From reproduction to a research question

A successful reproduction gives you a place from which to ask a better question. It does not establish that the next parameter value is unexplored, or that a discrepancy is new physics. This article turns one exact result into a small investigation whose outcomes can be interpreted.

Enter after one completed capstone: the free-chiral response, finite-rank Wilson loop, rank-one ABJM integral, Seiberg–Witten particle analysis, four-dimensional duality check, or SQED–XYZ deformation. You do not need to complete all six.

The worked example comes from the Seiberg–Witten capstone, but its integral and all required calculus are supplied below. Readers from other entrances can use it as a small model of the method, then return to their own quantity. The status of the exercise is reproduction followed by a training extension.

Write a question that admits a check

Suppose a low-order approximation disagrees with a numerical integral as a vacuum parameter changes. “Does the approximation fail?” is too vague: the answer depends on the required accuracy and the parameter range. A useful question is: How many terms certify an absolute error below a stated tolerance, and which part of that certificate deteriorates near a singular vacuum?

Here is a completed first research note. A convention package means the definitions that must be held together for two answers to describe the same quantity.

Item	Specification for this reproduction
Quantity	The dimensionless electric period $F(t)=a(u)/\sqrt{2u}$, in the real branch specified below
Model and regime	Pure four-dimensional N=2 $SU(2)$ Yang–Mills; real Coulomb vacuum $u>\Lambda^2>0$, fixed $\Lambda$
Conventions	First-paper period normalization $a\sim\sqrt{2u}$; positive electric period; $t=\Lambda^2/u$; no continuation around a singularity
Starting result	The exact integral in (D08.1), obtained from Seiberg–Witten (6.22); reproduce its weak-limit normalization before comparing corrections
Independent check	Compare direct integration with a series whose omitted positive terms have a derived bound; check the endpoint value separately
Controlled variation and status	Increase $t$ at fixed theory, first keeping truncation order fixed and then increasing it. This tests an approximation and its error estimate; it is a training problem, with no novelty claim

This note is already more informative than a plot labeled “exact versus approximate.” It says which vacuum moves, which definition of “exact” is used, and what would count as a disagreement.

Supply enough physics to reproduce the number

The theory has eight real supercharges and gauge rank one. Its generic Coulomb vacuum preserves all eight and has a low-energy Abelian gauge field. The variable $u$ is a gauge-invariant scalar expectation value of mass dimension two; $\Lambda$ is the fixed strong-coupling scale of dimension one. The period $a$ is a special coordinate entering the two-derivative low-energy action, defined by integrating a specified differential around a specified cycle. It is not by itself a duality-frame-independent observable.

For this exercise, its exact definition on the chosen real ray is enough:

$$a(u)=\frac{\sqrt2}{\pi}\int_{-\Lambda^2}^{\Lambda^2} \frac{\sqrt{u-x}}{\sqrt{\Lambda^4-x^2}}\,dx, \qquad F(t)=\frac1\pi\int_0^\pi\sqrt{1-t\cos\theta}\,d\theta, \qquad t=\frac{\Lambda^2}{u}\in(0,1). \tag{D08.1}$$

All square roots in these real integrals are positive. The substitution $x=\Lambda^2\cos\theta$, with its reversed integration limits, gives the second expression. The factor counting both sides of the original contour is already included in the first. Seiberg–Witten (6.21)–(6.24) specifies that factor and the weak limit. The period calculation derives the curve, cycles, differential equation, and effective coupling; those constructions are not additional requirements for this error calculation.

We have fixed the normalization by the finite singularity at $u=\Lambda^2$ and by $a/\sqrt{2u}\to1$. Replacing $\Lambda$ by another subtraction scale without a translation changes the problem. Increasing $t$ here moves between vacua of the same theory; it does not add a mass term or change the number of supercharges.

The two-derivative description is exact in its stated sense. It does not compute every correlation function, and one must retain additional light fields at a singular vacuum. No numerical period alone establishes a new stable particle.

Turn a series into an error certificate

For $|t|<1$, expand the square root under the integral. The expansion is uniformly convergent on every closed interval $|t|\leq t_{\max}<1$, which justifies termwise integration there. Odd cosine moments vanish by $\theta\mapsto\pi-\theta$; even moments are $\binom{2n}{n}/4^n$. Thus

$$\begin{aligned} F(t)&=1-\sum_{n=1}^{\infty}c_n t^{2n},\\ c_n&=-\binom{1/2}{2n}\frac{\binom{2n}{n}}{4^n}>0, \qquad \frac{c_{n+1}}{c_n}=\frac{16n^2-1}{16(n+1)^2}<1,\\ c_1&=\frac1{16},\qquad c_2=\frac{15}{1024},\qquad c_3=\frac{105}{16384}. \end{aligned} \tag{D08.2}$$

The binomial symbol with upper entry $1/2$ means the product $(1/2)(1/2-1)\cdots(1/2-j+1)/j!$. The ratio follows by canceling adjacent products; it gives both a simple implementation and a useful inequality.

Let $F_p=1-\sum_{n=1}^{p}c_nt^{2n}$, where $p$ counts retained correction terms. Because every omitted coefficient is positive and decreases with $n$,

$$0<c_{p+1}t^{2p+2} \leq F_p(t)-F(t) \leq \frac{c_{p+1}t^{2p+2}}{1-t^2}, \qquad 0<t<1. \tag{D08.3}$$

The lower bound is the first omitted term. For the upper bound, replace every later coefficient by $c_{p+1}$ and sum the geometric series. This is an absolute error bound, not a fitted estimate and not a relative error percentage. It also predicts the sign: the finite polynomial lies above the exact integral.

At $t=1/2$, keeping $p=2$ gives

$$F_2=0.98345947265625,\qquad 0.0001001358032227<F_2-F<0.0001335144042969. \tag{D08.4}$$

Direct integration gives $F=0.9833426507751652$, hence a residual about $0.0001168218810848$, inside that interval. Even exact arithmetic applied to $F_2$ cannot reduce this discrepancy.

Here is the same check at four inputs. The decimal entries are rounded numerical values; the inequalities come from (D08.3).

$t$	Integral $F(t)$	Residual $F_2-F$	Certified upper bound	Smallest $p$ for which this bound is below $10^{-8}$
$0.2$	$0.997476142939$	$4.19561\times10^{-7}$	$4.27246\times10^{-7}$	3
$0.5$	$0.983342650775$	$1.16822\times10^{-4}$	$1.33514\times10^{-4}$	8
$0.8$	$0.951263039116$	$2.73696\times10^{-3}$	$4.66667\times10^{-3}$	23
$0.95$	$0.920207466153$	$1.14550\times10^{-2}$	$4.83177\times10^{-2}$	87

The bound column is rounded for display; use (D08.3) for certification. The final column is a sufficient stopping rule for this particular bound. It need not be the smallest order that actually achieves the tolerance.

To reproduce the table, compute the coefficients by their rational recurrence, sum through the chosen order, and integrate (D08.1) directly. Repeat the integral at higher precision and with a different subdivision. Then compare it with the certified series interval. Agreement between two calls to the same special-function routine would be a weaker check.

In the computation supporting this example, direct quadrature was also compared with a positive Laplace-integral representation and an elliptic-integral evaluation at 70-digit working precision. A 40-to-70-digit repeat tests arithmetic stability. These are numerical checks of the implementation; the proof of (D08.3) supplies the truncation guarantee.

Change one input and interpret what happens

Fix $\Lambda$, the branch and the tolerance. Vary $t$, leaving the model unchanged. First hold $p=2$ fixed; next increase $p$ until (D08.3) meets the tolerance. The table answers two different questions: how a chosen approximation deteriorates, and how much work a chosen certificate requires.

A residual outside the bounds is useful evidence against the implementation or its assumptions. Before assigning a physical meaning to it, check the factor of two, the units of $u$, the argument $t$, and whether the integral and polynomial use the same branch. If the residual stays inside the bounds, the disagreement is already explained by omitted terms.

Near $t=1$, the denominator of the upper bound tends to zero. This does not prove that $F$ diverges. In fact,

$$F(1)=\frac{\sqrt2}{\pi}\int_0^\pi\sin(\theta/2)\,d\theta =\frac{2\sqrt2}{\pi}. \tag{D08.5}$$

This agrees with the scale-restored endpoint $a(\Lambda^2)=4\Lambda/\pi$ in Seiberg–Witten (6.28). The derivative has a logarithmic singularity, as the final exercise shows. An approximation can therefore converge in value while giving poor control of a derivative. A physical response calculation must be tested for the quantity actually being differentiated.

The next training question is now precise: Can an error estimate remain useful on the entire interval $0\leq t\leq1$, and what changes if the target is $F'$ rather than $F$? The exercises answer the first part and expose the obstacle to the second. This is a known-function investigation, not a proposed discovery about the theory.

Check the literature before naming an open problem

Literature check: 20 October 2026. The period, its weak expansion, and its singular endpoint are established results. A targeted search for exact Seiberg–Witten periods and strong-coupling expansions leads beyond the original paper to D’Hoker, Dumitrescu and Nardoni, 2208.11502v1. Their section 2.4 compares with known $SU(2)$ periods, explicitly translating scales, cycles and continuation paths. Their broader $SU(N)$ study also distinguishes exact expansions from numerical evidence and conjectures. Its introduction is a useful example of how to state those different statuses; it is not evidence that every question labeled open in 2022 remains open today.

For another capstone, read the matching primary passage before changing the input:

Starting point	Passage and the comparison it teaches
Curved-background response	Closset et al. v3, (8.1)–(8.3): specify the supersymmetric source family and the current normalization before interpreting a derivative
Four-dimensional duality	Intriligator–Seiberg v1, (5.5)–(5.6): meson normalization includes a matching scale, and scale-matching phases matter under deformation
Large-rank gauge theory	Minahan v3, (5.3)–(5.7): color counting and operator normalization are part of the limit; fixed-length conclusions need care when length grows
ABJM	Aharony et al. v4, introduction: specify the compact group and opposite integer levels; increasing rank at fixed level and at fixed rank-to-level ratio are different limits

These are reading directions, not extra prerequisites or claims about the latest frontier in every subject. Return to your completed capstone for the actual conventions and established benchmark.

For a proposed extension, record the model, quantity and regime in your literature search, including alternate names and normalization choices. Read the nearest primary results and their assumptions; follow later work that bears on the same claim. State a dated, limited conclusion such as “this calculation reproduces the stated special case” or “the proposed extension has not yet been checked against the relevant literature.” Failure to find a paper does not establish novelty.

Exercises

1. Distinguish a change of units from a change of vacuum

Two calculations use $(\Lambda,u,a)$ and $(c\Lambda,c^2u,ca)$, with $c>0$. Determine what happens to $t$ and $F$. Now consider replacing $u$ by $2u$ while holding $\Lambda$ fixed. Which comparison requires a new integral evaluation? Explain why a factor-of-two disagreement in $a$, with $u,\Lambda$ unchanged, is not evidence for a different vacuum.

Hint 1

Track the mass dimensions before substituting numerical values.

Hint 2

Both $t$ and $F$ are dimensionless ratios. The contour’s two equal legs supply a separate normalization check.

Solution

The simultaneous rescaling leaves $t'=(c\Lambda)^2/(c^2u)=t$ and $F'=ca/\sqrt{2c^2u}=F$. It can represent a change of units and leaves the dimensionless reproduction unchanged. The dimensional integral scales by $c$, as required.

At fixed $\Lambda$, sending $u$ to $2u$ sends $t$ to $t/2$. Its new result is $a(2u)=\sqrt{4u}\,F(t/2)$; one cannot merely multiply the old answer by $\sqrt2$, because $F$ changes. This moves along the chosen vacuum ray.

If $u,\Lambda$ and the positive branch are unchanged, (D08.1) defines a unique number. A factor of two signals a different normalization or an error, such as counting only one contour leg. The weak limit $a/\sqrt{2u}\to1$ detects that error. Neither comparison establishes the existence of an additional particle.

2. Diagnose a suspicious discrepancy

At $t=4/5$, a calculation reports $F_2-F_{\rm num}=0.00273696$. Repeating the integral at higher precision changes the displayed result by less than $10^{-10}$. Another implementation reports a residual $-0.0027$. Use an exact interval to assess both results. Write a one-sentence next task that changes only the approximation order.

Hint 1

The first omitted coefficient is $105/16384$, and $t^6=(4/5)^6$.

Hint 2

Precision stability does not certify that the input formula is correct. The sign of the omitted tail supplies independent information.

Solution

Equation (D08.3) gives the exact bracket

$$\frac{21}{12500} \leq F_2-F \leq \frac{7}{1500},$$

or $0.00168\leq F_2-F\leq0.004666\ldots$. The positive reported discrepancy lies inside it; its size is explained by truncation. The precision repeat supports numerical stability but cannot reduce that truncation error. The negative discrepancy is impossible for the stated integral and polynomial. Check the subtraction order and normalization before changing the theory.

A controlled next task is: “At the same $t=4/5$, branch and integral precision, increase $p$ until the upper bound is below $10^{-8}$, and verify the integral remains inside the resulting interval.” The recurrence first certifies this at $p=23$. This is a sufficient certified order, not a claim that every lower order fails the actual tolerance.

3. Replace a failing bound and test a derivative

Show that $c_n\leq1/[8n(n+1)]$. Deduce a bound for $F_p-F$ valid on all of $0\leq t\leq1$, including the endpoint. Then differentiate (D08.1) for $t<1$ and examine the region near $\theta=0$ as $t=1-\delta\to1$. Does uniform convergence of the values justify a uniformly accurate derivative from a fixed polynomial?

Hint 1

The proposed coefficient bound is exact at $n=1$. Compare $c_{n+1}/c_n$ with $n/(n+2)$, and telescope $1/[n(n+1)]$.

Hint 2

Near the endpoint, $1-t\cos\theta=\delta+\theta^2/2+\cdots$. Integrating its inverse square root produces a logarithm.

Solution

The difference of the ratios is positive:

$$\frac{n}{n+2}-\frac{16n^2-1}{16(n+1)^2} =\frac{17n+2}{16(n+1)^2(n+2)}>0.$$

Starting from $c_1=1/16$, induction proves the coefficient bound. Since $t^{2n}\leq1$,

$$0\leq F_p(t)-F(t) \leq\frac18\sum_{n=p+1}^\infty \left(\frac1n-\frac1{n+1}\right) =\frac1{8(p+1)}. \tag{D08.6}$$

The summable coefficient majorant establishes uniform convergence of the series on the closed interval. Its endpoint agrees with the continuous integral by dominated convergence. This bound is slower than the geometric one away from the endpoint, but remains finite where that bound fails.

For $t<1$, differentiation is justified on each compact subinterval:

$$F'(t)=-\frac1{2\pi}\int_0^\pi \frac{\cos\theta}{\sqrt{1-t\cos\theta}}\,d\theta.$$

The small-$\theta$ contribution is $-[1/(2\pi)]\int_0^\eta d\theta/\sqrt{\delta+\theta^2/2}$ for a fixed small positive $\eta$. Its logarithmic term gives

$$F'(1-\delta)=\frac{\log\delta}{2\pi\sqrt2}+O(1) \longrightarrow-\infty. \tag{D08.7}$$

The difference between the exact integrand and this local approximation contributes a bounded remainder; the part with $\theta\geq\eta$ is regular. In contrast, each fixed $F_p'$ is finite at $t=1$; for example $F_2'(1)=-47/256$. Uniform value convergence therefore does not justify exchanging differentiation with the endpoint limit. A derivative project needs its own error analysis or an expansion adapted to that singularity.

The transferable result is a concrete research note: a known quantity, fixed conventions, a checked baseline, one changed input, and a test that separates possible explanations. Apply that sequence to your completed capstone before enlarging the physical question.

For a geometric first project, deform a supersymmetric torus while tracking its holomorphic data. For an exact-observable project, construct the A-twist and compare mode counting with a regulated free-chiral determinant.

Further reproductions test the full phase of a duality, boundary supersymmetry and gluing, or global regularity on a spindle. Each starts with an explicit model and a changed assumption whose effect can be checked.

Other starting points compare global gauge groups, extract a finite-angle defect response, or derive a protected bootstrap constraint. Choose the observable whose prerequisites you have completed, and keep its physical and numerical error tests separate.

For a spectral project, compare coordinate waves and operator mixing or ABJM spectral data and omitted exponential sectors. Their finite calculations distinguish a physical change from a truncation or normalization error.

To study source responses, compare deformed spheres and integrated CFT data with framed defect observables. For a spectral stability project, recover BPS indices across a wall while keeping the physical chamber and formal truncation distinct.

Two further projects connect four-dimensional dynamics to new mathematical descriptions: match an instanton coefficient to a quantum period or sum the finite-radius correction to a Coulomb-branch metric. Each fixes the boundary conditions or global periods before comparing formulas.