In standard auction theory, the golden rule for a buyer in a second-price auction is simple: bid your true utility. If an advertising impression is worth $0.05 to you, you should bid exactly $0.05. Bidding higher risks paying more than the item is worth; bidding lower simply reduces your chances of winning a profitable item. This elegant result relies on the assumption of quasi-linear utility—the idea that your satisfaction is simply the value of the item minus the money you spend, with no strings attached.
In the real world of digital advertising, federated learning marketplaces, and mobile marketing, this assumption falls apart. Bidders rarely operate with infinite wallets or isolated decisions. Instead, they are bound by global, long-term constraints: a hard daily budget, a target Cost-Per-Click (CPC), a maximum average Cost-Per-Action (CPA), or a strict Return-on-Spend (ROI) requirement.
When these constraints are introduced, truthful bidding is no longer optimal. If you have a tight budget and bid your true value on early items, you might exhaust your funds on mediocre opportunities, leaving nothing for highly lucrative items that appear later.
Remarkably, across a wide variety of constraints, the mathematically optimal response to this complex, multi-round optimization problem collapses into an incredibly elegant strategy: linear bid scaling. Instead of bidding your true value $v$, you bid a scaled version $\alpha \cdot v$, where $\alpha$ is a uniform multiplier adjusted dynamically over time.
The Intuition of the Shadow Price
To understand why a simple linear multiplier is the optimal way to handle global constraints, it helps to look at the problem through the lens of constrained optimization.
Suppose an automated agent wants to maximize the total value of the items it wins over a campaign lifetime, subject to a total budget constraint. This can be framed as a packing problem. If the agent knew every future opportunity and its market price in advance, it could solve this globally. Because it operates in real time, it must make split-second decisions on an item-by-item basis.
By applying the Lagrangian multiplier method, as detailed by Jun Wang 0012 et al. in their comprehensive overview of real-time bidding systems, Display Advertising with Real-Time Bidding (RTB) and Behavioural Targeting, we can transform this constrained problem into an unconstrained one. The budget constraint is replaced by a penalty term proportional to the cost of the item. The agent's objective for a single auction changes from maximizing net value ($v - \text{price}$) to maximizing a modified utility:
$$\text{Utility} = v - \lambda \cdot \text{price}$$
Here, $\lambda$ represents the shadow price of the budget—the marginal value of an extra dollar of budget. If your budget is overflowing, an extra dollar is worth nothing to you, so $\lambda = 0$, and you bid your true value. If your budget is extremely tight, an extra dollar is highly valuable, so $\lambda$ becomes large, heavily penalizing expensive auctions.
To maximize this modified utility in a second-price auction, the agent should choose to win if and only if $v - \lambda \cdot \text{price} > 0$, which algebraically rearranges to:
$$\frac{1}{\lambda} \cdot v > \text{price}$$
Because a bidder wins a second-price auction whenever their bid exceeds the market price, the agent can perfectly implement this optimal decision rule by submitting a bid equal to:
$$\text{Bid} = \alpha \cdot v \quad \text{where} \quad \alpha = \frac{1}{\lambda}$$
This is why linear scaling works. The multiplier $\alpha$ acts as a global regulator that translates a long-term, macro-level constraint into an instantaneous, micro-level bidding decision.
[Global Constraint: Budget / ROI]
│
▼ (Lagrangian Optimization)
[Shadow Price (λ)]
│
▼ (Inversion: α = 1/λ)
[Linear Multiplier (α)]
│
▼ (Per-Query Scaling)
[Final Bid = α × Value]
Adapting to Diverse Constraints
The beauty of value-proportional scaling is its versatility. Different business objectives alter the definition of the shadow price, but the core mechanism remains a linear adjustment of the valuation.
Cost-Per-Action (CPA) Constraints
In many campaigns, advertisers care less about total budget depletion and more about efficiency, such as maintaining a target average cost per conversion. In Cost per action constrained auctions, Benjamin Heymann proves that when bidders are bound by an average CPA target across a sequence of second-price auctions, the optimal strategy shifts away from truthful bidding to a linear multiplier of the item's private value. The multiplier is jointly determined by the target cost and the expected level of market competition, allowing the bidder to balance high-cost, high-value acquisitions with cheaper, lower-value wins to hit the average target.
Dual Constraints (Budget + CPC)
Real-world systems often have to balance multiple constraints simultaneously. For instance, an advertiser might want to maximize conversions while staying under a total budget and keeping the average Cost-Per-Click (CPC) below a specific threshold. Xun Yang et al. tackle this in Bid Optimization by Multivariable Control in Display Advertising. By deriving a bidding mechanism from a primal-dual optimization framework that accounts for both boundaries, the authors show that the optimal bid becomes a weighted combination of the predicted value and the target cost, ensuring the system filters traffic effectively without violating either constraint. To refine this traffic filtering further, Bo Yang et al. demonstrate in Lightweight Auto-bidding based on Traffic Prediction in Live Advertising that incorporating both upper and lower bounds for CPC allows auto-bidding algorithms to actively filter out low-quality traffic while preserving budget compliance in high-frequency live streaming environments.
Return-on-Spend (ROI) Constraints
When the goal is to maintain a specific ratio of revenue to expenditure, agents can use online learning to tune their scaling factors. In On the Coordination of Value-Maximizing Bidders, Yanru Guan et al. utilize mirror descent algorithms to dynamically adjust a bidding multiplier based on historical performance. This approach allows agents to maximize total conversion value while strictly adhering to a predefined return-on-spend constraint across multiple auction rounds.
Why Simplicity Wins: Linear vs. Non-Linear Strategies
Given the complexity of modern ad exchanges, one might wonder if a non-linear bidding function—such as a deep neural network mapping impression features directly to a bid—would perform better. However, both theoretical and empirical evidence favor linear scaling.
First, linear functions are remarkably robust against market volatility. Tengyun Wang et al. demonstrate in A Revenue-Maximizing Bidding Strategy for Demand-Side Platforms that linear bidding functions frequently outperform non-linear alternatives in generalized second-price auctions. Simple proportional bidding remains highly effective because it avoids overfitting to transient market fluctuations while directly optimizing the core value-to-cost ratio.
Second, linear scaling is an inherent behavioral equilibrium for agents facing financial friction. In Non-quasi-linear Agents in Quasi-linear Mechanisms, Moshe Babaioff et al. study what happens when participants have non-linear payment costs, such as those arising from strict budgets or varying opportunity costs of capital. The authors prove that in standard auctions designed for linear agents, these non-linear participants can optimize their utility by uniformly scaling their entire valuation profile by a single scalar factor. This uniform shading strategy simplifies a highly complex, non-linear decision-making process into a single-parameter optimization problem, while still guaranteeing at least 50% of the maximum possible welfare and revenue for the auctioneer.
Beyond Advertising: Broad Applications
The mathematical principle of using linear scaling and dual variables to manage resource constraints extends far beyond traditional display advertising.
- Mobile Payment Marketing: In Joint Incentive Optimization of Customer and Merchant in Mobile Payment Marketing, Li Yu et al. formulate real-time budget allocation for digital coupons and merchant commissions as a constrained optimization problem. By using the Lagrangian multiplier method to derive shadow prices, they enable millisecond-level decision-making for high-volume requests, balancing global marketing budgets with local incentive decisions.
- Federated Learning Marketplaces: In competitive machine learning environments where multiple data consumers compete for data owners, Han Yu et al. introduce a framework called Fed-Bidder in Utility-Maximizing Bidding Strategy for Data Consumers in Auction-based Federated Learning. This framework employs Lagrangian multipliers to derive optimal bidding values, balancing expected data utility against financial limits to provide closed-form solutions for data acquisition.
- Mechanism Design and Procurement: From the perspective of the auction designer, understanding linear scaling allows for the creation of simpler mechanisms. Alon Eden et al. show in From Welfare to Utility: Generalized Objectives in Budget-Feasible Procurement that when budget constraints only need to be met in expectation, using virtual costs and Lagrangian multipliers transforms constrained procurement problems into unconstrained ones by penalizing budget violations. Similarly, Shuchi Chawla et al. in Buy-Many Mechanisms for Many Unit-Demand Buyers utilize a Lagrangian profit transformation to convert revenue maximization under supply constraints into a profit maximization framework, assigning virtual production costs to items to simplify pricing mechanisms.
A Unified Paradigm
As the digital advertising ecosystem has transitioned from manual keyword bidding to fully automated systems, linear bid scaling has emerged as the foundational architecture for modern auto-bidding engines.
As synthesized by Gagan Aggarwal et al. in their comprehensive survey, Auto-bidding and Auctions in Online Advertising: A Survey, optimal bidding strategies for automated agents consistently converge to a value-proportional formula. Whether an advertiser is optimizing for clicks, revenue, or conversions, the core engine operates by scaling the estimated value of an impression by a factor derived from the dual variables of their constraints.
By separating the estimation of value (handled by machine learning models predicting user behavior) from the enforcement of constraints (handled by adjusting the linear multiplier $\alpha$), linear bid scaling provides an elegant, robust, and mathematically optimal bridge between high-level business goals and real-time execution.
Key Foundational Work
Cost per action constrained auctions (2019) — contributed 3 concept elements to this direction.
Bid Optimization by Multivariable Control in Display Advertising (2019) — contributed 2 concept elements to this direction.
Display Advertising with Real-Time Bidding (RTB) and Behavioural Targeting (2016) — contributed 1 concept elements to this direction.