Willy Wonka Mechanisms

Thomas Archbold Bart de Keijzer Carmine Ventre

King’s College London

United Kingdom
{thomas.archbold,bart.de keijzer,carmine.ventre}@kcl.ac.uk

February 13, 2024

arXiv:2402.08314v1 [cs.GT] 13 Feb 2024

Abstract
Bounded rationality in mechanism design aims to ensure incentive-compatibility for agents who
are cognitively limited. These agents lack the contingent reasoning skills that traditional mecha-
nism design assumes, and depending on how these cognitive limitations are modelled this alters the
class of incentive-compatible mechanisms. In this work we design mechanisms without any “obvi-
ous” manipulations for several auction settings that aim to either maximise revenue or minimise
the compensation paid to the agents. A mechanism without obvious manipulations is said to be
not obviously manipulable (NOM), and assumes agents act truthfully as long as the maximum and
minimum utilities from doing so are no worse than the maximum and minimum utilities from lying,
with the extremes taken over all possible actions of the other agents. We exploit the definition of
NOM by introducing the concept of golden tickets and wooden spoons, which designate bid profiles
ensuring the mechanism’s incentive-compatibility for each agent. We then characterise these “Willy
Wonka” mechanisms, and by carefully choosing the golden tickets and wooden spoons we use this to
design revenue-maximising auctions and frugal procurement auctions.

1 Introduction
In algorithmic mechanism design we are tasked with designing systems that elicit some information from
a set of agents in order to return an outcome from some feasible set. Each agent holds some piece of
private information reflecting the true state of the world, and agents are assumed to be selfish, meaning
that they may lie about their piece of information (also known as their type) if it is beneficial to do so.
The mechanisms we design must therefore be robust against the selfish actions of the agents and carefully
consider the agents’ incentives. These incentives are modelled by solution concepts. Strategyproofness,
otherwise known as dominant-strategy incentive-compatibility, is a solution concept which stipulates that
for each agent it is a dominant strategy to report her true type to the mechanism over any deviation.
This is desirable from a theoretical standpoint since if a mechanism is strategyproof then it guarantees
that regardless of the joint action profile of the other players an agent will always maximise her utility
by reporting her type truthfully rather than by lying to the mechanism. Strategyproofness is a viable
solution concept only when agents are perfectly rational: in order to conclude that acting truthfully is
indeed optimal an agent must have a detailed knowledge of the mechanism and correctly reason about
all possible states that may result from her actions by considering the actions of the other agents. This
may place too high a cognitive burden on the agents to be useful in practical mechanisms.
Bounded rationality has seen growing interest in the mechanism design literature and considers agents
who are cognitively limited. [Li, 2017] introduced a strengthening of strategyproofness known as “obvious
strategyproofness” which assumes that agents will only tell the truth if it is obvious to do so. Roughly a
mechanism is obviously strategyproof (OSP) if the minimum utility that an agent can achieve from telling
the truth is no worse than the maximum utility that she can achieve from reporting her type dishonestly.
Troyan and Morrill take a more optimistic approach towards cognitively-limited agents, motivated by
evidence [Pathak and Sönmez, 2008, Dur et al., 2018] gathered from mechanisms in practice for school-
choice and two-sided matching that agents in manipulable (i.e., non-strategyproof) mechanisms might
fail to recognise when it is beneficial to lie, and introduce a weaker version of incentive compatibility
for these agents known as “non-obvious manipulability” [Troyan and Morrill, 2020]. Under this solution
concept it is assumed that agents will only lie if it is obvious, meaning that the maximum (respectively,

1
minimum) utility resulting from a lie is strictly greater than the maximum (respectively, minimum)
utility resulting from reporting truthfully to the mechanism; otherwise, agents are assumed to report
their types truthfully. The extrema of the agents’ utility functions are taken over all possible joint profiles
of the other agents. When a mechanism has no such “obvious manipulations” then it is said to be not
obviously manipulable (NOM).
This definition is motivated by empirical evidence [Pathak and Sönmez, 2008, Dur et al., 2018] from
school choice and two-sided matching that people facing manipulable (i.e., non-strategyproof) mecha-
nisms fail to recognise when to lie. Students participating in the Boston Mechanism, for example, can
guarantee a spot at their second-choice school by misreporting her preferences, whereas hospitals may
end up being allocated a doctor it finds undesirable when attempting to manipulate the Deferred Ac-
ceptance algorithm. With NOM thus defined they apply it to a variety of mechanism design settings
both with and without monetary transfers, including school choice, two-sided matching, auctions, and
bilateral trade.
In revenue maximisation, we are interested in designing incentive-compatible mechanisms that also
provide guarantees on the profits that can be extracted from the bidders. To analyse the performance
of a mechanism in a prior-free setting we can compare the revenue that the mechanism extracts from
the bidders against a variety of benchmarks. For example, the “optimal omniscient benchmark” is a
measure of how much profit the optimal mechanism extracts from the bidders when it knows the bids in
advance, while the “optimal fixed-price benchmark” measures the revenue that can be extracted when
each bidder is offered the same “take-it-or-leave-it” price. Other benchmarks can be defined to reflect,
for example, the revenue extracted by an auction that sets a single price and must sell at least some
number of copies m of the item, or the revenue extracted by an auction that must offer a monotone price
vector with respect to the ordering of the bidders according to their valuation. [Goldberg et al., 2006]
introduced the notion of “competitive” auctions, adapted from the analysis of online algorithms, in the
setting of prior-free auctions for digital goods. An auction is said to be competitive to some benchmark if
on every bid profile the revenue extracted by the mechanism is within a constant factor of that specified
by the benchmark. They focus on dominant strategy incentive compatibility and show that no truthful
auction in this setting is competitive to even the optimal fixed-price benchmark, much less the optimal
omniscient benchmark. To this end, in this paper we study the extent to which we can design competitive
auctions for digital goods and a variety of other settings by relaxing the strategyproofness requirement
to non-obvious manipulability.
Related Work. [Troyan and Morrill, 2020] introduce non-obvious manipulability for direct-revelation
mechanisms and provide a characterisation, similarly to [Li, 2017], for such mechanisms as those whose
profitable deviations may be recognised by a cognitively limited agent that is unable to engage in con-
tingent reasoning. They apply this to the settings of school choice, two-sided matching, auctions, and
bilateral trade and use their framework to classify mechanisms in these settings as either obviously
manipulable or not obviously manipulable.
NOM has since been studied in a range of different contexts. [Aziz and Lam, 2021] look at obvious
manipulations in the context of voting and study the conditions necessary for certain voting rules to
be NOM. They also look at computational issues related to computing obvious manipulations when
they exist, reducing the problem to that of determining whether there exists a way for a set of voters
to vote, given a set of votes that have already been submitted, in order to elect a given candidate,
yielding a polynmomial-time algorithm for the k-approval voting rule. [Arribillaga and Bonifacio, 2024]
also study the non-obvious manipulability of voting rules and restrict attention to “tops-only” rules,
which only consider each agent’s top preference when selecting an outcome. They first characterise the
rules without obvious manipulations in this general setting, then restrict attention to two subclasses of
these rules known as median voter schemes and voting-by-committees to provide a more fine-grained
classification of the set of tops-only rule which are NOM.
[Ortega and Segal-Halevi, 2022] apply non-obvious manipulability to indirect mechanisms, specifi-
cally cake-cutting, and show that, unlike strategyproofness, NOM is compatible with a notion of fairness
known as proportionality. [Psomas and Verma, 2022] also apply NOM to the setting of fair division and
study deterministic mechanisms for allocating indivisible goods to agents with additive valuations. They
focus on envy-freeness up to one good (EF1) and show that while NOM mechanisms exist for maximis-
ing social welfare, the same is not true for maximising egalitarian or Nash welfare. They also reduce
the problem of designing NOM and EF1 mechanisms to that of designing EF1 algorithms, where the
reductions preserve the efficiency guarantees.
[Archbold et al., 2023a] study how to design NOM mechanisms using monetary transfers, provid-
ing characterisations for the class of allocation functions that are implementable in general domains,

2
and then by focusing on single-parameter domains they recover an analogue to the monotonicity con-
dition for allocation functions that are truthfully implementable. They apply this to bilateral trade,
and while any efficient, individually rational, budget balanced mechanism was known to be obviously
manipulable, this issue persists even for approximate budget balance. This line of work is extended in
[Archbold et al., 2023b] to study the class of allocation functions implementable with payments as indi-
rect NOM mechanisms. They prove an analogous result to the revelation principle for single-parameter
agents: for any allocation function implementable by an indirect mechanism there is an equivalent direct
mechanism that implements it.
Our Contribution. In this paper we study the extent to which it is possible to design NOM
mechanisms that perform well for revenue maximisation and frugal procurement. In the former setting
we wish to allocate goods to the agents with the aim of maximising the social welfare as well as the
revenue extracted from the bidders. In the latter we instead allocate chores and must compensate the
bidders for being selected, and we wish to design an incentive-compatible mechanism while minimising
the sum of payments made to the bidders.
Starting with goods auctions in which agents are single-parameter we identify two necessary and
sufficient conditions for mechanisms to be NOM in this context. These conditions state that for each
agent and each valuation there must be some valuation profile of the other bidders in which she wins the
auction for free and in which she loses. We refer to these bid profiles and “golden ticket” and “wooden
spoon” profiles. For single-parameter settings this characterisation holds for any allocation and payment
functions with non-zero approximation guarantee to the optimal social welfare.

Main Result 1 (informal). An auction that α-approximates social welfare and is β-

competitive against social welfare for any α, β > 0 is not obviously manipulable if and
only if it is a Willy Wonka mechanism.

We can show a slightly weaker result when generalising from binary allocation settings to more general
outcome spaces. Now the golden ticket profile is any input to the mechanism where an agent receives
her most valued allocation for free, and this is both necessary and sufficient whenever the allocation
function is maximal-in-range. We show that the wooden spoon profile, again an input resulting in the
agent losing the auction, is also sufficient to satisfy the worst-case NOM constraints.
We refer to mechanisms satisfying these golden ticket and wooden spoon properties as “Willy Wonka”
mechanisms. Given black box access to an α-approximate (or maximal-in-range) allocation function we
then design basic payment rules that result in high revenue for the auctioneer, which sacrifices the
revenue from the lowest winning bidder in the approximately optimal allocation in order to satisfy the
golden ticket property. Specifically, given an input bid profile we first compute the approximately optimal
allocation and charge each winning bidder first-price payments, except the lowest winning bidder whose
payment is set to zero. Of course, if we were to do this on each allocation then we might sacrifice the
revenue from the only winning bidder (depending on the feasible set of allocations) and end up with zero
revenue. It suffices to have only one bid profile where this occurs in order to satisfy the best-case NOM
constraints. Therefore we can choose the bid profile corresponding to an allocation that maximises the
number of winners as the golden ticket and thus minimise the amount of revenue we sacrifice. For all
remaining bid profiles we will simply recover the optimal revenue.

Main Result 2 (informal). There is a mechanism that is not obviously manipulable

and α(1 − 1/τ )-competitive to the optimal revenue, where α denotes the approximation
guarantee of the allocation function and τ the maximum number of winning bidders in
an optimal allocation.

Finally we turn our attention to procurement auctions where agents incur costs for being included in
an allocation and agents must therefore be compensated. Now we aim to minimise the social cost, and
we want to design the payments to implement the socially optimal allocation as a NOM mechanism. We
want our payment rules to be frugal, meaning they should not overpay the agents to achieve incentive-
compatibility. To this end we apply our golden ticket and wooden spoon properties to this setting. The
former states that for each agent, on top of being compensated her bid (due to individual rationality),
there must be a profile where she is paid an additional and costly sum of money that is sure to maximise
her utility. The latter, again, states that each must also lose the auction. Our first mechanism is a
natural analogue to the Willy Wonka mechanisms given for revenue maximisation, but we show that its

3
performance with respect to frugality against the optimal solution can go to zero. We then amend the
payment rule to modify the (costly) golden ticket profiles and show that the resulting mechanism never
overpays the agents by more than a factor of two with respect to the second-lowest social cost.

Main Result 3 (informal). There is a procurement auction that is not obviously

manipulable and pays at most twice the cost of the second-best solution.

We note that we change the benchmark from the optimum social cost to that of the second best
social cost. For a given agent, if they can never be allocated at the same time as another agent, then
the frugality ratio with respect to the optimal will go to zero if the auction is NOM, and is therefore
unavoidable. We can sidestep this issue by benchmarking the mechanism against the cost of the next
best solution, and show that now whenever an agent can only be the sole winner of the allocation our
mechanism does not overpay at all with respect to this cost. We conclude the paper with a brief discussion
on the areas for further research that our work leaves open.

2 Preliminaries
General setup. For a natural number k, let [k] denote the set {1, 2, . . . , k}. We consider the familiar
mechanism design setting where there is a set [n] of agents (or equally, bidders) where each agent
i has a piece of private information ti (also known as her type) taken from some domain Di . The
set of type profiles is denoted by D = ×i∈[n] Di . We also refer to D as the set of bid profiles. For
any b = (b1 , b2 , . . . , bn ) ∈ D we write (ti , b−i ) to denote the bid profile obtained by replacing the ith
coordinate of b with ti ∈ Di .
Auctions. We consider mechanisms that use monetary transfers. In this setting an auction (mech-
anism) is a tuple M = (a, p) consisting of an allocation function a : D → A, where A is some set of
feasible allocations, and a payment function p : D → Rn . On input b the output of M is denoted
M (b). An agent’s type describes the utility she gets from each allocation, and we can think of agent
i’s type as a function ti : A → R. For a given allocation A ∈ A we assume A can be decomposed into
the vector (A1 , A2 , . . . , An ), where Ai denotes what is allocated to agent i under A. We also assume
each agent derives utility only with regard to her own allocation, hence for each agent i and allocation
A we have ti (A) = ti (Ai ). We will therefore denote by Ai the set of all unique allocations to agent i.
Moreover we denote by µ(A) the set of bidders allocated under the allocation A. We will also refer to
this set as the set of “winners”. The notation 1[·] represents the indicator function, and we will use
1[i ∈ µ(A)] = 1 if i ∈ µ(A) and 1[i ∈ µ(A)] = 0 otherwise. Given a mechanism M and agent i with
type ti we write i’s utility on outcome M (b) as ui (M (b) ; ti ). All auctions we consider are assumed to
be individually rational, meaning that for each agent i with type ti and every partial profile b−i it holds
that ui (M (ti , b−i ) ; ti ) ≥ 0.
Goods and chores. Depending on the setting, we assume each agent receives either a non-negative
or non-positive utility for each feasible allocation in A. We refer to the former as a goods auction and
the latter as a procurement auction, or equivalently, an auction in which allocations represent (bundles
of) chores. In a goods auction we extract payments from the bidders and we assume our auctions make
no positive transfers, meaning no bidder is paid money in addition to their allocation. Conversely, for a
procurement auction agents incur a cost for being allocated and hence must be compensated, and here
individual rationality implies that all payments to agents are non-negative. We will formally introduce
these notions, including those related to the optimality of an auction’s allocation and payment functions,
in their respective Sections 3 and 4 as needed.
Incentive compatibility. We are interested in designing incentive-compatible mechanisms for
agents with a particular form of imperfect rationality whereby they are only able to compare outcomes
of the mechanism at the extremes of their utility function. A mechanism M = (a, p) is not obviously
manipulable (NOM) if the following two conditions hold for every i ∈ [n]:
sup{ ui (M (ti , b−i ) ; ti ) } ≥ sup{ ui (M (bi , b−i ; ti ) }, (1)
b−i b−i

inf { ui (M (ti , b−i ) ; ti ) } ≥ inf { ui (M (bi , b−i ; ti ) }. (2)

b−i b−i

If (1) holds then M is best-case not obviously manipulable (BNOM) and if (2) holds then it is worst-case
not obviously manipulable (WNOM). We also use the term incentive compatible to refer to a mechanism

4
that is NOM. Any misreport bi such that either (1) or (2) does not hold is said to be an obvious
manipulation of M .
The structure of the feasible set. The structure of A reflects the auction setting under con-
sideration. We provide several examples here for clarity of exposition. A set system is a tuple (E, S)
where E is referred to as the ground set and S ⊆ 2E the family of subsets. In general in a binary
allocation unit-demand setting each bidder can either be allocated or unallocated in each allocation A,
and the feasible set of allocations A in these settings can be described by the set S in the set system
([n], S). This set S will vary depending on the type of auction being run, and can encompass, for ex-
ample, digital goods auctions [Goldberg et al., 2006], knapsack auctions [Aggarwal and Hartline, 2006],
and spanning tree auctions [Quadir, 2017]. For these settings we can represent each allocation by the
vector A ∈ {0, 1}n, where Ai = 1 if i is allocated and Ai = 0 otherwise. In a multi-unit auction there are
k copies of a single item and each agent may therefore be allocated up to k copies (and note that agent
types may not be linear functions in the number of items they are allocated). In this case each allocation
n
can be represented by a vector A P∈ {0, . . . , k} , where Ai describes the number of copies allocated to
agent i (and clearly we require i∈[n] Ai ≤ k for A to be a feasible allocation). As a final example,
consider a combinatorial auction in which there is a set S of items and each allocation distributes a
(possibly empty) bundle of items to each agent. Here for each allocation A the entry Ai ⊆ S describes
the bundle of items allocated to i, and therefore the feasible set of allocations can be represented by the
set of all A such that Ai ⊆ S for each i, and Ai ∩ Aj = for each distinct i, j ∈ [n]. Regardless of the
specific setting we assume that for each player there is some feasible allocation in A where they do not
win, otherwise clearly there is nothing to be done from the perspective of realigning incentives.
Bidding languages. In general we assume that each agent i reports her type directly to the
mechanism by specifying a real number for each unique personal allocation Ai ∈ Ai . To avoid cluttering
the notation we will drop the i subscript from Ai . Therefore agent i’s type is represented by the vector
ti = (tA A
i )A∈Ai . In Section 3 we assume that for each A ∈ A we have ti ∈ [0, h] and in Section 4 we assume
A
that ti ∈ [−h, 0] for some h ∈ R+ . In auctions based on set systems each agent needs only to report a
single number to the mechanism, and as such we refer to them in this case as being single-parameter.
When this is not the case then agents are in general said to be multiparameter.

3 Revenue Maximisation
In this section we consider goods auctions, where for each agent i and each allocation A ∈ A we have
ti (A) ≥ 0. We therefore refer to each agent’s type as her valuation function and denote the valuation
function for agent i as vi : A → R≥0 . This is simply a notational change to differentiate between the two
settings of revenue maximisation and frugality. In this setting the auction mechanism extracts payments
from bidders in exchange for their allocation, hence for an auction mechanism M = (a, p) and input bid
profile b the utility that agent i with valuation vi will receive is denoted ui (M (b) ; vi ) = vi (ai (b)) − pi (b).
In this setting we consider allocation functions that (approximately) maximise the social welfare. Fix
aPvaluation profile v. The social welfare of an allocation A with respect to v is simply SW(A, v) =
i∈[n] vi (Ai ). An allocation function a is said to α-approximate the social welfare for α ∈ (0, 1) if
SW(a(v), v) ≥ αW ∗ for every v, where W ∗ = maxA∈A SW(A, v) denotes the optimal social welfare of
an allocation with respect to v.
We study the revenue we can extract from the bidders while ensuring the mechanism P is NOM. The
revenue of an auction (a, p) on bid profile b is simply the sum of the payments i∈[n] pi (b). The
performance of a mechanism’s payment function will be measured with respect to some benchmark. Let
X (b) be the revenue of some payment benchmark X when given bid profile P b. A payment function p is
said to be β-competitive against X for β ∈ (0, 1) if for all b it holds that i∈[n] pi (b) ≥ βX (b).

3.1 Binary allocation

We begin by characterising the class of auctions which are NOM for single-parameter agents binary
allocation settings. Our characterisation shows that two properties based on the outcomes of the mech-
anism, known as “golden ticket” and “wooden spoon” profiles, are necessary and sufficient in order for
the mechanism to be NOM. The former states that, for each bidder and each bid, it is possible, given
some bid profile of the other agents, to win the auction for free. It is straightforward to see that this
ensures that (1) is satisfied, since there is a bid profile where each agent receives her highest possible
utility (assuming no positive transfers from the mechanism to the agent). On the other hand, as long

5
as the allocation function has a positive welfare guarantee then the golden ticket property is necessary
when the mechanism is BNOM, otherwise she may submit a bid below the assumed minimum payment
price and get strictly greater utility in the best case than when bidding truthfully.
Lemma 1 (Golden ticket for binary allocation). Let M = (a, p) be an auction where the allocation
function has any positive approximation α > 0 to the optimal social welfare. Then M is BNOM if and
only if for every agent i ∈ [n] with valuation vi there exists a bid profile such that i wins the auction for
free.
Proof. ( =⇒ ) Begin by observing that i must always be able to win the auction. On the bid profile
(vi , 0, . . . , 0) a positive approximation to the optimal social welfare is only possible if i is allocated.
Suppose for contradiction that whenever i wins the auction she pays at least some positive price p. Her
best utility when bidding vi truthfully is at most vi − p. Now let bi be any misreport such that 0 < bi < p
and note that p ≤ vi by individual rationality. Take the profile (bi , 0, . . . , 0) and by the same argument
i must win. Her utility for this misreport is at least vi − bi ≥ vi − p, breaking BNOM. Therefore i must
be able to win the auction for free.
( ⇐= ) Since i can win the auction for free when bidding vi her best truthful utility is exactly vi . This
is clearly unbeatable by any misreport if M makes no positive transfers, hence the auction is BNOM.
Note that as a corollary of Lemma 1 we have that for NOM auctions in which for each allocation the
set of winning bidders is a singleton, such as single-item auctions, then the revenue guarantee against
the optimal goes to zero when taken over all possible inputs. To verify that this is the case, consider the
Vickrey auction, which is strategyproof and hence also NOM. The revenue obtained by this mechanism
is simply the second highest bid, which is zero whenever there is a single agent i with a positive bid bi
and the remaining agents bid zero. The optimal revenue is bi , resulting in a 0-approximation.
Analogously to the golden ticket, the wooden spoon property states that for each bidder and each
bid, it is possible to lose the auction. This guarantees that (2) is satisfied, since now there is a bid profile
where each agent receives zero utility and by individual rationality this is the lowest possible when
bidding truthfully. Since this holds for each bid then the worst utility of a dishonest bid can certainly be
no greater than zero. Now to prove the necessity of the wooden spoon we assume the payment function
extracts any positive factor of revenue from the bidders and show that if the agent always wins the
auction then it is possible to underbid and get a strictly greater worst case utility versus a truthful bid.
Lemma 2 (Wooden spoon for binary allocation). Let M = (a, p) be an auction where the payment
function guarantees to extract any positive factor β > 0 of the sum of bids in revenue from the bidders.
Then M is WNOM if and only if for every agent i ∈ [n] with valuation vi there exists a bid profile such
that i loses the auction.
Proof. ( =⇒ ) Suppose bidder i always wins the auction. When bidding bi < vi her worst utility is at
least vi − bi . On the profile (vi , 0, . . . , 0) then i must win the auction and pay at least βvi meaning her
worst truthful utility is at most (1 − β)vi . Taking a misreport bi < βvi results in a worst case utility of
vi − bi > (1 − β)vi , breaking WNOM. Therefore i must lose the auction.
( ⇐= ) Take bidder i with value vi . Since i can lose then her worst utility when bidding vi truthfully
is at exactly 0. If i were to underbid then the same argument gives her a worst dishonest utility of 0,
while submitting an overbid bi can clearly yield a negative utility if i were to win and pay bi > vi . Hence
M is WNOM.
With Lemmas 1 and 2 we have a full characterisation for NOM auctions in a range of binary allocation
settings. Namely, as long the allocation function has some positive approximation guarantee to the
optimal social welfare and the payment function always extracts some positive revenue from the bidders,
then for each agent there must be both a bid profile where she wins the auction for free and a bid profile
where she loses. This motivates the following definition for auctions that satisfy these criteria.
Definition 1 (Willy Wonka mechanism for binary allocation). A Willy Wonka mechanism for binary
allocation settings is a mechanism M = (a, p) whose allocation and payment functions satisfy the follow-
ing:
1. For each i and each bi there is a bid profile b = (bi , b−i ) where i wins the auction for free.
2. For each i and each bi there is a bid profile b = (bi , b−i ) where i loses the auction.

6
 Given a bid bi of player i we refer to the golden ticket profile for bi as the profile b−i such that i wins
for free on b = (bi , b−i ). Likewise we refer to the wooden spoon profile for bi as the profile b−i such that
i loses the auction on b = (bi , b−i ). Therefore as long as each player i has both a golden ticket profile
and a wooden spoon profile for each bid bi then the resulting mechanism is NOM. In the following we
will assume that the wooden spoon profile is ensured by the feasible set, meaning that for a sufficiently
approximately-optimal allocation function and each bid of an agent i there is a bid profile of the other
agents where i must lose the auction.
Our idea is to design Willy Wonka mechanisms that sacrifice a small portion of the revenue in order
to achieve incentive compatibility. Ideally, for each bidder i and each bid bi we want to find a bid profile
b−i and an allocation A where i wins that maximises the social welfare of the bidders excluding i, since
we can charge first-price payments to the winners and award i her allocation for free. This b−i would
then be the golden ticket profile for i with bid bi . In other words for each i with bid bi the optimal Willy
Wonka mechanism would compute an allocation that obtains revenue

max { max SW(A, b−i ) }. (3)

A : i∈µ(A) b−i

Suppose we have black box access to an α-approximate allocation function. In order to tackle finding
a good approximation to the optimal revenue above, consider the Willy Wonka mechanism that first
computes an approximately optimal allocation and then charges first-price payments to all but the
lowest winning bidder, who is allocated at no charge. Then the revenue extracted by the mechanism is
simply
α · SW(A, b) − min bi .
i∈µ(A)

A problem with this mechanism is that it specifies a golden ticket profile for every input b. Now
if there is any allocation A such that |µ(A)| = 1 then we lose all of the revenue, resulting in a 0-
approximation. We amend this in Mechanism 1 by only allocating the lowest bidder for free when the
cardinality of the winning set is maximal over all allocations in the image of the allocation function.

Mechanism 1 Willy Wonka auction for binary allocation.

Input: Bid profile b ∈ D.
Output: Allocation a(b) ∈ A ⊆ {0, 1}n, payments p(b) ∈ Rn≥0 .
Let A be an α-optimal allocation w.r.t. b and let µ = µ(A).
for each bidder i ∈ [n] do
ai (b) ← Ai
pi (b) ← bi · Ai
if A ∈ argmax{ |A′ | : A′ ∈ image(a) } then
Let i ∈ argmin{ bj : j ∈ µ }
pi (b) ← 0
return a(b), p(b)

We note a few assumptions made by the mechanism in order to simplify its description. Firstly, we
assume that in each bid profile corresponding to a winning set of maximum cardinality there is a distinct
lowest winning bidder in the optimal allocation. If lowest winning bids are tied then in order to obtain
the revenue bound of Theorem 1 we would want for only one of these to be allocated for free. We can
break ties arbitrarily when this is the case, and since there is at least one profile where each player is the
unique lowest winning bidder for each possible then Lemma 1 ensures the mechanism is BNOM.
Again we note that this payment rule taken together with the allocation algorithm does not necessarily
define any wooden spoon profiles. In most cases the wooden spoon profiles should be ensured by the
welfare approximation of the allocation function being high enough and by the feasible set of allocations.
In other words, if it is possible to lose the auction then WNOM is guaranteed by Lemma 2. In settings
where this does not hold and it is always possible to allocate a given agent, then in order to achieve
WNOM we must specify these profiles explicitly. In this special case then, again, in order to achieve the
revenue bound of Theorem 1 we have to take care that the golden ticket profile of one agent does not
coincide with the wooden spoon profile of another agent.
Theorem 1. Mechanism 1 is not obviously manipulable and is α(1 − 1/τ )-competitive to the optimal
revenue, where τ = max{ |µ(A)| : A ∈ image(a) } denotes the cardinality of the largest winning set under
a.

7
Proof. The mechanism is NOM by Lemmas 1 and 2. It charges first-price payments to all winning bidders
apart from the lowest when the number of winners is maximal, meaning the revenue is a (1 − 1/τ ) factor
of the social welfare of the returned allocation. Since the allocation function is α-approximate then the
social welfare of the allocation is at least an α fraction of the optimal, giving the bound.
In the next section we generalise the arguments in Lemmas 1 and 2 to handle general outcome spaces,
with an additional condition on the allocation function.

3.2 General outcome spaces

In this section we assume A represents more general classes of allocations and that agents are multipa-
rameter. Recall that each agent i’s type is now a vector (viA )A∈A where viA ∈ [0, h] represents the utility
i receives from allocation A. In order to derive similar golden ticket and wooden spoon properties to the
previous section we impose an additional requirement on the allocation function of the mechanism that
requires the allocation to be able maximise over a fixed range of allocations. An allocation function a
is said to be maximal-in-range if there exists some subset R ⊆ A such that for all b ∈ D the function
outputs a(b) ∈ argmaxA∈R SW(A, b).
Lemma 3 (Golden ticket for general outcome spaces). Let M = (a, p) be an maximal-in-range auction
with range R ⊆ A. Then M is BNOM if and only if for every agent i ∈ [n] with valuation vi = (viA )A∈A
there exists a bid profile b such that a(b) ∈ argmaxA∈R viA and pi (b) = 0.
′
Proof. Let A ∈ argmaxA′ ∈R viA be i’s highest-valued allocation amongst those in R and first note that
it must always be possible for i to be allocated A. Take the bid profile (vi , 0, . . . , 0), where 0 = (0, . . . , 0)
denotes the all-zero type. Since M is maximal-in-range and A ∈ R then M must return the allocation
A.
( =⇒ ) Now suppose that whenever the allocation A is returned i must pay some positive price p.
Her best truthful utility is at most viA − p. Let bi be any misreport of agent i with 0 < bA i < p and
A′
bA
i = max A ′ ∈R b
i . Note that b A
i is an underbid for allocation A since p ≤ vi
A
by individual rationality.
Now take the profile (bi , 0, . . . , 0). By the same argument as above i must be allocated A and pay some
amount p′ where 0 < p′ < bA A A A
i . Her best utility for this misreport is at least vi − bi > vi − p and hence
strictly larger than her best utility for reporting vi truthfully, breaking BNOM. Hence i must win A for
free under some bid profile.
( ⇐= ) Since i wins A for free then bidding vi truthfully can result in a utility of viA , which is clearly
unbeatable since M makes no positive transfers. Therefore M is BNOM.
We must strengthen the requirement of the allocation function being α-optimal to being maximal-
in-range. This ensures that bidder i is allocated her most valuable allocation in the range of the auction
when the other bidders are excluded from winning. Without this requirement then the auction could
return any allocation A that i values at least an α fraction as much as her favourite allocation, which
may not imply incentive-compatibility. In this case the performance of the allocation function can be
expressed as
argmaxA∈R SW(A, b)
α = min . (4)
b argmaxA∈A SW(A, b)

The wooden spoon property remains exactly the same and using the exact very simple argument
from the proof of Lemma 2 we can show that if each agent can lose the auction then it must be WNOM.
Lemma 4 (Wooden spoon for general outcome spaces). Let M = (a, p) be an auction for multiparameter
agents that is β-competitive against the social welfare. If for every agent i ∈ [n] with valuation vi =
(viA )A∈A there exists a bid profile b such that ai (b) = ∅ then M is WNOM.
We note that unlike Lemmas 1 to 3 this is not a complete characterisation since we do not have the
necessity of the wooden spoon for auctions in these settings to be WNOM. However, when designing our
NOM auctions in the following we will only need the sufficiency of the golden tickets and wooden spoons
to show the resulting mechanisms are incentive-compatible.
Definition 2 (Willy Wonka mechanism for general outcome spaces). A Willy Wonka mechanism for
general outcome spaces is a mechanism M = (a, p) whose allocation and payment functions satisfy the
following:

8
 1. For each i and each bi there is a bid profile b = (bi , b−i ) where i is allocated argmaxA∈A bA
i for
free.
2. For each i and each bi there is a bid profile b = (bi , b−i ) where i loses the auction.
We only need to modify Mechanism 1 slightly in order to apply to general outcome spaces, which we
provide in Mechanism 2. We now assume the allocation function a is maximal-in-range as per Lemma 3,
but as before in Mechanism 2 we compute an approximately optimal allocation and then charge first-
price payments in most cases. When the allocation has maximal cardinality then we sacrifice the revenue
from the lowest winning bidder in order to satisfy the golden ticket property.

Mechanism 2 Willy Wonka auction for general outcome spaces.

Input: Bid profile b ∈ D.
Output: Allocation a(b) ∈ A, payments p(b) ∈ Rn≥0 .
Let A ∈ argmax{ SW(A′ , b) : A′ ∈ R } and let µ = µ(A).
for each bidder i ∈ [n] do
ai (b) ← Ai
pi (b) ← bA
i · 1[i ∈ µ]
if A ∈ argmax{ |A′ | : A′ ∈ R } then
Let i ∈ argmin{ bA j : j ∈ µ}
pi (b) ← 0
return a(b), p(b)

Again it is assumed that the wooden spoon profiles arise from the feasible set itself, that is, for each
agent there is some bid profile such that it is not feasible to allocate the agent. If this is not the case
we would have to artificially exclude the agent from the winning set. Naturally Mechanism 2 obtains a
similar revenue guarantee to Mechanism 1 of α(1 − 1/τ ) where τ is the size of the maximum cardinality
winning set for any allocation returned by the mechanism, and α again corresponds to the approximation
guarantee of the allocation function with respect to the socially optimal allocation and takes the form
given in (4).
Theorem 2. Mechanism 2 is not obviously manipulable and is α(1 − 1/τ )-competitive to the optimal
revenue, where τ = max{ |µ(A)| : A ∈ R } denotes the cardinality of the largest winning set under the
maximal-in-range allocation function a.

4 Frugal Procurement
In this section we consider procurement auctions, where for each agent i and each allocation A ∈ A we
have ti (A) ≤ 0. For clarity of exposition we now introduce for each agent i a cost function and denote
it ci : A → R≥0 , setting ci = −ti . Now we have a procurement auction that allocates chores to the
agents, hence the agents must be compensated for taking on their allocation. For a procurement auction
M = (a, p) given the input b the utility that agent i with cost ci will receive is denoted ui (M (b) ; ci ) =
pi (b) − ci (ai (b)).
We are now interested in allocation functions that (approximately) minimise the social cost of resulting
allocation.
P Fix a cost profile c. The social cost of an allocation A with respect to c is SC(A) =
i∈[n] ci (Ai ). An allocation function a is said to α-approximate the social cost for α ≥ 1 if SC(a(c), c) ≤
αC ∗ for every c, where C ∗ = minA∈A SC(A, c) denotes the social cost of an optimal allocation with
respect to c.
We now need to compensate agents for their allocations. All of our mechanisms must be individually
rational, hence each bidder must always be paid at least the cost she reports for the resulting allocation.
We want to study the extent to which our mechanism overpays with respect to some benchmark. Let
X be some payment benchmark for a procurement auction and X (b) denote the sum of payments made
P the bidders on bid profile b. A payment function p is said to be β-frugal against X for β ≥ 1 if
to
i∈[n] pi (b) ≤ βX (b) for all b. We refer to β as the frugality ratio of the payment function (and equally,
the mechanism) with respect to X .
We think of the auctions as allocating bundles of chores, and to maintain individual rationality we
must now instead reimburse or compensate the bidders allocated a chore with a monetary transfer.
For consistency we still refer to the set of such bidders as the set of winners, and every other bidder is
therefore still considered a loser of the auction. We also restrict our attention to single-parameter agents.

9
Lemma 5 (Golden ticket for binary allocation). Let M = (a, p) be a procurement auction that α-
approximates the social cost for any α > 0. Then M is BNOM if and only if for every agent i ∈ [n] with
cost ci there exists a bid profile where i wins the auction and is paid h.
Proof. ( =⇒ ) Suppose for contradiction that M only ever pays bidder i at most p < h when she bids
ci . By individual rationality we must have p ≥ ci . Now take an overbid bi such that p < bi < h/α. Let
b = (bi , b−i ) be a bid profile such that bj > αbi for each j 6= i. On b the chore must get allocated to i
in order to preserve the α guarantee to the optimal social cost. Since i is allocated the chore she must
be compensated at least bi , resulting in a best case dishonest utility of at least bi − ci which is strictly
greater than her best case truthful utility of p − ci , thus violating BNOM.
( ⇐= ) If there exists a bid profile where i wins and is paid h then when i has cost ci and bids
truthfully her best utility is h − ci ≥ 0. Since i can never submit an overbid greater than h then no
misreport can get her a greater utility, thus M is BNOM.
Lemma 6 (Wooden spoon for binary allocation). Let M = (a, p) be a procurement auction that is
β-frugal against the social cost for any β > 0. Then M is WNOM if and only if for every agent i ∈ [n]
with cost ci there exists a bid profile such that i loses the auction.
Proof. ( =⇒ ) Assume i always wins the auction. Then on an overbid bi > ci she would get utility
at least bi − ci . Now take the bid profile (ci , h, . . . , h) such that (n − 1)h/β > ci where only i may
win the auction. The β-frugality of the auction means that i cannot be compensated more than βci
meaning her best truthful utility is at most (β − 1)ci . Taking bi > βci results in a worst utility of at least
bi − ci > (β − 1)ci thus breaking the WNOM constraints. Therefore i must be able to lose the auction.
( ⇐= ) By a similar argument to the proof of Lemma 5 if the monetary payment is independent of
the bid then i’s worst truthful and dishonest utilities are both 0, satisfying WNOM.
We use the above conditions to design frugal Willy Wonka mechanisms for procurement auctions.
Similarly to (3) fix a player i with bid bi . On the golden ticket profile for i the optimal Willy Wonka
mechanism for frugal procurement should select an allocation A that compensates the agents by

min { min SC(A, b−i ) + h },

A : i∈µ(A) b−i

since i will be paid h while the remaining winners are paid their bid. To obtain a good approximation
to the above, again consider a mechanism with black box access to an α-approximate allocation function
such that on input b the mechanism selects an α-optimal allocation A that pays the highest winner h and
the remaining winners their bid. Now the highest winning bidder (instead of the lowest) is the recipient
of the golden ticket, since this minimises the total payment made on the given instance. Then the total
cost on allocation A is equal to
α · SC(A, b) + h − max bi .
i∈µ(A)

Again, we need not designate a golden ticket profile (giving the highest winning bidder a payment of
h) for every bid profile b as this will degrade the frugality ratio with respect to the sum of (winning) bids.
Consider an analogous approach to Mechanisms 1 and 2 for designing frugal Willy Wonka procurement
auctions, where the golden ticket profile occurs for each bidder when they have the highest reported cost
in an (approximately) optimal allocation of maximum winning set cardinality. Suppose A contains all
allocations with singleton winning sets. Now fix bidder i and take b = (bi , h, . . . , h) for some bi . Observe
that the optimal cost of allocating only i is bi , and at least h for any other allocation. For a given α
then a sufficiently small bi means i must be the sole winner. If the mechanism paid h to bidder i then
the resulting frugality ratio would be infinite when we take bi to zero.
We provide an alternative way of specifying the golden ticket profiles of each bidder in Mechanism 3
and show that although its frugality ratio with respect to the optimal can still be infinite, it will never
overpay the bidders by more than twice the cost of the second-best solution, provided there exists an
allocation for each agent where they are not the sole winner. Let k = maxA : i∈µ(A) |µ(A)| denote the
maximum cardinality of winning set containing i for any allocation in A, and let S be the corresponding
winning set of any such allocation. Given bid bi of player i define the golden ticket profile such that for
all j 6= i we have bj = h/(k − 1) whenever j ∈ S and bj = h otherwise. Denote this golden ticket for bi
as γi (bi ).
We note that the frugality ratio for Mechanism 3 can vary depending on the size k of the set S. If
k > 1 then the total payments made on the bid profile (bi , γi (bi )) is h + (k − 1)h/(k − 1) = 2h. When

10
Mechanism 3 Willy Wonka procurement auction.
Input: Bid profile b ∈ D.
Output: Allocation a(b) ∈ A ⊆ {0, 1}n, payments p(b) ∈ Rn≥0 .
Let A be an α-optimal allocation w.r.t. b and let µ = µ(A).
for each bidder i ∈ [n] do
k ← maxA : i∈µ(A) |µ(A)| and h′ ← h/(k − 1) (when k > 1)
γi (bi ) ← (h′ , . . . , h′ , h, . . . , h)
| {z } | {z }
(k−1) (n−k)

ai (b) ← Ai
pi (b) ← bi · Ai
if b−i = γi (bi ) then
pi (b) ← 0
return a(b), p(b)

bi = 0 then the optimum cost is h, giving a frugality ratio of 2. If k = 1 then the sum of payments is h
while the optimum is zero, meaning the frugality ratio is infinite.
Instead let us benchmark the mechanism again the cost of the second-best solution. We denote by
FR(2) (b) the frugality ratio of a given mechanism on input b with respect to the cost of the second-best
solution: P
(2) i∈[n] pi (b)
FR (b) = ,
minA { SC(A, b) : SC(A, b) > C ∗ }
where C ∗ = minA SC(A, b) denotes the cost of an optimum allocation on b. We then take FR(2) =
maxb FR(2) (b). With the following we show that Mechanism 3 has good performance with respect to
FR(2) as long as there exists for each agent an allocation where they are not the only winning bidder.
Claim 1. The frugality ratio FR(2) for Mechanism 3 is equal to 2 whenever k > 1, and 1 otherwise.
Proof. Let A = a(b) be the allocation returned by a on input b = (bi , γi (bi )). For all cases where k ≥ 1
then the cost of the second-best solution is at least h. When k > 1 then if i ∈ µ(A) the total sum of
payments is at most h + (k − 1)h/(k − 1) = 2h, otherwise if i ∈ / µ(A) then the total sum of payments is
simply h. When k = 1 then γi (bi ) = (h, . . . , h) by definition, and whoever wins is paid h, either because i
wins the golden ticket or because some bidder j 6= i wins and must be compensated at least her bid.

5 Conclusion
In this work we study NOM auctions for revenue-maximisation and frugal procurement. We characterise
the set of incentive-compatible mechanisms by introducing two design criteria that guarantee the resulting
mechanism is NOM (in the right context), namely golden tickets and wooden spoons. The former ensures
a profile where each bidder can maximise her utility when reporting her type truthfully, and the latter
ensures her worst utility when bidding truthfully is zero, and taken together they ensure the best-case
and worst-case NOM constraints are always satisfied. Given black-box access to an (approximately)
optimal allocation function we then show that we can design simple payment rules that provide good
performance with respect to either maximising the sum of payments extracted from the bidders or
minmising the compensation paid. These payment rules work by selecting, for an optimal allocation
with the maximum (across all allocations) number of winning bidders, a single bidder whose utility
function we will maximise.
There are several directions for further research left open by our work. The performance of the
payment rules we design is dependent on the number of winners in an optimal allocation. Can we design
payment rules whose performance does not depend on the structure of the feasible set, to better deal
with instances where this results in poor performance for the mechanism? We also note that, given
a bid bi of player i, the best and worst outcomes of the mechanism always occur on the same profile,
regardless of her true type. These are the golden ticket and wooden spoon profiles, respectively, and
the resulting mechanisms are therefore single-line [Archbold et al., 2023a]. These payment schemes are
simple to state. Are golden ticket and wooden spoon payment rules the only types of single-line payment
rules leading to good performance from the mechanism? NOM also permits much more general payment

11
schemes – can we design non single-line payment rules that yield better performance that can be achieved
by a single-line mechanism? Finally, can we design (efficient) allocation functions ourselves for general
settings so that our mechanisms do not rely on them as black boxes, and can we use in conjunction with
an appropriate payment rule to yield a NOM mechanism?

References
[Aggarwal and Hartline, 2006] Aggarwal, G. and Hartline, J. D. (2006). Knapsack auctions. In Proceed-
ings of the Seventeenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2006, Miami,
Florida, USA, January 22-26, 2006, pages 1083–1092. ACM Press.
[Archbold et al., 2023a] Archbold, T., de Keijzer, B., and Ventre, C. (2023a). Non-obvious manipula-
bility for single-parameter agents and bilateral trade. In Agmon, N., An, B., Ricci, A., and Yeoh,
W., editors, Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent
Systems, AAMAS 2023, London, United Kingdom, 29 May 2023 - 2 June 2023, pages 2107–2115.
ACM.
[Archbold et al., 2023b] Archbold, T., de Keijzer, B., and Ventre, C. (2023b). Non-obvious manipulabil-
ity in extensive-form mechanisms: The revelation principle for single-parameter agents. In Proceedings
of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI 2023, 19th-25th
August 2023, Macao, SAR, China, pages 2503–2510. ijcai.org.
[Arribillaga and Bonifacio, 2024] Arribillaga, R. P. and Bonifacio, A. G. (2024). Obvious manipulations
of tops-only voting rules. Games and Economic Behavior, 143:12–24.
[Aziz and Lam, 2021] Aziz, H. and Lam, A. (2021). Obvious manipulability of voting rules. In Fotakis,
D. and Insua, D. R., editors, Algorithmic Decision Theory - 7th International Conference, ADT 2021,
Toulouse, France, November 3-5, 2021, Proceedings, volume 13023 of Lecture Notes in Computer
Science, pages 179–193. Springer.
[Dur et al., 2018] Dur, U., Hammond, R. G., and Morrill, T. (2018). Identifying the harm of manipulable
school-choice mechanisms. American Economic Journal: Economic Policy, 10(1):187–213.
[Goldberg et al., 2006] Goldberg, A. V., Hartline, J. D., Karlin, A. R., Saks, M., and Wright, A. (2006).
Competitive auctions. Games and Economic Behavior, 55(2):242–269.
[Li, 2017] Li, S. (2017). Obviously strategy-proof mechanisms. American Economic Review,
107(11):3257–3287.
[Ortega and Segal-Halevi, 2022] Ortega, J. and Segal-Halevi, E. (2022). Obvious manipulations in cake-
cutting. Soc. Choice Welf., 59(4):969–988.
[Pathak and Sönmez, 2008] Pathak, P. A. and Sönmez, T. (2008). Leveling the playing field: Sincere
and sophisticated players in the boston mechanism. American Economic Review, 98(4):1636–1652.
[Psomas and Verma, 2022] Psomas, A. and Verma, P. (2022). Fair and efficient allocations without
obvious manipulations. In NeurIPS.
[Quadir, 2017] Quadir, A. (2017). Spanning tree auctions: A complete characterization. Math. Soc. Sci.,
86:1–8.
[Troyan and Morrill, 2020] Troyan, P. and Morrill, T. (2020). Obvious manipulations. Journal of Eco-
nomic Theory, 185:104970.

12