Alphabeta Math
Pipeline-generated
How statement and proof provenance work

The first chip identifies the source of the statement or construction; the second identifies the source of its local proof or verification.

  • Literature-sourced: the exact statement appears in a cited source; only wording and notation differ.
  • AI-adapted: a semantically identical restatement of literature-sourced material, modulo indexing, notation, and boundary cases adopted by the library.
  • AI-generated: a genuinely novel statement formulated by AI, with no source for the claim itself.

These labels describe origin, not correctness: citations and verification chips remain separate evidence.

Outward rounded interval and ball arithmetic — Examples

1 · Prerequisites

2 · Summary

These examples show how repeated variables widen a natural interval extension, why dividing [1,2] by [−1,1] has no finite sound hull, and how certified range reduction encloses sine near an exact multiple of π.

3 · Logical flowchart

4 · Definitions, theorems and proofs

None yet.

5 · Examples, counterexamples and false statements

ExampleConstruction: AI-generatedVerification: AI-generatedOpen item page →

Division of [1,2] by [-1,1] has no finite hull

Example

Take numerator interval X=[1,2] and divisor interval Y=[−1,1]. The exact quotient set over y≠0 is unbounded above and below. Therefore no finite interval is a sound hull; the permitted results are a sign split or an entire status.

Facts & Assumptions

Given: X=[1,2], Y=[−1,1], and the quotient set {x/y:x∈X, y∈Y∖{0}}.

[L1]

Outward-rounded interval, rectangle and ball operations: division with a denominator interval containing zero returns split-domain or a non-finite status, never a finite quotient interval.

[L2]

Outward-rounded operations are inclusion-isotone and enclose DAG evaluations: a finite interval result must contain every exact quotient in its domain.

Verification

1.1givenchooseconstructalgebra

For any M>0, choose a real t>M+1, set x=1 and y=1/t. Then x∈[1,2], y∈(0,1]⊂Y, and x/y=t>M. Replacing y by −1/t gives a quotient −t<−M. Thus the quotient set is unbounded in both directions.

2.1step 1.1L1L2algebra∎

A finite interval has finite lower and upper endpoints, so choose M larger than their absolute values in step 1.1; the resulting exact quotient lies outside that interval, contradicting the required inclusion [L2]. The branch in [L1] avoids that contradiction by declining a finite hull.

ExampleConstruction: AI-generatedVerification: AI-generatedjudge pass (gpt-6-luna)Open item page →

Natural interval subtraction loses repeated-input correlation

Example

Let the single real input x range over X=[1,2], and evaluate the expression DAG f(x)=x−x by its natural interval extension. The exact range is f(X)={0}, while the interval computation returns X−X=[−1,1]. Both endpoints are sound, but the output is wider because the two occurrences of x are evaluated as independent members of X.

Facts & Assumptions

Given: The real interval X=[1,2], the expression f(x)=x−x, and the outward subtraction rule.

[L1]

Outward-rounded interval, rectangle and ball operations: for [a,b]−[c,d], the interval result before final endpoint rounding is [a−d,b−c].

[L2]

Outward-rounded operations are inclusion-isotone and enclose DAG evaluations: natural interval evaluation encloses every exact DAG value.

Verification

1.1givenL1algebra

Substituting a=c=1 and b=d=2 into [L1] gives [1,2]−[1,2]=[1−2,2−1]=[−1,1].

2.1givenL2algebra∎

For every real x∈[1,2], the two DAG nodes have the same input value, so f(x)=x−x=0, and at x=1 this value is attained. Thus the exact range is {0}, which is contained in the computed interval by [L2].

ExampleConstruction: AI-generatedVerification: AI-generatedjudge pass (gpt-6-luna)Open item page →

Certified sine range near a multiple of pi

Example

Use truncation index N=12 (13 retained terms) and the finite format p=128, emin⁡=−512, emax⁡=512. Form the Machin interval P from exact rational partial sums and the alternating-series tail bounds, and round only its final two endpoints outward. The formula in the trigonometric lemma gives width⁡(P)≤32(1/5)2727(1−1/25)+8(1/239)2727(1−1/2392)+2⋅2−126<10−18. Let X=2P, request width ε=1, and use total box-split budget B=0. The certified sine evaluator uses the canonical cell k=4 and the range reduction centered at 2π.

Facts & Assumptions

Given: The exact-rational Machin interval P, the interval X=2P, the 128-bit dyadic format, N=12, ε=1, and B=0.

[L1]

Range-reduced exponential, logarithm and arctangent series have certified tails: for t∈[0,1/2], the series through index N has error at most t2N+3/((2N+3)(1−t2)).

[L2]

Trigonometric identities, Machin's formula and reduced Taylor tails: Machin's identity is exact, and a sine Taylor polynomial on a reduced interval plus its explicit tail encloses the exact sine range.

[L3]

Budgeted range evaluation terminates and certifies only sound enclosures: a returned certified enclosure contains the exact range and meets the requested width.

[L4]

Continuous functions have an oriented Riemann integral: if a continuous integrand lies between constants m and M on a forward interval of length d, its integral lies between md and Md.

Verification

1.1givenL1L2L4algebra

Apply [L1] at 1/5 and 1/239 with N=12. The arctangent interval widths are at most twice the displayed tails; multiplying by 16 and 4 gives the two rational terms in the bound. By [L4], the integral gives 13/20≤arctan⁡1≤9/10<1: its integrand is at least 4/5 on [0,1/2] and at least 1/2 on [1/2,1], while it is at most 1 on the first half and at most 4/5 on the second. Thus 13/5≤π≤18/5<4. Before rounding, the Machin interval has width less than 2⋅10−19+10−30 by the tail estimates below, so its exact endpoints lie in [5/2,37/10] using 13/5≤π≤18/5. The directed conversions stay in that interval and therefore each add at most 2−126 to the width. For the first tail, 32(1/5)27/(27(1−1/25))=800/(648⋅527)<2⋅10−19, using 4000⋅1018<648⋅527. The second tail is less than 239−27<16−27=2−108<10−30. Therefore exact rational comparison gives W:=width⁡(P)<10−18 and π∈P.

2.1step 1.1L2algebra

Each x∈X has form x=2p with p∈P, so ∣x−2π∣=2∣p−π∣≤2W<2⋅10−18<inf⁡(P)/4≤π/4; inf⁡(P)≥5/2 by step 1.1. Thus X lies inside the canonical k=4 cell. If P=[p−,p+], the actual cell formula Ck=kP/2+[−1/4,1/4]P gives sup⁡C3=7p+/4 and inf⁡C5=5p−/2−p+/4=9p−/4−W/4. Since p−>2 and W<10−18, both p−>7W and p−>9W hold. Consequently 2p−−sup⁡C3=(p−−7W)/4>0 and inf⁡C5−2p+=(p−−9W)/4>0. Thus X=2P meets neither neighboring enclosure. All farther cells are separated as well by monotonicity of their endpoints in the integer cell index, while C4 contains X. Hence the only included candidate is k=4. The interval residual computed by the evaluator is R=X−4(P/2)=[−2W,2W], which contains each exact reduced argument and lies inside the certified Taylor range.

3.1step 2.1givenL2L3algebra∎

Put r0=2W<2⋅10−18. Bound each odd Taylor term on R=[−r0,r0] by the symmetric exact rational interval [−r02k+1/(2k+1)!,r02k+1/(2k+1)!], sum these intervals exactly, and add the geometric tail before outward conversion. The polynomial interval has total width at most 2∑k=012r02k+1/(2k+1)!≤2r0/(1−r02)<6⋅10−18, since r02<1/3 gives (1−r02)−1<3/2. Since r0<1/2, the displayed tail formula gives Ts<2r027/27!<2r02<10−18, so the symmetric tail width is below 2⋅10−18. Only four directed endpoint conversions occur in this exact-rational path (two for P and two for the sine enclosure); their total width increase is at most 4⋅2−126, below the conservative allowance 200⋅2−126. The total width is therefore below 1. With B=0, the budgeted evaluator checks this finite enclosure and returns certified; [L3] confirms the exact range and width claim. No system sine routine is used.

Sources