Common Pitfalls in Path Integrals

A path integral is not specified by the symbol

$$\int\mathcal Dq\, e^{iS[q]/\hbar}$$

alone. One must also state what quantity is being computed, which histories are integrated, the boundary or contour conditions, the regulator, the discretization and operator-ordering convention, and the normalization.

Path-integral normalization and ordering

Pitfall	Warning sign	First repair
Undefined measure	Only $\mathcal Dq$ is written	Restore a time slice, lattice, mode cutoff, or another regulator
Missing normalization	Wrong dimensions, delta-function limit, or composition law	Restore short-time prefactors and any division by $Z[0]$
Hidden ordering choice	Noncommuting $p$ and $q$ occur	Derive the discretization from the operator expression
Unstated mathematical status	A formal continuum expression is called a probability measure	Say whether the result is exact, regulated, asymptotic, or heuristic
Phase treated as probability	$e^{iS/\hbar}$ is described as a likelihood	Add amplitudes first and form probabilities afterward
Automatic Wick rotation	$t=-i\tau$ is used without analytic or boundary data	Start from the operator kernel and identify the continuation domain
Method chosen by habit	A simple spectral problem becomes a difficult functional integral	Compare operator, differential-equation, numerical, and path-integral routes

The quickest universal check is to return to a regulated expression. For a Cartesian particle with

$$H = \frac{p^2}{2m} + V(q),$$

one standard finite-slice kernel is

$$K_N(q_f,T;q_i,0) = \left( \frac{m}{2\pi i\hbar\epsilon} \right)^{N/2} \int \prod_{j=1}^{N-1}dq_j \times \exp\left[ \frac{i}{\hbar} S_N(q_N,\ldots,q_0) \right],$$

where

$$\epsilon=\frac{T}{N}, \qquad q_0=q_i, \qquad q_N=q_f.$$

At finite $N$, the integration variables, normalization, and discretized action are visible. Most continuum-notation mistakes become obvious here.

The Measure Is Not an Ordinary Measure

The notation $\mathcal Dq$ is shorthand for a limiting construction, not an infinite-dimensional version of ordinary Lebesgue measure. In the finite-slice expression, only the $N-1$ intermediate coordinates are integrated:

$$\prod_{j=1}^{N-1}dq_j.$$

The endpoints are fixed by the propagator. The product of differentials is also accompanied by the short-time factors shown above. Calling the product of differentials alone “the measure” hides part of the definition.

There is no nontrivial, locally finite, translation-invariant Lebesgue measure on the infinite-dimensional path spaces used here. The real-time integral is generally oscillatory and is often defined through time slicing, an $i\varepsilon$ prescription, analytic continuation, or another regulator rather than as integration against a positive measure.

The Euclidean case is better behaved in many standard models, but it still does not justify treating paths like smooth classical curves. For the free Euclidean particle, the rigorous probabilistic construction is related to Wiener measure; typical paths are continuous but nowhere differentiable. The formal term

$$\int d\tau\, \frac{m}{2} \left( \frac{dq}{d\tau} \right)^2$$

must therefore be understood through the regulated construction rather than by evaluating an ordinary derivative on a typical path.

Changing variables can also change the measure. Curvilinear coordinates, configuration-space geometry, constraints, coherent states, fermionic variables, and gauge redundancy can introduce Jacobians, boundary terms, or determinants. Transforming only the action while leaving $\mathcal Dq$ untouched is not generally valid. Propagators in Multiple Dimensions gives the neighboring coordinate-space warning.

For a proposed continuum expression, first identify its finite-dimensional approximation and endpoint data. These determine what is actually being integrated.

Normalization Is Part of the Definition

For a one-dimensional coordinate kernel, the short-time factors are required by two operator identities. First, the equal-time limit must be

$$\lim_{T\to0^+} K(q_f,T;q_i,0) = \delta(q_f-q_i).$$

Second, kernels must compose:

$$\int dq\, K(q_f,t_f;q,t) K(q,t;q_i,t_i) = K(q_f,t_f;q_i,t_i).$$

These conditions fail if one keeps only $e^{iS/\hbar}$ and discards the prefactor. Dimensional analysis catches the same problem: in one dimension,

$$[K]=\text{length}^{-1}.$$

The phase is dimensionless and cannot supply this dimension.

There are $N$ short-time kernels but only $N-1$ intermediate integrations. The resulting factor

$$\left( \frac{m}{2\pi i\hbar\epsilon} \right)^{N/2}$$

is therefore not optional decoration. Its phase convention also belongs to the real-time prescription.

Normalization appears again in generating functionals. Source derivatives of $Z[J]$ often need division by $Z[0]$:

$$\langle\mathcal T q(t_1)q(t_2)\rangle = \frac{1}{Z[0]} \left( \frac{\hbar}{i} \right)^2 \frac{\delta^2Z[J]} {\delta J(t_1)\delta J(t_2)} \bigg|_{J=0},$$

for the corresponding real-time convention. A constant can be dropped only after showing that it cancels in the normalized observable being calculated.

Diagnostic: Check dimensions, the equal-time delta function, the composition law, and any normalization by a vacuum amplitude or partition function.

Discretization Encodes Operator Ordering

Classical functions commute; quantum operators need not. The classical expression $qp$ does not distinguish

$$\hat q\hat p, \qquad \hat p\hat q, \qquad \frac12 \left( \hat q\hat p+\hat p\hat q \right),$$

yet

$$\hat q\hat p-\hat p\hat q = i\hbar.$$

A continuum action obtained from a classical Hamiltonian symbol can therefore lose information about the quantum operator.

Time slicing restores that information. In a phase-space path integral, whether $H(p,q)$ is evaluated at $q_j$, $q_{j+1}$, or a midpoint is tied to how operators were ordered in each short-time matrix element. Midpoint prescriptions are often associated with Weyl ordering, but that statement has hypotheses and does not license replacing every discretization by a midpoint rule.

For the elementary Cartesian Hamiltonian $p^2/(2m)+V(q)$, common discretizations converge to the same standard kernel under suitable assumptions. Ordering subtleties become sharper for position-dependent kinetic terms, curved coordinates, coherent-state path integrals, spin systems, or constrained dynamics. In those settings, apparently small changes can generate finite quantum corrections.

The robust direction of reasoning is

$$\text{operator expression} \longrightarrow \text{short-time matrix element} \longrightarrow \text{discretized path integral}.$$

Running this logic backward from an unqualified continuum action can be ambiguous.

Diagnostic: Identify the operator whose matrix element is being represented and show which finite-slice rule follows from it. Trotter Product Formula supplies the underlying operator limit.

Formal Does Not Mean Rigorously Defined

Path integrals appear at several different levels of mathematical control. These should not be blended:

• an exact operator identity, such as a composition law;
• a finite-dimensional time-sliced integral;
• a continuum limit proved under stated hypotheses;
• a formal generating expression used to organize perturbation theory;
• an asymptotic saddle-point expansion;
• a numerical discretization with cutoff and convergence tests.

For many semibounded Schrödinger operators, Euclidean kernels admit rigorous constructions related to the Feynman–Kac formula. The admissible potential class, operator domain, boundary conditions, and measure still matter. Singular potentials, unbounded-below actions, complex terms, constraints, and field-theory continuum limits require additional care.

Real-time path integrals are more delicate because the weight does not provide absolute damping. Their finite-dimensional approximants are Fresnel-type oscillatory integrals, and the continuum expression may be interpreted distributionally or through analytic continuation. A successful formal calculation can be physically correct without itself being a proof of existence.

The right response is not to discard formal path integrals. It is to label the mathematical status honestly and keep the regulator close enough that normalization, contour choices, and limiting operations can be checked.

Diagnostic: Ask what converges, in which sense, and under which assumptions. If the answer is only “the continuum notation suggests,” classify the step as formal or asymptotic.

Real-Time Weights Are Not Probabilities

For a real action,

$$\left| e^{iS[q]/\hbar} \right| = 1.$$

The factor $e^{iS/\hbar}$ is a phase, not a positive likelihood. It cannot say that one path is intrinsically more probable than another. Histories contribute amplitudes, and amplitudes interfere.

For two coarse-grained alternatives with amplitudes $A_1$ and $A_2$,

$$\begin{aligned} P &= \left| A_1+A_2 \right|^2 \\ &= |A_1|^2 + |A_2|^2 + 2\operatorname{Re} \left( A_1A_2^* \right). \end{aligned}$$

The interference term would be absent if $A_1$ and $A_2$ were treated as classical probabilities. It is the main reason the real-time sum over histories cannot be read as a stochastic ensemble.

Stationary phase does not change this interpretation. A classical path is a stationary point of the phase, not a maximum of a path probability. Several stationary paths can contribute, and their amplitudes still interfere.

This oscillatory structure is also the origin of severe numerical cancellation in direct real-time Monte Carlo methods. Replacing the phase by its absolute value changes the theory rather than solving the convergence problem.

Diagnostic: Locate the final operation that turns amplitudes into the requested probability, rate, trace, or expectation value. If none appears, a probability interpretation is premature.

Euclidean Continuation Has Hypotheses

For a time-independent Hamiltonian bounded below, the imaginary-time kernel is defined by the operator

$$K_E(q_f,\tau;q_i,0) = \langle q_f\rvert e^{-H\tau/\hbar} \lvert q_i\rangle, \qquad \tau\gt0.$$

For a purely discrete spectrum,

$$K_E(q_f,\tau;q_i,0) = \sum_n \psi_n(q_f) \psi_n^*(q_i) e^{-E_n\tau/\hbar}.$$

A continuous spectrum requires the corresponding spectral integral. This formula explains relative damping and identifies the analytic data that must be continued to recover a real-time kernel. Merely replacing $t$ by $-i\tau$ in a final formula does not establish that singularities, contours, domains, and boundary prescriptions continue correctly.

Boundary conditions are especially easy to lose. A fixed-endpoint Euclidean kernel is not a thermal partition function. The thermal trace requires an interval of length $\beta\hbar$, identifies the endpoints, and integrates over their common value:

$$Z_\beta = \int dq\, K_E(q,\beta\hbar;q,0).$$

Euclidean damping also does not guarantee a positive probability measure in every theory. Magnetic, Berry-phase, topological, chemical-potential, or fermionic contributions can leave a complex weight. Even when a positive Euclidean measure exists, reconstructing real-time spectral information from approximate Euclidean data can be ill-conditioned.

Diagnostic: Begin with the operator or spectral object, state the continuation domain and boundary conditions, and distinguish fixed endpoints, vacuum projection, and thermal traces.

Use the Simplest Effective Formulation

Path integrals are a formulation of quantum mechanics, not a requirement that every problem be solved by functional integration. Their strengths are especially clear for semiclassical saddle structure, tunneling, sources and correlation functions, many degrees of freedom, symmetry actions, topology, and the bridge to fields.

Other formulations are often shorter for routine questions:

Question	Productive first method
Low-lying spectrum of a one-dimensional potential	Schrödinger eigenvalue problem or matrix diagonalization
Exact evolution in a small finite Hilbert space	Matrix exponential or ordinary differential equation
Conservation law from a commutator	Operator algebra
Short-time splitting error	Trotter or Magnus analysis
Quadratic propagator	Kernel differential equation or Gaussian path integral
Several semiclassical trajectories	Stationary-phase path integral
Barrier exponent	WKB or Euclidean saddle, chosen to fit the observable
Time-ordered correlators and source insertions	Generating functional

Method choice should follow the observable and the structure of the Hamiltonian. Using operator methods to establish normalization or spectra and path integrals to expose saddle sectors is often more efficient than insisting on one language throughout.

Diagnostic: State what the path-integral formulation makes easier in this problem. If the answer is only “it is another way to write quantum mechanics,” test a simpler method first.

References

• R. P. Feynman and A. R. Hibbs, Quantum Mechanics and Path Integrals, McGraw-Hill, 1965.
• J. R. Klauder and B.-S. Skagerstam, Coherent States: Applications in Physics and Mathematical Physics, World Scientific, 1985.
• H. Kleinert, Path Integrals in Quantum Mechanics, Statistics, Polymer Physics, and Financial Markets, 5th ed., World Scientific, 2009.
• L. S. Schulman, Techniques and Applications of Path Integration, Dover, 2005.
• B. Simon, Functional Integration and Quantum Physics, 2nd ed., AMS Chelsea, 2005.
• J. Zinn-Justin, Path Integrals in Quantum Mechanics, Oxford University Press, 2005.

Exercises

1. Repair the following incomplete time-sliced free-particle kernel:

$$\widetilde K_N = \int dq_1\cdots dq_{N-1}\, \exp\left[ \frac{i}{\hbar} \sum_{j=0}^{N-1} \frac{m(q_{j+1}-q_j)^2}{2\epsilon} \right].$$

Solution

The endpoints and time step must be stated:

$$q_0=q_i, \qquad q_N=q_f, \qquad \epsilon=\frac{T}{N}.$$

Each short-time free kernel contributes a normalization factor. The repaired expression is

$$K_N(q_f,T;q_i,0) = \left( \frac{m}{2\pi i\hbar\epsilon} \right)^{N/2} \int dq_1\cdots dq_{N-1} \times \exp\left[ \frac{i}{\hbar} \sum_{j=0}^{N-1} \frac{m(q_{j+1}-q_j)^2}{2\epsilon} \right].$$

A complete real-time definition also specifies the branch or $i\varepsilon$ prescription for the Fresnel factors. The normalization is required for the delta-function limit and gives the kernel dimension $\text{length}^{-1}$.

2. Why can $e^{iS[q]/\hbar}$ not be normalized into a probability density over real-time paths by dividing by its integral?

Solution

For real $S[q]$, the modulus of the weight is one:

$$\left| e^{iS[q]/\hbar} \right| = 1.$$

It is complex rather than nonnegative, so it does not satisfy the defining properties of a probability density. Its integral is an oscillatory amplitude and can exhibit cancellations or even vanish; dividing by it does not turn the local complex phase into a positive measure.

Quantum probabilities arise after amplitudes for alternatives have been combined and the appropriate absolute square, trace, or expectation value has been formed.

3. Explain why the classical symbol $qp$ does not determine a unique quantum path integral.

Solution

Classically, $qp=pq$. Quantum mechanically,

$$[\hat q,\hat p] = i\hbar,$$

so

$$\hat q\hat p \neq \hat p\hat q.$$

The symmetrized operator

$$\frac12 \left( \hat q\hat p+\hat p\hat q \right)$$

is different from either ordered product by a finite multiple of $i\hbar$. The individual products $\hat q\hat p$ and $\hat p\hat q$ are not symmetric on the usual common domain; if a self-adjoint dilation generator is intended, this is an additional reason to use the symmetric combination. A continuum classical symbol does not record which operator was intended. A short-time matrix element and its discretization rule must supply that information.

4. Distinguish a fixed-endpoint Euclidean kernel from a thermal partition function.

Solution

The fixed-endpoint kernel is

$$K_E(q_f,\tau;q_i,0) = \langle q_f\rvert e^{-H\tau/\hbar} \lvert q_i\rangle.$$

Its paths satisfy

$$q(0)=q_i, \qquad q(\tau)=q_f.$$

The thermal partition function is a trace:

$$Z_\beta = \operatorname{Tr}e^{-\beta H} = \int dq\, \langle q\rvert e^{-\beta H} \lvert q\rangle.$$

Thus its Euclidean interval has length $\beta\hbar$, its endpoints are identified, and the common endpoint is integrated:

$$Z_\beta = \int dq\, K_E(q,\beta\hbar;q,0).$$

The two objects differ in both boundary conditions and physical interpretation.

5. Choose a productive first method for each task: finding the two lowest energies of a one-dimensional anharmonic oscillator, extracting the leading tunneling exponent in a high double well, and deriving time-ordered coordinate correlators with many insertions.

Solution

For two low-lying energies, direct matrix diagonalization or a Schrödinger boundary-value solver is usually the cleanest first method. It yields approximate eigenvalues and normalized eigenvectors and provides a benchmark for approximations.

For the leading high-barrier tunneling exponent, either WKB under the barrier or a Euclidean instanton is natural. WKB is often shortest in one dimension; the instanton formulation is especially useful when the saddle structure or extension to many degrees of freedom matters.

For many time-ordered insertions, introduce a source and use a generating functional. Functional derivatives then organize the correlators more efficiently than repeated operator insertions. The answer still depends on the state, contour, ordering convention, and normalization by the source-free functional.