Searcharxiv⌕ Search

arXiv subjects

Linda Cai

Publications and source records attributed to Linda Cai.

16 recordsLinked to original sources
arXiv2025

Cost-aware Stopping for Bayesian Optimization

In automated machine learning, scientific discovery, and other applications of Bayesian optimization, deciding when to stop evaluating expensive black-box functions in a cost-aware manner is an important but underexplored practical consideration. A natural performance metric for this purpose is the cost-adjusted simple regret, which explicitly captures the trade-off between solution quality and cumulative evaluation cost. Existing stopping rules for Bayesian optimization are either heuristic, or are theoretically grounded but designed to optimize simple regret without accounting for evaluation costs; as a result, they provide no guarantees against unnecessary evaluations when costs are high. We propose a principled cost-aware stopping rule for Bayesian optimization that adapts to varying evaluation costs without heuristic tuning. Our rule is grounded in a theoretical connection to state-of-the-art cost-aware acquisition functions, namely the Pandora's Box Gittins Index (PBGI) and log expected improvement per cost (LogEIPC). When paired with either acquisition function, we prove that the resulting policy satisfies a theoretical guarantee bounding the expected cost-adjusted simple regret. Across synthetic tasks and empirical benchmarks including hyperparameter optimization and neural architecture size search, pairing our stopping rule with PBGI or LogEIPC usually matches or outperforms other acquisition-function--stopping-rule pairs in terms of cost-adjusted simple regret.

cs.LG↗
arXiv2025

Competition Complexity in Multi-Item Auctions: Beyond VCG and Regularity

We quantify the value of the monopoly's bargaining power in terms of competition complexity--that is, the number of additional bidders the monopoly must attract in simple auctions to match the expected revenue of the optimal mechanisms (c.f., Bulow and Klemperer, 1996, Eden et al., 2017)--within the setting of multi-item auctions. We show that for simple auctions that sell items separately, the competition complexity is $Θ(\frac{n}α)$ in an environment with $n$ original bidders under the slightly stronger assumption of $α$-strong regularity, in contrast to the standard regularity assumption in the literature, which requires $Ω(n \cdot \ln \frac{m}{n})$ additional bidders (Feldman et al., 2018). This significantly reduces the value of learning the distribution to design the optimal mechanisms, especially in large markets with many items for sale. For simple auctions that sell items as a grand bundle, we establish a constant competition complexity bound in a single-bidder environment when the number of items is small or when the value distribution has a monotone hazard rate. Some of our competition complexity results also hold when we compete against the first best benchmark (i.e., optimal social welfare).

cs.GT↗
arXiv2025

Automatic Raman Measurements in a High-Throughput Bioprocess Development Lab

This study presents a collection of physical devices and software services that fully automate Raman spectra measurements for liquid samples within a robotic facility. This method is applicable to various fields, with demonstrated efficacy in biotechnology, where Raman spectroscopy monitors substrates, metabolites, and product-related concentrations. Our system specifically measures 50 $\micro L$ samples using a liquid handling robot capable of taking 8 samples simultaneously. We record multiple Raman spectra for 10s each. Furthermore, our system takes around 20s for sample handling, cleaning, and preparation of the next measurement. All spectra and metadata are stored in a database, and we use a machine learning model to estimate concentrations from the spectra. This automated approach enables gathering spectra for various applications under uniform conditions in high-throughput fermentation processes, calibration procedures, and offline evaluations. This allows data to be combined to train sophisticated machine learning models with improved generalization. Consequently, we can develop accurate models more quickly for new applications by reusing data from prior applications, thereby reducing the need for extensive calibration data.

q-bio.QM↗
arXiv2024

Faster Diffusion Sampling with Randomized Midpoints: Sequential and Parallel

Sampling algorithms play an important role in controlling the quality and runtime of diffusion model inference. In recent years, a number of works~\cite{chen2023sampling,chen2023ode,benton2023error,lee2022convergence} have proposed schemes for diffusion sampling with provable guarantees; these works show that for essentially any data distribution, one can approximately sample in polynomial time given a sufficiently accurate estimate of its score functions at different noise levels. In this work, we propose a new scheme inspired by Shen and Lee's randomized midpoint method for log-concave sampling~\cite{ShenL19}. We prove that this approach achieves the best known dimension dependence for sampling from arbitrary smooth distributions in total variation distance ($\widetilde O(d^{5/12})$ compared to $\widetilde O(\sqrt{d})$ from prior work). We also show that our algorithm can be parallelized to run in only $\widetilde O(\log^2 d)$ parallel rounds, constituting the first provable guarantees for parallel sampling with diffusion models. As a byproduct of our methods, for the well-studied problem of log-concave sampling in total variation distance, we give an algorithm and simple analysis achieving dimension dependence $\widetilde O(d^{5/12})$ compared to $\widetilde O(\sqrt{d})$ from prior work.

cs.LG↗
arXiv2024

Profitable Manipulations of Cryptographic Self-Selection are Statistically Detectable

Cryptographic Self-Selection is a common primitive underlying leader-selection for Proof-of-Stake blockchain protocols. The concept was first popularized in Algorand [CM19], who also observed that the protocol might be manipulable. [FHWY22] provide a concrete manipulation that is strictly profitable for a staker of any size (and also prove upper bounds on the gains from manipulation). Separately, [YSZ23, BM24] initiate the study of undetectable profitable manipulations of consensus protocols with a focus on the seminal Selfish Mining strategy [ES14] for Bitcoin's Proof-of-Work longest-chain protocol. They design a Selfish Mining variant that, for sufficiently large miners, is strictly profitable yet also indistinguishable to an onlooker from routine latency (that is, a sufficiently large profit-maximizing miner could use their strategy to strictly profit over being honest in a way that still appears to the rest of the network as though everyone is honest but experiencing mildly higher latency. This avoids any risk of negatively impacting the value of the underlying cryptocurrency due to attack detection). We investigate the detectability of profitable manipulations of the canonical cryptographic self-selection leader selection protocol introduced in [CM19] and studied in [FHWY22], and establish that for any player with $α< \frac{3-\sqrt{5}}{2} \approx 0.38$ fraction of the total stake, every strictly profitable manipulation is statistically detectable. Specifically, we consider an onlooker who sees only the random seed of each round (and does not need to see any other broadcasts by any other players). We show that the distribution of the sequence of random seeds when any player is profitably manipulating the protocol is inconsistent with any distribution that could arise by honest stakers being offline or timing out (for a natural stylized model of honest timeouts).

cs.GT↗
arXiv2024

Bundling in Oligopoly: Revenue Maximization with Single-Item Competitors

We consider a principal seller with $m$ heterogeneous products to sell to an additive buyer over independent items. The principal can offer an arbitrary menu of product bundles, but faces competition from smaller and more agile single-item sellers. The single-item sellers choose their prices after the principal commits to a menu, potentially under-cutting the principal's offerings. We explore to what extent the principal can leverage the ability to bundle product together to extract revenue. Any choice of menu by the principal induces an oligopoly pricing game between the single-item sellers, which may have multiple equilibria. When there is only a single item this model reduces to Bertrand competition, for which the principal's revenue is $0$ at any equilibrium, so we assume that no single item's value is too dominant. We establish an upper bound on the principal's optimal revenue at every equilibrium: the expected welfare after truncating each item's value to its revenue-maximizing price. Under a technical condition on the value distributions -- that the monopolist's revenue is sufficiently sensitive to price -- we show that the principal seller can simply price the grand-bundle and ensure (in any equilibrium) a constant approximation to this bound (and hence to the optimal revenue). We also show that for some value distributions violating our conditions, grand-bundle pricing does not yield a constant approximation to the optimal revenue in any equilibrium.

cs.GT↗
arXiv2023

Optimal Stopping with Multi-Dimensional Comparative Loss Aversion

Despite having the same basic prophet inequality setup and model of loss aversion, conclusions in our multi-dimensional model differs considerably from the one-dimensional model of Kleinberg et al. For example, Kleinberg et al. gives a tight closed-form on the competitive ratio that an online decision-maker can achieve as a function of $λ$, for any $λ\geq 0$. In our multi-dimensional model, there is a sharp phase transition: if $k$ denotes the number of dimensions, then when $λ\cdot (k-1) \geq 1$, no non-trivial competitive ratio is possible. On the other hand, when $λ\cdot (k-1) < 1$, we give a tight bound on the achievable competitive ratio (similar to Kleinberg et al.). As another example, Kleinberg et al. uncovers an exponential improvement in their competitive ratio for the random-order vs. worst-case prophet inequality problem. In our model with $k\geq 2$ dimensions, the gap is at most a constant-factor. We uncover several additional key differences in the multi- and single-dimensional models.

cs.GT↗
arXiv2023

Recent Developments in Pandora's Box Problem: Variants and Applications

In 1979, Weitzman introduced Pandora's box problem as a framework for sequential search with costly inspections. Recently, there has been a surge of interest in Pandora's box problem, particularly among researchers working at the intersection of economics and computation. This survey provides an overview of the recent literature on Pandora's box problem, including its latest extensions and applications in areas such as market design, decision theory, and machine learning.

cs.GT↗
arXiv2023

Selling to Multiple No-Regret Buyers

We consider the problem of repeatedly auctioning a single item to multiple i.i.d buyers who each use a no-regret learning algorithm to bid over time. In particular, we study the seller's optimal revenue, if they know that the buyers are no-regret learners (but only that their behavior satisfies some no-regret property -- they do not know the precise algorithm/heuristic used). Our main result designs an auction that extracts revenue equal to the full expected welfare whenever the buyers are "mean-based" (a property satisfied by standard no-regret learning algorithms such as Multiplicative Weights, Follow-the-Perturbed-Leader, etc.). This extends a main result of [BMSW18] which held only for a single buyer. Our other results consider the case when buyers are mean-based but never overbid. On this front, [BMSW18] provides a simple LP formulation for the revenue-maximizing auction for a single-buyer. We identify several formal barriers to extending this approach to multiple buyers.

cs.GT↗
arXiv2023

Baechi: Fast Device Placement of Machine Learning Graphs

Machine Learning graphs (or models) can be challenging or impossible to train when either devices have limited memory, or models are large. To split the model across devices, learning-based approaches are still popular. While these result in model placements that train fast on data (i.e., low step times), learning-based model-parallelism is time-consuming, taking many hours or days to create a placement plan of operators on devices. We present the Baechi system, the first to adopt an algorithmic approach to the placement problem for running machine learning training graphs on small clusters of memory-constrained devices. We integrate our implementation of Baechi into two popular open-source learning frameworks: TensorFlow and PyTorch. Our experimental results using GPUs show that: (i) Baechi generates placement plans 654 X - 206K X faster than state-of-the-art learning-based approaches, and (ii) Baechi-placed model's step (training) time is comparable to expert placements in PyTorch, and only up to 6.2% worse than expert placements in TensorFlow. We prove mathematically that our two algorithms are within a constant factor of the optimal. Our work shows that compared to learning-based approaches, algorithmic approaches can face different challenges for adaptation to Machine learning systems, but also they offer proven bounds, and significant performance benefits.

cs.DC↗
arXiv2022

Pandora's Problem with Nonobligatory Inspection: Optimal Structure and a PTAS

Weitzman introduced Pandora's box problem as a mathematical model of sequential search with inspection costs, in which a searcher is allowed to select a prize from one of $n$ alternatives. Several decades later, Doval introduced a close version of the problem, where the searcher does not need to incur the inspection cost of an alternative, and can select it uninspected. Unlike the original problem, the optimal solution to the nonobligatory inspection variant is proved to need adaptivity, and by recent work of [FLL22], finding the optimal solution is NP-hard. Our first main result is a structural characterization of the optimal policy: We show there exists an optimal policy that follows only two different pre-determined orders of inspection, and transitions from one to the other at most once. Our second main result is a polynomial time approximation scheme (PTAS). Our proof involves a novel reduction to a framework developed by [FLX18], utilizing our optimal two-phase structure. Furthermore, we show Pandora's problem with nonobligatory inspection belongs to class NP, which by using the hardness result of [FLL22], settles the computational complexity class of the problem. Finally, we provide a tight 0.8 approximation and a novel proof for committing policies [BK19] (informally, the set of nonadaptive policies) for general classes of distributions, which was previously shown only for discrete and finite distributions [GMS08].

cs.DS↗
arXiv2021

99% Revenue with Constant Enhanced Competition

The enhanced competition paradigm is an attempt at bridging the gap between simple and optimal auctions. In this line of work, given an auction setting with $m$ items and $n$ bidders, the goal is to find the smallest $n' \geq n$ such that selling the items to $n'$ bidders through a simple auction generates (almost) the same revenue as the optimal auction. Recently, Feldman, Friedler, and Rubinstein [EC, 2018] showed that an arbitrarily large constant fraction of the optimal revenue from selling $m$ items to a single bidder can be obtained via simple auctions with a constant number of bidders. However, their techniques break down even for two bidders, and can only show a bound of $n' = n \cdot O(\log \frac{m}{n})$. Our main result is that $n' = O(n)$ bidders suffice for all values of $m$ and $n$. That is, we show that, for all $m$ and $n$, an arbitrarily large constant fraction of the optimal revenue from selling $m$ items to $n$ bidders can be obtained via simple auctions with $O(n)$ bidders. Moreover, when the items are regular, we can achieve the same result through auctions that are prior-independent, {\em i.e.}, they do not depend on the distribution from which the bidders' valuations are sampled.

cs.GT↗
arXiv2021

The Short-Side Advantage in Random Matching Markets

A breakthrough of Ashlagi, Kanoria, and Leshno [AKL17] found that imbalance in the number of agents on either side of a random matching market has a profound effect on the market's expected characteristics. Specifically, across all stable matchings, the "long side" (i.e. the side with a greater number of agents) receives significantly worse matches in expectation than the short side. Intuitively, this occurs because an agent on the long side is essentially unneeded to create a stable matching -- a matching could form almost as easily without them. Thus, an agent on the long side has very little market power, and must settle for a match which is not much better than a random assignment. We provide a new and simpler proof for a result of [AKL17] which formalizes this intuition.

cs.GT↗
arXiv2020

Implementation in Advised Strategies: Welfare Guarantees from Posted-Price Mechanisms when Demand Queries are NP-hard

State-of-the-art posted-price mechanisms for submodular bidders with $m$ items achieve approximation guarantees of $O((\log \log m)^3)$ [Assadi and Singla, 2019]. Their truthfulness, however, requires bidders to compute an NP-hard demand-query. Some computational complexity of this form is unavoidable, as it is NP-hard for truthful mechanisms to guarantee even an $m^{1/2-\varepsilon}$-approximation for any $\varepsilon > 0$ [Dobzinski and Vondrák, 2016]. Together, these establish a stark distinction between computationally-efficient and communication-efficient truthful mechanisms. We show that this distinction disappears with a mild relaxation of truthfulness, which we term implementation in advised strategies, and that has been previously studied in relation to "Implementation in Undominated Strategies" [Babaioff et al, 2009]. Specifically, advice maps a tentative strategy either to that same strategy itself, or one that dominates it. We say that a player follows advice as long as they never play actions which are dominated by advice. A poly-time mechanism guarantees an $α$-approximation in implementation in advised strategies if there exists poly-time advice for each player such that an $α$-approximation is achieved whenever all players follow advice. Using an appropriate bicriterion notion of approximate demand queries (which can be computed in poly-time), we establish that (a slight modification of) the [Assadi and Singla, 2019] mechanism achieves the same $O((\log \log m)^3)$-approximation in implementation in advised strategies.

cs.GT↗
arXiv2019

Representing All Stable Matchings by Walking a Maximal Chain

The seminal book of Gusfield and Irving [GI89] provides a compact and algorithmically useful way to represent the collection of stable matches corresponding to a given set of preferences. In this paper, we reinterpret the main results of [GI89], giving a new proof of the characterization which is able to bypass a lot of the "theory building" of the original works. We also provide a streamlined and efficient way to compute this representation. Our proofs and algorithms emphasize the connection to well-known properties of the deferred acceptance algorithm.

cs.GT↗