Asymptotic Scales, Remainders, Uniformity, and Optimal Truncation

An expansion is asymptotic only after its limiting process, scale, region, and remainder have been specified. For every fixed number of retained terms, the remainder must be smaller than the last retained scale in the stated limit. Uniformity strengthens this statement by requiring one estimate to work over a whole parameter set. Optimal truncation is stronger again: it needs information about how the remainder bound depends on the truncation order. The symbol $\sim$ by itself supplies none of that growing-order control.

This page develops those distinctions and then proves them in a zero-dimensional Euclidean $\phi^4$ integral. The example gives a large-parameter expansion, an exact finite-order remainder bound, factorially growing coefficients, and a justified least-term prescription for that bound. It is a controlled analogue of a perturbative field integral, not a theorem about continuum QFT.

Helpful background. Limits, Completeness, and Modes of Convergence supplies the pointwise and uniform convergence language used in remainder estimates.

Asymptotic scales and remainder data

Start with a dimensionless parameter $\epsilon\to0^+$; other limits can be reduced to this form, for example by setting $\epsilon=1/\Lambda$ when $\Lambda\to+\infty$. A complete asymptotic statement must say:

• how the parameter approaches its limit;
• which auxiliary variables are fixed and which are allowed to vary;
• the real interval or complex sector on which the estimate holds;
• the norm in which the remainder is measured;
• and, for complex powers or logarithms, which branch is used.

For functions $g$ and $h$, with $h$ nonzero sufficiently near the limit,

$$\begin{aligned} g(\epsilon)=o\!\left(h(\epsilon)\right) &\quad\Longleftrightarrow\quad \frac{g(\epsilon)}{h(\epsilon)}\longrightarrow0,\\ g(\epsilon)=O\!\left(h(\epsilon)\right) &\quad\Longleftrightarrow\quad \left|\frac{g(\epsilon)}{h(\epsilon)}\right| \text{ is bounded.} \end{aligned}$$

Little-$o$ is a limiting statement. Big-$O$ asserts the existence of a bound, but the hidden constant and the neighborhood on which it holds still matter. Neither notation is a numerical error estimate until those quantities are known.

An ordered family $\{\varphi_n\}_{n\geq0}$ is an asymptotic scale in the declared limit and region if

$$\varphi_{n+1}(\epsilon) =o\!\left(\varphi_n(\epsilon)\right) \qquad(n=0,1,2,\ldots).$$

The standard power scale is $1,\epsilon,\epsilon^2,\ldots$ as $\epsilon\to0$. At a logarithmic threshold, or in an exponential problem, a different scale may be more natural. The ordering is part of the claim; a list of functions is not an asymptotic scale independently of a limit and region.

Poincaré expansions are fixed-order statements

Define the $N$-term partial sum and its remainder by

$$\begin{aligned} S_N(\epsilon) &=\sum_{n=0}^{N-1}a_n\varphi_n(\epsilon),\\ R_N(\epsilon) &=f(\epsilon)-S_N(\epsilon). \end{aligned}$$

Then

$$f(\epsilon)\sim \sum_{n=0}^{\infty}a_n\varphi_n(\epsilon)$$

means that, for every fixed $N\geq1$,

$$R_N(\epsilon) =o\!\left(\varphi_{N-1}(\epsilon)\right) \qquad(\epsilon\to0^+).$$

The quantifiers are decisive: choose $N$, hold it fixed, and only then take the limit. Because the expansion is asserted to every order,

$$R_N(\epsilon) =a_N\varphi_N(\epsilon) +o\!\left(\varphi_N(\epsilon)\right),$$

so $R_N=O(\varphi_N)$ at each fixed order. For a finite expansion known only through $\varphi_{N-1}$, however, the definition gives only $R_N=o(\varphi_{N-1})$ unless a next-scale estimate is proved separately.

For a fixed nonvanishing scale and a fixed approach region, the coefficients are unique. They can be recovered recursively:

$$a_N= \lim_{\epsilon\to0^+} \frac{ f(\epsilon)-\displaystyle\sum_{n=0}^{N-1} a_n\varphi_n(\epsilon)} {\varphi_N(\epsilon)}.$$

The function represented by the coefficients is not unique. If $A>0$, then

$$e^{-A/\epsilon}=o(\epsilon^N) \qquad\text{for every fixed }N.$$

Thus $f(\epsilon)$ and $f(\epsilon)+C e^{-A/\epsilon}$ have the same asymptotic power expansion as $\epsilon\to0^+$. The extra term is beyond all algebraic orders. This observation alone does not identify a saddle, a new physical sector, or a summation prescription.

The fixed-order definition, coefficient uniqueness, and this beyond-all-orders nonuniqueness are developed in Hunter 2004, Chapter 2, pp. 19–27, PDF.

For complex $\epsilon$, even this elementary statement needs a sector. On the closed subsector

$$|\arg\epsilon|\leq\frac{\pi}{2}-\delta, \qquad 0<\delta<\frac{\pi}{2},$$

one has

$$\left|e^{-A/\epsilon}\right| \leq \exp\!\left( -\frac{A\sin\delta}{|\epsilon|} \right),$$

uniformly. On the boundary $\arg\epsilon=\pm\pi/2$, its modulus is $1$, so the exponentially small estimate fails. Sector boundaries are therefore mathematical data, not decorative qualifiers.

Asymptotic does not mean convergent

Convergence and asymptoticity take different limits:

• convergence studies $S_N(\epsilon)$ as $N\to\infty$ with $\epsilon$ fixed;
• asymptoticity studies $S_N(\epsilon)$ as $\epsilon\to0$ with $N$ fixed.

An asymptotic series may converge, but it need not. If it diverges for every nonzero $\epsilon$, its partial sums can still approximate $f$ increasingly well for several orders before eventually getting worse. Conversely, a convergent series centered at one point need not be useful asymptotically in a different limit.

Accordingly,

$$f(\epsilon)\sim\sum_{n=0}^{\infty}a_n\epsilon^n$$

does not assert equality to an infinite sum. The right-hand side is a formal record of a family of remainder statements unless convergence or some separate summation procedure has been established.

Uniformity names what is allowed to vary

Suppose $f(\epsilon,y)$ depends on an auxiliary parameter $y\in Y$. With a scale independent of $y$, the expansion is uniform on $Y$ if, for every fixed $N\geq1$,

$$\sup_{y\in Y} \left| \frac{R_N(\epsilon,y)} {\varphi_{N-1}(\epsilon)} \right| \longrightarrow0.$$

A common stronger result is a bound

$$\sup_{y\in Y}|R_N(\epsilon,y)| \leq C_{N,Y}|\varphi_N(\epsilon)|,$$

where the constant is independent of $y$. It may still depend strongly on $N$, on $Y$, or on the distance from a sector boundary.

The elementary function

$$F(\epsilon,y)=\frac{1}{1+\epsilon y}, \qquad \epsilon>0,\quad y\geq0,$$

exposes the distinction. The finite geometric identity gives

$$F(\epsilon,y) =\sum_{n=0}^{N-1}(-\epsilon y)^n +\frac{(-\epsilon y)^N}{1+\epsilon y}.$$

On every fixed interval $0\leq y\leq M$,

$$\sup_{0\leq y\leq M}|R_N(\epsilon,y)| \leq(\epsilon M)^N,$$

so the power expansion is uniform there. It is not uniform on $[0,\infty)$. Already at leading order,

$$\sup_{y\geq0} \left|F(\epsilon,y)-1\right|=1$$

for every $\epsilon>0$. The distinguished scaling $y=c/\epsilon$ makes the reason visible:

$$F(\epsilon,c/\epsilon)=\frac{1}{1+c},$$

which is not $1+o(1)$. Near such a transition, the original ordering of terms has ceased to be uniform and a rescaled or matched description is needed.

For the corresponding need for uniform integral approximations near transition points, see Temme 1995, pp. 395–399.

The same caution applies in the complex plane. An expansion may be uniform on every closed subsector while its constants diverge as the subsector approaches a boundary. Powers such as $z^\alpha$ and terms containing $\log z$ also require a declared branch. Uniform control of the function itself does not automatically justify termwise differentiation or integration over an unbounded domain; those operations need derivative estimates, domination, or an appropriate norm bound.

What controls a truncation error

Poincaré asymptoticity controls each fixed $N$ separately. It does not say how the constants in

$$|R_N(\epsilon)|\leq C_N|\varphi_N(\epsilon)|$$

grow with $N$. Consequently, it says nothing by itself about a choice $N=N(\epsilon)\to\infty$. A least-term rule becomes justified only when a theorem or an explicit calculation relates the remainder to the terms over the relevant growing range of $N$.

This distinction between a formal fixed-order series and a justified least-term prescription is illustrated in Mariño 2026, § 2.1, pp. 2–4, PDF.

It helps to separate three objects:

1. the actual optimal index, which minimizes $|R_N|$ but is usually unknown;
2. the index that minimizes a proved upper bound $B_N$;
3. the least term of the formal series, which is only a proxy unless linked to the remainder.

For example, suppose one has proved, uniformly in the required region and for the needed range of $N$,

$$|R_N(\epsilon)| \leq B_N(\epsilon) =K\,\Gamma(N+\beta)(A\epsilon)^N, \qquad K>0,\quad A>0.$$

Here $K$, $A$, and the real number $\beta$ are fixed, and $N+\beta>0$ throughout the range under consideration.

The ratio of consecutive bounds is

$$\frac{B_{N+1}}{B_N} =A\epsilon(N+\beta).$$

The smallest bound therefore lies near

$$N_*(\epsilon)\simeq \frac{1}{A\epsilon}-\beta.$$

Stirling’s formula then gives

$$B_{N_*} =O\!\left( (A\epsilon)^{1/2-\beta} e^{-1/(A\epsilon)} \right).$$

The exponentially small scale follows from the assumed $N$-dependent bound, not from the definition of an asymptotic expansion. Without such control, even decreasing terms can give a misleading picture of the remainder; NIST DLMF 2026, § 2.11(i) gives explicit warnings and counterexamples.

A controlled zero-dimensional scalar benchmark

Consider the normalized Euclidean integral

$$Z(\Lambda) = \sqrt{\frac{\Lambda}{2\pi}} \int_{\mathbb R} \exp\!\left[ -\Lambda\left( \frac{\phi^2}{2}+\frac{\phi^4}{24} \right) \right]\mathrm d\phi, \qquad \Lambda\to+\infty.$$

Here $\Lambda>0$ is dimensionless. After $x=\sqrt{\Lambda}\,\phi$,

$$Z(\Lambda) = \frac{1}{\sqrt{2\pi}} \int_{\mathbb R} e^{-x^2/2} e^{-x^4/(24\Lambda)} \mathrm dx.$$

Equivalently, if $X$ is a standard normal random variable,

$$Z(\Lambda) =\mathbb E\!\left[ e^{-X^4/(24\Lambda)} \right].$$

Expanding the second exponential through $N-1$ and evaluating the Gaussian moments gives

$$\begin{aligned} Z(\Lambda) &= \sum_{n=0}^{N-1} (-1)^n c_n\Lambda^{-n} +R_N(\Lambda),\\ c_n &= \frac{\mathbb E[X^{4n}]}{24^n n!} = \frac{(4n)!} {96^n n!(2n)!}. \end{aligned}$$

Thus

$$Z(\Lambda) \sim 1-\frac{1}{8\Lambda} +\frac{35}{384\Lambda^2} -\frac{385}{3072\Lambda^3} +\cdots.$$

This is not merely a formal integration. For $u\geq0$, Taylor’s integral remainder is

$$e^{-u} = \sum_{n=0}^{N-1}\frac{(-u)^n}{n!} + \frac{(-u)^N}{(N-1)!} \int_0^1 (1-s)^{N-1}e^{-su}\,\mathrm ds.$$

Substitute $u=X^4/(24\Lambda)$ and take the Gaussian expectation. The integrand in the remainder has a fixed sign and $e^{-su}\leq1$, so for every integer $N\geq1$ and every $\Lambda>0$,

$$0\leq (-1)^N R_N(\Lambda) \leq c_N\Lambda^{-N}.$$

This exact estimate proves the fixed-order asymptotic expansion and, in this example, bounds the error by the first omitted term.

Now let $T_n=c_n\Lambda^{-n}$ be the magnitude of the $n$th term. Directly from the factorial formula,

$$\frac{T_{n+1}}{T_n} = \frac{(4n+1)(4n+3)} {24(n+1)\Lambda} \sim \frac{2n}{3\Lambda}.$$

For every fixed $\Lambda$, the terms eventually grow, so the formal series diverges. The first-omitted-term bound is smallest near

$$N_*\sim\frac{3\Lambda}{2}.$$

Stirling’s formula makes the corresponding scale explicit:

$$c_n \sim \frac{1}{\pi\sqrt2} \left(\frac{2}{3}\right)^n \Gamma(n).$$

Choosing the nearest integer to $3\Lambda/2$ therefore gives

$$|R_{N_*}(\Lambda)| \leq T_{N_*} \sim \sqrt{\frac{2}{3\pi\Lambda}}\, e^{-3\Lambda/2}.$$

The final $\sim$ describes the upper bound $T_{N_*}$ as $\Lambda\to+\infty$; it does not assert that the actual remainder is asymptotic to that bound. This is precisely the extra evidence that the generic least-term slogan lacks.

The ordinary integral is often called zero-dimensional $\phi^4$ theory. More concretely, $\mathbb E[X^{4n}]=(4n-1)!!$ counts Gaussian Wick pairings, while $1/[n!(4!)^n]$ is the factor obtained by expanding $n$ indistinguishable quartic vertices. The coefficients therefore reproduce the combinatorics of perturbative vacuum diagrams, so the example is genuinely QFT-facing; compare Zinn-Justin 2021, § 7.4, p. 132. But it has no spacetime modes, ultraviolet limit, renormalization scheme, or physical observable beyond the finite integral. The interpretation of control parameters, saddle sectors, loop counting, and failure conditions belongs to Saddles, Control Parameters, and Loop Counting.

A reporting checklist

A trustworthy asymptotic approximation should make the following information recoverable:

1. Limit: Which dimensionless quantity tends to zero or infinity, and along what path?
2. Scale: Which ordered functions define “successively smaller”?
3. Region: Which interval, parameter set, or closed complex subsector is covered?
4. Uniformity: Which variables are fixed, and which vary under a supremum or norm?
5. Remainder: Is the claim little-$o$, big-$O$, or an explicit inequality?
6. Order dependence: Are constants known when $N$ grows, or only for each fixed $N$?
7. Branches and boundaries: Which branch is used, and where can the estimate fail?

If an “optimal” truncation is quoted, also state whether it minimizes the true error, a proved bound, or only the displayed term magnitudes.

Common pitfalls

Reading $\sim$ as an infinite equality. A Poincaré series records fixed-order remainder behavior. It need not converge or define a unique function.

Letting the truncation order drift silently. A proof for every fixed $N$ does not cover $N=N(\epsilon)$. Growing-order truncation requires bounds uniform in the relevant range of $N$.

Dropping the uniformity set. Pointwise validity at every fixed auxiliary parameter does not imply one estimate works over an unbounded or transition region. Test distinguished scalings in which the auxiliary parameter depends on the small parameter.

Treating hidden constants as universal. A constant in $O(\varphi_N)$ may depend on $N$, the parameter set, a branch, or the distance from a sector boundary.

Using the first omitted term without a theorem. The zero-dimensional example has a sign-definite integral remainder, so the rule is proved there. A generic asymptotic series does not inherit that bound.

Overinterpreting the benchmark. A finite-dimensional scalar integral illustrates perturbative combinatorics and large-order control; it does not establish continuum-QFT existence, renormalized error bounds, or physical nonperturbative sectors.

Check your understanding

1. Retrieve the definition

State the asymptotic-scale condition and the remainder condition for an $N$-term Poincaré expansion. Which quantity is held fixed in the defining limit?

Solution

The scale satisfies $\varphi_{n+1}=o(\varphi_n)$. With

$$S_N=\sum_{n=0}^{N-1}a_n\varphi_n, \qquad R_N=f-S_N,$$

the expansion means $R_N=o(\varphi_{N-1})$ for every fixed $N\geq1$. The truncation order is fixed while the asymptotic parameter approaches its limit; no $N\to\infty$ assertion follows.

2. Find information invisible to all powers

For $A>0$, prove that $e^{-A/\epsilon}=o(\epsilon^N)$ as $\epsilon\to0^+$ for every fixed $N$. What changes if $\epsilon$ approaches zero in the complex plane?

Solution

Set $t=A/\epsilon$. Then

$$\epsilon^{-N}e^{-A/\epsilon} =A^{-N}t^Ne^{-t}\longrightarrow0 \qquad(t\to+\infty).$$

For complex $\epsilon$, decay along an approach requires $\operatorname{Re}(A/\epsilon)\to+\infty$. This holds uniformly on any closed subsector $|\arg\epsilon|\leq\pi/2-\delta$, but it fails on the imaginary-axis boundary. Hence the power coefficients can be unique even though the represented function is not.

3. Diagnose nonuniformity

For

$$F(\epsilon,y)=\frac{1}{1+\epsilon y},$$

derive the exact $N$-term geometric remainder. Decide whether the expansion is uniform on $[0,M]$ and on $[0,\infty)$.

Solution

The finite identity is

$$F(\epsilon,y) =\sum_{n=0}^{N-1}(-\epsilon y)^n +\frac{(-\epsilon y)^N}{1+\epsilon y}.$$

For fixed $M$, the remainder is bounded by $(\epsilon M)^N$, so the expansion is uniform on $[0,M]$. On $[0,\infty)$, the leading error has supremum $1$ for every $\epsilon>0$, so uniformity fails. The scaling $y=c/\epsilon$ keeps $\epsilon y$ of order one and exposes the transition.

4. Transfer the method to the scalar benchmark

Rescale the zero-dimensional integral, compute $c_1$ and $c_2$, derive $T_{n+1}/T_n$, and locate the least first-omitted-term bound.

Solution

The substitution $x=\sqrt{\Lambda}\phi$ turns the integral into a standard-normal expectation. From

$$c_n=\frac{(4n)!}{96^n n!(2n)!},$$

one obtains

$$c_1=\frac18, \qquad c_2=\frac{35}{384}.$$

For $T_n=c_n\Lambda^{-n}$,

$$\frac{T_{n+1}}{T_n} = \frac{(4n+1)(4n+3)} {24(n+1)\Lambda} \sim\frac{2n}{3\Lambda}.$$

The bound is therefore least near $N_*\sim3\Lambda/2$. Taylor’s sign-definite integral remainder is what converts this term calculation into a rigorous error bound. The conclusion applies to this regulated zero-dimensional model, not automatically to continuum QFT.

Where the method continues

This page supplied a criterion and an error discipline, not a mechanism for deriving coefficients from a general integral. Laplace Method and Steepest Descent develops that mechanism, including saddle selection and contour geometry. Stationary Phase, Coalescing Saddles, and Stokes Phenomena treats oscillatory integrals and changing critical-point structure, while WKB, Eikonal Expansions, and Turning Points carries the same error discipline into differential equations. Large-Order Growth and the Borel Transform develops the later summation questions that a divergent series raises.

References

• John K. Hunter, Lecture Notes on Asymptotics, Chapter 2: “Asymptotic Expansions”, PDF, University of California, Davis (2004), pp. 19–27. Definitions, coefficient uniqueness, beyond-all-orders nonuniqueness, convergence versus asymptoticity, and nonuniform limits.

• Marcos Mariño, “Resurgent Methods in String Theory”, PDF, 2026 ICTP lecture notes, §2.1, pp. 2–4. Formal asymptotic series, optimal truncation, and the quartic-integral example.

• NIST Digital Library of Mathematical Functions, version 1.2.7, released June 15, 2026, National Institute of Standards and Technology, §2.1 “Definitions and Elementary Properties” and §2.11(i) “Numerical Use of Asymptotic Expansions”. Structural definitions, uniformity conventions, and limitations of least-term error estimates.

• N. M. Temme, “Uniform Asymptotic Expansions of Integrals: A Selection of Problems”, Journal of Computational and Applied Mathematics 65 (1995), 395–417; especially pp. 395–399. Remainders, least-term truncation, and the need for uniform approximations near transitions.

• Jean Zinn-Justin, Quantum Field Theory and Critical Phenomena, 5th ed., Oxford University Press (2021), §7.4, p. 132. Zero-dimensional $\phi^4$ integrals and perturbative diagram weights.