Submodular Functions, Optimization, and Applications to Machine Learning
|
|
- Claire Fitzgerald
- 5 years ago
- Views:
Transcription
1 Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture 5 Prof. Jeff Bilmes University of Washington, Seattle Department of Electrical Engineering April 9th, 2018 Clockwise from top left:v Lásló Lovász Jack Edmonds Satoru Fujishige George Nemhauser Laurence Wolsey András Frank Lloyd Shapley H. Narayanan Robert Bixby William Cunningham William Tutte Richard Rado Alexander Schrijver Garrett Birkhoff Hassler Whitney Richard Dedekind Prof. Jeff Bilmes + f (A) + f (B) f (A [ B) = f (Ar ) + 2f (C ) + f (Br ) = f (Ar ) +f (C ) + f (Br ) f (A \ B) = f (A \ B) EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F1/66 (pg.1/187)
2 Logistics Review Cumulative Outstanding Reading Read chapter 1 from Fujishige s book. Read chapter 2 from Fujishige s book. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F2/66 (pg.2/187)
3 Logistics Review Announcements, Assignments, and Reminders Homework 1 out, due Monday, 4/9/ :59pm electronically via our assignment dropbox ( If you have any questions about anything, please ask then via our discussion board ( Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F3/66 (pg.3/187)
4 Logistics Class Road Map - EE563 Review L1(3/26): Motivation, Applications, & Basic Definitions, L2(3/28): Machine Learning Apps (diversity, complexity, parameter, learning target, surrogate). L3(4/2): Info theory exs, more apps, definitions, graph/combinatorial examples L4(4/4): Graph and Combinatorial Examples, Matrix Rank, Examples and Properties, visualizations L5(4/9): More Examples/Properties/ Other Submodular Defs., Independence, Matroids L6(4/11): L7(4/16): L8(4/18): L9(4/23): L10(4/25): L11(4/30): L12(5/2): L13(5/7): L14(5/9): L15(5/14): L16(5/16): L17(5/21): L18(5/23): L (5/28): Memorial Day (holiday) L19(5/30): L21(6/4): Final Presentations maximization. Last day of instruction, June 1st. Finals Week: June 2-8, Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F4/66 (pg.4/187)
5 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.5/187)
6 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.6/187)
7 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Weighted set cover f(a) =w( S a2a U a), othercoverfunctions,cut functions. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.7/187)
8 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Weighted set cover f(a) =w( S a2a U a), othercoverfunctions,cut functions. Matrix rank r(a), the dimensionality of the vector space spanned by the set of vectors {x a } a2a. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.8/187)
9 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Weighted set cover f(a) =w( S a2a U a), othercoverfunctions,cut functions. Matrix rank r(a), the dimensionality of the vector space spanned by the set of vectors {x a } a2a. Adding modular functions to submodular functions preserves submodularity. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.9/187)
10 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Weighted set cover f(a) =w( S a2a U a), othercoverfunctions,cut functions. Matrix rank r(a), thedimensionalityofthevectorspacespannedby the set of vectors {x a } a2a. Adding modular functions to submodular functions preserves submodularity. Conic P mixtures: if i 0 and f i :2 V! R is submodular, then so is i if i. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.10/187)
11 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Weighted set cover f(a) =w( S a2a U a), othercoverfunctions,cut functions. Matrix rank r(a), thedimensionalityofthevectorspacespannedby the set of vectors {x a } a2a. Adding modular functions to submodular functions preserves submodularity. Conic P mixtures: if i 0 and f i :2 V! R is submodular, then so is i if i. Restrictions: f 0 (A) =f(a \ S) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.11/187)
12 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Weighted set cover f(a) =w( S a2a U a), othercoverfunctions,cut functions. Matrix rank r(a), thedimensionalityofthevectorspacespannedby the set of vectors {x a } a2a. Adding modular functions to submodular functions preserves submodularity. Conic P mixtures: if i 0 and f i :2 V! R is submodular, then so is i if i. Restrictions: f 0 (A) =f(a \ S) max: f(a) = max j2a c j and facility location. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.12/187)
13 Logistics Summary submodular properties Review c(a), numberofconnectedcomponentsinducedbya E(G) is supermodular. f(x) =m 1 X X M1 X submodular iff off-diagonal elements of M non-positive. Weighted set cover f(a) =w( S a2a U a), othercoverfunctions,cut functions. Matrix rank r(a), thedimensionalityofthevectorspacespannedby the set of vectors {x a } a2a. Adding modular functions to submodular functions preserves submodularity. Conic P mixtures: if i 0 and f i :2 V! R is submodular, then so is i if i. Restrictions: f 0 (A) =f(a \ S) max: f(a) = max j2a c j and facility location. Log determinant f(a) = log det( A ) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F5/66 (pg.13/187)
14 - Concave over non-negative modular Let m 2 R E + be a non-negative modular function, and over R. Definef :2 E! R as aconcavefunction f(a) = (m(a)) (5.1) then f is submodular. # E " :v lrt othkcttt ) m ( = L MIAI ) min ) ' III. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F6/66 (pg.14/187)
15 Concave over non-negative modular Let m 2 R E + be a non-negative modular function, and over R. Definef :2 E! R as aconcavefunction then f is submodular. f(a) = (m(a)) (5.1) a #... atc b btc Proof. Given A B E \ v, wehave0applea = m(a) apple b = m(b), and 0 apple c = m(v). Forg concave, we have (a + c) (a) (b + c) (b), and thus (m(a)+m(v)) (m(a)) (m(b)+m(v)) (m(b)) (5.2) Aformofconverseistrueaswell. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F6/66 (pg.15/187)
16 m ( A) =/ A ) lvl =3 OH tea, N µ 1011AM
17 - normchtel - monotone Concave composed with non-negative modular Theorem Given a ground set V.Thefollowingtwoareequivalent: 1 For all modular functions m :2 V! R +,thenf :2 V! R defined as f(a) = (m(a)) is submodular 2 : R +! R is concave. If is non-decreasing concave w. (0) = 0, thenf is polymatroidal. fl 0/7=0 non - dlcmcsihy - Submodulm. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F7/66 (pg.16/187)
18 Concave composed with non-negative modular Theorem Given a ground set V.Thefollowingtwoareequivalent: 1 For all modular functions m :2 V! R +,thenf :2 V! R defined as f(a) = (m(a)) is submodular 2 : R +! R is concave. If is non-decreasing concave w. (0) = 0, thenf is polymatroidal. Sums of concave over modular functions are submodular KX f(a) = i(m i (A)) (5.3) i=1 HAl=Ifvy m( Ethan Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F7/66 (pg.17/187)
19 Concave composed with non-negative modular Theorem Given a ground set V.Thefollowingtwoareequivalent: 1 For all modular functions m :2 V! R +,thenf :2 V! R defined as f(a) = (m(a)) is submodular 2 : R +! R is concave. If is non-decreasing concave w. (0) = 0, thenf is polymatroidal. Sums of concave over modular functions are submodular KX f(a) = i(m i (A)) (5.3) i=1 Very large class of functions, including graph cut, bipartite neighborhoods, set cover (Stobbe & Krause 2011), and feature-based submodular functions (Wei, Iyer, & Bilmes 2014). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F7/66 (pg.18/187)
20 Concave composed with non-negative modular Theorem Given a ground set V.Thefollowingtwoareequivalent: 1 For all modular functions m :2 V! R +,thenf :2 V! R defined as f(a) = (m(a)) is submodular 2 : R +! R is concave. If is non-decreasing concave w. (0) = 0, thenf is polymatroidal. Sums of concave over modular functions are submodular KX f(a) = i(m i (A)) (5.3) i=1 Very large class of functions, including graph cut, bipartite neighborhoods, set cover (Stobbe & Krause 2011), and feature-based submodular functions (Wei, Iyer, & Bilmes 2014). However, Vondrak showed that a graphic matroid rank function over K 4 (we ll define this after we define matroids) are not members. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F7/66 (pg.19/187)
21 Monotonicity Definition Afunctionf :2 V! R is monotone nondecreasing (resp. monotone increasing) ifforalla B, wehavef(a) apple f(b) (resp. f(a) <f(b)). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F8/66 (pg.20/187)
22 Monotonicity Definition Afunctionf :2 V! R is monotone nondecreasing (resp. monotone. O increasing) ifforalla B, wehavef(a) apple f(b) (resp. f(a) <f(b)). Definition Afunctionf :2 V! R is monotone nonincreasing (resp. monotone decreasing) ifforalla B, wehavef(a) f(b) (resp. f(a) >f(b)). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F8/66 (pg.21/187)
23 Composition of non-decreasing submodular and non-decreasing concave Theorem Given two functions, one defined on sets and another continuous valued one: f :2 V! R (5.4) : R! R (5.5) the composition formed as h = f :2 V! R (defined as h(s) = (f(s))) isnondecreasingsubmodular,if is non-decreasing concave and f is nondecreasing submodular. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F9/66 (pg.22/187)
24 Monotone difference of two functions Let f and g both be submodular functions on subsets of V and let (f g)( ) be either monotone non-decreasing or monotone non-increasing Then h :2 V! R defined by is submodular. Proof. h(a) =min(f(a),g(a)) (5.6) If h(a) agrees with f on both X and Y (or g on both X and Y ), and since h(x)+h(y )=f(x)+f(y ) f(x [ Y )+f(x \ Y ) (5.7) or h(x)+h(y )=g(x)+g(y ) g(x [ Y )+g(x \ Y ), (5.8) the result (Equation 5.6 being submodular) follows since f(x)+f(y ) g(x)+g(y ) min(f(x [ Y ),g(x [ Y )) + min(f(x \ Y ),g(x \ Y )) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F10/66 (pg.23/187) (5.9)...
25 Monotone difference of two functions...cont. Otherwise, w.l.o.g., h(x) =f(x) and h(y ) =g(y ), giving h(x)+h(y )=f(x)+g(y) f(x [ Y )+f(x \ Y )+g(y) f(y ) (5.10) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F11/66 (pg.24/187)
26 - Monotone difference of two functions...cont. Otherwise, w.l.o.g., h(x) =f(x) and h(y ) =g(y ), giving h(x)+h(y )=f(x)+g(y) f(x [ Y )+f(x \ Y )+g(y) f(y ) (5.10) f ( XVY ) - g ( xvy ) Z f 14 ) Assume the case where f g is monotone non-decreasing Hence, f(x [ Y )+g(y) f(y ) g(x [ Y ) giving - g ( Y ) h(x)+h(y ) g(x [ Y )+f(x \ Y ) h(x [ Y )+h(x \ Y ) (5.11) What is an easy way to : prove the case where f non-increasing? ring g is monotone Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F11/66 (pg.25/187)
27 non Saturation via the min( ) function non - chlrebsirz Let f :2 V! R be a monotone increasing mm or decreasing - submodular function and let be a constant. Then the function h :2 V! R defined by non - ihcrlaskr h(a) =min(, f(a)) (5.12) is submodular. mink,x ). if f is non. - Mercury h( 4=2 = flu ) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F12/66 (pg.26/187)
28 Saturation via the min( ) function Let f :2 V! R be a monotone increasing or decreasing submodular function and let be a constant. Then the function h :2 V! R defined by is submodular. Proof. h(a) =min(, f(a)) (5.12) For constant k, wehavethat(f k) is non-decreasing (or non-increasing) so this follows from the previous result. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F12/66 (pg.27/187)
29 Saturation via the min( ) function Let f :2 V! R be a monotone increasing or decreasing submodular function and let be a constant. Then the function h :2 V! R defined by is submodular. Proof. h(a) =min(, f(a)) (5.12) For constant k, wehavethat(f k) is non-decreasing (or non-increasing) so this follows from the previous result. Note also, g(a) =min(k, a) for constant k is a non-decreasing concave function, so when f is monotone nondecreasing submodular, we can use the earlier result about composing a monotone concave function with a monotone submodular function to get a version of this. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F12/66 (pg.28/187)
30 More on Min - the saturate trick In general, the minimum of two submodular functions is not submodular (unlike concave functions, closed under min). I *.me#antis@ Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F13/66 (pg.29/187)
31 More on Min - the saturate trick In general, the minimum of two submodular functions is not submodular (unlike concave functions, closed under min). However, when wishing to maximize two monotone non-decreasing submodular functions f,g, wecandefinefunctionh :2 V! R as h (A) = 1 2 min(, f(a)) + min(, g(a)) (5.13) then h is submodular, and h (A) if and only if both f(a) and g(a), forconstant 2 R. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F13/66 (pg.30/187)
32 More on Min - the saturate trick In general, the minimum of two submodular functions is not submodular (unlike concave functions, closed under min). However, when wishing to maximize two monotone non-decreasing submodular functions f,g, wecandefinefunctionh :2 V! R as h (A) = 1 2 min(, f(a)) + min(, g(a)) surrogut for mm ( FIA ), g ( A ) ) (5.13) then h is submodular, and h (A) if and only if both f(a) and g(a), forconstant 2 R. This can be useful in many applications. An instance of a submodular surrogate (where we take a non-submodular problem and find a submodular one that can tell us something about it). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F13/66 (pg.31/187)
33 Arbitrary functions: difference between submodular funcs. Theorem Given an arbitrary set function h, itcanbeexpressedasadifference between two submodular functions (i.e., 8h 2 2 V! R, 9f,g s.t. 8A, h(a) =f(a) g(a) where both f and g are submodular). Proof. Let h be given and arbitrary, and define: If = min X,Y :X6 Y,Y 6 X h(x)+h(y ) h(x [ Y ) h(x \ Y ) 0 then h is submodular, so by assumption <0. (5.14) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F14/66 (pg.32/187)
34 Arbitrary functions: difference between submodular funcs. Theorem Given an arbitrary set function h, itcanbeexpressedasadifference between two submodular functions (i.e., 8h 2 2 V! R, 9f,g s.t. 8A, h(a) =f(a) g(a) where both f and g are submodular). Proof. Let h be given and arbitrary, and define: = min X,Y :X6 Y,Y 6 X h(x)+h(y ) h(x [ Y ) h(x \ Y ) (5.14) If 0 then h is submodular, so by assumption <0. Nowletf be an arbitrary strict submodular function and define = min f(x)+f(y ) f(x [ Y ) f(x \ Y ). (5.15) X,Y :X6 Y,Y 6 X Strict means that >0. FLA ) = MAT... Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F14/66 (pg.33/187)
35 Arbitrary functions as difference between submodular funcs....cont. Define h 0 :2 V! R as h 0 (A) =h(a)+ f(a) (5.16) Then h 0 is submodular (why?), and h = h 0 (A) between two submodular functions as desired. f(a), adifference Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F15/66 (pg.34/187)
36 Gain We often wish to express the gain of an item j 2 V in context A, namely f(a [{j}) f(a). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F16/66 (pg.35/187)
37 Gain We often wish to express the gain of an item j 2 V in context A, namely f(a [{j}) f(a). This is called the gain and is used so often, there are equally as many ways to notate this. I.e., you might see: f(a [{j}) f(a) = j (A) (5.17) = A (j) (5.18) = r j f(a) (5.19) = f({j} A) (5.20) = f(j A) (5.21) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F16/66 (pg.36/187)
38 Gain We often wish to express the gain of an item j 2 V in context A, namely f(a [{j}) f(a). This is called the gain and is used so often, there are equally as many ways to notate this. I.e., you might see: We ll use f(j A). f(a [{j}) f(a) = j (A) (5.17) fljlaltgljltf = A (j) (5.18) = r j f(a) (5.19) = f({j} A) (5.20) = f(j A) (5.21) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F16/66 (pg.37/187)
39 - FIA Gain Vii We often wish to express the gain of an item j 2 V in context A, namely f(a [{j}) f(a). This is called the gain and is used so often, there are equally as many ways to notate this. I.e., you might see: A f(a [{j}) f(a) = j (A) (5.17) = A (j) (5.18) ) flj zfljln ) = r j f(a) (5.19) = f({j} A) (5.20) fl ;) z f ( ) furthµ At ;) f- Ii + ) HA ) Zflatj ) = f(j A) (5.21) We ll use f(j A). Submodularity s diminishing returns definition can be stated as saying that f(j A) is a monotone non-increasing function of A, since f(j A) f(j B) whenever A B (conditioning reduces valuation). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F16/66 (pg.38/187)
40 - Non norm. IAU Cose. fifth'la ) flj )+fla ) Z f ( At ;) tf( ) Doe ) this dyim svbnodulmity? fancy
41 Gain Notation It will also be useful to extend this to sets. Let A, B be any two sets. Then So when j is any singleton f(a B), f(a [ B) f(b) (5.22). f(j B) =f({j} B) =f({j}[b) f(b) (5.23) Bt ; = Busj ) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F17/66 (pg.39/187)
42 Gain Notation It will also be useful to extend this to sets. Let A, B be any two sets. Then So when j is any singleton f(a B), f(a [ B) f(b) (5.22) f(j B) =f({j} B) =f({j}[b) f(b) (5.23) Inspired from information theory notation and the notation used for conditional entropy H(X A X B )=H(X A,X B ) H(X B ). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F17/66 (pg.40/187)
43 . Totally normalized functions Any normalized submodular function g (even non-monotone) can be represented as a sum of a polymatroid (normalized monotone non-decreasing submodular) function ḡ and a modular function m g. min f ) ( VIAI ) E 2 Ha. A E v =f( v EIA mouth 2 function. h (A) A) + A h h(v1a7=f( HA ) +2 Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F18/66 (pg.41/187)
44 Totally normalized functions Any normalized submodular function g (even non-monotone) can be represented as a sum of a polymatroid (normalized monotone non-decreasing submodular) function ḡ and a modular function m g. Given arbitrary normalized submodular g :2 V! R, constructa function ḡ :2 V! R as follows: X ḡ(a) =g(a) g(a V \{a}) =g(a) m g (A) (5.24) a2a where m g (A), P a2a g(a V \{a}) is a modular function. g ( and Eglalt ) ttc via Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F18/66 (pg.42/187)
45 Totally normalized functions Any normalized submodular function g (even non-monotone) can be represented as a sum of a polymatroid (normalized monotone non-decreasing submodular) function ḡ and a modular function m g. Given arbitrary normalized submodular g :2 V! R, constructa function ḡ :2 V! R as follows: X ḡ(a) =g(a) g(a V \{a}) =g(a) m g (A) (5.24) a2a where m g (A), P a2a g(a V \{a}) is a modular function. ḡ is normalized since ḡ(;) =0. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F18/66 (pg.43/187)
46 Totally normalized functions Any normalized submodular function g (even non-monotone) can be represented as a sum of a polymatroid (normalized monotone non-decreasing submodular) function ḡ and a modular function m g. Given arbitrary normalized submodular g :2 V! R, constructa function ḡ :2 V! R as follows: X ḡ(a) =g(a) g(a V \{a}) =g(a) m g (A) (5.24) a2a where m g (A), P a2a g(a V \{a}) is a modular function. ḡ is normalized since ḡ(;) =0. ḡ is monotone non-decreasing since for v/2 A V : ḡ(v A) =g(v A) g(v V \{v}) 0 (5.25) Ty ( v1a7 70 man. nonmoving. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F18/66 (pg.44/187)
47 Totally normalized functions Any normalized submodular function g (even non-monotone) can be represented as a sum of a polymatroid (normalized monotone non-decreasing submodular) function ḡ and a modular function m g. Given arbitrary normalized submodular g :2 V! R, constructa function ḡ :2 V! R as follows: X ḡ(a) =g(a) g(a V \{a}) =g(a) m g (A) (5.24) a2a where m g (A), P a2a g(a V \{a}) is a modular function. ḡ is normalized since ḡ(;) =0. ḡ is monotone non-decreasing since for v/2 A V : ḡ(v A) =g(v A) g(v V \{v}) 0 (5.25) ḡ is called the totally normalized version of g. ) - I ( v1v\{ v } ) = 0 # Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F18/66 (pg.45/187)
48 Totally normalized functions Any normalized submodular function g (even non-monotone) can be represented as a sum of a polymatroid (normalized monotone non-decreasing submodular) function ḡ and a modular function m g. Given arbitrary normalized submodular g :2 V! R, constructa function ḡ :2 V! R as follows: X ḡ(a) =g(a) g(a V \{a}) =g(a) m g (A) (5.24) a2a where m g (A), P a2a g(a V \{a}) is a modular function. ḡ is normalized since ḡ(;) =0. ḡ is monotone non-decreasing since for v/2 A V : ḡ(v A) =g(v A) g(v V \{v}) 0 (5.25) ḡ is called the totally normalized version of g. Then g(a) =ḡ(a)+m g (A). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F18/66 (pg.46/187)
49 Arbitrary function as difference between two polymatroids Any normalized function h (i.e., h(;) =0) can be represented as a difference not only between submodular, but between polymatroid (normalized monotone non-decreasing submodular) functions. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F19/66 (pg.47/187)
50 Arbitrary function as difference between two polymatroids Any normalized function h (i.e., h(;) =0) can be represented as a difference not only between submodular, but between polymatroid (normalized monotone non-decreasing submodular) functions. Given submodular f and g, let f and ḡ be them totally normalized. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F19/66 (pg.48/187)
51 Arbitrary function as difference between two polymatroids Any normalized function h (i.e., h(;) =0) canberepresentedasa difference not only between submodular, but between polymatroid (normalized monotone non-decreasing submodular) functions. Given submodular f and g, let f and ḡ be them totally normalized. Given arbitrary h = f g where f and g are normalized submodular, h = f g = f + m f (ḡ + m g ) (5.26) = f ḡ +(m f m g ) (5.27) = f ḡ + m f h (5.28) = f + m + f g (ḡ +( m f g ) + ) (5.29) where m + i is the positive part of modular function m. Thatis, m + (A) = P a2a m(a)1(m(a) > 0). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F19/66 (pg.49/187)
52 Arbitrary function as difference between two polymatroids Any normalized function h (i.e., h(;) =0) canberepresentedasa difference not only between submodular, but between polymatroid (normalized monotone non-decreasing submodular) functions. Given submodular f and g, let f and ḡ be them totally normalized. Given arbitrary h = f g where f and g are normalized submodular, h = f g = f + m f (ḡ + m g ) (5.26) = f ḡ +(m f m g ) (5.27) = f ḡ + m f h (5.28) = f + m + f g (ḡ +( m f g ) + ) (5.29) where m + is the positive part of modular function m. Thatis, m + (A) = P a2a m(a)1(m(a) > 0). Both f + m + f g and ḡ +( m f g) + are polymatroid functions! Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F19/66 (pg.50/187)
53 Arbitrary function as difference between two polymatroids Any normalized function h (i.e., h(;) =0) canberepresentedasa difference not only between submodular, but between polymatroid (normalized monotone non-decreasing submodular) functions. Given submodular f and g, let f and ḡ be them totally normalized. Given arbitrary h = f g where f and g are normalized submodular, h = f g = f + m f (ḡ + m g ) (5.26) = f ḡ +(m f m g ) (5.27) = f ḡ + m f h (5.28) = f + m + f g (ḡ +( m f g ) + ) (5.29) where m + is the positive part of modular function m. Thatis, m + (A) = P a2a m(a)1(m(a) > 0). Both f + m + f g and ḡ +( m f g) + are polymatroid functions! Thus, any function can be expressed as a difference between two, not only submodular (DS), but polymatroid functions. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F19/66 (pg.51/187)
54 Two Equivalent Submodular Definitions Definition (submodular concave) Afunctionf :2 V! R is submodular if for any A, B V,wehavethat: f(a)+f(b) f(a [ B)+f(A \ B) (5.8) An alternate and (as we will soon see) equivalent definition is: Definition (diminishing returns) Afunctionf :2 V! R is submodular if for any A B V,and v 2 V \ B, wehavethat: f(a [{v}) f(a) f(b [{v}) f(b) (5.9) The incremental value, gain, or cost of v decreases (diminishes) as the context in which v is considered grows from A to B. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F20/66 (pg.52/187)
55 Submodular Definition: Group Diminishing Returns An alternate and equivalent definition is: Definition (group diminishing returns) Afunctionf :2 V! R is submodular if for any A B V,and C V \ B, wehavethat: f(a [ C) f(a) f(b [ C) f(b) (5.30) This means that the incremental value or gain of set C decreases as the context in which C is considered grows from A to B (diminishing returns) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F21/66 (pg.53/187)
56 Submodular Definition Basic Equivalencies We want to show that Submodular Concave (Definition 5.4.1), Diminishing Returns (Definition 5.4.2), and Group Diminishing Returns (Definition 5.4.1) are identical. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F22/66 (pg.54/187)
57 Submodular Definition Basic Equivalencies We want to show that Submodular Concave (Definition 5.4.1), Diminishing Returns (Definition 5.4.2), and Group Diminishing Returns (Definition 5.4.1) are identical. We will show that: Submodular Concave ) Diminishing Returns Diminishing Returns ) Group Diminishing Returns Group Diminishing Returns ) Submodular Concave Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F22/66 (pg.55/187)
58 Submodular Concave ) Diminishing Returns f(s)+f(t ) f(s [ T )+f(s \ T ) ) f(v A) f(v B),A B V \ v. Assume Submodular concave, so 8S, T we have f(s)+f(t ) f(s [ T )+f(s \ T ). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F23/66 (pg.56/187)
59 Submodular Concave ) Diminishing Returns f(s)+f(t ) f(s [ T )+f(s \ T ) ) f(v A) f(v B),A B V \ v. Assume Submodular concave, so 8S, T we have f(s)+f(t ) f(s [ T )+f(s \ T ). Given A, B and v 2 V such that: A B V \{v}, wehavefrom submodular concave that: f(a + v)+f(b) f(b + v)+f(a) (5.31) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F23/66 (pg.57/187)
60 Submodular Concave ) Diminishing Returns f(s)+f(t ) f(s [ T )+f(s \ T ) ) f(v A) f(v B),A B V \ v. Assume Submodular concave, so 8S, T we have f(s)+f(t ) f(s [ T )+f(s \ T ). Given A, B and v 2 V such that: A B V \{v}, wehavefrom submodular concave that: Rearranging, we have f(a + v)+f(b) f(b + v)+f(a) (5.31) f(a + v) f(a) f(b + v) f(b) (5.32) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F23/66 (pg.58/187)
61 Diminishing Returns ) Group Diminishing Returns f(v S) f(v T ),S T V \ v ) f(c A) f(c B),A B V \ C. Let C = {c 1,c 2,...,c k }.Thendiminishing returns implies f(a [ C) = f(a [ C) = f(a) kx 1 f(a [{c 1,...,c i}) i=1 kx f(a [{c 1...c i}) f(a [{c 1...c i 1}) = i=1 kx f(c i B [{c 1...c i 1}) = i=1 = f(b [ C) i=1 kx 1 f(b [{c 1,...,c i}) i=1 = f(b [ C) f(b) - f(a [{c 1,...,c i}) (5.33) f(a) (5.34) kx f(c i A [{c 1...c i 1}) (5.35) i=1 kx f(b [{c 1...c i}) f(b [{c 1...c i 1}) (5.36) f(b [{c 1,...,c i}) f(b) (5.37) (5.38) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F24/66 (pg.59/187)
62 Group Diminishing Returns ) Submodular Concave f(u S) f(u T ),S T V \ U ) f(a)+f(b) f(a [ B)+f(A \ B). Assume group diminishing returns. Assume A 6= B otherwise trivial. Define A 0 = A \ B, C = A \ B, andb 0 = B. ThensinceA 0 B 0, giving or f(a 0 + C) f(a 0 ) f(b 0 + C) f(b 0 ) (5.39) f(a 0 + C)+f(B 0 ) f(b 0 + C)+f(A 0 ) (5.40) f(a \ B + A \ B)+f(B) f(b + A \ B)+f(A \ B) (5.41) which is the same as the submodular concave condition f(a)+f(b) f(a [ B)+f(A \ B) (5.42) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F25/66 (pg.60/187)
63 Submodular Definition: Four Points Definition ( singleton, or four points ) Afunctionf :2 V! R is submodular if f for any A V,andany a, b 2 V \ A, wehavethat: f(a [{a})+f(a [{b}) f(a [{a, b})+f(a) (5.43) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F26/66 (pg.61/187)
64 - f( flats Submodular Definition: Four Points Definition ( singleton, or four points ) Afunctionf :2 V! R is submodular if f for any A V,andany a, b 2 V \ A, wehavethat: f(a [{a})+f(a [{b}) f(a [{a, b})+f(a) (5.43) This follows immediately from diminishing returns. fl Ata ) - A) f 2 ( Atsta ) f ( AIA ) 2 t ( altts ) ) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F26/66 (pg.62/187)
65 ... Submodular Definition: Four Points Definition ( singleton, or four points ) Afunctionf :2 V! R is submodular if f for any A V,andany a, b 2 V \ A, wehavethat: f(a [{a})+f(a [{b}) f(a [{a, b})+f(a) (5.43) This follows immediately from diminishing returns. Toachievediminishing returns, assumea B with B \ A = {b 1,b 2,...,b k }.Then f ( AIA ) 7- f ( al Ath ) Zfka ) Ats, th f(a + a) f(a) f(a + b 1 + a) f(a + b 1 ) (5.44) f(a + b 1 + b 2 + a) f(a + b 1 + b 2 ) (5.45)... ) z z fla / b ) (5.46) f(a + b b k + a) f(a + b b k ) (5.47) = f(b + a) f(b) (5.48) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F26/66 (pg.63/187)
66 Submodular on Hypercube Vertices Test submodularity via values on verticies of hypercube. Example: with V = n =2,thisis easy: With V = n =3,abitharder ana.tt?:tttntyng, D How many inequalities? Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F27/66 (pg.64/187)
67 Submodular Concave Diminishing Returns, in one slide. Theorem Given function f :2 V! R, then f(a)+f(b) f(a [ B)+f(A \ B) for all A, B V (SC) if and only if f(v X) f(v Y ) for all X Y V and v/2 Y (DR) Proof. (SC))(DR): Set A X [{v}, B Y.ThenA [ B = Y [{v} and A \ B = X and f(a) f(a \ B) f(a [ B) f(b) implies (DR). (DR))(SC): Order A \ B = {v 1,v 2,...,v r } arbitrarily. For i 2 1:r, f(v i (A \ B) [{v 1,v 2,...,v i 1 }) f(v i B [{v 1,v 2,...,v i 1 }). Applying telescoping summation to both sides, we get: rx rx f(v i (A \ B) [{v 1,v 2,...,v i 1 }) f(v i B [{v 1,v 2,...,v i 1 }) i=1 i=1 ) f(a) f(a \ B) f(a [ B) f(b) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F28/66 (pg.65/187)
68 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.66/187)
69 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.67/187)
70 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.68/187)
71 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) f(j S) f(j S [{k}), 8S V with j 2 V \ (S [{k}) (5.57) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.69/187)
72 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) f(j S) f(j S [{k}), 8S V with j 2 V \ (S [{k}) (5.57) f(a [ B A \ B) apple f(a A \ B)+f(B A \ B), 8A, B V (5.58) Conditional subadihut > Strong subadditivit, Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.70/187)
73 . Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) f(j S) f(j S [{k}), 8S V with j 2 V \ (S [{k}) (5.57) f(a [ B A \ B) apple f(a A \ B)+f(B A \ B), 8A, B V (5.58) f(t ) apple f(s)+ X X f(j S) f(j S [ T {j}), 8S, T V j2t \S j2s\t LB + ± ) fct = sut HSH ) = f ( ) f ( s ) + f ( Tls ft ) + 1 f Cs ) + us (5.59) ) f ( t ) EFLJ ) + UD - LB Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.71/187)
74 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) f(j S) f(j S [{k}), 8S V with j 2 V \ (S [{k}) (5.57) f(a [ B A \ B) apple f(a A \ B)+f(B A \ B), 8A, B V (5.58) f(t ) apple f(s)+ X X f(j S) f(j S [ T {j}), 8S, T V j2t \S j2s\t (5.59) f(t ) apple f(s)+ X f(j S), 8S T V (5.60) j2t \S Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.72/187)
75 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) f(j S) f(j S [{k}), 8S V with j 2 V \ (S [{k}) (5.57) f(a [ B A \ B) apple f(a A \ B)+f(B A \ B), 8A, B V (5.58) f(t ) apple f(s)+ X X f(j S) f(j S [ T {j}), 8S, T V j2t \S j2s\t (5.59) f(t ) apple f(s)+ X f(j S), 8S T V (5.60) f(t ) apple f(s) j2t \S X f(j S \{j})+ X f(j S \ T ) 8S, T V j2s\t j2t \S (5.61) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.73/187)
76 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) f(j S) f(j S [{k}), 8S V with j 2 V \ (S [{k}) (5.57) f(a [ B A \ B) apple f(a A \ B)+f(B A \ B), 8A, B V (5.58) f(t ) apple f(s)+ X X f(j S) f(j S [ T {j}), 8S, T V j2t \S j2s\t (5.59) f(t ) apple f(s)+ X f(j S), 8S T V (5.60) f(t ) apple f(s) f(t ) apple f(s) j2t \S X f(j S \{j})+ X f(j S \ T ) 8S, T V j2s\t j2t \S (5.61) X f(j S \{j}), 8T S V (5.62) j2s\t Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F30/66 (pg.74/187)
77 Equivalent Definitions of Submodularity We ve already seen that Eq Eq Eq Eq Eq Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F31/66 (pg.75/187)
78 Equivalent Definitions of Submodularity We ve already seen that Eq Eq Eq Eq Eq We next show that Eq ) Eq ) Eq ) Eq Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F31/66 (pg.76/187)
79 Approach To show these next results, we essentially first use: f(s [ T )=f(s)+f(t S) apple f(s)+upper-bound (5.63) and f(t )+lower-bound apple f(t )+f(s T)=f(S [ T ) (5.64) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F32/66 (pg.77/187)
80 Approach To show these next results, we essentially first use: f(s [ T )=f(s)+f(t S) apple f(s)+upper-bound (5.63) and f(t )+lower-bound apple f(t )+f(s T)=f(S [ T ) (5.64) leading to f(t )+lower-bound apple f(s)+upper-bound (5.65) or f(t ) apple f(s) +upper-bound lower-bound (5.66) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F32/66 (pg.78/187)
81 Eq ) Eq Let T \ S = {j 1,...,j r } and S \ T = {k 1,...,k q }. First, we upper bound the gain of T in the context of S: or f(s [ T ) f(s) = = rx f(s [{j 1,...,j t }) f(s [{j 1,...,j t 1 }) t=1 rx f(j t S [{j 1,...,j t t=1 = X j2t \S 1 }) apple (5.67) rx f(j t S) (5.68) t=1 f(j S) (5.69) f(t S) apple X j2t \S f(j S) (5.70) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F33/66 (pg.79/187)
82 Eq ) Eq Let T \ S = {j 1,...,j r } and S \ T = {k 1,...,k q }. Next, lower bound S in the context of T : qx f(s [ T ) f(t )= [f(t [{k 1,...,k t }) f(t [{k 1,...,k t 1 })] t=1 (5.71) qx qx = f(k t T [{k 1,...,k t }\{k t }) f(k t T [ S \{k t }) t=1 t=1 = X j2s\t (5.72) f(j S [ T \{j}) (5.73) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F34/66 (pg.80/187)
83 Eq ) Eq Let T \ S = {j 1,...,j r } and S \ T = {k 1,...,k q }. So we have the upper bound f(t S) =f(s [ T ) f(s) apple X j2t \S f(j S) (5.74) and the lower bound f(s T )=f(s [ T ) f(t ) X j2s\t f(j S [ T \{j}) (5.75) This gives upper and lower bounds of the form f(t )+lower bound apple f(s [ T ) apple f(s)+upper bound, (5.76) and combining directly the left and right hand side gives the desired inequality. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F35/66 (pg.81/187)
84 Eq ) Eq This follows immediately since if S T,thenS \ T = ;, andthelastterm of Eq vanishes. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F36/66 (pg.82/187)
85 Many (Equivalent) Definitions of Submodularity f(a) + f(b) f(a [ B)+f(A \ B), 8A, B V (5.54) f(j S) f(j T ), 8S T V, with j 2 V \ T (5.55) f(c S) f(c T ), 8S T V, with C V \ T (5.56) f(j S) f(j S [{k}), 8S V with j 2 V \ (S [{k}) (5.57) f(a [ B A \ B) apple f(a A \ B)+f(B A \ B), 8A, B V (5.58) f(t ) apple f(s)+ X X f(j S) f(j S [ T {j}), 8S, T V j2t \S j2s\t (5.59) f(t ) apple f(s)+ X f(j S), 8S T V (5.60) f(t ) apple f(s) f(t ) apple f(s) j2t \S X f(j S \{j})+ X f(j S \ T ) 8S, T V j2s\t j2t \S (5.61) X f(j S \{j}), 8T S V (5.62) j2s\t Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F37/66 (pg.83/187)
86 Eq ) Eq Here, we set T = S [{j, k}, j/2 S [{k} into Eq to obtain f(s [{j, k}) apple f(s)+f(j S)+f(k S) (5.77) = f(s)+f(s + {j}) f(s)+f(s + {k}) f(s) (5.78) = f(s + {j})+f(s + {k}) f(s) (5.79) = f(j S)+f(S + {k}) (5.80) giving f(j S [{k}) =f(s [{j, k}) f(s [{k}) (5.81) apple f(j S) (5.82) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F38/66 (pg.84/187)
87 Submodular Concave Why do we call the f(a)+f(b) f(a [ B)+f(A \ B) definition of submodularity, submodular concave? Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F39/66 (pg.85/187)
88 Submodular Concave Why do we call the f(a)+f(b) f(a [ B)+f(A \ B) definition of submodularity, submodular concave? Acontinuoustwicedifferentiablefunctionf : R n! R is concave if f r 2 f 0 (the Hessian matrix is nonpositive definite). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F39/66 (pg.86/187)
89 Submodular Concave Why do we call the f(a)+f(b) f(a [ B)+f(A \ B) definition of submodularity, submodular concave? Acontinuoustwicedifferentiablefunctionf : R n! R is concave if f r 2 f 0 (the Hessian matrix is nonpositive definite). Define a discrete derivative or difference operator defined on discrete functions f :2 V! R as follows: (r B f)(a), f(a [ B) f(a \ B) =f B (A \ B) (5.83) read as: the derivative of f at A in the direction B. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F39/66 (pg.87/187)
90 Submodular Concave Why do we call the f(a)+f(b) f(a [ B)+f(A \ B) definition of submodularity, submodular concave? Acontinuoustwicedifferentiablefunctionf : R n! R is concave if f r 2 f 0 (the Hessian matrix is nonpositive definite). Define a discrete derivative or difference operator defined on discrete functions f :2 V! R as follows: (r B f)(a), f(a [ B) f(a \ B) =f B (A \ B) (5.83) read as: the derivative of f at A in the direction B. Hence, if A \ B = ;, then(r B f)(a) =f(b A). Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F39/66 (pg.88/187)
91 Submodular Concave Why do we call the f(a)+f(b) f(a [ B)+f(A \ B) definition of submodularity, submodular concave? Acontinuoustwicedifferentiablefunctionf : R n! R is concave if f r 2 f 0 (the Hessian matrix is nonpositive definite). Define a discrete derivative or difference operator defined on discrete functions f :2 V! R as follows: : 0 (r B f)(a), f(a [ B) f(a \ B) =f B (A \ B) (5.83) read as: the derivative of f at A in the direction B. Hence, if A \ B = ;, then(r B f)(a) =f(b A). Consider a form of second derivative or 2nd difference: (r B f)(a) z } { (r C r B f)(a) =r C [ f(a [ B) f(a \ B) ] (5.84) =(r B f)(a [ C) (r B f)(a \ C) (5.85) = f(a [ B [ C) f((a [ C) \ B) f((a \ C) [ B)+f((A \ C) \ B) (5.86) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F39/66 (pg.89/187)
92 Submodular Concave If the second difference operator everywhere nonpositive: f(a [ B [ C) f((a [ C) \ B) f((a \ C) [ B)+f(A \ C \ B) apple 0 (5.87) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F40/66 (pg.90/187)
93 Submodular Concave If the second difference operator everywhere nonpositive: f(a [ B [ C) f((a [ C) \ B) then we have the equation: f((a \ C) [ B)+f(A \ C \ B) apple 0 (5.87) f((a [ C) \ B)+f((A \ C) [ B) f(a [ B [ C)+f(A \ C \ B) (5.88) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F40/66 (pg.91/187)
94 Submodular Concave If the second difference operator everywhere nonpositive: f(a [ B [ C) f((a [ C) \ B) then we have the equation: f((a \ C) [ B)+f(A \ C \ B) apple 0 (5.87) f((a [ C) \ B)+f((A \ C) [ B) f(a [ B [ C)+f(A \ C \ B) (5.88) Define A 0 =(A [ C) \ B and B 0 =(A \ C) [ B. Thentheabove implies: f(a 0 )+f(b 0 ) f(a 0 [ B 0 )+f(a 0 \ B 0 ) (5.89) and note that A 0 and B 0 so defined can be arbitrary. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F40/66 (pg.92/187)
95 Submodular Concave If the second difference operator everywhere nonpositive: f(a [ B [ C) f((a [ C) \ B) then we have the equation: f((a \ C) [ B)+f(A \ C \ B) apple 0 (5.87) f((a [ C) \ B)+f((A \ C) [ B) f(a [ B [ C)+f(A \ C \ B) (5.88) Define A 0 =(A [ C) \ B and B 0 =(A \ C) [ B. Thentheabove implies: f(a 0 )+f(b 0 ) f(a 0 [ B 0 )+f(a 0 \ B 0 ) (5.89) and note that A 0 and B 0 so defined can be arbitrary. One sense in which submodular functions are like concave functions. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F40/66 (pg.93/187)
96 Submodular Concave C C A B A B (a) A 0 =(A [ C) \ B (b) B 0 =(A \ C) [ B Figure: A figure showing A 0 [ B 0 = A [ B [ C and A 0 \ B 0 = A \ C \ B. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F41/66 (pg.94/187)
97 Submodular Concave C C A B A B (a) A 0 =(A [ C) \ B (b) B 0 =(A \ C) [ B Figure: A figure showing A 0 [ B 0 = A [ B [ C and A 0 \ B 0 = A \ C \ B. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F41/66 (pg.95/187)
98 Submodularity and Concave This submodular/concave relationship is more simply done with singletons. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F42/66 (pg.96/187)
99 Submodularity and Concave This submodular/concave relationship is more simply done with singletons. Recall four points definition: A function is submodular if for all X V and j, k 2 V \ X f(x + j)+f(x + k) f(x + j + k)+f(x) (5.90) f ( Al + FIB ) z fltub ) + fans ) # Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F42/66 (pg.97/187)
100 Submodularity and Concave This submodular/concave relationship is more simply done with singletons. Recall four points definition: A function is submodular if for all X V and j, k 2 V \ X f(x + j)+f(x + k) f(x + j + k)+f(x) (5.90) This gives us a simpler notion corresponding to concavity. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F42/66 (pg.98/187)
101 Submodularity and Concave This submodular/concave relationship is more simply done with singletons. Recall four points definition: A function is submodular if for all X V and j, k 2 V \ X f(x + j)+f(x + k) f(x + j + k)+f(x) (5.90) This gives us a simpler notion corresponding to concavity. Define gain as r j (X) =f(x + j) f(x), aformofdiscretegradient. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F42/66 (pg.99/187)
102 Submodularity and Concave This submodular/concave relationship is more simply done with singletons. Recall four points definition: A function is submodular if for all X V and j, k 2 V \ X f(x + j)+f(x + k) f(x + j + k)+f(x) (5.90) This gives us a simpler notion corresponding to concavity. Define gain as r j (X) =f(x + j) f(x), aformofdiscretegradient. Trivially becomes a second-order condition, akin to concave functions: AfunctionissubmodularifforallX V and j, k 2 V,wehave: r j r k f(x) apple 0 (5.91) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F42/66 (pg.100/187)
103 Example: Rank function of a matrix Consider the following 4 8 matrix, so V = {1, 2, 3, 4, 5, 6, 7, 8} = x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 Let A = {1, 2, 3}, B = {3, 4, 5}, C = {6, 7}, A r = {1}, B r = {5}. Then r(a) =3, r(b) =3, r(c) =2. r(a [ C) =3, r(b [ C) =3. r(a [ A r )=3, r(b [ B r )=3, r(a [ B r )=4, r(b [ A r )=4. r(a [ B) =4, r(a \ B) =1 <r(c) =2. 6= r(a)+r(b) =r(a [ B)+r(C) >r(a [ B)+r(A \ B) = 5 Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F43/66 (pg.101/187)
104 On Rank Let rank :2 V! Z + be the rank function. 0 Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F44/66 (pg.102/187)
105 On Rank Let rank :2 V! Z + be the rank function. In general, rank(a) apple A, and vectors in A are linearly independent if and only if rank(a) = A. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F44/66 (pg.103/187)
106 On Rank Let rank :2 V! Z + be the rank function. In general, rank(a) apple A, and vectors in A are linearly independent if and only if rank(a) = A. If A, B are such that rank(a) = A and rank(b) = B, with A < B, thenthespacespannedbyb is greater, and we can find a vector in B that is linearly independent of the space spanned by vectors in A. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F44/66 (pg.104/187)
107 On Rank Let rank :2 V! Z + be the rank function. In general, rank(a) apple A, and vectors in A are linearly independent if and only if rank(a) = A. If A, B are such that rank(a) = A and rank(b) = B, with A < B, thenthespacespannedbyb is greater, and we can find a vector in B that is linearly independent of the space spanned by vectors in A. To stress this point, note that the above condition is A < B, not A B which is sufficient (to be able to find an independent vector) but not required. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F44/66 (pg.105/187)
108 On Rank Let rank :2 V! Z + be the rank function. In general, rank(a) apple A, and vectors in A are linearly independent if and only if rank(a) = A. If A, B are such that rank(a) = A and rank(b) = B, with A < B, thenthespacespannedbyb is greater, and we can find a vector in B that is linearly independent of the space spanned by vectors in A. To stress this point, note that the above condition is A < B, not A B which is sufficient (to be able to find an independent vector) but not required. In other words, given A, B with rank(a) = A &rank(b) = B, then A < B,9an b 2 B such that rank(a [{b}) = A +1. Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F44/66 (pg.106/187)
109 Spanning trees/forests 1 1 o 2 We are given a graph G =(V,E), andconsidertheedgese = E(G) as an index set. Consider the V E incidence matrix of undirected graph G, which is the matrix X G =(x v,e ) v2v (G),e2E(G) where ( 1 if v 2 e x v,e = (5.92) 0 if v/2 e B C 7@ A (5.93) Prof. Jeff Bilmes EE563/Spring 2018/Submodularity - Lecture 5 - April 9th, 2018 F45/66 (pg.107/187)
Submodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture 4 http://www.ee.washington.edu/people/faculty/bilmes/classes/ee563_spring_2018/ Prof. Jeff Bilmes University
More informationSubmodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture http://www.ee.washington.edu/people/faculty/bilmes/classes/ee_spring_0/ Prof. Jeff Bilmes University of
More informationSubmodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture 3 http://j.ee.washington.edu/~bilmes/classes/ee596b_spring_2014/ Prof. Jeff Bilmes University of Washington,
More informationSubmodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture http://www.ee.washington.edu/people/faculty/bilmes/classes/eeb_spring_0/ Prof. Jeff Bilmes University of
More informationSubmodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture 12 http://www.ee.washington.edu/people/faculty/bilmes/classes/ee596b_spring_2016/ Prof. Jeff Bilmes University
More informationEE595A Submodular functions, their optimization and applications Spring 2011
EE595A Submodular functions, their optimization and applications Spring 2011 Prof. Jeff Bilmes University of Washington, Seattle Department of Electrical Engineering Spring Quarter, 2011 http://ssli.ee.washington.edu/~bilmes/ee595a_spring_2011/
More informationSubmodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture 17 http://www.ee.washington.edu/people/faculty/bilmes/classes/ee563_spring_2018/ Prof. Jeff Bilmes University
More informationEE595A Submodular functions, their optimization and applications Spring 2011
EE595A Submodular functions, their optimization and applications Spring 2011 Prof. Jeff Bilmes University of Washington, Seattle Department of Electrical Engineering Winter Quarter, 2011 http://ee.washington.edu/class/235/2011wtr/index.html
More informationSubmodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture 13 http://www.ee.washington.edu/people/faculty/bilmes/classes/ee563_spring_2018/ Prof. Jeff Bilmes University
More informationSubmodular Functions, Optimization, and Applications to Machine Learning
Submodular Functions, Optimization, and Applications to Machine Learning Spring Quarter, Lecture 14 http://www.ee.washington.edu/people/faculty/bilmes/classes/ee596b_spring_2016/ Prof. Jeff Bilmes University
More informationCS599: Convex and Combinatorial Optimization Fall 2013 Lecture 24: Introduction to Submodular Functions. Instructor: Shaddin Dughmi
CS599: Convex and Combinatorial Optimization Fall 2013 Lecture 24: Introduction to Submodular Functions Instructor: Shaddin Dughmi Announcements Introduction We saw how matroids form a class of feasible
More informationLearning with Submodular Functions: A Convex Optimization Perspective
Foundations and Trends R in Machine Learning Vol. 6, No. 2-3 (2013) 145 373 c 2013 F. Bach DOI: 10.1561/2200000039 Learning with Submodular Functions: A Convex Optimization Perspective Francis Bach INRIA
More information1 Submodular functions
CS 369P: Polyhedral techniques in combinatorial optimization Instructor: Jan Vondrák Lecture date: November 16, 2010 1 Submodular functions We have already encountered submodular functions. Let s recall
More informationApproximating Submodular Functions. Nick Harvey University of British Columbia
Approximating Submodular Functions Nick Harvey University of British Columbia Approximating Submodular Functions Part 1 Nick Harvey University of British Columbia Department of Computer Science July 11th,
More informationA combinatorial algorithm minimizing submodular functions in strongly polynomial time
A combinatorial algorithm minimizing submodular functions in strongly polynomial time Alexander Schrijver 1 Abstract We give a strongly polynomial-time algorithm minimizing a submodular function f given
More informationEE596A Submodularity Functions, Optimization, and Application to Machine Learning Fall Quarter 2012
EE596A Submodularity Functions, Optimization, and Application to Machine Learning Fall Quarter 2012 Prof. Jeff Bilmes University of Washington, Seattle Department of Electrical Engineering Spring Quarter,
More informationLearning symmetric non-monotone submodular functions
Learning symmetric non-monotone submodular functions Maria-Florina Balcan Georgia Institute of Technology ninamf@cc.gatech.edu Nicholas J. A. Harvey University of British Columbia nickhar@cs.ubc.ca Satoru
More informationAn Introduction to Submodular Functions and Optimization. Maurice Queyranne University of British Columbia, and IMA Visitor (Fall 2002)
An Introduction to Submodular Functions and Optimization Maurice Queyranne University of British Columbia, and IMA Visitor (Fall 2002) IMA, November 4, 2002 1. Submodular set functions Generalization to
More informationApplications of Submodular Functions in Speech and NLP
Applications of Submodular Functions in Speech and NLP Jeff Bilmes Department of Electrical Engineering University of Washington, Seattle http://ssli.ee.washington.edu/~bilmes June 27, 2011 J. Bilmes Applications
More informationh(x) lim H(x) = lim Since h is nondecreasing then h(x) 0 for all x, and if h is discontinuous at a point x then H(x) > 0. Denote
Real Variables, Fall 4 Problem set 4 Solution suggestions Exercise. Let f be of bounded variation on [a, b]. Show that for each c (a, b), lim x c f(x) and lim x c f(x) exist. Prove that a monotone function
More information1 Continuous extensions of submodular functions
CS 369P: Polyhedral techniques in combinatorial optimization Instructor: Jan Vondrák Lecture date: November 18, 010 1 Continuous extensions of submodular functions Submodular functions are functions assigning
More informationDiscrete Optimization in Machine Learning. Colorado Reed
Discrete Optimization in Machine Learning Colorado Reed [ML-RCC] 31 Jan 2013 1 Acknowledgements Some slides/animations based on: Krause et al. tutorials: http://www.submodularity.org Pushmeet Kohli tutorial:
More informationEE512 Graphical Models Fall 2009
EE512 Graphical Models Fall 2009 Prof. Jeff Bilmes University of Washington, Seattle Department of Electrical Engineering Fall Quarter, 2009 http://ssli.ee.washington.edu/~bilmes/ee512fa09 Lecture 3 -
More informationOn Matroid and Polymatroid Connectivity
Louisiana State University LSU Digital Commons LSU Doctoral Dissertations Graduate School 2014 On Matroid and Polymatroid Connectivity Dennis Wayne Hall II Louisiana State University and Agricultural and
More informationSubmodular Functions Properties Algorithms Machine Learning
Submodular Functions Properties Algorithms Machine Learning Rémi Gilleron Inria Lille - Nord Europe & LIFL & Univ Lille Jan. 12 revised Aug. 14 Rémi Gilleron (Mostrare) Submodular Functions Jan. 12 revised
More informationA submodular-supermodular procedure with applications to discriminative structure learning
A submodular-supermodular procedure with applications to discriminative structure learning Mukund Narasimhan Department of Electrical Engineering University of Washington Seattle WA 98004 mukundn@ee.washington.edu
More information1 Matroid intersection
CS 369P: Polyhedral techniques in combinatorial optimization Instructor: Jan Vondrák Lecture date: October 21st, 2010 Scribe: Bernd Bandemer 1 Matroid intersection Given two matroids M 1 = (E, I 1 ) and
More information1. Bounded linear maps. A linear map T : E F of real Banach
DIFFERENTIABLE MAPS 1. Bounded linear maps. A linear map T : E F of real Banach spaces E, F is bounded if M > 0 so that for all v E: T v M v. If v r T v C for some positive constants r, C, then T is bounded:
More informationMath 320: Real Analysis MWF 1pm, Campion Hall 302 Homework 7 Solutions Please write neatly, and in complete sentences when possible.
Math 320: Real Analysis MWF 1pm, Campion Hall 302 Homework 7 Solutions Please write neatly, and in complete sentences when possible. Do the following problems from the book: 4.2.1, 4.2.3, 4.2.6, 4.2.8,
More informationRealization of set functions as cut functions of graphs and hypergraphs
Discrete Mathematics 226 (2001) 199 210 www.elsevier.com/locate/disc Realization of set functions as cut functions of graphs and hypergraphs Satoru Fujishige a;, Sachin B. Patkar b a Division of Systems
More informationOn Submodular and Supermodular Functions on Lattices and Related Structures
On Submodular and Supermodular Functions on Lattices and Related Structures Dan A. Simovici University of Massachusetts Boston Department of Computer Science Boston, USA dsim@cs.umb.edu Abstract We give
More informationMODULE 8 Topics: Null space, range, column space, row space and rank of a matrix
MODULE 8 Topics: Null space, range, column space, row space and rank of a matrix Definition: Let L : V 1 V 2 be a linear operator. The null space N (L) of L is the subspace of V 1 defined by N (L) = {x
More information6.046 Recitation 11 Handout
6.046 Recitation 11 Handout May 2, 2008 1 Max Flow as a Linear Program As a reminder, a linear program is a problem that can be written as that of fulfilling an objective function and a set of constraints
More informationDRAFT - Math 101 Lecture Note - Dr. Said Algarni
2 Limits 2.1 The Tangent Problems The word tangent is derived from the Latin word tangens, which means touching. A tangent line to a curve is a line that touches the curve and a secant line is a line that
More informationEconomics 204 Fall 2011 Problem Set 2 Suggested Solutions
Economics 24 Fall 211 Problem Set 2 Suggested Solutions 1. Determine whether the following sets are open, closed, both or neither under the topology induced by the usual metric. (Hint: think about limit
More informationMaximization of Submodular Set Functions
Northeastern University Department of Electrical and Computer Engineering Maximization of Submodular Set Functions Biomedical Signal Processing, Imaging, Reasoning, and Learning BSPIRAL) Group Author:
More informationThe Simplex Method for Some Special Problems
The Simplex Method for Some Special Problems S. Zhang Department of Econometrics University of Groningen P.O. Box 800 9700 AV Groningen The Netherlands September 4, 1991 Abstract In this paper we discuss
More informationFactors in Graphs With Multiple Degree Constraints
Factors in Graphs With Multiple Degree Constraints Richard C. Brewster, Morten Hegner Nielsen, and Sean McGuinness Thompson Rivers University Kamloops, BC V2C0C8 Canada October 4, 2011 Abstract For a graph
More informationMATHEMATICAL ENGINEERING TECHNICAL REPORTS. On the Pipage Rounding Algorithm for Submodular Function Maximization A View from Discrete Convex Analysis
MATHEMATICAL ENGINEERING TECHNICAL REPORTS On the Pipage Rounding Algorithm for Submodular Function Maximization A View from Discrete Convex Analysis Akiyoshi SHIOURA (Communicated by Kazuo MUROTA) METR
More informationMean Value Theorem. Increasing Functions Extreme Values of Functions Rolle s Theorem Mean Value Theorem FAQ. Index
Mean Value Increasing Functions Extreme Values of Functions Rolle s Mean Value Increasing Functions (1) Assume that the function f is everywhere increasing and differentiable. ( x + h) f( x) f Then h 0
More informationLecture 10 (Submodular function)
Discrete Methods in Informatics January 16, 2006 Lecture 10 (Submodular function) Lecturer: Satoru Iwata Scribe: Masaru Iwasa and Yukie Nagai Submodular functions are the functions that frequently appear
More information9. Submodular function optimization
Submodular function maximization 9-9. Submodular function optimization Submodular function maximization Greedy algorithm for monotone case Influence maximization Greedy algorithm for non-monotone case
More informationCSE541 Class 22. Jeremy Buhler. November 22, Today: how to generalize some well-known approximation results
CSE541 Class 22 Jeremy Buhler November 22, 2016 Today: how to generalize some well-known approximation results 1 Intuition: Behavior of Functions Consider a real-valued function gz) on integers or reals).
More informationAlgebra II. Paulius Drungilas and Jonas Jankauskas
Algebra II Paulius Drungilas and Jonas Jankauskas Contents 1. Quadratic forms 3 What is quadratic form? 3 Change of variables. 3 Equivalence of quadratic forms. 4 Canonical form. 4 Normal form. 7 Positive
More information116 Problems in Algebra
116 Problems in Algebra Problems Proposer: Mohammad Jafari November 5, 011 116 Problems in Algebra is a nice work of Mohammad Jafari. Tese problems have been published in a book, but it is in Persian (Farsi).
More information1 Overview. 2 Multilinear Extension. AM 221: Advanced Optimization Spring 2016
AM 22: Advanced Optimization Spring 26 Prof. Yaron Singer Lecture 24 April 25th Overview The goal of today s lecture is to see how the multilinear extension of a submodular function that we introduced
More informationStatic Problem Set 2 Solutions
Static Problem Set Solutions Jonathan Kreamer July, 0 Question (i) Let g, h be two concave functions. Is f = g + h a concave function? Prove it. Yes. Proof: Consider any two points x, x and α [0, ]. Let
More informationSubmodularity, matroids, and the common colouring problem
Submodularity, matroids, and the common colouring problem Richard Santiago Torres Master of Arts Department of Mathematics and Statistics McGill University Montreal, Quebec August 13, 2015 A thesis submitted
More informationM-convex Functions on Jump Systems A Survey
Discrete Convexity and Optimization, RIMS, October 16, 2012 M-convex Functions on Jump Systems A Survey Kazuo Murota (U. Tokyo) 121016rimsJUMP 1 Discrete Convex Analysis Convexity Paradigm in Discrete
More informationContinuity. Chapter 4
Chapter 4 Continuity Throughout this chapter D is a nonempty subset of the real numbers. We recall the definition of a function. Definition 4.1. A function from D into R, denoted f : D R, is a subset of
More informationOptimizing Ratio of Monotone Set Functions
Optimizing Ratio of Monotone Set Functions Chao Qian, Jing-Cheng Shi 2, Yang Yu 2, Ke Tang, Zhi-Hua Zhou 2 UBRI, School of Computer Science and Technology, University of Science and Technology of China,
More informationWeek 5: Functions and graphs
Calculus and Linear Algebra for Biomedical Engineering Week 5: Functions and graphs H. Führ, Lehrstuhl A für Mathematik, RWTH Aachen, WS 07 Motivation: Measurements at fixed intervals 1 Consider a sequence
More informationMinimization and Maximization Algorithms in Discrete Convex Analysis
Modern Aspects of Submodularity, Georgia Tech, March 19, 2012 Minimization and Maximization Algorithms in Discrete Convex Analysis Kazuo Murota (U. Tokyo) 120319atlanta2Algorithm 1 Contents B1. Submodularity
More informationRevisiting the Greedy Approach to Submodular Set Function Maximization
Submitted to manuscript Revisiting the Greedy Approach to Submodular Set Function Maximization Pranava R. Goundan Analytics Operations Engineering, pranava@alum.mit.edu Andreas S. Schulz MIT Sloan School
More informationDarboux functions. Beniamin Bogoşel
Darboux functions Beniamin Bogoşel Abstract The article presents the construction of some real functions with surprising properties, including a discontinuous solution for the Cauchy functional equation
More informationMinimization of Symmetric Submodular Functions under Hereditary Constraints
Minimization of Symmetric Submodular Functions under Hereditary Constraints J.A. Soto (joint work with M. Goemans) DIM, Univ. de Chile April 4th, 2012 1 of 21 Outline Background Minimal Minimizers and
More informationConvex Optimization Notes
Convex Optimization Notes Jonathan Siegel January 2017 1 Convex Analysis This section is devoted to the study of convex functions f : B R {+ } and convex sets U B, for B a Banach space. The case of B =
More informationDefinition: For nonempty sets X and Y, a function, f, from X to Y is a relation that associates with each element of X exactly one element of Y.
Functions Definition: A relation is a correspondence between two sets. If x and y are two elements in these sets and if a relation exists between x and y, then we say that x corresponds to y or that y
More informationInverse Function Theorem writeup for MATH 4604 (Advanced Calculus II) Spring 2015
Inverse Function Theorem writeup for MATH 4604 (Advanced Calculus II) Spring 2015 Definition: N := {1, 2, 3,...}. Definition: For all n N, let i n : R n R n denote the identity map, defined by i n (x)
More informationSuper edge-magic labeling of graphs: deficiency and maximality
Electronic Journal of Graph Theory and Applications 5 () (017), 1 0 Super edge-magic labeling of graphs: deficiency and maximality Anak Agung Gede Ngurah a, Rinovia Simanjuntak b a Department of Civil
More information1 Maximizing a Submodular Function
6.883 Learning with Combinatorial Structure Notes for Lecture 16 Author: Arpit Agarwal 1 Maximizing a Submodular Function In the last lecture we looked at maximization of a monotone submodular function,
More informationWord Alignment via Submodular Maximization over Matroids
Word Alignment via Submodular Maximization over Matroids Hui Lin, Jeff Bilmes University of Washington, Seattle Dept. of Electrical Engineering June 21, 2011 Lin and Bilmes Submodular Word Alignment June
More informationDifferentiation. f(x + h) f(x) Lh = L.
Analysis in R n Math 204, Section 30 Winter Quarter 2008 Paul Sally, e-mail: sally@math.uchicago.edu John Boller, e-mail: boller@math.uchicago.edu website: http://www.math.uchicago.edu/ boller/m203 Differentiation
More informationThe Steiner Network Problem
The Steiner Network Problem Pekka Orponen T-79.7001 Postgraduate Course on Theoretical Computer Science 7.4.2008 Outline 1. The Steiner Network Problem Linear programming formulation LP relaxation 2. The
More informationSubmodular Function Minimization Based on Chapter 7 of the Handbook on Discrete Optimization [61]
Submodular Function Minimization Based on Chapter 7 of the Handbook on Discrete Optimization [61] Version 4 S. Thomas McCormick June 19, 2013 Abstract This survey describes the submodular function minimization
More informationMonotone Submodular Maximization over a Matroid
Monotone Submodular Maximization over a Matroid Yuval Filmus January 31, 2013 Abstract In this talk, we survey some recent results on monotone submodular maximization over a matroid. The survey does not
More informationOptimization of Submodular Functions Tutorial - lecture I
Optimization of Submodular Functions Tutorial - lecture I Jan Vondrák 1 1 IBM Almaden Research Center San Jose, CA Jan Vondrák (IBM Almaden) Submodular Optimization Tutorial 1 / 1 Lecture I: outline 1
More informationSubmodular Functions: Extensions, Distributions, and Algorithms A Survey
Submodular Functions: Extensions, Distributions, and Algorithms A Survey Shaddin Dughmi PhD Qualifying Exam Report, Department of Computer Science, Stanford University Exam Committee: Serge Plotkin, Tim
More information2.3 The Chain Rule and Inverse Functions
2.3 The Chain Rule and Inverse Functions 2.3. The Chain Rule and its Applications Theorem 2.3. (Chain Rule). If f(x) and g(x) are differentiable, then the derivative of the composite function (f g)(x)
More informationPacking Arborescences
Egerváry Research Group on Combinatorial Optimization Technical reports TR-2009-04. Published by the Egerváry Research Group, Pázmány P. sétány 1/C, H1117, Budapest, Hungary. Web site: www.cs.elte.hu/egres.
More informationA Distributed Newton Method for Network Utility Maximization, II: Convergence
A Distributed Newton Method for Network Utility Maximization, II: Convergence Ermin Wei, Asuman Ozdaglar, and Ali Jadbabaie October 31, 2012 Abstract The existing distributed algorithms for Network Utility
More informationStrictly Positive Definite Functions on a Real Inner Product Space
Strictly Positive Definite Functions on a Real Inner Product Space Allan Pinkus Abstract. If ft) = a kt k converges for all t IR with all coefficients a k 0, then the function f< x, y >) is positive definite
More informationa i,1 a i,j a i,m be vectors with positive entries. The linear programming problem associated to A, b and c is to find all vectors sa b
LINEAR PROGRAMMING PROBLEMS MATH 210 NOTES 1. Statement of linear programming problems Suppose n, m 1 are integers and that A = a 1,1... a 1,m a i,1 a i,j a i,m a n,1... a n,m is an n m matrix of positive
More informationMatroids and submodular optimization
Matroids and submodular optimization Attila Bernáth Research fellow, Institute of Informatics, Warsaw University 23 May 2012 Attila Bernáth () Matroids and submodular optimization 23 May 2012 1 / 10 About
More informationAdvanced Combinatorial Optimization Updated April 29, Lecture 20. Lecturer: Michel X. Goemans Scribe: Claudio Telha (Nov 17, 2009)
18.438 Advanced Combinatorial Optimization Updated April 29, 2012 Lecture 20 Lecturer: Michel X. Goemans Scribe: Claudio Telha (Nov 17, 2009) Given a finite set V with n elements, a function f : 2 V Z
More informationAlgebraic Geometry Spring 2009
MIT OpenCourseWare http://ocw.mit.edu 18.726 Algebraic Geometry Spring 2009 For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms. 18.726: Algebraic Geometry
More informationPart II: Integral Splittable Congestion Games. Existence and Computation of Equilibria Integral Polymatroids
Kombinatorische Matroids and Polymatroids Optimierung in der inlogistik Congestion undgames im Verkehr Tobias Harks Augsburg University WINE Tutorial, 8.12.2015 Outline Part I: Congestion Games Existence
More informationDiscrete DC Programming by Discrete Convex Analysis
Combinatorial Optimization (Oberwolfach, November 9 15, 2014) Discrete DC Programming by Discrete Convex Analysis Use of Conjugacy Kazuo Murota (U. Tokyo) with Takanori Maehara (NII, JST) 141111DCprogOberwolfLong
More information1/12/05: sec 3.1 and my article: How good is the Lebesgue measure?, Math. Intelligencer 11(2) (1989),
Real Analysis 2, Math 651, Spring 2005 April 26, 2005 1 Real Analysis 2, Math 651, Spring 2005 Krzysztof Chris Ciesielski 1/12/05: sec 3.1 and my article: How good is the Lebesgue measure?, Math. Intelligencer
More informationCalculus. Weijiu Liu. Department of Mathematics University of Central Arkansas 201 Donaghey Avenue, Conway, AR 72035, USA
Calculus Weijiu Liu Department of Mathematics University of Central Arkansas 201 Donaghey Avenue, Conway, AR 72035, USA 1 Opening Welcome to your Calculus I class! My name is Weijiu Liu. I will guide you
More information2.2. Show that U 0 is a vector space. For each α 0 in F, show by example that U α does not satisfy closure.
Hints for Exercises 1.3. This diagram says that f α = β g. I will prove f injective g injective. You should show g injective f injective. Assume f is injective. Now suppose g(x) = g(y) for some x, y A.
More informationEconomics 204 Summer/Fall 2011 Lecture 5 Friday July 29, 2011
Economics 204 Summer/Fall 2011 Lecture 5 Friday July 29, 2011 Section 2.6 (cont.) Properties of Real Functions Here we first study properties of functions from R to R, making use of the additional structure
More informationImplicit Functions, Curves and Surfaces
Chapter 11 Implicit Functions, Curves and Surfaces 11.1 Implicit Function Theorem Motivation. In many problems, objects or quantities of interest can only be described indirectly or implicitly. It is then
More informationMATHEMATICAL PROGRAMMING I
MATHEMATICAL PROGRAMMING I Books There is no single course text, but there are many useful books, some more mathematical, others written at a more applied level. A selection is as follows: Bazaraa, Jarvis
More informationTheorem 5.3. Let E/F, E = F (u), be a simple field extension. Then u is algebraic if and only if E/F is finite. In this case, [E : F ] = deg f u.
5. Fields 5.1. Field extensions. Let F E be a subfield of the field E. We also describe this situation by saying that E is an extension field of F, and we write E/F to express this fact. If E/F is a field
More informationInstructions. 2. Four possible answers are provided for each question and only one of these is correct.
Instructions 1. This question paper has forty multiple choice questions. 2. Four possible answers are provided for each question and only one of these is correct. 3. Marking scheme: Each correct answer
More informationFinal Review Sheet. B = (1, 1 + 3x, 1 + x 2 ) then 2 + 3x + 6x 2
Final Review Sheet The final will cover Sections Chapters 1,2,3 and 4, as well as sections 5.1-5.4, 6.1-6.2 and 7.1-7.3 from chapters 5,6 and 7. This is essentially all material covered this term. Watch
More informationLearning Submodular Functions
Learning Submodular Functions Maria-Florina Balcan Nicholas J. A. Harvey Abstract This paper considers the problem of learning submodular functions. A problem instance consists of a distribution on {0,
More informationM2PM1 Analysis II (2008) Dr M Ruzhansky List of definitions, statements and examples Preliminary version
M2PM1 Analysis II (2008) Dr M Ruzhansky List of definitions, statements and examples Preliminary version Chapter 0: Some revision of M1P1: Limits and continuity This chapter is mostly the revision of Chapter
More informationReview of Multi-Calculus (Study Guide for Spivak s CHAPTER ONE TO THREE)
Review of Multi-Calculus (Study Guide for Spivak s CHPTER ONE TO THREE) This material is for June 9 to 16 (Monday to Monday) Chapter I: Functions on R n Dot product and norm for vectors in R n : Let X
More informationContinuity. Chapter 4
Chapter 4 Continuity Throughout this chapter D is a nonempty subset of the real numbers. We recall the definition of a function. Definition 4.1. A function from D into R, denoted f : D R, is a subset of
More informationLecture 5: Functions : Images, Compositions, Inverses
Lecture 5: Functions : Images, Compositions, Inverses 1 Functions We have all seen some form of functions in high school. For example, we have seen polynomial, exponential, logarithmic, trigonometric functions
More informationA Min-Max Theorem for k-submodular Functions and Extreme Points of the Associated Polyhedra. Satoru FUJISHIGE and Shin-ichi TANIGAWA.
RIMS-1787 A Min-Max Theorem for k-submodular Functions and Extreme Points of the Associated Polyhedra By Satoru FUJISHIGE and Shin-ichi TANIGAWA September 2013 RESEARCH INSTITUTE FOR MATHEMATICAL SCIENCES
More informationLecture 2: Review of Prerequisites. Table of contents
Math 348 Fall 217 Lecture 2: Review of Prerequisites Disclaimer. As we have a textbook, this lecture note is for guidance and supplement only. It should not be relied on when preparing for exams. In this
More informationVector Spaces. Addition : R n R n R n Scalar multiplication : R R n R n.
Vector Spaces Definition: The usual addition and scalar multiplication of n-tuples x = (x 1,..., x n ) R n (also called vectors) are the addition and scalar multiplication operations defined component-wise:
More informationSuppose R is an ordered ring with positive elements P.
1. The real numbers. 1.1. Ordered rings. Definition 1.1. By an ordered commutative ring with unity we mean an ordered sextuple (R, +, 0,, 1, P ) such that (R, +, 0,, 1) is a commutative ring with unity
More information1.3 Limits and Continuity
.3 Limits and Continuity.3. Limits Problem 8. What will happen to the functional values of as x gets closer and closer to 2? f(x) = Solution. We can evaluate f(x) using x values nearer and nearer to 2
More informationLecture 2: Linear Algebra Review
EE 227A: Convex Optimization and Applications January 19 Lecture 2: Linear Algebra Review Lecturer: Mert Pilanci Reading assignment: Appendix C of BV. Sections 2-6 of the web textbook 1 2.1 Vectors 2.1.1
More informationImmerse Metric Space Homework
Immerse Metric Space Homework (Exercises -2). In R n, define d(x, y) = x y +... + x n y n. Show that d is a metric that induces the usual topology. Sketch the basis elements when n = 2. Solution: Steps
More information