For and , independence gives the cumulative distribution function
The cumulative distribution function is zero below and one above . Differentiating on the interior gives
up to irrelevant endpoint values. There are no endpoint atoms. This maximum is an order statistic.
Split the unit square into and . The maximum is on the first triangle and on the second, so
The shared boundary has zero probability, so it is not counted twice in this joint density calculation.
For , decreasing the second parameter by one gives another valid beta distribution density and
Normalization therefore gives
The restriction is essential for this finite expected value; when , the singularity near instead makes the expectation infinite.
Multiplying the beta distribution density by gives a constant multiple of the normalized density with parameters :
Integrating and using the fact that the latter probability density function integrates to one gives
This is a normalization-ratio argument, requiring no explicit integration of the polynomial density.
At , gradient descent remains a plausible large-scale choice; dense Newton method and full-memory BFGS are generally impractical. A dense double-precision matrix alone contains entries, about TB. Dense factorization also has order arithmetic work. Full BFGS avoids factorization but retains quadratic storage and update work, so it does not solve that scaling problem.
Gradient descent uses linear working storage in , though the cost of reading the data and evaluating the gradient is still for dense . Sparsity or streaming can be important. Limited-memory BFGS, a distinct limited-memory variant, stores only a fixed number of recent update pairs and uses storage and update work, making it a realistic alternative to full BFGS.[ref]
The available Hessian matrix factorization also permits a matrix-free Hessian-vector product:
It costs for dense data without forming . A suitable iterative, safeguarded Newton method can exploit this structure, so the conclusion rules out the dense implementation rather than every possible second-order algorithm.
These methods seek a local solution and require convergence checks appropriate to nonconvex optimization. The formula does not even ensure a finite minimizer exists: with one positive coefficient and nonzero , the positive exponential loss approaches zero along an unbounded direction without attaining it. Dimension alone cannot supply a convergence guarantee.
At , gradient descent and full BFGS are practical candidates; a dense Newton method can also be feasible, but its extra work and indefinite curvature need attention. The number of examples , data sparsity and conditioning still matter.
For dense , a gradient evaluation costs arithmetic, with working storage beyond the data and accumulation buffers. Gradient descent is inexpensive per iteration and, with an appropriate line search, follows a descent direction, but poor conditioning can make it slow.
Forming the Hessian matrix explicitly costs arithmetic, and a dense Newton method solve costs . A -square double-precision matrix uses about MB, so storage alone is manageable, but repeated factorization may be expensive. Because the Hessian matrix can be indefinite or singular, an unmodified Newton step need not decrease the loss; a modified Hessian, damping or a trust region is appropriate.
Full BFGS uses gradients and an curvature approximation, avoiding explicit second derivatives. Its rank-two updates and direction calculation cost per step and require roughly the same matrix storage. A positive-definite initial approximation with safeguarded curvature updates and a suitable line search preserves descent directions. For this scale it is often a good compromise, without guaranteeing a global minimum.[ref]
Differentiating the gradient term by term gives the Hessian matrix
Take row of to be . Then
Indeed, entry of this product is , exactly the entry of the displayed sum of outer products. Explicitly,
These coefficients may have either sign, so a positive-definite Hessian matrix cannot be assumed. This is an instance of the Hessian of a sum of functions of linear forms.
Use a column gradient. By the multivariable chain rule,
The scalar coefficients multiply the constant vectors . In particular , although explicit differentiation of is not required.
About five steps are certified by the uniform worst-case estimate; the actual iteration reaches double-precision accuracy in about four. Put . The initial enclosure gives , and repeated squaring yields
For this is about , while for it is below . Thus five ideal-arithmetic steps push this bound well below the approximately relative resolution of double-precision floating-point format.
The bound is conservative. Factoring identifies
High-precision evaluation of the actual successive errors gives approximately , , , and . Thus the fourth step would already exceed double-precision accuracy in exact arithmetic. In floating-point evaluation, cancellation in the residual and other roundoff errors prevent the subsequent theoretical digits from being realized.
On , and , since . Hence increases and has minimum . Also
with equality at . Therefore
There is a unique root in : , , and the function is strictly decreasing. Moreover , so .
It remains to justify that the iteration stays where the error estimate applies. Because is a concave function, its graph lies below each tangent line. At any , the tangent has nonpositive value at , negative slope, and nonnegative value at . Its zero therefore satisfies
By mathematical induction, the Newton root-finding iteration stays in and decreases to some limit . Continuity in implies , so uniqueness gives .
Now the supplied error bound applies at every step:
This proves at least quadratic convergence, with an invariant interval and a nonvanishing derivative explicitly established.
The Newton root-finding iteration is
It is defined when . Finding an exact zero of a function is a stopping condition, rather than a reason to continue dividing.
The extreme value theorem guarantees extrema on the closed interval. Compare its endpoints and the stationary points :
Thus
The two-point answer in each case also follows from the even function symmetry.
The first three derivatives factor as
Let . Sign analysis of the factors gives
All displayed endpoints are excluded because a factor vanishes there; the extra zero of at does not create a positive interval.
On the panel , the maximum fourth derivative is at the left endpoint because it is decreasing. Thus
The geometric series and the sum of consecutive integers give
Substitution proves
For , at least one panel's maximum is smaller than the global maximum , so this bound is strictly tighter than the global bound. For , the two bounds coincide.
Take to be a positive even integer, as required by composite Simpson's rule. There are two-strip panels, each of width . Applying the supplied panel bound with the global fourth-derivative minimum and maximum , and summing, gives
Here is the nonnegative absolute error. In this example Simpson's rule overestimates the integral, so . Each panel has that same sign because its fourth derivative is nonnegative; cancellation is not being assumed.[ref]
False. It suffices to find a subsequence along which the ratio is unbounded. Set . The last factors of are at least , while the first factors have product at least one. Hence
This contradicts the eventual uniform bound required by Big O notation. The proof is an elementary factorial comparison, independent of Stirling's approximation.
True. For every , , and raising both nonnegative sides to the th power gives
Thus the Big O notation condition holds with constant and threshold .
False. For , every factor from through is at least , so
No fixed multiple of eventually bounds the factorial, contradicting the definition of Big O notation. No use of Stirling's approximation is required.
The displayed relation is not a partial order on the printed . Since the empty set belongs to , vacuous truth makes
hold for every finite , including nonempty . This violates the defining condition of an antisymmetric relation. It also violates transitivity: holds, but . Consequently the usual unique-maximum and unique-least upper bound terminology for a partially ordered set does not apply as claimed.
The available inputs still determine what happens if one applies the formulas in (b) literally to this relation:
FamilyElements satisfying the maximum formulaElements satisfying the least-upper-bound formula
None
and and
None and
For , no nonempty finite set can lie above all triples of arbitrarily large numbers, while does so vacuously. Thus is its sole literal upper bound, but is not a member of . For each of the two bounded families, the literal upper bounds are and all nonempty finite sets whose elements are at least . Among these, exactly and lie below every such upper bound. Both belong to , but neither belongs to its three-element subfamily.
For comparison, a natural correction is to retain the given rule for nonempty sets and declare to be a bottom element: no nonempty set is then below . Under this explicitly corrected partial order, has neither a maximum nor a least upper bound; has maximum and least upper bound ; and has no maximum but has least upper bound . A nonempty upper bound must have every element at least as large as every element of the bounded family. These corrected answers are conditional on the stated repair, not answers obtained by silently excluding the empty set from the source.

There are unlisted articles, also show them or only show them.