Algebraic Property Testing Madhu Sudan MIT CSAIL Joint work with Tali Kaufman (IAS).
Download ReportTranscript Algebraic Property Testing Madhu Sudan MIT CSAIL Joint work with Tali Kaufman (IAS).
Algebraic Property Testing Madhu Sudan MIT CSAIL Joint work with Tali Kaufman (IAS). Classical Data Processing Big computers Small Data Modern Data Processing Small computers Enormous Data Needs new algorithmic paradigm Imbalance not a question of technology : I.e., not because computing speeds are growing less fast than memory capacity. Imbalance is a function of expectations: E.g., Users expect to be able to “analyze” the WWW, using a laptop. But WWW includes millions other such laptops. Need: Sublinear time algorithms That “estimate” rather than “compute” some given function. Property Testing Dat a: f : D ! R f given by a sampling box x Property: F µ f f : D ! Rg f f(x) q-query Test : Samples f -box q t imes. Accept s if f 2 F Hope: Reject s if f 6 2F Impossible wit h q ¿ jD j Reject s if f is ±-far from F f ±-close t o F if 9 g 2 F s.t . Pr x 2 D [f (x) 6 = g(x)] · ± Property Testing Dat a: f : D ! R f given by a sampling box x Property: F µ f f : D ! Rg f f(x) q-query Test : Samples f -box q t imes. Accept s if f 2 F Hope: Reject s if f 6 2F Impossible wit h q ¿ jD j Reject s if f is ±-far from F f ±-close t o F if 9 g 2 F s.t . Pr x 2 D [f (x) 6 = g(x)] · ± Example: Linearity Testing f f : Fn ! F2 g 2 [Blum, Luby, Rubinfeld ’90] D = Fn ; R = F2 2 Property = Linearity F = f f j8x; yf (x) + f (y) = f (x + y)g Test: Pick random x; y Accept if f (x) + f (y) = f (x + y) F Non-trivial analysis: f ±-far from F ) reject w.p. 2±=9 Example: Linearity Testing f f : Fn ! F2 g 2 [Blum, Luby, Rubinfeld ’90] D = Fn ; R = F2 2 Property = Linearity F = f f j8x; yf (x) + f (y) = f (x + y)g Test: Pick random x; y Accept if f (x) + f (y) = f (x + y) F Non-trivial analysis: f ±-far from F ) reject w.p. 2±=9 Property Testing: Abbreviated History Prehistoric: Statistical sampling E.g., “Is mean/median at least X”. Linearity Testing [BLR’90], Multilinearity Testing [Babai, Fortnow, Lund ’91]. Graph/Combinatorial Property Testing [Goldreich, Goldwasser, Ron ’94]. E.g., Is a graph “close” to being 3-colorable. Algebraic Testing [GLRSW,RS,FS,AKKLR,KR,JPSZ] Is multivariate function a polynomial (of bounded degree). Graph Testing [Alon-Shapira, AFNS, Borgs et al.] Characterizes graph properties that are testable. This Talk Abstracting Algebraic Property Testing Generic If F µ FnTheorem: ! F is closed under addit ion, and under a± ne t ranformat ions of t he coordinat es, and is locally charact erized t hen it is t est able. Motivations: Generalizes, unifies previous algebraic works Initiates systematic study of testability for algebraic properties Sheds light on testing and invariances of properties. Property Testing vs. “Statistics” Classical Statistics (Mean, Median, Quantiles): Also run in sublinear time. So what’s special about “linearity testing”? Classical statistics work on “symmetric” properties: F closed under arbit rary permut at ion on D . Linear functions closed under much smaller group of permutations: F closed under linear maps L : Fn ! Fn 2 2 jD j l og j D j such maps vs. jD j!. Simlarly, graph properties have nice invariances Is testing a corollary of invariance? Example 1: D = R; F 1 = f I djI d(x) = xg Invariant group trivial. Testing easy. Invariant group still trivial. No O(1) local tests. D = f 1; : : : ; ng; R = f 1; : : : ; n 9 g; Example 2: F = f f jx < y ) f (x) < f (y)g 2 Conclusion: Testing not necessarily a consequence of invariance. If we believe this to be the case for the linearity test, must prove it!! Is testing a corollary of invariance? Example 1: D = R; F 1 = f I djI d(x) = xg Invariant group trivial. Testing easy. Invariant group still trivial. No O(1) local tests. D = f 1; : : : ; ng; R = f 1; : : : ; n 9 g; Example 2: F = f f jx < y ) f (x) < f (y)g 2 Conclusion: Testing not necessarily a consequence of invariance. If we believe this to be the case for the linearity test, must prove it!! Part II: Formal Definitions & Results Linear Invariance F is Linear Invariant if ² F is a linear subspace (of Fj Fj n ) ² f 2 F and L : Fn ! Fn linear ) f ± L 2 F (A± ne Invariance de¯ned similarly) Examples: ² ² ² ² Linear funct ions, n-variat e polynomials of degree · d, homogenous polynomialsof degree d, F1 + F2 Testing, constraints, characterizations ² Suppose F has a k-query t est . ² T hen members of F sat isfy a k-local const raint . Const raint : C= (x 1 ; : : : ; x k 2 Fn ; subspace V µ Fk ) 8 f 2 F ; f sat is¯es C i.e., hf (x 1 ); : : : ; f (x k )i 2 V ² E.g., in t he linear case: f (®) + f (¯) = f (® + ¯) C = (®; ¯; ® + ¯; V = f 000; 011; 101; 110g) Charact erizat ion: C = f C1 ; : : : ; Cm g f 2 F , f sat is¯es C1 ; : : : ; Cm : (Linear-Invariant) Algebraic Characterizations Characterizations require many constraints! Linear (affine) invariance turns one constraint into many. C = (x 1 ; : : : ; x k ; V ) const raint and L linear (a± ne) ) C ± L = (L (x 1 ); : : : ; L (x k ); V ) is also a const raint . (Linearity) (L (®); L (¯); Example: L (® + ¯); V ) const raint for linearity Algebraic Charact erizat ion: Single const raint C s.t . f C ± L jL linear (a± ne) g charact erize F . Main Theorems T heorem: F a± ne-invariant and has k-local algebraic charact erizat ion, implies it has a k-query property t est . ² Uni¯es, simpli¯es, and ext ends previous algebraic t est s T heorem: F a± ne-invariant and has k-local const raint ) F has f F (k)-local algebraic charact erizat ion. Pictorially For a± ne-invariant propert ies F k-query t est Theorem 1 k! k By defn. k-local alg. char. k-local charact erizat ion By defn. k-local const raint Theorem 2 k ! f F (k) X k! k (with Grigorescu) Pictorially For a± ne-invariant propert ies F k-query t est Theorem 1 k! k By defn. k-local alg. char. k-local charact erizat ion By defn. k-local const raint Theorem 2 k ! f F (k) X k! k (with Grigorescu) Part III: BLR (and our) analysis BLR Analysis: Outline ² Have f s.t . Pr x ;y [f (x) + f (y) 6 = f (x + y)] = ± < 2=9. Want t o show f close t o some g 2 F . ² De¯ne g(x) = most likely y f f (x + y) ¡ f (y)g. ² If f close t o F t hen g will be in F and close t o f . ² But if f not close? g may not even be uniquely de¯ned! ² St eps: ¡ St ep 0: Prove f close t o g ¡ Step 1: Prove “most likely” is overwhelming majority. ¡ St ep 2: Prove t hat g is in F . BLR Analysis: Step 0 ² De¯ne g(x) = most likely y f f (x + y) ¡ f (y)g. Claim: Pr x [f (x) 6 = g(x)] · 2± ¡ Let B = f xj Pr y [f (x) 6 = f (x + y)f (y)] ¸ ¡ Pr x ;y [linearity t est reject s jx 2 B ] ¸ ) Pr x [x 2 B ] · 2± ¡ If x 6 2 B t hen f (x) = g(x) 1 2 1g 2 Vot ex (y) BLR Analysis: Step 1 ² De¯ne g(x) = most likely y f f (x + y) ¡ f (y)g. ² Suppose for some x, 9 two equally likely values. Presumably, only one leads t o linear x, so which one? ² If we wish t o show g linear, t hen need t o rule out t his case. Lemma: 8 x, Pr y ;z [Vot ex (y) 6 = Vot ex (z))] · 4± Vot ex (y) BLR Analysis: Step 1 ² De¯ne g(x) = most likely y f f (x + y) ¡ f (y)g. ² Suppose for some x, 9 two equally likely values. Presumably, only one leads t o linear x, so which one? ² If we wish t o show g linear, t hen need t o rule out t his case. Lemma: 8 x, Pr y ;z [Vot ex (y) 6 = Vot ex (z))] · 4± Vot ex (y) BLR Analysis: Step 1 ² De¯ne g(x) = most likely y f f (x + y) ¡ f (y)g. Lemma: 8 x, Pr y ;z [Vot ex (y) 6 = Vot ex (z))] · 4± ? f (z) f (y) ¡ f (x + y) f (y + z) ¡ f (y + 2z) ¡ f (x + z) ¡ f (2y + z) f (x + 2y + 2z) Prob. Row/ column sum non-zero · ±. Vot ex (y) BLR Analysis: Step 1 ² De¯ne g(x) = most likely y f f (x + y) ¡ f (y)g. Lemma: 8 x, Pr y ;z [Vot ex (y) 6 = Vot ex (z))] · 4± ? f (z) f (y) ¡ f (x + y) f (y + z) ¡ f (y + 2z) ¡ f (x + z) ¡ f (2y + z) f (x + 2y + 2z) Prob. Row/ column sum non-zero · ±. BLR Analysis: Step 2 (Similar) Lemma: If ± < g(x) f (z) 1 , 20 t hen 8 x; y, g(x) + g(y) = g(x + y) g(y) ¡ g(x + y) f (y + z) ¡ f (y + 2z) ¡ f (x + z) ¡ f (2y + z) f (x + 2y + 2z) Prob. Row/ column sum non-zero · 4±. Our Analysis: Outline ² f s.t . Pr L [hf (L (x 1 ); : : : ; f (L (x k ))i 2 V ] = ± ¿ 1. ² De¯ne g(x) = ® t hat maximizes Pr f L j L ( x ) = x g [h®; f (L (x 2 )); : : : ; f (L (x k ))i 2 V ] 1 ² St eps: ¡ St ep 0: Prove f close t o g ¡ Step 1: Prove “most likely” is overwhelming majority. ¡ St ep 2: Prove t hat g is in F . Our Analysis: Outline ² f s.t . Pr L [hf (L (x 1 ); : : : ; f (L (x k ))i 2 V ] = ± ¿ 1. ² De¯ne g(x) = ® t hat maximizes Pr f L j L ( x ) = x g [h®; f (L (x 2 )); : : : ; f (L (x k ))i 2 V ] 1 Same as before ² St eps: ¡ St ep 0: Prove f close t o g ¡ Step 1: Prove “most likely” is overwhelming majority. ¡ St ep 2: Prove t hat g is in F . Vot ex (L ) Matrix ² De¯ne Magic? g(x) = ® t hat maximizes Pr f L j L ( x )= x g 1 [h®; f (L (x 2 )); : : : ; f (L (x k ))i 2 V ] Lemma: 8 x, Pr L ;K [Vot ex (L ) 6 = Vot ex (K ))] · 2(k ¡ 1)± x K (x 2 ) .. . K (x k ) L (x 2 ) ? ¢¢¢ L (x k ) Matrix Magic? x K (x 2 ) .. . L (x 2 ) ? ¢¢¢ L (x k ) K (x k ) ² Want marked rows t o be random const raint s. ² Suppose x 1 ; : : : ; x ` linearly independent ; and rest dependent on t hem. Matrix Magic? ` Fill with random entries Fill so as to make constraints x ` L (x 2 ) ¢¢¢ Linear algebra implies final columns are also constraints. L (x k ) K (x 2 ) .. . K (x k ) ² Suppose x 1 ; : : : ; x ` linearly independent ; and rest dependent on t hem. Matrix Magic? ` Fill with random entries Fill so as to make constraints x ` L (x 2 ) ¢¢¢ Linear algebra implies final columns are also constraints. L (x k ) K (x 2 ) .. . K (x k ) ² Suppose x 1 ; : : : ; x ` linearly independent ; and rest dependent on t hem. Conclusions Linear/Affine-invariant properties testable if they have local constraints. Gives clean generalization of linearity and lowdegree tests. Future work: What kind of invariances lead to testability (from characterizations)?