From Matrix-Vector Multiplication to Matrix-Matrix Multiplication

Transcription

1 Week4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication here are a LO of programming assignments this week hey are meant to help clarify slicing and dicing hey show that the right abstractions in the mathematics, when reflected in how we program, allow one to implement algorithms very quickly hey help you understand special properties of matrices Practice as much as you think will benefit your understanding of the material here is no need to do them all! 41 Opening Remarks 411 Predicting the Weather View at edx he following table tells us how the weather for any day eg, today predicts the weather for the next day eg, tomorrow: omorrow oday sunny cloudy rainy sunny cloudy rainy his table is interpreted as follows: If today is rainy, then the probability that it will be cloudy tomorrow is 06, etc 117

2 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 118 Homework 4111 If today is cloudy, what is the probability that tomorrow is sunny? cloudy? rainy? View at edx Homework 4112 If today is sunny, what is the probability that the day after tomorrow is sunny? cloudy? rainy? ry this! If today is cloudy, what is the probability that a week from today it is sunny? cloudy? rainy? hink about this for at most two minutes, and then look at the answer When things get messy, it helps to introduce some notation View at edx Let χ k s Let χ k c Let χ k r denote the probability that it will be sunny k days from now on day k denote the probability that it will be cloudy k days from now denote the probability that it will be rainy k days from now he discussion so far motivate the equations χ k+1 s 04 χ k s + 03 χ k c + 01 χ k r χ k+1 c 04 χ k s + 03 χ k c + 06 χ k r χ k+1 r 02 χ k s + 04 χ k c + 03 χ k r he probabilities that denote what the weather may be on day k and the table that summarizes the probabilities are often represented as a state vector, x k, and transition matrix, P, respectively: x k χ k s χ k c χ k r and P he transition from day k to day k + 1 is then written as the matrix-vector product multiplication χ k+1 s χ k+1 c χ k+1 r or x k+1 Px k, which is simply a more compact representation way of writing the system of linear equations What this demonstrates is that matrix-vector multiplication can also be used to compactly write a set of simultaneous linear equations χ k s χ k c χ k r

3 41 Opening Remarks 119 Assume again that today is cloudy so that the probability that it is sunny, cloudy, or rainy today is 0, 1, and 0, respectively: χ 0 s 0 x 0 χ 0 c 1 χ 0 r 0 If we KNOW today is cloudy, then the probability that is is sunny today is zero, etc Ah! Our friend the unit basis vector reappears! hen the vector of probabilities for tomorrow s weather, x 1, is given by χ 1 s χ s χ 1 c χ c χ 1 r χ r Ah! Pe 1 p 1, where p 1 is the second column in matrix P You should not be surprised! he vector of probabilities for the day after tomorrow, x 2, is given by χ 2 s χ 1 s χ 2 c χ 1 c χ 2 r χ 1 r Repeating this process preferrably using Python rather than by hand, we can find the probabilities for the weather for the next seven days, under the assumption that today is cloudy: x k k View at edx Homework 4113 Follow the instructions in the above video View at edx We could build a table that tells us how to predict the weather for the day after tomorrow from the weather today:

4 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 120 Day after omorrow sunny cloudy rainy oday sunny cloudy rainy One way you can do this is to observe that χ 2 s χ 2 c χ 2 r χ 1 s χ 1 c χ 1 r χ 0 s χ 0 c χ 0 r Q χ 0 s χ 0 c χ 0 r, where Q is the transition matrix that tells us how the weather today predicts the weather the day after tomorrow Well, actually, we don t yet know that applying a matrix to a vector twice is a linear transformation We ll learn that later this week Now, just like P is simply the matrix of values from the original table that showed how the weather tomorrow is predicted from today s weather, Q is the matrix of values for the above table Homework 4114 Given omorrow oday sunny cloudy rainy sunny cloudy rainy fill in the following table, which predicts the weather the day after tomorrow given the weather today: Day after omorrow sunny cloudy rainy oday sunny cloudy rainy Now here is the hard part: Do so without using your knowledge about how to perform a matrix-matrix multiplication, since you won t learn about that until later this week May we suggest that you instead use MALAB to perform the necessary calculations

5 41 Opening Remarks Outline 41 Opening Remarks Predicting the Weather Outline What You Will Learn Preparation Partitioned Matrix-Vector Multiplication ransposing a Partitioned Matrix Matrix-Vector Multiplication, Again Matrix-Vector Multiplication with Special Matrices ranspose Matrix-Vector Multiplication riangular Matrix-Vector Multiplication Symmetric Matrix-Vector Multiplication Matrix-Matrix Multiplication Product Motivation From Composing Linear ransformations to Matrix-Matrix Multiplication Computing the Matrix-Matrix Product Special Shapes Cost Enrichment Markov Chains: heir Application Wrap Up Homework Summary 155

6 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication What You Will Learn Upon completion of this unit, you should be able to Apply matrix vector multiplication to predict the probability of future states in a Markov process Make use of partitioning to perform matrix vector multiplication ranspose a partitioned matrix Partition conformally, ensuring that the size of the matrices and vectors match so that matrix-vector multiplication works ake advantage of special structures to perform matrix-vector multiplication with triangular and symmetric matrices Express and implement various matrix-vector multiplication algorithms using the FLAME notation and FlamePy Make connections between the composition of linear transformations and matrix-matrix multiplication Compute a matrix-matrix multiplication Recognize scalars and column/row vectors as special cases of matrices Compute common vector-vector and matrix-vector operations as special cases of matrix-matrix multiplication Compute an outer product xy as a special case of matrix-matrix multiplication and recognize that he rows of the resulting matrix are scalar multiples of y he columns of the resulting matrix are scalar multiples of x

7 42 Preparation Preparation 421 Partitioned Matrix-Vector Multiplication View at edx Motivation Consider A A 00 a 01 A a 10 α 11 a , x 0 2 y 0 x χ 1 3, and y ψ 1, 4 5 where y 0, R 2 hen y Ax means that y y 0 A 00 a 01 A 02 x 0 A 00 x 0 + a 01 χ 1 + A 02 ψ 1 a 10 α 11 a 12 χ 1 a 10 x 0 + α 11 χ 1 + a 12 x 2 A 20 x 0 + a 21 χ 1 + A

8 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication Homework 4211 Consider A and x , and partition these into submatrices regions as follows: A 00 a 01 A 02 a 10 α 11 a 12 and x 0 χ 1, where A 00 R 3x3, x 0 R 3, α 11 is a scalar, and χ 1 is a scalar Show with lines how A and x are partitioned: Homework 4212 With the partitioning of matrices A and x in the above exercise, repeat the partitioned matrixvector multiplication, similar to how this unit started heory Let A R m n, x R n, and y R m Partition A A 0,0 A 0,1 A 0,N 1 A 1,0 A 1,1 A 1,N 1, x x 0 x 1, and y y 0 y 1 A M 1,0 A M 1,1 A M 1,N 1 x N 1 y M 1 where m m 0 + m m M 1, m i 0 for i 0,,M 1, n n 0 + n n N 1, n j 0 for j 0,,N 1, and

9 42 Preparation 125 A i, j R m i n j, x j R n j, and y i R m i If y Ax then A 0,0 A 0,1 A 0,N 1 A 1,0 A 1,1 A 1,N 1 x 0 x 1 A M 1,0 A M 1,1 A M 1,N 1 x N 1 A 0,0 x 0 + A 0,1 x A 0,N 1 x N 1 A 1,0 x 0 + A 1,1 x A 1,N 1 x N 1 A M 1,0 x 0 + A M 1,1 x A M 1,N 1 x N 1 In other words, N 1 y i j0 A i, j x j his is intuitively true and messy to prove carefully herefore we will not give its proof, relying on the many examples we will encounter in subsequent units instead If one partitions matrix A, vector x, and vector y into blocks, and one makes sure the dimensions match up, then blocked matrix-vector multiplication proceeds exactly as does a regular matrix-vector multiplication except that individual multiplications of scalars commute while in general individual multiplications with matrix and vector blocks submatrices and subvectors do not he labeling of the submatrices and subvectors in this unit was carefully chosen to convey information Consider A 00 a 01 A 02 A a 10 α 11 a 12 he letters that are used convey information about the shapes For example, for a 01 and a 21 the use of a lowercase Roman letter indicates they are column vectors while the s in a 10 and a 12 indicate that they are row vectors Symbols α 11 and χ 1 indicate these are scalars We will use these conventions consistently to enhance readability Notice that the partitioning of matrix A and vectors x and y has to be conformal he simplest way to understand this is that matrix-vector multiplication only works if the sizes of matrices and vectors being multiply match So, a partitioning of A, x, and y, when performing a given operation, is conformal if the suboperations with submatrices and subvectors that are encountered make sense 422 ransposing a Partitioned Matrix View at edx

10 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 126 Motivation Consider his example illustrates a general rule: When transposing a partitioned matrix matrix partitioned into submatrices, you transpose the matrix of blocks, and then you transpose each block Homework 4221 Show, step-by-step, how to transpose heory Let A R m n be partitioned as follows: A A 0,0 A 0,1 A 0,N 1 A 1,0 A 1,1 A 1,N 1 A M 1,0 A M 1,1 A M 1,N 1, where A i, j R m i n j hen A A 0,0 A 1,0 A M 1,0 A 0,1 A 1,1 A M 1,1 A 0,N 1 A 1,N 1 A M 1,N 1 ransposing a partitioned matrix means that you view each submatrix as if it is a scalar, and you then transpose the matrix as if it is a matrix of scalars But then you recognize that each of those scalars is actually a submatrix and you also transpose that submatrix

11 42 Preparation 127 Special cases We now discuss a number of special cases that you may encounter Each submatrix is a scalar If A α 0,0 α 0,1 α 0,N 1 α 1,0 α 1,1 α 1,N 1 α M 1,0 α M 1,1 α M 1,N 1 then A α 0,0 α 1,0 α M 1,0 α 0,1 α 1,1 α M 1,1 α 0,0 α 1,0 α M 1,0 α 0,1 α 1,1 α M 1,1 α 0,N 1 α 1,N 1 α M 1N 1 α 0,N 1 α 1,N 1 α M 1,N 1 his is because the transpose of a scalar is just that scalar he matrix is partitioned by rows If A ã 0 ã 1, ã m 1 where each ã i is a row of A, then A ã 0 ã 1 ã m 1 ã 0 ã 1 ã m 1 ã 0 ã 1 ã m 1 his shows that rows of A, ã i, become columns of A : ã i he matrix is partitioned by columns If A a 0 a 1 a n 1, where each a j is a column of A, then A his shows that columns of A, a j, become rows of A : a j a 0 a 1 a n 1 a 0 a 1 a n blocked partitioning If A A L A BL A R A BR,

12 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 128 then A A L A R A BL A BR 3 3 blocked partitioning If then A Anyway, you get the idea!!! A 00 a 01 A 02 A a 10 α 11 a 12, A 00 a 01 A 02 A 00 a 10 A 20 a 10 α 11 a 12 a 01 α 11 a 21 A 02 a 12 A 22 A 00 a 10 A 20 a 01 α 11 a 21 A 02 a 12 A 22

13 42 Preparation 129 Homework 4222 ranspose the following matrices: For any matrix A R m n, A A A 423 Matrix-Vector Multiplication, Again View at edx Motivation In the next few units, we will modify the matrix-vector multiplication algorithms from last week so that they can take advantage of matrices with special structure eg, triangular or symmetric matrices Now, what makes a triangular or symmetric matrix special? For one thing, it is square For another, it only requires one triangle of a matrix to be stored It was for this reason that we ended up with algorithm skeletons that looked like the one in Figure 41 when we presented algorithms for triangularizing or symmetrizing a matrix

14 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 130 Algorithm: [A] : ALGORIHM SKELEONA Partition A A L A BL wherea L is 0 0 while ma L < ma do A R A BR Repartition A L A BL A R A BR whereα 11 is 1 1 A 00 a 01 A 02 a 10 α 11 a 12 Continue with endwhile A L A BL A R A BR A 00 a 01 A 02 a 10 α 11 a 12 Figure 41: Code skeleton for algorithms when matrices are triangular or symmetric Now, consider a typical partitioning of a matrix that is encountered in such an algorithm: A 00 a 01 A 02 a 10 α 01 a 12, where each represents an entry in the matrix in this case 6 6 If, for example, the matrix is lower triangular, A 00 a 01 A 02 a 10 α 11 a , 0 0 0

15 42 Preparation 131 Algorithm: y : MVMUL N UNB VAR1BA, x, y Partition A A L A R, A BL A BR x x, y y where A L is 0 0, x, y are 0 1 while ma L < ma do Repartition A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 Algorithm: y : MVMUL N UNB VAR2BA, x, y Partition A A L A R, A BL A BR x x, y y where A L is 0 0, x, y are 0 1 while ma L < ma do Repartition A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 y 0 : χ 1 a 01 + y 0 ψ 1 : a 10 x 0 + α 11 χ 1 + a 12 + ψ 1 ψ 1 : χ 1 α 11 + ψ 1 : χ 1 a 21 + Continue with A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 endwhile Continue with A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 endwhile Figure 42: Alternative algorithms for matrix-vector multiplication then a 01 0, A 02 0, and a 12 0 Remember: the 0 is a matrix or vector of appropriate size If instead the matrix is symmetric with only the lower triangular part stored, then a 01 a 10 a10, A 02 A 20, and a 12 a 21 he above observation leads us to express the matrix-vector multiplication algorithms for computing y : Ax + y given in Figure 42 Note: For the left algorithm, what was previously the current row in matrix A, a 1, is now viewed as consisting of three parts: a 1 a 10 α 11 a 12 while the vector x is now also partitioned into three parts: x x 0 χ 1 As we saw in the first week, the partitioned dot product becomes a 1 x a 10 α 11 a 12 x 1 x 0 χ 1 x 1 a 10x 0 + α 11 χ 1 + a 12,

16 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 132 which explains why the update ψ 1 : a 1 x + ψ 1 is now ψ 1 : a 10x 0 + α 11 χ 1 + a 12 + ψ 1 Similar, for the algorithm on the right, based on the matrix-vector multiplication algorithm that uses the AXPY operations, we note that y : χ 1 a 1 + y is replaced by y 0 ψ 1 : χ 1 a 01 α 11 + y 0 ψ 1 a 21 which equals his explains the update y 0 ψ 1 : χ 1 a 01 + y 0 χ 1 α 11 + ψ 1 χ 1 a 21 + y 0 : χ 1 a 01 + y 0 ψ 1 : χ 1 α 11 + ψ 1 : χ 1 a 21 + Now, for matrix-vector multiplication y : Ax + y, it is not beneficial to break the computation up in this way ypically, a dot product is more efficient than multiple operations with the subvectors Similarly, typically one AXPY is more efficient then multiple AXPYs But the observations in this unit lay the foundation for modifying the algorithms to take advantage of special structure in the matrix, later this week Homework 4231 Implement routines [ y out ] Mvmult n unb var1b A, x, y ; and [ y out ] Mvmult n unb var2b A, x, y that compute y : Ax + y via the algorithms in Figure Matrix-Vector Multiplication with Special Matrices 431 ranspose Matrix-Vector Multiplication View at edx

17 43 Matrix-Vector Multiplication with Special Matrices 133 Algorithm: y : MVMUL UNB VAR1A, x, y Partition A A L A R, y y where A L is m 0 and y is 0 1 while my < my do Repartition A L A R A 0 a 1 A 2 ψ 1 : a 1 x + ψ 1 Continue with A L A R A 0 a 1 A 2 endwhile, y, y y 0 ψ 1 y 0 ψ 1 Algorithm: y : MVMUL UNB VAR2A, x, y Partition A A, x x A B where A is 0 n and x is 0 1 while ma < ma do Repartition A A B y : χ 1 a 1 + y Continue with A A B endwhile A 0 a 1 A 2 A 0 a 1 A 2,, x x x 0 χ 1 x 0 χ 1 Figure 43: Algorithms for computing y : A x + y Motivation Let A and x hen A x he thing to notice is that what was a column in A becomes a row in A Algorithms Let us consider how to compute y : A x + y It would be possible to explicitly transpose matrix A into a new matri using, for example, the transpose function you wrote in Week 3 and to then compute y : Bx + y his approach has at least two drawbacks: You will need space for the matri Computational scientists tend to push the limits of available memory, and hence are always hesitant to use large amounts of space that isn t absolutely necessary ransposing A into B takes time A matrix-vector multiplication requires 2mn flops ransposing a matrix requires 2mn memops mn reads from memory and mn writes to memory Memory operations are very slow relative to floating point operations So, you will spend all your time transposing the matrix Now, the motivation for this unit suggest that we can simply use columns of A for the dot products in the dot product based algorithm for y : Ax+y his suggests the algorithm in FLAME notation in Figure 43 left Alternatively, one can exploit the

18 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 134 fact that columns in A become rows of A to change the algorithm for computing y : Ax + y that is based on AXPY operations into an algorithm for computing y : A x + y, as shown in Figure 43 right Implementation Homework 4311 Implement the routines [ y out ] Mvmult t unb var1 A, x, y ; and [ y out ] Mvmult t unb var2 A, x, y that compute y : A x + y via the algorithms in Figure 43 Homework 4312 Implementations achieve better performance finish faster if one accesses data consecutively in memory Now, most scientific computing codes store matrices in column-major order which means that the first column of a matrix is stored consecutively in memory, then the second column, and so forth Now, this means that an algorithm that accesses a matrix by columns tends to be faster than an algorithm that accesses a matrix by rows hat, in turn, means that when one is presented with more than one algorithm, one should pick the algorithm that accesses the matrix by columns Our FLAME notation makes it easy to recognize algorithms that access the matrix by columns For the matrix-vector multiplication y : Ax + y, would you recommend the algorithm that uses dot products or the algorithm that uses axpy operations? For the matrix-vector multiplication y : A x+y, would you recommend the algorithm that uses dot products or the algorithm that uses axpy operations? he point of this last exercise is to make you aware of the fact that knowing more than one algorithm can give you a performance edge Useful if you pay $30 million for a supercomputer and you want to get the most out of its use 432 riangular Matrix-Vector Multiplication View at edx Motivation Let U R n n be an upper triangular matrix and x R n be a vector Consider U 00 u 01 U 02 x Ux u 10 υ 11 u 12 χ U 20 u 21 U ,

19 43 Matrix-Vector Multiplication with Special Matrices 135 where s indicate components of the result that aren t important in our discussion right now We notice that u 10 0 a vector of two zeroes and hence we need not compute with it heory If U U L U BL U R U BR U 00 u 01 U 02 u 10 υ 11 u 12 U 20 u 21 U 22, where U L and U 00 are square matrices hen U BL 0, u 10 0, U 20 0, and u 21 0, where 0 indicates a matrix or vector of the appropriate dimensions U L and U BR are upper triangular matrices We will just state this as intuitively obvious Similarly, if L L L L BL L R L BR L 00 l 01 L 02 l 10 λ 11 l 12 L 20 l 21 L 22, where L L and L 00 are square matrices, then L R 0, l 01 0, L 02 0, and l12 0, where 0 indicates a matrix or vector of the appropriate dimensions L L and L BR are lower triangular matrices Algorithms Let us start by focusing on y : Ux + y, where U is upper triangular he algorithms from the previous section can be restated as in Figure 44, replacing A by U Now, notice the parts in gray Since u 10 0 and u 21 0, those computations need not be performed! Bingo, we have two algorithms that take advantage of the zeroes below the diagonal We probably should explain the names of the routines: RMVP UN UNB VAR1: riangular matrix-vector multiply plus y, with upper triangular matrix that is not transposed, unblocked variant 1 Yes, a bit convoluted, but such is life Homework 4321 Write routines [ y out ] rmvp un unb var1 U, x, y ; and [ y out ] rmvp un unb var2 U, x, y that implement the algorithms in Figure 44 that compute y : Ux + y Homework 4322 Modify the algorithms in Figure 45 so that they compute y : Lx + y, where L is a lower triangular matrix: Just strike out the parts that evaluate to zero We suggest you do this homework in conjunction with the next one Homework 4323 Write the functions [ y out ] rmvp ln unb var1 L, x, y ; and [ y out ] rmvp ln unb var2 L, x, y that implement thenalgorithms for computing y : Lx + y from Homework 4322

20 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 136 Algorithm: y : RMVP UN UNB VAR1U, x, y Partition U U L U R, U BL U BR x x, y y where U L is 0 0, x, y are 0 1 while mu L < mu do Repartition U U 00 u 01 U 02 L U R u 10 υ 11 u 12, U BL U BR U 20 u 21 U 22 x x 0 χ 1, y y 0 ψ 1 Algorithm: y : RMVP UN UNB VAR2U, x, y Partition U U L U R, U BL U BR x x, y y where U L is 0 0, x, y are 0 1 while mu L < mu do Repartition U U 00 u 01 U 02 L U R u 10 υ 11 u 12, U BL U BR U 20 u 21 U 22 x x 0 χ 1, y y 0 ψ 1 y 0 : χ 1 u 01 + y 0 ψ 1 : u 10 x 0+ υ 11 χ 1 + u 12 + ψ 1 ψ 1 : χ 1 υ 11 + ψ 1 : χ 1 u 21 + Continue with U L U R U BL U BR x U 00 u 01 U 02 u 10 υ 11 u 12, U 20 u 21 A 22 x 0 χ 1, y y 0 ψ 1 Continue with U L U R U BL U BR x U 00 u 01 U 02 u 10 υ 11 u 12, U 20 u 21 A 22 x 0 χ 1, y y 0 ψ 1 endwhile endwhile Figure 44: Algorithms for computing y : Ux + y, where U is upper triangular Homework 4324 Modify the algorithms in Figure 46 to compute x : Ux, where U is an upper triangular matrix You may not use y You have to overwrite x without using work space Hint: hink carefully about the order in which elements of x are computed and overwritten You may want to do this exercise hand-in-hand with the implementation in the next homework Homework 4325 Write routines [ x out ] rmv un unb var1 U, x ; and [ x out ] rmv un unb var2 U, x that implement the algorithms for computing x : Ux from Homework 4324 Homework 4326 Modify the algorithms in Figure 47 to compute x : Lx, where L is a lower triangular matrix You may not use y You have to overwrite x without using work space Hint: hink carefully about the order in which elements of x are computed and overwritten his question is VERY tricky You may want to do this exercise hand-in-hand with the implementation in the next homework

21 43 Matrix-Vector Multiplication with Special Matrices 137 Algorithm: y : RMVP LN UNB VAR1L, x, y Partition L L L L R, L BL L BR x x, y y where L L is 0 0, x, y are 0 1 while ml L < ml do Repartition L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 Algorithm: y : RMVP LN UNB VAR2L, x, y Partition L L L L R, L BL L BR x x, y y where L L is 0 0, x, y are 0 1 while ml L < ml do Repartition L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 y 0 : χ 1 l 01 + y 0 ψ 1 : l 10 x 0 + λ 11 χ 1 + l 12 + ψ 1 ψ 1 : χ 1 λ 11 + ψ 1 : χ 1 l 21 + Continue with L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 endwhile Continue with L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 endwhile Figure 45: Algorithms to be used in Homework 4322 Homework 4327 Write routines [ y out ] rmv ln unb var1 L, x ; and [ y out ] rmv ln unb var2 L, x that implement the algorithms from Homework 4326 for computing x : Lx Homework 4328 Develop algorithms for computing y : U x + y and y : L x + y, where U and L are respectively upper triangular and lower triangular Do not explicitly transpose matrices U and L Write routines [ y out ] rmvp ut unb var1 U, x, y ; and [ y out ] rmvp ut unb var2 U, x, y [ y out ] rmvp lt unb var1 L, x, y ; and [ y out ] rmvp ln unb var2 L, x, y that implement these algorithms

22 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 138 Algorithm: y : RMVP UN UNB VAR1U, x, y Partition U U L U R, U BL U BR x x, y y where U L is 0 0, x, y are 0 1 while mu L < mu do Repartition U U 00 u 01 U 02 L U R u 10 υ 11 u 12, U BL U BR U 20 u 21 U 22 x x 0 χ 1, y y 0 ψ 1 Algorithm: y : RMVP UN UNB VAR2U, x, y Partition U U L U R, U BL U BR x x, y y where U L is 0 0, x, y are 0 1 while mu L < mu do Repartition U U 00 u 01 U 02 L U R u 10 υ 11 u 12, U BL U BR U 20 u 21 U 22 x x 0 χ 1, y y 0 ψ 1 y 0 : χ 1 u 01 + y 0 ψ 1 : u 10 x 0+ υ 11 χ 1 + u 12 + ψ 1 ψ 1 : χ 1 υ 11 + ψ 1 : χ 1 u 21 + Continue with U L U R U BL U BR x U 00 u 01 U 02 u 10 υ 11 u 12, U 20 u 21 A 22 x 0 χ 1, y y 0 ψ 1 Continue with U L U R U BL U BR x U 00 u 01 U 02 u 10 υ 11 u 12, U 20 u 21 A 22 x 0 χ 1, y y 0 ψ 1 endwhile endwhile Figure 46: Algorithms to be used in Homework 4324 Homework 4329 Develop algorithms for computing x : U x and x : L x, where U and L are respectively upper triangular and lower triangular Do not explicitly transpose matrices U and L Write routines [ y out ] rmv ut unb var1 U, x ; and [ y out ] rmv ut unb var2 U, x [ y out ] rmv lt unb var1 L, x ; and [ y out ] rmv ln unb var2 L, x that implement these algorithms Cost Let us analyze the algorithms for computing y : Ux + y he analysis of all the other algorithms is very similar For the dot product based algorithm, the cost is in the update ψ 1 : υ 11 χ 1 + u 12 + ψ 1 which is typically computed in two steps: ψ 1 : υ 11 χ 1 + ψ 1 ; followed by a dot product ψ 1 : u 12 + ψ 1

23 43 Matrix-Vector Multiplication with Special Matrices 139 Algorithm: y : RMVP LN UNB VAR1L, x, y Partition L L L L R, L BL L BR x x, y y where L L is 0 0, x, y are 0 1 while ml L < ml do Repartition L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 Algorithm: y : RMVP LN UNB VAR2L, x, y Partition L L L L R, L BL L BR x x, y y where L L is 0 0, x, y are 0 1 while ml L < ml do Repartition L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 y 0 : χ 1 l 01 + y 0 ψ 1 : l 10 x 0 + λ 11 χ 1 + l 12 + ψ 1 ψ 1 : χ 1 λ 11 + ψ 1 : χ 1 l 21 + Continue with L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 endwhile Continue with L L 00 l 01 L 02 L L R l10 λ 11 l 12, L BL L BR L 20 l 21 L 22 x x 0 χ 1, y y 0 ψ 1 endwhile Figure 47: Algorithms to be used in Homework 4326 Now, during the first iteration, u 12 and are of length n 1, so that that iteration requires 2n n flops for the first step During the kth iteration starting with k 0, u 12 and are of length n k 1 so that the cost of that iteration is 2n k flops hus, if A is an n n matrix, then the total cost is given by n 1 k0 n 1 [2n k] 2 k0 n k 2n + n n k1 k 2n + 1n/2 flops Recall that we proved in the second week that n i1 i nn+1 2 Homework Compute the cost, in flops, of the algorithm for computing y : Lx + y that uses AXPY s

24 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 140 Homework As hinted at before: Implementations achieve better performance finish faster if one accesses data consecutively in memory Now, most scientific computing codes store matrices in column-major order which means that the first column of a matrix is stored consecutively in memory, then the second column, and so forth Now, this means that an algorithm that accesses a matrix by columns tends to be faster than an algorithm that accesses a matrix by rows hat, in turn, means that when one is presented with more than one algorithm, one should pick the algorithm that accesses the matrix by columns Our FLAME notation makes it easy to recognize algorithms that access the matrix by columns For example, in this unit, if the algorithm accesses submatrix a 01 or a 21 then it accesses columns If it accesses submatrix a 10 or a 12, then it accesses the matrix by rows For each of these, which algorithm accesses the matrix by columns: For y : Ux + y, RSVP UN UNB VAR1 or RSVP UN UNB VAR2? Does the better algorithm use a dot or an axpy? For y : Lx + y, RSVP LN UNB VAR1 or RSVP LN UNB VAR2? Does the better algorithm use a dot or an axpy? For y : U x + y, RSVP U UNB VAR1 or RSVP U UNB VAR2? Does the better algorithm use a dot or an axpy? For y : L x + y, RSVP L UNB VAR1 or RSVP L UNB VAR2? Does the better algorithm use a dot or an axpy? 433 Symmetric Matrix-Vector Multiplication View at edx Motivation Consider A 00 a 01 A 02 a 10 α 11 a 12, Here we purposely chose the matrix on the right to be symmetric We notice that a 10 a 01, A 20 A 02, and a 12 a 21 A moment of reflection will convince you that this is a general principle, when A 00 is square Moreover, notice that A 00 and A 22 are then symmetric as well heory Consider A A L A BL A R A BR where A L and A 00 are square matrices If A is symmetric then A L, A BR, A 00, and A 22 are symmetric; a 10 a 01 and a 12 a 21 ; and A 00 a 01 A 02 a 10 α 11 a 12,

25 43 Matrix-Vector Multiplication with Special Matrices 141 Algorithm: y : SYMV U UNB VAR1A, x, y Partition A A L A R, A BL A BR x x, y y where A L is 0 0, x, y are 0 1 while ma L < ma do Repartition A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 Algorithm: y : SYMV U UNB VAR2A, x, y Partition A A L A R, A BL A BR x x, y y where A L is 0 0, x, y are 0 1 while ma L < ma do Repartition A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 ψ 1 : a 01 {}}{ a 10 x 0 + α 11 χ 1 + a 12 + ψ 1 y 0 : χ 1 a 01 + y 0 ψ 1 : χ 1 α 11 + ψ 1 : χ 1 a 21 }{{} + Continue with A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 endwhile a 12 Continue with A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 endwhile Figure 48: Algorithms for computing y : Ax + y where A is symmetric, where only the upper triangular part of A is stored A 20 A 02 We will just state this as intuitively obvious Algorithms Consider computing y : Ax+y where A is a symmetric matrix Since the upper and lower triangular part of a symmetric matrix are simply the transpose of each other, it is only necessary to store half the matrix: only the upper triangular part or only the lower triangular part In Figure 48 we repeat the algorithms for matrix-vector multiplication from an earlier unit, and annotate them for the case where A is symmetric and only stored in the upper triangle he change is simple: a 10 and a 21 are not stored and thus For the left algorithm, the update ψ 1 : a 10 x 0 +α 11 χ 1 +a 12 +ψ 1 must be changed to ψ 1 : a 01 x 0 +α 11 χ 1 +a 12 +ψ 1 For the algorithm on the right, the update : χ 1 a 21 + must be changed to : χ 1 a 12 + or, more precisely, : χ 1 a 12 + since a 12 is the label for part of a row

26 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 142 Algorithm: y : SYMV L UNB VAR1A, x, y Partition A A L A R, A BL A BR x x, y y where A L is 0 0, x, y are 0 1 while ma L < ma do Repartition A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 Algorithm: y : SYMV L UNB VAR2A, x, y Partition A A L A R, A BL A BR x x, y y where A L is 0 0, x, y are 0 1 while ma L < ma do Repartition A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 y 0 : χ 1 a 01 + y 0 ψ 1 : a 10 x 0 + α 11 χ 1 + a 12 + ψ 1 ψ 1 : χ 1 α 11 + ψ 1 : χ 1 a 21 + Continue with A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 endwhile Continue with A A 00 a 01 A 02 L A R a 10 α 11 a 12, A BL A BR x x 0 χ 1, y y 0 ψ 1 endwhile Figure 49: Algorithms for Homework 4332 Homework 4331 Write routines [ y out ] Symv u unb var1 A, x, y ; and [ y out ] Symv u unb var2 A, x, y that implement the algorithms in Figure 48 Homework 4332 Modify the algorithms in Figure 49 to compute y : Ax + y, where A is symmetric and stored in the lower triangular part of matrix You may want to do this in conjunction with the next exercise Homework 4333 Write routines [ y out ] Symv l unb var1 A, x, y ; and [ y out ] Symv l unb var2 A, x, y that implement the algorithms from the previous homework

27 44 Matrix-Matrix Multiplication Product 143 Homework 4334 Challenge question! As hinted at before: Implementations achieve better performance finish faster if one accesses data consecutively in memory Now, most scientific computing codes store matrices in column-major order which means that the first column of a matrix is stored consecutively in memory, then the second column, and so forth Now, this means that an algorithm that accesses a matrix by columns tends to be faster than an algorithm that accesses a matrix by rows hat, in turn, means that when one is presented with more than one algorithm, one should pick the algorithm that accesses the matrix by columns Our FLAME notation makes it easy to recognize algorithms that access the matrix by columns he problem with the algorithms in this unit is that all of them access both part of a row AND part of a column So, your challenge is to devise an algorithm for computing y : Ax + y where A is symmetric and only stored in one half of the matrix that only accesses parts of columns We will call these variant 3 hen, write routines [ y out ] Symv u unb var3 A, x, y ; and [ y out ] Symv l unb var3 A, x, y Hint: Let s focus on the case where only the lower triangular part of A is stored If A is symmetric, then A L + L where L is the lower triangular part of A and L is the strictly lower triangular part of A Identify an algorithm for y : Lx + y that accesses matrix A by columns Identify an algorithm for y : L x + y that accesses matrix A by columns You now have two loops that together compute y : Ax + y L + L x + y Lx + L x + y Can you merge the loops into one loop? 44 Matrix-Matrix Multiplication Product 441 Motivation View at edx he first unit of the week, in which we discussed a simple model for prediction the weather, finished with the following exercise:

28 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 144 Given omorrow oday sunny cloudy rainy sunny cloudy rainy fill in the following table, which predicts the weather the day after tomorrow given the weather today: Day after omorrow sunny cloudy rainy oday sunny cloudy rainy Now here is the hard part: Do so without using your knowledge about how to perform a matrix-matrix multiplication, since you won t learn about that until later this week he entries in the table turn out to be the entries in the transition matrix Q that was described just above the exercise: χ 2 s χ 1 s χ 2 c χ 1 c χ 2 r χ 1 r χ 0 s χ 0 s χ 0 c Q χ 0 c, χ 0 r χ 0 r Now, those of you who remembered from, for example, some other course that χ s χ 0 c χ r χ 0 s χ 0 c χ 0 r would recognize that Q And, if you then remembered how to perform a matrix-matrix multiplication or you did P * P in Python, you would have deduced that Q

29 44 Matrix-Matrix Multiplication Product 145 hese then become the entries in the table If you knew all the above, well, GOOD FOR YOU! However, there are all kinds of issues that one really should discuss How do you know such a matrix exists? Why is matrix-matrix multiplication defined this way? We answer that in the next few units 442 From Composing Linear ransformations to Matrix-Matrix Multiplication View at edx Homework 4421 Let L A : R k R m and L B : R n R k both be linear transformations and, for all x R n, define the function L C : R n R m by L C x L A L B x L C x is a linear transformations Always/Sometimes/Never Now, let linear transformations L A, L B, and L C be represented by matrices A R m k, B R k n, and C R m n, respectively You know such matrices exist since L A, L B, and L C are linear transformations hen Cx L C x L A L B x ABx he matrix-matrix multiplication product is defined as the matrix C such that, for all vectors x, Cx ABx he notation used to denote that matrix is C A B or, equivalently, C AB he operation AB is called a matrix-matrix multiplication or product If A is m A n A matrix, B is m B n B matrix, and C is m C n C matrix, then for C AB to hold it must be the case that m C m A, n C n B, and n A m B Usually, the integers m and n are used for the sizes of C: C R m n and k is used for the other size : A R m k and B R k n : n k n k B m C m A Homework 4422 Let A R m n A A is well-defined By well-defined we mean that A A makes sense In this particular case this means that the dimensions of A and A are such that A A can be computed Always/Sometimes/Never Homework 4423 Let A R m n AA is well-defined Always/Sometimes/Never 443 Computing the Matrix-Matrix Product View at edx he question now becomes how to compute C given matrices A and B For this, we are going to use and abuse the unit basis vectors e j Consider the following Let

30 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 146 C R m n, A R m k, and B R k n ; and C AB; and L C : R n R m equal the linear transformation such that L C x Cx; and L A : R k R m equal the linear transformation such that L A x Ax L B : R n R k equal the linear transformation such that L B x Bx; and e j denote the jth unit basis vector; and c j denote the jth column of C; and b j denote the jth column of B hen c j Ce j L C e j L A L B e j L A Be j L A b j Ab j From this we learn that If C AB then the jth column of C, c j, equals Ab j, where b j is the jth column of B Since by now you should be very comfortable with partitioning matrices by columns, we can summarize this as c 0 c 1 c n 1 C AB A b 0 b 1 b n 1 Ab 0 Ab 1 Ab n 1 Now, let s expose the elements of C, A, and B C γ 0,0 γ 0,1 γ 0,n 1 γ 1,0 γ 1,1 γ 1,n 1, A α 0,0 α 0,1 α 0,k 1 α 1,0 α 1,1 α 1,k 1, γ m 1,0 γ m 1,1 γ m 1,n 1 α m 1,0 α m 1,1 α m 1,k 1 β 0,0 β 0,1 β 0,n 1 β 1,0 β 1,1 β 1,n 1 and B β k 1,0 β k 1,1 β k 1,n 1 We are going to show that γ i, j which you may have learned in a high school algebra course k 1 α i,p β p, j, p0

31 44 Matrix-Matrix Multiplication Product 147 We reasoned that c j Ab j : γ 0, j γ 1, j γ i, j α 0,0 α 0,1 α 0,k 1 α 1,0 α 1,1 α 1,k 1 α i,0 α i,1 α i,k 1 β 0, j β 1, j β k 1, j γ m 1, j α m 1,0 α m 1,1 α m 1,k 1 Here we highlight the ith element of c j, γ i, j, and the ith row of A We recall that the ith element of Ax equals the dot product of the ith row of A with the vector x hus, γ i, j equals the dot product of the ith row of A with the vector b j : γ i, j k 1 α i,p β p, j p0 Let A R m k, B R k n, and C R m n hen the matrix-matrix multiplication product C AB is computed by γ i, j k 1 α i,p β p, j α i,0 β 0, j + α i,1 β 1, j + + α i,k 1 β k 1, j p0 As a result of this definition Cx ABx ABx and can drop the parentheses, unless they are useful for clarity: Cx ABx and C AB Homework 4431 Compute Q P P We emphasize that for matrix-matrix multiplication to be a legal operations, the row and column dimensions of the matrices must obey certain constraints Whenever we talk about dimensions being conformal, we mean that the dimensions are such that the encountered matrix multiplications are valid operations Homework 4432 Let A and B Compute AB BA Homework 4433 Let A R m k and B R k n and AB BA A and B are square matrices Always/Sometimes/Never

32 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 148 Homework 4434 Let A R m k and B R k n AB BA Always/Sometimes/Never Homework 4435 Let A,B R n n AB BA Always/Sometimes/Never Homework 4436 A 2 is defined as AA Similarly A k A k A k 1 A for k > 0 A k is well-defined only if A is a square matrix AA A }{{} k occurrences of A Consistent with this, A 0 I so that rue/false Homework 4437 Let A, B,C be matrix of appropriate size so that ABC is well defined ABC is well defined Always/Sometimes/Never 444 Special Shapes View at edx We now show that if one treats scalars, column vectors, and row vectors as special cases of matrices, then many all? operations we encountered previously become simply special cases of matrix-matrix multiplication In the below discussion, consider C AB where C R m n, A R m k, and B R k n m n k 1 scalar multiplication 1 C 1 A 1 B In this case, all three matrices are actually scalars: γ 0,0 α 0,0 β 0,0 α 0,0 β 0,0 so that matrix-matrix multiplication becomes scalar multiplication Homework 4441 Let A 4 and B 3 hen AB n 1,k 1 SCAL 1 B m C m A

33 44 Matrix-Matrix Multiplication Product 149 Now the matrices look like γ 0,0 γ 1,0 α 0,0 α 1,0 β 0,0 α 0,0 β 0,0 α 1,0 β 0,0 β 0,0 α 0,0 β 0,0 α 1,0 β 0,0 α 0,0 α 1,0 γ m 1,0 α m 1,0 α m 1,0 β 0,0 β 0,0 α m 1,0 α m 1,0 In other words, C and A are vectors, B is a scalar, and the matrix-matrix multiplication becomes scaling of a vector 1 Homework 4442 Let A 3 and B 4 hen AB 2 m 1,k 1 SCAL n 1 C 1 A n 1 B Now the matrices look like γ 0,0 γ 0,1 γ 0,n 1 α 0,0 β 0,0 β 0,1 β 0,n 1 α 0,0 β 0,0 β 0,1 β 0,n 1 α 0,0 β 0,0 α 0,0 β 0,1 α 0,0 β 0,n 1 In other words, C and B are just row vectors and A is a scalar he vector C is computed by scaling the row vector B by the scalar A Homework 4443 Let A 4 and B hen AB m 1,n 1 DO 1 C k 1 A k B he matrices look like γ 0,0 α 0,0 α 0,1 α 0,k 1 β 0,0 β 1,0 k 1 α 0,p β p,0 p0 β k 1,0 In other words, C is a scalar that is computed by taking the dot product of the one row that is A and the one column that is B 2 Homework 4444 Let A and B 1 hen AB 0

34 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 150 k 1 outer product n n 1 B m C m A γ 0,0 γ 0,1 γ 0,n 1 γ 1,0 γ 1,1 γ 1,n 1 α 0,0 α 1,0 β 0,0 β 0,1 β 0,n 1 γ m 1,0 γ m 1,1 γ m 1,n 1 α m 1,0 α 0,0 β 0,0 α 0,0 β 0,1 α 0,0 β 0,n 1 α 1,0 β 0,0 α 1,0 β 0,1 α 1,0 β 0,n 1 α m 1,0 β 0,0 α m 1,0 β 0,1 α m 1,0 β 0,n 1 Homework 4445 Let A and B 1 2 hen AB

35 44 Matrix-Matrix Multiplication Product 151 Homework 4446 Let a and b C 1 c 0 c 1 2 and C ab Partition C by columns and by rows: and C c 0 c 1 c 2 hen c c C c c 1 3 c C rue/false rue/false rue/false rue/false rue/false rue/false rue/false Homework 4447 Fill in the boxes:

36 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 152 Homework 4448 Fill in the boxes: n 1 matrix-vector product m C m A k k B γ 0,0 γ 1,0 γ m 1,0 α 0,0 α 0,1 α 0,k 1 α 1,0 α 1,1 α 1,k 1 α m 1,0 α m 1,1 α m 1,k 1 β 0,0 β 1,0 β k 1,0 We have studied this special case in great detail o emphasize how it relates to have matrix-matrix multiplication is computed, consider the following: γ 0,0 γ i,0 γ m 1,0 α 0,0 α 0,1 α 0,k 1 α i,0 α i,1 α i,k 1 α m 1,0 α m 1,1 α m 1,k 1 β 0,0 β 1,0 β k 1,0 m 1 row vector-matrix product 1 C n 1 A k k B n γ 0,0 γ 0,1 γ 0,n 1 α 0,0 α 0,1 α 0,k 1 β 0,0 β 0,1 β 0,n 1 β 1,0 β 1,1 β 1,n 1 β k 1,0 β k 1,1 β k 1,n 1

37 44 Matrix-Matrix Multiplication Product 153 so that γ 0, j k 1 p0 α 0,pβ p, j o emphasize how it relates to have matrix-matrix multiplication is computed, consider the following: γ 0,0 γ 0, j γ 0,n 1 α 0,0 α 0,1 α 0,k 1 β 0,0 β 0, j β 0,n 1 β 1,0 β 1, j β 1,n 1 β k 1,0 β k 1, j β k 1,n 1 Homework 4449 Let A and B hen AB Homework Let e i R m equal the ith unit basis vector and A R m n hen e i A ǎ i, the ith row of A Always/Sometimes/Never Homework Get as much practice as you want with the MALAB script in LAFF-20xM/Programming/Week04/PracticeGemmm If you understand how to perform a matrix-matrix multiplication, then you know how to perform all other operations with matrices and vectors that we have encountered so far 445 Cost View at edx Consider the matrix-matrix multiplication C AB where C R m n, A R m k, and B R k n Let us examine what the cost of this operation is: We argued that, by definition, the jth column of C, c j, is computed by the matrix-vector multiplication Ab j, where b j is the jth column of B Last week we learned that a matrix-vector multiplication of a m k matrix times a vector of size k requires 2mk floating point operations flops C has n columns since it is a m n matrix Putting all these observations together yields a cost of n 2mk 2mnk flops

38 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 154 ry this! Recall that the dot product of two vectors of size k requires approximately 2k flops We learned in the previous units that if C AB then γ i, j equals the dot product of the ith row of A and the jth column of B Use this to give an alternative justification that a matrix multiplication requires 2mnk flops 45 Enrichment 451 Markov Chains: heir Application Matrices have many real world applications As we have seen this week, one noteworthy use is connected to Markov chains here are many, many examples of the use of Markov chains You can find a brief look at some significant applications in HE FIVE GREAES APPLICAIONS OF MARKOV CHAINS by Philipp von Hilgers and Amy N Langville 46 Wrap Up 461 Homework Homework 4611 Let A R m n and x R n hen Ax x A Always/Sometimes/Never Homework 4612 Our laff library has a routine that has the following property laff gemv trans, alpha, A, x, beta, y laff gemv No transpose, alpha, A, x, beta, y computes y : αax + βy laff gemv ranspose, alpha, A, x, beta, y computes y : αa x + βy he routine works regardless of whether x and/or y are column and/or row vectors Our library does NO include a routine to compute y : x A What call could you use to compute y : x A if y is stored in yt and x in xt? laff gemv No transpose, 10, A, xt, 00, yt laff gemv No transpose, 10, A, xt, 10, yt laff gemv ranspose, 10, A, xt, 10, yt laff gemv ranspose, 10, A, xt, 00, yt Homework 4613 Let A Compute A 2 A 3 For k > 1, A k

39 46 Wrap Up 155 Homework 4614 Let A A 2 A 3 For n 0, A 2n For n 0, A 2n+1 Homework 4615 Let A A 2 A 3 For n 0, A 4n For n 0, A 4n+1 Homework 4616 Let A be a square matrix If AA 0 the zero matrix then A is a zero matrix AA is often written as A 2 rue/false Homework 4617 here exists a real valued matrix A such that A 2 I Recall: I is the identity Homework 4618 here exists a matrix A that is not diagonal such that A 2 I rue/false rue/false 462 Summary Partitioned matrix-vector multiplication A 0,0 A 0,1 A 0,N 1 A 1,0 A 1,1 A 1,N 1 A M 1,0 A M 1,1 A M 1,N 1 x 0 x 1 x N 1 A 0,0 x 0 + A 0,1 x A 0,N 1 x N 1 A 1,0 x 0 + A 1,1 x A 1,N 1 x N 1 A M 1,0 x 0 + A M 1,1 x A M 1,N 1 x N 1 ransposing a partitioned matrix A 0,0 A 0,1 A 0,N 1 A 1,0 A 1,1 A 1,N 1 A 0,0 A 1,0 A M 1,0 A 0,1 A 1,1 A M 1,1 A M 1,0 A M 1,1 A M 1,N 1 A 0,N 1 A 1,N 1 A M 1,N 1

40 Week 4 From Matrix-Vector Multiplication to Matrix-Matrix Multiplication 156 Composing linear transformations Let L A : R k R m and L B : R n R k both be linear transformations and, for all x R n, define the function L C : R n R m by L C x L A L B x hen L C x is a linear transformations Matrix-matrix multiplication AB A b 0 b 1 b n 1 Ab 0 Ab 1 Ab n 1 If C γ 0,0 γ 0,1 γ 0,n 1 γ 1,0 γ 1,1 γ 1,n 1, A α 0,0 α 0,1 α 0,k 1 α 1,0 α 1,1 α 1,k 1, γ m 1,0 γ m 1,1 γ m 1,n 1 α m 1,0 α m 1,1 α m 1,k 1 β 0,0 β 0,1 β 0,n 1 β 1,0 β 1,1 β 1,n 1 and B β k 1,0 β k 1,1 β k 1,n 1 then C AB means that γ i, j k 1 p0 α i,pβ p, j A table of matrix-matrix multiplications with matrices of special shape is given at the end of this week Outer product Let x R m and y R n hen the outer product of x and y is given by xy Notice that this yields an m n matrix: xy χ 0 χ 1 ψ 0 ψ 1 χ 0 χ 1 ψ 0 ψ 1 ψ n 1 χ m 1 ψ n 1 χ m 1 χ 0 ψ 0 χ 0 ψ 1 χ 0 ψ n 1 χ 1 ψ 0 χ 1 ψ 1 χ 1 ψ n 1 χ m 1 ψ 0 χ m 1 ψ 1 χ m 1 ψ n 1

41 46 Wrap Up 157 m n k Shape Comment C 1 A 1 B Scalar multiplication 1 B m 1 1 m C m A Vector times scalar scalar times vector 1 n 1 n 1 C 1 A n 1 B Scalar times row vector 1 1 k 1 C k 1 A k B Dot product with row and column n n 1 B m n 1 m C m A Outer product k m 1 k m C m A k B Matrix-vector multiplication 1 n k n 1 C k 1 A k n B Row vector times matrix multiply