Steffensen's inequality

Bounds a capped, fixed-mass weighting of a monotone function between its integrals over equal-length endpoint intervals.

Version

v1 · 2026-10-17 · History

Domain-specific #

14314

Domain group

Formal Sciences

Origin domain

Mathematics

Subdomains

Real Analysis, Integral Inequalities → Mathematics

Aliases

Steffensen integral inequality

Core Idea

Steffensen's inequality bounds a weighted integral when the function being weighted is monotone and the weight can vary in shape but has a fixed total mass and never exceeds one. For a real-valued, integrable, nonincreasing function \(f\) on \([a,b]\), an integrable weight \(g\) with \(0\leq g\leq 1\), and \(k=\int_a^b g(t)\,dt\), the statement is^[1]

\[ \int_{b-k}^{b} f(t)\,dt \;\leq\; \int_a^b f(t)g(t)\,dt \;\leq\; \int_a^{a+k} f(t)\,dt. \]

The rightmost interval of length \(k\) supplies the lower bound, and the leftmost interval of the same length supplies the upper bound. These are sharp: choosing \(g\) as the indicator of either endpoint interval attains the corresponding equality. When \(f\) is nondecreasing, both inequalities reverse. The inspected published statement does not require \(f\) to be nonnegative; that condition in the frozen screening seed is an unnecessary restriction for this finite-integral form.^[1]

Structural Signature

Sig role-phrases: ordered interval; monotone integrand; unit-capped weight; conserved weight mass; weighted target integral; equal-mass endpoint comparators.

• Ordered interval. \([a,b]\) gives a left and a right end and a length scale. Without that order, the classical leftmost and rightmost length-\(k\) comparison intervals cannot be named.^[1]
• Monotone integrand. A nonincreasing \(f\) makes moving an admissible amount of weight left weakly increase the weighted integral and moving it right weakly decrease it. For a nondecreasing \(f\), the direction reverses. Mere integrability without the relevant monotonicity does not supply these bounds.^[1]
• Unit-capped weight. \(g\) may be zero or one, or may take fractional values, but it stays in \([0,1]\). This cap prevents a fixed amount of weight from being compressed into less than its own length \(k\). A height-two weight can violate the claimed upper bound.
• Conserved weight mass. \(k=\int_a^b g\) fixes how much weight is being compared. Since \(0\leq g\leq 1\), \(0\leq k\leq b-a\); the two endpoint intervals of length \(k\) therefore fit inside the domain.^[1]
• Weighted target. \(\int_a^b fg\) is the unknown or variable-shape quantity being bounded. The theorem is a relation among integrals, not a pointwise estimate for \(f\) or \(g\) separately.^[1]
• Endpoint comparators. The integrals of \(f\) over \([b-k,b]\) and \([a,a+k]\) are the extremal lower and upper cases for a decreasing \(f\). They use the same mass \(k\) as \(g\), and endpoint-indicator weights attain them.^[1]

What It Is Not

It is not a generic promise that any known total weight mass bounds an integral. The pointwise cap \(g\leq 1\), monotonicity of \(f\), and the particular equal-mass endpoint comparisons do the work. For example, on \([0,1]\), let \(f(x)=1-x\) and \(g=2\) on \([0,1/4]\), zero elsewhere. Then \(k=1/2\), but the weighted integral is \(7/16\), greater than the purported upper endpoint value \(3/8\). The classical hypothesis has been broken, so the classical conclusion fails.

Nor does the theorem say that \(f\) must be positive. Adding a constant \(c\) to \(f\) adds \(ck\) to each of the three displayed integrals and leaves their comparisons unchanged. A formulation that insists on nonnegativity is a valid special case, not the full identity stated in the inspected source.^[1]

Scope of Application

The literal scope is real analysis on an ordered interval: an integrable monotone function is paired with a measurable, integrable, unit-capped weight. An indicator weight selects a possibly scattered measurable region; a fractional weight assigns partial intensity across the interval. Both are genuine instances of the same formal rule when the hypotheses hold. They should not be conflated with claims that a particular actuarial or probability study used the theorem.^[1]

Steffensen's original 1918 paper has an actuarial framing in its title, and a later published article explicitly attributes the classical inequality to him. This source pass inspected the later article's full theorem statement but only bibliographic metadata for the 1918 original. It therefore does not establish a specific actuarial valuation formula, original proof passage, or detailed historical application.^[2][1]

Clarity

The mass \(k\) is a length in the endpoint comparisons because \(g\) can place at most one unit of weight at any location. It is not necessarily a probability, a fraction of players, or an observed duration. If \(g\) is an indicator, \(k\) is the measure of the selected part of the interval. If \(g\) is fractional, it is the integral of those fractional values. The same \(k\) determines both bounds.^[1]

The direction depends on the order of \(f\). For decreasing \(f\), the largest values lie toward \(a\), so left packing gives the upper bound; for increasing \(f\), the larger values lie toward \(b\), so both comparison directions reverse. Keeping that orientation explicit prevents an apparently plausible but false use of the displayed inequality.^[1]

Manages Complexity

There may be many admissible profiles of \(g\) with the same total mass. The theorem avoids optimizing over each profile one by one: it replaces them with two endpoint-indicator weights whose integrals are often easier to compute. The bounds are sharp for the stated class, even though a particular non-extremal \(g\) may produce a value strictly inside them.^[1]

The compression has a limit. Knowing more about the actual shape of \(g\) may narrow the estimate, while knowing only its cap and mass leaves the interval between the two endpoint values. That observation describes how much information the inequality uses; it is not evidence of an intrinsic cost or opposing design objective.

Abstract Reasoning

To apply the inequality, first check the ordered interval and finite integrals. Determine whether \(f\) is nonincreasing or nondecreasing. Verify \(0\leq g\leq 1\) almost everywhere, compute \(k=\int_a^b g\), and then evaluate the integrals over the leftmost and rightmost intervals of length \(k\). Only after those checks does the bound license a conclusion about \(\int_a^b fg\).^[1]

The source gives an exact reason for the upper gap when \(f\) decreases:^[1]

\[ \int_a^{a+k} f-\int_a^b fg =\int_a^{a+k}\bigl(f(t)-f(a+k)\bigr)\bigl(1-g(t)\bigr)\,dt +\int_{a+k}^b\bigl(f(a+k)-f(t)\bigr)g(t)\,dt. \]

Each product on the right is nonnegative under the hypotheses. The corresponding identity at \(b-k\) proves the lower gap. This decomposition separates the two reasons the estimate works: \(f\)'s order fixes the sign of each value difference, and \(g\)'s cap fixes the sign of \(1-g\). Remove either condition and the sign argument can fail.^[1]

Knowledge Transfer

The same check transfers between two mathematically unlike weight objects. A set indicator records inclusion or exclusion of positions; a continuously fractional \(g\) can weight every interior point by a different amount. Both enter through the unit cap and total mass, so neither changes the theorem's identity. The transfer is literal within the theorem's analytic domain, and the examples below are constructed to show it rather than to document an external field's adoption.^[1]

The 2015 research article develops Taylor-formula generalizations for \(n\)-convex functions. Those results demonstrate a research lineage, but their added hypotheses and formulas should not be silently folded into the classical two-sided statement here.^[1]

Examples

Canonical — a disconnected selected region

On \([0,1]\), choose \(f(x)=1-x\). Let \(g(x)=1\) on \([0,1/4]\cup[3/4,1]\) and zero elsewhere. This selector has mass \(k=1/2\). Direct integration gives the endpoint values \(1/8\) and \(3/8\), while the selected-region integral is \(1/4\). Hence \(1/8\leq 1/4\leq 3/8\). This is a constructed instance of the theorem, not a reported actuarial or statistical case.^[1]

Mapped back: the ordered interval is \([0,1]\); the monotone integrand is \(1-x\); the unit-capped weight is the disconnected indicator; the conserved mass is \(1/2\); the weighted target is \(1/4\); and the endpoint comparators are \(1/8\) and \(3/8\).

Applied mathematical illustration — fractional weighting

Keep \(f(x)=1-x\) on \([0,1]\), but use \(g(x)=x\). This weight is neither a selected set nor an endpoint indicator: it is fractional at every interior point. Again \(k=1/2\). The weighted integral is \(\int_0^1 x(1-x)\,dx=1/6\), between the same endpoint bounds \(1/8\) and \(3/8\). This is a second constructed mathematical realization; calling \(g\) an acceptance probability would be an interpretation, not evidence of a documented probability application.^[1]

Mapped back: the ordered interval and monotone integrand are unchanged; the unit-capped weight now varies continuously; its mass remains \(1/2\); the target is \(1/6\); and the comparators remain \(1/8\) and \(3/8\). The same relation survives a substantive change in how the weight fills its role.

Structural Tensions

The checked theorem establishes sharp extremal bounds and clear failure conditions, but it does not establish a necessary trade-off between opposed objectives. More knowledge of \(g\) can tighten an estimate; the classical statement deliberately works with only its cap and mass. That is a distinction in available information, not an intrinsic pressure pair. Likewise, the unit cap and monotonicity are hypotheses, not costs that must be balanced against each other.

The diagnostic question is therefore which assumption is doing the work in the proposed use? If a weight exceeds one, the constructed \(7/16>3/8\) case shows the cap cannot be skipped. If \(f\) is nonmonotone, take \(f(x)=x(1-x)\) and an indicator of \([1/4,3/4]\): the middle integral \(11/96\) exceeds both endpoint half-interval integrals \(1/12\). Such cases call for a different result, not an unqualified invocation of Steffensen's inequality.

Structural–Framed Character

The theorem is structural: its truth turns on order, integration, a pointwise cap, conserved mass, and two equal-length comparisons, not on what a particular \(f\) or \(g\) represents. Its evaluative weight is limited to mathematical correctness and sharpness; usefulness for a chosen problem is a separate judgment. Its human-practice dependence lies in proving and applying it, while the formal implication itself is independent of a particular observer or institution.^[1]

Its institutional origin is a named 1918 publication, although the detailed original text was not inspected here. Its vocabulary travel is literal among analytic problems satisfying the same hypotheses and unproved for the seed's proposed actuarial and probability stories as documented applications. Its import-versus-recognition test is whether the complete integral relation is present: a generic mention of “rearranging weight” or of Steffensen's name does not establish this precise bound.^[2][1]

Its character: a sharp, hypothesis-bound real-analysis inequality that substitutes equal-mass endpoint integrals for an otherwise unspecified capped weight profile.

Structural Core vs. Domain Accent

The potentially wider skeleton is extremal placement of bounded mass along an order. Here the domain-specific identity is more exact: integrable functions on a real interval, monotonicity of \(f\), \(0\leq g\leq 1\), \(k=\int g\), and the two length-\(k\) endpoint integrals. Changing those roles may produce a useful related inequality, but it is no longer this classical statement without a new proof.^[1]

The sources establish reuse across indicator and fractional weight profiles within analysis. They do not by themselves establish a distinct cross-domain Prime for extremal mass placement. A broader pattern would need its own admission and unlike evidence; the present entry remains a named mathematical rule.

Instantiates / Related Primes

This entry is part of Boundedness and is part of Order.

The reviewed graph places Order and Boundedness inside this theorem as strict constituents. Order supplies the real interval, endpoints, and monotone comparison; the unit cap on \(g\) supplies a bounded weight whose complement \(1-g\) remains nonnegative in the gap proof. The arrows run from this specialist theorem to those more foundational Primes. Constraint and Optimization are related ways to describe admissible weights and extrema, but their full live signatures are not required by every statement of this theorem. The current Chebyshev's Inequality and Jensen's Inequality entries have different formal kernels; a shared word “inequality” is not an identity or parent proof.

Relationships to Other Abstractions

Current abstraction Steffensen's inequality Domain-specific

Parents (2) — more general patterns this builds on

• Steffensen's inequality is part of Boundedness Prime

  The unit bound on g is an identity-bearing constituent that makes fixed mass k pack into an endpoint interval of length k.

• Steffensen's inequality is part of Order Prime

  The ordered real interval and monotone comparison are identity-bearing constituents of Steffensen’s endpoint bounds.

Hierarchy paths (4) — routes to 4 parentless roots

• Steffensen's inequality → Boundedness

Show alternative paths (3)

Neighborhood in Abstraction Space

Steffensen's inequality sits in a sparse region of the domain-specific corpus (76^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Numerical Root-Finding & Quadrature Methods (8 abstractions)

Nearest neighbors

• Average Order of an Arithmetic Function — 0.85
• Càdlàg Function — 0.84
• Riemann–Liouville integral — 0.83
• Jensen's Inequality — 0.82
• Maharam Algebra — 0.81

Computed from structural-signature embeddings · 2026-10-17

Not to Be Confused With

• A bound for arbitrary weights: the height-two constructed case breaks the classical upper estimate.
• A result for arbitrary \(f\): the nonmonotone middle-interval case breaks the classical endpoint comparison.
• A nonnegative-function theorem only: the checked classical restatement does not impose \(f\geq0\); integrability and monotonicity suffice for the finite-integral form.^[1]
• Steffensen's numerical method: that name appears in the catalog's Aitken acceleration discussion and concerns fixed-point root finding, not this integral inequality.
• Every later Steffensen-type generalization: altered weight, measure, convexity or differentiability hypotheses may support a different formula; this entry states the classical two-sided relation.^[1]

References

[1] Josip Pečarić, Anamarija Perušić Pribanić, and Ksenija Smoljak Kalamir, Generalizations of Steffensen's Inequality via Taylor's Formula, Journal of Inequalities and Applications 2015, article 207, Introduction Theorem 1.1 and equations (1.1)–(1.3). Full publisher HTML inspected for the classical theorem and later extensions. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y

[2] J. F. Steffensen, On Certain Inequalities Between Mean Values, and Their Application to Actuarial Problems, Scandinavian Actuarial Journal 1918(1), 82–97, DOI 10.1080/03461238.1918.10405302. Publisher-domain bibliographic record inspected; original full text and detailed actuarial applications were not verified. The title does not contain the later eponym. registry ↩a ↩b