00:35:50,030
to know the sequence of moves that happened three

729
00:35:50,030 --> 00:35:52,730
hours ago to reach this board state. The fact

730
00:35:52,730 --> 00:35:55,190
that the white knight captured a black pawn on

731
00:35:55,190 --> 00:35:57,949
turn four has absolutely no bearing on the geometry

732
00:35:57,949 --> 00:36:01,280
of the board on turn 40. The present bored state

733
00:36:01,280 --> 00:36:03,960
completely shields the future from the past.

734
00:36:04,079 --> 00:36:06,800
The history is completely irrelevant once the

735
00:36:06,800 --> 00:36:09,059
present is known. That is an excellent physical

736
00:36:09,059 --> 00:36:12,380
representation. To formalize this, the authors

737
00:36:12,380 --> 00:36:15,239
introduced transition kernels, specifically for

738
00:36:15,239 --> 00:36:18,139
time -homogeneous Markov chains. A transition

739
00:36:18,139 --> 00:36:21,239
kernel is essentially the engine of the process.

740
00:36:21,320 --> 00:36:23,400
It's a function that takes your current state

741
00:36:23,789 --> 00:36:26,329
the current chessboard, and outputs a probability

742
00:36:26,329 --> 00:36:29,050
distribution for all possible states you could

743
00:36:29,050 --> 00:36:33,090
land on next. And time homogeneous means the

744
00:36:33,090 --> 00:36:35,829
engine doesn't degrade, right? The rules of chess

745
00:36:35,829 --> 00:36:39,150
don't suddenly change on turn 50. The transition

746
00:36:39,150 --> 00:36:41,789
probabilities remain locked, even as the state

747
00:36:41,789 --> 00:36:44,630
bounces around. Exactly. And once we understand

748
00:36:44,630 --> 00:36:47,369
processes that forget the past, we are ready

749
00:36:47,369 --> 00:36:49,989
to look at processes that enforce absolute fairness

750
00:36:49,989 --> 00:36:52,730
in the present, which takes us to Chapter 5.

751
00:36:53,539 --> 00:36:57,780
If Markov processes are about amnesia, martingales

752
00:36:57,780 --> 00:36:59,940
are about fairness in an unpredictable world.

753
00:37:00,380 --> 00:37:03,280
The text defines a martingale as a real valued

754
00:37:03,280 --> 00:37:06,659
stochastic process modeling a fair betting game.

755
00:37:06,780 --> 00:37:09,280
And the word fair has a very strict mathematical

756
00:37:09,280 --> 00:37:11,940
definition here. It means that the house doesn't

757
00:37:11,940 --> 00:37:14,840
have an edge and neither do you. At any given

758
00:37:14,840 --> 00:37:17,710
time, T. If you take the conditional expectation

759
00:37:17,710 --> 00:37:20,969
of the process at a future time, given the entire

760
00:37:20,969 --> 00:37:23,550
filtration of knowledge you have right now, that

761
00:37:23,550 --> 00:37:26,769
expectation must equal the exact value of the

762
00:37:26,769 --> 00:37:29,670
process right now. Your expected future profit

763
00:37:29,670 --> 00:37:33,190
is always precisely zero. If I am holding a stock

764
00:37:33,190 --> 00:37:36,570
worth $100 and its price follows a true martingale

765
00:37:36,570 --> 00:37:39,250
process, and you ask me what I expect it to be

766
00:37:39,250 --> 00:37:41,829
worth tomorrow, given all the market data in

767
00:37:41,829 --> 00:37:44,539
my filtration today, The only mathematically

768
00:37:44,539 --> 00:37:48,780
valid answer is $100. Exactly. It might jump

769
00:37:48,780 --> 00:37:52,519
to $150. It might crash to $50. But the mathematical

770
00:37:52,519 --> 00:37:56,159
expectation of the change is zero. This is a

771
00:37:56,159 --> 00:37:58,719
devastating concept for gamblers and day traders.

772
00:37:58,860 --> 00:38:02,280
Because to analyze these fair games, the text

773
00:38:02,280 --> 00:38:05,219
introduces stopping times. This is the only valid

774
00:38:05,219 --> 00:38:07,519
way you are allowed to interact with a martingale.

775
00:38:07,849 --> 00:38:10,110
the authors use a really great analogy for stopping

776
00:38:10,110 --> 00:38:13,349
times they describe it as an alarm system for

777
00:38:13,349 --> 00:38:16,369
a catastrophe let's say you are monitoring a

778
00:38:16,369 --> 00:38:19,530
chemical reaction a valid stopping time is programming

779
00:38:19,530 --> 00:38:21,829
an alarm to ring the moment the temperature hits

780
00:38:21,829 --> 00:38:25,170
exactly 100 degrees why is that valid because

781
00:38:25,170 --> 00:38:27,130
the alarm uses only the information currently

782
00:38:27,130 --> 00:38:29,590
available in your filtration at that exact second

783
00:38:29,590 --> 00:38:32,849
the sensor feels 100 degrees the alarm rings

784
00:38:32,849 --> 00:38:35,369
you are making a decision based purely on the

785
00:38:35,369 --> 00:38:37,889
present And the crucial distinction is understanding

786
00:38:37,889 --> 00:38:40,650
what is not a stopping time. Right. You cannot

787
00:38:40,650 --> 00:38:43,650
program an alarm to ring exactly one hour before

788
00:38:43,650 --> 00:38:46,670
the temperature hits 100 degrees. Right. The

789
00:38:46,670 --> 00:38:48,130
system doesn't know what the temperature will

790
00:38:48,130 --> 00:38:51,190
be in an hour. Exactly. You cannot reach outside

791
00:38:51,190 --> 00:38:53,670
your filtration into the future to trigger today's

792
00:38:53,670 --> 00:38:56,889
alarm. In the stock market, sell my shares the

793
00:38:56,889 --> 00:39:00,030
day before the market crashes is not a stopping

794
00:39:00,030 --> 00:39:02,170
time. Because you are demanding knowledge outside

795
00:39:02,170 --> 00:39:04,940
your sigma algebra. Now let's connect the fair

796
00:39:04,940 --> 00:39:07,619
game of the martingale to the strict rules of

797
00:39:07,619 --> 00:39:10,619
the stopping time. The text builds to Dube's

798
00:39:10,619 --> 00:39:12,320
optional stopping theorem, which is one of the

799
00:39:12,320 --> 00:39:15,159
most crushing proofs in all of probability. I

800
00:39:15,159 --> 00:39:17,920
want to test this. Because people try to beat

801
00:39:17,920 --> 00:39:21,010
martingales all the time. Let's use the classic

802
00:39:21,010 --> 00:39:23,289
martingale betting strategy. I go to a casino

803
00:39:23,289 --> 00:39:26,750
to play a fair coin flip game. I bet $1. If I

804
00:39:26,750 --> 00:39:29,769
lose, my next bet is $2. If I lose again, $4.

805
00:39:29,809 --> 00:39:31,809
I keep doubling my bet. The classic doubling

806
00:39:31,809 --> 00:39:35,289
strategy. The moment I win, I will cover all

807
00:39:35,289 --> 00:39:37,409
my previous losses plus walk away with exactly

808
00:39:37,409 --> 00:39:40,730
$1 in profit. I stop playing. Since I'm guaranteed

809
00:39:40,730 --> 00:39:42,869
to win eventually, haven't I just guaranteed

810
00:39:42,869 --> 00:39:45,510
a profit from a fair game? Haven't I mathematically

811
00:39:45,510 --> 00:39:48,829
beaten the Martingale? It feels like a paradox,

812
00:39:49,030 --> 00:39:51,849
doesn't it? You've created a stopping time, stop

813
00:39:51,849 --> 00:39:54,389
when I win, that seems to guarantee a positive

814
00:39:54,389 --> 00:39:57,590
expectation. But Dube's theorem explains exactly

815
00:39:57,590 --> 00:40:00,469
why this is a mathematical illusion. Do prove

816
00:40:00,469 --> 00:40:03,349
that if you play a fair martingale, the expected

817
00:40:03,349 --> 00:40:05,309
value of your wealth at the stopping time is

818
00:40:05,309 --> 00:40:07,949
exactly equal to your starting wealth, meaning

819
00:40:07,949 --> 00:40:10,869
expected profit is zero, provided certain conditions

820
00:40:10,869 --> 00:40:13,150
are met. What conditions? The conditions that

821
00:40:13,150 --> 00:40:15,570
completely shatter your doubling strategy. The

822
00:40:15,570 --> 00:40:17,730
theorem holds if your stopping time is bounded,

823
00:40:17,949 --> 00:40:19,869
meaning you must stop playing before a fixed

824
00:40:19,869 --> 00:40:22,670
time limit, or if your total wealth is bounded,

825
00:40:22,829 --> 00:40:25,710
meaning you have a finite bankroll. Ah. In your

826
00:40:25,710 --> 00:40:27,510
doubling strategy, if you hit a losing streak

827
00:40:27,510 --> 00:40:30,329
of 20 flips, your next bet requires over a million

828
00:40:30,329 --> 00:40:33,110
dollars. Unless you possess literally infinite

829
00:40:33,110 --> 00:40:35,190
capital and infinite time, you will eventually

830
00:40:35,190 --> 00:40:37,230
hit your bankroll limit and lose everything,

831
00:40:37,510 --> 00:40:40,389
which perfectly mathematically balances out the

832
00:40:40,389 --> 00:40:42,230
high probability of winning just one dollar.

833
00:40:42,949 --> 00:40:45,110
Dube's theorem guarantees that a fair game cannot

834
00:40:45,110 --> 00:40:48,230
be beaten by clever scheduling. It's like a law

835
00:40:48,230 --> 00:40:51,190
of conservation for uncertainty. You cannot generate

836
00:40:51,190 --> 00:40:54,820
certainty out of true randomness. Wow. And the

837
00:40:54,820 --> 00:40:57,079
authors use martingales as a skeleton key to

838
00:40:57,079 --> 00:41:00,079
unlock other parts of math. They show how martingale

839
00:41:00,079 --> 00:41:02,699
convergence theorems actually prove the strong

840
00:41:02,699 --> 00:41:05,800
law of large numbers. They even connect it back

841
00:41:05,800 --> 00:41:08,420
to measure theory with the radon -nicodem theorem.

842
00:41:08,820 --> 00:41:12,059
That is a deep, deep connection. That is. The

843
00:41:12,059 --> 00:41:14,340
radon -nicodem derivative is essentially the

844
00:41:14,340 --> 00:41:16,980
mathematical exchange rate between two different

845
00:41:16,980 --> 00:41:19,179
probability measures on the exact same space.

846
00:41:19,500 --> 00:41:21,500
If you want to change how you measure probability,

847
00:41:22,190 --> 00:41:24,949
Radon Nicodem tells you how to re -weight the

848
00:41:24,949 --> 00:41:28,050
outcomes. The text demonstrates that this abstract

849
00:41:28,050 --> 00:41:30,449
density function can actually be discovered by

850
00:41:30,449 --> 00:41:33,110
taking the limit of a Martingale process. It

851
00:41:33,110 --> 00:41:35,889
just weaves the entire book together. Which perfectly

852
00:41:35,889 --> 00:41:38,869
sets the stage for the climax of the text, chapter

853
00:41:38,869 --> 00:41:41,889
six and chapter seven. We're going to take everything

854
00:41:41,889 --> 00:41:44,630
we've built, Martingales, Markov properties,

855
00:41:44,809 --> 00:41:47,309
continuous time, and push it to the absolute

856
00:41:47,309 --> 00:41:50,610
limit. We enter the world of the Wiener process.

857
00:41:51,130 --> 00:41:54,309
or Brownian motion. This is the most famous stochastic

858
00:41:54,309 --> 00:41:57,150
process in existence. Historically, it was modeled

859
00:41:57,150 --> 00:41:59,829
after a physical phenomenon observed by botanist

860
00:41:59,829 --> 00:42:03,510
Robert Brown. He saw microscopic specks of pollen

861
00:42:03,510 --> 00:42:06,889
suspended in water jittering erratically, battered

862
00:42:06,889 --> 00:42:09,130
by millions of invisible water molecules from

863
00:42:09,130 --> 00:42:11,570
every direction. The mathematical model of this,

864
00:42:11,690 --> 00:42:14,570
the Wiener process, is the ultimate chimera.

865
00:42:15,059 --> 00:42:17,539
It sits at the exact intersection of all our

866
00:42:17,539 --> 00:42:21,010
major concepts. It is a Gaussian process. because

867
00:42:21,010 --> 00:42:22,869
its changes over time are perfectly normally

868
00:42:22,869 --> 00:42:26,130
distributed. It is a martingale, because its

869
00:42:26,130 --> 00:42:29,489
expected future movement is perfectly zero. And

870
00:42:29,489 --> 00:42:32,010
it is a Markov process, because its future path

871
00:42:32,010 --> 00:42:34,190
depends only on where the pollen spec is right

872
00:42:34,190 --> 00:42:37,150
now, not the chaotic zigzag it took to get there.

873
00:42:37,429 --> 00:42:39,650
And the tech spends considerable time on its

874
00:42:39,650 --> 00:42:42,030
construction, detailing Levy's construction method

875
00:42:42,030 --> 00:42:44,750
using interpolation. But the most striking aspects

876
00:42:44,750 --> 00:42:47,960
of Brownian motion are its path properties. The

877
00:42:47,960 --> 00:42:50,039
text links the physical wandering of the process

878
00:42:50,039 --> 00:42:53,039
to the Laplacian operator in differential equations.

879
00:42:53,579 --> 00:42:56,139
Let's expand on that a bit. How does a random

880
00:42:56,139 --> 00:42:59,519
walk connect to differential equations? It connects

881
00:42:59,519 --> 00:43:02,199
through the heat equation. If you drop a cluster

882
00:43:02,199 --> 00:43:04,659
of particles at the origin and let them all undergo

883
00:43:04,659 --> 00:43:07,320
independent Brownian motion, the way they spread

884
00:43:07,320 --> 00:43:09,699
out, the way the probability density diffuses

885
00:43:09,699 --> 00:43:12,280
over space and time, is governed by the exact

886
00:43:12,280 --> 00:43:14,619
same differential equation that describes how

887
00:43:14,619 --> 00:43:18,340
heat dissipates through a solid metal rod. Randomness

888
00:43:18,340 --> 00:43:20,980
at the microscopic level creates perfectly predictable

889
00:43:20,980 --> 00:43:23,360
diffusion at the macroscopic level. But when

890
00:43:23,360 --> 00:43:26,039
you zoom in on the microscopic level, when you

891
00:43:26,039 --> 00:43:28,420
look at the physical path a single Brownian particle

892
00:43:28,420 --> 00:43:32,559
traces, the math completely breaks. The text

893
00:43:32,559 --> 00:43:35,340
highlights a wild, almost terrifying contradiction.

894
00:43:35,579 --> 00:43:39,119
The path of a Wiener process is continuous everywhere.

895
00:43:39,519 --> 00:43:42,219
The particle doesn't teleport. It draws a solid,

896
00:43:42,260 --> 00:43:45,480
unbroken line through space. But the path is

897
00:43:45,480 --> 00:43:49,250
differentiable. Nowhere. Nowhere. There is not

898
00:43:49,250 --> 00:43:52,010
a single point on that infinite curve where you

899
00:43:52,010 --> 00:43:55,300
can draw a tangent line. Try to visualize a curve

900
00:43:55,300 --> 00:43:58,199
that is continuous but differentiable nowhere.

901
00:43:58,780 --> 00:44:01,619
It is a line that is literally entirely made

902
00:44:01,619 --> 00:44:04,920
of corners. It changes direction infinitely fast.

903
00:44:05,460 --> 00:44:07,840
If you zoom in on a small section of the path,

904
00:44:07,940 --> 00:44:10,980
it doesn't smooth out. It just reveals more jagged

905
00:44:10,980 --> 00:44:13,880
fractal spikes. You zoom in a million times,

906
00:44:14,059 --> 00:44:17,840
it's still infinitely jagged. How can calculus

907
00:44:17,840 --> 00:44:21,110
possibly function on a shape like that? Our entire

908
00:44:21,110 --> 00:44:24,130
system of calculus, invented by Newton and Leibniz,

909
00:44:24,309 --> 00:44:27,510
relies on curves being smooth enough, locally

910
00:44:27,510 --> 00:44:30,289
straight enough, to calculate slopes and areas.

911
00:44:30,630 --> 00:44:34,130
You have hit the precipice. This is exactly where

912
00:44:34,130 --> 00:44:36,969
standard mathematics fails. Standard Riemann

913
00:44:36,969 --> 00:44:39,309
integration requires the curve to have bounded

914
00:44:39,309 --> 00:44:41,829
variation, meaning the total vertical distance

915
00:44:41,829 --> 00:44:44,730
it travels up and down must be finite. But because

916
00:44:44,730 --> 00:44:47,610
Brownian motion is all corners, its path zigzags

917
00:44:47,610 --> 00:44:50,230
up and down infinitely over any time interval.

918
00:44:50,449 --> 00:44:53,170
Its variation is infinite. The Riemann integral

919
00:44:53,170 --> 00:44:56,829
just shatters. Which necessitates the final chapter

920
00:44:56,829 --> 00:45:01,449
of the book, Chapter 7, Ito Calculus. We literally

921
00:45:01,449 --> 00:45:05,269
had to invent a bespoke calculus to tame infinite

922
00:45:05,269 --> 00:45:09,130
jaggedness. Kiyoshi Ito recognized that to integrate

923
00:45:09,130 --> 00:45:11,909
against a process with unbounded variation, like

924
00:45:11,909 --> 00:45:14,349
the Wiener process, we couldn't use geometric

925
00:45:14,349 --> 00:45:17,389
approximations. We had to rethink the integration

926
00:45:17,389 --> 00:45:19,750
process from the ground up, respecting the flow

927
00:45:19,750 --> 00:45:22,489
of information, the filtrations we established

928
00:45:22,489 --> 00:45:25,349
back in Chapter 2. The authors don't just hand

929
00:45:25,349 --> 00:45:27,750
you Ito's formula. They force you to build it

930
00:45:27,750 --> 00:45:30,510
step by step. They start by defining the integral

931
00:45:30,510 --> 00:45:34,309
for simple processes. A simple process is easy

932
00:45:34,309 --> 00:45:36,050
to integrate. Think of a staircase function.

933
00:45:36,570 --> 00:45:38,570
It stays completely flat at a constant value,

934
00:45:38,789 --> 00:45:41,489
then jumps to a new constant value, stays flat,

935
00:45:41,630 --> 00:45:44,590
and jumps again. To integrate this against Brownian

936
00:45:44,590 --> 00:45:46,650
motion, you just multiply the constant height

937
00:45:46,650 --> 00:45:48,670
of the stair by the change in the Brownian motion

938
00:45:48,670 --> 00:45:51,090
over that specific interval and sum them up.

939
00:45:51,170 --> 00:45:54,050
It's just multiplying rectangles. But Brownian

940
00:45:54,050 --> 00:45:57,849
motion isn't a staircase. It's a continuous fractal

941
00:45:57,849 --> 00:46:00,610
nightmare. So how do we bridge the gap between

942
00:46:00,610 --> 00:46:05,110
simple steps and infinite corners? We use a canonical

943
00:46:05,110 --> 00:46:07,559
approximation. We take our continuous process,

944
00:46:07,780 --> 00:46:10,280
and we approximate it with a sequence of finer

945
00:46:10,280 --> 00:46:13,400
and finer simple staircase processes. As the

946
00:46:13,400 --> 00:46:15,539
stairs get infinitely narrow, the approximation

947
00:46:15,539 --> 00:46:19,360
converges. But here's the genius of Ito, and

948
00:46:19,360 --> 00:46:21,179
it is arguably the most important conceptual

949
00:46:21,179 --> 00:46:24,460
leap in modern probability. When you are calculating

950
00:46:24,460 --> 00:46:28,199
the area of each little rectangle, you must strictly

951
00:46:28,199 --> 00:46:31,659
evaluate your function at the left endpoint of

952
00:46:31,659 --> 00:46:35,050
the time interval. The left endpoint. Let's unpack

953
00:46:35,050 --> 00:46:38,530
why that is so incredibly vital. In standard

954
00:46:38,530 --> 00:46:41,110
calculus, when you divide a curve into rectangles

955
00:46:41,110 --> 00:46:43,590
to find the area, it doesn't matter where you

956
00:46:43,590 --> 00:46:46,269
evaluate the height of the rectangle. You can

957
00:46:46,269 --> 00:46:48,010
use the left side, the right side, or the midpoint.

958
00:46:48,639 --> 00:46:51,760
As the rectangles get infinitely thin, they all

959
00:46:51,760 --> 00:46:54,579
converge to the exact same area. The geometry

960
00:46:54,579 --> 00:46:57,320
is forgiving. Standard calculus is forgiving

961
00:46:57,320 --> 00:46:59,820
because the path is deterministic and smooth.

962
00:47:00,360 --> 00:47:03,460
But in stochastic calculus, the path is random

963
00:47:03,460 --> 00:47:06,059
and jagged. The choice of endpoint isn't just

964
00:47:06,059 --> 00:47:08,260
a geometric convenience. It is a philosophical

965
00:47:08,260 --> 00:47:12,300
adherence to causality. The left endpoint represents

966
00:47:12,300 --> 00:47:15,340
the present. The right endpoint represents the

967
00:47:15,340 --> 00:47:18,679
future. Exactly. If you evaluate your function

968
00:47:18,679 --> 00:47:21,380
at the right endpoint or even the midpoint, you

969
00:47:21,380 --> 00:47:23,219
are using information from the future of the

970
00:47:23,219 --> 00:47:25,539
Brownian path to decide how much to integrate

971
00:47:25,539 --> 00:47:28,300
today. You are reaching outside your current

972
00:47:28,300 --> 00:47:30,420
filtration. You are breaking the stopping time

973
00:47:30,420 --> 00:47:32,920
rules we just established. And if you peek into

974
00:47:32,920 --> 00:47:35,309
the future, you break the fairness of the game.

975
00:47:35,559 --> 00:47:38,019
The integral would lose its martingale property.

976
00:47:38,360 --> 00:47:40,860
By strictly forcing the integral to evaluate

977
00:47:40,860 --> 00:47:44,119
only at the left endpoint, Ito guarantees that

978
00:47:44,119 --> 00:47:46,599
the integral only uses currently available information.

979
00:47:46,940 --> 00:47:49,900
It remains predictable. It remains a martingale.

980
00:47:50,250 --> 00:47:53,949
It is a stunning triumph of logic over geometry.

981
00:47:54,230 --> 00:47:57,329
And once this left endpoint integral is rigorously

982
00:47:57,329 --> 00:48:00,530
constructed, it yields the crown jewel of stochastic

983
00:48:00,530 --> 00:48:04,519
analysis, Ito's formula. This is the stochastic

984
00:48:04,519 --> 00:48:06,739
analog to the chain rule in standard calculus.

985
00:48:07,380 --> 00:48:10,000
Let's do the Taylor expansion logic on this because

986
00:48:10,000 --> 00:48:12,739
it explains everything. In standard calculus,

987
00:48:13,019 --> 00:48:16,000
if I want to know how a function f of x changes

988
00:48:16,000 --> 00:48:19,840
when I move forward by a tiny amount dx, I use

989
00:48:19,840 --> 00:48:24,159
a Taylor series. The change, df, is roughly equal

990
00:48:24,159 --> 00:48:27,280
to the first derivative, f prime of x, multiplied

991
00:48:27,280 --> 00:48:31,659
by dx, plus one half the second derivative, f

992
00:48:31,659 --> 00:48:35,619
double prime of x multiplied by dx squared. But

993
00:48:35,619 --> 00:48:38,719
in standard calculus, as dx gets infinitely small,

994
00:48:39,159 --> 00:48:42,099
dx squared becomes microscopic. It approaches

995
00:48:42,099 --> 00:48:44,699
zero so much faster than dx that it effectively

996
00:48:44,699 --> 00:48:47,280
vanishes. So we ignore the second derivative.

997
00:48:47,480 --> 00:48:49,679
The chain rule is just based on the first derivative.

998
00:48:49,980 --> 00:48:52,199
But Ito says we can't ignore it here because

999
00:48:52,199 --> 00:48:57,579
our dx is actually dw, a tiny change in Brownian

1000
00:48:57,579 --> 00:49:00,769
motion. And Brownian motion is so erratic, so

1001
00:49:00,769 --> 00:49:03,869
infinitely jagged, that the square of its changes

1002
00:49:03,869 --> 00:49:07,460
doesn't vanish. The text proves that dw squared

1003
00:49:07,460 --> 00:49:10,719
is actually equal to dt, the change in time itself.

1004
00:49:11,340 --> 00:49:14,519
That is the magic equation. dw squared equals

1005
00:49:14,519 --> 00:49:17,739
dt. The square of the randomness solidifies into

1006
00:49:17,739 --> 00:49:20,960
deterministic time. Because of this, that second

1007
00:49:20,960 --> 00:49:23,219
order term in the Taylor expansion, the one with

1008
00:49:23,219 --> 00:49:25,920
the second derivative and dw squared, doesn't

1009
00:49:25,920 --> 00:49:29,010
vanish to zero. It survives. It survives as dt.

1010
00:49:29,530 --> 00:49:33,389
So Ito's formula adds this extra term. You have

1011
00:49:33,389 --> 00:49:35,989
the standard first derivative part, but you must

1012
00:49:35,989 --> 00:49:39,090
add one half the second partial derivative. It

1013
00:49:39,090 --> 00:49:40,989
is the mathematical signature, the permanent

1014
00:49:40,989 --> 00:49:44,909
scar of infinite jaggedness. If you don't include

1015
00:49:44,909 --> 00:49:47,130
it, your equations will drift completely detached

1016
00:49:47,130 --> 00:49:50,349
from reality. And with Ito's formula, the text

1017
00:49:50,349 --> 00:49:52,789
finally unleashes the power to solve stochastic

1018
00:49:52,789 --> 00:49:56,190
differential equations, or SDEs. We are no longer

1019
00:49:56,190 --> 00:49:59,090
just observing randomness. We are building dynamic

1020
00:49:59,090 --> 00:50:02,050
engines with it. This is what runs the modern

1021
00:50:02,050 --> 00:50:05,269
world. When a quant on Wall Street prices a complex

1022
00:50:05,269 --> 00:50:07,809
options derivative, they are using geometric

1023
00:50:07,809 --> 00:50:11,230
Brownian motion, an SDE solved by Ito calculus.

1024
00:50:11,750 --> 00:50:14,289
When physicists model particle diffusion in a

1025
00:50:14,289 --> 00:50:17,920
reactor, they use SDEs. The text walks through

1026
00:50:17,920 --> 00:50:21,239
models like the Bessel process, showing how randomness

1027
00:50:21,239 --> 00:50:23,539
can be baked directly into the differential geometry

1028
00:50:23,539 --> 00:50:26,340
of the universe. We transition from the passive

1029
00:50:26,340 --> 00:50:28,860
paradoxes of measuring dust to engineering systems

1030
00:50:28,860 --> 00:50:31,139
that harness the unpredictability of reality.

1031
00:50:31,579 --> 00:50:34,320
What an incredible intellectual arc this book

1032
00:50:34,320 --> 00:50:37,739
provides. Let's recap the journey. We started

1033
00:50:37,739 --> 00:50:40,000
with a physical tape measure that completely

1034
00:50:40,000 --> 00:50:42,380
broke when trying to measure the infinite canter

1035
00:50:42,380 --> 00:50:45,280
dust and the Vitaly set. To prevent the math

1036
00:50:45,280 --> 00:50:47,679
from crashing, we built a quarantine zone called

1037
00:50:47,679 --> 00:50:50,840
a sigma algebra. We populated that safe zone

1038
00:50:50,840 --> 00:50:53,320
with stochastic processes. tracking uncertainty

1039
00:50:53,320 --> 00:50:55,440
through the expanding knowledge of filtrations.

1040
00:50:55,699 --> 00:50:58,780
We mathematically proved how random paths converge,

1041
00:50:59,199 --> 00:51:01,619
learned how to update our expectations using

1042
00:51:01,619 --> 00:51:04,900
the amnesia of Markov processes, and explored

1043
00:51:04,900 --> 00:51:07,039
the strict fairness of martingales and stopping

1044
00:51:07,039 --> 00:51:09,960
times. And finally, we confronted the infinitely

1045
00:51:09,960 --> 00:51:12,940
jagged reality of Brownian motion, forcing us

1046
00:51:12,940 --> 00:51:16,239
to invent a bespoke calculus, Ito calculus, just

1047
00:51:16,239 --> 00:51:19,400
to calculate a rate of change. It is a towering

1048
00:51:19,400 --> 00:51:21,739
architectural achievement by Croce and Botev

1049
00:51:21,739 --> 00:51:25,119
to fit this into a concise text. But I want to

1050
00:51:25,119 --> 00:51:27,460
leave you, the listener, with a final lingering

1051
00:51:27,460 --> 00:51:30,059
thought that builds on this entire edifice of

1052
00:51:30,059 --> 00:51:32,880
probability. We have seen that through the machinery

1053
00:51:32,880 --> 00:51:36,260
of filtrations and measure theory, we can perfectly

1054
00:51:36,260 --> 00:51:39,360
mathematically model uncertainty. The logic is

1055
00:51:39,360 --> 00:51:42,300
flawless. But it raises a profound philosophical

1056
00:51:42,300 --> 00:51:45,960
question. Does true randomness actually exist

1057
00:51:45,960 --> 00:51:49,059
in the physical universe? What do you mean? Like,

1058
00:51:49,099 --> 00:51:51,659
is the universe actually rolling dice? Exactly.

1059
00:51:51,880 --> 00:51:55,099
Or is probability simply the mathematics of our

1060
00:51:55,099 --> 00:51:58,260
own ignorance? Think about the filtration. It

1061
00:51:58,260 --> 00:52:00,599
represents incomplete knowledge. If we had a

1062
00:52:00,599 --> 00:52:02,980
perfect, ultimate filtration, if we knew the

1063
00:52:02,980 --> 00:52:05,260
exact position, velocity, and state of every

1064
00:52:05,260 --> 00:52:07,280
single particle and quantum field in the universe

1065
00:52:07,280 --> 00:52:10,579
at this exact moment, would the need for probability

1066
00:52:10,579 --> 00:52:13,510
completely vanish? Would the future be perfectly

1067
00:52:13,510 --> 00:52:16,329
deterministic, making the Markov transition kernels

1068
00:52:16,329 --> 00:52:20,349
collapse into absolute certainty? Is the vast,

1069
00:52:20,429 --> 00:52:22,730
complex architecture of stochastic processes

1070
00:52:22,730 --> 00:52:25,409
we just explored actually a feature of reality?

1071
00:52:25,849 --> 00:52:28,389
Or is it just a mirror reflecting the limits

1072
00:52:28,389 --> 00:52:31,429
of our own perception? Are we modeling the universe?

1073
00:52:31,730 --> 00:52:34,349
Or are we just modeling the boundaries of human

1074
00:52:34,349 --> 00:52:37,559
knowledge? That is an incredible question to

1075
00:52:37,559 --> 00:52:40,429
leave on. Thank you for joining us on this highly

1076
00:52:40,429 --> 00:52:43,349
rigorous, mathematically demanding journey. We

1077
00:52:43,349 --> 00:52:45,289
hope this deep dive proved that true insight

1078
00:52:45,289 --> 00:52:47,429
doesn't come from ignoring the complexities of

1079
00:52:47,429 --> 00:52:49,809
the foundations, but from wrestling with them.

1080
00:52:50,110 --> 00:52:52,449
Because sometimes, when the tape measure snaps

1081
00:52:52,449 --> 00:52:55,510
and the formulas break, that is when you finally

1082
00:52:55,510 --> 00:52:57,869
start to see how the mathematical universe really

1083
00:52:57,869 --> 00:52:58,190
works.