Exact Statements, Saddle Expansions, and Conditional Derivations

The word exact applies to a specified statement, not automatically to every interpretation attached to it. A localization identity may determine a protected boundary observable at finite coupling and finite rank, while reading that answer as a string amplitude still assumes a holographic dictionary. Conversely, a bulk saddle can reproduce genuinely coupling-dependent dynamics while controlling only a finite large-$N$, strong-coupling expansion. This page gives a reproducible method for keeping those logical layers separate.

Required background. Holographic Duality: Claims, Dictionaries, and Regimes fixes the theory pair, observable, and parameter map to which each statement refers.

Helpful background. Effective Field Theory as a Controlled Expansion and EFT Truncation Errors and Breakdown Diagnostics explain how a remainder turns a formal series into a controlled prediction; Numerical Self-Consistency and Error Budgets supplies the corresponding numerical check.

The logical type of a calculation

Suppose a boundary quantity $F_{\mathrm B}$ is compared with a bulk calculation. Four statement types can appear in the same argument, but they require different evidence and license different conclusions.

Exact model identity. Every object is independently defined in a stated model and the equality follows without truncating an expansion or leaving an uncontrolled step. “Exact” still retains the model, observable, global form, boundary conditions, and normalization among its hypotheses.

Protected-sector result. A symmetry, index, cohomology class, or localization supercharge restricts the observable enough to make a calculation exact. The conclusion concerns that protected sector. It does not say that a generic correlator, spectrum, or real-time process is protected.

Saddle or asymptotic result. The computation retains finitely many terms and carries an explicit remainder. With two control parameters, write a finite statement such as

$$\begin{aligned} F(N,\lambda) &=F_{G,\mathbf K}^{\mathrm{trunc}}(N,\lambda) +R_{G,\mathbf K}(N,\lambda),\\ F_{G,\mathbf K}^{\mathrm{trunc}}(N,\lambda) &=\sum_{g=0}^{G}\frac1{N^{2g}} \sum_{k=0}^{K_g}a_{gk}\lambda^{-p_{gk}}. \end{aligned}$$

Both sums terminate, and $R_{G,\mathbf K}$ contains every omitted contribution. A power counting may estimate its first omitted algebraic term, but it does not by itself bound omitted saddles, exponentially small sectors, Stokes ambiguities, or a nonuniform kinematic limit.

The displayed even powers of $1/N$ illustrate an oriented closed-string genus expansion. Boundaries, unoriented sectors, different large-$N$ limits, or differently normalized observables can organize their powers in another way; the finite-truncation and remainder logic is unchanged.

Conditional derivation. If a dictionary $\mathfrak D$ and a bulk approximation $A$ are premises, the logical output is

$$(\mathfrak D\ \text{and}\ A)\Longrightarrow P_{\mathrm{bulk}}.$$

The derivation can be flawless while the conclusion remains conditional. It proves neither $\mathfrak D$ nor the exact bulk statement obtained by deleting $A$.

A reproducible claim-classification method

The method is deliberately mechanical so that another reader can reproduce both the calculation and its claim ceiling.

1. Inputs. Fix the theory and global form, state or ensemble, boundary conditions, observable and normalization, kinematic domain, renormalization prescription, control parameters, order of limits, and every imported dictionary entry. Identify which inputs are definitions, proven results, or conjectures.
2. Procedure. Separate same-description operations from cross-dictionary steps. Inventory every contribution at the requested order, compute only a finite truncation, and attach a named remainder to each expansion. Propagate each premise to every downstream conclusion.
3. Output. Report the calculated quantity together with its domain, retained orders, remainder model, logical status, and strongest licensed interpretation. An equation without this surrounding record is not the output of the method.
4. Validation. Check dimensions, normalization, symmetries and Ward identities, analytic structure, a solvable limit, and stability under adding the next available order. A numerical residual and a truncation uncertainty are separate entries.
5. Independent comparison. Compare with a route that changes the computational mechanism—for example, localization versus a worldsheet saddle, or an integrated localization constraint versus a flat-space string-amplitude limit. Record shared premises so that two consequences of one dictionary are not miscounted as two proofs of that dictionary.
6. Stop rule. Stop or downgrade the claim when the expansion parameter is not small, the result is nonuniform in the stated kinematics, a competing saddle is unsuppressed, a required same-order term is missing, an independent check fails, or no defensible remainder can be assigned. Crossing a stop rule calls for a new calculation, not a larger adjective.

The practical difficulty is set by the first unresolved structure. Localization can make the final integral elementary while the localization proof is highly nontrivial; an unprotected correlator can require loop Witten diagrams, higher-derivative vertices, mixing data, and analytic continuation before its next order is complete.

Reusable claim and approximation ledger

Field	What to record	Acceptance question
Object	Theory, state, observable, normalization, and scheme	Could another researcher construct the same quantity?
Domain	Kinematics, boundary data, global form, and admitted sectors	Is every displayed equality evaluated on the same domain?
Exact input	Identity, theorem, Ward identity, or protected reduction with hypotheses	Which step, precisely, has no approximation?
Approximation	Control parameters, finite retained orders, and order of limits	Are all contributions of the claimed order present?
Remainder	First omitted terms, omitted saddles or sectors, and whether the estimate is a bound	What could change the answer without changing the retained coefficients?
Imported premise	Dictionary entry, analytic continuation, contour, or model identification	Would the conclusion survive if this premise were removed?
Validation	Internal checks and their quantitative tolerances	What observation would reveal a sign, normalization, or domain error?
Independent comparison	Second mechanism, plus assumptions shared with the first	Is the comparison evidentially independent at the claimed level?
Claim ceiling and stop rule	Strongest licensed sentence and the first condition that forces a downgrade	Does the prose remain no stronger than its least-controlled step?

Protected application: the half-BPS circular Wilson loop

Work in Euclidean $\mathcal N=4$ super-Yang–Mills theory with gauge group $\mathrm{SU}(N)$ and ‘t Hooft coupling $\lambda=g_{\mathrm{YM}}^2N$. For a circle $C$ and a fixed unit vector $n^I$ in the scalar-coupling space, the normalized fundamental half-BPS loop is

$$W_\circ =\frac1N\operatorname{tr}\,\mathcal P \exp\!\oint_C d\tau\, \left(iA_\mu\dot x^\mu+\lvert\dot x\rvert n^I\Phi_I\right).$$

The special contour and scalar coupling preserve enough supersymmetry for localization. The four-dimensional path integral reduces to a Gaussian matrix integral; in this specified theory and normalization, the result is exact at finite $N$ and finite $\lambda$. Pestun proves the localization statement and the absence of additional one-loop and instanton factors in the $\mathcal N=4$ case in Pestun 2012, §1, eqs. (1.1)–(1.4), and §§3–5, PDF. Evaluating the matrix integral gives

$$\begin{aligned} \left\langle W_\circ\right\rangle_{\mathrm{SU}(N)} &=\frac1N L_{N-1}^{(1)}\!\left(-\frac{\lambda}{4N}\right) \exp\!\left[\frac{\lambda}{8N}\left(1-\frac1N\right)\right],\\ \left\langle W_\circ\right\rangle_{\mathrm{planar}} &=\lim_{N\to\infty}\left\langle W_\circ\right\rangle =\frac{2I_1(\sqrt\lambda)}{\sqrt\lambda}. \end{aligned}$$

Here $L_{N-1}^{(1)}$ is a generalized Laguerre polynomial and $I_1$ is a modified Bessel function. The finite-$N$ $\mathrm{U}(N)$ result and its $\mathrm{SU}(N)$ trace-factor correction are given in Drukker and Gross 2001, §II.B, eqs. (2.25)–(2.27), and §V, eq. (5.1), printed pp. 2903–2904 and 2910. The equality is broad in $N$ and $\lambda$ but narrow in observable: a generic Wilson loop is not fixed by this formula.

For comparison with a weakly curved string worldsheet, take the planar result first and then $\lambda\to\infty$. A finite Poincaré truncation is

$$\left\langle W_\circ\right\rangle_{\mathrm{planar}} =\sqrt{\frac2\pi}\,\lambda^{-3/4}e^{\sqrt\lambda} \left[ 1-\frac{3}{8\sqrt\lambda}-\frac{15}{128\lambda} +R_W^{(2)}(\lambda) \right], \qquad R_W^{(2)}=\mathcal O(\lambda^{-3/2}).$$

This is not an infinite convergent series: its role is to approximate the exact Bessel function at a declared order on the positive real $\lambda$ axis. The exact expression supplies an unusually strong validation target—one can measure the actual truncation error instead of inferring it from the last term. Drukker and Gross explicitly discuss the asymptotic, non-Borel-summable strong-coupling expansion in Drukker and Gross 2001, §V, eq. (5.2), printed pp. 2910–2911.

The holographic step is logically separate. The proposed dictionary maps $W_\circ$ to a fundamental-string path integral ending on the boundary circle. In the regime $1\ll\lambda\ll N$, curvature and string-loop corrections are both suppressed because $\alpha'/L^2=\lambda^{-1/2}$ and $g_s=\lambda/(4\pi N)$. The renormalized classical disk has $S_{\mathrm{F1,cl}}=-\sqrt\lambda$, so its saddle gives $e^{\sqrt\lambda}$ and reproduces the leading term in $\log\langle W_\circ\rangle$ Drukker, Gross, and Ooguri 1999, §§II–III, PDF. The factor $\lambda^{-3/4}$ and later terms require fluctuation determinants, zero-mode measures, and higher worldsheet orders; their calculation is substantially more delicate. A ratio method with explicit zero-mode treatment recovers the next-to-leading field-theory result in Li and Medina-Rincon 2020, §§2–5, especially eqs. (5), (11), (32)–(46), and the matching in eqs. (45)–(46), PDF.

The licensed conclusion is therefore two-part: the gauge-theory formula is exact for the protected loop, while the string interpretation and each semiclassical worldsheet truncation remain conditional on the dictionary and their stated regimes. Exactness does not cross that logical boundary automatically.

Unprotected application: a dynamical four-point function

Now consider $\mathrm{SU}(N)$ $\mathcal N=4$ super-Yang–Mills and the dimension-two half-BPS scalar primary $S_2(x,Y)$ in the stress-tensor multiplet, normalized by

$$\left\langle S_2(x_1,Y_1)S_2(x_2,Y_2)\right\rangle =\frac{(Y_1\!\cdot Y_2)^2}{x_{12}^4}, \qquad Y_i\!\cdot Y_i=0.$$

The external dimensions are protected, but the unconstrained dynamical function in the four-point correlator is not. Calling the entire four-point function “protected” would confuse protected external data with coupling-dependent operator exchanges and OPE coefficients. Superconformal Ward identities isolate the dynamical function, whose reduced Mellin amplitude will be denoted $M_2(s,t)$; the definitions and normalization are fixed in Binder et al. 2019, §§2.1–2.2, eqs. (2.3)–(2.18), printed pp. 5–8.

Let $c=(N^2-1)/4$ and $u=4-s-t$. Take $N\to\infty$ at fixed $\lambda$, then take $\lambda\to\infty$, while holding the Mellin variables fixed and away from poles. At large but finite $N$, the corresponding weak-string-coupling window also requires $g_s=\lambda/(4\pi N)\ll1$. Through the first two stringy contact terms, the finite result in that paper’s convention is

$$\begin{aligned} M_2(s,t) &=\frac8c\Bigg[ \frac{1}{(s-2)(t-2)(u-2)} +\frac{15\zeta(3)}{\lambda^{3/2}}\\ &\qquad +\frac{315\zeta(5)}{4\lambda^{5/2}} \left(s^2+t^2+u^2-3\right) +R_\lambda(s,t;\lambda) \Bigg] +R_c(s,t;N,\lambda),\\ R_\lambda&=\mathcal O(\lambda^{-3}), \qquad R_c=\mathcal O(c^{-2})\quad\text{at fixed }\lambda. \end{aligned}$$

The first term is tree-level type-IIB supergravity on $\mathrm{AdS}_5\times S^5$. The $\lambda^{-3/2}$ and $\lambda^{-5/2}$ terms are the tree-level $R^4$ and $D^4R^4$ corrections; the absence of a $\lambda^{-2}$ term is itself a nontrivial result. The displayed formula and coefficients appear in Binder et al. 2019, §4, eqs. (4.1)–(4.13), printed pp. 18–21. The remainder labels express the expected next power-counting orders at fixed Mellin variables, not rigorous global bounds.

Several checks are visible without recomputing the correlator. The amplitude is dimensionless. Its displayed terms are symmetric under permutations of $s,t,u$, as required for four identical operators. The supergravity exchange term has poles, while the higher-derivative contact interactions give polynomials. At $R^4$ order, the localization constraint fixes the coefficient and the independent flat-space string-amplitude limit gives the same value; their agreement is a precision check beyond supergravity. At $D^4R^4$ order, neither route alone fixes the displayed $p=2$ coefficient: combining the localization constraint with the flat-space constraint yields it Binder et al. 2019, §§4.1–4.2, eqs. (4.5), (4.12), and (4.13), printed pp. 19–21.

Those checks do not make $M_2$ exact. At finite $N$, bulk loops enter through $R_c$, and their fixed-$\lambda$ estimate need not remain uniform as $\lambda$ grows. At finite $\lambda$, additional higher-derivative and nonperturbative sectors enter through $R_\lambda$ and terms invisible to its algebraic power counting. At large Mellin variables, the polynomial corrections grow and the fixed-kinematics derivative expansion is nonuniform. The calculation must stop when $g_s$ or $1/c$ is not small, when $\lambda^{-1/2}$ does not suppress successive terms, or when $\max(|s|,|t|,|u|)/\sqrt\lambda$ is no longer small. The strongest surviving statement is then a tree-level, low-Mellin, large-$N$, strong-coupling prediction—not an exact correlator.

The two applications in one ledger

Claim and approximation ledger for the protected loop and dynamical four-point function

Ledger field	Circular Wilson loop	Stress-tensor-multiplet four-point function
Boundary object	Normalized fundamental half-BPS circle in Euclidean SU(N) maximally supersymmetric Yang–Mills theory	Dynamical reduced Mellin amplitude of four normalized S₂ primaries
Exact input	Localization to a Gaussian matrix model and its finite-N, finite-λ evaluation	Superconformal kinematics, Ward identities, and protected two-point normalization
Approximation used here	Planar limit followed by a finite strong-coupling truncation for the worldsheet comparison	Leading inverse-c term through the first two stringy contact corrections at fixed Mellin variables
Imported dictionary	Boundary loop maps to a fundamental string ending on the same contour	Single-trace exchanges and Mellin contact terms map to supergravity and higher-derivative string interactions
Independent comparison	Exact matrix model versus the classical worldsheet and a one-loop ratio calculation	Integrated localization constraint versus flat-space string-amplitude limit
Uncertainty	No boundary-theory truncation in the exact formula; finite worldsheet, genus, and strong-coupling remainders after mapping	Higher-derivative, bulk-loop, nonperturbative, and nonuniform-kinematics effects
Licensed claim	Exact protected boundary observable; conditional holographic interpretation; controlled saddle statement at displayed order	Conditional unprotected correlator prediction through the displayed large-N and strong-coupling orders
Stop rule	Do not use the semiclassical string reading when curvature, string loops, or a required determinant are uncontrolled	Do not extrapolate through large Mellin variables or beyond the regime in which successive inverse-c and inverse-square-root-of-λ orders decrease

The contrast is the main lesson. The protected result is exact but probes a narrow observable. The unprotected correlator probes broader interacting dynamics but is only a controlled finite approximation. Neither fact ranks one as a proof of the complete duality.

Adversarial tests for overclaimed exactness

Invisible exponential sector. Let a calculation determine every displayed algebraic coefficient of $F_{\mathrm{alg}}(N,\lambda)$ and consider

$$\widetilde F(N,\lambda) =F_{\mathrm{alg}}(N,\lambda) +e^{-\kappa N}G(\lambda), \qquad \kappa>0.$$

For every fixed integer $m\geq0$,

$$\lim_{N\to\infty}N^m e^{-\kappa N}=0.$$

The added sector changes no coefficient in any algebraic $1/N$ test. Unless a theorem, exact definition, or separate nonperturbative calculation forces $G=0$, the strongest surviving conclusion is agreement through the controlled algebraic sector. The downgrade is caused by an unresolved alternative that the expansion cannot see.

Noncommuting limits. The elementary control function

$$q(N,\lambda)=\frac{N}{N+\lambda}$$

obeys

$$\lim_{N\to\infty}\lim_{\lambda\to\infty}q=0, \qquad \lim_{\lambda\to\infty}\lim_{N\to\infty}q=1.$$

This is not proposed as a holographic observable. It identifies the missing proof obligation: interchanging the large-$N$ and strong-coupling limits requires a uniform estimate in the parameter held back. Without one, the licensed statement is the ordered limit actually calculated, and the downgrade is caused by nonuniform convergence.

Common classification failures

Exact boundary answer, exact duality. An exact answer in one independently defined theory can be compelling evidence for a dictionary entry. It becomes an exact bulk result only conditionally on that dictionary unless the bulk object and the map have also been constructed and proved.

First omitted term, rigorous bound. Power counting supplies an expected scale when its coefficient assumptions and domain hold. It is not a deterministic bound unless those assumptions are themselves bounded and the relevant tail is controlled.

Protected external operators, protected correlator. The dimension and two-point normalization of $S_2$ are protected. Its separated four-point dynamical function contains unprotected coupling-dependent data.

Two calculations, two independent tests. Localization and a flat-space string limit use different computational mechanisms, but a holographic interpretation can remain common to both. Independence must be assessed at the level of the claim being tested.

Exercises

1. Weak-coupling normalization of the planar loop. Use

$$I_1(z)=\frac z2+\frac{z^3}{16}+\frac{z^5}{384}+\mathcal O(z^7)$$

to expand the exact planar Wilson loop through order $\lambda^2$.

Solution

Substitute $z=\sqrt\lambda$ into $2I_1(z)/z$:

$$\frac{2I_1(\sqrt\lambda)}{\sqrt\lambda} =1+\frac{\lambda}{8}+\frac{\lambda^2}{192} +\mathcal O(\lambda^3).$$

The constant term verifies the normalized trace convention, and the first two coefficients provide a weak-coupling check independent of the large-$\lambda$ truncation.

2. Classify a finite correlator truncation. A calculation reproduces the three displayed terms of $M_2$ but neither computes $R_c$ nor proves a bound on $R_\lambda$. What is the strongest justified statement?

Solution

It establishes the reduced Mellin amplitude at leading order in $1/c$ through order $\lambda^{-5/2}$, in the declared ordered limit and fixed low-Mellin domain. It does not establish the finite-$N$, finite-coupling correlator or a numerical coverage interval. The first omitted higher-derivative order and the $c^{-2}$ bulk-loop sector remain conditional uncertainty estimates.

3. Test an exactness claim. Show why agreement of all algebraic $1/N$ coefficients cannot distinguish $F$ from $F+e^{-\kappa N}G(\lambda)$, and state one kind of evidence that could.

Solution

Because $N^m e^{-\kappa N}\to0$ for every fixed $m$, the difference is smaller than every power of $1/N$ and contributes to no algebraic coefficient. An exact finite-$N$ identity, a convergent representation with a uniqueness theorem, or a separate nonperturbative calculation constraining $G$ could distinguish the two. More perturbative coefficients alone cannot.

Evidence cutoff. Literature-status examples use a cutoff of 25 July 2026; no current confidence assessment is implied. Generic expansion theory remains owned by the linked EFT pages, protected-sector techniques by the supersymmetry and duality treatment, and theorem-level existence or equivalence claims by Mathematical QFT.

The chapter overview contains the structure diagram and validity and failure diagram. They are embedded there once so that their shared chapter-level context is not repeated on every article.

For the chapter-wide comparison of assumptions, counterevidence, falsifiers, and claim ceilings, see the claim-domain table.

References

• Binder, Damon J., Shai M. Chester, Silviu S. Pufu, and Yifan Wang. 2019. “$\mathcal N=4$ Super-Yang–Mills Correlators at Strong Coupling from String Theory and Localization.” Journal of High Energy Physics 2019 (12), 119. DOI. Open PDF.
• Drukker, Nadav, and David J. Gross. 2001. “An Exact Prediction of $\mathcal N=4$ Supersymmetric Yang–Mills Theory for String Theory.” Journal of Mathematical Physics 42 (7), 2896–2914. DOI. Open PDF.
• Drukker, Nadav, David J. Gross, and Hirosi Ooguri. 1999. “Wilson Loops and Minimal Surfaces.” Physical Review D 60, 125006. DOI. Open PDF.
• Li, Botao, and Daniel Medina-Rincon. 2020. “On Precision Holography for the Circular Wilson Loop in $\mathrm{AdS}_5\times S^5$.” Physics Letters B 810, 135789. DOI. Open PDF.
• Pestun, Vasily. 2012. “Localization of Gauge Theory on a Four-Sphere and Supersymmetric Wilson Loops.” Communications in Mathematical Physics 313, 71–129. DOI. Open PDF.