Rank-Guaranteed Auctions¹¹1We thank Mohammad Akbarpour, Martin Bichler, Tilman Börgers, Kim-Sau Chung, Rahul Deb, Jason Hartline, Paul Klemplerer, Fuhito Kojima, Andrew Komo, Paul Milgrom, Michael Ostrovsky, Antonio Penta, Ning Sun, Satoru Takahashi, and Andrzej Skrzypacz for helpful discussions. This manuscript subsumes He et al., (2022).

There is a set $S$ of $M$ items to be sold to $N$ bidders. Let $\mathcal{N}=\{1,2,\ldots,N\}$. We write $b\subseteq S$ to denote a generic bundle of items. Let $\bm{v}^{n}=\{v_{b}^{n}\}_{b\subseteq S}$ denote the valuation vector of bidder $n$, where $v_{b}^{n}$ is bidder $n$’s valuation of bundle $b$. Valuations are normalized so that $v^{n}_{\emptyset}=0$ and $v_{b}^{n}\in[\underline{v},\mkern 1.5mu\overline{\mkern-1.5muv\mkern-1.5mu}\mkern 1.5mu]$ ($\underline{v}\geq 0$) for all $b\neq\emptyset$. A generic valuation profile is denoted by $\bm{v}=(\bm{v}^{1},\bm{v}^{2},\ldots,\bm{v}^{N})$. Let $\mathcal{M}\subseteq 2^{S}$ denote a menu of bundles chosen by the auctioneer. While $\mathcal{M}$ is a choice variable of the auctioneer, for now we take it as exogenously given; we defer the discussion of menu design to Section 4. Assume that $N\geq|\mathcal{M}|+1$. Let

denote the set of feasible allocations of bundles within the menu $\mathcal{M}$, i.e., all collections consisting of non-overlapping bundles.

We define the Combinatorial Ascending Auction (CASA) as follows. The auction has an iterative structure, with the “state of the auction” characterized by the identity of the leading bidder and the leading price for each bundle. Initially, the leading price for each bundle is zero and none of the bidders is a leading bidder for any bundle. Bidders take turns raising the bids on the bundles, which determines new leading bidders and leading prices. The process repeats itself until when there are no new bids on any bundle. At that point, the auction stops. The auctioneer chooses a feasible allocation to maximize revenue, taking the leading prices as the prices for the bundles. There is also an activity rule designed to ensure that bidding activity starts out high and declines during the auction as prices rise far enough to discourage some bidders from continuing.

Formally, let $P\subset\mathbb{R}^{+}$ be a finite grid of feasible bids with grid size $\epsilon$ and $\max P>\mkern 1.5mu\overline{\mkern-1.5muv\mkern-1.5mu}\mkern 1.5mu$.

1. (1)

   Initialization stage $t=0$. Define

   • •

     the leading bidder vector at stage $0$: $\bm{\phi}^{0}=(\phi^{0}_{b})_{b\in\mathcal{M}}=\bm{0}$,

   • •

     the leading price vector at stage $0$: $\bm{p}^{0}=(p^{0}_{b})_{b\in\mathcal{M}}=\bm{0}$,

   • •

     the set of active bidders at stage $0$: $\mathcal{N}^{0}=\{1,2,\ldots,N\}$.

2. (2)

   Bidding stage $t\geq 1$. An active bidder $n\in\mathcal{N}^{t-1}$ observes $(\bm{p}^{t-1},\{b\,|\,\phi_{b}^{t-1}=n\})$, and decides whether to quit, which bundles to bid on, and how much to bid.³³3The bidder selection rule and the observability of history is inconsequential for our analysis. For concreteness, we consider the selection rule that active bidders are cycled in ascending order according to their indices, and the observability of history is minimized to maximally protect privacy.

   • •

     Bidder $n$ may choose to quit by submitting an empty bid only if $\{b\,|\,\phi_{b}^{t-1}=n\}=\emptyset$, i.e., bidder $n$ is not a leading bidder for any bundle in stage $t-1$. Quitting is irreversible, that is, if bidder $n$ chooses to quit, then bidder $n$ becomes an inactive bidder and does not participate in future bidding rounds. Update:

     • –

       $\bm{\phi}^{t}=\bm{\phi}^{t-1}$, $\bm{p}^{t}=\bm{p}^{t-1}$, $\mathcal{N}^{t}=\mathcal{N}^{-1}\setminus\{n\}$.

   • •

     If bidder $n$ chooses not to quit, then she submits a bid—a nonempty set of bundle-price pairs $\{(b,p_{b})\}\subset\mathcal{M}\times P$, subject to the requirements that (1) Leading bids are binding: if bidder $n$ is the leading bidder at some bundle in stage $t-1$, then she must include that bundle in her bid with a bid that is weakly higher than the current leading price for that bundle, and (2) Minimum bid increment: if bidder $n$ would like to bid on some bundle for which she is not the leading bidder, then her bid for that bundle must be strictly higher than the current leading price for that bundle. Update:

     • –

       $\phi_{b}^{t}=n$ and $p_{b}^{t}=p_{b}$ for any bundle $b$ included in her bid,

     • –

       $\phi_{b^{\prime}}^{t}=\phi_{b^{\prime}}^{t-1}$ and $p_{b^{\prime}}^{t}=p_{b^{\prime}}^{t-1}$ for any bundle $b^{\prime}$ not included in her bid,

     • –

       $\mathcal{N}^{t}=\mathcal{N}^{t-1}$.

   Then, move on to the bidding stage $t+1$.

3. (3)

   Allocation. The auction ends (in stage $T$) when the leading prices stay constant for $N$ consecutive periods. The auctioneer chooses a feasible allocation to maximize

   $\displaystyle\max_{\bm{b}\in\mathcal{B}(\mathcal{M})}\,\sum_{b\in\bm{b}}\,p_{b}^{T}.$

   Denote the maximizer by $\bm{b}^{*}$. Each bundle $b\in\bm{b}^{*}$ is allocated to $\phi_{b}^{T}$ at the price $p_{b}^{T}$.

In words, the auction format of CASA runs parallel ascending auctions for each bundle $b\in\mathcal{M}$. Then, the items are allocated to maximize the total price. We discuss possible variants of CASA and its relation to existing auction formats in Section 2.4.

In this subsection, we study the strategic behavior of the bidders and establish the rank-guarantee property of CASA. We only assume minimal rationality on the part of the bidders—bidders are rational in the sense of not playing obviously dominated strategies. Formally we adopt the solution concept of implementation in strategies that are not obviously dominated (see Li, (2017) and Li and Dworczak, (2021)). We first sketch the intuition, and then provide the formal arguments.

At any history, consider a non-leading bidder’s choice as to whether to quit. Obviously, as quitting is irreversible, quitting the auction leads to a best possible outcome of a zero payoff. Suppose that there is some bundle for which the bidder’s valuation is higher than the current leading price for that bundle. Consider the following strategy where the bidder raises the price for this particular bundle and never revises her bid afterwards. Clearly, this continuing strategy guarantees a non-negative payoff for the bidder. Thus, at least for the purpose of deciding whether to quit, it is “obviously optimal” not to quit.

More formally, let $h=(t,(\mathcal{N}^{0},\ldots,\mathcal{N}^{t-1}),(\bm{p}^{0}\ldots,\bm{p}^{t-1}),(\bm{\phi}^{0},\ldots,\bm{\phi}^{t-1}))$ denote a history of the game in stage $t$, $H_{t}$ the set of such histories in stage $t$, and $H=\cup_{t\geq 0}H_{t}$. Suppose that bidder $n$ is the active bidder in some stage $t$ and $I_{n}$ is bidder $n$’s information set. Then the observed prices $\bm{p}$ and $n$’s leading bundles $\bm{b}$ are the same for all $h\in I_{n}$. Let $\mathcal{I}_{n}$ denote all information sets of $n$. Let $s_{n}:\mathcal{I}_{n}\to 2^{\mathcal{M}\times\mathbb{R}^{+}}$ denote bidder $n$’s (pure behavioral) strategy and $u_{n}(\bm{s},\bm{v}^{n}|h)$ the payoff to bidder $n$ given valuation vector $\bm{v}^{n}$, strategy profile $\bm{s}$, conditional on the current history $h$ and $n$ bidding in period $t$.

The first inequality is identical to the definition of the obvious dominance relation in Li, (2017), i.e., the best outcome under $s_{n}$ is weakly worse than the worst outcome under $s_{n}^{\prime}$. In addition, we require the dominated strategy to be non-equivalent in terms of the induced outcome to the strategy that dominates it. The second requirement guarantees that the set of non-obviously dominated strategies is non-empty. The earlier intuition then translates to:

• Proof.

To show that $s_{n}$ is obviously dominated, we explicitly construct another strategy $s_{n}^{\prime}$ that obviously dominates it. Consider the information set $I_{n}\in\mathcal{I}_{n}$ (with observed prices $\bm{p}$) such that $s_{n}(I_{n})=\emptyset$ and

• –

  $\exists\bm{p}^{\prime}\in P$, $\tilde{\bm{b}}\in\arg\max\limits_{\bm{b}\in\mathcal{B}(\mathcal{M})}\,\sum_{b^{\prime}\in\bm{b}}\,p^{\prime}_{b^{\prime}}$ and $\displaystyle b\in\tilde{\bm{b}}$ such that $\bm{p}^{\prime}\geq\bm{p}$ and $v^{n}_{b}>p^{\prime}_{b}>p_{b}$.

Obviously, the payoff from $s_{n}$ conditional on any history $h\in I_{n}$ is zero. Let $s_{n}^{\prime}$ be the same as $s_{n}$ before the information set $I_{n}$ and let bidder $n$ bid $s_{n}^{\prime}(I_{n})=(b,p^{\prime}_{b})$ and never revise her bid afterwards.

We first discuss a special case in which $|\mathcal{N}^{t-1}|=1$ and $n$ is not a current leading bidder (otherwise quitting the auction is not feasible for bidder $n$). Then, all the current prices must be $0$ and all the other bidders have quit (as this is the unique consistent history). Following the strategy $s_{n}^{\prime}$, the auction ends with bidder $n$ bidding $(b,p^{\prime}_{b})$, leading to a positive payoff of $v_{b}^{n}-p_{b}^{\prime}>0$ for bidder $n$.

Next, we consider the case in which $|\mathcal{N}^{t-1}|>1$. It is clear that bidder $n$ will get a nonnegative payoff regardless of the value profiles and bidding strategies of other bidders. We show that the best possible payoff for bidder $n$ following the strategy $s_{n}^{\prime}$ is positive. It suffices to consider the case in which any other subsequent active bidder bids $p^{\prime}_{b^{\prime}}$ for some bundle $b^{\prime}$ whenever possible, the auction ends with prices $\bm{p}^{\prime}$, leading to a positive payoff of $v_{b}^{n}-p_{b}^{\prime}>0$ for bidder $n$. ∎

Note that Lemma 1 suggests that our intuition is incomplete as we cannot fully rule out strategies that quit when the bidder’s value for some bundle is above the leading price for that bundle. To rule out all such strategies, it further requires the existence of some scenario under which bidder $n$ may be pivotal: a strictly profitable bid of $n$ may ever be selected in the end. Nevertheless, we show below that the non-pivotal bidders are inconsequential for our analysis.

Let $S^{n}_{NOD}(P)$ denote the set of non-obviously dominated strategies of $n$ given price grid $P$. Let $R(\bm{s},\bm{v})$ denote the revenue to seller given value profile $\bm{v}$ and strategy profile $\bm{s}$. Define

where the first limit inf is taken over $\epsilon\to 0$ and all $P$ with grid size $\epsilon$. That is, $\underline{R}_{CASA}(\bm{v})$ is the worst-case ex-post revenue from CASA under non-obviously dominated strategies in the limit where grid $P$ becomes dense.

Given the valuation profile $\bm{v}$ and menu $\mathcal{M}$, the $k^{th}$-guarantee is defined as the maximal revenue from feasible allocations within $\mathcal{M}$, taking the $k^{th}$-highest valuations for each bundle as the price.

Our key observation is that CASA achieves the $k^{th}$-guarantee as long as bidders avoid strategies that are obviously dominated. In other words, for CASA to achieve the $k^{th}$-guarantee, we only need minimal rationality on the part of the bidders.

• Proof.

Consider a grid $P$ with any grid size $\epsilon>0$. We show that as long as bidders avoid obviously dominated strategies, for any value profile $\bm{v}$,

Suppose to the contrary, this is not true. Then, there exists strategy profile $s$ (with $s_{n}$ being not obviously dominated for each bidder $n$), value profile $\bm{v}$, and $\delta>0$ such that

Evidently, there exists some bundle $b$ such that $p_{b}^{T}<v_{b}^{(k)}-\epsilon-\delta$, or equivalently, $p_{b}^{T}+\epsilon<v_{b}^{(k)}-\delta$. For each such bundle $b$, raise the price of bundle $b$ to (the closest price below) $v_{b}^{(k)}-\delta$ sequentially until we find the first pivotal bundle $\tilde{b}$ when $\max_{\bm{b}\in\mathcal{B}(\mathcal{M})}\sum_{b\in\bm{b}}p_{b}$ is strictly improved by such increase in prices.

Since $k=|\mathcal{M}|+1$ and $p_{\tilde{b}}^{T}<v_{\tilde{b}}^{(k)}-\epsilon-\delta$, there exists a bidder $n$ with $v_{\tilde{b}}^{n}\geq v_{\tilde{b}}^{(k)}$ that quits the auction before the auction ends. Let $h$ be the (on-path) history at which $n$ quits (note that at this history, $n$ must not be a leading bidder and $p_{\tilde{b}}(h)\leq p_{\tilde{b}}^{t}\leq v_{\tilde{b}}^{n}-\delta-\epsilon$, where $p_{\tilde{b}}(h)$ is the price of bundle $\tilde{b}$ at the history $h$). Let $h\in I_{n}$. Then, $s_{n}(I_{n})=\emptyset$. Let $\bm{p}^{\prime}$ be the raised prices and $\tilde{\bm{b}}\in\arg\max\limits_{\bm{b}\in\mathcal{B}(\mathcal{M})}\sum_{b^{\prime}\in\bm{b}}p^{\prime}_{b^{\prime}}$. Then $\tilde{b}\in\tilde{\bm{b}}$, and $\bm{p}^{\prime}$ satisfies the condition in Lemma 1. Therefore, $s_{n}$ is obviously dominated. We arrive at a contradiction. ∎

As we have pointed out following Lemma 1, elimination of obviously dominated strategies does not guarantee the ex-post prices to be above the $k^{th}$-highest values. To establish Theorem 1, we prove in addition that the behaviors of non-pivotal bidders are inconsequential. With Theorem 1, we say that CASA is a rank-guaranteed auction format as long as we are comfortable assuming that bidders avoid strategies that are obviously dominated.

CASA and its rank-guarantee become extremely simple when $|\mathcal{M}|=1$, where CASA reduces to the canonical English auction of the unique feasible bundle and the $k^{th}$-guarantee becomes the second highest value. In this special case, Theorem 1 shares the insight from Li, (2017) that English auction achieves the $2^{nd}$-guarantee via an obviously strategy-proof outcome. To understand our general result, we provide an alternative interpretation of this special case. The English auction made it obvious that a bidder should avoid losing the auction when there is a remaining surplus. Therefore, it maximally utilizes the competition among losers of the auction, pushing the price to the highest value among them. Meanwhile, it does not further screen the winner at all. CASA is exactly the multi-item generalization of this philosophy: it maximally invokes competition among losers by making it obviously suboptimal to lose with a remaining surplus. Meanwhile, it does not “care” at all which winner gets which bundle and how much extra he pays. As a result, the auctioneer extracts at least as much surplus as that from the losers only, which is exactly the rank-guarantee.