# Fundamental theorem

Fundamental theorems of Machine learning (with proofs) BS-4 MIPT course

## Motivation and syllabus

The goal of the course is to boost the quality of students bachelor and master thesis works; to make results of student scientific research well-founded. The course studies techniques and practice of theorem formulations and proofs in the field of machine learning.

Why one needs to convey an important message, a scientific result as a theorem?

1. Theorem is a most important message in the field of research.
2. Theorem present results in the language of mathematics by virtue of generality and rigor.
3. Theorems are at the heart of mathematics, they also play a central role in its aesthetics.

Theorems present the message immediately and left reasoning after. The direct narration put the reasoning first and left the results after.

• How direct narration transform to fast narration?
• How to find, state and prove theorems in our work?

This course show both narration styles. It refers to our educational study and our work experience:

1. Educational mimic progression
• Definition $$\to$$ (Axiom set) $$\to$$ Theorem $$\to$$ Proof $$\to$$ Corollaries $$\to$$ Examples $$\to$$ Impact to applications
2. Scientific discovery progression
• Application problems $$\to$$ Problem generalisations $$\to$$ Useful algebraic platform $$\to$$ Definitions $$\to$$ Axiom set

So in our practice we mimic the first part of progression, and then we learn to discover patterns and formulate theorems. The theoretical talks give us series of good examples.

## Lecture topics

1. Fundamental theorem of linear algebra S
2. Singular values decomposition and spectral theorem W
3. Gauss–Markov-(Aitken) theorem W
4. Principal component analysis W
5. Karhunen–Loève theorem W
6. Kolmogorov–Arnold representation theorem W
7. Universal approximation theorem by Cybenko W
8. Deep neural network theorem
9. No free lunch theorem by Wolpert W
10. RKHS by Aronszajn and Mercer's theorem W
11. Representer theorem by Schölkopf, Herbrich, and Smola W
12. Convolution theorem (FT, convolution, correlation with CNN examples) W
13. Fourier inversion theorem W
14. Wiener–Khinchin theorem about autocorrelation and spectral decomposition W
15. Parseval's theorem (and uniform, non-uniform convergence) W
16. Probably approximately correct learning with the theorem about compression means learnability
17. Bernstein–von Mises theorem W
18. Holland's schema theorem W
19. Variational approximation
20. Convergence of random variables and Kloek's theorem W
21. Exponential family of distributions and Nelder's theorem
22. Multi-armed bandit theorem
23. Copulas and Sklar's theorem W
24. Boosting theorem Freud, Shapire, 1996, 1995
25. Bootstrap theorem (statistical estimations): Ergodic theorem
26. SVM theorem

### Each lesson contains

1. A lecturer's talk on one of fundamental theorems ($$40' = 30' + 10'$$ discussion)
2. Two students' talks (each $$20' = 15' + 5'$$ discussion)

### Each student delivers two talks

1. On a theorem, which is formulated in a paper from the list of student thesis work's references
2. On a theorem, which is formulated and proved by the student

### It is welcome to

• Make variants of our own formulations and proofs
• Re-formulate significant messages of researchers and formulate these messages as theorems

### Plan of a talk

1. Introduction: the main message briefly
2. If necessary (it could be introduced during the talk)
1. Axiom sets
2. Definitions
3. Algebraic structures
4. Notations
3. Theorem formulation and exact proof
1. The author's variant of the proof could be ameliorated
4. Corollaries
5. Theorem significance and applications

### Typography

• set the font size $$\geqslant 14$$pt
• include plots, diagrams, freehand drawings

### Scoring

• Talks and text 0-4 points, according to comparison
• Out-of-schedule drops a half
• The exam 2 points: schemes of proof of various theorems
• time-limit test (as Physics state exam) and discussion
• theorem formulation and poof scheme are hand-written
• two random theorems from the list below, 10 min to write the text

## Theorem types

• Uniqueness, existence
• Universality
• Convergence
• Complexity
• Properties of estimations
• Bounds

## Schedule

Spring semester 2021

### Student talks

Speaker References Thesis work