Interactive Linear Algebra (2019)(textbooks.math.gatech.edu) |
Interactive Linear Algebra (2019)(textbooks.math.gatech.edu) |
• https://www.intmath.com/ - Interactive Mathematics Learn math while you play with it
• http://worrydream.com/LadderOfAbstraction/ - up and down the ladder of abstraction
• https://betterexplained.com/ - Intuitive guides to various things in math
• https://www.math3ma.com/blog/matrices-probability-graphs - Viewing Matrices & Probability as Graphs
• http://immersivemath.com/ila/index.html - immersive linear alg
"site:GitHub.com inurl:awesome linear algebra jupyter" lists a few awesome lists with interactive linear algebra resources: https://www.google.com/search?q=site%3Agithub.com+inurl%3Aaw...
3blue1brown's "Essence of linear algebra" playlist has some excellent tutorials with intuition-building visualizations built with manim: https://youtube.com/playlist?list=PLZHQObOWTQDPD3MizzM2xVFit...
> With microMathematics Plus, not only can you perform mathematical calculations in naturally readable form but you can also create and manage your own collection of interactive formulas!
[0] https://github.com/mkulesh/microMathematics
[1] https://f-droid.org/packages/com.mkulesh.micromath.plus
[2] https://play.google.com/store/apps/details?id=com.mkulesh.mi...
I think that textbooks, math textbooks in particular, are an example where print publishing does a disservice. (I'm counting PDFs here too.) By having to lay everything out in print form, you have to clutter up your explanations with examples and footnotes that take up physical room. Here, the examples are toggle-able. If I _want_ to explore an example, I can. But I don't need to. This kind of thing is especially helpful when reviewing content for, say, a test rather than learning it for the first time.
Also finding things in textbooks is a real pain. It's difficult to index things in a helpful way, so you just have these counting schemes in LaTeX that increment for every definition, theorem, etc. I'd love to be able to tag things then search the tags.
And that says nothing for when you want to explain something that's difficult with static images. Being able to interact with animations by zooming, panning, pausing, slowing down, speeding up, etc. is a boon. (I don't think I actually saw an example of a non-static image here, but I think my point still stands.)
All in all, I'd love to see more interactive textbooks. We've got this really expressive kind of document via the web. I think we should be taking advantage of it more.
The visualization helped me spot an unstable fixed point and understand the behaviour of the algorithm near eigenvalue clashes. The behaviour's quite sophisticated.
I think I wrote this there, but here goes again. The idea is that a positive-definite symmetric matrix can be visualized as an ellipse. This follows from the spectral theorem. Each iteration of the QR algorithm causes the ellipse to fall towards the x-axis, as if under the influence of gravity. The unstable fixed point corresponds to when the ellipse is standing up precariously, unable to fall in either direction. If you tilt it by just a bit, it will fall over (so the fixed point is unstable).
The case when the ellipse is nearly circular (corresponding to near eigenvalue clashes) causes the ellipse to fall over slowly. I think this also makes physical sense, if you think of it being under the influence of gravity. If you think of this near-circle as being a matrix, then this matrix is nearly equal to a scalar multiple of the identity matrix, so its eigenvalues are essentially known. The fact that the ellipse falls very slowly implies that the eigenvectors are unstable near eigenvalue clashes, but the eigenvalues are easy to find.
Note: The issues surrounding the unstable fixed point can be fixed using Wilkinson shifts. This makes each iteration into a discontinuous function, allowing all the fixed points to be stable. The issue surrounding instability of the eigenvectors near eigenvalue clashes cannot be fixed, as it's intrinsic to eigendecomposition (even of symmetric matrices). The latter difficulties can be dodged by slightly perturbing the matrix, but the resulting eigenvectors can be very different from the eigenvectors of the unperturbed matrix.
When I took it, it was purely proofs up and down on the blackboard, zero visualizations.
Most students just needed the brief intro. A substantial fraction (but not the majority) got the 2xx computational form. A very small number of us took the 4xx version, and since (by then) you understood the applications there was little need to focus on computations since the interesting (to the course) bit was the higher level understanding of LA.
What are some fun projects that uses LA for an individual? I'm thinking about things like generative art, if anyone knows of any artists that inspire them.
This perspective is most straightforward with a 1-dimensional domain, like an audio signal, or the temperature along a thin rod, because there's a clear way to organize all the values in a vector, e.g. all the audio samples are just in sequence. So if that's appealing to you at all, I would recommend that first. You can do filtering (like with an "equalizer") and add effects like echo/reverb as linear operations.
I think that it is particularly fun to think of some ground-truth signal that is corrupted by an echo in a room and a microphone that attenuates some frequencies, and trying to undo that corruption.
2 years ago https://news.ycombinator.com/item?id=21628449
And also, it's awesome!
I agree with your sentiment. Although, I believe there is bit of problem of generalizing it to a wider field. Surely, you can make explorables with Linear Algebra, Trigonometry, Geometry, Number Theory etc, but there still exists a lot of mathematical fields, where such explanation using visualization is pretty difficult. It takes a lot of time to come up good explanations, while even longer for interactive explorables.
> All in all, I'd love to see more interactive textbooks. We've got this really expressive kind of document via the web. I think we should be taking advantage of it more.
This is something I have spent the last 4 years of my life building. I designed an alternative way of writing interactive books that takes a different approach than using just explorables. It's a conversational way of learning subjects designed for auto-didacts.
You interact (converse?) with the book and learn gradually. You can of course embed explorables into it (which I am planning to for upcoming linear algebra / discrete mathematics course), but the general process can still be generalized to any field. And if you are a LaTeX fan, you are in luck as all your notes are compiled as a Tufte-Latex Notebook. The medium is called Primer, homage to the diamond age.
You can checkout Primer here https://primerlabs.io.
A comic-based guide to understand how Primer helps you learn effectively: https://primerlabs.io/comics/introducing-primer-comics/
I feel like teaching kids some amount of coding to be able to play with math could be very helpful.
Programming wasn't an end unto itself, but a means of exploring complex topics. Notebook-style programming is a great modern iteration of this, especially with the ability to produce interesting and complex visualizations of the underlying data and structures.
One reason I love them is that I can easily browse a shelf. Scrolling through a digital library doesn’t have the same sense of amusement.
What would be an example (in the area of learning) where it does not? In the age of pervasive computing printed text is just that - a printout. A screenshot, if you will. Not extremely useful. Even a basic 3D model of something (like a car engine) that you could simply rotate and zoom is immensely more informative than a static picture of the same object.
Text, such as a mathematical proof, too, could be made dynamic for the benefit of the student: you could hide or show references, insert notes, etc.
Static/printed content is merely an artifact of the past technology we continue clinging to for no good reason.
To say nothing about the environmental impact.
[1] https://www.amazon.com/Munsell-Soil-Book-Color-M50215B/dp/B0...
This sounds so much like a meeting of Lamport's idea of how to structure a proof (https://lamport.azurewebsites.net/pubs/proof.pdf) and the Stacks project (https://stacks.math.columbia.edu).
When I was a TA I loved giving this as an example to students. It's a great teaser for the rabbit hole that is random matrix theory.
[1] Mezzadri 2002, "How to Generate Random Matrices from the Classical Compact Groups" (http://www.ams.org/notices/200705/fea-mezzadri-web.pdf)
It is clear that you mean the ellipsoid as a set of points {x | x^T * A * x = 1} (or some other constant). There is another way in which all square matrices define an ellipsoid based on how the matrix transforms a unit sphere: {A*x | x^T * x = 1} (different matrices can map to the same ellipse here, however).
I always like to clarify which one.
(I unfortunately do not have intuition for the QR algorithm, but I am distracted by the description of an ellipse falling, as if it's rolling against the x-axis)
I actually had the latter one in mind: The one given by {A*x | x^T * x = 1}. I hadn't thought of your other suggestion.
> (I unfortunately do not have intuition for the QR algorithm, but I am distracted by the description of an ellipse falling, as if it's rolling against the x-axis)
I don't know if this helps: https://ibb.co/DRDLzgJ
Each iteration causes it to rotate around the origin. The big semi-axis of the ellipse makes a smaller angle with the x-axis on each iteration. When the semi-axes are parallel to the coordinate axes, the matrix is diagonal.
Given, obsidian's popularity, this is also definitely something on the roadmap. Ability to integrate what you learn directly into your knowledge-base (or second brain as some people call it), is an excellent idea.
However, don't have the manpower right now and much focus is to create courses right now.
Might be able do it, maybe within a year or so.
I am near certain that you did not mean this.
> When the semi-axes are parallel to the coordinate axes, the matrix is diagonal.
Let D =
the matrix
| 2 0 |
| 0 1 |
A = D * R would give you a locus of points (under {Ax | x^T x = 1}) that has the semi-axes parallel to the coordinate axes. In general, A is not diagonal.
For the condition you mention (axis-aligned semi-axes of ellipsoid iff matrix is diagonal) is true for {x | x^T * A * x = 1} .
A proof of convergence for the algorithm is only known for certain families of matrices. These include the positive-definite symmetric matrices, all symmetric matrices (once suitable improvements to the algorithm are made), the Hermitian matrices, and perhaps some others. But a proof that's valid for all matrices isn't known, even though the algorithm (when improved using Wilkinson shifts) appears to converge everywhere.
A = Q^T * D * Q
What I am still failing to understand is this sentence from the article:
> The basic QR algorithm can be visualized in the case where A is a positive-definite symmetric matrix.
It sounds like you can visualize the iterates A_{k} no matter what. Is the problem that there isn't a fixed point when the ellipsoid is axis-aligned?
The article has:
> Under certain conditions,[4] the matrices Ak converge to a triangular matrix, the Schur form of A.
So am I to understand that for A positive semi-definite, the A_{k} converges to a diagonal matrix?
> So am I to understand that for A positive semi-definite, the A_{k} converges to a diagonal matrix?
Yes. Each iteration of QR (and LR) turns a PSD matrix into another PSD matrix. A PSD matrix in Schur form is necessarily a diagonal matrix.
>> The basic QR algorithm can be visualized in the case where A is a positive-definite symmetric matrix.
> It sounds like you can visualize the iterates A_{k} no matter what. Is the problem that there isn't a fixed point when the ellipsoid is axis-aligned?
My visualization technique assumes that the matrix is PSD. In the PSD case, there is a one-to-one correspondence between ellipsoids and PSD matrices. I don't know what happens if you apply it to non-PSD matrices, but the one-to-one correspondence gets lost.