Posted Price Mechanisms for Online Allocation with

Diseconomies of Scale
Hossein Nekouyan Jazi∗ Bo Sun† Raouf Boutaba‡ Xiaoqi Tan§
arXiv:2502.02543v1 [cs.GT] 4 Feb 2025

Abstract
This paper addresses the online 𝑘-selection problem with diseconomies of scale (OSDoS), where
a seller seeks to maximize social welfare by optimally pricing items for sequentially arriving buyers,
accounting for increasing marginal production costs. Previous studies have investigated deterministic
dynamic pricing mechanisms for such settings. However, significant challenges remain, particularly in
achieving optimality with small or finite inventories and developing effective randomized posted price
mechanisms. To bridge this gap, we propose a novel randomized dynamic pricing mechanism for OS-
DoS, providing a tighter lower bound on the competitive ratio compared to prior work. Our approach
ensures optimal performance in small inventory settings (i.e., when 𝑘 is small) and surpasses existing
online mechanisms in large inventory settings (i.e., when 𝑘 is large), leading to the best-known posted
price mechanism for optimizing online selection and allocation with diseconomies of scale across vary-
ing inventory sizes.

1 Introduction
Online resource allocation has been widely studied in recent years and finds a broad range of applications
in cloud computing [Zha+15; ZLW17], network routing [AAP93; BN09; CST22], and various other online,
market-based Internet platforms. In this problem, most existing studies assume that the seller has a finite
inventory of resources before a stream of online buyers arrives, with the goal of maximizing social welfare
or profit from these resources. However, in real-world applications, sellers often face diseconomies of scale
in providing resources—meaning they incur increasing marginal costs for supplying each additional unit
of resource. For instance, in cloud computing systems, the power cost of servers increases superlinearly
as the utilization of computing resources grows [AAZ16]. Similarly, in network routing, congestion costs
(e.g., end-to-end delay) increase significantly with the rise in traffic intensity brought by users.
In this work, we study online resource allocation with increasing marginal production costs. In partic-
ular, we frame it as an online 𝑘-selection with diseconomies of scale (OSDoS) in a posted price mechanism:
A seller offers a certain item to buyers arriving one at a time in an online manner. Each buyer has a private
valuation 𝑣𝑡 for one unit of the item. The seller can produce 𝑘 units of the item in total; however, the
marginal cost of producing each unit increases as more units are produced. When the 𝑡-th buyer arrives,
the seller posts a price 𝑝𝑡 to the buyer, provided that fewer than 𝑘 units have already been produced and
allocated. If the buyer’s valuation 𝑣𝑡 exceeds 𝑝𝑡 , the buyer accepts the price and takes one unit of the item.
The objective is to maximize social welfare, defined as the sum of the utilities of all the buyers and the
revenue of the seller.
∗ University of Alberta. Email: [email protected]
† University of Waterloo. Email: [email protected]
‡ University of Waterloo. Email: [email protected]
§ University of Alberta. Email: [email protected]

1
 The incorporation of increasing marginal production costs in online resource allocation was first in-
troduced by [Blu+11] and later studied by [HK19] in online combinatorial auctions. Variants of OSDoS
have since been explored, including online convex packing and covering [Aza+16], online knapsack with
packing costs [Tan+20], and online selection with convex costs [Tan+23]. A key challenge in these prob-
lems is balancing pricing strategies. Setting prices too low early on may allocate many items to low-value
buyers, increasing production costs and lowering social welfare. Conversely, overly high prices can result
in missed opportunities to sell. Thus, pricing for 𝑘 units must carefully account for early-stage decisions
to avoid rapid growth in marginal production costs while maximizing efficiency.
To address this challenge, Huang et al. [HK19] developed optimal deterministic dynamic pricing mech-
anisms for fractional online combinatorial auctions with production costs and infinite capacity (𝑘 = ∞).
They extended this to the integral case using fractional pricing functions, achieving a competitive ratio
close to the fractional setting but with a nonzero additive loss. However, as the competitive ratio ap-
proaches the fractional lower bound, the additive loss grows unbounded, which is undesirable. To over-
come this, Tan et al. [Tan+23] studied online selection with convex costs and limited supply (𝑘 < ∞),
establishing a lower bound for the integral setting without additive loss. They further showed that the
competitive ratio of their deterministic posted price mechanism asymptotically converges to the lower
bound as k grows large. Recently, Sun et al. [Sun+24] proposed a randomized static pricing algorithm,
which samples a static price from a pre-determined distribution for OSDoS. This randomization improves
performance over the deterministic approach in small inventory settings but is not asymptotically optimal
and fails to converge to the lower bound from [Tan+23] as 𝑘 → ∞.
Despite previous efforts, two questions remain unresolved: First, how to derive a tight lower bound
for OSDoS in small inventory settings? Second, it remains an open question how to develop randomized
algorithms to solve OSDoS with tight guarantees, especially for settings when 𝑘 is small.
In this paper, we address these questions by deriving a new tight lower bound for the OSDoS problem,
achieving the best-known results in both small and asymptotically large inventory settings. Building on
this, we propose a novel randomized dynamic pricing algorithm that uses up to 𝑘 randomized prices. We
show that this algorithm is optimal for small inventories and outperforms existing designs from [Tan+23]
and [Sun+24] in large inventory settings.

1.1 Overview of Main Results and Techniques

The primary contribution of this paper is the development of novel posted price mechanisms using ran-
domized dynamic pricing schemes that extend the results in [Blu+11; HK19; Sun+24; Tan+23]. The pro-
posed scheme, r-Dynamic, sequentially updates the item’s price as new units are produced and sold. Specif-
ically, as the marginal production cost increases with each additional unit, r-Dynamic utilizes a different
cumulative distribution function (CDF) to independently randomize the price for each unit. The main
lower bound result is as follows:

Theorem 1 (Informal Statement of Theorem 3). Assume that buyers’ valuations are bounded within the
range [𝐿, 𝑈 ] and the cumulative cost of production up to the 𝑖-th unit is given by 𝑓 (𝑖). The seller can produce
a total of 𝑘 units. For any given 𝑘 ≥ 1, 𝑈 ≥ 𝐿 ≥ 1, and a cumulative production cost function 𝑓 , no online
algorithm can be (𝛼 ∗S (𝑘) − 𝜖)-competitive for any 𝜖 > 0, where S := {𝐿, 𝑈 , 𝑓 }.

We note that [Tan+23] also established a lower bound for the competitive ratio of online algorithms for
OSDoS, but it was derived by connecting the integral selection problem to its fractional counterpart. This
approach requires the cumulative production cost function 𝑓 to be defined not only at discrete points but
also for all fractional values in [0, 𝑘], leading to two issues: (i) assuming the availability of a continuous
cost function 𝑓 may be impractical for an inherently integral problem, and (ii) the lower bound is only

2
 5.2
r-Static
5.0 d-Dynamic
r-Dynamic

Competitve Ratio
4.8
Lower-Bound ( *)
4.6

4.4

4.2

4.0
5 10 15 20 25 30 35 40
Number of Items (k)
Figure 1: The blue curve (i.e., r-Dynamic) corresponds to the competitive ratio of Algorithm 1 that uses
randomized dynamic pricing. The red curve (i.e., d-Dynamic) and the yellow curve (i.e., r-Static) corre-
spond to the competitive ratios of the deterministic dynamic pricing mechanism developed by [Tan+23]
and the static randomized pricing mechanism by [Sun+24]. In this figure, we set 𝐿 = 1, 𝑈 = 10, and
𝑖2
𝑓 (𝑖) = 59 .

tight in large inventory settings, as it assumes 𝑘 → ∞.

In this paper, we address these two issues by deriving the lower bound 𝛼 ∗S (𝑘) via a totally different
approach. In particular, we do not rely on results in the fractional setting and only need to assume that the
cost function 𝑓 (𝑖) is defined at discrete points for 𝑖 ∈ {1, 2, · · · , 𝑘 }, leading to the tight lower bound 𝛼 ∗S (𝑘)
for all 𝑘 ≥ 1.

Theorem 2 (Informal Statement of Theorem 4). For any given 𝑘 ≥ 1, 𝑈 ≥ 𝐿 ≥ 1, and a cumulative
production cost function 𝑓 , there existsa randomized dynamic price mechanism (r-Dynamic) that achieves a
𝛼 ∗ (𝑘 )
competitive ratio of 𝛼 ∗S (𝑘) · exp S𝑘 . In addition, when 𝑘 = 2, r-Dynamic is 𝛼 ∗S (2)-competitive.

Due to the arbitrary nature of the cost function 𝑓 , neither our work nor [Tan+23] can derive a closed-
form expression for the competitive ratio, preventing a direct comparison between our r-Dynamic and the
deterministic dynamic pricing mechanism (d-Dynamic) in [Tan+23]. In Figure 1, we compare the asymp-
totic performance of r-Dynamic with d-Dynamic from [Tan+23] and the randomized static pricing mech-
anism (r-Static) in [Sun+24]. The results show that r-Dynamic significantly outperforms both d-Dynamic
and r-Static, converging faster to the lower bound as 𝑘 → ∞. Notably, for small 𝑘, r-Dynamic achieves
the lower bound when 𝑘 = 2. Beyond its strong theoretical guarantees, empirical results (Section 4.2)
further confirm that r-Dynamic consistently outperforms both d-Dynamic and r-Static, highlighting its
superiority over existing designs.
The key technical component in deriving the above lower and upper bounds is a new representative
function-based approach, which models the dynamics of any randomized online algorithm using a se-
quence of 𝑘 probability functions, {𝜓𝑖 }𝑖 ∈ [𝑘 ] . We design a family of hard instances and characterize the
performance of any 𝛼-competitive algorithm on these instances through a set of differential equations in-
volving {𝜓𝑖 }𝑖 ∈ [𝑘 ] . To determine the lower bound 𝛼 ∗S (𝑘) in Theorem 1, we compute the minimum 𝛼 for
which these equations have a feasible solution, namely valid probability functions {𝜓𝑖 }𝑖 ∈ [𝑘 ] . By reverse
engineering the equations, we derive inverse probability functions, {𝜙𝑖 }𝑖 ∈ [𝑘 ] , for pricing each unit, which
leads to r-Dynamic in Theorem 2.

3
1.2 Other Related Work
Online resource allocation—the process of assigning limited resources to a sequence of online requests to
maximize social welfare or profit—has been a central topic in computer science and operations research. In
addition to the previously mentioned related work, readers are referred to the survey by Gupta and Singla
on the secretary problem [GS21] for a detailed discussion of online allocation and selection in random-
order models. Significant advancements have also been made in studying the prophet inequality through
the lens of posted price mechanisms [Cor+19; Luc17] and in online matching with applications to Internet
advertising (e.g., [HTW24; Meh13]). Beyond the stochastic i.i.d. model in the prophet inequality, recent
work explores the correlated arrival model based on a Markov chain [Jia+23]. However, these studies focus
on variants of online allocation and selection without considering production costs. In contrast, our work
primarily examines the impact of increasing marginal production costs on online 𝑘-selection.
Recent years have seen efforts to study online allocation problems with various forms of production
costs in stochastic settings (e.g., [BMY15], [GMM18], [Bar+12], [Sek17]). For instance, [BMY15] examined
online allocation with economies of scale (decreasing marginal costs), proposing a constant-competitive
strategy for unit-demand customers with valuations sampled i.i.d. from an unknown distribution. In con-
trast, [Sek17] addressed Bayesian online allocation with convex production costs (diseconomies of scale),
developing posted price mechanisms with 𝑂 (1)-approximation for fractionally subadditive buyers and log-
arithmic approximations for subadditive buyers. Our study differs by focusing on OSDoS in adversarial
settings, assuming no knowledge of the arrival sequence beyond the finite support of valuations, making
these results not directly comparable to ours.
On the applied side, allocating limited resources under diseconomies of scale is common across various
online platforms. For example, in online cloud resource allocation [Zha+15], convex server costs model
energy consumption based on CPU utilization, while in online electric vehicle charging [STT18], electricity
generation costs are often modeled as nonlinear, typically quadratic.

2 Problem Statement and Assumptions

We formally define online 𝑘-selection with diseconomies of scale (OSDoS) as follows. Consider an online
market operating under posted price mechanisms. On the supply side, a seller can produce a total of 𝑘
units of an item, with increasing (or at least non-decreasing) marginal production costs. Let 𝒄 := {𝑐𝑖 } ∀𝑖 ∈ [𝑘 ]
represent the marginal production cost, where 𝑐𝑖 denotes the cost of producing the 𝑖-th unit, and 𝑐 1 ≤ 𝑐 2 ≤
Í
· · · ≤ 𝑐𝑘 . Define 𝑓 (𝑖) = 𝑖𝑗=1 𝑐 𝑗 as the cumulative production cost of the first 𝑖 units. On the demand side,
𝑇 buyers arrive sequentially, each demanding one unit of the item. Let 𝑣𝑡 denote the private valuation of
the 𝑡-th buyer. Once buyer 𝑡 arrives, a price 𝑝𝑡 is posted, and then the buyer decides to accept the price
and make a purchase if a non-negative utility is gained 𝑣𝑡 − 𝑝𝑡 ≥ 0, and reject it otherwise.
Let 𝑥𝑡 ∈ {0, 1} represent the decision of buyer 𝑡, where 𝑥𝑡 = 1 indicates a purchase and 𝑥𝑡 = 0 otherwise.
Í Í
Then buyer 𝑡 obtains a utility (𝑣𝑡 −𝑝𝑡 )𝑥𝑡 and the seller collects a total revenue of 𝑡 ∈ [𝑇 ] 𝑝𝑡 𝑥𝑡 − 𝑓 ( 𝑡 ∈ [𝑇 ] 𝑥𝑡 )
from all buyers. The goal of the online market is to determine the posted prices {𝑝𝑡 } ∀𝑡 ∈ [𝑇 ] to maximize
the social welfare, which is the sum of utilities of all the buyers and the revenue of the producer, i.e.,
Í Í Í Í Í
𝑡 ∈ [𝑇 ] 𝑥𝑡 · (𝑣 𝑡 − 𝑝𝑡 ) + 𝑡 ∈ [𝑇 ] 𝑥𝑡 · 𝑝𝑡 − 𝑓 ( 𝑡 ∈ [𝑇 ] 𝑥𝑡 ) = 𝑡 ∈ [𝑇 ] 𝑣 𝑡 𝑥𝑡 − 𝑓 ( 𝑡 ∈ [𝑇 ] 𝑥𝑡 ).
Let I = {𝑣 1, · · · , 𝑣𝑇 } denote an arrival instance of buyers. An optimal offline algorithm that knows all
the information of I can obtain the optimal social welfare OPT(I) by solving the following optimization
problem
∑︁ ∑︁  ∑︁
OPT(I) = max 𝑣 𝑡 𝑥𝑡 − 𝑓 𝑥𝑡 , s.t. 𝑥𝑡 ≤ 𝑘.
𝑥𝑡 ∈ {0,1} 𝑡 ∈ [𝑇 ] 𝑡 ∈ [𝑇 ] 𝑡 ∈ [𝑇 ]

4
However, in the online market, the posted price 𝑝𝑡 is determined without knowing the valuations of future
buyers {𝑣𝜏 }𝜏 >𝑡 . We aim to design an online mechanism to determine the posted prices such that the social
welfare achieved by the online mechanism, denoted by ALG(I), is competitive compared to OPT(I).
Specifically, an online algorithm is 𝛼-competitive if for any input instance I, the following inequality
holds:
OPT(I)
𝛼≥ ,
E[ALG(I)]
where the expectation of E[ALG(I)] is taken with respect to the randomness of the online algorithm. To
attain a bounded competitive ratio, we consider a constrained adversary model [Jia+21; Tan+23], where
the buyers’ valuations are assumed to be bounded.

Assumption 1. Buyers’ valuations are bounded in [𝐿, 𝑈 ], i.e., 𝑣𝑡 ∈ [𝐿, 𝑈 ], ∀𝑡 ∈ [𝑇 ].

The interval [𝐿, 𝑈 ] can be considered as the prediction interval that covers the valuations of all buy-
ers [Jia+21], and is known to the online algorithm. As shown in [Tan+23], the competitive analysis of
online algorithms for OSDoS depends on the relationship between buyers’ valuations and the production
cost function. For simplicity, we focus on the case where the production cost is always smaller than the
buyer’s valuation (𝑐𝑘 < 𝐿) and derive lower and upper bounds in Sections 3 and 4, respectively. In Appen-
dices I and J, we show that this assumption is without loss of generality, as our results extend naturally to
the general case.

3 Lower Bound for OSDoS: Hardness of Allocation with Diseconomies

of Scale
We first derive a tight lower bound for OSDoS, which informs the design of r-Dynamic (Algorithm 1) in
Section 4.

3.1 Lower Bound 𝛼 S∗ (𝑘)

Theorem 3 below formally states the lower bound 𝛼 ∗S (𝑘) for the competitive ratio of any online algorithm
for OSDoS.

Theorem 3 (Lower Bound). Given S = {𝐿, 𝑈 , 𝑓 } for the OSDoS problem with 𝑘 ≥ 1, no online algorithm,
including those with randomization, can achieve a competitive ratio smaller than 𝛼 ∗S (𝑘), where 𝛼 ∗S (𝑘) is the
solution to the following equation of 𝛼:

𝛼 𝛼
 𝛼
 𝛼  𝛼

𝑈 = 𝐿 − 𝑐𝑘 · 𝑒 𝑘 · (𝑘+1−𝑘¯ −𝜉 ) + 𝑐𝑘 · 𝑒 𝑘 · (𝑘 −𝑘¯ ) + 𝑐𝑘+1 · 1 − 𝑒 𝑘 · 𝑒 𝑘 · (𝑘 −1−𝑘¯ ) + · · · + 𝑐𝑘 · 1 − 𝑒 𝑘 . (1)
¯ ¯ ¯

In Eq. (1), 𝑘 ∈ [𝑘] denotes the smallest natural number such that
¯
∑︁𝑘
¯ (𝐿 − 𝑐 𝑖 ) ≥
1  ∑︁𝑘 
· 𝑘𝐿 − 𝑐𝑖 , (2)
𝑖=1 𝛼 𝑖=1

and 𝜉 ∈ (0, 1] denotes the unique solution to the following equation

1 Í𝑘 Í𝑘 −1
𝛼 · (𝑘𝐿 − 𝑖=1 𝑐 𝑖 )− ¯
𝑖=1 (𝐿 − 𝑐𝑖 )
𝜉= . (3)
𝐿 − 𝑐𝑘
¯

5
 Theorem 3 is our main result concerning the hardness of OSDoS. To prove Theorem 3, a key step is
to establish a set of necessary conditions that any 𝛼-competitive online algorithm must satisfy. A formal
proof will be provided in Section 3.3. Below, we offer several remarks to clarify the key intuitions.
• By the definition of 𝑘 in Eq. (2), 𝑘 represents the minimum number of units that any 𝛼-competitive
¯ ¯
deterministic algorithm, denoted by ALGd , must sell when faced with an arrival instance of 𝑘 identical
(𝐿) (𝐿)
buyers with valuation 𝐿, denoted by I𝑖𝑑𝑒𝑛 = {𝐿, · · · , 𝐿}. Under the instance I𝑖𝑑𝑒𝑛 , the maximum social
Í𝑘
welfare achievable by the offline optimal algorithm is 𝑘𝐿 − 𝑖=1 𝑐𝑖 . Therefore, ALGd must sell at least 𝑘
¯
units to ensure 𝛼-competitiveness, implying that 𝑘 is well-defined for all values of 𝛼 ≥ 1.
¯
• Eq. (3) demonstrates that 𝜉 is defined as the fraction of the 𝑘-th unit required to make Eq. (2) binding. We
¯
argue that 𝜉 ∈ (0, 1] is well-defined and always exists as long as there is an 𝛼-competitive randomized
algorithm, denoted as ALGr . Specifically, if a randomized algorithm ALGr is run on the same instance
(𝐿)
I𝑖𝑑𝑒𝑛 , ALGr must sell at least 𝑘 − 1 units plus a fraction 𝜉 of the 𝑘-th unit of the item, in expectation.
¯ ¯
• Note that, in general, a closed-form expression for the lower bound 𝛼 ∗S (𝑘) cannot be derived. This is
expected due to the arbitrary nature of the sequence of marginal production costs. However, because
of the monotonicity of 𝑘, 𝜉, and the right-hand side of Eq. (1) with respect to 𝛼, 𝛼 ∗S (𝑘) can be easily
¯
computed by solving Eq. (1) numerically using binary search.
In the next subsection, we construct a family of hard instances and introduce a novel representative
function-based approach to derive a system of differential equations, which are crucial to proving the lower
bound result in Theorem 3.

3.2 Representing Worst-Case Performance by (Probabilistic) Allocation Functions

3.2.1 Hard Instances {I𝑣(𝜖 ) } ∀𝑣 ∈ [𝐿,𝑈 ]

We introduce a family of hard instances based on the instance I (𝜖 ) defined as follows.

Definition 1 (Instance I (𝜖 ) ). For any given value of 𝜖 > 0, the instance I (𝜖 ) begins with 𝑘 identical
buyers, each having a valuation of 𝐿 during the initial stage. This is followed by a series of stages, each
consisting of 𝑘 identical buyers,  valuations incrementally increasing by 𝜖, starting from 𝐿 + 𝜖 and
 with
reaching the upper bound 𝐿 + 𝑈 𝜖−𝐿 · 𝜖. The instance I (𝜖 ) is mathematically defined as:
(     )
𝑈 −𝐿 𝑈 −𝐿
𝐿, . . . , 𝐿, 𝐿 + 𝜖, . . . , 𝐿 + 𝜖 , . . . , 𝐿 + 𝑗 · 𝜖, . . . , 𝐿 + 𝑗 · 𝜖 , . . . , 𝐿 + · 𝜖, . . . , 𝐿 + ·𝜖 ,
| {z } | {z } | {z } 𝜖 𝜖
𝑘 buyers 𝑘 buyers 𝑘 buyers in stage 𝐿+𝑗 ·𝜖 | {z }
𝑘 buyers

where 𝑗 ranges from 1 to ⌊(𝑈 − 𝐿)/𝜖⌋. Furthmore, let us define the set 𝑉 (𝜖 ) = {𝐿, 𝐿 +𝜖, . . . , 𝐿 + ⌊(𝑈 − 𝐿)/𝜖⌋ ·
𝜖} to contain all the possible valuations that buyers in the instance I (𝜖 ) may possess.

We refer to the 𝑘 buyers with valuation 𝑣 ∈ 𝑉 (𝜖 ) as stage-𝑣 arrivals in I (𝜖 ) . For any 𝑣 ∈ 𝑉 (𝜖 ) , let I𝑣(𝜖 )
denote all the buyers in I (𝜖 ) from the beginning up to stage-𝑣. For instance, if 𝑣 = 𝐿 +2𝜖, then I𝑣(𝜖 ) includes
the first 3𝑘 buyers in I (𝜖 ) with valuations 𝐿, 𝐿 + 𝜖, and 𝐿 + 2𝜖. Due to the online nature of the problem, we
emphasize that I (𝜖 ) may terminate at any stage 𝑣. In other words, there exists a family of hard instances,
{I𝑣(𝜖 ) } ∀𝑣 ∈𝑉 (𝜖 ) , induced by I (𝜖 ) . Here, I𝑣(𝜖 ) denotes the arrival instance of I (𝜖 ) that terminates at stage-𝑣.
Henceforth, we will use “instance I𝑣(𝜖 ) " and “instance I (𝜖 ) by the end of stage-𝑣" interchangeably.
Given any 𝛼-competitive algorithm ALG, an arbitrary instance from {I𝑣(𝜖 ) } ∀𝑣 ∈𝑉 (𝜖 ) may be the one that
ALG processes. Thus, for any 𝑣 ∈ 𝑉 (𝜖 ) , by the end of stage-𝑣 of I (𝜖 ) , ALG must achieve at least a 1/𝛼
Í
fraction of the optimal social welfare, 𝑘𝑣 − 𝑘𝑖=1 𝑐𝑖 , which is attained by rejecting all previous buyers except

6
for the last 𝑘 buyers with valuation 𝑣. Consequently, an 𝛼-competitive algorithm must ensure
  1  ∑︁𝑘 
ALG I𝑣(𝜖 ) ≥ · 𝑘𝑣 − 𝑐𝑖 , ∀𝑣 ∈ 𝑉 (𝜖 ) , (4)
𝛼 𝑖=1

where ALG(I𝑣(𝜖 ) ) denotes the expected performance of ALG under the instance I𝑣(𝜖 ) .

3.2.2 Representing ALG(I𝑣(𝜖 ) ) by Allocation Functions

For any randomized algorithm, we define 𝑘 + 1 states, {𝑞𝑖 } ∀𝑖 ∈ {0,··· ,𝑘 } , which represent the allocation be-
havior of the online algorithm at any stage of instance I (𝜖 ) , as follows:
• State 𝑞 0 corresponds to the situation where the online algorithm has not allocated any units.
• For all 𝑖 ∈ [𝑘], state 𝑞𝑖 represents that the online algorithm has allocated at least 𝑖 units of the item.
For all 𝑣 ∈ 𝑉 (𝜖 ) and 𝑖 ∈ {0, · · · , 𝑘 }, we define Ψ𝑖 (𝑣) : 𝑉 (𝜖 ) → {0, 1} such that Ψ𝑖 (𝑣) = 1 if the algorithm
is in state 𝑞𝑖 after processing all the buyers in I𝑣(𝜖 ) , and Ψ𝑖 (𝑣) = 0 otherwise. Specifically, Ψ𝑖 (𝑣) = 1 if the
online algorithm allocates at least 𝑖 units of the item at the end of stage 𝑣 in I (𝜖 ) , which occurs with some
probability depending on the algorithm’s randomness. Since the instance I (𝜖 ) is deterministically defined,
Ψ𝑖 (𝑣) is a binary random variable whose distribution depends solely on the algorithm’s randomness. This
leads to the definition of 𝝍 = {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] below.

Definition 2 (Allocation Functions). For any randomized online algorithm, let 𝝍 = {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] and 𝜓𝑖 :
𝑉 (𝜖 ) → [0, 1] represent the functions where 𝜓𝑖 (𝑣) = E[Ψ𝑖 (𝑣)], with the expectation taken over the ran-
domness of the algorithm.

Based on the definition above, we have 𝜓𝑖 (𝑣) = Pr(Ψ𝑖 (𝑣) = 1), where Ψ𝑖 (𝑣) = 1 indicates that the
algorithm is in state 𝑞𝑖 (i.e., at least 𝑖 units of the item have been allocated) after processing all buyers in
I𝑣(𝜖 ) (i.e., by the end of stage 𝑣 of instance I (𝜖 ) ). In this context, 𝜓𝑖 (𝑣) represents the probability that the
online algorithm has allocated at least 𝑖 units of the item by the end of stage 𝑣 in instance I (𝜖 ) . Therefore,
the term probabilistic allocation functions is used or simply allocation functions for brevity. We show that
𝜓𝑖 (𝑣) is monotonic in 𝑖 ∈ [𝑘].

Lemma 1 (Monotonicity). For any randomized online algorithm, 𝜓𝑖 (𝑣) ≥ 𝜓𝑖+1 (𝑣) holds for all 𝑖 ∈ [𝑘] and
𝑣 ∈ [𝐿, 𝑈 ].

The proof of the above lemma is given in Appendix A. Lemma 1 implies that it suffices to focus on
randomized algorithms whose allocation functions are from the following set
n o
Ω = 𝝍 𝜓𝑖 (𝑣) ∈ [0, 1],𝜓𝑖 (𝑣) ≥ 𝜓𝑖+1 (𝑣),𝜓𝑖 (𝑣) ≤ 𝜓𝑖 (𝑣 ′ ), ∀𝑖 ∈ [𝑘], 𝑣, 𝑣 ′ ∈ 𝑉 (𝜖 ) , and 𝑣 < 𝑣 ′ .

Next, we analyze how the allocation level of an 𝛼-competitive algorithm should evolve as new buyers
with higher valuations arrive in I (𝜖 ) . We argue that the expected performance of any online algorithm
under the instance I (𝜖 ) can be fully represented by the 𝑘 allocation functions {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] . Let ALG(I𝑣(𝜖 ) )
denote the expected objective value of the algorithm under instance I𝑣(𝜖 ) . Then ALG(I𝑣(𝜖 ) ) can be framed
using 𝝍 = {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] as follows.

Proposition 1 (Representation based on 𝝍). For any randomized algorithm ALG under the family of hard in-
stances {I𝑣(𝜖 ) } ∀𝑣 ∈𝑉 (𝜖 ) , its expected performance can be represented by its allocation functions {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] ∈

7
Ω as follows:
 𝑘
 ∑︁
ALG I𝐿(𝜖 ) = 𝜓𝑖(𝐿) · (𝐿 − 𝑐𝑖 ),
𝑖=1
𝑈 −𝐿
    ∑︁ 𝜖 ⌉ h
𝑘 ⌈ ∑︁  i
(𝜖 )
ALG I𝐿+𝑗 ·𝜖 = ALG I𝐿(𝜖 ) + (𝐿 + 𝑚 · 𝜖) · 𝜓𝑖 (𝐿 + 𝑚 · 𝜖) − 𝜓𝑖 (𝐿 + (𝑚 − 1) · 𝜖) ,
𝑖=1 𝑚=1
𝑈 − 𝐿 
∀𝑗 = 1, 2, . . . , .
𝜖
The above proposition relates the expected performance of an online algorithm to the set of allocation
functions {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] that capture its dynamics under hard instances {I𝑣(𝜖 ) } ∀𝑣 ∈𝑉 (𝜖 ) . The detailed proof
can be found in Appendix B.
Combining Proposition 1 and Eq. (4) gives the lemma below.
Lemma 2 (Necessary Conditions). If there exists an 𝛼-competitive algorithm for OSDoS, then there exists
𝑘 allocation functions {𝜓𝑖 }𝑖 ∈ [𝑘 ] ∈ Ω, where each function 𝜓𝑖 : [𝐿, 𝑈 ] → [0, 1] is continuous within its range
and also satisfies the following equation:
𝑘 𝑘 ∫ 𝑣 𝑘
!
∑︁ ∑︁ 1 ∑︁
𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖 (𝜂) ≥ · 𝑘𝑣 − 𝑐𝑖 , ∀𝑣 ∈ [𝐿, 𝑈 ]. (5)
𝑖=1 𝑖=1 𝜂=𝐿
𝛼 𝑖=1

The above result is derived based on the family of instances {I𝑣(𝜖 ) } ∀𝑣 ∈𝑉 (𝜖 ) when 𝜖 approaches to zero.
The proof is given in Appendix C. The lemma above provides a set of necessary conditions for the allocation
functions {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] induced by any 𝛼-competitive algorithm. Therefore, determining a tight lower bound
for OSDoS is equivalent to finding the lowest 𝛼 such that there exists a set of allocation functions in Ω
that satisfy Eq. (5).

3.3 Proof of Theorem 3

We now move on to prove Theorem 3. Based on the necessary conditions in Lemma 2, the lower bound
can be defined as
n o
𝛼 ∗S (𝑘) = inf 𝛼 ≥ 1 there exist a set of 𝑘 allocation functions {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] ∈ Ω that satisfy Eq. (5) .

Next, we show that it is possible to find a tight design of {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] that satisfies the necessary conditions
in Eq. (5) by equality, ultimately leading to Eq. (1) in Theorem 3.
For any 𝛼 ≥ 𝛼 ∗S (𝑘), let Γ (𝛼 ) denote the superset of the set of functions {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] ∈ Ω that satisfy Eq.
(5). Note that Γ (𝛼 ) ⊂ Ω holds for all 𝛼 ≥ 𝛼 ∗S (𝑘). Define 𝜒 (𝛼 ) (𝑣) : [𝐿, 𝑈 ] → [0, 𝑘] as
n∑︁𝑘 o
𝜒 (𝛼 ) (𝑣) = inf 𝜓𝑖 (𝑣) {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] ∈ Γ (𝛼 ) . (6)
𝑖=1

Based on the definition of 𝜒 (𝛼 ) , we construct a set of allocation functions {𝜓𝑖(𝛼 ) (𝑣)} ∀𝑖 ∈ [𝑘 ] as follows:
 
𝜓𝑖(𝛼 ) (𝑣) = 𝜒 (𝛼 ) (𝑣) − (𝑖 − 1) · 1 {𝑖 −1≤ 𝜒 (𝛼 ) (𝑣) ≤𝑖 } + 1 { 𝜒 (𝛼 ) (𝑣) >𝑖 } , ∀𝑣 ∈ [𝐿, 𝑈 ], ∀𝑖 ∈ [𝑘], (7)

where 1 {𝐴} is the standard indicator function, equal to 1 if 𝐴 is true and 0 otherwise. In the following
lemma, we argue that the set of functions {𝜓𝑖(𝛼 ) (𝑣)} ∀𝑖 ∈ [𝑘 ] is a feasible solution to Eq. (5) and satisfies it as
an equality.

8
Lemma 3. For any 𝛼 ≥ 𝛼 ∗S (𝑘), the functions {𝜓𝑖(𝛼 ) } ∀𝑖 ∈ [𝑘 ] satisfy Eq. (5) as an equality.

The detailed proof for the above lemma is in Appendix D. Following the definition of {𝜓𝑖(𝛼 ) (𝑣)} ∀𝑖 ∈ [𝑘 ] ,
we observe that these functions exhibit the following property:

Lemma 4. For any 𝑖 ∈ [𝑘] and 𝑣 ∈ [𝐿, 𝑈 ], if 𝜓𝑖(𝛼 ) (𝑣) ∈ (0, 1) holds, then 𝜓 𝑗(𝛼 ) (𝑣) = 1 for all 𝑗 = 1, · · · , 𝑖 − 1
and 𝜓 𝑗(𝛼 ) (𝑣) = 0 for all 𝑗 = 𝑖 + 1, · · · , 𝑘.

Lemma 4 asserts that if the online algorithm inducing {𝜓 (𝛼 ) } ∀𝑖 begins allocating unit 𝑖 with some
positive probability to buyers in stage-𝑣 of I (𝜖 ) , then the algorithm must have already allocated all units
𝑗 < 𝑖 with probability one to buyers arriving at or before stage-𝑣 of I (𝜖 ) . Furthermore, if the algorithm
has not allocated unit 𝑖 with probability one by the end of stage-𝑣, then all units 𝑗 > 𝑖 remain in the system
with probability one at the end of stage-𝑣. Given that the marginal cost for each additional unit of resource
increases, the algorithm should only produce and allocate a new unit once all previously produced units
have been fully allocated.
According to Lemma 3, the inequality in Eq. (5) can be replaced with an equality. By combining Lemma
3 with Lemma 4, we conclude that there exists a unique set of functions that satisfy Eq. (5) as an equality
and also fulfill the property stated in Lemma 4. Proposition 2 below formally states this result.

Proposition 2. For any 𝛼 ≥ 𝛼 ∗S (𝑘), there exist a set of allocation functions {𝜓𝑖(𝛼 ) } ∀𝑖 ∈ [𝑘 ] ∈ Ω that satisfy Eq.
(5) by equality:

𝜓𝑖(𝛼 ) (𝑣) = 1, 𝑖 = 1, . . . , 𝑘 − 1,
(  𝑣−𝑐¯ 
𝑘 𝑘
𝜉 + · ln ¯ 𝑣 ∈ [𝐿, 𝑢𝑘 ],
𝜓𝑘(𝛼 ) (𝑣) = 𝛼 𝐿−𝑐𝑘
¯ ¯
¯ 1 𝑣 > 𝑢𝑘 ,
¯

0

 𝑣 ≤ ℓ𝑖 ,
 
𝜓𝑖(𝛼 ) (𝑣) = 𝛼𝑘 · ln ℓ𝑣−𝑐


𝑖
𝑖 −𝑐 𝑖
𝑣 ∈ [ℓ 𝑖 , 𝑢𝑖 ], 𝑖 = 𝑘 + 1, . . . , 𝑘 − 1,

 ¯

1
 𝑣 ≥ 𝑢 𝑖 ,
(
(𝛼 )
0 𝑣 ≤ ℓ𝑘 ,
𝜓𝑘 (𝑣) = 𝑘 
𝑣−𝑐𝑘

𝛼 · ln ℓ𝑘 −𝑐𝑘 𝑣 ∈ [ℓ𝑘 , 𝑈 ],

where the intervals {[ℓ𝑖 , 𝑢𝑖 ]} ∀𝑖 are specified by

𝛼
𝑢𝑘 = ℓ𝑘+1 = (𝐿 − 𝑐𝑘 ) · 𝑒 (1−𝜉 ) · 𝑘 + 𝑐𝑘 , (8)
¯ ¯ ¯ ¯
𝛼/𝑘
𝑢𝑖 = ℓ𝑖+1 = (ℓ𝑖 − 𝑐𝑖 ) · 𝑒 + 𝑐𝑖 ∀𝑖 = 𝑘 + 1, . . . , 𝑘. (9)
¯
Recall that the parameters 𝑘 and 𝜉 are defined in Eq. (2) and Eq. (3), respectively. Once 𝛼 is given,
¯
both 𝑘 and 𝜉 can be uniquely determined. Therefore, the set of allocation functions {𝜓𝑖(𝛼 ) } ∀𝑖 ∈ [𝑘 ] given in
¯
Proposition 2 can also be explicitly computed once 𝛼 is given. The full proof of how to derive the explicit
designs of {𝜓𝑖(𝛼 ) } ∀𝑖 ∈ [𝑘 ] is given in Appendix E.
Putting together Eq. (8) and Eq. (9), we have
𝛼 𝛼 𝛼 𝛼 𝛼
𝑢𝑘 = (𝐿 − 𝑐𝑘 ) · 𝑒 𝑘 · (𝑘+1−𝑘¯ −𝜉 ) + 𝑐𝑘 · 𝑒 𝑘 · (𝑘 −𝑘¯ ) + 𝑐𝑘+1 · (1 − 𝑒 𝑘 ) · 𝑒 𝑘 · (𝑘 −1−𝑘¯ ) + · · · + 𝑐𝑘 · (1 − 𝑒 𝑘 ).
¯ ¯ ¯
Note that the right-hand side of the equation above is increasing in 𝛼. Therefore, as 𝛼 decreases, the value
of 𝑢𝑘 also decreases and will eventually fall below 𝑈 for a specific value of 𝛼. Consequently, according

9
Algorithm 1 Randomized Dynamic Pricing (r-Dynamic) for OSDoS
1: Input: pricing functions {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] ;
2: Initiate: index of the unit to be sold 𝜅 1 = 1;
3: Generate a random seed vector 𝒔 = {𝑠𝑖 } ∀𝑖 ∈ [𝑘 ] , each element sampled independently from uniform
distribution Unif(0, 1);
4: Set a price vector P = {𝑃𝑖 } ∀𝑖 ∈ [𝑘 ] , where 𝑃𝑖 = 𝜙𝑖 (𝑠𝑖 );
5: while buyer 𝑡 arrives do
6: if 𝜅𝑡 ≤ 𝑘 then:
7: Post the price 𝑝𝑡 = 𝑃𝜅𝑡 to buyer 𝑡;
8: if buyer 𝑡 accepts the price then
9: One unit is sold and set 𝑥𝑡 = 1;
10: end if
11: end if
12: Update 𝜅𝑡 +1 = 𝜅𝑡 + 𝑥𝑡 . ⊲ 𝑥𝑡 = 0 if buyer 𝑡 declines 𝑝𝑡 .
13: end while

to the definition of 𝜓𝑘(𝛼 ) in Proposition 2, 𝜓𝑘(𝛼 ) (𝑈 ) will exceed 1 (since 𝜓𝑘(𝛼 ) (𝑈 ) > 𝜓𝑘(𝛼 ) (𝑢𝑘 ), and based on
Eq. (9), 𝜓𝑘(𝛼 ) (𝑢𝑘 ) is equal to one). However, this will generate an infeasible allocation function 𝜓𝑘(𝛼 ) , as we
require that 𝜓𝑘(𝛼 ) (𝑣) ≤ 1 holds for all 𝑣 ∈ [𝐿, 𝑈 ]. As a result, for those values of 𝛼 where 𝑢𝑘 < 𝑈 , the set of
𝑘 allocation functions {𝜓𝑖(𝛼 ) } ∀𝑖 ∈ [𝑘 ] obtained in Proposition 2 becomes infeasible, meaning that 𝛼 must be
less than 𝛼 ∗S (𝑘). Therefore, 𝛼 ∗S (𝑘) is the value of 𝛼 for which 𝑢𝑘 = 𝑈 , and this gives Eq. (1) in Theorem 3.
Thus, we complete the proof of Theorem 3.

4 r-Dynamic: A Randomized Dynamic Posted Price Mechanisms

We propose a randomized dynamic pricing mechanism (r-Dynamic), as described in Algorithm 1, to solve
the OSDoS problem. Before the buyers arrive, r-Dynamic samples 𝑘 independent random prices {𝑃𝑖 } ∀𝑖 ∈ [𝑘 ] ,
where 𝑃𝑖 is the price for the 𝑖-th unit of the item. Specifically, for each unit 𝑖 ∈ [𝑘], a random seed 𝑠𝑖 is
drawn from the uniform distribution Unif(0, 1), and the random price is set as 𝑃𝑖 = 𝜙𝑖 (𝑠𝑖 ), where 𝜙𝑖 (𝑠𝑖 ) is
the pricing function designed for the 𝑖-th unit. r-Dynamic then posts the price of the available unit with
the smallest index from {𝑃𝑖 } ∀𝑖 ∈ [𝑘 ] to the online arriving buyers.
For all 𝑖 ∈ [𝑘], the pricing function 𝜙𝑖 : [0, 1] → [𝐿𝑖 , 𝑈𝑖 ] is constructed such that the 𝑘 price intervals
{[𝐿𝑖 , 𝑈𝑖 ]} ∀𝑖 ∈ [𝑘 ] span the entire range of [𝐿, 𝑈 ], where 𝐿 = 𝐿1 ≤ 𝑈 1 = 𝐿2 ≤ 𝑈 2 ≤ · · · ≤ 𝑈𝑘 −1 = 𝐿𝑘 ≤ 𝑈𝑘 = 𝑈 .
That is, the upper boundary of 𝜙𝑖 (i.e., the maximum price of 𝑃𝑖 ) is the lower boundary of 𝜙𝑖+1 (i.e., the
minimum price of 𝑃𝑖+1 ). As a result, the posted prices will always be non-decreasing (i.e., 𝑃 1 ≤ 𝑃 2 ≤ · · · ≤
𝑃𝑘 ), regardless of the realization of the random seeds {𝑠𝑖 } ∀𝑖 ∈ [𝑘 ] . This design ensures that units with higher
production costs are sold at higher prices, which is consistent with the natural pricing scheme where more
expensive units reflect higher production costs.

4.1 Asymptotic Optimality of r-Dynamic

We show that by carefully designing the pricing functions, r-Dynamic achieves an asymptotically optimal
competitive ratio.
 𝛼 ∗ (𝑘 ) 
Theorem 4. Given S = {𝐿, 𝑈 , 𝑓 } for the OSDoS problem with 𝑘 ≥ 1, r-Dynamic is 𝛼 ∗S (𝑘) · exp S𝑘 -

10
competitive when the pricing functions are given by

𝜙𝑖 (𝑠) = 𝐿, ∀𝑠 ∈ [0, 1], 𝑖 ∈ [𝑘 ∗ − 1],

( ¯
𝐿 𝑠 ∈ [0, 𝜉 ∗ ],
𝜙𝑘 ∗ (𝑠) = ∗ ∗
¯ (𝐿 − 𝑐𝑘 ∗ ) · 𝑒 (𝑠 −𝜉 ) ·𝛼 S (𝑘 )/𝑘 + 𝑐𝑘 ∗ 𝑠 ∈ [𝜉 ∗, 1],
¯ ¯
∗
𝜙𝑖 (𝑠) = (𝐿𝑖 − 𝑐𝑖 ) · 𝑒 𝑠 ·𝛼 S (𝑘 )/𝑘 + 𝑐𝑖 ,
∀𝑠 ∈ [0, 1], 𝑖 = 𝑘 ∗ + 1, . . . , 𝑘,
¯
where 𝑘 and 𝜉 are respectively the values of 𝑘 and 𝜉 defined in Theorem 3, corresponding to 𝛼 = 𝛼 ∗S (𝑘), and
∗ ∗
¯ ¯
the price intervals {[𝐿𝑖 , 𝑈𝑖 ]} ∀𝑖 ∈ [𝑘 ] are given as follows:
∗ ) ·𝛼 ∗ (𝑘 )/𝑘
𝑈𝑘 ∗ = 𝐿𝑘 ∗ +1 = (𝐿 − 𝑐𝑘 ∗ ) · 𝑒 (1−𝜉 S + 𝑐𝑘 ∗ , (10)
¯ ¯ ¯ ¯
𝛼 S∗ (𝑘 )/𝑘 ∗
𝑈𝑖 = 𝐿𝑖+1 = (𝐿𝑖 − 𝑐𝑖 ) · 𝑒 + 𝑐𝑖 , ∀𝑖 = 𝑘 + 1, . . . , 𝑘. (11)
¯
We provide a proof sketch of Theorem 4 in Section 4.3. At a high level, the design of the pricing func-
tions {𝜙𝑖 (𝑠)} ∀𝑖 ∈ [𝑘 ] is inspired by the dynamics of an 𝛼 ∗S (𝑘)-competitive algorithm on the arrival instance
I (𝜖 ) studied in the lower bound section. Essentially, the inverse of the pricing function 𝜙𝑖 (𝑠), defined as
𝜙𝑖−1 (𝑣) = sup{𝑠 : 𝜙𝑖 (𝑠) ≤ 𝑣 }, follows the same design as 𝜓𝑖(𝛼 ) (𝑣) in Proposition 2 when 𝛼 = 𝛼 ∗S (𝑘), namely,
(𝛼 S∗ (𝑘 ) )
𝜓𝑖 (𝑣) = sup{𝑠 : 𝜙𝑖 (𝑠) ≤ 𝑣 }.
Asymptotic optimality of r-Dynamic in general settings. Previous studies (e.g., [HK19; Tan+23])
have shown that 𝛼 ∗S (𝑘) remains bounded by a constant as 𝑘 → ∞. Thus, the competitive ratio of r-
Dynamic approaches 𝛼 ∗S (𝑘) as 𝑘 goes to infinity, meaning that r-Dynamic is asymptotically optimal.
Exact optimality of r-Dynamic when 𝑘 = 2. For the small inventory case of 𝑘 = 2, a tighter analysis
shows that r-Dynamic is 𝛼 ∗S (2)-competitive using the same design of pricing functions in Theorem 4,
where 𝛼 ∗S (2) is the lower bound obtained in Theorem 3 for 𝑘 = 2. This indicates that r-Dynamic is not just
asymptotically optimal, but also optimal in the small inventory setting when 𝑘 = 2. The corollary below
formalizes this result.
Corollary 1. Given S = {𝐿, 𝑈 , 𝑓 } for the OSDoS problem with 𝑘 = 2, r-Dynamic is 𝛼 ∗S (2)-competitive when
𝜙 1 : [0, 1] → [𝐿1, 𝑈 1 ] and 𝜙 2 : [0, 1] → [𝐿2, 𝑈 2 ] are designed as follows:
1 −𝑐 2
• If 𝛼 ∗S (2) ≥ 2𝐿−𝑐
𝐿−𝑐 1 , then:
(
𝐿 𝑠 ∈ [0, 𝜉 ∗ ],
𝜙 1 (𝑠) = (𝑠 −𝜉 ∗ ) ·𝛼 S∗ (2)/2
(𝐿 − 𝑐 1 ) · 𝑒 + 𝑐 1 𝑠 ∈ [𝜉 ∗, 1],
∗
𝜙 2 (𝑠) = (𝐿2 − 𝑐 2 ) · 𝑒 𝑠 ·𝛼 S (2)/2 + 𝑐 2 ∀𝑠 ∈ [0, 1].

In this case, the price intervals and 𝜉 ∗ are given by

∗ ) ·𝛼 ∗ (2)/2
𝐿1 = 𝐿, 𝑈 1 = 𝐿2 = (𝐿 − 𝑐 1 ) · 𝑒 (1−𝜉 S + 𝑐 1, 𝑈 2 = 𝑈 ,
1 (2𝐿 − 𝑐 1 − 𝑐 2 )
𝜉∗ = ∗ · .
𝛼 S (2) 𝐿 − 𝑐1
2𝐿−𝑐 1 −𝑐 2
• If 𝛼 ∗S (2) < 𝐿−𝑐 1 , then:

𝜙 1 (𝑠) = 𝐿, ∀𝑠 ∈ [0, 1],

(
𝐿 𝑠 ∈ [0, 𝜉 ∗ ],
𝜙 2 (𝑠) = ∗ ∗
(𝐿 − 𝑐 2 ) · 𝑒 (𝑠 −𝜉 ) ·𝛼 S (2)/2 + 𝑐 2 𝑠 ∈ [𝜉 ∗, 1].

11
 1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

CDF

CDF

CDF
0.4 0.4 0.4
x = * (10) x = * (10)
0.2 r-Dynamic 0.2 r-Dynamic 0.2 r-Dynamic
d-Dynamic d-Dynamic d-Dynamic
0.0
r-Static 0.0
r-Static 0.0
r-Static
4.00 4.25 4.50 4.75 5.00 5.25 5.50 5.75 2.50 2.75 3.00 3.25 3.50 3.75 4.00 4.25 1.0 1.2 1.4 1.6 1.8
Empirical Competitive Ratio Empirical Competitive Ratio Empirical Competitive Ratio
(a) Instance-Sorted (b) Instance-Low2High (c) Instance-IID

Figure 2: CDF plots of empirical competitive ratios of r-Dynamic (Algorithm 1), d-Dynamic [Tan+23] and
r-Static [Sun+24].

In this case, the price intervals and 𝜉 ∗ are given by

𝐿1 = 𝑈 1 = 𝐿2 = 𝐿, 𝑈 2 = 𝑈 ,
(2𝐿 − 𝑐 1 − 𝑐 2 )/𝛼 ∗S (2) − (𝐿 − 𝑐 1 )
𝜉∗ = .
𝐿 − 𝑐2
The proof of the corollary above is given in Appendix G. In the following two subsections, we first
evaluate the empirical performance of r-Dynamic and then provide a proof sketch of Theorem 4 to show
the asymptotic optimality of r-Dynamic.

4.2 Empirical Performance of r-Dynamic

We perform three experiments to evaluate the empirical performance of r-Dynamic and compare its per-
formance to two other algorithms, d-Dynamic [Tan+23] and r-Static [Sun+24]. Throughout the three
experiments, the setup S is fixed to be {𝐿 = 1, 𝑈 = 30, 𝑓 (𝑖) = 𝑖 2 /16} and 𝑘 = 10. To stimulate different
arrival patterns of buyers, we consider the following three types of instances:
• Instance-IID: We generate the valuations of 1000 buyers using the truncated normal distribution 𝑁 (15, 15) [1,30] .
• Instance-Sorted: We generate 1000 buyers using the same approach as Instance-IID, and sort these
buyers in increasing order by their valuations. This instance mimics the hard instance I (𝜖 ) .
• Instance-Low2High: We generate the valuations of 500 buyers using truncated normal distribution
𝑁 (7.5, 7.5) [1,30] . Following these 500 buyers, we generate another 500 buyers using distribution 𝑁 (22.5, 7.5) [1,30] .
Figure 2 presents the CDF plot of the empirical competitive ratios for the three algorithms r-Dynamic,
d-Dynamic, and r-Static, evaluated on 300 instances from each type of instance. In Figure 2(a), r-Dynamic
significantly outperforms the other two algorithms under Instance-Sorted. This is because the valua-
tions of online arrivals are increasing, similar to the hard instance I (𝜖 ) defined in Section 3.2. This result
confirms the superior performance of r-Dynamic under difficult instances compared to the other algo-
rithms. Additionally, Figure 2(a) demonstrates that r-Dynamic’s performance is very close to the lower
bound 𝛼 ∗S (10), suggesting that r-Dynamic may not only be asymptotically optimal in the large 𝑘 regime
but also near-optimal in the small 𝑘 regime. In Figure 2(b), Instance-Low2High consists of two phases:
low-valued buyers arriving first, followed by high-valued buyers. This instance is simpler than Instance-
Sorted, and the performance of all three algorithms improves, with r-Dynamic continuing to outperform
the others. Finally, in Figure 2(c), under Instance-IID, all algorithms achieve a competitive ratio close to
1, with r-Dynamic and d-Dynamic performing similarly. These results indicate that r-Dynamic’s advan-
tage is most evident on more challenging instances, particularly when low-valued buyers arrive before
high-valued ones.

12
4.3 Proof Sketch of Theorem 4
 𝛼 ∗ (𝑘 ) 
For an arbitrary arrival instance I = {𝑣𝑡 } ∀𝑡 ∈ [𝑇 ] , we prove that r-Dynamic is 𝛼 ∗S (𝑘)·exp S𝑘 -competitive
if the pricing functions {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] are designed according to Theorem 4.
Recall that P = {𝑃𝑖 }𝑖 ∈ [𝑘 ] is generated using the pricing functions {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] at the start of r-Dynamic
(line 3 of Algorithm 1). Hereafter, we will refer to Algorithm 1 as r-Dynamic(P) to indicate that the
algorithm is executed with the random price vector P. Based on the design of {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] in Theorem 4,
the first 𝑘 ∗ − 1 prices in P are all 𝐿’s (i.e., 𝑃 1 = · · · = 𝑃𝑘 ∗ −1 = 𝐿), the 𝑘 ∗ -th price 𝑃𝑘 ∗ is a random variable
¯ ¯
within [𝐿, 𝑈𝑘 ∗ ], and for all 𝑖 ∈ {𝑘 ∗ + 1, · · · , 𝑘], the 𝑖-th price 𝑃𝑖 is a random variable within [𝐿𝑖 , 𝑈𝑖 ]. Here,
¯ ¯
¯ ∗ ¯
the values of 𝑘 and {[𝐿𝑖 , 𝑈𝑖 ]} ∀𝑖 are all defined in Theorem 4.
¯
Let P denote the support of all possible values of the random price vector P:
∗
Ö
P = {𝐿}𝑘¯ −1 × [𝐿, 𝑈𝑘 ∗ ] × [𝐿𝑖 , 𝑈𝑖 ].
¯
𝑖 ∈ {𝑘 ∗ +1,··· ,𝑘 }
¯
Given a price vector P ∈ P, let 𝑊 (P) represent the total number of items allocated by r-Dynamic(P) under
the input instance I. Since P is a random variable, 𝑊 (P) is also a random variable. For clarity, we will
sometimes omit the price vector and refer to it simply as 𝑊 whenever the context is clear.
Let 𝜔 be the maximum value in the support of the random variable 𝑊 (i.e., 𝜔 is the maximum possible
value of 𝑊 (P) for all P ∈ P). Thus, 𝜔 is a deterministic value that depends only on the input instance I. In
addition, let 𝝅 ∈ P be a price vector such that r-Dynamic(𝝅) allocates the 𝜔-th item earlier than any other
price vector in the set P. That is, for all P ∈ P, r-Dynamic(P) allocates the 𝜔-th item no earlier than that
of r-Dynamic(𝝅). Let us define the set {𝜏𝑖 } ∀𝑖 ∈ [𝜔 ] so that 𝜏𝑖 is the arrival time of the buyer in the instance
I to whom r-Dynamic(𝝅) allocates the 𝑖-th unit. Note that for all 𝑖 ∈ {1, · · · , 𝜔 }, 𝜏𝑖 is a deterministic value
once 𝝅 and I are given. Let the random variable 𝑊 𝜏𝜔 (P) denote the total number of items allocated by
r-Dynamic(P) after the arrival of buyer 𝜏𝜔 in the instance I. The lemma below shows that the random
variable 𝑊 𝜏𝜔 (P) is always lower bounded by 𝜔 − 1.
Lemma 5. Given instance I, 𝑊 𝜏𝜔 (P) ≥ 𝜔 − 1 holds for all 𝑷 ∈ P.
Lemma 5 greatly simplifies the analysis of r-Dynamic since it implies that the support of the random
variable 𝑊 𝜏𝜔 consists only of two values: 𝜔 − 1 and 𝜔 (note that all 𝑊 ’s are upper bounded by 𝜔). The
intuition behind Lemma 5 is as follows. For all 𝑖 ∈ {1, · · · , 𝜔 }, recall that 𝜏𝑖 denotes the arrival time of the
buyer in the instance I who receives the 𝑖-th unit under r-Dynamic(𝝅). Upon the arrival of buyer 𝜏𝑖 , if the
number of items allocated by r-Dynamic(P) is less than 𝑖 − 1, then the current 𝜏𝑖 -th buyer will definitely
accept the price offered to her, ensuring that one more unit will be sold. As a result, at least 𝜔 − 1 items
will be allocated by the end of time 𝜏𝜔 . Lemma 5 thus follows.
The following two lemmas help us lower bound the expected performance ofr-Dynamic on input in-
stance I and upper bound the objective of the offline optimal algorithm, respectively.
Lemma 6. If a buyer in instance I arrives before time 𝜏𝜔 with a valuation within [𝐿𝜔 , 𝑈 ], then for all P ∈ P,
r-Dynamic(P) will allocate one unit of the item to that buyer.
Lemma 6 can be proved as follows. By definition, 𝜏𝜔 is the earliest time across all possible price vectors
in P that the production level exceeds 𝜔 − 1, causing the posted price to exceed 𝑈𝜔 −1 . Thus, for all possible
realization of P ∈ P, the posted prices by r-Dynamic remain below 𝑈𝜔 −1 before the arrival of buyer at time
𝜏𝜔 . Consequently, when a buyer with a valuation within [𝐿𝜔 , 𝑈 ] arrives before time 𝜏𝜔 , the buyer accepts
the price posted to him (since 𝐿𝜔 ≥ 𝑈𝜔 −1 ) and a unit of item will thus be allocated to this buyer.
Lemma 7. There are no buyers in instance I with a valuation within [𝑈𝜔 , 𝑈 ] arriving after time 𝜏𝜔 , namely,
the valuations of all buyers arrive after 𝜏𝜔 are less than 𝑈𝜔 .

13
 The above lemma can be proved by contradiction. If there exists a buyer arriving after time 𝜏𝜔 with a
valuation within [𝑈𝜔 , 𝑈 ], then there must exist a price vector in P, say P′ , such that r-Dynamic(P′ ) will
allocate more than 𝜔 units, contradicting the definition of 𝜔.
Applying Lemma 6 and observing that r-Dynamic sells at least 𝜔 −1 units, we can derive a lower bound
on the expected performance of r-Dynamic. Conversely, using the lemma 7 and the fact that for all P ∈ P,
the allocation level of r-Dynamic never exceeds 𝜔, we can upper bound the objective of the offline optimal
algorithm. The combination of these two bounds yields the final competitive ratio of r-Dynamic. For the
full proof of Theorem 4, refer to Appendix F.

5 Conclusions and Future Work

In this paper, we studied online 𝑘-selection with production costs that exhibit diseconomies of scale
(OSDoS) and developed novel randomized dynamic pricing mechanisms with the best-known competi-
tive ratios. Specifically, our randomized dynamic pricing scheme provides tight guarantees in both the
small and large inventory settings (i.e., small and large 𝑘), addressing the gap left by [Tan+23]. These
findings advance the theoretical understanding of OSDoS and offer practical insights for designing ran-
domized dynamic pricing mechanisms in online resource allocation problems with increasing marginal
production costs.
This work highlights several promising directions for future research. First, we conjecture that our
proposed randomized pricing mechanism is optimal for all 𝑘 ≥ 1. However, a more refined analysis is
required to establish or refute its optimality for 𝑘 ≥ 3. Additionally, extending our results to multi-resource
or combinatorial settings could reveal new insights into online resource allocation with diseconomies of
scale in more complex environments. Furthermore, it would be valuable to explore other metrics, such
as risk and fairness, in online allocation and selection to ensure that the developed randomized pricing
mechanisms not only maximize efficiency but also promote reliable and equitable outcomes.

Acknowledgments
Hossein Nekouyan Jazi and Xiaoqi Tan acknowledge support from the Alberta Machine Intelligence Insti-
tute (Amii), the Alberta Major Innovation Fund, and the NSERC Discovery Grant RGPIN-2022-03646. Bo
Sun and Raouf Boutaba acknowledge support from the NSERC Grant RGPIN-2019-06587.

References
[AAZ16] Matthew. Andrews, Spyridon. Antonakopoulos, and Lisa. Zhang. “Minimum-Cost Network
Design with (Dis)economies of Scale”. In: SIAM Journal on Computing 45.1 (2016), pp. 49–66.
[AAP93] Baruch Awerbuch, Yossi Azar, and Serge Plotkin. “Throughput-competitive on-line routing”.
In: Proceedings of 1993 IEEE 34th Annual Foundations of Computer Science. IEEE. 1993, pp. 32–
40.
[Aza+16] Yossi Azar et al. “Online algorithms for covering and packing problems with convex objec-
tives”. In: 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS). IEEE.
2016, pp. 148–157.
[Bar+12] Siddharth Barman et al. “Secretary problems with convex costs”. In: International Colloquium
on Automata, Languages, and Programming. Springer. 2012, pp. 75–87.

14
[BMY15] Avrim Blum, Yishay Mansour, and Liu Yang. “Online allocation and pricing with economies of
scale”. In: Web and Internet Economics: 11th International Conference, WINE 2015, Amsterdam,
The Netherlands, December 9-12, 2015, Proceedings 11. Springer. 2015, pp. 159–172.
[Blu+11] Avrim Blum et al. “Welfare and Profit Maximization with Production Costs”. In: Proceedings of
the 52nd Annual Symposium on Foundations of Computer Science (FOCS ’11). 2011, pp. 77–86.
isbn: 978-0-7695-4571-4.
[BN09] Niv Buchbinder and Joseph Naor. “Online primal-dual algorithms for covering and packing”.
In: Mathematics of Operations Research 34.2 (2009), pp. 270–286.
[CST22] Ying Cao, Bo Sun, and Danny HK Tsang. “Online network utility maximization: Algorithm,
competitive analysis, and applications”. In: IEEE Transactions on Control of Network Systems
10.1 (2022), pp. 274–284.
[Cor+19] Jose Correa et al. “Recent developments in prophet inequalities”. In: ACM SIGecom Exchanges
17.1 (2019), pp. 61–70.
[GMM18] Anupam Gupta, Ruta Mehta, and Marco Molinaro. “Maximizing Profit with Convex Costs
in the Random-order Model”. In: 45th International Colloquium on Automata, Languages, and
Programming (ICALP 2018). 2018.
[GS21] Anupam Gupta and Sahil Singla. Random-Order Models. Chapter 11 – Beyond the Worst-Case
Analysis of Algorithms. Cambridge University Press, 2021, pp. 234–258.
[HK19] Zhiyi Huang and Anthony Kim. “Welfare maximization with production costs: A primal dual
approach”. In: Games and Economic Behavior 118.C (2019), pp. 648–667.
[HTW24] Zhiyi Huang, Zhihao Gavin Tang, and David Wajc. “Online Matching: A Brief Survey”. In:
SIGecom Exch. 22.1 (Oct. 2024), pp. 135–158.
[Jia+23] Jianhao Jia et al. “Online resource allocation in Markov Chains”. In: Proceedings of the ACM
Web Conference 2023. 2023, pp. 3498–3507.
[Jia+21] Zhihao Jiang et al. “Online selection problems against constrained adversary”. In: International
Conference on Machine Learning. PMLR. 2021, pp. 5002–5012.
[Luc17] Brendan Lucier. “An economic view of prophet inequalities”. In: ACM SIGecom Exchanges 16.1
(2017), pp. 24–47.
[Meh13] Aranyak Mehta. “Online Matching and Ad Allocation”. In: Found. Trends Theor. Comput. Sci.
8.4 (Oct. 2013), pp. 265–368. issn: 1551-305X.
[Sek17] Shreyas Sekar. “Posted pricing sans discrimination”. In: Proceedings of the 26th International
Joint Conference on Artificial Intelligence. 2017, pp. 388–394.
[STT18] Bo Sun, Xiaoqi Tan, and Danny HK Tsang. “Eliciting multi-dimensional flexibilities from elec-
tric vehicles: A mechanism design approach”. In: IEEE Transactions on Power Systems 34.5
(2018), pp. 4038–4047.
[Sun+24] Bo Sun et al. “Static Pricing for Online Selection Problem and its Variants”. In: The 20th Con-
ference on Web and Internet Economics, WINE 2024, Edinburgh, United Kingdom Proceedings
,December 2-5, 2024. Springer. 2024.
[Tan+20] Xiaoqi Tan et al. “Mechanism Design for Online Resource Allocation: A Unified Approach”.
In: Proc. ACM Meas. Anal. Comput. Syst. (SIGMETRICS ’20) 4.2 (June 2020).
[Tan+23] Xiaoqi Tan et al. “Threshold Policies with Tight Guarantees for Online Selection with Convex
Costs”. In: WINE (2023).

15
[Zha+15] X. Zhang et al. “Online Auctions in IaaS Clouds: Welfare and Profit Maximization with Server
Costs”. In: Proceedings of the 2015 ACM SIGMETRICS International Conference on Measurement
and Modeling of Computer Systems (SIGMETRICS ’15). Portland, Oregon, USA, 2015, pp. 3–15.
isbn: 978-1-4503-3486-0.
[ZLW17] Z. Zhang, Z. Li, and C. Wu. “Optimal Posted Prices for Online Cloud Resource Allocation”. In:
Proc. ACM Meas. Anal. Comput. Syst. (SIGMETRICS ’17) 1.1 (June 2017). issn: 2476-1249.

A Proof of Lemma 1
The monotonicity of each function 𝜓𝑖 follows from the fact that if the random variable Ψ𝑖 (𝑣) is realized to
be equal to one for some 𝑣, then for all 𝑣 ′ > 𝑣, Ψ𝑖 (𝑣 ′ ) must also be equal to one (based on the definition
of Ψ𝑖 ). Since 𝜓𝑖 (𝑣) and 𝜓𝑖 (𝑣 ′ ) represent the expected values of Ψ𝑖 (𝑣) and Ψ𝑖 (𝑣 ′ ), respectively, this property
ensures that each𝜓𝑖 is increasing. Furthermore, by the definition of the state variables 𝑞𝑖 and 𝑞𝑖+1 , whenever
Ψ𝑖+1 (𝑣) = 1, the allocation must have reached at least 𝑖 + 1 units, which implies Ψ𝑖 (𝑣) = 1. Consequently,
it follows that 𝜓𝑖 (𝑣) ≥ 𝜓𝑖+1 (𝑣).

B Proof of Proposition 1
For any randomized algorithm ALG, let 𝐷 (𝐿) denote the number of units that ALG allocates under the
instance I𝐿(𝜖 ) (i.e., the instance I (𝜖 ) by the end of stage-𝐿). Thus, 𝐷 (𝐿) is a random variable taking values
from 0 to 𝑘. Based on definition of 𝐷 (𝐿), ALG(I𝐿(𝜖 ) ) can be computed as follows:

(𝐿)
" 𝐷∑︁
#
 
ALG I𝐿(𝜖 ) = E 𝐷 (𝐿) · 𝐿 − 𝑐𝑖 ,
𝑖=1

where the expectation is taken with respect to the randomness of 𝐷 (𝐿) (the distribution depends on the
randomness of the algorithm ALG). Let the indicator function 1 {𝐷 (𝐿)=𝑗 } = 1 if ALG allocates exactly 𝑗
units at the end of stage-𝐿, and 1 {𝐷 (𝐿)=𝑗 } = 0 otherwise. Based on definition of the random variables
{Ψ𝑖 (𝐿)} ∀𝑖 ∈ [𝑘 ] , we argue that:

1 {𝐷 (𝐿)=𝑗 } = Ψ𝑗 (𝐿) − Ψ𝑗+1 (𝐿), 1 ≤ 𝑗 ≤ 𝑘. (12)

Here, Ψ𝑘+1 (𝐿) = 0 always holds. To see why Eq. (12) is true, consider the case where the random variable
𝐷 (𝐿) = 𝑗, then:

Ψ𝑖 (𝐿) = 1, ∀𝑖 ≤ 𝑗,
Ψ𝑖 (𝐿) = 0, ∀𝑖 > 𝑗 .

From the equation above, we can observe that when the indicator function 1 {𝐷 (𝐿)=𝑗 } = 1, Ψ𝑗+1 (𝐿)−Ψ𝑗 (𝐿) =
1 holds. For the case when 1 {𝐷 (𝐿)=𝑗 } = 0, if 𝐷 (𝐿) < 𝑗, then Ψ𝑗 (𝐿) = Ψ𝑗+1 (𝐿) = 0 and 1 {𝐷 (𝐿)=𝑗 } =
Ψ𝑗 (𝐿) − Ψ𝑗+1 (𝐿) follows. For the case 𝐷 (𝐿) > 𝑗, the two equations Ψ𝑗 (𝐿) = Ψ𝑗+1 (𝐿) = 1 and 1 {𝐷 (𝐿)=𝑗 } =

16
Ψ𝑗 (𝐿) − Ψ𝑗+1 (𝐿) again follow. As a result, ALG(I𝐿(𝜖 ) ) can be computed as follows:
(𝐿)
" 𝐷∑︁
#
 
ALG I𝐿(𝜖 ) = E 𝐷 (𝐿) · 𝐿 − 𝑐𝑖
𝑖=1
𝑘 𝑗
!
∑︁   ∑︁
= E 1 {𝐷 (𝐿)=𝑗 } · 𝑗 · 𝐿 − 𝑐𝑖
𝑗=1 𝑖=1
𝑘 𝑗
!
∑︁ ∑︁
E Ψ𝑗 (𝐿) − Ψ𝑗+1 (𝐿) · 𝑗 · 𝐿 −
 
= 𝑐𝑖
𝑗=1 𝑖=1
𝑘 𝑗
!
∑︁
∑︁
= 𝜓 𝑗 (𝐿) − 𝜓 𝑗+1 (𝐿) · 𝑗 · 𝐿 − 𝑐𝑖
𝑗=1 𝑖=1
𝑘 𝑘
!
∑︁ ∑︁
= 𝜓 𝑗 (𝐿) · (𝐿 − 𝑐 𝑗 ) − 𝜓𝑘+1 (𝐿) · 𝑘 · 𝐿 − 𝑐𝑖
𝑗=1 𝑖=1
𝑘
∑︁
= 𝜓 𝑗 (𝐿) · (𝐿 − 𝑐 𝑗 ).
𝑗=1

Now, let us compute the objective of the 𝛼-competitive algorithm at the end of stage-𝑣, ∀𝑣 ∈ 𝑉 (𝜖 ) , such
that 𝑣 = 𝐿 + 𝑚 · 𝜖. Let the random variable 𝑋𝑖 (𝑣) be the value obtained from allocating the 𝑖-th unit of the
item at the end of some stage-𝑣 ∈ 𝑉 (𝜖 ) . It follows that
"𝑚 #
∑︁  
E[𝑋𝑖 (𝑣) − 𝑐𝑖 ] = 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + E (𝐿 + 𝑗 · 𝜖 − 𝑐𝑖 ) · Ψ𝑖 (𝐿 + 𝑗 · 𝜖) − Ψ𝑖 (𝐿 + ( 𝑗 − 1) · 𝜖)
𝑗=1
𝑚
∑︁
= 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝐿 + 𝑗 · 𝜖 − 𝑐𝑖 ) · E [Ψ𝑖 (𝐿 + 𝑗 · 𝜖) − Ψ𝑖 (𝐿 + ( 𝑗 − 1) · 𝜖)]
𝑗=1
𝑚
∑︁
= 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝐿 + 𝑗 · 𝜖 − 𝑐𝑖 ) · (𝜓𝑖 (𝐿 + 𝑗 · 𝜖) − 𝜓𝑖 (𝐿 + ( 𝑗 − 1) · 𝜖)) .
𝑗=1

where the first equality follows because if the 𝑖-th unit is allocated at some stage 𝐿 + 𝑗 ·𝜖, then the algorithm
must have sold at least 𝑖 units of the item by the end of 𝐿 + 𝑗 · 𝜖, leading to Ψ𝑖 (𝐿 + 𝑗 · 𝜖) = 1. Additionally,
if the 𝑖-th unit is allocated at stage 𝐿 + 𝑗 · 𝜖, then at stage 𝐿 + ( 𝑗 − 1) · 𝜖, the algorithm must have allocated
fewer than 𝑖 units, indicating that Ψ𝑖 (𝐿 + ( 𝑗 − 1) · 𝜖) = 0. Putting together the above results, it follows that:
 
ALG I𝑣(𝜖 )
𝑘
∑︁
= E[𝑋𝑖 (𝑣) − 𝑐𝑖 ]
𝑖=1
𝑘
∑︁ h 𝑚
∑︁  i
= 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝐿 + 𝑗 · 𝜖 − 𝑐𝑖 ) · 𝜓𝑖 (𝐿 + 𝑗 · 𝜖) − 𝜓𝑖 (𝐿 + ( 𝑗 − 1) · 𝜖) ,
𝑖=1 𝑗=1
𝑘 ∑︁
𝑚  
  ∑︁   𝑈 −𝐿
= ALG I𝐿(𝜖 ) + (𝐿 + 𝑗 · 𝜖 − 𝑐𝑖 ) · 𝜓𝑖 (𝐿 + 𝑗 · 𝜖) − 𝜓𝑖 (𝐿 + ( 𝑗 − 1) · 𝜖) , ∀𝑚 ∈ 1, . . . , ⌊ ⌋ .
𝑖=1 𝑗=1
𝜖

Proposition 1 thus follows.

17
C Proof of Lemma 2
Based on Proposition 1, for any online algorithm ALG, we have:
 𝑘
 ∑︁
ALG I𝐿(𝜖 ) = 𝜓𝑖(𝐿) · (𝐿 − 𝑐𝑖 ),
𝑖=1
   ∑︁ 𝑗 
𝑘 ∑︁ 
(𝜖 ) (𝜖 )
ALG I𝐿+𝑗 ·𝜖 = ALG I𝐿 + (𝐿 + 𝑚 · 𝜖) · 𝜓𝑖 (𝐿 + 𝑚 · 𝜖)
𝑖=1 𝑚=1
 𝑈 −𝐿
 
− 𝜓𝑖 (𝐿 + 𝑚 · 𝜖 − 𝜖) , ∀𝑗 = 1, 2, . . . , .
𝜖
As 𝜖 → 0, following the Riemann summation, it follows that:
    𝑘 ∫
∑︁ 𝑣 h i
ALG I𝑣(𝜖 ) = ALG I𝐿(𝜖 ) + (𝜂 − 𝑐𝑖 ) · 𝜓𝑖 (𝜂) − 𝜓𝑖 (𝜂 − 𝑑𝜂) , ∀𝑣 ∈ [𝐿, 𝑈 ].
𝑖=1 𝜂=𝐿

Based on above, the set of functions {𝜓𝑖 }𝑖 ∈ [𝑘 ] should be defined over the range [𝐿, 𝑈 ].
In the next step, we prove that the set of functions {𝜓𝑖 }𝑖 ∈ [𝑘 ] exists such that these set of functions are
continous within their range [𝐿, 𝑈 ]. For now, let us assume this claim holds. Then, it follows that :
   𝑘 ∫
 ∑︁ 𝑣
ALG I𝑣(𝜖 ) = ALG I𝐿(𝜖 ) + (𝜂 − 𝑐𝑖 ) · [𝜓𝑖 (𝜂) − 𝜓𝑖 (𝜂 − 𝑑𝜂)]
𝑖=1 𝜂=𝐿
  𝑘 ∫ 𝑣
∑︁
= ALG I𝐿(𝜖 ) + (𝜂 − 𝑐𝑖 ) · 𝑑𝜓 (𝜂).
𝑖=1 𝜂=𝐿

Following from Eq. (4), if there exists an 𝛼-competitive algorithm, then there should exists a set of functions
{𝜓𝑖 }𝑖 ∈ [𝑘 ] such that:
   𝑘 ∫
 ∑︁ 𝑣
ALG I𝑣(𝜖 ) = ALG I𝐿(𝜖 ) + (𝜂 − 𝑐𝑖 ) · 𝑑𝜓 (𝜂)
𝑖=1 𝜂=𝐿
1  ∑︁𝑘 
≥ · 𝑘𝑣 − 𝑐𝑖 , ∀𝑣 ∈ [𝐿, 𝑈 ].
𝛼 𝑖=1

Now, let us get back to prove that a set of functions {𝜓𝑖 }𝑖 ∈ [𝑘 ] exists corresponding to some online algorithm,
that all these functions are continuous within the range [𝐿, 𝑈 ]. Let ALG be an 𝛼-competitive algorithm.
For some 𝑣 ∈ (𝐿, 𝑈 ) and 𝑖 ∈ [𝑘], let the function 𝜓𝑖 (.) corresponding to ALG be non-continuous at 𝑣.
Let lim𝑥→𝑣 − 𝜓𝑖 (𝑣) = 𝜈 and 𝜓𝑖 (𝑣) = lim𝑥→𝑣 + 𝜓𝑖 (𝑣) = 𝜈 + 𝛿, for some 𝛿 > 0. Then the algorithm must
be selling at least in expectation a 𝛿-fraction of the 𝑖-th unit to the buyers with valuation 𝑣 in instance
I. Conversely, for ALG to be 𝛼-competitive, the expected objective of the algorithm before the arrival
of buyers with valuation 𝑣, ALG(I𝑣(𝜖− ) ), must be at least equal to 𝛼1 · OPT(I𝑣(𝜖− ) ) = 𝛼1 · OPT(I𝑣(𝜖 ) ), where
OPT(I𝑣(𝜖 ) ) denotes the objective value of the offline optimal algorithm on the hard instance I (𝜖 ) up to the
end of stage-𝑣. It can be seen that selling in expectation at least a 𝛿 fraction of the 𝑖-th unit is unnecessary
and ALG could save this fraction of the unit and sell it to buyers with higher valuations. In other words,
we can construct another online algorithm, say ALG, d that follows ALG up to the arrival of buyers with
valuation 𝑣, but sells the 𝛿-fraction of the 𝑖-th unit to buyers with valuation strictly greater than 𝑣 instead.
d will obtain a better objective value with its 𝜓ˆ𝑖 being continuous at 𝑣. Lemma 1
It is easy to see that ALG
follows by repeating the same process for any other discontinuous point of 𝜓𝑖 (𝑣).

18
D Proof of Lemma 3
For any 𝑣 ∈ [𝐿, 𝑈 ], let us define 𝐶 𝑣 as follows:
𝑘 ∫
∑︁ 𝑣
𝐶 𝑣 = 𝐶𝐿 + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖𝛼 (𝜂), ∀𝑣 ∈ (𝐿, 𝑈 ],
𝑖=1 𝜂=𝐿
𝑘
∑︁
𝐶𝐿 = 𝜓𝑖𝛼 (𝐿) · (𝐿 − 𝑐𝑖 ).
𝑖=1

To prove Lemma 3, we need to first prove the feasibility of {𝜓𝑖(𝛼 ) } ∀𝑖 ∈ [𝑘 ] , namely, 𝐶 𝑣 is greater than 𝛼1 · (𝑘 ·
Í
𝑣 − 𝑖 𝑐𝑖 ) for all 𝑣 ∈ [𝐿, 𝑈 ].
For some 𝑣 ∈ [𝐿, 𝑈 ], based on the definition of 𝜒 𝛼 (𝑣) in Eq. (6), there exist a set of functions {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ]
that satisfy Eq. (5) and in the meanwhile, for some arbitrary small value 𝜖, we have:
𝑘
∑︁
𝜒 𝛼 (𝑣) + 𝜖 ≥ 𝜓𝑖 (𝑣). (13)
𝑖=1

Next, using integration by parts, we have

𝑘 ∫
∑︁ 𝑣
𝐶 𝑣 = 𝐶𝐿 + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖𝛼 (𝜂)
𝑖=1 𝜂=𝐿
𝑘 𝑘 ∫ 𝑘
!
∑︁ ∑︁ 𝑣 ∑︁
= 𝐶𝐿 + 𝜓𝑖𝛼 (𝑣) · (𝑣 − 𝑐𝑖 ) − 𝜓𝑖𝛼 (𝐿) · (𝐿 − 𝑐𝑖 ) − 𝜓𝑖𝛼 (𝜂) 𝑑𝜂
𝑖=1 𝑖=1 𝜂=𝐿 𝑖=1
𝑘 ∫ 𝑘
!
∑︁ 𝑣 ∑︁
= 𝜓𝑖𝛼 (𝑣) · (𝑣 − 𝑐𝑖 ) − 𝜓𝑖𝛼 (𝜂) 𝑑𝜂
𝑖=1 𝜂=𝐿 𝑖=1
𝑘
! 𝑘 ∫ 𝑘
!
∑︁ ∑︁ 𝑣 ∑︁
=𝑣 · 𝜓𝑖𝛼 (𝑝) − 𝜓𝑖𝛼 (𝑣) · 𝑐𝑖 − 𝜓𝑖𝛼 (𝜂) 𝑑𝜂
𝑖=1 𝑖=1 𝜂=𝐿 𝑖=1
𝑘
∑︁ ∫ 𝑣
𝛼
= 𝑣 · 𝜒 (𝑣) − 𝜓𝑖𝛼 (𝑣) · 𝑐𝑖 − 𝜒 𝛼 (𝑣)𝑑𝜂,
𝑖=1 𝜂=𝐿

where the last equality follows the definition of {𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 ] in Eq. (7). Thus, we have

𝑘
∑︁ ∫ 𝑣
𝛼
𝐶 𝑣 = 𝑣 · 𝜒 (𝑣) − 𝜓𝑖𝛼 (𝑣) · 𝑐𝑖 − 𝜒 𝛼 (𝑣) · 𝑑𝜂,
𝑖=1 𝜂=𝐿
𝑘
∑︁ 𝑘
∑︁ ∫ 𝑣
≥𝑣· 𝜓𝑖 (𝑣) − 𝑣 · 𝜖 − 𝜓𝑖𝛼 (𝑣) · 𝑐𝑖 − 𝜒 𝛼 (𝑣) · 𝑑𝜂,
𝑖=1 𝑖=1 𝜂=𝐿
𝑘 ∫ 𝑘
! 𝑘
∑︁ 𝑣 ∑︁ ∑︁
≥𝑣· 𝜓𝑖 (𝑣) − 𝜓𝑖 (𝜂) · 𝑑𝜂 − 𝜓𝑖𝛼 (𝑣) · 𝑐𝑖 − 𝑣 · 𝜖, (14)
𝑖=1 𝜂=𝐿 𝑖=1 𝑖=1

where the first inequality follows Eq. (13) and the second inequality directly follows the definition of 𝜒 𝛼 (𝑣)
Í
(recall that 𝜒 𝛼 (𝑣) ≤ 𝑘𝑖=1 𝜓𝑖 (𝑣) holds for all 𝑣 ∈ [𝐿, 𝑈 ]).

19
 Í
By the definition of {𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 ] , we have 𝑖 ∈ [𝑘 ] 𝜓𝑖 (𝑣)
𝛼 = 𝜒 𝛼 (𝑣). Putting together the inequality 𝜒 𝛼 (𝑣) ≤
Í𝑘
𝑖=1 𝜓𝑖 (𝑣) and the fact that productions costs are increasing, we have
𝑘
∑︁ 𝑘
∑︁
𝜓𝑖𝛼 (𝑣) · 𝑐𝑖 ≤ 𝜓𝑖 (𝑣) · 𝑐𝑖 .
𝑖=1 𝑖=1

Putting together the above inequality and the right-hand-side of Eq. (14), it follows that:
𝑘
∑︁−1 𝑘
∑︁−1 ∫ 𝑣 𝑘
∑︁−1
𝐶𝑣 ≥ 𝑝 · 𝜓𝑖 (𝑣) − 𝜓𝑖 (𝜂) · 𝑑𝜂 − 𝜓𝑖 (𝑣) · 𝑐𝑖+1 − 𝑣 · 𝜖
𝑖=0 𝑖=0 𝜂=𝐿 𝑖=0
 
≥ ALG I𝑣(𝜖 ) − 𝑣 · 𝜖,

where ALG is the online algorithm corresponding to the set of allcation functions {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] and recall
that ALG(I𝑣(𝜖 ) ) is defined as follows:

  𝑘
∑︁
ALG I𝐿(𝜖 ) = 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ),
𝑖=1
    𝑘 ∫
∑︁ 𝑣
ALG I𝑣(𝜖 ) = ALG I𝐿(𝜖 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖 (𝜂), ∀𝑣 ∈ [𝐿, 𝑈 ].
𝑖=1 𝜂=𝐿

Since {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] satisfy Eq. (5), it follows that

 
𝐶 𝑣 ≥ ALG I𝑣(𝜖 ) − 𝑣 · 𝜖
𝑘
!
1 ∑︁
≥ · 𝑘 ·𝑣 − 𝑐𝑖 − 𝑣 · 𝜖, ∀𝑣 ∈ [𝐿, 𝑈 ].
𝛼 𝑖=1

By setting 𝜖 → 0, it follows that

𝑘
!
1 ∑︁
𝐶𝑣 ≥ · 𝑘 · 𝑣 − 𝑐𝑖 , ∀𝑣 ∈ [𝐿, 𝑈 ].
𝛼 𝑖=1

To complete the proof of Lemma 3, we also need to prove that the above inequality holds as an equality
for the set of functions {𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 ] . This can be proved by contradiction. Suppose that at some point
𝑣 ∈ [𝐿, 𝑈 ], the above equality does not hold, then there must exist another set of feasible functions, say
{𝜓ˆ𝑖 } ∀𝑖 ∈ [𝑘 ] , induced by a new algorithm, say ALG,
d that satisfy Eq. (5) and

𝑘
∑︁ 𝑘
∑︁
𝜓ˆ𝑖 (𝑣) < 𝜓𝑖𝛼 (𝑣).
𝑖=1 𝑖=1

We argue that the new set of functions {𝜓ˆ𝑖 } ∀𝑖 ∈ [𝑘 ] will allocate a smaller fraction of its total units to buy-
ers in I (𝜖 ) arriving at or before stage-𝑣 compared to {𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 ] . However, by still following the alloca-
d 𝑣(𝜖 ) ) will be exactly equal to 1 (𝑘 · 𝑣 − Í𝑘𝑖=1 𝑐𝑖 ). Given the definition of
tion functions {𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 ] , ALG(I
Í Í 𝛼
{𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 ] , we have 𝑘𝑖=1 𝜓𝑖𝛼 (𝑣) = 𝜒 𝛼 (𝑣), meaning that 𝑘𝑖=1 𝜓ˆ𝑖 (𝑣) < 𝜒 𝛼 (𝑣). However, this contradicts the
definition of 𝜒 𝛼 (𝑣). We thus complete the proof of Lemma 3.

20
E Proof of Proposition 2
From Lemma 3, we know that {𝜓𝑖𝛼 (𝑣)} ∀𝑖 ∈ [𝑘 ] satisfy Eq. (5) with an equality. Therefore, the set of allocation
functions {𝜓𝑖𝛼 (𝑣)} ∀𝑖 ∈ [𝑘 ] is a solution to the following system of equations:

𝑘
∑︁ 𝑘 ∫
∑︁ 𝑣
𝜓𝑖𝛼 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖𝛼 (𝜂)
𝑖=1 𝑖=1 𝜂=𝐿
1 ∑︁
= · (𝑘 · 𝑣 − 𝑐𝑖 ), ∀𝑖 ∈ [𝑘], 𝑣 ∈ [𝐿, 𝑈 ]. (15)
𝛼 𝑖

Also, based on Lemma 4, we argue that if the value of the function 𝜓𝑖𝛼 (𝑣) is changing at some value
𝑣 ∈ [𝐿, 𝑈 ] (i.e., 𝑑𝜓𝑖𝛼 (𝑣) ≠ 0), then the value of all the functions {𝜓 𝑗∗ (𝑣)} ∀ 𝑗 ∈ [𝑖 −1] are equal to one, and all the
functions in the set {𝜓 𝑗∗ (𝑣)} 𝑗 >𝑖 are equal to zero. Based on this property, we can assign an interval [ℓ𝑖 , 𝑢𝑖 ]
to each 𝜓𝑖𝛼 (𝑣). In the interval of [ℓ𝑖 , 𝑢𝑖 ], only the value of 𝜓𝑖𝛼 changes while the other functions {𝜓 𝑗𝛼 } ∀ 𝑗≠𝑖
in that interval are fixed to be one or zero. Additionally, the following relation exists between the start
and end points of these intervals:

𝐿 = ℓ1 ≤ 𝑢 1 = ℓ2 ≤ 𝑢 2 ≤ · · · ≤ ℓ𝑘 ≤ 𝑢𝑘 = 𝑈 .
Í Í
To satisfy the equality 𝑖 ∈ [𝑘 ] 𝜓𝑖𝛼 (𝐿) · (𝐿 −𝑐𝑖 ) = 𝛼1 · (𝑘 ·𝐿 − 𝑖 𝑐𝑖 ), the set of functions {𝜓𝑖𝛼 (𝑣)} ∀𝑖 ∈ [𝑘 −1] should
¯
be equal to one at the point 𝑣 = 𝐿. Thus, the explicit design of the functions {𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 −1] is as follows:
¯

𝜓𝑖(𝛼 ) (𝑣) = 1, 𝑖 = 1, . . . , 𝑘 − 1.
¯
Í 1 Í Í 1 Í
In the case that 𝑖 ∈ [𝑘 ] 𝐿 − 𝑐𝑖 < 𝛼 · (𝑘 · 𝐿 − 𝑖 𝑐𝑖 ), to satisfy 𝑖 ∈ [𝑘 ] 𝜓𝑖𝛼 (𝐿) · (𝐿 − 𝑐𝑖 ) = 𝛼 · (𝑘 · 𝐿 − 𝑖 𝑐 𝑖 ),
¯
we need to have:
Í 1 Í
𝑖 ∈ [𝑘 −1] (𝐿 − 𝑐𝑖 ) − 𝛼 · 𝑖 ∈ [𝑘 ] (𝐿 − 𝑐𝑖 )
𝜓𝑘𝛼 (𝐿) = ¯ = 𝜉.
¯ 𝐿 − 𝑐𝑘
¯

Since for all 𝑣 ∈ [ℓ𝑘 , 𝑢𝑘 ] with ℓ𝑘 = 𝐿, only the value of 𝜓𝑘𝛼 (𝑣) changes (i.e., 𝑑𝜓𝑖𝛼 (𝑣) = 0 for all 𝑖 ≠ 𝑘), it
¯ ¯ ¯ ¯ ¯
follows that:
𝑘
∑︁ 𝑘 ∫
∑︁ 𝑣
𝜓𝑖𝛼 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖𝛼 (𝜂)
𝑖=1 𝑖=1 𝜂=𝐿
𝑘
∑︁ ∫ 𝑣
= 𝜓𝑖𝛼 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑘 )𝑑𝜓𝑘∗ (𝜂), ∀𝑣 ∈ [𝐿, 𝑢𝑘 ].
𝜂=𝐿 ¯ ¯ ¯
𝑖=1

Based on the system of equations in Eq. (15), we need to have:

𝑘 ∫ 𝑣
∑︁ 1 ∑︁
𝜓𝑖𝛼 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑘 )𝑑𝜓𝑘∗ (𝜂) = · (𝑘 · 𝑣 − 𝑐𝑖 ), ∀𝑣 ∈ [ℓ𝑘 , 𝑢𝑘 ].
𝑖=1 𝜂=𝐿 ¯ ¯ 𝛼 𝑖
¯ ¯

Taking derivative w.r.t. 𝑣 from both sides of the equation above, we have

𝑘
(𝑣 − 𝑐𝑘 ) · 𝑑𝜓𝑘∗ (𝑣) = .
¯ ¯ 𝛼

21
Solving the above differential equation leads to
𝑘
𝜓𝑘∗ (𝑣) =


· ln 𝑣 − 𝑐𝑘 + 𝑄, ∀𝑣 ∈ [ℓ𝑘 , 𝑢𝑘 ],
¯ 𝛼 ¯ ¯ ¯

where 𝑄 is a constant. To find 𝑄, since 𝜓𝑘∗ (𝐿) = 𝜉, it follows that 𝑄 = 𝜉 − 𝑘

𝛼 · ln 𝐿 − 𝑐𝑘 . As a result, the
¯ ¯
explicit design of the function 𝜓𝑘𝛼 is as follows:
¯
(  𝑣−𝑐 
𝑘 𝑘
𝜉+ · ln ¯ 𝑣 ∈ [𝐿, 𝑢𝑘 ],
𝜓𝑘(𝛼 ) (𝑣) = 𝛼 𝐿−𝑐𝑘
¯ ¯
¯ 1 𝑣 > 𝑢𝑘 .
¯

To obtain the value of 𝑢𝑘 , we set 𝜓𝑘∗ (𝑢𝑘 ) = 1 (the function 𝜓𝑘∗ reaches its maximum). Consequently, it
¯ ¯ ¯ ¯
follows that:
𝛼
𝑢𝑘 = (𝐿 − 𝑐𝑘 ) · 𝑒 𝑘 · (1−𝜉 ) + 𝑐𝑘 .
¯ ¯ ¯

Using the same procedure as what has been applied to 𝜓𝑘𝛼 , for all the other functions {𝜓𝑖𝛼 (𝑣)} ∀𝑖>𝑘 , we have
¯ ¯

𝑘
∑︁ 𝑘 ∫
∑︁ 𝑣
𝜓 𝑗∗ (𝐿) · (𝐿 − 𝑐 𝑗 ) + (𝜂 − 𝑐 𝑗 )𝑑𝜓𝑖𝛼 (𝜂)
𝑗=1 𝑗=1 𝜂=𝐿

𝑘
∑︁ ∫ 𝑣
= 𝜓 𝑗∗ (𝐿) · (𝐿 − 𝑐 𝑗 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖𝛼 (𝜂), ∀𝑣 ∈ [ℓ𝑖 , 𝑢𝑖 ].
𝑗=1 𝜂=𝐿

Taking derivative w.r.t. 𝑣 from both sides of the equation above, it follows that:
𝑘
(𝑣 − 𝑐𝑖 ) · 𝑑𝜓𝑖𝛼 (𝑣) = .
𝛼
Solving the above differential equation leads to
𝑘
𝜓𝑖𝛼 (𝑣) = · ln(𝑣 − 𝑐𝑖 ) + 𝑄,
ˆ ∀𝑣 ∈ [ℓ𝑖 , 𝑢𝑖 ].
𝛼
Since 𝜓 ∗ (ℓ𝑖 ) = 0, we have 𝑄ˆ = − ln(ℓ𝑖 − 𝑐𝑖 ). The explicit design of the function 𝜓𝑖𝛼 is thus as follows:



 0 𝑣 ≤ ℓ𝑖 ,
 
𝜓𝑖(𝛼 ) (𝑣)
𝑘
 𝑣−𝑐𝑖
= 𝛼 · ln ℓ𝑖 −𝑐𝑖 𝑣 ∈ [ℓ𝑖 , 𝑢𝑖 ], 𝑖 = 𝑘 + 1, . . . , 𝑘 − 1

 ¯

1
 𝑣 ≥ 𝑢𝑖 .

For the function 𝜓𝑘𝛼 , since it is the last function, it follows that:
(
0 𝑣 ≤ ℓ𝑘 ,
𝜓𝑘(𝛼 ) (𝑣) = 𝑘

𝑣−𝑐𝑘

𝛼 · ln ℓ𝑘 −𝑐𝑘 𝑣 ∈ [ℓ𝑘 , 𝑈 ].

By setting 𝜓 ∗ (𝑢𝑖 ) = 1, it follows that:

𝛼
𝑢𝑖 = (ℓ𝑖 − 𝑐𝑖 ) · 𝑒 𝑘 + 𝑐𝑖 , 𝑘 + 1 ≤ 𝑖 ≤ 𝑘.
¯
Putting everything together, Proposition 2 follows.

22
F Full Proof of Theorem 4
In this section, we provide a complete proof of Theorem 5. We begin by introducing several important
notations and lemmas. Then, we break the problem into two independent subproblems based on the buy-
ers’ valuations in some arbitrary arrival instance I. For each case, we proceed to show how to upper
bound OPT(I), the objective of the optimal offline algorithm on I. We then proceed to lower bound the
expected performance of r-Dynamic on that instance, ALG(I). Ultimately, we combine everything and
obtain a performance guarantee for r-Dynamic under all adversarially chosen instances of OSDoS for that
subproblem.

F.1 Notations and Definitions

Consider an arbitrary arrival instance I = {𝑣𝑡 }𝑡 ∈ [𝑇 ] . Recall that the random price vector P = {𝑃 1, · · · , 𝑃𝑘 }
is generated using the pricing functions {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] at the beginning of r-Dynamic (line 3 of Algorithm 1).
In the following, we will refer to Algorithm 1 as r-Dynamic(P) to indicate that the algorithm is executed
with the random price vector being realized as P. Based on the design of {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] in Theorem 4, the first
𝑘 ∗ − 1 prices in P are all 𝐿’s (i.e., 𝑃 1 = · · · = 𝑃𝑘 ∗ −1 = 𝐿), the 𝑘 ∗ -th price 𝑃𝑘 ∗ is a random variable within
[𝐿, 𝑈𝑘 ∗ ], and for all 𝑖 ∈ {𝑘 ∗ + 1, · · · , 𝑘], we have 𝑃𝑖 ∈ [𝐿𝑖 , 𝑈𝑖 ] (recall that 𝑃𝑖 is also a random variable). Here,
the values of 𝑘 ∗ and {[𝐿𝑖 , 𝑈𝑖 ]} ∀𝑖 are all defined in Theorem 4.
¯
Let P denote the support of all possible values of the random price vector P:
∗
Ö
P = {𝐿}𝑘¯ −1 × [𝐿, 𝑈𝑘 ∗ ] × [𝐿𝑖 , 𝑈𝑖 ].
¯
𝑖 ∈ {𝑘 ∗ +1,··· ,𝑘 }
¯

Given a price vector realization P ∈ P, let𝑊 (P) represent the total number of items allocated by r-Dynamic(P)
under the input instance I. Since P is a random variable, 𝑊 (P) is also a random variable. For clarity, we
will sometimes omit the price vector and refer to it simply as 𝑊 whenever the context is clear.
Let 𝜔 denote the maximum value in the support of the random variable 𝑊 (i.e., 𝜔 is the maximum
possible value of 𝑊 (P) for all P ∈ P). Thus, 𝜔 is a deterministic value that depends only on the input
instance I. Furthermore, let 𝝅 ∈ P be a price vector such that r-Dynamic(𝝅) allocates the 𝜔-th item
earlier than any other price vector in the set P. That is, for all P ∈ P, r-Dynamic(P) allocates the 𝜔-th
item no earlier than that of r-Dynamic(𝝅).
Let us define the set {(𝜈𝑖 , 𝜏𝑖 )} ∀𝑖 ∈ [𝜔 ] so that 𝜏𝑖 is the arrival time of the buyer in the instance I to whom
r-Dynamic(𝝅) allocates the 𝑖-th unit and 𝜈𝑖 is its valuation. Note that for all 𝑖 ∈ {1, · · · , 𝜔 }, 𝜏𝑖 and 𝜈𝑖 are
deterministic values once 𝝅 and I are given.
We can derive the following inequality regarding 𝜈𝑖 :

𝜈𝑖 ≥ 𝐿𝑖 , ∀𝑖 ∈ [𝜔], (16)

where 𝐿𝑖 is the lower bound for the range of the pricing function 𝜙𝑖 , used to generate the random price
for the 𝑖-th unit. This inequality holds since the buyer arriving at time 𝜏𝑖 accepts the price posted for the
𝑖-th unit by r-Dynamic. The price for the 𝑖-th unit is at least equal to 𝐿𝑖 based on the design of the pricing
functions 𝜙𝑖 .
Let the random variable 𝑊 𝜏𝜔 (P) denote the total number of items allocated by r-Dynamic(P) after the
arrival of buyer 𝜏𝜔 in the instance I. The lemma below shows that the random variable 𝑊 𝜏𝜔 (P) is always
lower bounded by 𝜔 − 1.

Lemma 8. Given an arbitrary instance I, 𝑊 𝜏𝜔 (P) ≥ 𝜔 − 1 holds for all 𝑷 ∈ P.

23
Proof. If 𝜔 = 1, this lemma is trivial, so we consider the case where 𝜔 ≥ 2. Suppose before the arrival of
the buyer at time 𝜏2 , no items have been sold. From Eq. (16), we know that 𝜈 2 ≥ 𝐿2 . Additionally, based
on the design of the pricing functions 𝜙 1 (.) and 𝜙 2 (.), we have 𝐿2 ≥ 𝑈 1 . Consequently, it follows that
𝜈 2 ≥ 𝑈 1 . Since the realized price for the first unit under any sampled price vector will be at most 𝑈 1 (based
on design of the pricing function 𝜙 1 ), the buyer arriving at time 𝜏2 will accept the price for the first unit,
and the algorithm will sell the first item. Thus, for all possible price vector P, the value of the random
variable 𝑊 𝜏2 (P) is at least equal to one. By the same reasoning, if before the arrival of the buyer at time 𝜏3 ,
only one item has been sold, the buyer arriving at 𝜏3 will accept the price for the second unit, regardless
of its price, and the total number of items sold by r-Dynamic will increase to two. This reasoning can be
extended to the time 𝜏𝜔 . As a result, after the arrival of the buyer at time 𝜏𝜔 , r-Dynamic sells at least 𝜔 − 1
units and thereby the claim in the lemma follows. ■

Lemma 8 implies that the support of the random variable 𝑊 𝜏𝜔 consists only of two values: 𝜔 − 1 and
𝜔. This greatly simplifies the analysis of the algorithm.
The following two lemmas help us lower bound the expected performance of r-Dynamic under the
input instance I and upper bound the objective of the offline optimal algorithm given the instance I,
respectively.

Lemma 9. If a buyer in instance I arrives before time 𝜏𝜔 with a valuation within [𝐿𝜔 , 𝑈 ], then for all P ∈ P,
r-Dynamic(P) will allocate one unit of the item to that buyer.

Proof. According to the definition of 𝝅, 𝜏𝜔 is the earliest time across all possible price vectors in P that the
production level exceeds 𝜔 −1, causing the posted price to exceed 𝑈𝜔 −1 . Thus, for all possible realization of
P, the posted prices by r-Dynamic remain below 𝑈𝜔 −1 before the arrival of buyer at time 𝜏𝜔 . Consequently,
when a buyer with a valuation within [𝐿𝜔 , 𝑈 ] arrives before time 𝜏𝜔 , the buyer accepts the price posted to
him (since 𝐿𝜔 ≥ 𝑈𝜔 −1 ) and a unit of item will thus be allocated to this buyer. ■

Lemma 10. There are no buyers in instance I with a valuation within [𝑈𝜔 , 𝑈 ] arriving after time 𝜏𝜔 , namely,
the valuations of all buyers arrive after 𝜏𝜔 are less than 𝑈𝜔 .

Proof. If there exist a buyer with a valuation larger than 𝑈𝜔 arriving after the time 𝜏𝜔 , then there must
exist a price vector in P, say P′ , such that the number of units sold by r-Dynamic(P′ ) will exceed 𝜔. This
contradicts the definition of 𝜔. Thus, the lemma follows.1 ■

Given an instance I, let the set B ⊆ I contain the highest-valued buyers that the offline optimal
algorithm selects. We further divide B into two subsets: B1 and B2 . B1 comprises the highest-valued
buyers up to time 𝜏𝜔 , while B2 includes the remaining buyers in B who arrive at or after time 𝜏𝜔 . Let us
further partition B1 into two subsets: B1,1 and B1,2 . Here, B1,1 consists of buyers in B1 with valuations at
least 𝐿𝜔 , and B1,2 = B1 \ B1,1 comprises those with valuations strictly less than 𝐿𝜔 .
For the rest of the analysis, let us study the problem for two separate cases that may occur depending
on the instance I.

F.2 Case 1: Buyer 𝜏𝜔 Has the Highest Valuation

In this case, in the set B2 , no buyer has a valuation greater than 𝑈𝜔 −1 except for the buyer at time 𝜏𝜔 .
Therefore, the buyer at time 𝜏𝜔 possesses the highest valuation in the instance I.
1 In fact, such a price vector P′ for the initial 𝜔 units should have the same prices as the vector 𝝅 and for the (𝑖 + 1)-th unit,
P′ should be equal to 𝑈 𝜔 (i.e., 𝑃𝑖+1
′ = 𝑈 𝜔 ).

24
F.2.1 Bound OPT from Above for Case 1

The following upper bound can be derived for OPT(I), which denotes the objective value of the offline
optimal algorithm on instance I:
OPT(I)
|B|
∑︁
= 𝑉 (B1 ) + 𝑉 (B2 ) − 𝑐𝑖
𝑖=1
|B|
∑︁
≤ 𝑉 (B1 ) + (|B2 | − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
|B|
∑︁
= 𝑉 (B1,1 ) + 𝑉 (B1,2 ) + (|B2 | − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
≤ |B1,1 | · 𝑈𝜔 −1 + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 )
|B|
∑︁
+ |B1,2 | · 𝑈𝜔 −1 + (|B2 | − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
𝑘
∑︁
≤ (𝑘 − 1) · 𝑈𝜔 −1 + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 ) + 𝜈𝜏𝜔 − 𝑐𝑖 ,
𝑖=1

where the first inequality directly follows the condition of Case 1. The second inequality follows the
definition of B1,1 and B1,2 . Finally, the third inequality follows the fact that we only focus on the case
when 𝑐𝑘 < 𝐿.

F.2.2 Bound ALG from Below for Case 1

Moving forward, we focus on establishing a lower bound on the performance of r-Dynamic under the
arrival instance I. Let the random variables {𝑋𝑖 } ∀𝑖 ∈ [𝑘 ] represent the value obtained by r-Dynamic from
allocating the 𝑖-th unit of the item. Given the input instance I, let E[ALG(I)] denote the expected per-
formance of r-Dynamic. Therefore, we have:
E[ALG(I)]
" 𝑘 #
∑︁
=E (𝑋𝑖 − 𝑐𝑖 ) · 1{i-th item is sold under price vector P} ,
𝑖=1
𝜔
∑︁−1
≥ E[𝑋𝑖 − 𝑐𝑖 ]
𝑖=1
𝜔
∑︁−1 𝜔
∑︁−1
= E[𝑋𝑖 ] − 𝑐𝑖
𝑖=1 𝑖=1
𝜔
∑︁−1 ∫ 1 𝜔
∑︁−1
≥ 𝜙𝑖 (𝜂)𝑑𝜂 + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 ) − 𝑐𝑖 .
𝑖=1 0 𝑖=1

In the equations above, all expectations are taken with respect to the randomness of the price vector P.
The first inequality follows Lemma 8, indicating that under any price vector P, r-Dynamic sells at least
𝜔 − 1 units. The first term in the second inequality follows due to the independent sampling used to set
the price of the 𝑖-th unit using the pricing function 𝜙𝑖 , and the second term follows Lemma 9.

25
 Let us define 𝜓𝑖 (𝑣) = sup{𝑠 : 𝜙𝑖 (𝑠) ≤ 𝑣 } for all 𝑖 ∈ [𝑘]. From the definition of {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] in Theorem 4,
it follows that:

E[ALG(I)]
𝜔
∑︁−1 ∫ 1 𝜔
∑︁−1


≥ 𝜙𝑖 (𝜂)𝑑𝜂 − 𝑐𝑖 + 𝑉 (B1,1 ) − |B1 | · 𝐿𝜔
𝑖=1 0 𝑖=1
𝜔
∑︁ 𝜔
∑︁−1 ∫ 𝑈𝜔 −1


= 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖+1 ) + (𝜂 − 𝑐𝑖+1 )𝑑𝜓𝑖 (𝜂) + 𝑉 (B1,1 ) − |B1 | · 𝐿𝜔 .
𝑖=1 𝑖=1 𝜂=𝐿

Furthermore, it is evident that based on the design of {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] with 𝛼 = 𝛼 ∗S (𝑘), the set of functions
{𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] follows the same design as {𝜓𝑖𝛼 (𝑣)} ∀𝑖 ∈ [𝑘 ] given in Proposition 2. As a result, it follows that:

𝑘
∑︁ 𝜔
∑︁−1 ∫ 𝑈𝜔 −1
𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖+1 ) + (𝜂 − 𝑐𝑖+1 )𝑑𝜓𝑖 (𝜂) + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 )
𝑖=1 𝑖=1 𝜂=𝐿
!
1 ∑︁

≥ ∗ · 𝑘 · 𝑈𝜔 −1 − 𝑐𝑖 + 𝑉 (B1,1 ) − |B1 | · 𝐿𝜔 .
𝛼 S (𝑘) 𝑖

F.2.3 Putting Everything Together for Case 1

Putting together the lower bound and upper bound derived for the expected objective value of r-Dynamic
and the offline optimal algorithm, it follows that:

OPT(I)
E[ALG(I)]
Í
(𝑘 − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 + (𝑉 (B1,1 ) − |B1,1 | · 𝐿𝜔 ) − 𝑘𝑖=1 𝑐𝑖
≤ 1 Í
𝛼 S∗ (𝑘 ) · (𝑘 · 𝑈𝜔 −1 − 𝑖 𝑐 𝑖 ) + (𝑉 (B1,1 ) − |B1,1 | · 𝐿𝜔 )
Í
(𝑘 − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 − 𝑘𝑖=1 𝑐𝑖
≤ 1 Í
𝛼 S∗ (𝑘 ) · (𝑘 · 𝑈𝜔 −1 − 𝑖 𝑐 𝑖 )
 
∗ 𝜈𝜏𝜔 − 𝑈𝜔 −1
= 𝛼 S (𝑘) · 1 +
𝑘 · 𝑈𝜔 −1 − 𝐶
 
∗ 𝑈𝜔 − 𝑈𝜔 −1
≤ 𝛼 S (𝑘) · 1 +
𝑘 · 𝑈𝜔 −1 − 𝐶
𝛼 ∗ (𝑘 )
S
≤ 𝛼 ∗S (𝑘) · 𝑒 𝑘 .
𝛼 ∗ (𝑘 )
S
In the equation above, the last inequality is due to the fact that 𝑈𝑈𝜔𝜔 −𝑈 𝜔 −1
−1 −𝑐 𝜔
= 𝑈𝜔 −𝑐 𝜔
𝑈𝜔 −1 −𝑐 𝜔 −1 ≤ 1+𝑒 𝑘 , where
the last inequality follows the design in Eq. (10).

F.3 Case 2: Buyer 𝜏𝜔 Does Not Have the Highest Valuation

In the set of buyers B2 , there are other buyers with valuation greater than 𝑈𝜔 −1 besides the buyer at time
𝜏𝜔 . Let 𝜆 denote the value of the highest buyer in B2 along with the value of buyer at time 𝜏𝜔 . First, let
us consider the case that 𝜆 ≤ 𝜈𝜏𝜔 . The proof for the case that 𝜆 > 𝜈𝜏𝜔 follows exactly the same as the
following case.

26
F.3.1 Bound OPT from Above for Case 2

Following the same approach as the previous Case 1, let us first upper bound the objective of the offline
optimal algorithm on instance I:

OPT(I)
|B|
∑︁
= 𝑉 (B1 ) + 𝑉 (B2 ) − 𝑐𝑖
𝑖=1
|B|
∑︁
≤ 𝑉 (B1 ) + (|B2 | − 1) · 𝜆 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
|B|
∑︁
≤ 𝑉 (B1,1 ) + 𝑉 (B1,2 ) + (|B2 | − 1) · 𝜆 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
≤ |B1,1 | · 𝑈𝜔 −1 + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 )+
|B|
∑︁
|B1,2 | · 𝑈𝜔 −1 + (|B2 | − 1) · 𝜆 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
𝑘
∑︁
≤ (𝑘 − 1) · 𝜆 + 𝜈𝜏𝜔 + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 ) − 𝑐𝑖 .
𝑖=1

F.3.2 Bound ALG from Below for Case 2

To establish a lower bound on the performance of r-Dynamic in this case, let us consider the following
lemma:

Lemma 11. If the random price of the 𝜔-th unit is realized to be less than 𝜆 and further assume that 𝜆 ≤ 𝜈𝜏𝜔 ,
then the number of items allocated by r-Dynamic in the end is equal to 𝜔.

Proof. Under any price realization, as established by Lemma 8, it is proven that after the arrival of the
buyer at time 𝜏𝜔 , the number of allocated units is at least 𝜔 − 1. If the price of the 𝜔-th unit is realized to
be less than 𝜆, then upon the arrival of the buyer with valuation 𝜆 at some time after 𝜏𝜔 , the buyer will
accept the price if the 𝜔-th unit has not already been sold. ■

Next, we obtain a lower bound on the performance of r-Dynamic as follows:

E[ALG(I)]
" 𝑘 #
∑︁
=E (𝑋𝑖 − 𝑐𝑖 ) · 1{i-th item is sold under pricie vector P}
𝑖=1
𝜔
∑︁−1
≥ E[𝑋𝑖 − 𝑐𝑖 ] + E[𝑋𝜔 − 𝑐 𝜔 |𝑃𝜔 ≤ 𝜆]
𝑖=1
𝜔
∑︁−1 ∫ 1 ∫ −1 (𝜆)
𝜙𝜔 𝜔
∑︁−1
≥ 𝜙𝑖 (𝜂)𝑑𝜂 + 𝜙𝜔 (𝜂)𝑑𝜂 − 𝜙𝜔−1 (𝜆) · 𝑐𝜔 − 𝑐𝑖 + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 ).
𝑖=1 0 0 𝑖=1

In the equations above, all expectations are taken with respect to the randomness of the price vector P ∈ P.
The first inequality follows Lemma 11, where 𝑃𝜔 denotes the 𝜔-element of the random price vector P that

27
r-Dynamic posts for the 𝜔-th unit. The second inequality is true because of the independent sampling that
is used to set the random price of the 𝑖-th unit using 𝜙𝑖 and Lemma 9.
Let us define 𝜓𝑖 (𝑣) = sup{𝑠 : 𝜙𝑖 (𝑠) ≤ 𝑣 }, 𝑖 ∈ [𝑘]. From the definition of {𝜙𝑖 }𝑖 ∈ [𝑘 ] in Theorem 4, it
follows that:
𝜔
∑︁−1 ∫ 1 ∫ −1 (𝜆)
𝜙𝜔 𝜔
∑︁−1
𝜙𝑖 (𝜂)𝑑𝜂 + 𝜙𝜔 (𝜂)𝑑𝜂 − 𝜙𝜔−1 (𝜆) · 𝑐𝜔 − 𝑐𝑖 + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 )
𝑖=1 0 0 𝑖=1
𝑘
∑︁ 𝜔 ∫
∑︁ 𝜆
= 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖+1 ) + (𝜂 − 𝑐𝑖+1 )𝑑𝜓𝑖 (𝜂) + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 ).
𝑖=1 𝑖=1 𝜂=𝐿

Furthermore, it is evident that the set of functions {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] follows the same design as {𝜓𝑖𝛼 (𝑣)}𝑖 ∈ [𝑘 ]
given in Lemma 2 (recall that {𝜓𝑖𝛼 (𝑣)}𝑖 ∈ [𝑘 ] are based on {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] ). As a result, it follows that:

𝑘
∑︁ 𝜔 ∫
∑︁ 𝜆
𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖+1 ) + (𝜂 − 𝑐𝑖+1 )𝑑𝜓𝑖 (𝜂) + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 )
𝑖=1 𝑖=1 𝜂=𝐿
!
1 ∑︁
≥ ∗ · 𝑘 ·𝜆− 𝑐𝑖 ) + (𝑉 (B1 − |B1 | · 𝐿𝜔 ).
𝛼 S (𝑘) 𝑖

F.3.3 Putting Everything Together for Case 2

Putting together the above lower and upper bounds, it follows that:

OPT(I)
E[ALG(I)]
Í
(𝑘 − 1) · 𝜆 + 𝜈𝜏𝜔 + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 ) − 𝑘𝑖=1 𝑐𝑖
≤ 1 Í𝑘
𝛼 S∗ (𝑘 ) · (𝑘 · 𝜆 − 𝑖=1 𝑐 𝑖 ) + (𝑉 (B1 ) − |B1 | · 𝐿𝜔 )
Í
(𝑘 − 1) · 𝜆 + 𝜈𝜏𝜔 − 𝑘𝑖=1 𝑐𝑖
≤ 1 Í𝑘
𝛼 ∗ (𝑘 ) · (𝑘 · 𝜆 − 𝑖=1 𝑐 𝑖 )
S
𝜈𝜏𝜔 − 𝜆
= 𝛼 ∗S (𝑘) · (1 + )
𝑘 · 𝜆 −𝐶
𝑈𝜔 − 𝑈𝜔 −1
≤ 𝛼 ∗S (𝑘) · (1 + )
𝑘 · 𝑈𝜔 −1 − 𝐶
𝛼 ∗ (𝑘 )
S
≤ 𝛼 ∗S (𝑘) · 𝑒 𝑘 .

We thus complete the proof of Theorem 4.

Remark 1. Theorem 4 argues that r-Dynamic is asymptotically optimal. We emphasize that our analysis
of Theorem 4 is not tight because it does not differentiate between the sample paths of r-Dynamic when the
algorithm sells 𝜔 − 1 units and those when it sells 𝜔 units. As a result, our analysis considers that r-Dynamic
sells 𝜔 − 1 units of the item on all sample paths.2 However, in the subsequent analysis for the case of 𝑘 = 2, we
can enumerate all the scenarios and therefore do not require such a reduction. For this reason, we can prove in
the next section that r-Dynamic is indeed optimal for 𝑘 = 2 (see the proof of Corollary 1 next).
2 We conjecture that r-Dynamic is optimal even in the small inventory regime if a tighter analysis is performed.

28
G Proof of Corollary 1
In this section, we prove that for an arbitrary instance I, the expected performance of r-Dynamic, denoted
I)
as E[ALG(I)], is at least OPT(
𝛼 ∗ (2) . S
Let 𝑣 1∗, 𝑣 2∗ denote the two highest valuations in the instance I (we omit the proof for the trivial case
with only one buyer in I). Depending on the values of 𝑣 1∗ and 𝑣 2∗ , the following three cases occur. In each
OPT( I ) 𝑣1∗ +𝑣2∗ −𝑐 1 −𝑐 2
scenario, we prove that E[ALG(I)] ≥ 𝛼 S∗ (2) = 𝛼 S∗ (2) holds.
Case I: 𝑣 1∗ ≤ 𝑣 2∗ ≤ 𝑈 1 . Let random variables 𝑋 1 and 𝑋 2 denote the valuations of the buyers that purchase
the first and second unit of the item, respectively. Then, it follows that

E[ALG(I)]
= E𝒔∼𝑈 2 (0,1) [𝑋 1 + 𝑋 2 − 𝑐 1 − 𝑐 2 ]
∫ 𝜙 −1 (𝑣 ∗ )
1 1
≥ (𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1 + (𝑣 2∗ − 𝑐 1 ) · (𝜙 1−1 (𝑣 2∗ ) − 𝜙 1−1 (𝑣 1∗ )),
𝑠 1 =0
∫ 𝜙 1−1 (𝑣1∗ ) ∫ 𝜙 1−1 (𝑣2∗ )
≥ (𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1 + (𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1
𝑠 1 =0 𝑠 1 =𝜙 1−1 (𝑣1∗ )
∫ 𝜙 1−1 (𝑣2∗ )
= (𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1
𝑠 1 =0
𝑣 1∗ + 𝑣 2∗ − 𝑐 1 − 𝑐2
≥
𝛼 ∗S (2)
OPT(I)
= ,
𝛼 ∗S (2)

where the first two terms in the first inequality arise from the fact that if the realized price for the first
unit of the item, denoted as 𝑃 1 = 𝜙 1 (𝑠 1 ), is set below 𝑣 1∗ , then in the worst-case scenario, the value obtained
from the first item will be at least equal to 𝜙 1 (𝑠 1 ). The subsequent two terms are included because if the
price for the first item falls within the range from 𝑣 1∗ to 𝑣 2∗ , then the first item is allocated to the buyer
whose valuation is 𝑣 2∗ . The second inequality follows since 𝜙 1 (𝑠 1 ) is an non-decreasing function. The third
inequality follows from the design of 𝜙 1 (𝑠 1 ) in Theorem 1.
Case II: 𝑣 1∗ ≤ 𝑈 1 = 𝐿2 ≤ 𝑣 2∗ ≤ 𝑈 . In this case, we have

E[ALG(I)]
= E𝒔∼𝑈 2 (0,1) [𝑋 1 − 𝑐 1 ] + E𝒔∼𝑈 2 (0,1) [𝑋 2 − 𝑐 2 ]
∫ 𝜙 −1 (𝑣 ∗ )
1 1
≥ (𝜙 1 (𝑠 1 ) − 𝑐 1 )𝑑𝑠 1 + (𝑣 2∗ − 𝑐 1 ) · (1 − 𝜙 1−1 (𝑣 1∗ )) + (𝑣 2∗ − 𝑐 2 ) · 𝜙 1−1 (𝑣 1∗ ) · 𝜙 2−1 (𝑣 2∗ )
𝑠 1 =0
2 · 𝑣 1∗ − 𝑐 1 − 𝑐 2
= + (𝑣 2∗ − 𝑐 1 ) · (1 − 𝜙 1−1 (𝑣 1∗ )) + (𝑣 2∗ − 𝑐 2 ) · 𝜙 1−1 (𝑣 1∗ ) · 𝜙 2−1 (𝑣 2∗ ).
𝛼 ∗S (2)
OPT( I ) 𝑣1∗ +𝑣2∗ −𝑐 1 −𝑐 2
To prove E[ALG(I)] ≥ 𝛼 S∗ (2) = 𝛼 S∗ (2) , we define the following function

2 · 𝑣 1∗ − 𝑐 1 − 𝑐 2
𝐺 (𝑣 1∗, 𝑣 2∗ ) = + (𝑣 2∗ − 𝑐 1 ) · (1 − 𝜙 1−1 (𝑣 1∗ ))+
𝛼 ∗S (2)
∗ −1 ∗ −1 ∗ 𝑣 1∗ + 𝑣 2∗ − 𝑐 1 − 𝑐 2
(𝑣 2 − 𝑐 2 ) · 𝜙 1 (𝑣 1 ) · 𝜙 2 (𝑣 2 ) − .
𝛼 ∗S (2)

29
Then the goal is to prove 𝐺 (𝑣 1∗, 𝑣 2∗ ) ≥ 0 in its domain 𝐿 ≤ 𝑣 1∗ ≤ 𝑈 1 and 𝐿2 ≤ 𝑣 2∗ ≤ 𝑈 . The proposition below
formally states this result.
Proposition 3. For all 𝑣 1∗ ∈ [𝐿1, 𝑈 1 ] and 𝑣 2∗ ∈ [𝐿2, 𝑈 2 ], we have 𝐺 (𝑣 1∗, 𝑣 2∗ ) ≥ 0.

We deferred the proof of the above proposition to Appendix H. The idea is to simply prove that
𝐺 (𝑣 1∗, 𝑣 2∗ ) ≥ 0 holds at all extreme points within its domain.
Case III: 𝐿2 ≤ 𝑣 1∗ ≤ 𝑣 2∗ . In this case, we show that we can lower bound the expected performance of
r-Dynamic as follows:
E[ALG(I)]
= E𝒔∼𝑈 2 (0,1) [𝑋 1 − 𝑐 1 ] + E𝒔∼𝑈 2 (0,1) [𝑋 2 − 𝑐 2 ]
∫ 1 ∫ 𝜙 −1 (𝑣 ∗ )
2 1
≥ (𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1 + (𝜙 2 (𝑠 2 ) − 𝑐 2 ) · 𝑑𝑠 2 + (𝑣 2∗ − 𝑐 2 ) · (𝜙 2−1 (𝑣 2∗ ) − 𝜙 2−1 (𝑣 1∗ ))
𝑠 1 =0 𝑠 2 =0
∫ 1 ∫ 𝜙 2−1 (𝑣1∗ ) ∫ 𝜙 2−1 (𝑣2∗ )
≥ (𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1 + (𝜙 2 (𝑠 2 ) − 𝑐 2 ) · 𝑑𝑠 2 + (𝜙 2 (𝑠 2 ) − 𝑐 2 ) · 𝑑𝑠 2
𝑠 1 =0 𝑠 2 =0 𝑠 2 =𝜙 2−1 (𝑣1∗ )
∫ 1 ∫ 𝜙 2−1 (𝑣2∗ )
= (𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1 + (𝜙 2 (𝑠 2 ) − 𝑐 2 ) · 𝑑𝑠 2
𝑠 1 =0 𝑠 2 =0
∫ 𝜙 2−1 (𝑣2∗ )
2 · 𝑈1 − 𝑐 1 − 𝑐 2
= + (𝜙 2 (𝑠 2 ) − 𝑐 2 ) · 𝑑𝑠 2
𝛼 ∗S (2) 𝑠 2 =0
2 · 𝑣 2∗ − 𝑐 1 − 𝑐 2
=
𝛼 ∗S (2)
OPT(I)
≥ ∗ ,
𝛼 S (2)
where the first term in the first inequality arises from the fact that if the realized price for the first unit of
the item, denoted as 𝑃 1 = 𝜙 1 (𝑠 1 ), is set below 𝐿2 , then in the worst-case scenario, the value obtained from
the first item will be at least equal to 𝜙 1 (𝑠 1 ). The second and third terms follow the same reasoning. The
second inequality follows the fact that 𝜙 2 (𝑠 2 ) is non-decreasing. The third and forth equalities follow the
design of 𝜙 1 (𝑠 1 ) and 𝜙 2 (𝑠 2 ) in Theorem 1.
Combining the analysis of the above three cases, Corollary 1 follows.

H Proof of Proposition 3
𝜕𝐺 (𝑣1∗ ,𝑣2∗ )
We first evaluate the value of 𝐺 (𝑣 1∗, 𝑣 2∗ ) at its critical points, that is, at the points where 𝜕𝑣1∗ = 0 and
𝜕𝐺 (𝑣1∗ ,𝑣2∗ )
𝜕𝑣2∗ = 0, and show that 𝐺 (𝑣 1∗, 𝑣 2∗ ) ≥ 0 holds at these critical points. After that, the proposition follows
by evaluating the values of 𝐺 (𝑣 1∗, 𝑣 2∗ ) at the four boundary hyperplanes of its domain.
𝜕𝐺 (𝑣 ∗ ,𝑣 ∗ )
First, let us compute 𝜕𝑣1∗ 2 . It follows that:
1

𝜕𝐺 (𝑣 1∗, 𝑣 2∗ ) 1 2 𝑣 2∗ − 𝑐 1 2 𝑣 2∗ − 𝑐 2 −1 ∗
= − · + · · 𝜙 (𝑣 ).
𝜕𝑣 1∗ 𝛼 ∗S (2) 𝛼 ∗S (2) 𝑣 1∗ − 𝑐 1 𝛼 ∗S (2) 𝑣 1∗ − 𝑐 1 2 2
Setting the right-hand side of above equation to be zero, we have
𝑣 1∗ − 𝑐 1
𝜙 2−1 (𝑣 2∗ ) · (𝑣 2∗ − 𝑐 2 ) = 𝑣 2∗ − 𝑐 1 − .
2

30
 𝜕𝐺 (𝑣1∗ ,𝑣2∗ )
Using the equation above, we then compute 𝐺 (𝑣 1∗, 𝑣 2∗ ) at the points that 𝜕𝑣1∗ = 0, it follows that:

𝑣 1∗ − 𝑣 2∗ 𝑣 1∗ − 𝑐 1
𝐺 (𝑣 1∗, 𝑣 2∗ ) = + (𝑣 ∗
− 𝑐 ) · (1 − 𝜙 −1 ∗
(𝑣 )) + (𝑣 ∗
− 𝑐 − ) · 𝜙 1−1 (𝑣 1∗ )
𝛼 ∗S (2) 2 1 1 1 2 1
2
𝑣∗ − 𝑣∗ 𝑣 ∗ − 𝑐 1 −1 ∗
= 1∗ 2 + (𝑣 2∗ − 𝑐 1 ) − 1 · 𝜙 1 (𝑣 1 )
𝛼 S (2) 2
𝑣∗ − 𝑣∗ 𝑣 ∗ − 𝑐1
≥ 1∗ 2 + (𝑣 2∗ − 𝑐 1 ) − 1
𝛼 S (2) 2
1 1 1 𝑐1
=𝑣 1∗ · ( ∗ − ) + 𝑣 2∗ · (1 − ∗ )−
𝛼 S (2) 2 𝛼 S (2) 2
∗
𝑣 − 𝑐1
≥ 1
2
>0,

leading to the conclusion that 𝐺 (𝑣 1∗, 𝑣 2∗ ) ≥ 0 holds at its critical points.

Next, we consider the boundary hyperplanes and prove that 𝐺 (𝑣 1∗, 𝑣 2∗ ) is positive in all four boundary
planes given below:
• 𝐺 (𝐿1, 𝑣 2∗ ), ∀𝑣 2∗ ∈ [𝐿2, 𝑈 2 ].
• 𝐺 (𝑈 1, 𝑣 2∗ ), ∀𝑣 2∗ ∈ [𝐿2, 𝑈 2 ].
• 𝐺 (𝑣 1∗, 𝐿2 ), ∀𝑣 1∗ ∈ [𝐿1, 𝑈 1 ].
• 𝐺 (𝑣 1∗, 𝑈 2 ), ∀𝑣 1∗ ∈ [𝐿1, 𝑈 1 ].
We start with the first one 𝐺 (𝐿1, 𝑣 2∗ ):

2 · 𝐿 − 𝑐1 − 𝑐2 𝐿 + 𝑣 2∗ − 𝑐 1 − 𝑐 2
𝐺 (𝐿, 𝑣 2∗ ) = + (𝑣 ∗
− 𝑐 ) −
𝛼 ∗S (2) 𝛼 ∗S (2)
2 2

𝑣∗ − 𝐿
=(𝑣 2∗ − 𝑐 2 ) − 2∗
𝛼 S (2)
≥0, ∀𝐿2 ≤ 𝑣 2∗ ≤ 𝑈 2,

where the equations above follow since 𝐿 ≥ 𝑐 2 holds (the assumption that the marginal production costs
are always less than the valuations).
For the second one 𝐺 (𝑈 1, 𝑣 2∗ ):

2 · 𝑈1 − 𝑐 1 − 𝑐 2 𝑈 1 + 𝑣 2∗ − 𝑐 1 − 𝑐 2
𝐺 (𝑈 1, 𝑣 2∗ ) = + (𝑣 ∗
− 𝑐 ) · 𝜙 −1 ∗
(𝑣 ) −
𝛼 ∗S (2) 𝛼 ∗S (2)
2 2 2 2

𝑣 ∗ − 𝑈1
=(𝑣 2∗ − 𝑐 2 ) · 𝜙 2−1 (𝑣 2∗ ) − 2 ∗
𝛼 S (2)
≥0, ∀𝐿2 ≤ 𝑣 2∗ ≤ 𝑈 2 = 𝑈 .
∫ 𝜙 −1 (𝑣 ∗ ) 𝑣2∗ −𝑈 1
The equations above follow since (𝑣 2∗ − 𝑐 2 ) · 𝜙 2−1 (𝑣 2∗ ) ≥ 2
𝑠 2 =0
(𝜙 2 (𝑠 2 ) − 𝑐 2 ) · 𝑑𝑠 2 ≥ 2 · 𝛼 S∗ (2) based on the
definition of 𝜙 2 (𝑠).

31
 For the third one 𝐺 (𝑣 1∗, 𝐿2 ):
2 · 𝑣 1∗ − 𝑐 1 − 𝑐 2 𝑣 1∗ + 𝐿2 − 𝑐 1 − 𝑐 2
𝐺 (𝑣 1∗, 𝐿2 ) = −1 ∗
+ (𝐿2 − 𝑐 1 ) · 𝜙 1 (𝑣 1 ) −
𝛼 ∗S (2) 𝛼 ∗S (2)
𝐿2 − 𝑣 ∗
=(𝐿2 − 𝑐 1 ) · (1 − 𝜙 1−1 (𝑣 1∗ )) − ∗ 1
𝛼 S (2)
≥0, ∀𝐿1 ≤ 𝑣 1∗ ≤ 𝑈 1,
∫ 𝜙 −1 (𝐿2 ) 𝐿2 −𝑣1∗
where the above equation follows since (𝐿2 − 𝑐 1 ) · (1 − 𝜙 1−1 (𝑣 1∗ )) ≥ 1
𝑠 1 =𝜙 1−1 (𝑣1∗ )
(𝜙 1 (𝑠 1 ) − 𝑐 1 ) · 𝑑𝑠 1 ≥ 2 · 𝛼 S∗ (2)
(based on the definition of 𝜙 1 (𝑠)).
Finally, for the last one 𝐺 (𝑣 1∗, 𝑈 2 ):
2 · 𝑣 1∗ − 𝑐 1 − 𝑐 2 𝑣 1∗ + 𝑈 2 − 𝑐 1 − 𝑐 2
𝐺 (𝑣 1∗, 𝑈 2 ) = + (𝑈 − 𝑐 ) · (1 − 𝜙 −1 ∗
(𝑣 )) + (𝑈 − 𝑐 ) · 𝜙 −1 ∗
(𝑣 ) −
𝛼 ∗S (2) 𝛼 ∗S (2)
2 1 1 1 2 2 1 1

𝑈2 − 𝑣 ∗
≥(𝑈 2 − 𝑐 2 ) − ∗ 1
𝛼 S (2)
≥0, ∀𝐿 = 𝐿1 ≤ 𝑣 1∗ ≤ 𝑈 1,

where the equations above follow since 𝑣 1∗ ≥ 𝑐 2 holds (again, the assumption that the marginal production
costs are always less than the valuations).
Combining all the above analysis, we thus complete the proof of Proposition 3.

I Extension of the Lower Bound Results to General Production Cost

Functions
In this section, we extend our lower bound result in Theorem 3, originally developed for the high-value
case,3 to general cumulative production cost functions.
Before presenting the main theorem on obtaining a lower bound for general cost functions, let us
introduce some notations. Define 𝑓 ∗
(𝑣) : [𝐿, 𝑈 ] → R as the conjugate of the total production cost function,
where 𝑓 ∗ (𝑣) = max𝑖 ∈ [𝑘 ] 𝑣 · 𝑖 − 𝑓 (𝑖) . Additionally, let 𝑔(𝑣) be defined as
∑︁
𝑔(𝑣) = (𝑓 ∗ ) ′ (𝑣) = 1 {𝑣 ≥𝑐𝑖 } ,
𝑖 ∈ [𝑘 ]

where 1 {𝐴} is the standard indicator function. Let 𝑘 denote the smallest natural number such that:
¯
𝑘
¯
∑︁ 1
(𝐿 − 𝑐𝑖 ) > · 𝑓 ∗ (𝐿).
𝑖=1
𝛼

Following Theorem 3, we also define 𝜉 as follows:

Í𝑘 −1
1
𝛼 · 𝑓 ∗ (𝐿) − 𝑖=1
¯ (𝐿 − 𝑐 𝑖 )
𝜉= .
𝐿 − 𝑐𝑘
¯
Theorem 5 below extends our lower bound results to settings with general cost functions.
3 This corresponds to the case when 𝑐 < 𝐿, or equivalently, the lowest possible valuation 𝐿 is no less than the highest marginal
𝑘
production cost 𝑐𝑘 .

32
Theorem 5. Given S = {𝐿, 𝑈 , 𝑓 } for the OSDoS problem with 𝑘 ≥ 1 and general production cost functions 𝑓 ,
no online algorithm, including those with randomization, can achieve a competitive ratio smaller than 𝛼 ∗S (𝑘),
where 𝛼 ∗S (𝑘) is the solution to the following system of equations of 𝛼:
∫ 𝑢𝑘
¯ 𝑔(𝜂)
𝑑𝜂 = 1 − 𝜉, (17)
𝜂=𝐿 𝛼 · (𝜂 − 𝑐 𝑘 )
∫ 𝑢𝑖 ¯
¯ 𝑔(𝜂)
𝑑𝜂 = 1, 𝑢𝑖 = ℓ𝑖+1, 𝑖 = 𝑘 + 1, . . . , 𝑘, (18)
𝜂=ℓ𝑖 𝛼 · (𝜂 − 𝑐 𝑖 ) ¯
𝑢𝑘 = 𝑈 . (19)

Proof. The proof proceeds similarly to the proof of Theorem 3 until the derivation of Eq. (5). Given the
arrival instance I (𝜖 ) up to the end of stage-𝑣, the objective of the offline optimal algorithm equals 𝑓 ∗ (𝑣).
Therefore, we reformulate Eq. (4) as follows:
  1
ALG I𝑣(𝜖 ) ≥ · 𝑓 ∗ (𝑣), ∀𝑣 ∈ [𝐿, 𝑈 ].
𝛼
In the case of general production cost functions, we derive the following inequality to capture the produc-
tion level changes of an 𝛼-competitive algorithm:
𝑘 𝑘 ∫ 𝑣
∑︁ ∑︁ 1 ∗
𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖 (𝜂) ≥ · 𝑓 (𝑣). (20)
𝑖=1 𝑖=1 𝜂=𝐿 𝛼

In addition, we define 𝛼 ∗S (𝑘) as follows:

n
𝛼 ∗S (𝑘) = inf 𝛼 ≥ 1 there exist a set of 𝑘 allocation
o
functions {𝜓𝑖 (𝑣)} ∀𝑖 ∈ [𝑘 ] ∈ Ω that satisfy Eq. (20) .

From this point onward, the proof continues in the same manner as the proof of Theorem 3. Let us define
the function 𝜒 𝛼 (𝑣) : [𝐿, 𝑈 ] → [0, 𝑘] and the set of functions {𝜓𝑖𝛼 (𝑣)}𝑖 ∈ [𝑘 ] as specified in Eq. (6) and Eq. (2).
Consequently, Lemma 3 holds as long as we have increasing marginal production costs (i.e., diseconomies
of scale) and Lemma 4 that follows the definition of {𝜓𝑖𝛼 (𝑣)}𝑖 ∈ [𝑘 ] holds in this case as well.
The primary distinction between the two proofs arises in the following proposition, which gives an
explicit design of the function {𝜓𝑖𝛼 } ∀𝑖 ∈ [𝑘 ] by replacing the inequality with an equality in Eq. (20).
Proposition 4. For any 𝛼 ≥ 𝛼 ∗S (𝑘), there exist a unique set of functions {𝜓𝑖𝛼 (𝑣)} ∀𝑖 ∈ [𝑘 ] that satisfy Eq. (20)
with an equality:

𝜓𝑖𝛼 (𝑣) = 1, ∀𝑣 ∈ [𝐿, 𝑈 ], 1 ≤ 𝑖 ≤ 𝑘 − 1,

¯


 0 𝑣 ≤ ℓ𝑘 ,

 ∫𝑣 𝑔 (𝜂 ) ¯
𝜓𝑘 (𝑣) = 𝜉 + 𝜂=𝐿 𝛼 · (𝜂−𝑐𝑖 ) 𝑑𝜂, 𝑣 ∈ [𝐿, 𝑢𝑘 ],
𝛼
¯ 
 ¯
1
 𝑣 ≥ 𝑢𝑘 ,
¯


 0 𝑣 ≤ ℓ𝑖 ,
∫ 𝑣
 𝑔 (𝜂 )
𝜓𝑖𝛼 (𝑣) = 𝜂=ℓ𝑖 𝛼 · (𝜂−𝑐𝑖 ) 𝑑𝜂, 𝑣 ∈ [ℓ𝑖 , 𝑢𝑖 ], , 𝑖 = 𝑘 + 1, . . . , 𝑘 − 1.

 ¯
1
 𝑣 ≥ 𝑢𝑖 ,
(
0 𝑣 ≤ ℓ𝑘 ,
𝜓𝑘𝛼 (𝑣) = ∫ 𝑣 𝑔 (𝜂 )
𝜂=ℓ 𝛼 · (𝜂−𝑐𝑘 )
𝑑𝜂, 𝑣 ∈ [ℓ𝑘 , 𝑈 ],
𝑘

33
where the intervals are specified by:
∫ 𝑢𝑘
¯ 𝑔(𝜂)
𝑑𝜂 = 1 − 𝜉, (21)
𝜂=𝐿 𝛼 · (𝜂 − 𝑐𝑘 )
∫ 𝑢𝑖 ¯
¯ 𝑔(𝜂)
𝑑𝜂 = 1, 𝑢𝑖 = ℓ𝑖+1, ∀𝑖 = 𝑘 + 1, . . . , 𝑘. (22)
𝜂=ℓ𝑖 𝛼 · (𝜂 − 𝑐𝑖 ) ¯

In the proposition above, for any given 𝛼 ≥ 𝛼 ∗S (𝑘), the values of 𝑢𝑖 and ℓ𝑖 can be determined. We begin
by solving Eq. (21) to find the value of 𝑢𝑘 , and then proceed to find the value of other variables {𝑢𝑖 } ∀𝑖
¯
using Eq. (22).
Based on the above proposition, as the value of 𝛼 decreases, the value of 𝑢𝑘 also decreases. Again,
following the same reasoning as the proof of Theorem 3, the lower bound 𝛼 ∗S (𝑘) is the value of 𝛼 for
which 𝑢𝑘 computed above is equal to 𝑈 . We thus complete the proof of Theorem 5. ■

J Extension of the Upper Bound Results to General Production Cost

Functions
In this section, we extend the randomized dynamic pricing scheme r-Dynamic, originally developed for
the high-value case, to general cumulative production cost functions.
Theorem 6. Given S = {𝐿, 𝑈 , 𝑓 } for the OSDoS problem with 𝑘 ≥ 1, r-Dynamic (Algorithm 1) is max𝑖 ∈ [𝑘 ] 𝛼 ∗S (𝑘)·
𝑖 −𝑐 𝑖
(1+ 𝑓 𝑈∗ (𝑈 𝑖 −1 )
)-competitive for the following design of the pricing functions {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] , where 𝛼 ∗S (𝑘) is the lower
bound obtained in Theorem 5:

𝜙𝑖 (𝑠) = 𝐿, ∀𝑠 ∈ [0, 1], 𝑖 ∈ [𝑘 ∗ − 1],

( ¯
𝐿 𝑠 ∈ [0, 𝜉 ∗ ],
𝜙𝑘 ∗ (𝑠) =
¯ 𝜓𝑘−1
∗ (𝑠) 𝑠 ∈ (𝜉 ∗, 1],
¯
𝜙𝑖 (𝑠) = 𝜓𝑖−1 (𝑠), ∀𝑠 ∈ [0, 1], 𝑖 = 𝑘 ∗ + 1, . . . , 𝑘,
¯
where the set of functions {𝜓𝑖 } ∀𝑖 ∈ {𝑘 ∗,··· ,𝑘 } are defined as follows:
¯
∫ 𝑣
𝑔(𝜂)
𝜓𝑘 ∗ (𝑣) = 𝜉 ∗ + 𝑑𝜂, ∀𝑣 ∈ [𝐿, 𝑈𝑘 ∗ ],
¯ 𝜂=𝐿 𝛼 · (𝜂 − 𝑐 𝑖 ) ¯
∫ 𝑣
𝑔(𝜂)
𝜓𝑖 (𝑣) = 𝑑𝜂, ∀𝑣 ∈ [𝐿𝑖 , 𝑈𝑖 ], 𝑖 = 𝑘 ∗ + 1, . . . , 𝑘;
𝜂=ℓ𝑖 𝛼 · (𝜂 − 𝑐 𝑖 ) ¯

the parameters 𝑘 ∗ and 𝜉 ∗ are respectively the values of 𝑘 and 𝜉 defined in Appendix I, corresponding to 𝛼 =
¯ ¯
𝛼 ∗S (𝑘), and the price intervals {[𝐿𝑖 , 𝑈𝑖 ]} ∀𝑖 ∈ [𝑘 ] are given as follows:
∫ 𝑈𝑘 ∗
¯ 𝑔(𝜂)
𝑑𝜂 = 1 − 𝜉,
𝜂=𝐿 𝛼 · (𝜂 − 𝑐𝑘 )
∫ ¯
𝑈𝑖
𝑔(𝜂)
𝑑𝜂 = 1, 𝑢𝑖 = ℓ𝑖+1, ∀𝑖 = 𝑘 ∗ + 1, . . . , 𝑘.
𝜂=𝐿𝑖 𝛼 · (𝜂 − 𝑐𝑖 ) ¯

Proof. The proof will follow the same process as the proof in Appendix F. So we only provide a brief proof
sketch.

34
 Consider an arbitrary arrival instance I = {𝑣𝑡 }𝑡 ∈ [𝑇 ] . Recall that the random price vector P = {𝑃 1, · · · , 𝑃𝑘 }
is generated using the pricing functions {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] at the beginning of r-Dynamic (line 3 of Algorithm 1).
Let us define the random variable 𝑊 (P), the variable 𝜔 and the price vector 𝝅, the set {𝜈𝑖 , 𝜏𝑖 } ∀𝑖 ∈ [𝜔 ] , and
𝑊 𝜏𝜔 (P) in the same fashion as in Appendix F.
Following the same reasoning, the property in Eq. (16) can be derived for {𝜈𝑖 } ∀𝑖 ∈ [𝜔 ] , and the lemmas
8, 9, and 10 follow as well.
We also define B ⊆ I, as before, to be the set of highest-valued buyers to whom the offline optimal
algorithm allocates a unit of the item in instance I. We further divide B into two subsets: B1 and B2 , as
done in the previous proof. Additionally, we partition B1 into two subsets: B1,1 and B1,2 , as before.
We continue our analysis for two separate cases that can arise depending on the instance I. In this
proof, we only provide the proof for the first case and the proof of the second case follows similarly as
Appendix F.
Case 1: In this case, no buyer in B2 has a valuation greater than 𝑈𝜔 −1 except for the buyer at time 𝜏𝜔 .
Therefore, the buyer at time 𝜏𝜔 possesses the highest valuation in instance I. The following upper bound
can be derived for OPT(I), which denotes the objective value of the offline optimal algorithm:
|B|
∑︁
OPT(I) = 𝑉 (B1 ) + 𝑉 (B2 ) − 𝑐𝑖
𝑖=1
|B|
∑︁
≤ 𝑉 (B1 ) + (|B2 | − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
|B|
∑︁
= 𝑉 (B1,1 ) + 𝑉 (B1,2 ) + (|B2 | − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
≤ |B1,1 | · 𝑈𝜔 −1 + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 )
|B|
∑︁
+ |B1,2 | · 𝑈𝜔 −1 + (|B2 | − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔 − 𝑐𝑖
𝑖=1
= (|B1,1 | + |B1,2 | + |B2 | − 1) · 𝑈𝜔 −1 + 𝜈𝜏𝜔
|B|
∑︁
+ (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 ) − 𝑐𝑖
𝑖=1
≤ 𝑓 ∗ (𝑈𝜔 −1 ) + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 ) + 𝜈𝜏𝜔 − 𝑐 𝜔 ,

where the first inequality follows the condition of Case 1. The second inequality follows the definition
of the sets B1,1 and B1,2 . Finally, the third inequality follows since based on definition of 𝑓 ∗ , we have
Í | B | −1
(|B1,1 | + |B1,2 | + |B2 | − 1) · 𝑈𝜔 −1 − 𝑖=1 𝑐𝑖 ≤ 𝑓 ∗ (𝑈𝜔 −1 ).
Moving forward, we can lower bound the expected performance of r-Dynamic under I, denoted by
E[ALG(I)], using the same approach as before.
𝜔
∑︁−1 ∫ 1 𝜔
∑︁−1
E[ALG(I)] ≥ 𝜙𝑖 (𝜂)𝑑𝜂 + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 ) − 𝑐𝑖 .
𝑖=1 0 𝑖=1

35
 Based on the definition of {𝜙𝑖 } ∀𝑖 ∈ [𝑘 ] , we have:
𝜔
∑︁−1 ∫ 1 𝜔
∑︁−1
E[ALG(I)] ≥ 𝜙𝑖 (𝜂)𝑑𝜂 − 𝑐𝑖 + (𝑉 (B1,1 ) − |B1 | · 𝑈𝜔 −1 )
𝑖=1 0 𝑖=1
𝑘
∑︁ 𝜔
∑︁−1 ∫ 𝑈𝜔 −1
= 𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖 (𝜂) + (𝑉 (B1,1 ) − |B1 | · 𝑈𝜔 −1 ).
𝑖=1 𝑖=1 𝜂=𝐿

Furthermore, based on the design of {𝜓𝑖 } ∀𝑖 ∈ [𝑘 ] in Theorem 6, we have

𝑘
∑︁ 𝜔
∑︁−1 ∫ 𝑈𝜔 −1
𝜓𝑖 (𝐿) · (𝐿 − 𝑐𝑖 ) + (𝜂 − 𝑐𝑖 )𝑑𝜓𝑖 (𝜂) + (𝑉 (B1 ) − |B1 | · 𝑈𝜔 −1 )
𝑖=1 𝑖=1 𝜂=𝐿
1 ∗
≥ 𝑓 (𝑈𝜔 −1 ) + (𝑉 (B1 ) − |B1 | · 𝑈𝜔 −1 ).
𝛼 ∗S (𝑘)

Putting together the above lower and upper bounds, it follows that:

OPT(I) 𝑓 ∗ (𝑈𝜔 −1 ) + (𝑉 (B1,1 ) − |B1,1 | · 𝑈𝜔 −1 ) + 𝜈𝜏𝜔 − 𝑐 𝜔

≤ ∗
E[ALG(I)] 1
𝛼 ∗ (𝑘 ) 𝑓 (𝑈𝜔 −1 ) + (𝑉 (B1 ) − |B1 | · 𝑈𝜔 −1 )
S
𝑓 ∗ (𝑈𝜔 −1 ) + 𝜈𝜏𝜔 − 𝑐 𝜔
≤ 1 ∗
𝛼 S∗ (𝑘 ) 𝑓 (𝑈𝜔 −1 )
 
𝜈𝜏 − 𝑐 𝜔
=𝛼 ∗S (𝑘) · 1 + ∗ 𝜔
𝑓 (𝑈𝜔 −1 )
 
∗ 𝑈𝜔 − 𝑐 𝜔
≤𝛼 S (𝑘) · 1 + ∗
𝑓 (𝑈𝜔 −1 )
 
∗ 𝑈𝑖 − 𝑐 𝑖
≤ max 𝛼 S (𝑘) · 1 + ∗ .
𝑖 ∈ [𝑘 ] 𝑓 (𝑈𝑖 −1 )

Case 2: In the set of buyers B2 , there are other buyers with valuations greater than 𝑈𝜔 −1 besides the
buyer at time 𝜏𝜔 . The proof in this case follows the same structure as the proof above and the proof in
Appendix F. ■

36