Coalition influence and resilient functions #
A Boolean function f on the cube ι → Bool is undetermined at x by the bits outside a
coalition Q (Cslib.BitString.IsUndetermined) if two strings that agree with x outside Q
take different values: after the bits outside Q are fixed to those of x, the coalition can
still decide the outcome. The influence of Q under a distribution D
(Cslib.BitString.coalitionInfluence) is the probability, for x drawn from D, that f is
undetermined at x by the bits outside Q. Under the uniform distribution this is the influence
I_Q(f) of Chattopadhyay–Zuckerman, Definition 1.6, which is the
coalitional influence of O'Donnell, Exercise 9.25; under a distribution D it is
the influence I_{Q,D}(f) of Chattopadhyay–Zuckerman, Definition 1.9.
Chattopadhyay and Zuckerman fix the bits outside Q by a distribution on those bits alone. Here
D is a distribution on all of the cube and the event ignores the bits in Q
(Cslib.BitString.isUndetermined_congr), so coalitionInfluence D Q f is their I_{Q,D'}(f) for
the marginal D' of D outside Q. The marginal of a k-wise independent distribution is again
k-wise independent, and a k-wise independent distribution outside Q extends by uniform bits on
Q, so the worst case over k-wise independent distributions is the same in both readings; this
equivalence is not formalized.
A function is (q, ε)-resilient (Cslib.BitString.IsResilient) if every coalition of q bits
has influence at most ε under the uniform distribution
Chattopadhyay–Zuckerman, Definition 1.7, and t-independent
(q, ε)-resilient (Cslib.BitString.IsKWiseResilient) if this holds under every t-wise
independent distribution Chattopadhyay–Zuckerman, Definition 1.10.
Chattopadhyay and Zuckerman define the largest influence I_q(f) and I_{q,t}(f) and bound it by
ε; the predicates state the bound for every coalition and distribution directly.
Equivalently, f is undetermined at x by the bits outside Q when its restriction fixing the
bits outside Q to those of x is not constant (Cslib.BitString.isUndetermined_iff_outside).
For a monotone function, f is undetermined at x by the bits outside Q exactly when setting
all of Q to true gives true and setting all of Q to false gives false
(Cslib.BitString.isUndetermined_iff_of_monotone). The influence of a single bit is the
influence of Cslib.BitString.influence (Cslib.BitString.coalitionInfluence_singleton)
O'Donnell, Exercise 9.25(a).
Main definitions #
Cslib.BitString.IsUndetermined:fis undetermined atxby the bits outsideQ.Cslib.BitString.coalitionInfluence: the influence of a coalition under a distribution.Cslib.BitString.IsResilient:(q, ε)-resilient functions.Cslib.BitString.IsKWiseResilient:t-independent(q, ε)-resilient functions.
Main results #
Cslib.BitString.isUndetermined_iff_outside: undetermined functions are those with a non-constant restriction.Cslib.BitString.isUndetermined_iff_of_monotone: undetermined monotone functions.Cslib.BitString.coalitionInfluence_uniformOfFintype: the uniform influence as an average.Cslib.BitString.coalitionInfluence_nonneg,Cslib.BitString.coalitionInfluence_le_one,Cslib.BitString.coalitionInfluence_mono: influences are probabilities, monotone in the coalition.Cslib.BitString.coalitionInfluence_singleton: the influence of a single bit.Cslib.BitString.isResilient_const: constant functions are resilient against every coalition.
References #
Undetermined functions #
The function f is undetermined at x by the bits outside the coalition Q if two strings
agreeing with x outside Q take different values under f
Chattopadhyay–Zuckerman, Definitions 1.6 and 1.9.
Instances For
A larger coalition leaves the function undetermined wherever a smaller one does.
A constant function is undetermined nowhere.
A function is undetermined at x by the bits outside Q exactly when its restriction fixing
the bits outside Q to those of x (Cslib.BitString.PartialAssignment.outside) is not
constant.
Undetermined monotone functions. A monotone function is undetermined at x by the bits
outside Q exactly when setting all of Q to true gives true and setting all of Q to
false gives false.
A function is undetermined at x by the bits other than i exactly when flipping bit i
of x changes its value.
Coalition influence #
The influence of the coalition Q on f under the distribution D: the probability, for
x drawn from D, that f is undetermined at x by the bits outside Q. This is I_{Q,D}(f)
Chattopadhyay–Zuckerman, Definition 1.9 for the marginal of D
outside Q; under the uniform distribution it is I_Q(f)
Chattopadhyay–Zuckerman, Definition 1.6, the coalitional influence
of O'Donnell, Exercise 9.25.
Instances For
Under the uniform distribution, the influence of a coalition is the fraction of the strings at which the function is undetermined by the bits outside it.
Influences are nonnegative O'Donnell, Exercise 9.25(b).
Influences are at most one O'Donnell, Exercise 9.25(b).
A larger coalition has at least the influence of a smaller one.
A coalition that leaves the function undetermined everywhere has influence one.
A coalition that leaves the function determined everywhere has influence zero.
The influence of a single bit under the uniform distribution is its influence
Cslib.BitString.influence on the function read as signs, the probability that flipping it
changes the value O'Donnell, Exercise 9.25(a).
Resilient functions #
A function is (q, ε)-resilient if every coalition of q bits has influence at most ε
under the uniform distribution
Chattopadhyay–Zuckerman, Definition 1.7. For q > card ι there is no
such coalition and the predicate holds vacuously, while their largest influence I_q(f) is
undefined.
Instances For
A function is t-independent (q, ε)-resilient if every coalition of q bits has
influence at most ε under every t-wise independent distribution
Chattopadhyay–Zuckerman, Definition 1.10, with D a distribution on
all bits. This is equivalent to their reading, with a distribution on the bits outside Q, by an
argument not formalized here (see the module docstring). For q > card ι there is no coalition of
q bits and the predicate holds vacuously, while their largest influence I_{q,t}(f) is
undefined.
Instances For
A t-independent resilient function is resilient, since the uniform distribution is t-wise
independent.
A constant function is t-independent (q, 0)-resilient for all t and q.
A constant function is (q, 0)-resilient for every q.