This is a repository for the course 18.06: Linear Algebra at MIT in Fall 2022. See other branches of this repository for previous semesters.
Instructor: Prof. Steven G. Johnson. Course administrator: Sergei Korotkikh.
Lectures: MWF11 in 26-100. Handwritten notes are posted online, along with video recordings (linked below) and other materials (slides, further reading) in the lecture summaries below.
Exams: 11am in 26-100, on 10/7, 11/14, & 12/9. Final exam: date TBA.
Recitations:
- R01,R02 — Chirag Falor: T9 in 2-143, T10 in 2-146 (office hours Mon 3pm via Zoom, Wed 3pm in 2-449).
- R03,R05 — Melissa Sherman-Bennett: T11 in 2-147, T12 in 2-147 (office hours Wed 10am via Zoom, Thurs 10am in 2-136).
- R04 — Sergei Korotkikh: T11 in 2-146 (office hours Tues 6pm via Zoom, Thurs 6pm in 2-242).
- R06,R09 — Victor Rong: T12 in 2-146, T1 in 2-146 (office hours Mon 8pm via Zoom, Tues 2pm in 2-361).
- R07,R08 — Mitchell Harris: T12 in 2-361, T1 in 2-142 (office hours Mon 2pm via Zoom, Fri 2pm in 32-D707).
- R10,R11 — Ishan Levy: T1 in 2-136, T2 in 2-142 (office hours Thurs 10:30am via Zoom, Wed 2pm in 2-390).
- R12,R13 — Gefei Dang: T2 in 2-146, T3 in 2-142 (office hours Thurs 4pm via Zoom, Wed 11am in 2-239).
Undergraduate Assistants: Raza Abbas, Alvin Chen, Christina Mirro, and Daniel Villagran. Email them at 1806fall22_ua ατ mit.edu for 1-on-1 technical help with Julia or other questions that don't work well over Piazza etc.
Resources: Piazza discussion forum, math learning center, TSR^2 study/resource room, pset partners.
This document is a brief summary of what was covered in each 18.06 lecture, along with links and suggestions for further reading. It is not a good substitute for attending lecture, but may provide a useful study guide. (You can also look at the analogous summaries from Spring 2022.)
- course overview/syllabus
- pset 1: due Friday Sep 16 at 11am (submit your solutions on Gradescope).
- video
Slides giving the syllabus and the "big picture" of what 18.06 is about. Introduction to thinking about matrices as linear operations, not just as "bags of numbers".
Further reading: Strang, chapter 1, and section 8.1 on linear transformations. 3blue1brown has a nice video on matrix multiplication as composition of linear transformations. If you've forgotten the basics of how to multiply matrices by vectors or matrices by matrices, google for some tutorial material online (e.g. Khan academy) and do a quick brush-up.
- handwritten notes: see link above (at beginning)
- video
Gaussian elimination for Ax=b: I started with the grade-school/high-school viewpoint of writing out three equations in three unknowns, adding/subtracting multiples of equations until we were left with one equation in one unknown. Then, I wrote the same equations in matrix form, and renamed this process "Gaussian elimination": we add/subtract multiples of matrix rows to introduce zeros below the diagonal, i.e. to make the matrix upper triangular U. We then do the same row operations to the right hand side b to get a new vector c. Finally, we solve Ux=c for x by working from bottom (1 equation in 1 variable) to top, a process called "backsubstitution".
To do the same operations to both A and b, a useful trick for hand calculations is to augment the matrix with a new column representing the right-hand side, forming [A b] before starting Gaussian elimination.
What comes next? The problem with expressing Gaussian elimination this way, as operations on individual numbers in the matrix, is that it is impossible to follow the process in detail for anything except a very tiny matrix. We need a higher-level "algebraic" way to express the process, both to help us understand it and also to help us to use it (e.g. to perform additional algebraic transformations afterwards). To do this, we want to express the process, not as operations on individual numbers, but as matrix operations.
Rewrote Gaussian elimination in matrix form: we multiply a matrix A on the left by a sequence of lower-triangular "elimination matrices" Eₙ to arrive at an upper-triangular matrix U = EA. To solve Ax=b, we can think of the earlier process as multiplying both sides on the left by E, the linear operator representing the composition (product) of all of the elimination steps, yielding Ux=EAx on the left and c=Eb on the right.
We're not done: it turns out to be even more fruitful to reverse the process, and write A = LU: L represents the operations required to turn the matrix U back into A, and turns out toe be a lower-triangular matrix whose entries are just a record of the elimination steps. This LU factorization is extremely useful and important because it allows us to replace a complicated matrix A with two much simpler (triangular) ones. For example, solving Ax=b turns into LUx=b, and we can do this just by two "triangular" solves. More on this next time.
Further reading: Textbook sections 2.1, 2.2, 2.3. Strang lecture 2 video. And there is a Gaussian-elimination Julia notebook that covers the same steps in Julia form. See also "The key reason why A = LU" in section 2.6 of the textbook.
- video (only): see the spring 2022 recording, lecture 3
- handwritten notes
- Matrix inverse and LU notebook
(Prof. Johnson is sick and so we will use the recorded lecture from spring.)
Showed that Gaussian elimination can be viewed as LU factorization:
- Gaussian elimination A ⟿ U=EA (without row swaps) can be thought of as A=LU: factorizing A into a product of two simpler (triangular) matrices (L=lower, U=upper). U is the matrix that you normally get when you do elimination by hand, and L (the inverse of the elimination steps L=E⁻¹, a lower-triangular matrix with 1's on the diagonal) is essentially a "record" of the elimination steps.
L is the matrix that "reverses" Gaussian elimination: it tells you how to get A back from L. Despite this, I showed in lecture that L is actually easier to get than E: all you do is make a diagonal matrix of 1's, and then fill in the multipliers from the elimination steps (flipping subtraction to addition) below the diagonal. So, L just requires bookkeeping, and no computation.
Computing U is hard (elimination is a lot of work, even for a computer), but once you have U and L then many things that you might want to do with A become easy.
- For example, suppose you want to solve Ax=b, given A=LU. Write LUx=b=L(Ux), and let y=Ux. Then Ly=b, and we can solve for y by forward-substitution. Given y, we can then solve Ux=y by back-substitution. Both of these steps are easy because the systems are triangular.
- Moreover, solving Ly=b turns out to exactly correspond to applying the elimination steps from A ⟿ U to b. (The 1's on L's diagonal mean that there are no divisions required, either.)
- This means that we can re-use L and U to solve Ax=b for many right-hand sides. In contrast, if you "augment" A with b and then do elimination (A|b)⟶(U|y), you get the same new right-hand side y but you haven't kept a record of the elimination steps, so if you have a new right-hand side you might naively repeat the whole elimination process (hard!) rather than solving Ly=b (easy!).
- More generally, whenever you have A as a product of "simpler" matrices (e.g. triangular, diagonal, …), you can solve Ax=b by a sequence of "simpler" solves.
Introduction to the concept of a matrix inverse more generally as the matrix that reverses the action of a linear operator. Key ideas from the notebook:
- A⁻¹ is the matrix that does the "reverse" of A: A⁻¹(Ax)=x for any x. It also follows that A is the reverse of A⁻¹: A(A⁻¹x)=x for any x, i.e. (A⁻¹)⁻¹=A.
- That is, if Ax=b, then A⁻¹b=x (for any x). (Equivalently, it gives the solution to Ax=b.)
- It only exists for square, nonsingular matrices A. (i.e. an m×m matrix A must give m nonzero pivots when you do elimination.)
- Equivalently, A⁻¹ is the matrix for which A⁻¹A = A⁻¹A = I (the m×m identity matrix).
- I is an identity matrix, the matrix that gives Ix=x for any x or IA=A and AI=A for any A. There are m×m identity matrices for all m, and when we write "I" we usually infer from context how big an I we mean.
In the next lecture (which will start with the end of this notebook), we will look at calculating inverses more generally (although it turns out that this is something that you should almost never do explicitly, even on a computer!).
Further reading: Textbook sections 2.5, 2.6. Strang lecture 4 video and lecture 3 video. See also "The key reason why A = LU" in section 2.6 of the textbook.
(No live tutorial since Prof. Johnson is sick.)
A basic overview of the Julia programming environment for numerical computations that we will use in 18.06 for simple computational exploration. This (Zoom-based) tutorial will cover what Julia is and the basics of interaction, scalar/vector/matrix arithmetic, and plotting — we'll be using it as just a "fancy calculator" and no "real programming" will be required.
- Tutorial materials (and links to other resources)
If possible, try to install Julia on your laptop beforehand using the instructions at the above link. Failing that, you can run Julia in the cloud (see instructions above).
Reviewed matrix inverses and key properties thereof.
Went through how to explicitly compute A⁻¹ by solving AA⁻¹ = I. Essentially, this is just solving Ax=b multiple times, where b is each column of I. A common way to organize this for hand calculation (ugh) is the Gauss–Jordan algorithm (on a 3×3 example that can also be found in the textbook): If we do row operations on A to get I, then the same row operations on I give A⁻¹! To carry this out by hand, we augment (A|I), do ordinary Gaussian elimination to get (U|C), and then do elimination "upwards" to get (I|A⁻¹).
Matrix inverses are mainly a conceptual tool that we use to move matrices around symbolically in equations. Once you are through with your algebraic manipulations, you might end up with an expression like A⁻¹b — but when it comes time to actually compute the answer, you should read "A⁻¹b" as "solve Ax=b for x by the best available method".
Further reading: Textbook sections 2.5, 2.6. Strang video lecture 3.
- Video recording
- pset 1 solutions
- pset 2: due Mon Sep 26 (Friday is a holiday).
- LU factorization for real
Brief review of previous topics in LU factorization with some more examples in the notebook:
- How the L matrix entries are just the multipliers from Gaussian elimination. No extra work is required!
- How in practice, one rarely "augments" the matrix with the right-hand side. Instead, you compute A=LU, substitute this into Ax=b=LUx, let c=Ux, solve Lc=b, then solve Ux=c. In particular, solving Lc=b is exactly the same as performing the Gaussian-elimination steps on c. (The "augmented" method is a little easier for human bookkeeping, but has essentially no advantage for the computer.)
Some new information about LU to complete the story:
- Given A=LU, you can efficiently solve multiple right-hand sides, or equivalently the matrix equation AX=B.
- How row swaps lead to the factorization PA=LU: in practice, the computer almost always does row swaps, and always gives you a permutation matrix P (or its equivalent).
We apply PA=LU to Ax=b in much the same way as for LU; the only difference is that we have to first apply the permutation P to b.
Permutation matrices P are a great example of a linear operator that is often easier to understand (and more efficient) if you don't write it as a matrix, but instead write it as a "vector" p of the permuted indices 1…n in the new order. Then Px is just x[p] in Julia (and very similarly in Matlab and Numpy): just make a new vector by extracting the components p₁,p₂,… of x.
Further reading: Textbook sections 2.7 (on permutations; we will talk about transposes soon), and 11.1. Strang video lecture 4 and video lecture 5. For 18.06, I don't expect you to know the details of how the permutation P in PA=LU is constructed even though you don't know the permutation in advance … you only need to know how to use PA=LU if it (or something similar) is given to you … but if you are interested this "partial pivoting" algorithm is described in lecture 21 of Numerical Linear Algebra by Trefethen and Bau, or in many other textbooks on numerical linear algebra.
- handwritten notes
- Computational complexity
Complexity of matrix operations: why matrix × vector or backsubstitution scale like n² for n×n matrices, while matrix × matrix or Gaussian elimination (LU factorization) scale like n³. Matrices much bigger than a few thousand square quickly become impractical, and really large problems are only tractable because they have special structure like sparsity.
The next few weeks will be devoted to problems arising from singular and non-square matrices. A "singular" square matrix is one for which we "run out of pivots" when doing elimination (we hit zeros on the diagonal we can't remove with row swaps). Non-square matrices are either "tall" (arising in overdetermined systems with more equations than unknowns, leading to fitting problems and approximate solutions) or "wide" (arising in underdetermined systems with more unknowns than equations, leading to freedom in the choice of solution and regularization techniques to impose "priors" on this choice). A key technique that will help us understand these equations is to break vectors into simpler/smaller vectors. To do this we must first broaden our concept of a "vector".
Introduced vector spaces (informally, a set V of anything you can add x±y and multiply by scalars αx) and subspaces (a subset of V such that vector operations stay in the subspace). Examples of vector spaces include real n-component vectors (ℝⁿ, or ℂⁿ for complex numbers), real m×n matrices, functions f(x) (ℝ↦ℝ, real numbers to real numbers). Examples of subspaces includes planes or lines through the origin in ℝ³, or the origin itself. The goal of this is to break vector spaces into smaller pieces that we can still do linear algebra on (hence the need for a subspace, not just any arbitrary subset). Subspaces are especially important to help us understand the solutions (if any) of Ax=b for non-square matrices A.
Further reading: Textbook sections 2.6 ("The cost of elimination") and 11.1. Section 3.1 and 3.2 on vector spaces and subspaces.