Diminishing Utility

Diminishing marginal utility means that an additional unit brings a smaller increase in utility when someone already has more. Utility represents how outcomes are valued in a decision model. If the first serving relieves hunger and the next adds less satisfaction, total satisfaction can still rise even as the extra benefit declines. “Diminishing” describes the increment, not necessarily the total.

Total utility and the next unit

For a differentiable utility function \(U(x)\), marginal utility is \(U\prime(x)\). Positive marginal utility means that more is preferred; a declining marginal utility means the curve is concave. The familiar smooth, strictly diminishing case has \(U\prime(x)>0\) and \(U\prime\prime(x)<0\). Weak concavity also permits flat portions of the marginal-utility curve. If an extra serving actually makes someone worse off, marginal utility has become negative, a stronger claim than diminishing positive benefits.

Consider the illustrative model \(U(x)=\log x\) for positive wealth. Its slope is \(1/x\), so the extra utility of another dollar declines as wealth rises. Increasing wealth from 1,000 to 2,000 adds \(\log 2\), about 0.693, while increasing it from 50,000 to 51,000 adds \(\log(1.02)\), about 0.0198. These are implications of an assumed preference model, not measured amounts of happiness or comparisons between two people.

Why curvature matters under uncertainty

Expected utility evaluates an uncertain choice by weighting the utility of each outcome by its probability. That differs from calculating the average monetary outcome and then taking its utility. A curved function makes the order of those operations matter.

Suppose a person with log utility faces a 50–50 chance of ending with 5,000 or 15,000. Expected wealth is 10,000. Expected utility is half log(5,000) plus half log(15,000). To find the guaranteed wealth with the same utility, exponentiate that average: the certainty equivalent is the square root of 5,000 × 15,000, about 8,660. Under this model, a guaranteed 9,000 is preferable to the gamble even though it has lower wealth than the gamble’s average.

The difference between expected wealth and the certainty equivalent is the risk premium: about 1,340 in this example. The low outcome reduces utility more than the equally distant high outcome increases it. More generally, concavity gives U(E[X]) ≥ E[U(X)], so an expected-utility decision maker with an increasing concave utility function weakly prefers a guaranteed expected outcome to the gamble. A linear utility function produces equality; convex utility reverses the preference where the relevant expectations exist.

This connection to risk aversion depends on the expected-utility framework. It does not establish that everyone uses logarithmic utility, or that a fitted curve captures every choice. A mean-preserving spread reduces expected utility for a concave function, but a larger variance alone does not order every pair of outcome distributions. The full distribution and the assumed utility function matter. The calculated risk premium is therefore model-specific, not a universal insurance price.

Related ideas in machine learning

Diminishing returns in a learning curve concern performance gained from additional resources, while marginal utility concerns the value assigned to outcomes. The two may be connected through a decision objective, but they are not interchangeable. Additional training data can yield smaller accuracy improvements; whether those improvements are worth buying depends on their value and the cost of collecting the data. Learning curves can also have irregularities or thresholds, so diminishing gains should be checked rather than assumed.

Similarly, an additional model in an ensemble may contribute little if it repeats existing errors. A model that makes different errors can contribute much more. Features can complement one another, so the next feature’s benefit need not be smaller than the last one’s. Compare incremental benefit under the actual combination of inputs and models.

Ranking metrics use a related but distinct construction. DCG weights a result at rank i by 1/log2(i+1), so equally relevant results contribute less at later ranks. The position discount is decreasing and convex for positive i; the cumulative gain from adding equally relevant results has diminishing increments. Calling the discount itself an increasing concave utility function would confuse the weight with the total it helps construct.

A decision objective should express the consequences that matter. Squared error represents a symmetric convex penalty on prediction error, not diminishing utility of wealth. With calibrated binary probabilities and only constant false-positive and false-negative costs, choose the positive action when its expected cost is lower: the threshold is C_FP/(C_FP + C_FN), assuming positive total cost. Equal costs give 0.5. If consequences change, revisit the threshold; if probabilities become unreliable, repair their estimation or calibration as well. The utility or cost model explains the desired decision, while the probability model supplies evidence about possible outcomes.

References

• Stanford Encyclopedia of Philosophy, Normative Theories of Rational Choice: Expected Utility