Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Linear algebra in your daily (digital) life
Andrew Schultz
Wellesley College
December 11, 2017
Andrew Schultz Linear algebra in your daily (digital) life Google’s PageRank algorithm Who cares about orthonormality? Filtering noise Image compression Who cares about change of basis? Image compression Sound compression
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Outline of the talk
Who cares about eigenvalues?
Andrew Schultz Linear algebra in your daily (digital) life Who cares about orthonormality? Filtering noise Image compression Who cares about change of basis? Image compression Sound compression
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Outline of the talk
Who cares about eigenvalues? Google’s PageRank algorithm
Andrew Schultz Linear algebra in your daily (digital) life Filtering noise Image compression Who cares about change of basis? Image compression Sound compression
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Outline of the talk
Who cares about eigenvalues? Google’s PageRank algorithm Who cares about orthonormality?
Andrew Schultz Linear algebra in your daily (digital) life Who cares about change of basis? Image compression Sound compression
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Outline of the talk
Who cares about eigenvalues? Google’s PageRank algorithm Who cares about orthonormality? Filtering noise Image compression
Andrew Schultz Linear algebra in your daily (digital) life Image compression Sound compression
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Outline of the talk
Who cares about eigenvalues? Google’s PageRank algorithm Who cares about orthonormality? Filtering noise Image compression Who cares about change of basis?
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Outline of the talk
Who cares about eigenvalues? Google’s PageRank algorithm Who cares about orthonormality? Filtering noise Image compression Who cares about change of basis? Image compression Sound compression
Andrew Schultz Linear algebra in your daily (digital) life V is the vertex set for G A ⊆ V × V is the set of arcs of G Want to rank vertices (using R : V → R) should depend only on structure of G shouldn’t produce many ties should be equitable should be stable under “attack” should be computable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph
1 o 2 ( 3 ^ h @I
Ù 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life Want to rank vertices (using R : V → R) should depend only on structure of G shouldn’t produce many ties should be equitable should be stable under “attack” should be computable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph V is the vertex set for G A ⊆ V × V is the set of arcs of G
1 o 2 ( 3 ^ h @I
Ù 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life should depend only on structure of G shouldn’t produce many ties should be equitable should be stable under “attack” should be computable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph V is the vertex set for G A ⊆ V × V is the set of arcs of G
Want to rank vertices (using R : V → ) 1 o 2 ( 3 R ^ h @I
Ù 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life shouldn’t produce many ties should be equitable should be stable under “attack” should be computable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph V is the vertex set for G A ⊆ V × V is the set of arcs of G
Want to rank vertices (using R : V → ) 1 o 2 ( 3 R ^ h @I should depend only on structure of G Ù 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life should be equitable should be stable under “attack” should be computable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph V is the vertex set for G A ⊆ V × V is the set of arcs of G
Want to rank vertices (using R : V → ) 1 o 2 ( 3 R ^ h @I should depend only on structure of G Ù shouldn’t produce many ties 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life should be stable under “attack” should be computable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph V is the vertex set for G A ⊆ V × V is the set of arcs of G
Want to rank vertices (using R : V → ) 1 o 2 ( 3 R ^ h @I should depend only on structure of G Ù shouldn’t produce many ties 4 ' 5 should be equitable
Andrew Schultz Linear algebra in your daily (digital) life should be computable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph V is the vertex set for G A ⊆ V × V is the set of arcs of G
Want to rank vertices (using R : V → ) 1 o 2 ( 3 R ^ h @I should depend only on structure of G Ù shouldn’t produce many ties 4 ' 5 should be equitable should be stable under “attack”
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The general constraints
Suppose that G is a directed graph V is the vertex set for G A ⊆ V × V is the set of arcs of G
Want to rank vertices (using R : V → ) 1 o 2 ( 3 R ^ h @I should depend only on structure of G Ù shouldn’t produce many ties 4 ' 5 should be equitable should be stable under “attack” should be computable
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 1: R i = j=1 Li,j
Page 3 wins (score 3), Pages 1,2,5 tie for second (score 2), Page 4 loses
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
Form matrix L so that 1 , if j / i L = i,j 0 , else
Andrew Schultz Linear algebra in your daily (digital) life 01100 Pn Scheme 1: R i = Li,j 00101 j=1 01011 Page 3 wins (score 3), Pages 1,2,5 00000 tie for second (score 2), Page 4 loses 10100
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 1: R i = j=1 Li,j
Page 3 wins (score 3), Pages 1,2,5 tie for second (score 2), Page 4 loses
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else 01100 00101 01011 00000 10100
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 1: R i = j=1 Li,j
Page 3 wins (score 3), Pages 1,2,5 tie for second (score 2), Page 4 loses
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else 01100 00101 01011 00000 10100
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 1: R i = j=1 Li,j
Page 3 wins (score 3), Pages 1,2,5 tie for second (score 2), Page 4 loses
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else 01100 00101 01011 00000 10100
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 1: R i = j=1 Li,j
Page 3 wins (score 3), Pages 1,2,5 tie for second (score 2), Page 4 loses
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else 01100 00101 01011 00000 10100
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 1: R i = j=1 Li,j
Page 3 wins (score 3), Pages 1,2,5 tie for second (score 2), Page 4 loses
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else 01100 00101 01011 00000 10100
Andrew Schultz Linear algebra in your daily (digital) life Page 3 wins (score 3), Pages 1,2,5 tie for second (score 2), Page 4 loses
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else 01100 Pn Scheme 1: R i = Li,j 00101 j=1 01011 00000 10100
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs The first approach
j / i is a vote for i from j
1 o 2 ( 3 ^ h @I Form matrix L so that Ù 1 , if j / i 4 ' 5 L = i,j 0 , else 01100 Pn Scheme 1: R i = Li,j 00101 j=1 01011 Page 3 wins (score 3), Pages 1,2,5 00000 tie for second (score 2), Page 4 loses 10100
Andrew Schultz Linear algebra in your daily (digital) life Evil users can manipulate results Not equitable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with first approach
Potential for lots of ties
Andrew Schultz Linear algebra in your daily (digital) life Evil users can manipulate results Not equitable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with first approach
01100 Potential for lots of ties 0 0101 0 1 0 1 1 0 0 0 0 0 10100
Andrew Schultz Linear algebra in your daily (digital) life Not equitable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with first approach
Potential for lots of ties 1 o 2 ( 3 ^ h @I Evil users can manipulate results Ù 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life Not equitable
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with first approach
20 200 Potential for lots of ties Evil users can manipulate ( 1 o 2 h 3 results _ ?I Ù 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with first approach
0 1 1 0 0 Potential for lots of ties 0 0 1 0 1 Evil users can manipulate 0 1 0 1 1 results 0 0 0 0 0 Not equitable 10100
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with first approach
0 1 1 0 0 Potential for lots of ties 0 0 1 0 1 Evil users can manipulate 0 1 0 1 1 results 0 0 0 0 0 Not equitable 10100
Andrew Schultz Linear algebra in your daily (digital) life Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300 Pn Scheme 2: R i = j=1 Wi,j
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 2: R i = j=1 Wi,j
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5 Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 2: R i = j=1 Wi,j
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5 Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 2: R i = j=1 Wi,j
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5 Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 2: R i = j=1 Wi,j
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5 Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 2: R i = j=1 Wi,j
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5 Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300
Andrew Schultz Linear algebra in your daily (digital) life Pn Scheme 2: R i = j=1 Wi,j
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5 Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Updating our approach
Each page gets a total of 1 vote:
1 o 2 ( 3 X ^ h @I `(j) = Li,j i Ù 4 ' 5 Form matrix W so that 00 .50 .3300 1 , if j / i 000 .3300 .5 `(j) Wi,j = 00 .5010 .5 0 , else 00000 100 .3300 Pn Scheme 2: R i = j=1 Wi,j
Andrew Schultz Linear algebra in your daily (digital) life Gives importance to links from unimportant pages
Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with second approach
Evil users can really manipulate results
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with second approach
Evil users can really manipulate results Gives importance to links from unimportant pages
Andrew Schultz Linear algebra in your daily (digital) life Who cares about eigenvalues? Who cares about matrix factorization? Who cares about change of basis?
Voting in Directed Graphs Problems with second approach
1 o 2 ( 3 ^ h @I
Ù Evil users can really manipulate 4 ' 5 results 0 0.5 0.33 0 0 Gives importance to links from 0 0 0.33 0 0.5 unimportant pages 0 0.5 010 .5 0 0 0 0 0 1 0 0.33 0 0
Andrew Schultz Linear algebra in your daily (digital) life 0 0.5 0.33 0 0 0 0 0.33 0 0.5 0 0.5 010 .5 0 0 0 0 0 1 0 0.33 0 0
0.41 0.47 0.52 is 1-eigenvector 0 0.58
Want