I remember that when I taught correlation in probability there were some surprising examples of correlation, zero and nonzero. Of course, that was 6 years ago (yikes!), so I'd have to look at my notes to remember.
@TedShifrin I once saw some notes where positively correlated is drawn as a straight line with positive slope, negatively correlated with negative slope, and 0 correlation with 0 slope
@LeakyNun you are not wrong about it, but it does not seem to be appropriate. If the answers were bad some comments would have been nice, but I got no comments.
A random theorem I learnt 2 days ago in my statistics class: If $P$ is a symmetric projection matrix and $Q$ is a symmetric positive semidefinite matrix then $I - P - Q$ is positive semidefinite iff $PQ = QP = 0$
@TedShifrin that is a valid point. Upper bound on $F^\perp$ means a set $Z$ such that $F^\perp\subset Z$. I was just trying to give a hint there, not a complete solution.
So we're talking about three downvotes total? That's not huge numbers. That question is interesting, but you haven't convinced me to switch from the usual definition.
If $P$ is an orthoprojection, $Q$ some self-adjoint positive guy, $P + Q$ will be dominated by the identity operator only if $P$ is projection onto the kernel of $Q$, or something of this sort.
I know that sometimes a person will, having seen a question or answer of yours, look to see what other questions or answers there are. This often results in a bunch of votes.
I have a problem that I think is interesting. I want to design a function whose gradients will lead directly to the local minima from anywhere on the function in $k$ steps.
However, I feel like such a function may trivially already exist since I'm at the beginning of my research. Does anyone know of one?
The idea is that if you take the gradient from a point outside of $[-1, 1]$, you can be arbitrarily far away and still get to at least within $[-1, 1]$ in finite steps; ideally to the nearest extrema.
So to refine what I'm talking about. I want to, while in $(-\infty, -1)$ searching for the minimum, land in $[-1, 1]$ in $k$ steps. While in $(1, \infty)$ searching for the maximum, land in $[-1, 1]$ in $k$ steps.
xf'(x) is another way of solving the problem I'm trying to solve if I wasn't in control of the function. It makes gradients farther away travel faster towards the center.
Hmm... that would make having the asymptote impossible without truncating the function to some (-u, u).
This might very well be bust.
I was kind of hoping for an activation function which punished inputs outside of the neighborhood, as opposed to tapering off normally towards -1 and 1 on the respective sides.
But with the properties I've mentioned, all that leaves is something like this
Red dotted lines would be parabola whereas the function I was hoping to have was in solid black
I can easily hack an application together that just does descent by the parabolic sections instead of the underlying function, but I worry that will have more issues.
@LeakyNun @TedShifrin @supinf thanks for humoring me on this