1. Some 3CNF Properties are Hard to Test Eli Ben-Sasson Harvard & MIT Prahladh Harsha MIT Sofya Raskhodnikova MIT

2. June 10, 2003-- STOC '03 --2 Property - Definition Property – Set of Strings Property – Set of Stringse.g.: 10000111111100 00111110001101 10101100100001  Triangle Free Graphs 1. 2. Satisfying assignments of a fixed CNF Æ ( : e i,j Ç : e j,k Ç : e k,i )

3. June 10, 2003-- STOC '03 --3 Property Testing - Definition [ Rubinfeld Sudan 96] [Goldreich Goldwasser Ron 98] V Classical Verifier Probabilistic Tester

4. June 10, 2003-- STOC '03 --4 Definition - Contd YES: Verifier accepts with probability 1 YES: Verifier accepts with probability 1 FAR from YES: Accepts with low probability FAR from YES: Accepts with low probability FAR – terms of hamming distance Extensions 2 sided Error 2 sided Error Adaptive Questions Adaptive Questions YES FAR from YES

5. June 10, 2003-- STOC '03 --5 Property Testing - Uses Naturally arise in contexts of PCPs Naturally arise in contexts of PCPs Massive Data Sets Massive Data Sets Eg: WWW, DNA samples, High Resolution Images Eg: WWW, DNA samples, High Resolution Images Large Access Time Large Access Time Need to efficiently check if data satisfies certain properties Need to efficiently check if data satisfies certain properties

6. June 10, 2003-- STOC '03 --6 Testable Properties Testable with constant number of queries bipartiteness [GGR 98] bipartiteness [GGR 98] k-colorability [GGR 98] k-colorability [GGR 98] membership in a regular language [AKNS 99] membership in a regular language [AKNS 99] Testing if function is linear [BLR 90] Testing if function is linear [BLR 90]

7. June 10, 2003-- STOC '03 --7 Non-Testable Properties [GGR 98] prove there exist properties not testable even with a linear number of queries. (Probabilistic Construction) [GGR 98] prove there exist properties not testable even with a linear number of queries. (Probabilistic Construction) Explicit Linear Lower bounds Explicit Linear Lower bounds 3 Colorable Bounded Degree Graphs [BOT 02] 3 Colorable Bounded Degree Graphs [BOT 02] Polynomials of degree n/2 represented as function evaluation [Sudan] Polynomials of degree n/2 represented as function evaluation [Sudan]

8. June 10, 2003-- STOC '03 --8 Testable Properties – Local Views Property is testable ) string far from property has lot of local views showing violation. Property is testable ) string far from property has lot of local views showing violation. Lower Bounds of [BOT 02] and [Sudan] exploit the fact that there are no small local views showing violation. Lower Bounds of [BOT 02] and [Sudan] exploit the fact that there are no small local views showing violation. traingle free graphs

9. June 10, 2003-- STOC '03 --9 Properties as CNF formulae Strings (length n) Has property ? 001101…..1X 010101…..0£  Each property can be represented as a CNF formula. Triangle Free Graphs Æ ( : e i,j Ç : e j,k Ç : e k,i )

10. June 10, 2003-- STOC '03 --10 CNF Property Testing CNF Property Testing: CNF Property Testing: For a fixed CNF (i.e., property), given an assignment, is it For a fixed CNF (i.e., property), given an assignment, is it A satisfying assignment? Or A satisfying assignment? Or Far from satisfying? Far from satisfying? Note: Different from the testing if CNF is satisfiable or far from satisfiable. Note: Different from the testing if CNF is satisfiable or far from satisfiable.

11. June 10, 2003-- STOC '03 --11 Bounds for CNF Property Testing Some CNF properties - hard to test [GGR 98] Some CNF properties - hard to test [GGR 98] 2CNF Property Testing: Testable with O( p n) queries [FLNRRS 02] 2CNF Property Testing: Testable with O( p n) queries [FLNRRS 02] What about kCNFs (k > 2)? What about kCNFs (k > 2)? “Possibly testable”: there exist “witness” of size k that falsifies kCNF. “Possibly testable”: there exist “witness” of size k that falsifies kCNF.

12. June 10, 2003-- STOC '03 --12 3CNF Property Testing - Hard Main Theorem: There exist 3CNF formulae that require linear number of queries, even with adaptive 2-sided error tests.

13. June 10, 2003-- STOC '03 --13 kCNF kLIN 3CNF: (x 1 Ç : x 2 Ç x 4 ) Æ ( : x 2 Ç x 3 Ç x 1 )  (x 25 Ç : x 10 ) + +3LIN: (x 3 © x 5 © x 1 ) Æ ( x 2 © x 3 © x 1 )  (x 23 © x 11 ) Advantages: Can use Linear Algebra

14. June 10, 2003-- STOC '03 --14 Linear Properties Defined by linear constraints. Testing membership in linear space. Defined by linear constraints. Testing membership in linear space. Variables Constraints V – set of vectors that satisfy all constraints. Right degree · k ) V can be represented by kLIN x1x1 x2x2 x3x3 xnxn X 1 © x 2 © x 4 = 0 (mod 2)

15. June 10, 2003-- STOC '03 --15 Lower Bound Proof (Linear property) 1. For linear property, adaptivity and 2-sided error does not help. 2. Prove sufft. properties for V to be hard for 1- sided non-adaptive tests. 3. Prove random linear spaces satisfy above properties. 4. k large in Step 3. Reduce k ! 3.

16. June 10, 2003-- STOC '03 --16 Adaptivity and 2-sided Error Theorem: For testing linear properties, adaptivity and 2- sided error do not help. Key Idea: Accept only if no constraints are violated. Accept only if no constraints are violated. To check if a linear constraint is satisfied, the order of checking the variables is immaterial. To check if a linear constraint is satisfied, the order of checking the variables is immaterial.

17. June 10, 2003-- STOC '03 --17 Lower Bound Proof Want to prove : 8 prob. Tests T, 9 string x far from having property, Pr[ T accepts x] is high. Sufft. to prove : (by Yao’s MinMax Principle) 9 bad distribution B of strings far from property, 8 deterministic tests T, Pr x à B [ T accepts x ] is high

18. June 10, 2003-- STOC '03 --18 Bad Distribution - Defintion Distribution B: uniformly pick a basis constraint c, uniformly pick a vector that falsifies only c. variablesconstraints linearly independent constraints falsified constraint 1 1 1 0

19. June 10, 2003-- STOC '03 --19 Sufficient Properties – Hard to Test If basis constraints satisfy, Property 1:  -separatedness Property 1:  -separatedness Property 2: (q,  )-locality Property 2: (q,  )-locality then, linear space is hard to test.

20. June 10, 2003-- STOC '03 --20 Property 1:  -Separatedness  -separated: Any string x that falsifies exactly one basis constraint has large weight.  -separated: Any string x that falsifies exactly one basis constraint has large weight. falsified constraint 1 1 1 0 w(1110 ) - large  -separatedness + All strings in B (bad distribution) are far from linear space V.

21. June 10, 2003-- STOC '03 --21 How can a test detect a string is from the bad distribution B? (q,  ) locality: Any dual constraint, that is a sum of at least  n basis constraints, depends on more than q variables. Property 2: (q,  ) – Locality x 2 + x 3 = 0 (mod 2) Dual Constraint, which is a sum of large number of basis constraints, depends on few variables.

22. June 10, 2003-- STOC '03 --22 Probabilistic Construction Properties 1 & 2 are expansion-like properties. Hence, random LDPC codes satisfy Properties 1 and 2. Properties 1 & 2 are expansion-like properties. Hence, random LDPC codes satisfy Properties 1 and 2. However, k (max. right degree) – large. However, k (max. right degree) – large. Reduction k ! 3: Reduction k ! 3: (x 1 © x 2 ©  x d ) (x 1 © x 2 ©  x d )+ (x 1 © x 2 ©  x d/2 © z) and (x d/2+1 © x d/2+2 ©  x d © z) This reduction preserves properties 1 and 2.

23. June 10, 2003-- STOC '03 --23 Summarizing…. For testing membership in a linear space, adaptivity and 2-sided error do not help. For testing membership in a linear space, adaptivity and 2-sided error do not help. Random LDPC codes are hard to test even with a linear number of queries. Random LDPC codes are hard to test even with a linear number of queries. Finally, Finally, There exist properties describable by 3CNFs that are hard to test with linear number of queries, even for adaptive 2-sided error tests.