1 Statement and intuition

Jensen’s inequality connects the value of a convex (or concave) function applied to an average with the average of the function applied to the points being averaged. In probability language, it compares \(f(\mathbb{E}[X])\) to \(\mathbb{E}[f(X)]\). Intuitively, convexity means that “mixing” inputs increases the function value relative to evaluating the function at the mixed (averaged) point.

1.1 One-variable convex form

Let \(f:I\to \mathbb{R}\) be convex on an interval \(I\subseteq \mathbb{R}\). If \(X\) is a random variable such that \(\mathbb{P}(X\in I)=1\) and the expectation \(\mathbb{E}[f(X)]\) is finite (or in an appropriate integrability class), then

\[ f(\mathbb{E}[X]) \le \mathbb{E}[f(X)]. \] If \(f\) is concave, the inequality reverses.

More generally, for a probability measure \(\mu\) supported on \(I\), one may write \[ f\!\left(\int_I x\,d\mu(x)\right)\le \int_I f(x)\,d\mu(x). \]

1.2 Concave function variant

For concave \(f\), convexity arguments apply to \(-f\), yielding \[ f(\mathbb{E}[X]) \ge \mathbb{E}[f(X)]. \] Thus, concave functions reward averaging rather than penalize it.

1.3 Geometric interpretation (chords and tangents)

Convexity admits a geometric description: the graph of a convex function lies below its chords. Concretely, for \(x,y\in I\) and \(t\in[0,1]\), \[ f(tx+(1-t)y) \le tf(x)+(1-t)f(y). \] Jensen’s inequality can be viewed as an extension of repeatedly applying this chord inequality across more than two points, or across a whole distribution. A related viewpoint uses tangents: for convex \(f\), every supporting line underestimates the graph, and averaging preserves this ordering.

1.4 Equality conditions

Equality holds when the randomness does not “create curvature effects.” Typical sufficient conditions are:

  • Degenerate distribution: If \(X\) is almost surely constant, then \(f(\mathbb{E}[X])=\mathbb{E}[f(X)]\).
  • Affine behavior on the relevant range: If \(f\) is affine on the convex hull of the support of \(X\), then equality holds.
  • Strict convexity: If \(f\) is strictly convex, equality forces \(X\) to be constant almost surely (under standard assumptions ensuring the expectation is meaningful).

2 Discrete formulation

Jensen’s inequality is often introduced in a finite setting, where it follows directly from convexity and weighted averages.

2.1 Weighted averages

Let \(x_1,\dots,x_n\in I\) and let \(w_1,\dots,w_n\ge 0\) with \(\sum_{i=1}^n w_i=1\). For convex \(f\), \[ f\!\left(\sum_{i=1}^n w_i x_i\right)\le \sum_{i=1}^n w_i f(x_i). \] This is the core finite form; continuous and probabilistic versions can be obtained by limits.

2.2 Finite probability distributions

If a random variable \(X\) takes values \(x_i\) with probabilities \(w_i=\mathbb{P}(X=x_i)\), then \(\mathbb{E}[X]=\sum_i w_i x_i\) and \(\mathbb{E}[f(X)]=\sum_i w_i f(x_i)\). The discrete weighted inequality becomes \[ f(\mathbb{E}[X])\le \mathbb{E}[f(X)]. \]

2.3 Deriving Jensen from convexity

One derivation proceeds by expressing the weighted average as iterated convex combinations. For example, combine \(x_1\) and \(x_2\) using a weight \(w_1/(w_1+w_2)\), apply convexity, then combine the result with \(x_3\), and continue. Each step preserves the inequality direction, yielding the full \(n\)-point statement.

3 Integral and measure-theoretic formulation

Measure-theoretic formulations express Jensen’s inequality for general probability measures, including cases where \(X\) is not finitely supported.

3.1 Expectation form under a probability measure

Let \((\Omega,\mathcal{F},\mathbb{P})\) be a probability space and \(X:\Omega\to I\) be measurable. Under suitable integrability conditions, \[ f(\mathbb{E}[X]) \le \mathbb{E}[f(X)]. \] Equivalently, for a probability measure \(\mu\) on \(I\), \[ f\!\left(\int_I x\,d\mu(x)\right)\le \int_I f(x)\,d\mu(x). \]

3.2 Integrability requirements

The inequality requires:

- Finite mean: \(\mathbb{E}[X]\) must be well-defined (typically \(\mathbb{E}[X]<\infty\)).
  • Finite expected output: \(\mathbb{E}[f(X)]\) must be meaningful, often requiring \(\mathbb{E}[f(X)]\in \mathbb{R}\) or at least not of the indeterminate form \(\infty-\infty\).
  • Convex function domain: \(f(X)\) must be defined almost surely, hence \(X\) is essentially supported in the interval where \(f\) is convex.

When \(f\) is extended-real-valued or the integrals may diverge, one typically states Jensen’s inequality under conventions ensuring the expression is not undefined.

3.3 Jensen for integrals with weights

A common continuous analogue uses weighted integrals. If \(g\ge 0\) is an integrable weight on a domain \(D\) and \(Z=\int_D g\) is finite and positive, then for convex \(f\), \[ f\!\left(\frac{1}{Z}\int_D x\,g(x)\,dx\right)\le \frac{1}{Z}\int_D f(x)\,g(x)\,dx, \] whenever both integrals exist. This formulation highlights that the inequality is fundamentally about averaging.

3.4 Extension to signed/extended-real settings (when applicable)

In more advanced settings, \(f\) may take values in \((-\infty,\infty]\) or \([-\infty,\infty)\), and the measure may involve extended integrals. Care is needed to avoid undefined combinations such as \(\infty-\infty\). Under appropriate assumptions (e.g., ensuring one side is not of the indeterminate type), Jensen’s inequality can still be stated in a meaningful extended sense. These variants are common in convex analysis and optimization, where constraints may naturally lead to extended-real values.

4 Conditional Jensen’s inequality

Jensen’s inequality remains valid when averaging is performed conditionally, replacing ordinary expectation by conditional expectation with respect to a sigma-algebra.

4.1 Formulation with conditional expectation

Let \(f\) be convex and \(X\) integrable. For a sub-sigma-algebra \(\mathcal{G}\subseteq \mathcal{F}\), \[ f\!\left(\mathbb{E}[X\mid \mathcal{G}]\right)\le \mathbb{E}[f(X)\mid \mathcal{G}] \quad\text{almost surely.} \] For concave \(f\), the inequality reverses.

4.2 Applications to iterated expectations

Conditional Jensen is frequently used with the tower property. If one applies Jensen conditionally and then takes expectations again, it yields strengthened bounds compared with applying Jensen directly. This technique is a standard tool in stochastic analysis, where one controls functions of random variables through their conditional means.

4.3 Convexity through sigma-algebras

The conditional form can be interpreted as a statement about how conditional averaging interacts with convexity. The sigma-algebra \(\mathcal{G}\) represents the information available before averaging. Jensen’s inequality ensures that, given that information, replacing \(X\) by its conditional mean does not increase the convex function’s expected value.

5 Multivariate Jensen’s inequality

Jensen’s inequality extends beyond one-dimensional inputs to vector-valued random variables and convex functions defined on vector spaces.

5.1 Convex functions on vector spaces

Let \(f: C\to \mathbb{R}\) be convex on a convex set \(C\subseteq \mathbb{R}^d\). For a random vector \(X\) taking values in \(C\) with \(\mathbb{E}[X]\in C\) and appropriate integrability, \[ f(\mathbb{E}[X]) \le \mathbb{E}[f(X)]. \] Here convexity means that for any \(x,y\in C\) and \(t\in[0,1]\), \[ f(tx+(1-t)y)\le tf(x)+(1-t)f(y). \]

5.2 Affine maps and preservation of convexity

If \(A\) is an affine map and \(f\) is convex, then \(f\circ A\) is convex on its domain. This interacts well with Jensen’s inequality because many distributions are studied after applying affine transformations, such as centering or projecting onto subspaces. In practice, verifying convexity along relevant affine directions often reduces multivariate problems to the general convexity definition.

5.3 Jensen with random vectors

With multivariate random vectors, Jensen’s inequality becomes a tool to bound nonlinear functions of expectations. It is used in deriving inequalities for norms, quadratic forms, and other convex objectives, where \(f(X)\) represents a performance measure and \(f(\mathbb{E}X)\) compares it to the performance at the averaged state.

Beyond being a standalone inequality, Jensen’s inequality connects to deeper principles in convex analysis and majorization theory.

6.1 Jensen and subgradients (supporting lines)

For convex \(f\), at interior points there exist supporting hyperplanes described via subgradients. This yields a characterization: if \(g\) is a subgradient at a point \(x_0\), then \[ f(x)\ge f(x_0)+g\cdot (x-x_0). \] Taking expectations of such inequalities and choosing \(x_0\) as \(\mathbb{E}[X]\) gives a route to Jensen’s inequality, and conversely, the validity of Jensen-type statements can imply supporting-line properties for convex functions.

6.2 Jensen via epigraphs

Convexity can also be formulated in terms of the epigraph \(\mathrm{epi}(f)=\{(x,r): r\ge f(x)\}\), which is a convex set. Jensen’s inequality can be interpreted as expressing that averaging points in a convex set preserves membership under convex combinations, which translates into the inequality between \(f(\mathbb{E}X)\) and \(\mathbb{E}f(X)\). This viewpoint is especially common in convex optimization.

6.3 Connections to Karamata’s inequality

Karamata’s inequality states that if one vector majorizes another, then sums of a convex function over the components satisfy an order relation. Jensen’s inequality can be seen as a special case or a closely related consequence: “averaging” operations correspond to particular majorization steps, and convex functions respond monotonically under these operations.

6.4 Relationship to majorization (overview)

Majorization formalizes the idea that one distribution is more “spread out” than another. Convex functions typically increase under such spread. In broad terms, Jensen’s inequality captures the monotonicity of convex functionals under the simplest averaging (from a distribution to its mean). Majorization extends this to richer families of redistribution operations.

7 Applications in analysis and probability

Jensen’s inequality appears throughout analysis because it turns convexity into quantitative bounds for expectations and integrals.

7.1 Moment and tail bounds (typical use cases)

Convexity-based bounds often lead to inequalities involving moments. For example, choosing \(f(x)=x^p\) for \(p\ge 1\) (on appropriate domains) relates \((\mathbb{E}X)^p\) to \(\mathbb{E}X^p\). Such relationships feed into tail estimates via auxiliary tools (e.g., Markov-type arguments), where moments control the likelihood of large deviations.

7.2 Entropy and information-theoretic inequalities (overview-level)

In information theory, many quantities are expressed through expectations of convex functions, such as logarithms or exponentials. While the full derivation of specific results depends on additional lemmas, Jensen’s inequality commonly serves as the foundational step for showing that certain divergences or entropy-related expressions are nonnegative or bounded. Its role is to compare a convex functional evaluated at an average with the average of the functional itself.

7.3 Inequalities from convexity (template approach)

A recurring template is:

  1. Identify a convex function \(f\) that matches the structure of the problem.
  2. Rewrite the target expression as \(f(\mathbb{E}[X])\) or \(\mathbb{E}[f(X)]\).
  3. Apply Jensen to obtain an inequality in the desired direction.
  4. Combine with monotonicity or other standard inequalities to reach the final bound.

This method is widely used in both deterministic analysis and stochastic settings.

7.4 Bounding integrals via convex envelopes

When integrals involve nonlinear terms, one can bound them by replacing the integrand with a convex envelope or by applying Jensen to conditional expectations inside the integral. This can simplify analysis by reducing the complexity of a nonlinear integral expression to something involving only averaged quantities.

8 Common examples

Illustrative choices of \(f\) show how Jensen’s inequality recovers familiar inequalities in algebra, analysis, and probability.

8.1 Convex power functions

For \(p\ge 1\), the map \(f(x)=x^p\) is convex on \([0,\infty)\). If \(X\ge 0\) and \(\mathbb{E}[X]\) exists, then \[ (\mathbb{E}[X])^p \le \mathbb{E}[X^p]. \] This provides a direct comparison between the \(p\)-th moment and the \(p\)-th power of the mean.

8.2 Exponential and logarithmic cases

The exponential function is convex on \(\mathbb{R}\), so for integrable \(X\), \[ \exp(\mathbb{E}[X]) \le \mathbb{E}[e^X]. \] Logarithms appear through convexity or concavity depending on the direction used; for concave \(\log\), \[ \log(\mathbb{E}[X]) \ge \mathbb{E}[\log X] \] for positive \(X\) under integrability conditions. Such forms are common in bounding products, geometric means, and likelihood-type expressions.

8.3 Reciprocal and quadratic examples

The reciprocal function \(f(x)=1/x\) is convex on \((0,\infty)\). Thus, for \(X>0\), \[ \frac{1}{\mathbb{E}[X]} \le \mathbb{E}\!\left[\frac{1}{X}\right]. \] Quadratic functions \(f(x)=x^2\) are convex on \(\mathbb{R}\), giving \[ (\mathbb{E}[X])^2 \le \mathbb{E}[X^2], \] which is closely related to basic variance decomposition.

8.4 Practical equality examples (when distributions concentrate)

In applications, equality (or near-equality) often occurs when the distribution is concentrated around a single value or when \(f\) behaves nearly affine over the region where \(X\) lives. For strictly convex \(f\), even small spread typically yields a strict inequality; conversely, when \(X\) takes only a couple of values and \(f\) is close to affine over their convex hull, the bound can be tight.

9 Extensions and generalizations

Several higher-level generalizations broaden Jensen’s inequality to operator settings, martingales, and more abstract functional forms.

9.1 Jensen for operator convex functions (high-level overview)

In operator theory, random variables may be replaced by self-adjoint operators and expectations by linear maps that preserve positivity and normalization. Operator convexity is a matrix-valued analogue of scalar convexity. Under appropriate conditions, a Jensen-type inequality holds in the form \[ f(\Phi(A)) \le \Phi(f(A)), \] where \(A\) is an operator, \(\Phi\) is a suitable positive unital map, and \(f\) is operator convex. This framework is used in matrix inequalities, quantum information, and semidefinite programming.

9.2 Jensen for martingales (high-level overview)

Martingale versions interpret conditional Jensen along the filtration of a stochastic process. If \((X_n)\) is a martingale and \(f\) is convex, then the process \(f(X_n)\) often forms a submartingale, reflecting the principle that convex transformations increase conditional expectation. The precise statement depends on integrability and measurability assumptions, but the guiding idea is the conditional Jensen inequality applied iteratively in time.

9.3 Functional forms and Jensen-type inequalities

Many generalized inequalities take the “Jensen template” but adapt the setting. Examples include:

  • inequalities for convex functions under nonstandard averaging operators (e.g., generalized means),
  • Jensen inequalities for random variables defined on more general spaces,
  • variants involving concave–convex duality (Legendre–Fenchel transforms) that re-express Jensen in terms of supporting functions.

These extensions preserve the central message: convexity turns averaging operations into monotone comparisons between function evaluations.