An inequality that states for any convex function φ and any random variable X (or any probability measure), φ(E[X]) ≤ E[φ(X)], provided the expectations exist. For concave φ the inequality direction reverses. It formalizes how nonlinear convex transformations interact with averaging.

Principle

Convexity forces the function of an average to lie below (or equal to) the average of the function; equivalently, averaging and convex nonlinearities do not commute and the expectation of a convex transform overdisperses relative to the transform at the mean.

Demonstration

Illustrative scenario: For a scalar random variable X with finite second moment and φ(x)=x^2 (convex), Jensen gives (E[X])^2 = φ(E[X]) ≤ E[φ(X)] = E[X^2], recovering the elementary variance lower bound Var(X) = E[X^2] − (E[X])^2 ≥ 0. For φ=log (concave), the inequality reverses: log E[X] ≥ E[log X].

Misapplication

Applying Jensen's inequality to a nonconvex function or neglecting integrability conditions (expectations not finite) or confusing the inequality direction for concave versus convex φ are common errors; another mistake is assuming equality holds except in trivial cases without checking whether φ is affine on the support or X is almost surely constant.

Consequence

Provides a systematic tool to derive bounds and approximations for expectations, moments, and information‑theoretic quantities (e.g., convexity of exponential or concavity of log), and underpins many inequalities and risk bounds in estimation and learning.

Reversal

If φ is concave the inequality direction is reversed; if expectations do not exist (infinite moments) or the measure is not probability, the basic statement needs modification (for example, using extended real values or conditional forms).

Boundary

Requires φ to be convex (or concave for the reversed form) on the relevant domain and E[|φ(X)|], E[|X|] to be finite as needed; does not apply directly when these integrability conditions fail or when X is not defined on a probability space.

Semantic Tension

Averaging (expectation) ↔ Nonlinearity (convex/concave transforms): Jensen quantifies the discrepancy caused by nonlinearity, forcing tradeoffs in approximation methods that replace expectations of transforms by transforms of expectations.

Synthesis

Jensen's inequality codifies the intuitive fact that 'the average of a convex transformation is at least the convex transformation of the average'; it is the foundational relation explaining how nonlinearity and uncertainty interact in probabilistic and statistical analysis.