For an autonomous scalar integral , the first variation gives the integrated Euler-Lagrange equation . If is and each fiber of is finite, the connected image of the interval under the continuous derivative lies in a finite set and is therefore a singleton. The fixed endpoints then determine exactly one affine solution. If is strictly convex, the strict Jensen's inequality applied to additionally makes this solution the unique global minimizer among fixed-endpoint Lipschitz functions. Without convexity the same unique smooth stationary solution can be a maximizer, for example at zero for and zero endpoint values.
Atom-probability error in a randomized endpoint rule is the change in included duration-model probability caused by using an inaccurate zero-atom probability in the randomization. Let , and hold the positive upper endpoint fixed at an actual cumulative distribution function value . If the true zero probability is but the rule uses a positive , its independent endpoint-inclusion probability is . Conditional on that chosen value,
This is the atom term isolated in distribution sensitivity of randomized zero-endpoint probabilities, with upper-endpoint error deliberately removed. Underestimating reduces inclusion; overestimating it increases inclusion. The independent auxiliary draw can be generated from a uniform distribution.[ref]
If comes from an independent calibration, conditional probabilities must also be averaged over that calibration. For an original example, take , , and . Suppose the calibration returns or with equal probability. Its estimate is unbiased, yet the resulting included probabilities are and , averaging about rather than .
More generally, for an unbiased positive estimate in the valid rule range, Jensen's inequality gives
This reciprocal-estimation bias in randomized atom inclusion comes from a nonlinear plug-in rule, even with independent calibration. It does not identify the size of the error without a calibration model. Values below , including zero, require an explicitly defined clipped or guarded rule; the displayed inverse formula cannot simply be used there.
If a validated upper bound is available and , choosing instead guarantees under the fixed-upper-endpoint and independence assumptions. It may be conservative rather than exact. Uncertainty in remains a separate error term, and these are probabilities for a duration distribution, not a removal of the physical zero-duration atom.
Common-neighbourhood sampling bound Created 2026-09-24 Updated 2026-11-05
Let a bipartite graph with nonempty parts have edge density of a bipartite graph at least , and let . Some ordered tuple of distinct vertices of has a common neighbourhood of size at least
Sample the tuple uniformly without replacement. A vertex belongs to its common neighbourhood with probability , where . This is at least . The function is convex, so averaging over and using Jensen's inequality proves the bound.
For fixed and , taking sufficiently large compared with makes this at least . The positive part in the formula is essential: without it, a negative base raised to an even power would assert a spurious positive common neighbourhood.
The current-transformer ratio correction factor multiplies a nominally scaled secondary-current indication to recover primary-current magnitude under stated calibration conditions. If is the marked primary-to-secondary ratio and is the true magnitude ratio, define
NIST distinguishes this factor from ratio correction, which is , and from the separately measured phase displacement.[ref] Correcting amplitude does not remove current-transformer phase displacement.
With the inferred-primary error convention , the exact relationship is
For an original example, , and give . Using nominal scaling alone gives 498 A. The approximate correction would instead give 499.992 A, leaving a small second-order difference.
A calibration correction has its own measurement uncertainty. Treating as an exact label and retaining correlation, first-order propagation gives
For a correction inferred from uncertain , its local sensitivity is . The GUM supplies the general sensitivity-and-covariance propagation rule used here.[ref]
Nonlinearity also affects averaging. Suppose a stipulated error takes values and with equal probabilities. Its mean is zero, but the mean correction factor is
The correction at the mean error is instead one. Convexity of the reciprocal on explains this Jensen's inequality effect. It is an illustrative probability model, not a claim about random drift of a particular transformer.
A correction factor should therefore travel with its current level, frequency, current-transformer burden and uncertainty information. A proposed check repeats the ratio comparison at selected operating conditions and tests whether one retained correction remains adequate. Applying more displayed digits cannot replace that validation.
EM likelihood monotonicity Created 2026-10-05 Updated 2026-10-14
For , the Jensen inequality gives
Thus any expectation-maximization algorithm M-step that increases cannot decrease the observed-data likelihood function.[ref]
For a short proof, assume common support and finite expressions, and keep the E-step conditional distribution fixed. The concavity of the logarithm and Jensen's inequality give
This evidence lower bound equals . At , the ratio inside the expectation is the constant , so equality holds. During the M-step, the second term is constant; increasing raises a lower bound that touches the current log-likelihood. Hence
Increasing likelihood values alone do not establish convergence of the statistical parameters or attainment of a global maximum. In a differentiable finite mixture model, the gradient of equals the candidate-parameter gradient of evaluated at the current fit. Thus an interior fixed point of an exact M-step is a stationary point, which need not be a local maximum; the Identical-mean EM trap is an explicit counterexample. Small changes in log-likelihood can also occur before the means have settled, so inspecting both is useful.
Entropy production Created 2026-09-28 Updated 2026-11-05
Entropy production is the entropy change of the system plus that of its environment. For a discrete-state stochastic process, the stochastic system entropy at the occupied state is . A single thermal reservoir contributes , with positive heat into the reservoir. Including the actual endpoint probabilities is essential even if the final ensemble has not reached thermal equilibrium.[ref]
Under the hypotheses of the integral fluctuation theorem,
Applying Jensen's inequality to the convex exponential function gives , hence . This constrains the mean, while a short individual sample trajectory may have negative entropy production. Equating a medium-only path-weight ratio to total entropy production would omit the stochastic system term.
A bound implies . It does not by itself imply . For , Jensen's inequality goes in the opposite direction. The missing contribution is the variance: .
For probability mass functions on symbols, set and . The entropy summand bound gives . The triangle inequality and Jensen's inequality give
At , use the continuous value zero. This proves uniform continuity of Shannon entropy on a fixed finite alphabet; the alphabet size appears explicitly in the bound.
Finite heat bath 2026-10-18
A finite heat bath has a finite whole-body heat capacity and therefore changes temperature when it absorbs heat . For a rigid closed system with no work, fixed composition and no phase transition, the first law of thermodynamics gives
If the heat capacity at constant volume is constant throughout this temperature interval, the integral becomes . This statement concerns an internal energy difference; need not be the body's absolute internal energy.
For an original example, put a rigid body 1 and a finite bath 2 into thermal contact inside an isolated system. Stipulate constant whole-body heat capacities and , with initial temperatures and . Each body remains internally uniform; the connection stores negligible energy; no external work, matter exchange or phase transition occurs. These are constructed model values. The argument in thermal equilibration of two finite bodies applies directly: equal final temperatures and conserved total internal energy give
Thus the heat transferred from body 1 to bath 2 is
Body 1 loses of internal energy and bath 2 gains exactly that amount. Body 1 cools by , while bath 2 warms by . Equal heat transfers produce unequal temperature changes because the heat capacities differ.
Figure 1.
Finite bath temperature changes and the integrated heat ledger
. Original constant-heat capacity model, not measured data. Left: accumulated heat leaving body 1 lowers its temperature and raises bath 2's temperature; the lines meet at and . Right: the shaded integrals have equal area , although their widths are and . The outgoing heat magnitude is shown positive; body 1's signed internal energy change is negative. Details.
The heat bath approximation can now be assessed against a chosen allowed temperature change . Constant requires , rather than merely calling the bath large. For this example, an allowed change of would fail because the computed change is . The exact bath shift and its ideal-reservoir limit are
Hold and both initial temperatures fixed in this limit. The transferred heat then approaches . The finite bath receives less heat because its warming reduces the available temperature difference. For heat capacities that vary, use the two integrals in the energy balance instead of this weighted average.
Conservation of energy also permits imagining the reverse transfer, so it cannot explain the direction of equilibration by itself. Entropy from heat capacity gives the additional test, using reversible comparison paths between equilibrium endpoints:
All ratios inside the natural logarithms use absolute temperatures. The actual exchange across unequal temperatures is irreversible, even though reversible paths calculate the endpoint entropy differences.
There is a short general proof of positivity. Put for positive constant heat capacities and positive initial temperatures. The strict concavity of the natural logarithm gives
with equality only for equal initial temperatures. In this proof all temperatures inside the logarithms mean their numerical values in one common absolute unit. This is Jensen's inequality applied to the same weighted average that conservation of energy selected. The direction of heat transfer and the stability of its endpoint also follow from entropy maximization under thermal contact.
Finite-sample bias in a negative multinomial probability estimator Created 2026-10-14 Updated 2026-11-05
For independent one-stop negative multinomial distribution observations, the natural estimates are and . Their sum is with strictly concave , so Jensen's inequality makes its expectation strictly below . Conditional category proportions split this bias between both estimates. With zero observed counts of a category, that estimate lies at zero; in a parameter space requiring strict positivity the likelihood supremum is not attained.
Otto Hölder's 1889 paper begins with a weighted-mean inequality for a function with increasing derivative. In current terminology, its opening statement is a weighted form of Jensen's inequality: applying a convex function after taking a weighted mean gives a lower bound for the weighted mean of its values. This connects product estimates to the geometry of convex functions, rather than treating them as isolated formulas.[ref]
For a biregular graph with left vertices of degree and right vertices of degree , let be the indicator vector of a uniformly selected independent set. Use base-two information entropy. Shearer inequality gives . Conditional on , each right vertex with an empty selected neighbourhood is an independent fair binary choice, and the other right vertices are forbidden. Writing gives . For the neighbourhood patterns, assign weight to the zero pattern and weight one to all others. Jensen's inequality implies . Since , the bound follows from . If divides , the disjoint union of copies of the complete bipartite graph attains equality: each component has independent sets.
Normalization of the reverse law gives an exact finite identity:
The sum notation describes discrete paths; the corresponding path integral has the same change-of-variable argument. Jensen's inequality for the exponential function then gives , hence . Equivalently this mean is the Kullback-Leibler divergence between the forward path law and the reverse path law pulled back by . Equality requires those two laws to agree. [ref]
Rare negative entropy production contributes disproportionately to this expectation through . Missing such events in a finite sample can spoil an estimate of the identity. Matching probability support is also essential: reverse paths absent from the forward law cannot be recovered by cancelling forward weights. This Second law of thermodynamics statement concerns an ensemble average; an individual negative value is compatible with it.
Jarzynski equality Created 2026-10-14 Updated 2026-11-05
Integrating the Crooks fluctuation theorem over work and using normalization of the reverse probability distribution gives
Consequently Jensen's inequality gives . Some realizations may nevertheless have . This is a work relation with equilibrium endpoint preparations, while the integral fluctuation theorem for total entropy production uses the actual final forward distribution. [ref]
Jensen gap Created 2026-10-14 Updated 2026-10-25
The Jensen gap is the nonnegative difference in Jensen's inequality, under its integrability assumptions. It quantifies how much applying a convex function before averaging differs from applying it afterward. For it is the variance of ; for an affine function it is zero. For general convex functions, it depends on both the function and the distribution of , and is not determined by the variance alone.
Jensen's inequality 2026-10-14
Jensen's inequality says that applying a convex function to a weighted mean cannot exceed the same weighted mean of its outputs. For nonnegative weights with and inputs in a convex set,
Nonnegative weights are essential: this is a convex combination, rather than an arbitrary affine combination. For two inputs it is precisely the defining inequality for a convex function; repeatedly grouping inputs gives the finite weighted version.
Take , weights and , and inputs and . Their weighted mean is , so the two sides are
The Jensen gap is . The inequality measures the effect of replacing distinct inputs by one representative input before applying a nonlinear function.
The probabilistic form uses an integrable random variable taking values in an interval where is a convex function. If is also an integrable random variable and belongs to that interval, then
The direction reverses for a concave function. For a strictly convex function, equality in the finite weighted form requires all inputs having positive weight to be equal; in the probabilistic form it requires to be constant almost surely. An affine function gives equality for any admissible distribution. [ref].
For the quadratic function and a random variable with a finite second moment, the difference between the two sides is exactly its variance:
Consider two equally likely values and . Their expected value stays at as increases, but their variance is . Thus the expected value of the square grows from to . For the exponential function, the same two-point distribution gives , while . Both are examples of Jensen's inequality, but their Jensen gaps depend differently on the spread of the inputs.
Figure 1.
Keeping the mean fixed while spreading the inputs increases the average of these convex functions
. The random variable takes and with equal probability. The quadratic function produces a Jensen gap of , exactly the variance. The exponential function produces a Jensen gap of , involving the hyperbolic cosine. The horizontal reference in each panel applies the function directly to the unchanged expected value . Details.
Another useful choice is the concave function on positive inputs. Its version of Jensen's inequality gives . Applying the increasing exponential function shows that the weighted geometric mean cannot exceed the weighted arithmetic mean. For the equally weighted inputs and , these are and . This recovers the arithmetic-geometric mean inequality and illustrates how choosing one function can generate an entire family of bounds.
A short proof idea explains the probabilistic form. If has a subgradient at , the supporting affine function gives . Taking the expected value cancels the linear term because . This argument applies, for example, to finite convex functions on ; more general domains of a function need attention to endpoints and the existence of a subgradient.
Johan Jensen's 1906 paper “Sur les fonctions convexes et les inégalités entre les valeurs moyennes” is an early source for the connection between convex functions and inequalities for means. [ref].
Johan Jensen Created 2026-10-13 Updated 2026-11-05
Johan Jensen was a mathematician and engineer whose work included convex functions, infinite series and complex analysis. His 1906 work on convex functions established the inequality now known as Jensen's inequality. He also related the zeros of a holomorphic function to the average of the logarithm of its absolute value around a circle.[ref] A basic instance of Jensen's inequality is for a real random variable with a finite second moment. Thus applying a convex function after averaging gives no larger a result than averaging after applying that function.
Figure 1.
Portrait of Johan Jensen
. Theodor Vilhelm Rieger; Royal Danish Library, Copenhagen, object166670; photograph dated 1902–1913. Public domain under Commons PD-Denmark50 and PD-US templates. Source and usage information on the linked Wikimedia file page. Source.
An arbitrary initial function has no well-defined nodal point samples. Cell averages or a bounded projection to the discrete approximation space supply grid data independently of its pointwise representative. Jensen's inequality bounds the cell-average discrete L2 norm by the continuous L2 norm. Combining consistent bounded initialization, stability, and density of smooth data can extend convergence to rough initial functions, without claiming their nonexistent classical derivatives.
Lyapunov moment inequality Created 2026-10-06 Updated 2026-10-14
The Lyapunov moment inequality says that higher positive absolute moments give no smaller absolute moment sizes on a probability space. For a real random variable and ,
allowing an infinite value. For the left expression is an Lp norm; below one it is a quasi-norm, so the usual triangle inequality terminology needs that distinction. The comparison uses a probability measure of total mass one; arbitrary unnormalized measure spaces need a factor depending on their total mass.
For a nonnegative count, the first moment size is its expected value, while the second is the square root of its second moment. Both ten-trial counts in second moment method have first moment size one. The independent count has second moment size , whereas the shared-outcome count has second moment size . The second moment distinguishes models with the same average because its square increases the influence of rare large counts.
More generally, for with a Bernoulli random variable of parameter ,
It rises from one at toward ten as . Replacing by a parameter gives , which decreases toward zero for a fixed positive exponent when , making the rare-event and large-exponent limits different questions. For a scaled Rademacher random variable taking values and two with equal probabilities, every absolute moment size is exactly two: increasing the exponent does not distinguish signs when the absolute value is constant.
Figure 1.
Lyapunov moment inequality reveals how higher moments emphasize large outcomes and when equality holds
. Left: the two original ten-trial models from second moment method have the same mean one, but different absolute moment sizes at exponents above one. The shared model has with . Right: equal masses at and two keep almost surely, giving equality for every pair of positive exponents. These are fully specified mathematical probability models. Details.
For finite , apply the reversed Jensen's inequality to the strictly concave function and the nonnegative random variable :
Taking the positive th root proves the claim. Equality in this finite case holds exactly when is constant almost surely. If the higher moment is infinite, the upper bound is infinite and remains valid. Exponent zero is excluded from the statement; logarithmic or geometric limiting summaries require separately stated conventions, especially when can vanish.
A finite-band transmittance averages different monochromatic attenuation rates. Its effective attenuation obeys , where the weighting is proportional to the transmitted spectrum. Strongly absorbing parts saturate first, so surviving radiation samples weaker absorption. Jensen's inequality gives . Equality throughout requires a spectrally constant coefficient.