In multi-agent systems, the most fundamental hurdle to collective decision-making is the problem of interpersonal utility comparison. If one agent reports a utility of 1,000 for a resource and another reports a utility of 10, it is impossible to know if the first agent truly wants it 100 times more, or if they are simply measuring their "happiness" on a different scale.
To solve this, researchers use utility normalization—a set of mathematical constraints that force diverse, subjective preferences into a shared, objective framework. By scaling individual valuations to a common baseline, we can design algorithms that are fair, efficient, and resistant to strategic manipulation.
The Normalization Toolbox
The core idea of normalization is to transform an agent's raw valuation function into a standardized representation. Several distinct "baselines" appear across the literature:
- Unit-Sum Normalization: Every agent’s total valuation for the entire set of items is scaled to sum to a fixed constant (usually 1). This "fixed budget" approach, explored by George Christodoulou and Vasilis Gkatzelis, ensures that no single agent can dominate the social welfare calculation by reporting arbitrarily high values.
- Sole-User Baseline: An agent’s satisfaction is measured relative to the maximum utility they could achieve if they had exclusive access to the resource. Michel Lemaître and Gérard Verfaillie used this to manage high-stakes satellite scheduling, ensuring that multi-agent compromises remain principled.
- Maximin Share (MMS) Calibration: In the fair division of indivisible goods, Hannaneh Akrami and Jugal Garg use normalization to set an agent's "fair share" to a fixed value (like 1). This simplifies the search for allocations where everyone receives at least a specific fraction of their proportional value.
Why Normalization Works: Strategic Robustness
Without normalization, resource allocation systems are vulnerable to utility inflation. A rational agent has every incentive to report that they value an item at "one billion" to ensure they receive it.
As Haris Aziz argues, enforcing a fixed sum for utilities prevents this strategic behavior. It equalizes the "voting power" of each participant. If an agent wants to report a very high value for one item, they must necessarily report lower values for others to stay within their normalized budget. This forces agents to reveal their true relative preferences rather than just their absolute greed.
Achieving Scale Invariance
A robust fairness mechanism should be scale-invariant: the final allocation should not change if an agent decides to measure their utility in dollars instead of cents.
Minming Li and Zhiyi Huang highlight that Nash Social Welfare (NSW)—the geometric mean of utilities—is naturally scale-invariant. Because the product of utilities is maximized, multiplying one agent's values by a constant doesn't change the optimal allocation. For other welfare functions, explicit normalization is required to achieve this same stability, ensuring that the system is not sensitive to the arbitrary units agents use to express their desires.
Overcoming the "Online" Impossibility
Normalization is particularly powerful in online settings, where items arrive one by one and must be assigned immediately. Normally, if you don't know what items are coming in the future, it is impossible to guarantee fairness.
However, Vasilis Gkatzelis and Xizhi Tan showed that if agent valuations are normalized (summing to 1), we can achieve a "best-of-both-worlds" result. They developed an algorithm that guarantees every agent their fair share while capturing 91.6% of the maximum possible social welfare. The normalization provides the algorithm with a "global" benchmark—even if it doesn't know the future items, it knows the total "value" each agent is entitled to, allowing it to set protective thresholds.
The Price of Fairness
Finally, normalization allows us to quantify the Price of Fairness—the efficiency we lose by insisting on equity. Nisarg Shah and Umang Bhaskar demonstrated that when valuations are normalized, the welfare loss is bounded by a factor of $\sqrt{n}$ (where $n$ is the number of agents). Without normalization, the discrepancy between how agents value the total resource pool can lead to a much higher, linear reduction in total social welfare.
By standardizing the "size" of each agent's preferences, normalization makes the trade-off between fairness and efficiency predictable and mathematically manageable.