Online sellers often hide part of their stock or tag items "only 3 left" because a visible shortage makes shoppers want the item more. This is artificial scarcity: limiting what customers can see in order to raise demand. The same sellers choose which products to display using greedy algorithms, which add one item at a time and keep whichever addition raises expected revenue most. Those algorithms carry worst-case guarantees, such as always reaching at least half of the best possible revenue, but only when revenue has diminishing returns in a suitable sense. No existing guarantee covers the case where the seller also picks how many units of each item to show and fewer visible units make the item more attractive. Scarcity signaling pushes against diminishing returns in two ways. A scarcer item draws more buyers, and its sales stop abruptly when the visible units run out, at which point demand spills over to the other items.
This note proposes to treat "which items, and how much visible stock of each" as one combinatorial selection problem. The central question is a threshold. As the scarcity effect grows stronger, at what point does the revenue function stop having the ordered diminishing-returns structure that greedy methods need? Below that point we want a greedy algorithm with a proven constant-factor guarantee. Above it we want a counterexample, and an empirical measurement of how badly greedy actually does. If the threshold sits well above the scarcity effects observed in real markets, practitioners can keep using fast greedy methods with a guarantee. If it sits below them, the greedy methods in use today are running without one.
Core Claim
Lift the problem to a ground set of pairs $(i, q)$, meaning "display item $i$ with $q$ visible units," where $q$ is drawn from a short ladder of display levels, the top level being "show everything, no badge." A feasible display picks at most one level per item and at most $k$ items. Customers choose by a multinomial logit (MNL) model, the standard choice model in which each option's purchase probability is proportional to its attractiveness score. The display level enters that score:
$$P(i \mid S) = \frac{v_i(q_i)}{1 + \sum_{(j,q_j) \in S} v_j(q_j)}, \qquad v_i(q) = \exp\big(\theta^\top x_i + \delta\, s(q)\big).$$
Here $S$ is the chosen set of pairs, $x_i$ are item features, $\theta$ are preference weights, $s(q)$ is a decreasing scarcity signal that equals zero at the top level, and $\delta \ge 0$ measures how strongly visible scarcity raises attractiveness. Over a season of customer arrivals, displayed units deplete. Once a unit cap $q_i$ is hit, item $i$ disappears and its would-be buyers re-choose among what remains. Expected revenue $R(S)$ sums each item's price $p_i$ times its expected units sold, which can never exceed $q_i$.
The tool we build on is the idea of submodular order functions from Submodular Order Functions and Assortment Optimization. That paper shows that MNL-type revenue, which lacks ordinary diminishing returns, still has diminishing returns along a fixed ordering of the products, typically by descending price. Along such an ordering, a threshold-based greedy method achieves a $0.5-\epsilon$ approximation under cardinality constraints, and a local search achieves $0.25$ under matroid constraints. A matroid is a general family of "pick at most this many from each group" constraints, and "one level per item" is one. The same paper proves that no polynomial-query algorithm can beat $0.5$ for this function class.
Our conjecture is that, in a fluid model where demand is replaced by its expected flow, $R$ remains a submodular order function under a lexicographic order: price descending, then display level from most to least visible. This should hold whenever $\delta$ is below a threshold $\delta^\star$ that depends on the ratio of adjacent prices and on how sharply sales are capped. Above $\delta^\star$, a cheap item with a scarcity badge can steal enough demand from an expensive item placed earlier in the order to invert marginal gains, and the order breaks. If the conjecture holds, the matroid guarantee of 0.25 transfers directly and the half-approximation transfers to the pure display-cap setting. If it fails, we want the explicit counterexample and a measurement of how fast greedy degrades.
Grounding in Prior Work
The behavioral premise comes from operations research. Dynamic Pricing and Inventory Management Under Inventory-Dependent Demand derives optimal "order-up-to / dispose-down-to / display-up-to" policies for a single product whose demand falls as visible stock rises. It shows that withholding stock supports higher prices, and it flags as open the case where scarcity signaling and stock-display effects act together. We take the display-up-to lever from this paper. We change the setting from one product solved by dynamic programming to a multi-product assortment, where dynamic programming does not scale and substitution between items matters.
The size of the scarcity effect comes from Daily Deals: Prediction, Social Diffusion, and Reputational Ramifications. That study regressed over 16,000 deals and found that a limited-quantity tag raised sales by 11% even when deals did not sell out. We use this figure to calibrate $\delta$, which gives the threshold $\delta^\star$ a real-world reference point.
Two assortment papers supply methods but stop short of this problem. Diversified Multinomial Logit Contextual Bandits inserts a diversity bonus into MNL choice and proves a greedy ratio of $1-1/(e+1)\approx 0.731$. That proof requires uniform revenues and unlimited stock. Our problem has neither, so that ratio should not be expected here. Combinatorial Assortment Optimization shows that offering the full catalog is a 4-approximation when items are well-priced, but its valuations are fixed and do not respond to scarcity.
The cautionary result is GKD-Recruiter. It shows that capped demand, its "saturation trap," breaks submodularity in crowdsourcing, and that greedy methods such as CELF then underperform. Its response is to drop guarantees in favor of reinforcement learning. Our sell-out caps create the same trap. We respond differently, by asking whether an ordering restores enough structure to keep a guarantee.
Research Program
The theoretical work has three parts. The first is to prove the submodular-order property of $R$ in the fluid model for $\delta \le \delta^\star$, with an explicit formula for $\delta^\star$. The second is to construct a matching instance that violates the property just above $\delta^\star$. The third is to bound the gap between the fluid model and the stochastic model, where arrivals are random. That gap should shrink like $1/\sqrt{\text{arrivals}}$ by standard concentration of sales counts around their expected flow.
On the algorithmic side, we will adapt Udwani's threshold-based greedy to the lifted pair ground set and add a cheap certificate. While the algorithm runs, it checks the ordered diminishing-returns inequality on every pair it evaluates. Each run therefore reports whether its guarantee actually held on that instance.
There is also a secondary question, motivated by The Groupon Effect on Yelp Ratings: A Root Cause Analysis. That study found that deal customers rated merchants 3.27 versus 3.73 stars, which it attributes to poorly matched buyers. Scarcity-driven demand may carry an endogenous mismatch cost. We will record each prefix of the greedy run, following the single-pass frontier construction of Computing Approximate Pareto Frontiers for Submodular Utility and Cost Tradeoffs, to trace the trade-off between revenue and expected mismatch. That paper assumes fixed, known costs. Here the cost depends on which set is chosen, so we report the trade-off curve empirically and do not claim a bicriteria guarantee for it.
Baselines
The naive reference point is to display the full catalog at full visibility with no scarcity badges. This is what a seller does without any optimization, and it is the setting in which Combinatorial Assortment Optimization proves a 4-approximation for well-priced items. Every method must beat it on expected revenue at matched $\delta$.
The strongest simple competitor is a revenue-ordered heuristic. It tries the top-$m$ items by price for every $m$, which is the classical optimal structure for uncapacitated MNL. For each item it then picks the display level independently, choosing whichever level maximizes that item's revenue in isolation. This is the main threat, because revenue-ordered sets are optimal when $\delta=0$ and caps do not bind.
The second strong competitor is plain lazy greedy (CELF) over the pairs, with no ordering. GKD-Recruiter reports that saturation-aware methods beat CELF by 21.41% on task satisfaction, so we expect it to suffer under binding caps.
We include one ablation that isolates the key ingredient: Udwani's threshold greedy run on the scarcity-blind model with $\delta=0$, whose chosen set is then evaluated in the true simulator. This tests whether modeling scarcity matters at all.
For small instances, roughly 12 items and 4 display levels, exhaustive enumeration gives the exact optimum. All methods share the same simulator, the same fitted $\theta$, and the same budget of revenue-oracle calls. Where tuning applies, such as threshold $\epsilon$ for ours and the level grid for the revenue-ordered heuristic, each method gets the same number of trials. Beating a baseline means a higher mean ratio to optimum, measured on matched instances with paired confidence intervals.
No reported result was found on Lacuna for revenue under scarcity-dependent display assortment, so there is no published number to beat. The relevant reported anchors are theoretical: $0.5-\epsilon$ and $0.25$ from Udwani, and $0.731$ from the bandit paper. The one reported empirical anchor is the 11% scarcity lift, which we use for calibration rather than as a target.
Why not just…?
Why not just use revenue-ordered assortments and set each item's display level separately? This is the real threat. When caps rarely bind and $\delta$ is small, it may match our method, because it amounts to our ordering with no interaction between items. Our hypothesis is that it fails exactly when a badge on a cheap item pulls demand away from a pricier one, an effect that per-item level selection cannot see. The $\delta$ sweep will settle this. If the revenue-ordered heuristic stays within 2% of optimum across the calibrated range of $\delta$, we will report that it is sufficient there, and the contribution narrows to the threshold theorem that explains why.
Why not just run CELF greedy over item–level pairs? It is fast, but it has no guarantee once sell-out caps make the objective non-submodular. Optimal Skill Selection for LLM Agents with Provable Bicriteria Guarantees finds that plain greedy hits the exact optimum in only about 44% of cases on its own capped problem, compared with 100% for prefix evaluation. Comparing CELF against exhaustive optima on small instances will show whether the same gap appears here.
Why not just train a reinforcement-learning policy, as GKD-Recruiter does? The reported runtime and quality gains of that approach are real. But it gives no worst-case bound, and it would need retraining whenever prices or $\delta$ change. We include a DQN-style learned policy on the semi-synthetic instances under the same oracle budget. Even if it matches us on average, the certificate tells sellers which instances they can trust.
Proposed Experiments
The first experiment uses synthetic MNL instances, generated in the style of the offline simulations in PASTA: Pessimistic Assortment Optimization. Instances have 12 items, so exact optima are computable, and $\delta$ is swept from 0 to well past the calibrated value. The primary metric is the worst-case and mean ratio of each method's revenue to the optimum. The decisive check is whether the measured breakpoint in greedy's worst-case ratio sits near the derived $\delta^\star$, and whether the run-time certificate fails on the instances where the ratio drops.
The second experiment is semi-synthetic. It uses item features built from the Rotten Tomatoes review vectors used by the diversified MNL bandit paper, with up to 200 items and $\delta$ calibrated so that a badge produces roughly an 11% sales lift. Here, with no exact optimum available, we compare methods against a best-of-many-local-search upper reference and report revenue, oracle calls, and wall-clock time.
The third experiment traces the revenue–mismatch curve from recorded greedy prefixes. A mismatched buyer is one drawn into a purchase only by the badge term $\delta s(q)$. We ask whether the frontier shows a knee, a point where the last increments of revenue come almost entirely from such buyers.
The proposal is wrong in an informative way if the scarcity order fails even at small $\delta$. That would mean the greedy guarantees cited in assortment practice do not survive the scarcity levers sellers actually use.
References
- Submodular Order Functions and Assortment Optimization
- Dynamic Pricing and Inventory Management Under Inventory-Dependent Demand
- Daily deals: prediction, social diffusion, and reputational ramifications
- Diversified Multinomial Logit Contextual Bandits
- Combinatorial Assortment Optimization
- GKD-Recruiter: Jointly Modeling Social and Task Heterogeneity for Spatial Crowdsourcing via Graph Knowledge Distillation
- The groupon effect on yelp ratings: a root cause analysis
- Computing Approximate Pareto Frontiers for Submodular Utility and Cost Tradeoffs
- Optimal Skill Selection for LLM Agents with Provable Bicriteria Guarantees
- PASTA: Pessimistic Assortment Optimization