You are on page 1of 3

MTH 301.

Extra credit Project


Spring 2013

Instructions
Write your names and IDs. NAMES: IDs:

This project is optional and will count for a maximum of ve extra points over 100 on your min (ex1, ex2, ex3) performance. You are allowed to work in pairs. Please specify the contributions of each partner precisely You have to type your answers in one document (no hand written answers accepted) and include the MATLAB code you have written Please hand in this project no later than the date of your nal exam Please site clearly the source of any online resources you have used Show all your work (to get partial credit).

We have shown in class that the Singular Value Decomposition (SVD) of a matrix A m n (for simplicity, assume m n) is A = U V T , where U, V are square orthogonal matrices and is an m n matrix that holds the ordered singular values of A : 1 2 min(m,n) 0 1. Show that A can be written as the linear combination of mn outer products matrices n A=
j =1 T j uj vj . r j =1 T j uj vj

If you know that rank (A) = r, show that this becomes

2. One of the key applications of the singular value decomposition is the construction of low-rank approximations to a matrix. One can approximate A by taking only a partial sum
k

A Ak =
j =1

T j uj vj

where k r. Show that rank (Ak ) = k . 0 1 3. Let A = 1 0 . Find the SVD of A by hand. Show all your steps. 1 1 Verify your answer in MATLAB. What is rank (A)? Compute a rank one approximation for this matrix. 4. As an illustration of the utility of low-rank matrix approximations, consider the compression of digital images. On a computer, an image is simply a matrix denoting pixel colors. For example, a grayscale image can be represented as a matrix whose entries are integers between 0 and 255 (for 256 shades of gray), denoting the shade of each pixel. Typically, such matrices can be well-approximated by low-rank matrices. Instead of storing the m n entries of the matrix A, one needs less. Show that if we choose a rank k approximation, then we only need to store k (m + n) + k numbers that make up the various j , vj uj , values in the sum 5. Experiment with a picture stored on your computer. Use MATLABs built in SVD operation to calculate a low rank approximation. Print a 1-rank, 5-rank, 10-rank approximation and comment on the savings for using such a low rank approximation

6. Compute which low rank approximation gives the best approximation while saving the most space. Be sure to print your picture and your low rank approximation and give an explanation on why you chose this specic low rank approximation.

You might also like