Case Study 2: When Perturbation Theory Fails — Strong Coupling and Non-Perturbative Physics

The Physical Situation

Perturbation theory is the most widely used computational tool in quantum mechanics. But it is not omnipotent. There are physically important phenomena that no finite order of perturbation theory can capture — phenomena that are, in a precise mathematical sense, invisible to the perturbation expansion. Understanding where perturbation theory fails is just as important as knowing how to use it.

This case study examines three classes of failure, each with deep physical significance:

1. Strong coupling: The perturbation is not small, and the series diverges rapidly.
2. Asymptotic divergence: The series has zero radius of convergence even for arbitrarily small perturbations.
3. Non-perturbative effects: Physical phenomena that are qualitatively absent from any truncation of the perturbation series.

Case A: The Quartic Oscillator at Strong Coupling

The Problem

Consider the anharmonic oscillator in natural units ($m = \omega = \hbar = 1$):

$$\hat{H} = \frac{1}{2}\hat{p}^2 + \frac{1}{2}\hat{x}^2 + \lambda\hat{x}^4$$

The perturbation expansion for the ground-state energy is:

$$E_0(\lambda) = \frac{1}{2} + \frac{3}{4}\lambda - \frac{21}{8}\lambda^2 + \frac{333}{16}\lambda^3 - \frac{30885}{128}\lambda^4 + \cdots$$

The coefficients grow factorially: $|a_k| \sim k!$, which means the series diverges for any $\lambda > 0$.

How Bad Is It?

For $\lambda = 0.01$ (weak coupling), the terms decrease at first:

The series gives six correct digits after just four terms — excellent agreement despite formal divergence.

For $\lambda = 1.0$ (strong coupling):

The series is oscillating wildly and diverging. Perturbation theory is useless here.

The Lesson

The perturbation expansion is a local tool — it captures the behavior of $E_0(\lambda)$ near $\lambda = 0$. For finite $\lambda$, it works only if enough terms are available before the factorial growth takes over. The smaller $\lambda$ is, the more terms you can use before the series deteriorates.

Case B: The Asymptotic Nature of Perturbation Series

Why Zero Radius of Convergence?

The zero radius of convergence for the quartic oscillator can be understood by a beautiful physical argument due to Dyson (1952).

Consider what happens if $\lambda < 0$. The potential becomes $V(x) = \frac{1}{2}x^2 + \lambda x^4$ with a negative quartic term. For large $|x|$, $V(x) \to -\infty$: the potential is unbounded below. This means:

• For $\lambda > 0$: The potential is confining, the spectrum is discrete, and $E_0(\lambda)$ is a well-defined real number.
• For $\lambda < 0$: No matter how small $|\lambda|$ is, the particle can tunnel through the potential barrier near $x = 0$ and escape to $x = \pm\infty$. There are no stable bound states.

Therefore $E_0(\lambda)$ is not analytic at $\lambda = 0$ — the function has different character for $\lambda > 0$ and $\lambda < 0$. A power series in $\lambda$ has a radius of convergence determined by the nearest singularity to the origin in the complex $\lambda$-plane. Since the singularity is at $\lambda = 0$ (in the direction $\text{Re}\,\lambda < 0$), the radius of convergence is zero.

The Optimal Truncation

Despite divergence, the asymptotic series is useful if truncated optimally. The rule of thumb is: stop adding terms when the terms start growing. For the quartic oscillator:

$$N^*(\lambda) \approx \frac{1}{3\lambda}$$

For $\lambda = 0.01$: $N^* \approx 33$. The error at optimal truncation is approximately $e^{-1/(3\lambda)} \approx e^{-33} \approx 10^{-15}$ — extraordinarily precise.

For $\lambda = 0.1$: $N^* \approx 3$. The error is $\sim e^{-3.3} \approx 0.04$ — about 4%. The window of usefulness is narrow.

For $\lambda = 1$: $N^* \approx 0$. Even the first correction makes things worse. Perturbation theory is useless.

The Stokes Phenomenon

Something mathematically subtle happens when $\lambda$ rotates in the complex plane from positive to negative values. The asymptotic expansion changes discontinuously — certain exponentially small terms "switch on" as one crosses specific lines in the complex plane (Stokes lines). This is the Stokes phenomenon, and it connects the perturbative coefficients to the non-perturbative physics. Understanding this connection is the subject of resurgence, an active area of mathematical physics.

Case C: Tunneling and Non-Perturbative Effects

The Double Well

Consider a particle in the symmetric double-well potential:

$$V(x) = g(x^2 - a^2)^2$$

For large $g$, each well is approximately harmonic with frequency $\omega = 2a\sqrt{2g/m}$. The low-lying energy levels come in nearly degenerate pairs: a symmetric state $|+\rangle$ and an antisymmetric state $|-\rangle$ with:

$$E_{\pm} = E_0 \pm \frac{\Delta}{2}$$

The splitting $\Delta$ is due to quantum tunneling through the barrier between the wells. In the WKB approximation:

$$\Delta \propto \omega\, e^{-S_{\text{inst}}/\hbar}$$

where $S_{\text{inst}} = \int_{-a}^{a} \sqrt{2mV(x)}\,dx$ is the instanton action — the classical action for the trajectory that traverses the barrier in imaginary time.

For the double-well potential:

$$S_{\text{inst}} = \frac{4}{3}ga^3\sqrt{2m} \propto \frac{1}{g} \cdot (\text{const})$$

Wait — let us be more careful. If we define the perturbation parameter as $\lambda = 1/g$ (weak coupling corresponds to large $g$, i.e., tall barrier), then:

$$\Delta \propto e^{-c/\lambda}$$

for some constant $c > 0$. This is non-perturbative: the function $e^{-c/\lambda}$ has a Taylor series $\sum_k a_k \lambda^k = 0$ at $\lambda = 0$ (all derivatives vanish). No finite number of terms in the perturbation series can approximate this exponential — it is identically zero at every order.

What Does This Mean?

It means that perturbation theory, applied to a single well, gives the energy levels of a particle trapped in one well. It correctly computes the zero-point energy, the anharmonic corrections, the shifts due to the shape of the bottom of the well — everything that depends analytically on the coupling constant. But it completely misses the fact that the particle can tunnel to the other well, producing a splitting of the energy levels.

This is not an academic curiosity. Non-perturbative effects include:

In each case, the physical effect is invisible to perturbation theory yet crucial for understanding the physics.

Resurgence: Connecting Perturbative and Non-Perturbative

The Modern Perspective

Since the late 2000s, the mathematical framework of resurgence has provided a systematic way to connect perturbative and non-perturbative contributions. The key insight is that the divergence pattern of the perturbation series (specifically, the large-order behavior of the coefficients) encodes information about the non-perturbative effects.

For the quartic oscillator, the large-order behavior of the perturbation coefficients is:

$$a_k \sim (-1)^{k+1} \frac{3^k \Gamma(k + 1/2)}{\pi^{3/2}}$$

The factor $3^k$ is related to the instanton action: $S_{\text{inst}} = 1/(3\lambda)$ (in appropriate units). This is not a coincidence — resurgence theory proves that the perturbative and non-perturbative sectors communicate, and the growth rate of the perturbation series directly determines the strength of the leading non-perturbative effect.

Borel Summation

One practical application is Borel summation, which can sometimes extract the correct answer from a divergent series:

1. Define the Borel transform: $\tilde{E}(t) = \sum_{k=0}^\infty \frac{a_k}{k!} t^k$. This series typically has a finite radius of convergence (because dividing by $k!$ tames the factorial growth).

2. Recover the original function by the Borel integral: $E_0(\lambda) = \int_0^\infty e^{-t} \tilde{E}(\lambda t)\,dt$.

For the quartic oscillator, $\tilde{E}(t)$ has singularities on the negative real $t$-axis (at $t = -1/3$), so the Borel integral along the positive real axis is well-defined and gives the correct answer.

This technique has been verified numerically to remarkable precision: Borel-resummed perturbation theory for the quartic oscillator reproduces the exact numerical eigenvalues to 50+ significant digits.

Practical Guidelines: When to Trust Perturbation Theory

Based on the examples in this case study, here are guidelines for assessing the reliability of perturbation theory:

Trust perturbation theory when:

1. The perturbation parameter is small: $\lambda \ll 1$ in appropriate units.
2. The matrix elements are small compared to level spacings: $|\langle m|\hat{H}'|n\rangle| \ll |E_n - E_m|$ for all relevant states.
3. The first few corrections decrease in magnitude: $|E^{(2)}| < |E^{(1)}|$ (or $|E^{(2)}|$ is reasonably small if $E^{(1)} = 0$).
4. The physics does not change qualitatively: The perturbation does not open new channels, create new bound states, or change the topology of the potential.

Be skeptical when:

1. The series oscillates with growing magnitude: Alternating-sign terms of increasing size signal asymptotic divergence.
2. Near-degeneracies are present: Small denominators amplify corrections.
3. The perturbation changes the potential qualitatively: For example, turning a confining potential into an unconfining one.
4. Tunneling or decay processes are important: These are non-perturbative by nature.
5. The coupling constant is of order unity or larger: Strong coupling requires non-perturbative methods (variational, numerical, lattice, etc.).

Discussion Questions

1. The perturbation series for the electron's anomalous magnetic moment ($g-2$) in QED has been computed to fifth order ($\alpha^5$, involving 12,672 Feynman diagrams). The series appears to converge beautifully, despite the expectation that it is asymptotic. Estimate how many orders would need to be computed before the series begins to diverge, given that the fine structure constant is $\alpha \approx 1/137$. (Hint: $N^* \sim c/\alpha$ for some $c$ of order unity.)

2. In quantum chromodynamics (QCD), the strong coupling constant $\alpha_s \sim 1$ at low energies. Explain why perturbation theory is useless for computing hadron masses, and why non-perturbative methods (lattice QCD) are essential.

3. The BCS theory of superconductivity predicts a gap $\Delta \propto e^{-1/(N_0 V)}$ where $N_0$ is the density of states and $V$ is the electron-phonon coupling. Why is this result inherently non-perturbative? What would happen if one tried to derive it from perturbation theory in $V$?

4. A student claims: "Since the perturbation series diverges, perturbation theory is fundamentally wrong and we should only use numerical methods." Argue against this claim. What advantages does perturbation theory offer even when the series diverges?

Key Takeaways

• Perturbation series in quantum mechanics are typically asymptotic, not convergent. They have zero radius of convergence due to singularities at $\lambda = 0^-$ (Dyson's argument).
• Despite divergence, asymptotic series are extremely useful when truncated optimally. The error scales as $\sim e^{-c/\lambda}$ — exponentially small for weak coupling.
• Non-perturbative effects (tunneling, instantons, confinement) scale as $e^{-c/\lambda}$ and are invisible to any finite order of perturbation theory.
• Resurgence provides a mathematical framework connecting perturbative and non-perturbative physics, showing that the divergence of the perturbation series encodes non-perturbative information.
• A physicist must know both how to use perturbation theory and when to stop trusting it. The ability to recognize non-perturbative physics is a hallmark of theoretical maturity.