« first day (3426 days earlier)      last day (1891 days later) » 

00:19
@Thorgott What does this have to do with pmf and pdf? You're talking about sampling.
I'm not talking about about sampling anywhere.
He's giving you an example of a random variable which is neither discrete nor continuous
As far as pmf and pdf go, both examples have neither.
Unless you act like a physicist and pretend that the Dirac delta distribution is an actual function
@Thorgott You're saying "first choose uniformly". Isn't this sampling?
00:25
I would call it "sampling" if you were to do this several times, I think
But we're not technically performing this action even once
We're just describing the behavior of the action were we to actually take it
"Sampling" is something you do in statistics. I'm saying "choose", because I'm trying to describe the random variable as outcome of a two-stage random experiment, which I thought might be more intuitive to you. Formally, it would suffice for me to just tell you the cdf.
It’s a random variable where the probability of being less than x is: 0 if $x\leq 0$, $x/6$ if $0<x<1/2$; $2/3$ if $x=2$; $2/3+(x-1)/3$ if $1<x<2$; 1 if $x\geq 2$
(Typing that out on mobile is a chore)
Anyways. The fact that the cdf is discontinuous at zero means it can’t have a pdf
You haven't defined probability for 1/2<x<1
And the fact that it’s not piecewise constant means it has no pmf either
Oops
I can’t seem to edit that on mobile, weird
Should have been $1<x<2$
mobile is temperamental
00:33
Also, should have been “discontinuous at $x=2$”
you have to tap the message and then quickly tap where the "edit" button pops up at the top of the screen, when it appears for a tenth of a second (at least for me)
You mean 0<x<1 ?
Blaaaargh yes
Here's a picture
back on laptop now, thank goodness
00:36
The jump discontinuity corresponds to a single point with positive probability ($1$ in this case)
That can never happen for a random variable with a pdf. (Why?)
Also you wrote x=2 instead of x=1 I think
Mobile is a pain
well, tbf, i'm typo prone regardless
i can just cover it up much better on laptop :P
I have a hat
00:40
Congrats
Awesome hat
I should note that examples of random variables with distributions that are neither continuous nor discrete can, in essence, not get much more complicated than the above. More concretely, every probability distribution can be decomposed into a discrete and a continuous part.
Even better, it can be decomposed into a discrete part, an absolutely continuous part (meaning it has a pdf), and a singular continuous part (which is potentially ugly).
The Cantor distribution being in the last category
(right?)
So, what the heck do people mean when they write p(x), p(X), p(X=x), P(x), P(X), P(X=x)? Why isn't P here a distribution?
00:45
P(X) is traditionally the probability of the event X. It should be a number, if I'm not mistaken
Yup, the Cantor distribution is singular continuous
Well, what people mean when they write something is something those people should specify at some point
Well, they specify it as a distribution. In many many cases! Maybe I am just talking to you mathematicians, who are nitpicking
there's a lot of cases where it's well-defined, yes
doesn't mean it's true for every case
and when it comes to proving things rigorously, you really do have to worry about that
(Also, I'll go on a limb and say that pdf and pmf, if they exist, are just convenient functions that allow you to compute those numbers)
@Semiclassical If a r.v. is discrete, can I say p(X) is a distribution over X?
In other words
00:48
$\mathbb{P}(X=x)$ is the probability that a random variable $X$ takes the value $x$. That notation is universal.
If a random variable X is discrete or continuous, can I use the notation p(X) to denote the probability distribution over X?
Every other thing you listed above seems ambiguous
@Thorgott Ok, I know that, but I don't care about this. I am talking about distributions
What is a distribution "over" X?
Yeah, probability theory has a lot of minutia to it
00:49
I don't understand what "over" is supposed to mean here
A distribution is usually represented as graph, where the integral of that graph is 1.
yeah. it's not a useful probability in some cases, to be clear: the probability that a uniform (0,1) variable has X=1/2 is zero. not very interesting
X is the domain of that graph
Measure theory generally is a very technical field
I don't see the problem with that
00:50
in the simple familiar cases, you can think of the distribution as a graph
No, you are talking about a pdf
doesn't mean you can always do it
@Semiclassical Ok, in the continuous case, that's the case. But I know this. Just consider pmf
(you can always think of the cdf as a graph, on the other hand.)
The "usual" way to represent a distribution would be the cdf
Cause every real-valued random variable has a cdf
(which can not be said about a pmf or pdf, as demonstrated by the earlier examples)
00:51
I know the difference between cdf, etc, etc
I don't give a damn about that
Like, a distribution might be something like $N(\mu,\sigma^2)$, but, it's distribution is not it's pdf, and it's pdf is not the probability of a particular event.
Listen, consider a discrete random variable
that has an associated pmf
So, do not consider continuous r.v.s
Ok?
sure
00:52
Can we use the notation p(X) to denote the pmf, where X is the random variable? Answer: yes, we can!
Answer: no, we cannot
typically I'd see that as $\{p_n\}$ where $\sum_n p_n=1$
And I don't see any problem with that
@Thorgott Why?
though I guess mine has in mind a pmf with integer outcomes
p(X) would almost be interchangeable with what you call the probability measure
00:54
The notation $p(X)$ is used to mean a function $p$ to an argument $X$.
wouldn't it be P(X=x) for that? It's a function of what point you choose
so $p(x)=P(X=x)$ would work
or $p_X(x)$
yes, that would be a valid way of defining the pmf
though the pmf in total is just "p" then, "p(x)" is a particular value it takes
P(X) simply means that the pmf is over some domain. P(X=x) is the probability (or mass, in this case) that X=x.
@Thorgott Yeah
that makes it seem like P is somehow a function of X
the domain? but $X$ is random variable
00:55
@Semiclassical What are talking about?
a random variable is certainly not a set
writing P(X)
I don't get it. You say you can write P(X=x), but not P(X)?
WTF
like $P(X)$ might be "The probability that $X$ takes any given value", so maybe $P(X)=1$
Why can't P(X) mean also mean $P_X$?
00:57
why can't "apple" mean "banana"?
because that's not how functions work?
@Thorgott Hahhhhh
no, I'm serious
that is exactly what you're asking
@Semiclassical How functions work?
function notation, at any rate
00:59
Ok, so why is the notation P(x) ok, but not P(X)?
because $x$ (presumably) is a real number and $X$ is a random variable
And what the heck does P(x) mean?
very different objects
probability that X=x.
So, P in P(x) is not a probability distribution?
01:00
no.
What is it??????????
A fucking probability measure
I wouldn't write p(x), personally. That's still a little ambiguous. I's say P(X=x)
So, what the heck is a probability distribution? Don't probability distributions integrate to 1???
No
pdfs integrate to 1
and pmfs????
01:02
sum to 1
pmfs sum to 1, but in measure theory you can equate the concepts of integration and summation
They should also integrate to 1, no???
however, cdfs dont integrate to 1
@Rithaniel OK, as if I didn't know this!!!!
their integrals always diverge
01:03
@Thorgott OK
So, what the heck is a probability distribution???
A measure on a measurable space such that the whole space has measure $1$.
Are you saying that a probability distribution is a probability measure???
Because I don't see any difference between your definition and what I recall to be a probability measure
Yes, these expressions are synonymous
Holy crap
So, a probability distribution has a measure of 1
But why doesn't it integrate to 1?
if you're doing lower-level probability, you wouldn't use that definition of course
01:06
The probability distribution itself doesn't have a measure
Aren't these two expressions equivalent?
It assigns the measure $1$ to the whole underlying space
@Thorgott You just said that probability measure is a synonym for probability distribution and you said that a probability distribution is a measure
Yes
and I said "a measure [...] such that the whole space has measure $1$"
I did not say "the measure has measure $1$"
cause that makes no sense
So, the probability distribution IS a measure, but does NOT POSSESS a measure
What does it mean for a measure to have a measure of 1?
01:08
Yes
That doesn't mean anything; it doesn't make sense
The Japanese and Korean term mu (Japanese: 無; Korean: 무) or Chinese wu (traditional Chinese: 無; simplified Chinese: 无), meaning "not have; without", is a key word in Buddhism, especially Zen traditions. == Etymology == Old Chinese *ma 無 is cognate with the Proto-Tibeto-Burman *ma "not". This reconstructed root is widely represented in Tibeto-Burman languages; for instance, ma means "not" in both Written Tibetan and Written Burmese. == Pronunciations == The Standard Chinese pronunciation of wú 無 "not; nothing" historically derives from (c. 7th century CE) Middle Chinese mju, (c. 3rd century CE)...
lol i think the amount of time you spent engaged in this discussion couldve been spent on reading the intro to some rigorous probability theory and i think you'll end up being more clear on how things work
Ok, so what does it mean for "a measure such that the space to have measure 1"???
loch is absolutely right
I don't have time for that
I still don't get you definition
"A measure on a measurable space such that the whole space has measure 1."
01:10
a measure is basically a function which assigns values to subsets in a consistent way. a probability measure is one which, among other things, assigns the value of 1 to the entire domain.
when you pick that as the subset to assign a value to
(don't ask me for details beyond that because I don't know them)
@Semiclassical Ok, this is consistent with my interpretation of probability distribution
I still don't get why I cannot say that a probability distribution integrates to 1
because that's you assigning a value to the measure, not the measure assigning a value to a subset
What does it mean to integrate a distribution?
Isn't a probability distribution a function? If it's a function, shouldn't integration be defined for it?
Well, that's a difficult question. Disappointingly (for you), the answer lies in measure theory
01:13
in the sense of measure theory, it's a function defined on subsets
not a function defined over the reals
@Semiclassical What do you mean? Do you mean that if I integrate something I am assigning a value? What's the problem with that?
(the more technical phrase is that it's a function on some sigma-algebra over the domain. yay.)
You don't integrate measures, you integrate functions with respects to measures.
@Semiclassical So, it can be integrated??
no. you integrate functions over the reals.
01:15
So, we cannot integrate a probability distribution
Ok!
This is a start to make things clearer
You can integrate the associated pdf or pmf, though
Yes, you can
Am I right?
Ok, so a probability distribution is a synonym for probability measure
Right?
(Well, you technically need to explain what it means to integrate a pmf, but this can be done with measure theory)
Yes
And a probability measure (or distribution) is a measure, of course, as the name suggests
Even though I am not sure what a measure is
01:17
But, in the case of a probability measure, it is required that "something" is 1
In terms of probability, the requirement is simply "the probability that anything happens at all is $1$"
A measure is a function over sets?
Yes, over a (specific type of) collection of sets
A probability measure (or distribution) is NOT a pdf or pmf
Yes
It is also not a cdf
However, the cdf, pdf or pmf (the latter only when they exist) each uniquely determine the distribution
01:22
A measure is not a function over reals. Right?
(Just to make sure)
The domain is not the reals, yes. The codomain is, however.
This is because a measure tells you the probability of a certain subset?
a probability measure does, yes
Ok, I am writing down a sequence of statements to make some sense of all of this
I have a book on measure theory and probability that I need to look at again over break
01:25
@Thorgott Why do you say "specific type of"?
Because you need to restrict yourself. Just trying to define the measure for the complete powerset is not going to work.
the technical term is sigma-algebra:
In mathematical analysis and in probability theory, a σ-algebra (also σ-field) on a set X is a collection Σ of subsets of X that includes X itself, is closed under complement, and is closed under countable unions. The definition implies that it also includes the empty subset and that it is closed under countable intersections. The pair (X, Σ) is called a measurable space or Borel space. A σ-algebra is a type of algebra of sets. An algebra of sets needs only to be closed under the union or intersection of finitely many subsets, which is a weaker condition.The main use of σ-algebras is in th...
The reasons for this are highly technical.
you could probably make a decent enough argument for why those restrictions make sense for probability, but I have no idea what motivates them in measure theory
After having read the first sentence of that Wikipedia entry, it seems that a sigma-algebra is just the powerset (if I recall correctly what a powerset is).
01:28
There is no translation-invariant non-trivial measure on $\mathcal{P}(\mathbb{R})$ @Semiclassical
@Thorgott However, you say that you cannot define a measure for the powerset
Even worse, in higher dimensions, you run into the Banach-Tarski Paradox
The Wikipedia page for that has a paragraph on the measure-theoretic implications, I think
blissful ignorance!
"If $X = \{a, b, c, d\}$, one possible $\sigma$-algebra on $X$ is $\Sigma = \{\emptyset, \{a, b\}, \{c, d\}, \{a, b, c, d\} \}$, where $\emptyset$ is the empty set."
from that wiki page
@nbro The powerset is a $\sigma$-algebra
01:29
so that's a set of subsets of X which is a sigma-algebra but is not the power set
Here's an example - throwing two coins
- Set (some people call sample space) $S = \{ HH, HT, TH, TT\}$
- A measure $\mu$ assigns a (non-negative) real number to every ("measurable" - in this case you can ignore this adjective) subset of your set. Morally, this corresponds to the 'probability' of that subset happening.

- For example (if the two coins are fair), $\mu(X) = 1$, $\mu(\{HH\}) = 1/4$, $\mu(\{HH, HT\}) = 1/2$ etc.

- A random variable is a (measurable) function $X: S \rightarrow \mathbb{R}$. An example of a random variable is "How many H's did you get". So $X(HH) = 2, X(HT) = 1, X
it's like rectangles and squares. not every rectangle is a square, but every square is a rectangle
sheesh. my brain today
Or PIDs and UFDs. Every PID is a UFD, but not every UFD is a PID
hmm, now I wonder
@Semiclassical So, a powerset is a sigma-algebra, but a sigma-algebra is not a powerset?
01:32
In probability theory, another motivation for $\sigma$-algebras is that they can be used to encode the amount of information one possesses
the power set on X is a sigma-algebra of X, but there are other sigma-algebras on X which are not the power set
Also, there are, in general, a lot of measures defined on the entire powerset (discrete measures can always be extended to the entire powerset)
But on uncountable spaces, most prominently $\mathbb{R}$, this is usually not possible
Or rather, I'm not sure if it's impossible. The set-theoretic side of this is ridiculously complicated.
So let's just say you will not be able to construct them.
there be dragons
Ok, I think I don't need all these details, but are you saying that you can define different measures for the same power-set?
more like, you can define measures where their action on some part of the power set isn't defined
for...reasons?
i get the sense that the reason one picks these definitions is because they ensure you can't generate icky things
taking the example I quoted before, where $\Sigma = \{\emptyset, \{a, b\}, \{c, d\}, \{a, b, c, d\} \}$
01:37
@loch This seems valuable information, but, even though I was familiar with this sample space and the idea of a measure that assigns a probability to subsets of the sample space, I need a little more time to understand why a random variable is a measureable function and what is its relationship with the measure
If I have a measure $\mu:\Sigma\to \mathbb{R}$, then it's meaningful to ask about $\mu(\{a,b\})$, but not about $\mu(\{a\})$ or $\mu(\{b\})$
that case is a bit artificial, though, for the reason Thorgott quoted before (discrete measures can always be extended to the entire powerset)
it's less artificial when my domain is uncountable, though?
@Semiclassical Because {a} is not part of the sample space?
well, I wasn't insisting on this being a probability measure specifically
but if you specify that, then I think that's right?
hmm, no
the sigma-algebra would be the equivalent of the event space in probability
even space = sample space?
no
if I roll a die, then the outcomes are getting 1 through 6
but there's a lot more possible events than that. one event would be getting an even roll
another would be to get 2,3,4
an event is a particular subset of outcomes
01:45
I finally managed to recover this statement about probability measures defined on all subsets of an uncountable set (with no single point having positive measure) from MO: "The existence of such a measure is equiconsistent to the existence of a measurable cardinal, one of the large cardinal notions, and if ZFC is consistent, cannot be proved in ZFC. (See the notion of real-valued measurable cardinal on the Wikipedia page.)"
like I said, there be dragons
I don't know what that means, but it certainly means we shouldn't bother
that's the technical part to do with things like $\sigma$-algebras. As an example of why the definitions there matter, consider the $\sigma$-algebra only consisting of the empty set and the entire space $S$.

i.e. your collection of subsets only has *two* subsets - $\emptyset$ and $S$. It turns out that the only measure you can define here has to be $\mu(\emptyset) = 0$, $\mu(S) = 1$.

now define $X$ as before. That's still a function $S\rightarrow \mathbb{R}$. However - it is no longer measurable, and you can kind of see why that's the case - We have $Pr(X=1) = \mu( \{HT, TH\} )$. But we n
Is there a term for a toplogy that is also a sigma algebra?
Discrete?
No wait
It isn't necessarily discrete
01:49
Yeah, like the example on the wikipedia page
I thought it might be discrete, too
Well, a subset will either be clopen or neither closed nor open.
Not sure what to make of that.
Well, actually, aren't all sigma algebras also toplogies?
Because the whole space and the empty set are included and you have countable unions and countable intersections (which includes finite intersections)
So a topology that is also a sigma algebra is just a sigma algebra
Topologies require arbitrary unions
So more than countable, right
Yup
So, actually, I think a lot of $\sigma$-algebras fail this
Say, the Borel-algebra
02:00
Well, perhaps we can say something about the cardinality of a sigma-topology, then
Well, maybe not, because arbitrary implies countable
@Rithaniel it seems this is what we're looking at: en.wikipedia.org/wiki/Alexandrov_topology
(Terminology is one of the biggest hurdles in topology. A lot of stuff has been studied, but hardly any of it has an easy or predictable name)
I found it by googling "topology that is closed under complements"
I was attempting "Sigma algebra with arbitrary unions"
Well, an Alexandrov metric space is discrete, I think. Check my logic, if you will:
every x in a metric space has arbitrarily small neighborhoods and every x in an Alexandrov space has a least neighborhood, and so this implies that every x in an Alexandrov metric space is a neighborhood of itself
02:20
Sounds right to me. Alternatively, consider $\{x\}=\cap_{n\in\mathbb{N}}B_{1/n}(x)$. So being closed under countable intersections actually suffices here.
Ah yeah, fair enough
I think the least neighborhood characterization of an Alexandrov space is the most easily understood
It's a pretty interesting characterization
02:37
here's a much duller question: How many sigma-algebras are possible for an n-element set?
 
3 hours later…
05:18
@Ultradark
I have to say guys I'm onna roll
0
Q: Equivalent statement to twin prime conjecture using elementary language.

Shine On You Crazy DiamondLet $X_k^2 = \{ n \in \Bbb{N} : n^{2} - 1 = q_1\cdots q_k$ for some primes $q_i\}$. Supposing the twin prime conjecture false is equivalent to supposing that $X_2^2$ is finite. This is because $n^2 - 1 = (n-1)(n+1) = pq \iff n$ is the unique number in between $p,p+2 = q$. If $X_2^2$ is finite ...

Second day, another proof of Twin primes. Lol, I shit them out
 
4 hours later…
09:20
@Thorgott Ah, ok
However, if I'm not wrong, it shows how one cannot omit to check the assumptions before applying the theorem, cuz they may fail. Instead, I'm looking for an example of let's say a two variable function that can be rewritten as an explicit function of one variable but that does not satisfy the assumptions of the theorem
In fact, if the theorem only provides sufficient conditions, we should say that there exists some implicit function on which, however, the theorem may fail to be applied
Is this possible?
 
2 hours later…
11:04
The twin prime conjecture question was deleted?
I don't know if trump being impeached is related to it I'm a believer in the idea that politics as we are informed is a puppet show performed according the wishes of a handful of billionaires, and people in public office actually hold little to no power over anything that happens
 
2 hours later…
13:22
@Shootforthemoon I think you might want to take a look at chapter 5.4
13:57
It's always difficult to search for things that don't fit a particular description
Like, right now, I want to know some ways to construct a ring out of two other rings without it just being the direct product of the two
Like, one I know is that if you have the integers and any other ring $R$ you can construct $\mathbb{Z}+R$ by letting $(n_1+r_1)+(n_2+r_2)=(n_1+n_2)+(r_1+r_2)$ and $(n_1+r_1)(n_2+r_2)=n_1n_2+(n_2r_1+n_1r_2+r_1r_2)$
But when I google "construct ring from two rings not direct product," I exclusively get descriptions of the direct product of two rings.
user131753
14:27
Let $M$ be a group. Let $N$ be a subgroup of $M$ such that if $a\in N$ then for all $g\in M$, whenever $g+a\in N$, $g\in N$. Does this type subgroup has a special name in literature?
That's true for every subgroup. If $a\in N$ and $g+a\in N$, then $g=(g+a)-a\in N$.
Is $\ln(x)\ln(y)=1$ a hyperbola?
I know that $xy=1$ is
Try graphing it
oh I think it's a hyperbola only in log-log space
$u=\ln(x)$ $v=\ln(y)$
user131753
15:08
@Thorgott Instead of $M$ being a group consider $M$ to be a monoid and $N$ to be its submonoid satisfying the property.
15:26
Same argument applies.
15:52
looking for a hint for proving:


$$\gcd \left( \left( n-2 \right) !-1,n \right) =
\cases{n&$n \in \mathbb{P}$\cr 1& $n \not\in \mathbb{P}$\cr}
$$
16:08
@Thorgott Thank you!
The example is the inverse of $y=x^3$, right?
Because it fails in the requirement that $F_x$ should be different from zero at the origin
user131753
16:23
@Thorgott I don't understand. If $a\in N$ then $-a$ mayn't exist always.
@Shootforthemoon the inverse of $x^3$ is not differentiable at $0$
@user170039 sorry, I was thinking about modules
@Thorgott mh, so even if the theorem cannot be applied, its consequence is not realized in the example
what about $(x-y)^2=0$ near (0,0) ?
as one user suggested
That should work
16:52
The map $(x,y)\mapsto(1/y,1/x)$ can be linearized by letting $u=\ln(x)$ and $v=\ln(y).$

Is it correct to conclude that $uv=1,$ in $\log-\log,$ space is a hyperbola, but in regular $x,y$ space it is not a hyperbola?
@Thorgott i.sstatic.net/B4NZI.png would it be wrong to apply the formula for the derivative of the implicit function in this case, before verifying the assumptions? Cuz it would simplify. I know we would obtain a 0/0, which is undetermined, and maybe I'm silly making collapse centuries of maths, but in fact I see no reason why we couldn't apply such a procedure, especially if we think that derivatives are just limits for x,y approaching the point
We would also obtain - (-3/3), which is 1 and is true for both the cases y=f(x) or x=g(y)

« first day (3426 days earlier)      last day (1891 days later) »