To me, the most exciting aspect of teaching mathematics using Lean is the immediate feedback. If a student's proof is wrong, it simply won't compile.
Previously, the only feedback students could receive would be from another human such as a TA, instructor, or other knowledgeable expert. Now they can receive rapid feedback from the Lean compiler.
In the future I hope there is an option for more instructive feedback from Lean's compiler in the spirit of how the Rust compiler offers suggestions to correct code. (This would probably require the use of dedicated LLMs though.)
But. I’m also thoughtful about proving things — my own math experience was decades ago, but I spent a lot of ‘slow thought’ time mulling over whatever my assignments were, trying things on paper, soaking up second hand smoke and coffee, your basic math paradise stuff.
I wonder if using Lean here could lead to some flailing / random checking / spewing. I haven’t worked with Lean much, although I did do a few rounds with coq five or ten years ago; my memory is that I mostly futzed with it and tried things.
Upshot - a solver might be great for a lot of things. But I wonder if it will cut out some of this slow thoughtful back-and-forth that leads to internalization, conceptualization, and sometimes new ideas. Any thoughts on this?
Jim Portegies (TU/e, Netherlands) and Jelle Wemmenhobe have done a lot of research on this, using their “waterproof” (controlled natural language compiled to coa) to test this directly in class. The results are very interesting, and indeed actively messing around is still a very important part of the learning experience, but you can see at least some benefits in also having a theorem prover to check if your proofs are correct.
What I was surprised is that the students learn some patterns of proof properly, but only if you make sure that they are explicitly exposed by the proof assistant (so the more automation the less learning also in this case).
In the future I hope there is an option for more instructive feedback from Lean's compiler in the spirit of how the Rust compiler offers suggestions to correct code.
This is how Acorn works, so that when proving fails but you are "close", you get suggestions in VS Code like:
Try this:
reduce(r.num, r.denom) = reduce(a, b)
cross_equals(a, b, r.num, r.denom)
r.denom * a = r.num * b
It doesn't use LLMs, though, there's a small local model running inside the VS Code extension. One day hopefully that small local model can be superhumanly strong. For more info: https://acornprover.org/docs/tutorial/proving-a-theorem/
This looks nice. I wasn't aware of Acorn--how much adoption does it have in the mathematics community (or formal methods / robotics / other communities)? I feel like most are rallying around Lean.
I'm super excited about this. Hope it gets moved to its own repo so it's easier to find and send to other people. I was always curious about math, and Tao's Analysis was the first textbook that really showed me how it can be constructed in a rigoruous way my programming brain was hoping for. Then I got a bit into Lean, and similarly it was very satisfying but Mathlib is a bit complicated to learn math concepts from. So it's nice to see a bridge from the book to the tool.
Same here. I learnt about convergence, Cauchy seq etc. it was published by a local non profit publisher -- Hindustan Book Agency so it was super affordable too.
It's nice to see theorem proving gain some momentum in a mainstream mathematics topic such as analysis.
In PLT, we already had a flagship book (The Formal Semantics of Programming Languages by Winskel) formally verified (kind of, it's not a 1-to-1 transcription) using Isabelle (http://concrete-semantics.org) back in the mid 2010s, when tools began to be very polished.
IMHO, if someone is interested in theorem proving, that's a much simpler starting point. Theorems in analysis are already pretty hard on their own.
I wouldn't be too surprised if PL proofs were simpler to start with. Part of what I hear people say is that they also are a lot more routine. Do structural induction, apply the IH to show an invariant holds, continue. I haven't done much theorem proving, nor have I done any "mathematical" (e.g. analysis) proofs with a theorem prover, but it makes me wonder how much skill transfer there is between them if "mathematical" proofs require a much different approach.
I will also mention Software Foundations in Rocq (perhaps there is a Lean port). I worked through some of the first parts of it and found it quite pleasant.
Kevin Buzzard said something like the PL proofs are about deep structures on simple objects (mostly integers), while modern math mainly concerns complex objects. If you already have the definitions, the properties usually don't involve a lot of recursion and case analyses.
It will be incredibly interesting to assess how the mainstream "textbook" approach to the subject compares to the one taken in the Mathlib. In general, formalized math libraries make it comparatively easier to state results with a maximum degree of generality and to refactor proof developments for straightforwardness and elegance.
(Refactoring is of course made easier because the system will always keep track of what follows logically from what. You don't have that when working with pen and paper, so opportunities to rework things are very often neglected.)
It is a natural question also to ask whether it makes sense to teach the Mathlib, "maximum generality" version of real analysis, or at least something approaching that, in an academic course. Same for any other field of proof-based math, of course.
Certainly not for introductory courses. There's already too much on the table - how to prove, how to program, and the base material itself.
I believe that's the experience of faculty who've tried as well - it's fine for advanced students, but a waste of instructional time for the average pupil.
As a mathematician and also someone who has programmed for quite awhile I think any programmatic formalism will fail at inculcating the underlying understating. My bias of course is that I learned mathematical concepts via academic papers.
I just feel that the overhead code presents is massive, since it often does not adhere to any semblance of style. I say say this as someone who has had to parse through other’s mathematical papers that were deemed incomprehensible. Code is 10x worse since there are virtually no standards with regards to comprehensibility.
Is there not an idiomatic way to write proofs in Lean/Coq/Agda though? Idiomatic in the sense that once you learn the common idioms/tactics proofs become a degree more readable.
He also has is own YouTube channel with a few videos where he uses Lean [1]. I don't know much about any of it, but seeing him at work, with and without LLMs, was cool.
I think this is a very nice project and a very nice approach for a foundational topic such as analysis.
A couple immediate worries:
1. Core analysis results in Mathlib work with limits in a general, unified way using the notion of filters. There are nevertheless some ad-hoc specializations of these results to their epsilon-delta form. I presume that Tao's Analysis uses a more traditional epsilon-delta approach.
2. Mathlib moves fast and breaks things. Things get renamed and refactored all the time. Downstream repos needs continual maintenance.
I have a pretty radical opinion that math education should be focused on building Computer Algebra Systems like Mathematica and Theorem provers like Lean with a heavy focus on visualization and practical applications. Taken to the extreme this could mean never doing paper math while being able to prove everything you learned in Lean.
I feel our current system's focus on endless manual calculations that seem so useless is boring and tedious and makes people hate math.
Formalization of many of the ideas from HoTT are currently happening in the Agda community. [1] It's out of my wheelhouse, so I don't know the exact motivations, but Agda is apparently a better way to formalize those ideas than in Lean.
Also, there's a new textbook coming out later this year that's a more modern update to the original HoTT book [2] which also has an Agda formalization. [3]
HoTT is a highly technical, highly niche topic and it doesn't make sense to tackle two ambitious projects at the same time like this. HoTT isn't even remotely close to being accepted as a reasonable standard, and it's kind of a non-starter for most people.
This is a bit like asking the developers of a Javascript framework why they they didn't write a framework for Elm or Haskell.
> HoTT is a highly technical, highly niche topic and it doesn't make sense to tackle two ambitious projects at the same time like this.
It's also a bit of a controversial topic in formalized mathematics. See the following relevant comments by Kevin Buzzard (an expert in Lean and also in the sort of abstract nonsense that might arguably benefit directly from HoTT) - Links courtesy of https://ncatlab.org/nlab/show/Kevin+Buzzard
A lot less work has gone into making HoTT theorem provers ergonomic to use. Also, the documentation is a lot more sparse. The benefits of HoTT are also unclear - it seems to only save work when dealing with very esoteric constructs in category theory.
Terrence Tao has a couple of analysis textbooks and this is his companion to the first of those books in Lean. He doesn’t have a type theory textbook, so that’s why no higher-order type theory - it’s not what he’s trying to do at all.
If HoTT is already proved and sets, categories, and types are already proven; I agree that it's not necessary to prove same in an applied analysis book; though it is another opportunity to verify HoTT in actual application domains.
Readit News
Previously, the only feedback students could receive would be from another human such as a TA, instructor, or other knowledgeable expert. Now they can receive rapid feedback from the Lean compiler.
In the future I hope there is an option for more instructive feedback from Lean's compiler in the spirit of how the Rust compiler offers suggestions to correct code. (This would probably require the use of dedicated LLMs though.)
But. I’m also thoughtful about proving things — my own math experience was decades ago, but I spent a lot of ‘slow thought’ time mulling over whatever my assignments were, trying things on paper, soaking up second hand smoke and coffee, your basic math paradise stuff.
I wonder if using Lean here could lead to some flailing / random checking / spewing. I haven’t worked with Lean much, although I did do a few rounds with coq five or ten years ago; my memory is that I mostly futzed with it and tried things.
Upshot - a solver might be great for a lot of things. But I wonder if it will cut out some of this slow thoughtful back-and-forth that leads to internalization, conceptualization, and sometimes new ideas. Any thoughts on this?
What I was surprised is that the students learn some patterns of proof properly, but only if you make sure that they are explicitly exposed by the proof assistant (so the more automation the less learning also in this case).
You can find a lot of the work summarized in Jelle’s PhD thesis at https://research.tue.nl/nl/publications/waterproof-transform...
This is how Acorn works, so that when proving fails but you are "close", you get suggestions in VS Code like:
It doesn't use LLMs, though, there's a small local model running inside the VS Code extension. One day hopefully that small local model can be superhumanly strong. For more info: https://acornprover.org/docs/tutorial/proving-a-theorem/Dead Comment
In PLT, we already had a flagship book (The Formal Semantics of Programming Languages by Winskel) formally verified (kind of, it's not a 1-to-1 transcription) using Isabelle (http://concrete-semantics.org) back in the mid 2010s, when tools began to be very polished.
IMHO, if someone is interested in theorem proving, that's a much simpler starting point. Theorems in analysis are already pretty hard on their own.
I will also mention Software Foundations in Rocq (perhaps there is a Lean port). I worked through some of the first parts of it and found it quite pleasant.
(Refactoring is of course made easier because the system will always keep track of what follows logically from what. You don't have that when working with pen and paper, so opportunities to rework things are very often neglected.)
It is a natural question also to ask whether it makes sense to teach the Mathlib, "maximum generality" version of real analysis, or at least something approaching that, in an academic course. Same for any other field of proof-based math, of course.
I believe that's the experience of faculty who've tried as well - it's fine for advanced students, but a waste of instructional time for the average pupil.
I just feel that the overhead code presents is massive, since it often does not adhere to any semblance of style. I say say this as someone who has had to parse through other’s mathematical papers that were deemed incomprehensible. Code is 10x worse since there are virtually no standards with regards to comprehensibility.
[1] https://www.youtube.com/@TerenceTao27
A couple immediate worries:
1. Core analysis results in Mathlib work with limits in a general, unified way using the notion of filters. There are nevertheless some ad-hoc specializations of these results to their epsilon-delta form. I presume that Tao's Analysis uses a more traditional epsilon-delta approach.
2. Mathlib moves fast and breaks things. Things get renamed and refactored all the time. Downstream repos needs continual maintenance.
https://link.springer.com/chapter/10.1007/978-981-19-7261-4_...
I feel our current system's focus on endless manual calculations that seem so useless is boring and tedious and makes people hate math.
Why no HoTT, though?
"Should Type Theory (HoTT) Replace (ZFC) Set Theory as the Foundation of Math?" https://news.ycombinator.com/item?id=43196452
Additional Lean resources from HN this week:
"100 theorems in Lean" https://news.ycombinator.com/item?id=44075061
"Google-DeepMind/formal-conjectures: collection of formalized conjectures in lean" https://news.ycombinator.com/item?id=44119725
Also, there's a new textbook coming out later this year that's a more modern update to the original HoTT book [2] which also has an Agda formalization. [3]
[1] https://martinescardo.github.io/HoTT-UF-in-Agda-Lecture-Note...
[2] https://www.cambridge.org/core/books/introduction-to-homotop...
[3] https://github.com/HoTT-Intro/Agda
This is a bit like asking the developers of a Javascript framework why they they didn't write a framework for Elm or Haskell.
It's also a bit of a controversial topic in formalized mathematics. See the following relevant comments by Kevin Buzzard (an expert in Lean and also in the sort of abstract nonsense that might arguably benefit directly from HoTT) - Links courtesy of https://ncatlab.org/nlab/show/Kevin+Buzzard
* Is HoTT the way to do mathematics? (2020) https://www.cl.cam.ac.uk/events/owls/slides/buzzard.pdf ("Nobody knows because nobody tried")
* Grothendieck's use of equality (2024) https://arxiv.org/abs/2405.10387 - Discussed here https://news.ycombinator.com/item?id=40414404
A lot less work has gone into making HoTT theorem provers ergonomic to use. Also, the documentation is a lot more sparse. The benefits of HoTT are also unclear - it seems to only save work when dealing with very esoteric constructs in category theory.
Sort of a weird question to ask imo.
Terrence Tao has a couple of analysis textbooks and this is his companion to the first of those books in Lean. He doesn’t have a type theory textbook, so that’s why no higher-order type theory - it’s not what he’s trying to do at all.
"Is this consistent with HoTT?" a tool could ask.