ComposableCore-sets for Determinant Maximization: A...
Transcript of ComposableCore-sets for Determinant Maximization: A...
![Page 1: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/1.jpg)
Composable Core-sets for Determinant Maximization:
A Simple Near-Optimal Algorithm
Sepideh Mahabadi
TTIC
Piotr IndykMIT
Shayan Oveis GharanU. of Washington
Alireza RezaeiU. of Washington
![Page 2: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/2.jpg)
Volume (Determinant) Maximization Problem
Input: a set of 𝑛𝑛 vectors V ∈ ℝ𝑑𝑑 and a parameter 𝑘𝑘 ≤ 𝑑𝑑,
𝒌𝒌 = 𝟐𝟐
![Page 3: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/3.jpg)
Volume (Determinant) Maximization Problem
Input: a set of 𝑛𝑛 vectors V ∈ ℝ𝑑𝑑 and a parameter 𝑘𝑘 ≤ 𝑑𝑑,Output: a subset 𝑆𝑆 ⊂ 𝑉𝑉 of size 𝑘𝑘 with the maximum volume
• Parallelepiped spanned by the points in 𝑆𝑆
𝒌𝒌 = 𝟐𝟐
![Page 4: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/4.jpg)
Volume (Determinant) Maximization Problem
Input: a set of 𝑛𝑛 vectors V ∈ ℝ𝑑𝑑 and a parameter 𝑘𝑘 ≤ 𝑑𝑑,Output: a subset 𝑆𝑆 ⊂ 𝑉𝑉 of size 𝑘𝑘 with the maximum volume
• Parallelepiped spanned by the points in 𝑆𝑆
𝒌𝒌 = 𝟐𝟐
![Page 5: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/5.jpg)
Determinant (Volume) Maximization Problem
Input: a set of 𝑛𝑛 vectors V ∈ ℝ𝑑𝑑 and a parameter 𝑘𝑘 ≤ 𝑑𝑑,Output: a subset 𝑆𝑆 ⊂ 𝑉𝑉 of size 𝑘𝑘 with the maximum volume
• Parallelepiped spanned by the points in 𝑆𝑆
Equivalent Formulation: Reuse 𝑉𝑉 to denote the matrix where its columns are the vectors in 𝑉𝑉
𝑣𝑣1 𝑣𝑣2 …𝑣𝑣𝑛𝑛
![Page 6: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/6.jpg)
Determinant (Volume) Maximization Problem
Input: a set of 𝑛𝑛 vectors V ∈ ℝ𝑑𝑑 and a parameter 𝑘𝑘 ≤ 𝑑𝑑,Output: a subset 𝑆𝑆 ⊂ 𝑉𝑉 of size 𝑘𝑘 with the maximum volume
• Parallelepiped spanned by the points in 𝑆𝑆
Equivalent Formulation: Reuse 𝑉𝑉 to denote the matrix where its columns are the vectors in 𝑉𝑉• Let 𝑀𝑀 be the gram matrix 𝑉𝑉𝑇𝑇𝑉𝑉 𝑀𝑀𝑖𝑖,𝑗𝑗 = 𝑣𝑣𝑖𝑖 · 𝑣𝑣𝑗𝑗
𝑣𝑣1𝑣𝑣2…𝑣𝑣𝑛𝑛
𝑣𝑣1 𝑣𝑣2 …𝑣𝑣𝑛𝑛 𝑣𝑣𝑖𝑖 ⋅ 𝑣𝑣𝑗𝑗𝑖𝑖
𝑗𝑗
![Page 7: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/7.jpg)
Determinant (Volume) Maximization Problem
Input: a set of 𝑛𝑛 vectors V ∈ ℝ𝑑𝑑 and a parameter 𝑘𝑘 ≤ 𝑑𝑑,Output: a subset 𝑆𝑆 ⊂ 𝑉𝑉 of size 𝑘𝑘 with the maximum volume
• Parallelepiped spanned by the points in 𝑆𝑆
Equivalent Formulation: Reuse 𝑉𝑉 to denote the matrix where its columns are the vectors in 𝑉𝑉• Let 𝑀𝑀 be the gram matrix 𝑉𝑉𝑇𝑇𝑉𝑉• Choose 𝑆𝑆 such that det(𝑀𝑀𝑆𝑆,𝑆𝑆) is maximized
𝑀𝑀𝑖𝑖,𝑗𝑗 = 𝑣𝑣𝑖𝑖 · 𝑣𝑣𝑗𝑗
det 𝑀𝑀𝑆𝑆,𝑆𝑆 = 𝑉𝑉𝑉𝑉𝑉𝑉 𝑆𝑆 2
𝑣𝑣1𝑣𝑣2…𝑣𝑣𝑛𝑛
𝑣𝑣1 𝑣𝑣2 …𝑣𝑣𝑛𝑛
![Page 8: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/8.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]
![Page 9: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/9.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]• Best algorithm: 𝑒𝑒𝑐𝑐-approximation [Nik’15]
![Page 10: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/10.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]• Best algorithm: 𝑒𝑒𝑐𝑐-approximation [Nik’15]• Greedy is a popular algorithm: achieves approximation factor 𝑘𝑘! 𝑈𝑈 ← ∅ For 𝑘𝑘 iterations, Add to 𝑈𝑈 the farthest point from the subspace spanned by 𝑈𝑈
𝒌𝒌 = 𝟐𝟐
![Page 11: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/11.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]• Best algorithm: 𝑒𝑒𝑐𝑐-approximation [Nik’15]• Greedy is a popular algorithm: achieves approximation factor 𝑘𝑘! 𝑈𝑈 ← ∅ For 𝑘𝑘 iterations, Add to 𝑈𝑈 the farthest point from the subspace spanned by 𝑈𝑈
𝒌𝒌 = 𝟐𝟐
![Page 12: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/12.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]• Best algorithm: 𝑒𝑒𝑐𝑐-approximation [Nik’15]• Greedy is a popular algorithm: achieves approximation factor 𝑘𝑘! 𝑈𝑈 ← ∅ For 𝑘𝑘 iterations, Add to 𝑈𝑈 the farthest point from the subspace spanned by 𝑈𝑈
𝒌𝒌 = 𝟐𝟐
![Page 13: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/13.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]• Best algorithm: 𝑒𝑒𝑐𝑐-approximation [Nik’15]• Greedy is a popular algorithm: achieves approximation factor 𝑘𝑘! 𝑈𝑈 ← ∅ For 𝑘𝑘 iterations, Add to 𝑈𝑈 the farthest point from the subspace spanned by 𝑈𝑈
𝒌𝒌 = 𝟐𝟐
![Page 14: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/14.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]• Best algorithm: 𝑒𝑒𝑐𝑐-approximation [Nik’15]• Greedy is a popular algorithm: achieves approximation factor 𝑘𝑘! 𝑈𝑈 ← ∅ For 𝑘𝑘 iterations, Add to 𝑈𝑈 the farthest point from the subspace spanned by 𝑈𝑈
𝒌𝒌 = 𝟐𝟐
![Page 15: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/15.jpg)
What is known?• Hard to approximate within a factor of 2𝑐𝑐𝑐𝑐 [CMI’13]• Best algorithm: 𝑒𝑒𝑐𝑐-approximation [Nik’15]• Greedy is a popular algorithm: achieves approximation factor 𝑘𝑘! 𝑈𝑈 ← ∅ For 𝑘𝑘 iterations, Add to 𝑈𝑈 the farthest point from the subspace spanned by 𝑈𝑈
• Greedy performs very well in practice
𝒌𝒌 = 𝟐𝟐
![Page 16: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/16.jpg)
Determinantal Point Processes (DPP)DPP: Very popular probabilistic model, where given a set of vectors 𝑉𝑉, samples any 𝑘𝑘-subset 𝑆𝑆 with
probability proportional to this determinant.
• Maximum a posteriori (MAP) decoding is determinant maximization
• Volume/determinant is a notion of diversity
• NeurIPS’18 Tutorial, Negative Dependence, Stable Polynomials, and All That, Jegelka, Sra• ICML’19 Workshop, Negative Dependence: Theory and Applications in Machine Learning, Gartrell,
Gillenwater, Kulesza, Mariet
![Page 17: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/17.jpg)
Application: Diversity Maximization
Given a set of objects, how to pick a few of them while maximizing diversity?
![Page 18: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/18.jpg)
Application: Diversity Maximization
• Searching
Given a set of objects, how to pick a few of them while maximizing diversity?
![Page 19: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/19.jpg)
Application: Diversity Maximization
• Searching
Given a set of objects, how to pick a few of them while maximizing diversity?
![Page 20: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/20.jpg)
Application: Diversity Maximization
Points in a high dimensional space
Objects (documents, images, etc)
Feature Vectors
![Page 21: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/21.jpg)
Application: Diversity Maximization
Input: a set of 𝑛𝑛 vectors V ⊂ ℝ𝑑𝑑 and a parameter 𝑘𝑘,
Goal: pick 𝒌𝒌 points while maximizing “diversity”.
𝒌𝒌 = 𝟑𝟑
![Page 22: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/22.jpg)
Determinantal Point Processes (DPP)DPP: Very popular probabilistic model, where given a set of vectors 𝑉𝑉, samples any 𝑘𝑘-subset 𝑆𝑆 with
probability proportional to this determinant.
• Maximum a posteriori (MAP) decoding is determinant maximization
• Volume/determinant is a notion of diversity
[MJK’17,GCGS’14] Video summarization[KT+’12, CGGS’15,KT’11] Document summarization[YFZ+’16] Tweet generation[LCYO’16] Object detection….
Applications
![Page 23: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/23.jpg)
Determinantal Point Processes (DPP)
• Most applications deal with massive data• Lots of effort for solving the problem in massive
data models of computation [MJK’17, WIB’14, PJG+’14, MKSK’13, MKBK’15, MZ’15, MZ’15, BENW’15]
• e.g. streaming, distributed, parallel
DPP: Very popular probabilistic model, where given a set of vectors 𝑉𝑉, samples any 𝑘𝑘-subset 𝑆𝑆 with
probability proportional to this determinant.
• Maximum a posteriori (MAP) decoding is determinant maximization
• Volume/determinant is a notion of diversity
[MJK’17,GCGS’14] Video summarization[KT+’12, CGGS’15,KT’11] Document summarization[YFZ+’16] Tweet generation[LCYO’16] Object detection….
Applications
![Page 24: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/24.jpg)
Determinantal Point Processes (DPP)
• Most applications deal with massive data• Lots of effort for solving the problem in massive
data models of computation [MJK’17, WIB’14, PJG+’14, MKSK’13, MKBK’15, MZ’15, MZ’15, BENW’15]
• e.g. streaming, distributed, parallel
DPP: Very popular probabilistic model, where given a set of vectors 𝑉𝑉, samples any 𝑘𝑘-subset 𝑆𝑆 with
probability proportional to this determinant.
• Maximum a posteriori (MAP) decoding is determinant maximization
• Volume/determinant is a notion of diversity
[MJK’17,GCGS’14] Video summarization[KT+’12, CGGS’15,KT’11] Document summarization[YFZ+’16] Tweet generation[LCYO’16] Object detection….
Applications
Composable Core-sets
![Page 25: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/25.jpg)
Core-sets
Solving the problem over 𝑼𝑼 gives a good approximation of solving the problem over 𝑽𝑽
Core-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
![Page 26: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/26.jpg)
Composable Core-setsCore-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
![Page 27: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/27.jpg)
Composable Core-setsCore-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
• Let 𝒇𝒇 be an optimization function
o E.g. 𝒇𝒇(𝑽𝑽): solution to 𝑘𝑘 determinant maximization
![Page 28: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/28.jpg)
Composable Core-setsCore-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
• Let 𝒇𝒇 be an optimization function
o E.g. 𝒇𝒇(𝑽𝑽): solution to 𝑘𝑘 determinant maximization
• Multiple data sets 𝑽𝑽𝟏𝟏,⋯ ,𝑽𝑽𝒎𝒎 and their coresets 𝑼𝑼𝟏𝟏 ⊂ 𝑽𝑽𝟏𝟏,⋯ ,𝑼𝑼𝒎𝒎 ⊂ 𝑽𝑽𝒎𝒎,
o 𝒇𝒇 𝑼𝑼𝟏𝟏 ∪⋯∪ 𝑼𝑼𝒎𝒎 approximates 𝒇𝒇 𝑽𝑽𝟏𝟏 ∪⋯∪ 𝑽𝑽𝒎𝒎 by a factor 𝜶𝜶
![Page 29: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/29.jpg)
Composable Core-setsCore-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
• Let 𝒇𝒇 be an optimization function
o E.g. 𝒇𝒇(𝑽𝑽): solution to 𝑘𝑘 determinant maximization
• Multiple data sets 𝑽𝑽𝟏𝟏,⋯ ,𝑽𝑽𝒎𝒎 and their coresets 𝑼𝑼𝟏𝟏 ⊂ 𝑽𝑽𝟏𝟏,⋯ ,𝑼𝑼𝒎𝒎 ⊂ 𝑽𝑽𝒎𝒎,
o 𝒇𝒇 𝑼𝑼𝟏𝟏 ∪⋯∪ 𝑼𝑼𝒎𝒎 approximates 𝒇𝒇 𝑽𝑽𝟏𝟏 ∪⋯∪ 𝑽𝑽𝒎𝒎 by a factor 𝜶𝜶
![Page 30: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/30.jpg)
Composable Core-setsCore-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
• Let 𝒇𝒇 be an optimization function
o E.g. 𝒇𝒇(𝑽𝑽): solution to 𝑘𝑘 determinant maximization
• Multiple data sets 𝑽𝑽𝟏𝟏,⋯ ,𝑽𝑽𝒎𝒎 and their coresets 𝑼𝑼𝟏𝟏 ⊂ 𝑽𝑽𝟏𝟏,⋯ ,𝑼𝑼𝒎𝒎 ⊂ 𝑽𝑽𝒎𝒎,
o 𝒇𝒇 𝑼𝑼𝟏𝟏 ∪⋯∪ 𝑼𝑼𝒎𝒎 approximates 𝒇𝒇 𝑽𝑽𝟏𝟏 ∪⋯∪ 𝑽𝑽𝒎𝒎 by a factor 𝜶𝜶
![Page 31: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/31.jpg)
Composable Core-sets
𝒇𝒇
Core-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
• Let 𝒇𝒇 be an optimization function
o E.g. 𝒇𝒇(𝑽𝑽): solution to 𝑘𝑘 determinant maximization
• Multiple data sets 𝑽𝑽𝟏𝟏,⋯ ,𝑽𝑽𝒎𝒎 and their coresets 𝑼𝑼𝟏𝟏 ⊂ 𝑽𝑽𝟏𝟏,⋯ ,𝑼𝑼𝒎𝒎 ⊂ 𝑽𝑽𝒎𝒎,
o 𝒇𝒇 𝑼𝑼𝟏𝟏 ∪⋯∪ 𝑼𝑼𝒎𝒎 approximates 𝒇𝒇 𝑽𝑽𝟏𝟏 ∪⋯∪ 𝑽𝑽𝒎𝒎 by a factor 𝜶𝜶
![Page 32: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/32.jpg)
Composable Core-setsCore-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
𝒇𝒇
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
• Let 𝒇𝒇 be an optimization function
o E.g. 𝒇𝒇(𝑽𝑽): solution to 𝑘𝑘 determinant maximization
• Multiple data sets 𝑽𝑽𝟏𝟏,⋯ ,𝑽𝑽𝒎𝒎 and their coresets 𝑼𝑼𝟏𝟏 ⊂ 𝑽𝑽𝟏𝟏,⋯ ,𝑼𝑼𝒎𝒎 ⊂ 𝑽𝑽𝒎𝒎,
o 𝒇𝒇 𝑼𝑼𝟏𝟏 ∪⋯∪ 𝑼𝑼𝒎𝒎 approximates 𝒇𝒇 𝑽𝑽𝟏𝟏 ∪⋯∪ 𝑽𝑽𝒎𝒎 by a factor 𝜶𝜶
![Page 33: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/33.jpg)
Composable Core-setsCore-sets [AHV’05]: a subset 𝑼𝑼 of the data 𝑽𝑽 that represents it well
Composable Core-sets have been studied for the diversity Maximization problems, for other
notions of diversity: minimum pairwise distance, sum of pairwise distances, etc.
Determinant maximization is a “higher order” notion of diversity
Composable Core-sets [AAIMV’13 and IMMM’14]:
The union of coresets is a coreset for the union
• Let 𝒇𝒇 be an optimization function
o E.g. 𝒇𝒇(𝑽𝑽): solution to 𝑘𝑘 determinant maximization
• Multiple data sets 𝑽𝑽𝟏𝟏,⋯ ,𝑽𝑽𝒎𝒎 and their coresets 𝑼𝑼𝟏𝟏 ⊂ 𝑽𝑽𝟏𝟏,⋯ ,𝑼𝑼𝒎𝒎 ⊂ 𝑽𝑽𝒎𝒎,
o 𝒇𝒇 𝑼𝑼𝟏𝟏 ∪⋯∪ 𝑼𝑼𝒎𝒎 approximates 𝒇𝒇 𝑽𝑽𝟏𝟏 ∪⋯∪ 𝑽𝑽𝒎𝒎 by a factor 𝜶𝜶
![Page 34: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/34.jpg)
Applications: Streaming Computation• Streaming Computation:
• Processing sequence of 𝑛𝑛 data elements “on the fly”• limited Storage
![Page 35: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/35.jpg)
Applications: Streaming Computation• Streaming Computation:
• Processing sequence of 𝑛𝑛 data elements “on the fly”• limited Storage
• Composable Core-set• Divide into chunks
𝒏𝒏 𝒏𝒏
![Page 36: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/36.jpg)
Applications: Streaming Computation• Streaming Computation:
• Processing sequence of 𝑛𝑛 data elements “on the fly”• limited Storage
• Composable Core-set• Divide into chunks• Compute Core-set for each chunk as it arrives
𝒏𝒏 𝒏𝒏
Core-set
![Page 37: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/37.jpg)
Applications: Streaming Computation• Streaming Computation:
• Processing sequence of 𝑛𝑛 data elements “on the fly”• limited Storage
• Composable Core-set • Divide into chunks• Compute Core-set for each chunk as it arrives
𝒏𝒏 𝒏𝒏
Core-set Core-set
![Page 38: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/38.jpg)
Applications: Streaming Computation• Streaming Computation:
• Processing sequence of 𝑛𝑛 data elements “on the fly”• limited Storage
• Composable Core-set • Divide into chunks• Compute Core-set for each chunk as it arrives• Space goes down from 𝑛𝑛 to 𝑛𝑛
𝒏𝒏 𝒏𝒏
Core-set Core-set
![Page 39: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/39.jpg)
Applications: Distributed Computation• Streaming Computation• Distributed System:
• Each machine holds a block of data.• A composable core-set is computed and sent to the server
Core-set
Data
Data
Data
Mapper
Mapper
Mapper
Reducer Solution
![Page 40: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/40.jpg)
Applications: Improving Runtime• Streaming Computation• Distributed System• Similar framework for improving the runtime
![Page 41: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/41.jpg)
Can we get a composable core-set of small size for the determinant maximization problem?
![Page 42: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/42.jpg)
Composable Core-sets for Volume Maximization
LP-based Optimal Approximation Algorithm of [IMOR’18]:
There exists a polynomial time algorithm for computing an �𝑶𝑶 𝒌𝒌 𝒌𝒌/𝟐𝟐 -composable
core-set of size �𝑶𝑶(𝒌𝒌) for the volume maximization problem.
[IMOR’18]Approximation �𝑶𝑶 𝒌𝒌 𝒌𝒌/𝟐𝟐
Core-set Size �𝑶𝑶(𝒌𝒌)Simple? ×
![Page 43: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/43.jpg)
Composable Core-sets for Volume Maximization
Lower bound [IMOR’18]:
Any composable core-set of size 𝒌𝒌𝑶𝑶(𝟏𝟏) for the volume maximization problem must
have an approximation factor of 𝛀𝛀(𝒌𝒌)𝒌𝒌𝟐𝟐(𝟏𝟏−𝒐𝒐 𝟏𝟏 ).
Lower Bound [IMOR’18]Approximation 𝛀𝛀 𝒌𝒌
𝒌𝒌𝟐𝟐−𝒐𝒐 𝒌𝒌 �𝑶𝑶 𝒌𝒌
𝒌𝒌𝟐𝟐
Core-set Size 𝒌𝒌𝑶𝑶 𝟏𝟏 �𝑶𝑶(𝒌𝒌)Simple? ×
![Page 44: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/44.jpg)
Our Results
Lower Bound [IMOR’18] Greedy Approximation 𝛀𝛀 𝒌𝒌
𝒌𝒌𝟐𝟐−𝒐𝒐 𝒌𝒌 �𝑶𝑶 𝒌𝒌
𝒌𝒌𝟐𝟐 𝑶𝑶(𝑪𝑪𝒌𝒌𝟐𝟐)
Core-set Size 𝒌𝒌𝑶𝑶 𝟏𝟏 �𝑶𝑶(𝒌𝒌) 𝒌𝒌Simple? × ✔
The widely used Greedy algorithm produces a composable core-set of size 𝑘𝑘 with
approximation factor 𝑂𝑂(𝐶𝐶𝑐𝑐2).
![Page 45: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/45.jpg)
Our Results
The Local Search Algorithm produces a composable core-set of size 𝑘𝑘 with
approximation factor 𝑂𝑂 𝑘𝑘 2𝑐𝑐.
Lower Bound [IMOR’18] Greedy Local SearchApproximation 𝛀𝛀 𝒌𝒌
𝒌𝒌𝟐𝟐−𝒐𝒐 𝒌𝒌 �𝑶𝑶 𝒌𝒌
𝒌𝒌𝟐𝟐 𝑶𝑶(𝑪𝑪𝒌𝒌𝟐𝟐) 𝑶𝑶 𝒌𝒌𝒌𝒌
Core-set Size 𝒌𝒌𝑶𝑶 𝟏𝟏 �𝑶𝑶(𝒌𝒌) 𝒌𝒌 𝒌𝒌Simple? × ✔ ✔
![Page 46: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/46.jpg)
This Talk
The Local Search Algorithm produces a composable core-set of size 𝑘𝑘 with
approximation factor 𝑂𝑂 𝑘𝑘 𝑐𝑐 for the volume maximization problem.
![Page 47: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/47.jpg)
This Talk
The Local Search Algorithm produces a composable core-set of size 𝑘𝑘 with
approximation factor 𝑂𝑂 𝑘𝑘 𝑐𝑐 for the volume maximization problem.
In comparison to the optimal core-set algorithm Approximation O 𝑘𝑘 𝒌𝒌 as opposed to 𝑂𝑂 𝑘𝑘 log 𝑘𝑘 𝒌𝒌/𝟐𝟐
Smaller Size 𝑘𝑘 as opposed to 𝑂𝑂 𝑘𝑘 log 𝑘𝑘 Simpler to implement (similar to Greedy)
Better performance in practice
![Page 48: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/48.jpg)
The Local Search Algorithm
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
![Page 49: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/49.jpg)
The Local Search Algorithm
𝒌𝒌 = 𝟑𝟑
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
![Page 50: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/50.jpg)
The Local Search Algorithm
𝒌𝒌 = 𝟑𝟑
𝑞𝑞
𝑝𝑝
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
![Page 51: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/51.jpg)
The Local Search Algorithm
𝒌𝒌 = 𝟑𝟑
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
![Page 52: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/52.jpg)
The Local Search Algorithm
𝒌𝒌 = 𝟑𝟑
𝑞𝑞
𝑝𝑝
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
![Page 53: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/53.jpg)
The Local Search Algorithm
𝒌𝒌 = 𝟑𝟑
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
![Page 54: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/54.jpg)
To bound the run time
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
Start with a crude approximation (Greedy algorithm)
If it increases by at least a factor of (1 + 𝜖𝜖)
![Page 55: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/55.jpg)
Checking the condition
𝑞𝑞𝑝𝑝
Input: a set 𝑉𝑉 of 𝑛𝑛 points and a parameter 𝑘𝑘
1. Start with an arbitrary subset of 𝑘𝑘 points 𝑆𝑆 ⊆ 𝑉𝑉
2. While there exists a point 𝑝𝑝 ∈ 𝑉𝑉 ∖ 𝑆𝑆 and 𝑞𝑞 ∈ 𝑆𝑆 s.t. replacing 𝑞𝑞 with 𝑝𝑝 increases the volume, then swap them, i.e., 𝑆𝑆 = 𝑆𝑆 ∪ 𝑝𝑝 ∖ {𝑞𝑞}
𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑,𝑯𝑯𝑺𝑺∖ 𝒒𝒒 > 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅(𝒒𝒒,𝑯𝑯𝑺𝑺∖ 𝒒𝒒 )
(𝒌𝒌 − 𝟏𝟏)-dimensional Subspace
![Page 56: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/56.jpg)
Main Lemma [informal]: Local Search preserves maximum distance to “all” subspaces of dimension 𝒌𝒌 − 𝟏𝟏
![Page 57: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/57.jpg)
𝑽𝑽 is the point set𝑺𝑺 = 𝐿𝐿𝑆𝑆 𝑉𝑉 is the core-set produced by local search
Main Lemma [informal]: Local Search preserves maximum distance to “all” subspaces of dimension 𝒌𝒌 − 𝟏𝟏
![Page 58: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/58.jpg)
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑞𝑞∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
𝑽𝑽 is the point set𝑺𝑺 = 𝐿𝐿𝑆𝑆 𝑉𝑉 is the core-set produced by local search
Main Lemma [informal]: Local Search preserves maximum distance to “all” subspaces of dimension 𝒌𝒌 − 𝟏𝟏
![Page 59: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/59.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
𝑝𝑝
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 60: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/60.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
𝑝𝑝
𝑮𝑮
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 61: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/61.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
• Assume for any 𝑞𝑞 ∈ 𝑆𝑆, 𝑑𝑑 𝑞𝑞,𝐺𝐺 ≤ 𝑥𝑥
𝑝𝑝
𝑮𝑮
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 62: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/62.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
• Assume for any 𝑞𝑞 ∈ 𝑆𝑆, 𝑑𝑑 𝑞𝑞,𝐺𝐺 ≤ 𝑥𝑥
𝑝𝑝
𝑮𝑮
≤ 𝑥𝑥≤ 𝑥𝑥
≤ 𝑥𝑥
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 63: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/63.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
• Assume for any 𝑞𝑞 ∈ 𝑆𝑆, 𝑑𝑑 𝑞𝑞,𝐺𝐺 ≤ 𝑥𝑥
• Goal: 𝒅𝒅 𝒑𝒑,𝑮𝑮 ≤ 𝟐𝟐𝒌𝒌𝟐𝟐
𝑝𝑝
𝑮𝑮
≤ 2𝑘𝑘𝑥𝑥≤ 𝑥𝑥
≤ 𝑥𝑥≤ 𝑥𝑥
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 64: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/64.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
• Assume for any 𝑞𝑞 ∈ 𝑆𝑆, 𝑑𝑑 𝑞𝑞,𝐺𝐺 ≤ 𝑥𝑥
• Goal:
• 𝐻𝐻 ≔ 𝐻𝐻𝑆𝑆 be the subspace passing through 𝑆𝑆
𝑯𝑯𝑝𝑝
𝑮𝑮
𝒅𝒅 𝒑𝒑,𝑮𝑮 ≤ 𝟐𝟐𝒌𝒌𝟐𝟐
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 65: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/65.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
• Assume for any 𝑞𝑞 ∈ 𝑆𝑆, 𝑑𝑑 𝑞𝑞,𝐺𝐺 ≤ 𝑥𝑥
• Goal:
• 𝐻𝐻 ≔ 𝐻𝐻𝑆𝑆 be the subspace passing through 𝑆𝑆
• Let 𝑝𝑝𝐻𝐻 be the projection of 𝑝𝑝 onto 𝐺𝐺
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
𝑮𝑮
𝒅𝒅 𝒑𝒑,𝑮𝑮 ≤ 𝟐𝟐𝒌𝒌𝟐𝟐
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 66: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/66.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
• Assume for any 𝑞𝑞 ∈ 𝑆𝑆, 𝑑𝑑 𝑞𝑞,𝐺𝐺 ≤ 𝑥𝑥
• Goal:
• 𝐻𝐻 ≔ 𝐻𝐻𝑆𝑆 be the subspace passing through 𝑆𝑆
• Let 𝑝𝑝𝐻𝐻 be the projection of 𝑝𝑝 onto 𝐺𝐺
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
𝑮𝑮
≤ 𝑘𝑘𝑥𝑥
≤ 𝑘𝑘𝑥𝑥
Lemma 1: 𝒅𝒅 𝒑𝒑,𝒑𝒑𝑯𝑯 ≤ 𝒌𝒌𝟐𝟐
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝒅𝒅 𝒑𝒑,𝑮𝑮 ≤ 𝟐𝟐𝒌𝒌𝟐𝟐
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 67: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/67.jpg)
Proof.
• Let 𝑝𝑝 ∈ 𝑉𝑉 be a point
• Let 𝐺𝐺 be a (𝑘𝑘 − 1)-dimensional subspace.
• Assume for any 𝑞𝑞 ∈ 𝑆𝑆, 𝑑𝑑 𝑞𝑞,𝐺𝐺 ≤ 𝑥𝑥
• Goal:
• 𝐻𝐻 ≔ 𝐻𝐻𝑆𝑆 be the subspace passing through 𝑆𝑆
• Let 𝑝𝑝𝐻𝐻 be the projection of 𝑝𝑝 onto 𝐺𝐺
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
𝑮𝑮
≤ 2𝑘𝑘𝑥𝑥
Lemma 1: 𝒅𝒅 𝒑𝒑,𝒑𝒑𝑯𝑯 ≤ 𝒌𝒌𝟐𝟐
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝒅𝒅 𝒑𝒑,𝑮𝑮 ≤ 𝟐𝟐𝒌𝒌𝟐𝟐
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 68: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/68.jpg)
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
![Page 69: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/69.jpg)
Proof.
• Since 𝐻𝐻 has dimension 𝑘𝑘, we can write 𝒑𝒑𝑯𝑯 = ∑𝒅𝒅=𝟏𝟏𝒌𝒌 𝜶𝜶𝒅𝒅𝒒𝒒𝒅𝒅
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
![Page 70: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/70.jpg)
Proof.
• Since 𝐻𝐻 has dimension 𝑘𝑘, we can write 𝒑𝒑𝑯𝑯 = ∑𝒅𝒅=𝟏𝟏𝒌𝒌 𝜶𝜶𝒅𝒅𝒒𝒒𝒅𝒅
• Let 𝑯𝑯−𝒅𝒅 ≔ 𝑯𝑯𝑺𝑺∖ 𝒒𝒒𝒅𝒅
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
𝑯𝑯−𝒅𝒅
𝑞𝑞𝑖𝑖
![Page 71: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/71.jpg)
Proof.
• Since 𝐻𝐻 has dimension 𝑘𝑘, we can write 𝒑𝒑𝑯𝑯 = ∑𝒅𝒅=𝟏𝟏𝒌𝒌 𝜶𝜶𝒅𝒅𝒒𝒒𝒅𝒅
• Let 𝑯𝑯−𝒅𝒅 ≔ 𝑯𝑯𝑺𝑺∖ 𝒒𝒒𝒅𝒅
• Since we did not choose 𝑝𝑝 in LS, 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒒𝒒𝒅𝒅,𝑯𝑯−𝒅𝒅
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
𝑯𝑯−𝒅𝒅
𝑞𝑞𝑖𝑖
![Page 72: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/72.jpg)
Proof.
• Since 𝐻𝐻 has dimension 𝑘𝑘, we can write 𝒑𝒑𝑯𝑯 = ∑𝒅𝒅=𝟏𝟏𝒌𝒌 𝜶𝜶𝒅𝒅𝒒𝒒𝒅𝒅
• Let 𝑯𝑯−𝒅𝒅 ≔ 𝑯𝑯𝑺𝑺∖ 𝒒𝒒𝒅𝒅
• Since we did not choose 𝑝𝑝 in LS, 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒒𝒒𝒅𝒅,𝑯𝑯−𝒅𝒅
• Since 𝑝𝑝𝐻𝐻 is a projection of 𝑝𝑝 onto 𝐻𝐻, 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑𝑯𝑯,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑,𝑯𝑯−𝒅𝒅
•
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝑯𝑯𝑝𝑝
𝑝𝑝𝐻𝐻
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
𝑯𝑯−𝒅𝒅
𝑞𝑞𝑖𝑖
![Page 73: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/73.jpg)
Proof.
• Since 𝐻𝐻 has dimension 𝑘𝑘, we can write 𝒑𝒑𝑯𝑯 = ∑𝒅𝒅=𝟏𝟏𝒌𝒌 𝜶𝜶𝒅𝒅𝒒𝒒𝒅𝒅
• Let 𝑯𝑯−𝒅𝒅 ≔ 𝑯𝑯𝑺𝑺∖ 𝒒𝒒𝒅𝒅
• Since we did not choose 𝑝𝑝 in LS, 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒒𝒒𝒅𝒅,𝑯𝑯−𝒅𝒅
• Since 𝑝𝑝𝐻𝐻 is a projection of 𝑝𝑝 onto 𝐻𝐻, 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑𝑯𝑯,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅(𝒑𝒑,𝑯𝑯−𝒅𝒅)
• Thus 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑𝑯𝑯,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒒𝒒𝒅𝒅,𝑯𝑯−𝒅𝒅
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝑯𝑯𝑝𝑝𝐻𝐻
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
𝑯𝑯−𝒅𝒅
𝑞𝑞𝑖𝑖
![Page 74: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/74.jpg)
Proof.
• Since 𝐻𝐻 has dimension 𝑘𝑘, we can write 𝒑𝒑𝑯𝑯 = ∑𝒅𝒅=𝟏𝟏𝒌𝒌 𝜶𝜶𝒅𝒅𝒒𝒒𝒅𝒅
• Let 𝑯𝑯−𝒅𝒅 ≔ 𝑯𝑯𝑺𝑺∖ 𝒒𝒒𝒅𝒅
• Since we did not choose 𝑝𝑝 in LS, 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒒𝒒𝒅𝒅,𝑯𝑯−𝒅𝒅
• Since 𝑝𝑝𝐻𝐻 is a projection of 𝑝𝑝 onto 𝐻𝐻, 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑𝑯𝑯,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅(𝒑𝒑,𝑯𝑯−𝒅𝒅)
• Thus 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒑𝒑𝑯𝑯,𝑯𝑯−𝒅𝒅 ≤ 𝒅𝒅𝒅𝒅𝒅𝒅𝒅𝒅 𝒒𝒒𝒅𝒅,𝑯𝑯−𝒅𝒅
• Thus |𝜶𝜶𝒅𝒅| ≤ 𝟏𝟏
Lemma 2: 𝒅𝒅 𝒑𝒑𝑯𝑯,𝑮𝑮 ≤ 𝒌𝒌𝟐𝟐
𝑯𝑯𝑝𝑝𝐻𝐻
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
𝑯𝑯−𝒅𝒅
𝑞𝑞𝑖𝑖
![Page 75: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/75.jpg)
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
![Page 76: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/76.jpg)
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
Assumption: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞𝑖𝑖 ,𝐺𝐺 ≤ 𝑥𝑥
![Page 77: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/77.jpg)
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
Assumption: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞𝑖𝑖 ,𝐺𝐺 ≤ 𝑥𝑥
Lemma2: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑝𝑝𝐻𝐻 ,𝐺𝐺 ≤ ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑞𝑞𝑖𝑖 ,𝐺𝐺) ≤ 𝑘𝑘 ⋅ 𝑥𝑥 ≤ 𝑘𝑘𝑥𝑥
![Page 78: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/78.jpg)
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
Assumption: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞𝑖𝑖 ,𝐺𝐺 ≤ 𝑥𝑥
Lemma2: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑝𝑝𝐻𝐻 ,𝐺𝐺 ≤ ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑞𝑞𝑖𝑖 ,𝐺𝐺) ≤ 𝑘𝑘 ⋅ 𝑥𝑥 ≤ 𝑘𝑘𝑥𝑥
Lemma 1: 𝑑𝑑 𝑝𝑝,𝑝𝑝𝐻𝐻 ≤ 𝑘𝑘𝑥𝑥
![Page 79: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/79.jpg)
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
Assumption: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞𝑖𝑖 ,𝐺𝐺 ≤ 𝑥𝑥
Lemma2: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑝𝑝𝐻𝐻 ,𝐺𝐺 ≤ ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑞𝑞𝑖𝑖 ,𝐺𝐺) ≤ 𝑘𝑘 ⋅ 𝑥𝑥 ≤ 𝑘𝑘𝑥𝑥
Lemma 1: 𝑑𝑑 𝑝𝑝,𝑝𝑝𝐻𝐻 ≤ 𝑘𝑘𝑥𝑥 Goal: 𝑑𝑑 𝑝𝑝,𝐺𝐺 ≤ 2𝑘𝑘𝑥𝑥
![Page 80: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/80.jpg)
Claim:We can write 𝑝𝑝𝐻𝐻 = ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑞𝑞𝑖𝑖 s.t. all 𝛼𝛼𝑖𝑖 ≤ 1
Assumption: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞𝑖𝑖 ,𝐺𝐺 ≤ 𝑥𝑥
Lemma2: 𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑝𝑝𝐻𝐻 ,𝐺𝐺 ≤ ∑𝑖𝑖=1𝑐𝑐 𝛼𝛼𝑖𝑖𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑞𝑞𝑖𝑖 ,𝐺𝐺) ≤ 𝑘𝑘 ⋅ 𝑥𝑥 ≤ 𝑘𝑘𝑥𝑥
Lemma 1: 𝑑𝑑 𝑝𝑝,𝑝𝑝𝐻𝐻 ≤ 𝑘𝑘𝑥𝑥 Goal: 𝑑𝑑 𝑝𝑝,𝐺𝐺 ≤ 2𝑘𝑘𝑥𝑥
Main Lemma [formal]:For any (𝑘𝑘 − 1)-dimensional subspace 𝐺𝐺, the maximum distance of the point set to 𝐺𝐺 is approximately preserved
max𝑠𝑠∈𝑆𝑆
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑 𝑞𝑞,𝐺𝐺 ≥1
2𝑘𝑘⋅ max𝑝𝑝∈𝑉𝑉
𝑑𝑑𝑖𝑖𝑑𝑑𝑑𝑑(𝑝𝑝,𝐺𝐺)
![Page 81: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/81.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Main TheoremLocal Search produces a core-set for volume maximization
![Page 82: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/82.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Main TheoremLocal Search produces a core-set for volume maximization
![Page 83: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/83.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝑉𝑉2
𝑉𝑉3
𝑉𝑉1
![Page 84: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/84.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝑉𝑉2
𝑉𝑉3
𝑉𝑉1
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 85: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/85.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟐𝟐
𝒐𝒐𝟑𝟑
𝒐𝒐𝟏𝟏
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 86: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/86.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟐𝟐
𝒐𝒐𝟑𝟑
𝒐𝒐𝟏𝟏
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 87: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/87.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟐𝟐
𝒐𝒐𝟑𝟑𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 88: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/88.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟐𝟐
𝒐𝒐𝟑𝟑
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 89: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/89.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟐𝟐
𝒐𝒐𝟑𝟑
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 90: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/90.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟑𝟑
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 91: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/91.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟑𝟑
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 92: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/92.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝒐𝒐𝟑𝟑
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 93: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/93.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 94: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/94.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 95: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/95.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝑉𝑉2
𝑉𝑉3
𝑉𝑉1
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 96: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/96.jpg)
Let 𝑽𝑽 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑺𝑺 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑶𝑶𝒑𝒑𝒅𝒅 = 𝒐𝒐𝟏𝟏, … ,𝒐𝒐𝒌𝒌 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
Main TheoremLocal Search produces a core-set for volume maximization
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
![Page 97: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/97.jpg)
Lose a factor of at most 2𝑘𝑘 at each iteration
Main TheoremLocal Search produces a core-set for volume maximization
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}Since local search preserve
maximum distances to subspaces
Let 𝑉𝑉 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑆𝑆 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑂𝑂𝑝𝑝𝑑𝑑 = 𝑉𝑉1, … , 𝑉𝑉𝑐𝑐 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
![Page 98: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/98.jpg)
Let 𝑉𝑉 = ⋃𝑖𝑖 𝑉𝑉𝑖𝑖 be the union of the point sets
Let 𝑆𝑆 = ⋃𝑖𝑖 𝑆𝑆𝑖𝑖 be the union of core-sets
Let 𝑂𝑂𝑝𝑝𝑑𝑑 = 𝑉𝑉1, … , 𝑉𝑉𝑐𝑐 ⊂ 𝑉𝑉 be the optimal subset of points maximizing the volume
𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑂𝑂𝑝𝑝𝑑𝑑
For 𝑖𝑖 = 1 𝑑𝑑𝑉𝑉 𝑘𝑘
• Let 𝑞𝑞𝑖𝑖 ∈ 𝑆𝑆 be the point that is farthest away from 𝐻𝐻𝑆𝑆𝑆𝑆𝑆𝑆∖ 𝑆𝑆𝑖𝑖
• 𝑆𝑆𝑉𝑉𝑉𝑉 ← 𝑆𝑆𝑉𝑉𝑉𝑉 ∪ 𝑞𝑞𝑖𝑖 ∖ {𝑉𝑉𝑖𝑖}
Lose a factor of at most 2𝑘𝑘 at each iteration
Total approximation factor 2𝑘𝑘 𝑐𝑐
Main TheoremLocal Search produces a core-set for volume maximization
![Page 99: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/99.jpg)
Empirical Results
Data Set• MNIST, with number of parts = 10• MNIST, with number of parts = 50• GENES, with number of parts = 10
Process- Partition the data set randomly into parts- Compute a core-set using one of the algorithms: Greedy, Local
Search, LP-Based algorithm of [IMOR’18]- Use greedy on the union of the coresets
![Page 100: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/100.jpg)
Local Search vs Greedy
Improvement of the solution of Local Search over Greedy On average, 1.2%, 2.5%, and 9.6%
improvement Some cases up to 58% improvement
Ratio of runtime of Local Search over Greedy On average, 6 times slower
![Page 101: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/101.jpg)
Local Search vs. LP-based Algorithm of [IMOR’18]
Improvement of the solution of Local Search over [IMOR’18] On average, 1.4%, 1.8%, and 7.3%
improvement Some cases up to 63% improvement
Ratio of runtime of Local Search over [IMOR’18] For lower values of k, Local Search is up to
50 times faster.
![Page 102: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/102.jpg)
[IMOR’18] Greedy Local SearchApproximation 𝑂𝑂 𝑘𝑘 log𝑘𝑘 𝑐𝑐/2 𝑂𝑂(𝐶𝐶𝑐𝑐2) 𝑂𝑂 𝑘𝑘𝑐𝑐
Core-set Size 𝑂𝑂(𝑘𝑘 log𝑘𝑘) 𝑘𝑘 𝑘𝑘Simple? × ✔ ✔
Empirical Approximation Performs Best
Empirical Runtime Slowest Fastest 4 times slower than Greedy.
Summary• Volume/Determinant Maximization Problem• Notion of composable core-sets• Algorithms that find composable core-sets for volume/determinant maximization
![Page 103: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/103.jpg)
[IMOR’18] Greedy Local SearchApproximation 𝑂𝑂 𝑘𝑘 log𝑘𝑘 𝑐𝑐/2 𝑂𝑂(𝐶𝐶𝑐𝑐2) 𝑂𝑂 𝑘𝑘𝑐𝑐
Core-set Size 𝑂𝑂(𝑘𝑘 log𝑘𝑘) 𝑘𝑘 𝑘𝑘Simple? × ✔ ✔
Empirical Approximation Performs Best
Empirical Runtime Slowest Fastest 4 times slower than Greedy.
Conclusion• Local Search might be the right algorithm to use in massive data models of computation.
Summary• Volume/Determinant Maximization Problem• Notion of composable core-sets• Algorithms that find composable core-sets for volume/determinant maximization
![Page 104: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/104.jpg)
[IMOR’18] Greedy Local SearchApproximation 𝑂𝑂 𝑘𝑘 log𝑘𝑘 𝑐𝑐/2 𝑂𝑂(𝐶𝐶𝑐𝑐2) 𝑂𝑂 𝑘𝑘𝑐𝑐
Core-set Size 𝑂𝑂(𝑘𝑘 log𝑘𝑘) 𝑘𝑘 𝑘𝑘Simple? × ✔ ✔
Empirical Approximation Performs Best
Empirical Runtime Slowest Fastest 4 times slower than Greedy.
Conclusion• Local Search might be the right algorithm to use in massive data models of computation.
Open Problem• Tight analysis of Greedy: does it also provide approximation 𝑘𝑘O(k)?
Summary• Volume/Determinant Maximization Problem• Notion of composable core-sets• Algorithms that find composable core-sets for volume/determinant maximization
![Page 105: ComposableCore-sets for Determinant Maximization: A …mahabadi/slides/Det-ICML.pdfComposableCore-sets for Determinant Maximization: A Simple Near -Optimal Algorithm Sepideh Mahabadi.](https://reader033.fdocuments.net/reader033/viewer/2022042010/5e71f1781578275091135e07/html5/thumbnails/105.jpg)
[IMOR’18] Greedy Local SearchApproximation 𝑂𝑂 𝑘𝑘 log𝑘𝑘 𝑐𝑐/2 𝑂𝑂(𝐶𝐶𝑐𝑐2) 𝑂𝑂 𝑘𝑘𝑐𝑐
Core-set Size 𝑂𝑂(𝑘𝑘 log𝑘𝑘) 𝑘𝑘 𝑘𝑘Simple? × ✔ ✔
Empirical Approximation Performs Best
Empirical Runtime Slowest Fastest 4 times slower than Greedy.
Summary• Volume/Determinant Maximization Problem• Notion of composable core-sets• Algorithms that find composable core-sets for volume/determinant maximization
Conclusion• Local Search might be the right algorithm to use in massive data models of computation.
Open Problem• Tight analysis of Greedy: does it also provide approximation 𝑘𝑘O(k)?
Thank you!