improved np inapproximability for 2 variable linear
play

Improved NP-inapproximability for 2-variable linear equations Johan - PowerPoint PPT Presentation

May 11, 2023 •241 likes •1.24k views

Improved NP-inapproximability for 2-variable linear equations Johan Hstad Sangxia Huang Rajsekar Manokaran KTH KTH KTH Ryan ODonnell John Wright CMU CMU 2Lin x 1 = x 5 x 10 = -x 3 x 61 = -x 24 ... x 48 = -x 5 (x i = -1,1) 2Lin

  1. Improved NP-inapproximability for 2-variable linear equations Johan Håstad Sangxia Huang Rajsekar Manokaran KTH KTH KTH Ryan O’Donnell John Wright CMU CMU

  2. 2Lin x 1 = x 5 x 10 = -x 3 x 61 = -x 24 ... x 48 = -x 5 (x i = -1,1)

  3. 2Lin 2Lin( 2 ) ∈ 2Lin( q ) ≈ UniqueGames x 1 = x 5 x 10 = -x 3 x 61 = -x 24 ... x 48 = -x 5 (x i = -1,1)

  4. 2Lin 2Lin( 2 ) ∈ 2Lin( q ) ≈ UniqueGames x 1 = x 5 (Actually, simplest case of UG ) x 10 = -x 3 x 61 = -x 24 ... x 48 = -x 5 (x i = -1,1)

  5. 2Lin 2Lin( 2 ) ∈ 2Lin( q ) ≈ UniqueGames x 1 = x 5 (Actually, simplest case of UG ) x 10 = -x 3 x 61 = -x 24 Folklore wisdom: get 2Lin( 2 ) right ... and 2Lin( q ) will follow. x 48 = -x 5 (x i = -1,1)

  6. Known results Suppose val( I ) = α . Can we guarantee a solution of value C*α ?

  7. Known results Suppose val( I ) = α . Can we guarantee a solution of value C*α ? [GW] : .878-approx algorithm

  8. Known results Suppose val( I ) = α . Can we guarantee a solution of value C*α ? [GW] : .878-approx algorithm [KKMO] + [MOO] : (.878+ ε )-approx UG -hard

  9. Known results Suppose val( I ) = α . Can we guarantee a solution of value C*α ? [GW] : .878-approx algorithm [KKMO] + [MOO] : (.878+ ε )-approx UG -hard [Håstad] + [TSSW] : 16 / 17 ≈ .941-approx NP -hard

  10. Known results Suppose val( I ) = α . Can we guarantee a solution of value C*α ? [GW] : .878-approx algorithm [KKMO] + [MOO] : (.878+ ε )-approx UG -hard [Håstad] + [TSSW] : 16 / 17 ≈ .941-approx NP -hard seems we’re close, right?

  11. A different perspective... Suppose val( I ) = (1 - ε ). Can we guarantee a solution of value (1 - C*ε )?

  12. A different perspective... Suppose val( I ) = (1 - ε ). Can we guarantee a solution of value (1 - C*ε )? Def: Such an algo. gives an ( ε , C*ε )-approx.

  13. A different perspective... Suppose val( I ) = (1 - ε ). Can we guarantee a solution of value (1 - f(ε) )? Def: Such an algo. gives an ( ε , f(ε) )-approx.

  14. A different perspective... Suppose val( I ) = (1 - ε ). Can we guarantee a solution of value (1 - f(ε) )? Def: Such an algo. gives an ( ε , f(ε) )-approx. Usually called “Min-2Lin(2)-Deletion”. Let me just call this 2Lin .

  15. Unratio state of affairs [easy] : ( ε , ε )-approx NP -hard

  16. Unratio state of affairs [easy] : ( ε , ε )-approx NP -hard [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard

  17. Unratio state of affairs [easy] : ( ε , ε )-approx NP -hard [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [KKMO] + [MOO] : ( ε , O( ε 1/2 ))-approx UG -hard

  18. Unratio state of affairs [easy] : ( ε , ε )-approx NP -hard [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [KKMO] + [MOO] : ( ε , O( ε 1/2 ))-approx UG -hard [GW] : ( ε , O( ε 1/2 ))-approx algorithm

  19. Unratio state of affairs [easy] : ( ε , ε )-approx NP -hard [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [KKMO] + [MOO] : ( ε , O( ε 1/2 ))-approx UG -hard [GW] : ( ε , O( ε 1/2 ))-approx algorithm asymptotically off from the truth

  20. Unratio state of affairs [easy] : ( ε , ε )-approx NP -hard [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [KKMO] + [MOO] : ( ε , O( ε 1/2 ))-approx UG -hard [GW] : ( ε , O( ε 1/2 ))-approx algorithm asymptotically off from the truth [Rao] : If ( ε , O(f( q )* ε 1/2 ))-approx is NP -hard for 2Lin( q ), for f( q ) = Ω(1), then UG is true.

  21. This work [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard

  22. This work [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [Us] : ( ε , 11 / 8 * ε )-approx NP -hard

  23. This work [Håstad] + [TSSW] : ( ε , 1.25* ε )-approx NP -hard [Us] : ( ε , 1.375* ε )-approx NP -hard

  24. This work [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [Us] : ( ε , 11 / 8 * ε )-approx NP -hard

  25. This work [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [Us] : ( ε , 11 / 8 * ε )-approx NP -hard Cons: - Still haven’t proven UniqueGames . 😟

  26. This work [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [Us] : ( ε , 11 / 8 * ε )-approx NP -hard Cons: - Still haven’t proven UniqueGames . 😟 Pros: - First improvement since 1997.

  27. This work [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [Us] : ( ε , 11 / 8 * ε )-approx NP -hard Cons: - Still haven’t proven UniqueGames . 😟 Pros: - First improvement since 1997. - Study new type of “gadget reduction”

  28. This work [Håstad] + [TSSW] : ( ε , ⁵ / ₄ * ε )-approx NP -hard [Us] : ( ε , 11 / 8 * ε )-approx NP -hard ( and more! ) Cons: - Still haven’t proven UniqueGames . 😟 Pros: - First improvement since 1997. - Study new type of “gadget reduction”

  29. Proving (ε, ⁵ / ₄ *ε)-hardness Standard two-step plan.

  30. Proving (ε, ⁵ / ₄ *ε)-hardness Standard two-step plan. [Håstad]: Given 3Lin instance I , NP -hard to distinguish - Yes : val( I ) ≥ (1 - ε ) - No : val( I ) ≤ (½ + ε )

  31. Proving (ε, ⁵ / ₄ *ε)-hardness Standard two-step plan. [Håstad]: Given 3Lin instance I , NP -hard to distinguish - Yes : val( I ) ≥ (1 - ε ) - No : val( I ) ≤ (½ + ε ) (In our language, ( ε , ½ - ε )-approxing 3Lin is NP -hard.)

  32. Proving (ε, ⁵ / ₄ *ε)-hardness Standard two-step plan. [Håstad]: Given 3Lin instance I , NP -hard to distinguish - Yes : val( I ) ≥ (1 - ε ) - No : val( I ) ≤ (½ + ε ) (In our language, ( ε , ½ - ε )-approxing 3Lin is NP -hard.) Step 2: gadget reduce 3Lin to 2Lin [TSSW]

  33. Proving (ε, ⁵ / ₄ *ε)-hardness Standard two-step plan. [Håstad]: Given 3Lin instance I , NP -hard to distinguish - Yes : val( I ) ≥ (1 - ε ) - No : val( I ) ≤ (½ + ε ) (In our language, ( ε , ½ - ε )-approxing 3Lin is NP -hard.) Step 2: gadget reduce 3Lin to 2Lin [TSSW] (see also [OW12] )

  34. 3Lin x 1 x 3 x 5 = 1 x 10 x 16 x 3 = -1 ... x 47 x 11 x 98 = -1

  35. 3Lin x 1 x 3 x 5 = 1 x 10 x 16 x 3 = -1 ... x 47 x 11 x 98 = -1

  36. 3Lin x 10 x 1 x 3 x 5 = 1 x 16 x 10 x 16 x 3 = -1 x 3 ... x 47 x 11 x 98 = -1

  37. (aux vars) 3Lin x 10 x 1 x 3 x 5 = 1 x 16 x 10 x 16 x 3 = -1 x 3 ... x 47 x 11 x 98 = -1

  38. (aux vars) 3Lin x 10 x 1 x 3 x 5 = 1 x 16 x 10 x 16 x 3 = -1 x 3 ... x 47 x 11 x 98 = -1

  39. 2Lin (aux vars) 3Lin gadget x 10 x 1 x 3 x 5 = 1 x 10 = - x 3 x 16 x 10 x 16 x 3 = -1 y 61 = -y 24 x 3 ... ... x 47 x 11 x 98 = -1 x 16 = -y 5

  40. 2Lin (aux vars) 3Lin gadget x 10 x 1 x 3 x 5 = 1 x 10 = - x 3 x 16 x 10 x 16 x 3 = -1 y 61 = -y 24 x 3 ... ... x 47 x 11 x 98 = -1 x 16 = -y 5 Final 2Lin inst: union all the gadgets

  41. 2Lin (aux vars) 3Lin gadget x 10 x 1 x 3 x 5 = 1 x 10 = - x 3 x 16 x 10 x 16 x 3 = -1 y 61 = -y 24 x 3 ... ... x 47 x 11 x 98 = -1 x 16 = -y 5 Final 2Lin inst: union all the gadgets The hope: - x i ’s satisfy 3Lin eq’n ⇒ good assgn to y i ’s

  42. 2Lin (aux vars) 3Lin gadget x 10 x 1 x 3 x 5 = 1 x 10 = - x 3 x 16 x 10 x 16 x 3 = -1 y 61 = -y 24 x 3 ... ... x 47 x 11 x 98 = -1 x 16 = -y 5 Final 2Lin inst: union all the gadgets The hope: - x i ’s satisfy 3Lin eq’n ⇒ good assgn to y i ’s - x i ’s don’t ⇒ no good assgn to y i ’s

  43. Gadgets Def: A ( c , s )-gadget

  44. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c )

  45. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c ) - x i ’s don’t ⇒ no assgn to y i ’s beats value (1 - s )

  46. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c ) - x i ’s don’t ⇒ no assgn to y i ’s beats value (1 - s ) [TSSW] : there is a 3Lin-to-2Lin ( ¼ , ⅜ )-gadget

  47. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c ) - x i ’s don’t ⇒ no assgn to y i ’s beats value (1 - s ) [TSSW] : there is a 3Lin-to-2Lin ( ¼ , ⅜ )-gadget ( ε , ½ - ε )-hardess for 3Lin

  48. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c ) - x i ’s don’t ⇒ no assgn to y i ’s beats value (1 - s ) [TSSW] : there is a 3Lin-to-2Lin ( ¼ , ⅜ )-gadget ( ε , ½ - ε )-hardess for 3Lin ⇒ - Yes case : ¼

  49. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c ) - x i ’s don’t ⇒ no assgn to y i ’s beats value (1 - s ) [TSSW] : there is a 3Lin-to-2Lin ( ¼ , ⅜ )-gadget ( ε , ½ - ε )-hardess for 3Lin ⇒ - Yes case : ¼ - No case : ½ * (¼ + ⅜)

  50. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c ) - x i ’s don’t ⇒ no assgn to y i ’s beats value (1 - s ) [TSSW] : there is a 3Lin-to-2Lin ( ¼ , ⅜ )-gadget ( ε , ½ - ε )-hardess for 3Lin ⇒ - Yes case : ¼ - No case : ½ * (¼ + ⅜) = 5 / 16

  51. Gadgets Def: A ( c , s )-gadget - x i ’s satisfy 3Lin eq’n ⇒ an assgn to y i ’s of value (1 - c ) - x i ’s don’t ⇒ no assgn to y i ’s beats value (1 - s ) [TSSW] : there is a 3Lin-to-2Lin ( ¼ , ⅜ )-gadget ( ε , ½ - ε )-hardess for 3Lin ⇒ - Yes case : ¼ - No case : ½ * (¼ + ⅜) = 5 / 16 = 5 / 4 * ¼

  52. How do you find gadgets? Gadgets are just 2Lin instances, so can just monkey around with small instances.

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

Improved Inapproximability for TSP The Role of Bounded Occurrence CSPs Michael Lampis KTH Royal

Improved Inapproximability for TSP The Role of Bounded Occurrence CSPs Michael Lampis KTH Royal

Improved Inapproximability for TSP The Role of Bounded Occurrence CSPs Michael Lampis KTH Royal Institute of Technology January 22, 2013 The Story Good research involves good storytelling Mike Fellows Improved Inapproximability for TSP 2 /

1.48k views • 124 slides
Improved Inapproximability for TSP Michael Lampis KTH Royal Institute of Technology August 15,

Improved Inapproximability for TSP Michael Lampis KTH Royal Institute of Technology August 15,

Improved Inapproximability for TSP Michael Lampis KTH Royal Institute of Technology August 15, 2012 The Traveling Salesman Problem Input: An edge-weighted graph G ( V, E ) Objective: Find an ordering of the vertices v 1 , v 2 , . . . ,

1.01k views • 64 slides
Baby steps towards TSP inapproximability Michael Lampis KTH Royal Institute of Technology May

Baby steps towards TSP inapproximability Michael Lampis KTH Royal Institute of Technology May

Baby steps towards TSP inapproximability Michael Lampis KTH Royal Institute of Technology May 3, 2013 Acknowledgements The material in this talk is based on the following papers: Improved Inapproximability for TSP, APPROX12

1.98k views • 132 slides
Multiple Linear Regression Often more than one predictor variable can be used to predict the

Multiple Linear Regression Often more than one predictor variable can be used to predict the

ST 370 Probability and Statistics for Engineers Multiple Linear Regression Often more than one predictor variable can be used to predict the value of a response variable. The basic approach of the simple linear regression model may be extended

204 views • 8 slides
1 Example : Example: Medical researchers have noted that adolescent females are more likely

1 Example : Example: Medical researchers have noted that adolescent females are more likely

Regression analysis is used to investigate whether there is a linear STAT E-150 relationship between two quantitative variables. Statistical Methods The variable we want to predict is the response variable ; the variable we use for this

366 views • 11 slides
Variable Ranges in Linear Constraints Salvatore RUGGIERI and Fred MESNARD Dipartimento di

Variable Ranges in Linear Constraints Salvatore RUGGIERI and Fred MESNARD Dipartimento di

Variable Ranges in Linear Constraints Salvatore RUGGIERI and Fred MESNARD Dipartimento di Informatica, Universit` a di Pisa, Italy LIM-IREMIA, universit e de la R eunion, France 1/22 Introduction: example We add variable ranges to linear

506 views • 25 slides
Linear Statistics Jung-Keun Lee and Woo-Hwan Kim ETRI Our contribution improved and extended

Linear Statistics Jung-Keun Lee and Woo-Hwan Kim ETRI Our contribution improved and extended

FSE 2020 Multiple Linear Cryptanalysis Using Linear Statistics Jung-Keun Lee and Woo-Hwan Kim ETRI Our contribution improved and extended approach of multiple linear cryptanalysis[BCQ04] (exploit dominant statistically independent linear

714 views • 30 slides
Improved Differential-Linear Cryptanalysis of 7-round Chaskey with Partitioning Gatan Leurent

Improved Differential-Linear Cryptanalysis of 7-round Chaskey with Partitioning Gatan Leurent

Chaskey ARX Cryptanalysis Improved Differential-Linear Conclusion Improved Differential-Linear Cryptanalysis of 7-round Chaskey with Partitioning Gatan Leurent Inria, Paris Eurocrypt 2016 m 0 m 1 m 2 K K K Gatan

540 views • 36 slides
Inapproximability of Congestion Games Alexander Skopalik, Berthold V ocking Department of

Inapproximability of Congestion Games Alexander Skopalik, Berthold V ocking Department of

Inapproximability of Congestion Games Alexander Skopalik, Berthold V ocking Department of Computer Science RWTH Aachen Warwick 2007 Alexander Skopalik, Berthold V ocking Inapproximability of Congestion Games Network Congestion Games

733 views • 38 slides
Improved Heuristics for Short Linear Programs Thomas Peyrin Quan Quan Tan Nanyang Technological

Improved Heuristics for Short Linear Programs Thomas Peyrin Quan Quan Tan Nanyang Technological

Improved Heuristics for Short Linear Programs Thomas Peyrin Quan Quan Tan Nanyang Technological University CHES 2020 Contributions of this paper: A new algorithm that finds good implementations of linear systems, to reduce the number of XOR

538 views • 28 slides
An Improved Regret Bound for Thompson Sampling in the Gaussian Linear Bandit Setting Cem

An Improved Regret Bound for Thompson Sampling in the Gaussian Linear Bandit Setting Cem

An Improved Regret Bound for Thompson Sampling in the Gaussian Linear Bandit Setting Cem Kalkanl, Ayfer Ozg ur Stanford University ISIT, June 2020 An Improved Regret Bound for Thompson Sampling in the Gaussian Linear Bandit Setting 1

379 views • 13 slides
Session 06 Generalized Linear Models 1 Nature of the generalization Single response variable,

Session 06 Generalized Linear Models 1 Nature of the generalization Single response variable,

Session 06 Generalized Linear Models 1 Nature of the generalization Single response variable, y Some candidate predictor variables x 1 , x 2 , , x p . The distribution of y can only depend on the predictors through a single linear

376 views • 25 slides
An improved primal simplex algorithm and column generation for degenerate linear programs

An improved primal simplex algorithm and column generation for degenerate linear programs

An improved primal simplex algorithm and column generation for degenerate linear programs Abdelmoutalib Metrane Ecole Polytechnique and GERAD Montreal Canada Column Generation 2008 Introduction Improved Primal Simplex Column generation

490 views • 35 slides
Logistic Regression Linear regression is designed for a quantitative response variable; in the

Logistic Regression Linear regression is designed for a quantitative response variable; in the

ST 370 Probability and Statistics for Engineers Logistic Regression Linear regression is designed for a quantitative response variable; in the model equation Y = 0 + 1 x + , the random noise term is usually assumed to be at least

322 views • 15 slides
Subselect 0.9-99 subsets in multivariate linear models A. PEDRO DUARTE SI LVA ( 1 ) ( * ) THE

Subselect 0.9-99 subsets in multivariate linear models A. PEDRO DUARTE SI LVA ( 1 ) ( * ) THE

Subselect 0.9-99: Selecting variable Subselect 0.9-99 subsets in multivariate linear models A. PEDRO DUARTE SI LVA ( 1 ) ( * ) THE PROBLEM : Finding a k-variable subset that is JORGE CADI MA ( 2 ) MANUEL MI NHOTO ( 3 ) a good surrogate for a

368 views • 4 slides
Optimal Inapproximability of Max CSPs over large alphabet Pasin Manurangsi 1 Preetum Nakkiran 2

Optimal Inapproximability of Max CSPs over large alphabet Pasin Manurangsi 1 Preetum Nakkiran 2

Optimal Inapproximability of Max CSPs over large alphabet Pasin Manurangsi 1 Preetum Nakkiran 2 Luca Trevisan 1 1 UC Berkeley 2 Harvard RANDOM-APPROX 2016 1/17 Max k -CSP R Maximum Constraint Satisfaction Problem: Variables take values in

1.08k views • 61 slides
Procurement Division Evaluation by Best Value for Money Methodology 1 Best Value for Money

Procurement Division Evaluation by Best Value for Money Methodology 1 Best Value for Money

Procurement Division Evaluation by Best Value for Money Methodology 1 Best Value for Money Best value principle as per UN Financial Regulations & Rules: 5.12 Due consideration when exercising the procurement function Move from

579 views • 9 slides
Some Experiences with Rapid Evidence Services Targeting Health-System Policymakers &

Some Experiences with Rapid Evidence Services Targeting Health-System Policymakers &

Some Experiences with Rapid Evidence Services Targeting Health-System Policymakers & Stakeholders Global Evidence Synthesis Initiative / Cochrane Learning Live Webinar John N. Lavis, MD, PhD Canada Research Chair in Evidence-Informed Health

391 views • 10 slides
Comparing Temporal Graphs with Time Warping Malte Renken Algorithmics and Computational

Comparing Temporal Graphs with Time Warping Malte Renken Algorithmics and Computational

0/18 Comparing Temporal Graphs with Time Warping Malte Renken Algorithmics and Computational Complexity, TU Berlin, Germany 8. July 2019 Joint work with Vincent Froese, Brijnesh Jain and Rolf Niedermeier. 1/18 Temporal Graphs t t t t

908 views • 72 slides
Time Michael George November 10, 2005 Michael George Time Overview Optimal Clock

Time Michael George November 10, 2005 Michael George Time Overview Optimal Clock

Overview Optimal Clock Synchronization Probablistic Clock Synchronization Conclusions Time Michael George November 10, 2005 Michael George Time Overview Optimal Clock Synchronization Probablistic Clock Synchronization Conclusions The

1.1k views • 71 slides
Predicting biathlon shooting performance using machine learning Thomas Maier 1 , Daniel Meister 2

Predicting biathlon shooting performance using machine learning Thomas Maier 1 , Daniel Meister 2

Predicting biathlon shooting performance using machine learning Thomas Maier 1 , Daniel Meister 2 , Severin Trsch 1 , Jon Peter Wehrlin 1 1 Eidgenssische Hochschule fr Sport Magglingen EHSM 2 Datahouse AG Introduction Shooting is crucial

610 views • 29 slides
Einf uhrung in Pragmatik und Diskurs Implicatures Ivana Kruijff-Korbayov a

Einf uhrung in Pragmatik und Diskurs Implicatures Ivana Kruijff-Korbayov a

E R S V I T I N A U S S S I A S R N A E V I Einf uhrung in Pragmatik und Diskurs Implicatures Ivana Kruijff-Korbayov a korbay@coli.uni-sb.de http://www.coli.uni-saarland.de/courses/pd/ Summer Semester 2005

815 views • 50 slides
Performance Densities in Elite Sports Backgrounds Practice Gerard Sierksma University of

Performance Densities in Elite Sports Backgrounds Practice Gerard Sierksma University of

Performance Densities in Elite Sports Backgrounds Practice Gerard Sierksma University of Groningen ORTEC-Sports April 12 th , 2018 End Three parts: 1. Backgrounds 2. Practice 3. Future and research problems Progress of the Mens World

1.06k views • 66 slides
Competition performances are related to technique test measures in elite rifle shooters Simo

Competition performances are related to technique test measures in elite rifle shooters Simo

Competition performances are related to technique test measures in elite rifle shooters Simo Ihalainen KIHU Research Institute for Olympic Sports University of Jyvskyl, Department of Biology of Physical Activity 16.9.2016 BACKGROUND

270 views • 11 slides

More recommend