## Small-Bias Probability Spaces: Efficient Constructions and Applications (1993)

### Cached

### Download Links

- [www.wisdom.weizmann.ac.il]
- [www.lirmm.fr]
- DBLP

### Other Repositories/Bibliography

Venue: | SIAM J. Comput |

Citations: | 263 - 15 self |

### BibTeX

@ARTICLE{Naor93small-biasprobability,

author = {Joseph Naor and Moni Naor},

title = {Small-Bias Probability Spaces: Efficient Constructions and Applications},

journal = {SIAM J. Comput},

year = {1993},

volume = {22},

pages = {838--856}

}

### Years of Citing Articles

### OpenURL

### Abstract

We show how to efficiently construct a small probability space on n binary random variables such that for every subset, its parity is either zero or one with "almost" equal probability. They are called ffl-biased random variables. The number of random bits needed to generate the random variables is O(log n + log 1 ffl ). Thus, if ffl is polynomially small, then the size of the sample space is also polynomial. Random variables that are ffl-biased can be used to construct "almost" k-wise independent random variables where ffl is a function of k. These probability spaces have various applications: 1. Derandomization of algorithms: many randomized algorithms that require only k- wise independence of their random bits (where k is bounded by O(log n)), can be derandomized by using ffl-biased random variables. 2. Reducing the number of random bits required by certain randomized algorithms, e.g., verification of matrix multiplication. 3. Exhaustive testing of combinatorial circui...

### Citations

9352 | Introduction to Algorithms - Cormen, Leiserson, et al. - 1989 |

3316 | A method for obtaining digital signatures and public-key cryptosystems
- Rivest, Shamir, et al.
- 1978
(Show Context)
Citation Context ... that a and N are integers such that (N; a) = 1. The above procedure can be applied to verify n equalities of the type x a i = y i mod N . This may be used for instance to check n given RSA equations =-=[46]-=- and thus amortize the cost of verifying signatures. 20 8 Vector sets for exhaustive testing A problem that has received much attention in the fault diagnostic literature is that of generating a small... |

2282 |
An introduction to probability theory and its applications
- Feller
- 1971
(Show Context)
Citation Context ...f the sample space needed for constructing k-wise ffi -dependent random variables is O(k + log log n + log 1 ffi ). 5 A moment inequality Basic tools in probabilistic analysis are moment inequalities =-=[25]-=- that bound the deviation of a random variable from its expected value. More specifically, let y be a random variable such that 14 E[y] = 0, then, Prob[jyjs]sE[jyj k ] k Let b 1 ; : : : ; b n be rando... |

2177 |
The Theory of Error-Correcting Codes
- MacWilliams, Sloane
- 1977
(Show Context)
Citation Context ...the distance between any two codewords is at least fi \Delta m. The weight of a codeword is the number of non-zero symbols that it contains. For more details on linear codes the reader is referred to =-=[40]-=-. 3 Constructing the probability distribution In this section we show how to construct a probability space\Omega of f0; 1g random variables x = x 1 ; : : : ; x n which are ffl-biased. The cardinality ... |

1764 | Probabilistic Method
- Alon, Spencer
- 1992
(Show Context)
Citation Context ... ). The discrepancy of a subset S i , \Delta(S i ; x), with respect to x, is defined as P j2S i x j . The discrepancy of the set family S is defined as, \Delta(S; x) = max S i \Delta(S i ; x) Spencer =-=[50, 9]-=- has shown that for each family S, there exists a 2-coloring x such that \Delta(S; x)s6 p n. This result is the best possible up to constant factors but it has the drawback of being non-constructive, ... |

658 | Communication Complexity
- Kushilevitz, Nisan
- 1997
(Show Context)
Citation Context ...small bias probability spaces have been discovered. Alon [4] used them to obtain an NC 1 algorithm for the parallel version of Beck's algorithm for the Lovasz local lemma. Feder, Kushilevitz and Naor =-=[24]-=- have applied them to amortize the communication complexity of equality. Blum et al. [15] used the hash function construction to authenticate memories. Kushilevitz and Mansour [31] used small bias pro... |

391 | A simple parallel algorithm for the maximal independent set problem
- Luby
- 1985
(Show Context)
Citation Context ...o the problem of reducing the number of random bits in a randomized algorithm is to show that limited independence of the random variables suffices to assure a high probability of success. (See e.g., =-=[5, 36, 38]-=-). However, in our case, it is not clear from the proof of Proposition 3.1 how many vectors r remain distinguishers with respect to the vector v when the entries of R are not completely independent. M... |

326 | Designing programs that check their work
- Blum, Kannan
- 1995
(Show Context)
Citation Context ...r A = B. We would like a single pass on the elements of A and B and the amount of memory should be O(log n). Here again a method that achieves constant probability error is known (see Blum and Kannan =-=[16]-=- for a description). We will show that it is possible to achieve any ffl error while using only O(log n+ log 1 ffl ) bits of memory. Let v A and v B denote the incidence vectors of A and B. We can thi... |

289 |
Probabilistic construction of deterministic algorithms: approximating packing integer programs
- Raghavan
- 1988
(Show Context)
Citation Context ...ties (sequentially), Spencer devised a polynomial-time deterministic algorithm which guarantees a 2-coloring x such that \Delta(S; x) = O( p n log n). This was improved to O( p ffi log n) by Raghavan =-=[45]-=-. For parallel algorithms we cannot guarantee as small a discrepancy as in the sequential case. However, we can come arbitrarily close to the sequential 16 bounds by computing a -good coloring. Defini... |

285 | Hardness vs Randomness - Nisan, Wigderson - 1994 |

280 | Simple constructions of almost kwise independent random variables
- Alon, Goldreich, et al.
- 1990
(Show Context)
Citation Context ...constructed small bias probability spaces for nonbinary random variables. Even et al. [23] constructed ffi-dependent non-uniform probability spaces based on small bias probability spaces. Alon et al. =-=[8]-=- provided simple and different constructions for small bias probability spaces. Their constructions are based on quadratic characters and linear feedback shift registers. Schulman [48] constructed eff... |

224 |
A fast and simple randomized parallel algorithm for the maximal independent set problem
- Alon, Babai, et al.
- 1986
(Show Context)
Citation Context ...o the problem of reducing the number of random bits in a randomized algorithm is to show that limited independence of the random variables suffices to assure a high probability of success. (See e.g., =-=[5, 36, 38]-=-). However, in our case, it is not clear from the proof of Proposition 3.1 how many vectors r remain distinguishers with respect to the vector v when the entries of R are not completely independent. M... |

191 | Learning decision trees using the Fourier spectrum
- Kushilevitz, Mansour
- 1991
(Show Context)
Citation Context ...ushilevitz and Naor [24] have applied them to amortize the communication complexity of equality. Blum et al. [15] used the hash function construction to authenticate memories. Kushilevitz and Mansour =-=[31]-=- used small bias probability spaces to derandomize their decision tree learning algorithm. Alon et al. [7] have used the polylogarithmic size construction of log log n-wise 1= log n-bias probablity sp... |

188 | How to recycle random bits
- Impagliazzo, Zuckerman
- 1989
(Show Context)
Citation Context ...between an algorithm that is completely deterministic, and a randomized algorithm that consumes many bits. Examples of previous work in reducing the number of random bits in randomized algorithms are =-=[2, 11, 20, 28, 32, 33, 43, 47, 49]-=-. Karloff and Raghavan [33] for example, studied several randomized algorithms for selection and sorting and showed that they can be run successfully when only O(log n) random bits are available. In S... |

157 | R.: The bit extraction problem or t-resilient functions
- Chor, Goldreich, et al.
- 1985
(Show Context)
Citation Context ... was to require only limited independence among the random variables (as opposed to full independence). Our approach is different; it is based on the equivalence of the following two conditions: (See =-=[18]-=- and [52]). 1. The random variables are independent and for all i, Prob[x i = 0] = Prob[x i = 1]. 2. For every subset S ` f1; : : : ; ng, it is equally likely that the parity of the subset (i.e., the ... |

121 | Construction of asymptotically good, low-rate error-correcting codes through pseudo-random graphs
- Alon, Bruck, et al.
- 1992
(Show Context)
Citation Context ...re more complicated and require exponentiation. Using our techniques we can actually enhance the error-correction of a linear code without decreasing the rate by much. This is further investigated in =-=[6]-=-. 3.2 Sampling the family F The problem of obtaining the vectors r 1 ; : : : ; r l in Stage 2 with the desired property can be abstracted in the following way. Suppose that there is a universe, and we... |

106 | Checking the correctness of memories
- Blum, Evans, et al.
- 1994
(Show Context)
Citation Context ...algorithm for the parallel version of Beck's algorithm for the Lovasz local lemma. Feder, Kushilevitz and Naor [24] have applied them to amortize the communication complexity of equality. Blum et al. =-=[15]-=- used the hash function construction to authenticate memories. Kushilevitz and Mansour [31] used small bias probability spaces to derandomize their decision tree learning algorithm. Alon et al. [7] ha... |

101 |
Explicit construction of linear sized superconcentrators
- Gabber, Galil
- 1981
(Show Context)
Citation Context ...t least one of the l samples is a member of S is at least 1 \Gamma /p 2jsjs0 ! l : Constructions for regular graphs of constant degree such that the second eigenvalue is bounded away froms0 are known =-=[27, 29, 37]-=-. For degree regular graphs of degree d,s0 = d and the value ofscan be almost 2 p d. For a given expander G, let ff G = 1 2 /s0 ! 2 and let l = log 1 ffl log ip 2j js0 j : In our case, fi, the precent... |

100 | On the power of two-point based sampling
- Chor, Goldreich
- 1989
(Show Context)
Citation Context ...ies of r are chosen pairwise independently, no distinguisher will be generated. Though limited independence is not sufficient for our purposes, we make use of it in two ways: one is that suggested in =-=[17]-=- that if we sample a universe looking for elements of some fixed subset, and if the expected number of elements we hit is greater than 1, then by making our choices only pairwise independent, we are n... |

90 | Dispersers, deterministic amplification, and weak random sources
- Cohen, Wigderson
- 1989
(Show Context)
Citation Context ...between an algorithm that is completely deterministic, and a randomized algorithm that consumes many bits. Examples of previous work in reducing the number of random bits in randomized algorithms are =-=[2, 11, 20, 28, 32, 33, 43, 47, 49]-=-. Karloff and Raghavan [33] for example, studied several randomized algorithms for selection and sorting and showed that they can be run successfully when only O(log n) random bits are available. In S... |

84 | A Fast Parallel Algorithm for the Maximal Independent Set Problem
- Karp, Wigderson
- 1985
(Show Context)
Citation Context ...o the problem of reducing the number of random bits in a randomized algorithm is to show that limited independence of the random variables suffices to assure a high probability of success. (See e.g., =-=[5, 36, 38]-=-). However, in our case, it is not clear from the proof of Proposition 3.1 how many vectors r remain distinguishers with respect to the vector v when the entries of R are not completely independent. M... |

72 |
Explicit expanders and the Ramanujan conjectures
- Lubotzky, Phillips, et al.
- 1986
(Show Context)
Citation Context ...t least one of the l samples is a member of S is at least 1 \Gamma /p 2jsjs0 ! l : Constructions for regular graphs of constant degree such that the second eigenvalue is bounded away froms0 are known =-=[27, 29, 37]-=-. For degree regular graphs of degree d,s0 = d and the value ofscan be almost 2 p d. For a given expander G, let ff G = 1 2 /s0 ! 2 and let l = log 1 ffl log ip 2j js0 j : In our case, fi, the precent... |

62 | A parallel algorithmic version of the local lemma
- Alon
- 1991
(Show Context)
Citation Context ...t, where addition is bitwise XoR. 10 Further results Since the preliminary version of this paper appeared in [42], several new applications of small bias probability spaces have been discovered. Alon =-=[4]-=- used them to obtain an NC 1 algorithm for the parallel version of Beck's algorithm for the Lovasz local lemma. Feder, Kushilevitz and Naor [24] have applied them to amortize the communication complex... |

59 |
Lectures on the Probabilistic Method
- Spencer, Ten
- 1987
(Show Context)
Citation Context ... ). The discrepancy of a subset S i , \Delta(S i ; x), with respect to x, is defined as P j2S i x j . The discrepancy of the set family S is defined as, \Delta(S; x) = max S i \Delta(S i ; x) Spencer =-=[50, 9]-=- has shown that for each family S, there exists a 2-coloring x such that \Delta(S; x)s6 p n. This result is the best possible up to constant factors but it has the drawback of being non-constructive, ... |

54 | Fast probabilistic algorithms - Freivalds - 1979 |

51 | Fast learning of k-term DNF formulas with queries
- Blum, Rudich
- 1995
(Show Context)
Citation Context ...struction of log log n-wise 1= log n-bias probablity space in order to derandomize an algorithm for all pairs shortest path whose running time is almost that of matrix multiplication. Blum and Rudich =-=[14]-=- have used the construction of (n; k)-universal sets of Section 8 in order to derandomize a k-term DNF polynomial time learning algorithm, for k which is logarithmic in the number of variables. Azar, ... |

51 | The probabilistic method yields deterministic parallel algorithms
- Motwani, Naor, et al.
- 1989
(Show Context)
Citation Context ... space. We exemplify this by providing a polynomial size sample space for the set balancing problem; this also yields an NC 1 algorithm. A previous approach to derandomizing the set balancing problem =-=[13, 39]-=- was to first construct an n O(logn) sample space, and then conduct a binary search for a good point. As a result, their time bounds are worse. Another problem we address is finding a heavy codeword i... |

47 |
Families of k-independent sets
- Kleitman, Spencer
- 1973
(Show Context)
Citation Context ...llection of n subsets of a ground set which is as small as possible such that the intersection of any k subsets or their complement is non-empty. This was called k-independent by Kleitman and Spencer =-=[35]-=-. The best known bounds for constructing (n; k) universal sets are given in [51] and [3]. The connection between k-wise ffi-dependent probability spaces and small (n; k)-universal sets is made in the ... |

41 |
adversaries and computation
- Vazirani, “Randomness
- 1986
(Show Context)
Citation Context ...equire only limited independence among the random variables (as opposed to full independence). Our approach is different; it is based on the equivalence of the following two conditions: (See [18] and =-=[52]-=-). 1. The random variables are independent and for all i, Prob[x i = 0] = Prob[x i = 1]. 2. For every subset S ` f1; : : : ; ng, it is equally likely that the parity of the subset (i.e., the number of... |

38 |
Expanders obtained from affine transformations
- Jimbo, Maruoka
- 1985
(Show Context)
Citation Context ...t least one of the l samples is a member of S is at least 1 \Gamma /p 2jsjs0 ! l : Constructions for regular graphs of constant degree such that the second eigenvalue is bounded away froms0 are known =-=[27, 29, 37]-=-. For degree regular graphs of degree d,s0 = d and the value ofscan be almost 2 p d. For a given expander G, let ff G = 1 2 /s0 ! 2 and let l = log 1 ffl log ip 2j js0 j : In our case, fi, the precent... |

37 |
Vector sets for exhaustive testing of logic circuits
- Seroussi, Bshouty
- 1988
(Show Context)
Citation Context ...nerating a small set of vectors, T ae f0; 1g n , such that for any k-subset of indices S = fi 1 ; i 2 ; : : : i k g, the projection of T on the indices S contains all possible 2 k configurations. See =-=[51]-=- for a bibliography on the problem. Such a set of vectors is called (n; k)-universal. The motivation for this problem is that such a test set allows exhaustive testing of a circuit where each componen... |

35 |
Deterministic simulation
- Ajtai, Komlos, et al.
- 1987
(Show Context)
Citation Context ...mber of the desired set. A straightforward implementation would require kl random bits. Several papers have addressed the question of achieving the probability error while requiring fewer random bits =-=[1, 20, 28, 32, 47, 49]-=-. We consider the method of [1] which is used by [20, 28]: For a graph G = (V; E), consider any 1-1 correspondence f : V ! f0; 1g k , the different assignments to the random bits of the sampling algor... |

32 | Approximations of general independent distributions
- Even, Goldreich, et al.
- 1992
(Show Context)
Citation Context ...earning algorithm, for k which is logarithmic in the number of variables. Azar, Motwani and Naor [10] defined and constructed small bias probability spaces for nonbinary random variables. Even et al. =-=[23]-=- constructed ffi-dependent non-uniform probability spaces based on small bias probability spaces. Alon et al. [8] provided simple and different constructions for small bias probability spaces. Their c... |

32 |
Expanders, randomness, or time versus space
- Sipser
- 1988
(Show Context)
Citation Context ...between an algorithm that is completely deterministic, and a randomized algorithm that consumes many bits. Examples of previous work in reducing the number of random bits in randomized algorithms are =-=[2, 11, 20, 28, 32, 33, 43, 47, 49]-=-. Karloff and Raghavan [33] for example, studied several randomized algorithms for selection and sorting and showed that they can be run successfully when only O(log n) random bits are available. In S... |

31 |
Realistic analysis of some randomized algorithms
- Bach
- 1991
(Show Context)
Citation Context |

29 |
Simulating (log c n)-wise independence in NC
- Berger, Rompel
- 1991
(Show Context)
Citation Context ... space. We exemplify this by providing a polynomial size sample space for the set balancing problem; this also yields an NC 1 algorithm. A previous approach to derandomizing the set balancing problem =-=[13, 39]-=- was to first construct an n O(logn) sample space, and then conduct a binary search for a good point. As a result, their time bounds are worse. Another problem we address is finding a heavy codeword i... |

28 |
Explicit construction of exponential sized families of k-independent sets
- Alon
- 1986
(Show Context)
Citation Context ...ction of any k subsets or their complement is non-empty. This was called k-independent by Kleitman and Spencer [35]. The best known bounds for constructing (n; k) universal sets are given in [51] and =-=[3]-=-. The connection between k-wise ffi-dependent probability spaces and small (n; k)-universal sets is made in the next proposition. Proposition 8.1 If\Omega is a k-wise ffi -dependent probability space ... |

28 |
Randomized algorithms and pseudorandom numbers
- Karlo, Raghavan
- 1993
(Show Context)
Citation Context |

28 |
Hardness vs
- Nisan, Wigderson
- 1994
(Show Context)
Citation Context |

25 | A Time-Space Trade-Off - Pippenger - 1978 |

24 |
Deterministic Simulation of Probabilistic Constant Depth Circuits
- Ajtai, Wigderson
- 1985
(Show Context)
Citation Context |

19 |
Efficient parallel algorithms for edge coloring problems
- Karloff, Shmoys
- 1987
(Show Context)
Citation Context ...rithm to find a -good coloring for any S: pick a random x uniformly at random from f0; 1g n . It turns out that for any S, the 2-coloring is -good with sufficiently high probability. (See for example =-=[34]-=-). This algorithm was derandomized by [13, 39] by proving that log n-wise independence suffices, and then showing how to conduct a binary search in a sample space of size n O(logn) . (The method of co... |

15 |
Sample spaces uniform on neighborhoods
- Schulman
- 1992
(Show Context)
Citation Context ...ces. Alon et al. [8] provided simple and different constructions for small bias probability spaces. Their constructions are based on quadratic characters and linear feedback shift registers. Schulman =-=[48]-=- constructed efficiently probability spaces with known dependencies. Acknowledgement We would like to thank Shuki Bruck for his contribution to Section 3.1.2. We would also like to thank Noga Alon, Yi... |

12 | Constructing Ramsey graphs from small probability spaces - Naor - 1992 |

8 |
On Using Deterministic Functions to Reduce Randomness in Probabilistic Algorithms
- Santha
- 1987
(Show Context)
Citation Context |

8 | The Theory ofError-Correcting Codes - MACWILLIAMS, SLOANE - 1977 |

7 |
On the randomness complexity of algorithms
- Peralta
- 1990
(Show Context)
Citation Context ...ications are discussed in Section 9. In Section 10 we briefly survey recent papers that have used the constructions described in this paper since it appeared in [42]. Independent of our work, Peralta =-=[44]-=- has considered ffl-bias probability spaces as well, and showed some applications to number theoretic algorithms. His construction is based on quadratic residues and Weil's Theorem. 2 Preliminaries an... |

3 |
Fast probabilistic algorithms, Springer Verlag
- Freivalds
- 1979
(Show Context)
Citation Context ...nd C, how 2 can one verify that A \Delta B = C without resorting to matrix multiplication. We show how to do that in O(n 2 ) time using O(log n) random bits, thus improving on a previous algorithm of =-=[26]-=- that required O(n) random bits. We next show how Adi Shamir's Boolean matrix multiplication algorithm can be implemented by using O(log n) random bits without an increase in the probability of error.... |

2 |
Witnesses for matrix multiplication and for shortest paths
- Alon, Galil, et al.
- 1992
(Show Context)
Citation Context .... [15] used the hash function construction to authenticate memories. Kushilevitz and Mansour [31] used small bias probability spaces to derandomize their decision tree learning algorithm. Alon et al. =-=[7]-=- have used the polylogarithmic size construction of log log n-wise 1= log n-bias probablity space in order to derandomize an algorithm for all pairs shortest path whose running time is almost that of ... |

2 |
A class of asymptotically good algebraic codes
- JUSTESEN
- 1972
(Show Context)
Citation Context ... maps f0; 1g n words into f0; 1g m codewords and its minimum distance is fi \Delta m. Linear codes for which m = O(n) and fi is some constant exist and are constructible. (For example, Justesen codes =-=[30]-=-, [40]). Let G be the generator matrix of this code. For any non-zero f0; 1g n vector v, v \Delta G contains at least fim non-zero entries. (By the above proposition). Hence, if we choose a column in ... |