text
stringlengths
128
2.05k
Theorem 5 For every instance of the box problem with exogenous ordering, there is a [MATH] -thresholding policy whose expected net value is at least half that of Weitzman’s optimal search procedure (which endogenously selects the ordering of the boxes).
Proof. Lemma reduces the analysis of [MATH] -thresholding policies to a question about prophet inequalities. In particular, the expected net value of running a
[MATH] -thresholding policy is equal to the expected covered call value of the random element selected from the sequence [MATH] by a particular threshold stopping rule. Since Samuel-Cahn, ’s 1984 ) prophet inequality Theorem
above) implies that threshold stopping rules can always attain at least half the expectation of the maximum random variable in the sequence, it follows that there is a [MATH] -thresholding policy whose expected net value is at least half the expectation of the maximum covered call value.
ensures that the latter is an upper bound on the expected net value of Weitzman’s optimal search procedure. 4.3 An approximately optimal mechanism
Recall that for a half-infinite interval [MATH] or [MATH] the mechanism [MATH] is defined to be the single proposal mechanism whose eligible set consists of solutions [MATH] that are feasible and satisfy [MATH]
Theorem 6 There exists a choice of [MATH] such that the expected net value of mechanism [MATH] — i.e., the principal’s value for adopting the agent’s proposal, if adopted, minus combined cost of all the alternatives explored — is at least half of the expected net value the principal could achieve by performing the opti...
Proof. Convert the delegated search problem into a box problem with exogenous order, where the order is defined by sorting the solutions [MATH] in non-increasing order of the agent’s priority value [MATH] and the value [MATH] inside box [MATH] is defined to be
[MATH] if [MATH] turns out to be feasible, 0 otherwise. According to Theorem there exists a choice of [MATH] such that the [MATH] -thresholding policy with target set [MATH]
attains at least half the expected net value of the optimal search procedure. This thresholding policy goes through boxes in the given order, i.e. descending [MATH] , and opens only those with [MATH] , selecting the first one such that [MATH] . Note that among the boxes which the policy opens, the first one with [MATH]
is also the first one corresponding to a feasible [MATH] This is because an infeasible [MATH] has [MATH] hence [MATH] , whereas a feasible [MATH] has [MATH] hence [MATH]
Recall from Section 4.1 that the agent’s best response to mechanism [MATH] is to go through the elements of [MATH] in decreasing order of [MATH] , stopping and proposing the first one that is discovered to be feasible. This is exactly the behavior of the [MATH] -thresholding policy with target set [MATH] as derived in ...
[MATH] coupled with the agent’s best response behavior emulates the [MATH] -thresholding policy which attains at least half the expected net value of the optimal search procedure.
4.4 Limiting the number of samples In some cases the number of distinct potential solutions, [MATH] , may be prohibitively large, and the agent may only have the power to explore the feasibility of a limited number of them, [MATH] . In this case, if the principal were to conduct the search autonomously without delegati...
The key observation is the following lemma, which provides a useful upper bound on the value of running the optimal search procedure.
Lemma 4 In the box problem with [MATH] boxes, if the searcher is limited to open at most [MATH] boxes before claiming a prize, then the expected net value of any search procedure is bounded above by [MATH] where
[MATH] is the random set of boxes that the procedure opens. Proof. Sum up the inequality ( ) over all boxes and note that [MATH] for [MATH] to derive
[EQUATION] The lemma follows by noting that [MATH] because [MATH] Lemma 5 There exists a (non-random) set [MATH] of cardinality [MATH] such that [MATH] where [MATH] is the random set of solutions explored by the optimal search procedure subject to a contraint of exploring at most [MATH] solutions.
Proof. The problem of adaptively exploring a random set [MATH] of at most [MATH] solutions to maximize [MATH] is a special case of the stochastic monotone submodular function maximization problem studied by Asadpour and Nazerzadeh, ( 2016 in which the role of the monotone submodular function
[MATH] is played by the function [MATH] and role of the matroid constraint is played by the cardinality constraint that at most [MATH] elements may be probed. Theorem 1 of Asadpour and Nazerzadeh, ( 2016 which asserts that the adaptivity gap of stochastic monotone submodular maximization is [MATH] , specializes in the ...
Theorem 7 Consider delegated search in the binary model with a constraint that no more than [MATH] solutions can be examined for feasibility. There exists a mechanism that attains at least
[MATH] fraction of the expected net value of the optimal search procedure subject to the same limitation of examining at most [MATH] solutions.
Proof. According to Lemmas and , there is an [MATH] -element set [MATH] such that the optimal search procedure that is limited to explore only solutions in [MATH] is able to attain at least [MATH] fraction of the expected net value of the optimal search procedure that is limited to examine at most [MATH] solutions but ...
during its search. When the set of solutions is restricted to [MATH] , the constraint that at most [MATH] solutions can be examined becomes irrelevant since [MATH] only has [MATH] elements. Thus,
Theorem guarantees the existence of a delegated search mechanism that is at least half as good as the optimal search procedure limited to [MATH] , and is consequently at least
[MATH] times as good as the optimal search procedure limited to examine at most [MATH] solutions. Moreover, by applying the algorithm in
Asadpour and Nazerzadeh, ( 2016 used to prove Lemma , we can implement this policy in polynomial time with a loss of a further additive [MATH]
in the approximation ratio, thus obtaining a bound of [MATH] efficiently. Acknowledgements. This work was supported in part by NSF grants CCF-1512964, CCF-1740822, and SES-1741441, a grant from the MacArthur Foundation, and a Simons Investigator Award. The authors would like to thank Brendan Lucier, Jens Ludwig, Sendhi...
Appendix A Proofs of Prophet Inequalities In this appendix, we provide proofs of the three main prophet inequalities used in this work,
Theorems and A.1 Threshold stopping rules for independent distributions In this section we provide a proof of the following theorem from Section
3.1 . The theorem (stated in a different form) was originally proven by Samuel-Cahn, ( 1984 ; we provide a proof here for the sake of making the paper self-contained.
Theorem 8 There is a prophet inequality with factor [MATH] for [MATH] Proof. Consider any distribution [MATH] in [MATH] and let [MATH] denote independent random variables representing the [MATH] -coordinates of random samples from [MATH] , respectively. Note that the subscripts on the variables [MATH]
represent the distributions from which they were sampled, not necessarily the order in which they arrive, since their corresponding time coordinates [MATH] may not be in ascending order. However, this issue will be immaterial in the proof because our argument is insensitive to the arrival order of [MATH]
Let [MATH] and choose a threshold [MATH] defining two semi-infinite intervals [EQUATION] such that [MATH] and [MATH] Let [MATH] be the solution to the equation [MATH] and let [MATH] denote the random variable defined by the following sampling process: first choose the interval
[MATH] with probability [MATH] and [MATH] with probability [MATH] Then apply the threshold stopping rule with eligible set [MATH] to select an element [MATH] and let [MATH] denote the [MATH] -coordinate of that element. We will prove that [MATH] Since [MATH] is a convex combination of the expected value of applying the...
To compare [MATH] with [MATH] we reason as follows. First, we have the following easy upper bound on [MATH] [EQUATION] To put a lower bound on [MATH] , let [MATH] denote an indicator random variable for the event that [MATH] Note that this event, when it happens, implies that
[MATH] [EQUATION] where the last line follows because [MATH] again by our construction of the random set [MATH] . Finally, the theorem follows by combining ( ) with ( ).
A.2 Threshold stopping rules for i.i.d. atomless distributions This section provides a proof of the following lemma, which is equivalent to Theorem
from Section 3.1 Lemma 6 If [MATH] is a sequence of i.i.d. random variables, each sampled from an atomless distributions, and [MATH]
is a threshold stopping rule such that [MATH] then [EQUATION] Proof. Let [MATH] . We have [EQUATION] To compare the integrands at any specified [MATH] , we consider the cases [MATH] and [MATH] separately. (The case [MATH] is omitted because it contributes zero to both integrals.) When [MATH] the inequality [MATH]
is satisfied whenever [MATH] . Since we are assuming [MATH] are identically distributed, and the threshold stopping rule has the same behavior at every point in time, we have
[EQUATION] Hence the probability that the stopping rule does not stop at any time [MATH] is [EQUATION] and therefore for [MATH] [EQUATION]
Meanwhile, for [MATH] , the probability that the stopping rule stops at time [MATH] and selects an element of value greater than [MATH] is
[MATH] . Summing over [MATH] we have [EQUATION] Using the fact that [MATH] , we find that [EQUATION] Since we have proven that [MATH]
for all [MATH] , we may integrate over [MATH] and combine with ( )-( to obtain ( ). A.3 Oblivious stopping rules applied to i.i.d. atomless product distributions
Finally, in this subsection we furnish the proof of Theorem from Section 3.1 . The oblivious stopping rule that fulfills a prophet inequality with factor
[MATH] is more complicated than the threshold stopping rules analyzed earlier. It is defined by first solving a differential equation
[EQUATION] with initial condition [MATH] to define a function [MATH] mapping [MATH] to [MATH] . The constant [MATH] is chosen so that [MATH] . Since the differential equation ( ) implies
[EQUATION] for all [MATH] such that [MATH] is defined, the boundary condition [MATH] requires the equation [EQUATION] to be satisfied, and we treat this equation as the definition of [MATH]
Denoting the cumulative distribution functions of [MATH] and [MATH] by [MATH] and [MATH] , respectively, we will be analyzing the oblivious stopping rule which selects the earliest [MATH]
satisfying [EQUATION] We will prove that this rule satisfies a prophet inequality with a specific factor [MATH] to be determined later. (Equation ( 26 ) below defines [MATH] .) Let [MATH] denote the
[MATH] -coordinate of the element [MATH] selected by the oblivious stopping rule, and let [MATH] denote the random variable [MATH] . As in the proof of Lemma
, we will make use of the equations [EQUATION] which reduces the task of proving a prophet inequality with factor [MATH] to the task of proving
[EQUATION] At this point a couple of observations will slightly simplify the analysis. If we change coordinates to replace [MATH] with [MATH] this has no effect on the behavior of the stopping rule, and of course if has no effect on [MATH] , so we are free to adopt this reparameterization and assume henceforth that [MA...
is uniformly distributed in [MATH] . In particular, this means the stopping rule simplifies to choosing the first [MATH] such that
[MATH] The next simplification comes from introducing the variable [MATH] and writing the event [MATH] in the form [MATH] The probability of this event is
[EQUATION] The function [MATH] will be appearing frequently throughout this calculation so we will assign it a name: [MATH] . (This name is inspired by the fact that [MATH]
as [MATH] for any fixed [MATH] , so [MATH] is a degree- [MATH] polynomial approximation to the exponential function.) We have derived that for [MATH]
[EQUATION] Now what about [MATH] We can calculate this probability by integrating, with respect to [MATH] the probability that the stopping rule selects a point in [MATH] In order for [MATH] to be this point, the following things must happen.
1. [MATH] . This has probability [MATH] 2. [MATH] 3. [MATH] 4. None of the points [MATH] , for [MATH] satisfy the stopping condition with [MATH]
The conjunction of the first three events has probability [MATH] Let [MATH] so that [MATH] is the probability that a specific point [MATH] satisfies [MATH] along with the stopping condition [MATH] The fourth event above has probability
[MATH] Hence [EQUATION] We are left with proving that for a suitable choice of [MATH] , the inequality [EQUATION] holds for all [MATH] At this point a useful identity comes to our aid.
Lemma 7 The functions [MATH] and [MATH] defined above satisfy the equation [EQUATION] for every [MATH] Proof. The proof is a brief but fairly opaque calculation using the differential equation ( ) that defines the function [MATH] . First, observe that [MATH]
satisfies the equation [MATH] because if one differentiates both sides of ( one obtains [EQUATION] Second, observe that the function [MATH] is constant: its derivative satisfies
[EQUATION] Since [MATH] , we may conclude that [MATH] for all [MATH] , which means [EQUATION] Integrating both sides of ( 13 ) we obtain
[EQUATION] Let [MATH] . The function [MATH] is increasing in [MATH] so [MATH] equals [MATH] for [MATH] , and it equals [MATH] for [MATH] Hence,
[EQUATION] which matches the equation claimed in the lemma statement. Compared to the equation ( 12 asserted by Lemma the inequality ( 11 ) that we need to prove differs in three important respects.
1. The [MATH] function appearing on both sides of ( 11 ) has been replaced by the exponential function in ( 12 ). 2. The left side of ( 12 ) has an additional [MATH] term that does not correspond to any term on the left side of ( 11 ).
3. The constant on the left side of ( 12 is [MATH] instead of [MATH] . In fact, we will be choosing [MATH] to be very close to [MATH] but slightly smaller. The difference between [MATH]
and [MATH] will be necessary to compensate for error terms arising from substituting [MATH] for [MATH] and from the additional [MATH] term on the left side.
The process of accounting for these differences between ( 11 ) and ( 12 begins with the following technical lemma that provides an upper bound on the value of
[MATH] for [MATH] Lemma 8 If [MATH] then [MATH] for all [MATH] Proof. First, recall from equation ( 10 ) that [MATH] As a function of [MATH] , the value of
[MATH] is monotonically decreasing, so the fact that [MATH] (as can be verified numerically) implies that [MATH] . Since [MATH] is increasing in [MATH] , this also implies
[MATH] for all [MATH] Now, as [MATH] is increasing in [MATH] , to prove the inequality [MATH] it suffices to show that [MATH] . Recalling equation ( 13 ), we have
[EQUATION] where the final inequality used the fact that [MATH] Taking the logarithm of both sides, we find that [MATH] as desired.
The next lemma bounds the multiplicative error in using [MATH] to approximate [MATH] Lemma 9 For [MATH] and [MATH] [EQUATION] Proof.
Recalling the definition of the [MATH] function and using the inquality [MATH] , we have [MATH] which furnishes the first inequality in the lemma statement. To prove the second inequality, note that
[EQUATION] where the first and last inequalities made use of the fact that [MATH] , and the second inequality depends on the fact that [MATH] , so that the exponent [MATH] is non-positive.
The following simple inequality will be used twice in the sequel. Lemma 10 For all [MATH] [EQUATION] Proof. The function [MATH] attains its minimum value at [MATH] , where its value is strictly positive. Since this function is the derivative of [MATH] , we may conclude that the latter function is strictly increasing, a...
Combining the steps above, we can obtain a useful lower bound on the right side of ( 11 ). Lemma 11 For [MATH] [EQUATION] Proof.
Lemma justifies applying Lemma to the factor [MATH] in the integrand, which implies [EQUATION] where the second line is derived from Lemma
Now, the function [MATH] is an increasing function of [MATH] in the range [MATH] , as can be verified by computing that the derivative is [MATH]
and observing that all three factors are strictly positive. Hence, for [MATH] we have [EQUATION] where the second inequality makes use of Lemma
10 Rearranging terms in ( 19 ) we obtain [EQUATION] and substituting this into the right side of ( 18 ) we obtain [EQUATION] The observation that [MATH]
concludes the proof. Next we need a lemma accounting for the discrepancy between the function [MATH] appearing on the left side of ( 11 ) and the function [MATH]
in ( 17 ). Lemma 12 For [MATH] [EQUATION] Proof. If [MATH] then [MATH] so both [MATH] and [MATH] are in the interval [MATH] , and the lemma follows easily in this case. Assume henceforth that [MATH]
For [MATH] the inequality [MATH] implies that [MATH] Integrating from [MATH] to [MATH] , we derive [EQUATION] Multiplying both sides by [MATH] we obtain
[EQUATION] Now exponentiate both sides of ( 24 ). [EQUATION] where the final inequality used the fact that [MATH] Lemma 10 implies that [MATH] so
[EQUATION] and the lemma follows upon rearranging terms. Combining equation ( 17 ) with Lemma 12 we obtain the inequality [EQUATION]
which proves ( 11 ) with [EQUATION] since [MATH] Finally, let us bound the difference between [MATH] and the constant [MATH] defined by the equation
[EQUATION] Perform the substitution [MATH] to deduce that [EQUATION] Set [MATH] , which implies [MATH] Noting that [EQUATION] and that [MATH] is a monotonically decreasing function of [MATH] we deduce that [MATH] which implies
[MATH] Hence [MATH] Appendix B Tightness of bounds In this section we prove that the bounds in each part of Theorem are tight under their respective assumptions.
Proposition 1 For any [MATH] there exists a joint distribution of the agent’s and principal’s utilities such that the principal’s expected utility is always less than [MATH]
regardless of the choice of mechanism. Proof. For a parameter [MATH] define the joint distribution of [MATH] to be a mixture of uniform distributions on two rectangles. The first is the rectangle
[MATH] the second is the rectangle [MATH] The pair [MATH] is drawn from the uniform distribution on [MATH] with probability [MATH] and from the uniform distribution on [MATH] with probability [MATH] When we draw [MATH] i.i.d. samples [MATH]
from this joint distribution, the random variable [MATH] is uniformly distributed in [MATH] with probability [MATH] , and it is uniformly distributed in [MATH] with probability
[MATH] . Therefore, letting [MATH] [EQUATION] Now we bound the principal’s expected utility under any mechanism. By Lemma it is enough to consider single proposal mechanisms. Let [MATH] denote the eligible set of the mechanism, and let [MATH] denote the probability that a random sample from the distribution of [MATH] b...
then the agent will propose an element of [MATH] . This event [MATH] happens with probability [MATH] and the principal’s expected utility conditional on [MATH]
is bounded above by [MATH] . The other case in which the principal gets non-zero utility is if all [MATH] samples belong to the complement of [MATH] , but they don’t all belong to the complement of [MATH] . The probability of this event [MATH] is
[MATH] and principal’s expected utility conditional on [MATH] is bounded above by [MATH] . Hence, denoting the principal’s utility by [MATH]