Path Integrals — Mathematical Foundations

⚡ Key Takeaways

Path integrals define quantum amplitudes and statistical partition functions as sums over all trajectories, weighted by $e^{iS/\hbar}$ (real time) or $e^{-S/\hbar}$ (imaginary time). The central result: the propagator equals a functional integral $\int \mathcal{D}x\, e^{iS[x]/\hbar}$ , which is rigorously a measure only in the Euclidean (Wiener) case. Applies across quantum mechanics, QFT, statistical mechanics, and stochastic analysis. Formulated by Dirac (1933) and Feynman (1948), made rigorous by Kac (1949).

💡 The Core Insight

Quantum evolution over a finite time is the composition of infinitely many infinitesimal evolutions. Inserting complete sets of position eigenstates at each time slice converts the operator $e^{-iHT/\hbar}$ into a multi-dimensional integral over intermediate positions; the continuum limit of these slices is the path integral. The phase of each path is its classical action $S[x]=\int L\,dt$ , so quantum mechanics becomes "democratic interference among histories," with the classical trajectory emerging by stationary phase as $\hbar\to 0$ .

DefinitionTime-Sliced Propagator

Let $H = \frac{p^2}{2m} + V(x)$ act on $L^2(\mathbb{R}^d)$ . The propagator (integral kernel of $e^{-iHT/\hbar}$ ) is defined as the $N\to\infty$ limit of the time-sliced approximation with $\epsilon = T/N$ , $x_0=x_i$ , $x_N=x_f$ :

1

\quad K(x_f,T;x_i,0) = \lim_{N\to\infty} \left(\frac{m}{2\pi i\hbar\epsilon}\right)^{Nd/2} \int \prod_{k=1}^{N-1} d^dx_k \; \exp\!\left[\frac{i}{\hbar}\sum_{k=1}^{N}\epsilon\, L\!\left(\frac{x_k-x_{k-1}}{\epsilon}, x_k\right)\right]

Derivation hint: Insert $N-1$ resolutions of identity $\int|x\rangle\langle x|d^dx$ between factors of $e^{-iH\epsilon/\hbar}$ , then use the Trotter formula $e^{-iH\epsilon/\hbar}\approx e^{-iT\epsilon/\hbar}e^{-iV\epsilon/\hbar}$ and the Gaussian Fourier transform of the free kernel. Here $L(\dot x, x) = \frac{1}{2}m\dot x^2 - V(x)$ is the Lagrangian, $\hbar$ is the reduced Planck constant, and the prefactor $(m/2\pi i\hbar\epsilon)^{d/2}$ per slice normalizes the free-particle Gaussian.

DefinitionFormal Functional Integral

In continuum notation one writes the same object as

2

\quad K(x_f,T;x_i,0) = \int_{x(0)=x_i}^{x(T)=x_f} \mathcal{D}x(t)\; \exp\!\left[\frac{i}{\hbar} S[x]\right], \qquad S[x]=\int_0^T L(\dot x, x)\, dt

Derivation hint: This is shorthand for [1]; $\mathcal{D}x$ absorbs the limit of the product measure together with the $N$ -dependent normalization prefactor. No genuine $\sigma$ -additive measure $\mathcal{D}x$ on path space exists for real $\hbar$ — the expression is defined by its time-sliced limit.

TheoremTrotter Product Formula

Let $A, B$ be self-adjoint operators on a Hilbert space such that $A+B$ is (essentially) self-adjoint on $D(A)\cap D(B)$ . Then for all $\psi$ in the Hilbert space,

3

\quad e^{-it(A+B)} \psi = \lim_{N\to\infty} \left(e^{-itA/N}\, e^{-itB/N}\right)^N \psi

Derivation hint: For bounded operators expand both sides to first order in $t/N$ ; the error per step is $O((t/N)^2)$ so the total error is $O(t^2/N)\to 0$ . The unbounded case is Kato's theorem, proved via the Trotter–Kato approximation theorem for strongly continuous semigroups. This is the rigorous engine converting the operator identity into equation [1].

TheoremFeynman–Kac Formula

Let $V$ be bounded below and continuous, and let $(b_t)_{t\ge 0}$ be $d$ -dimensional Brownian motion started at $x$ under Wiener measure. Then the imaginary-time (Euclidean) semigroup $e^{-tH/\hbar}$ has kernel given rigorously by an expectation:

4

\quad \left(e^{-tH}\psi\right)(x) = \mathbb{E}_x\!\left[\exp\!\left(-\int_0^t V(b_s)\,ds\right)\psi(b_t)\right]

Derivation hint: Apply Trotter [3] to $H=-\tfrac12\Delta+V$ ; the free factor $e^{t\Delta/2N}$ generates the Brownian transition density (heat kernel), and the product over slices converges to the time-ordered Wiener integral $\mathbb{E}_x[\cdots]$ by dominated convergence using boundedness of $V$ from below.

ProofFrom Operator Evolution to the Path Integral

Step 1 — Slice the evolution. Write $T=N\epsilon$ and use the group property

e^{-iHT/\hbar} = \left(e^{-iH\epsilon/\hbar}\right)^N.

Step 2 — Insert position resolutions. Between each pair of factors insert $\mathbb{1}=\int d^dx_k\,|x_k\rangle\langle x_k|$ , giving

K = \int \prod_{k=1}^{N-1} d^dx_k \prod_{k=1}^{N} \langle x_k| e^{-iH\epsilon/\hbar}|x_{k-1}\rangle.

Step 3 — Approximate the short-time kernel. By the Trotter formula [3], to leading order in $\epsilon$ ,

\langle x_k| e^{-iH\epsilon/\hbar}|x_{k-1}\rangle \approx \langle x_k| e^{-i T_{\rm kin}\epsilon/\hbar} e^{-iV\epsilon/\hbar}|x_{k-1}\rangle = e^{-iV(x_{k-1})\epsilon/\hbar}\langle x_k| e^{-i\hat p^2\epsilon/2m\hbar}|x_{k-1}\rangle.

Step 4 — Compute the free kernel. Insert momentum states and complete the Gaussian:

\langle x_k| e^{-i\hat p^2\epsilon/2m\hbar}|x_{k-1}\rangle = \left(\frac{m}{2\pi i\hbar\epsilon}\right)^{d/2}\exp\!\left[\frac{im(x_k-x_{k-1})^2}{2\hbar\epsilon}\right].

Step 5 — Assemble the action. Multiply the $N$ factors:

K = \left(\frac{m}{2\pi i\hbar\epsilon}\right)^{Nd/2}\!\int\!\prod_{k}d^dx_k\,\exp\!\left[\frac{i}{\hbar}\sum_k\!\left(\frac{m(x_k-x_{k-1})^2}{2\epsilon}-\epsilon V(x_{k-1})\right)\right].

The bracket is the Riemann sum of $\int_0^T(\tfrac12 m\dot x^2 - V)\,dt = S[x]$ .

Step 6 — Continuum limit. Taking $N\to\infty$ formally yields equation [2]. Rigor requires either Wick rotation $t\to -i\tau$ (yielding the convergent Wiener integral and [4]) or the theory of oscillatory integrals on infinite-dimensional spaces (Albeverio–Høegh-Krohn). $\blacksquare$

∎

RP

Richard P. Feynman

1918–1988

Feynman developed the path-integral formulation in his 1942 Princeton thesis and the 1948 Reviews of Modern Physics paper "Space-Time Approach to Non-Relativistic Quantum Mechanics," inspired by a cryptic 1933 remark of Dirac that $\langle x_{t+\epsilon}|x_t\rangle$ "corresponds to" $e^{iL\epsilon/\hbar}$ . Feynman recast all of quantum mechanics as a sum over histories, gave the diagrammatic perturbation theory that bears his name, and — crucially for foundations — emphasized that the dominant paths are nowhere differentiable, with $|\Delta x| \sim \sqrt{\Delta t}$ , anticipating the Brownian roughness that Kac would make precise. He shared the 1965 Nobel Prize with Schwinger and Tomonaga.

🧠 Intuition

Imagine evolving a wavefunction through time as a relay of infinitely many tiny hops. At each hop the amplitude to go from $x_{k-1}$ to $x_k$ is a complex number of unit magnitude whose phase is the little bit of action accrued. To get the total amplitude you multiply phases along a path (adding actions) and then add the resulting complex numbers over all paths. Most paths are wild and their phases scramble to zero by destructive interference — except near the classical trajectory, where $S$ is stationary and neighbouring paths share nearly equal phase, so they add coherently. As $\hbar\to 0$ this stationary-phase peak sharpens to a delta function on the classical path: that is how Newton emerges from quantum democracy. In imaginary time the phases become real exponential weights $e^{-S_E/\hbar}$ , turning interference into ordinary probability and the path integral into a genuine average over Brownian paths.

Figure

Sum over histories between fixed endpoints; the classical path (gold) is the stationary-phase trajectory among nearby quantum paths.

ExampleFree Particle Propagator

Given: $V=0$ , mass $m$ , propagate from $x_i$ to $x_f$ in time $T$ in $d=1$ .

Step 1 — Stationary path. The classical path is the straight line; the classical action is

S_{\rm cl} = \frac{m(x_f-x_i)^2}{2T}.

Step 2 — Gaussian fluctuations. Because $L$ is quadratic, the fluctuation integral is exactly the product of slice Gaussians, giving the prefactor $\sqrt{m/2\pi i\hbar T}$ .

Step 3 — Assemble.

K = \left(\frac{m}{2\pi i\hbar T}\right)^{1/2}\exp\!\left[\frac{i m(x_f-x_i)^2}{2\hbar T}\right].

Numerical check: Take $m=9.11\times10^{-31}\,$ kg (electron), $x_f-x_i = 1\,$ nm $=10^{-9}\,$ m, $T=10^{-15}\,$ s, $\hbar=1.055\times10^{-34}\,$ J·s. Phase $= \dfrac{m(\Delta x)^2}{2\hbar T} = \dfrac{(9.11\times10^{-31})(10^{-18})}{2(1.055\times10^{-34})(10^{-15})} \approx 4.32\,$ rad.

\boxed{K \approx \left(\tfrac{m}{2\pi i\hbar T}\right)^{1/2} e^{\,i(4.32)}, \quad \text{prefactor magnitude}\ \approx 3.7\times10^{10}\ \mathrm{m}^{-1/2}}

ExampleImaginary-Time Harmonic Oscillator Partition Function

Given: $H=\frac{p^2}{2m}+\frac12 m\omega^2 x^2$ . Compute $Z=\mathrm{Tr}\,e^{-\beta H}$ via the Euclidean path integral over periodic paths $x(\beta\hbar)=x(0)$ .

Step 1 — Mode expansion. Expand in Matsubara modes $\omega_n = 2\pi n/(\beta\hbar)$ . The Euclidean action's quadratic form has eigenvalues $m(\omega_n^2+\omega^2)$ .

Step 2 — Functional determinant.

Z = \prod_{n=-\infty}^{\infty}\frac{1}{\sqrt{(\omega_n^2+\omega^2)/\omega_n^2}}\ \propto\ \frac{1}{2\sinh(\beta\hbar\omega/2)}.

The product is evaluated using $\prod_{n\ge1}(1+x^2/n^2)=\sinh(\pi x)/(\pi x)$ .

Step 3 — Final result.

Z = \frac{1}{2\sinh(\beta\hbar\omega/2)} = \frac{e^{-\beta\hbar\omega/2}}{1-e^{-\beta\hbar\omega}}.

Numerical check: $\hbar\omega = 0.025\,$ eV (typical phonon), $T=300\,$ K so $\beta = 1/k_BT$ , $k_BT=0.0259\,$ eV. Then $\beta\hbar\omega = 0.025/0.0259 = 0.965$ .

\boxed{Z = \frac{1}{2\sinh(0.483)} = \frac{1}{2(0.502)} \approx 0.996}

matching the exact harmonic partition function, dimensionless.

⚠️ Common mistake

" $\mathcal{D}x$ is a measure on path space." It is not, for real time. The Cameron theorem shows there is no complex measure of finite total variation on $C[0,T]$ giving the Feynman integral; the would-be "Lebesgue measure on $\mathbb{R}^\infty$ " does not exist (no translation-invariant $\sigma$ -finite measure on an infinite-dimensional Banach space). The symbol [2] is rigorous only as the time-sliced limit [1], as an oscillatory integral (Albeverio–Høegh-Krohn / Fresnel integrals), or after Wick rotation to the Wiener measure.

⚠️ Common mistake

"Discretization is unambiguous." It is not. For Hamiltonians with momentum-dependent or position-dependent mass terms, different discretizations (mid-point vs. end-point evaluation of $V$ , ordering of $\hat x,\hat p$ ) yield genuinely different continuum limits differing by $O(\hbar)$ quantum corrections. The mid-point (Weyl-ordering) prescription is required to reproduce the symmetric operator ordering. Path integrals "remember" operator ordering through the discretization.

Counterexample

Non-existence of Feynman measure. Suppose a countably additive complex measure $\mu$ on $C[0,T]$ existed with $\int e^{(i/\hbar)\int(\frac12 m\dot x^2)dt}\,d\mu = K_{\rm free}$ . Cameron (1960) proved that any candidate constructed as a limit of finite-dimensional Gaussians with imaginary covariance has total variation diverging like $\prod_k |2\pi i\hbar\epsilon/m|^{-1/2}\to\infty$ . Hence no finite-variation measure reproduces the free propagator. The Euclidean analogue does exist — the Wiener measure — precisely because its covariance is real and positive, making the infinite product converge. This sharp contrast is the central foundational fact.

connection

The path integral sits at the crossroads of several theories. Via Wick Rotation it connects to the Wiener Process and the Feynman–Kac Formula, identifying quantum amplitudes with diffusion expectations. Through stationary phase it reproduces the WKB Approximation and the Hamilton–Jacobi Equation in the classical limit. Its functional-determinant evaluation links to Zeta Function Regularization and the Heat Kernel Expansion. In field theory it generalizes to the Generating Functional $Z[J]$ , from which Feynman Diagrams arise by functional differentiation, and to Lattice Gauge Theory as a rigorous regularization. The measure-theoretic obstruction connects to Gaussian Measures on Banach Spaces and the Cameron–Martin Theorem.

🔬 Where this appears

⏳ History

- 1923 (Wiener). Norbert Wiener constructs rigorous Brownian-motion measure on continuous path space — the eventual Euclidean foundation.

- 1933 (Dirac). Dirac's paper "The Lagrangian in Quantum Mechanics" notes that the short-time amplitude "corresponds to"

\exp(iL\,dt/\hbar)

.

- 1942 (Feynman). Feynman's Princeton PhD thesis develops the least-action principle in quantum mechanics into a sum over paths.

- 1948 (Feynman). "Space-Time Approach to Non-Relativistic Quantum Mechanics" publishes the path-integral formulation.

- 1949 (Kac). Mark Kac, hearing Feynman lecture, proves the rigorous Feynman–Kac formula by Wick rotation to Wiener measure.

- 1957 (Gelfand–Yaglom). Method for functional determinants of Gaussian path integrals via an ordinary differential equation.

- 1960 (Cameron). R. H. Cameron proves the non-existence of a countably additive Feynman measure of finite variation.

- 1972 (Nelson). Edward Nelson rigorously establishes the Trotter-formula route and stochastic mechanics.

- 1976 (Albeverio–Høegh-Krohn). Rigorous theory of oscillatory ("Fresnel") integrals on infinite-dimensional spaces, giving the real-time integral a precise meaning.

open-problem

The real-time measure problem remains unresolved in full generality. While Euclidean path integrals are rigorous for a wide class of potentials, and oscillatory-integral and analytic-continuation methods handle quadratic-plus-Fresnel-class perturbations, there is no universally accepted measure-theoretic foundation for real-time path integrals with generic interactions, especially in interacting relativistic QFT (the constructive field theory program succeeded for $\phi^4_2,\phi^4_3$ but $\phi^4_4$ is believed trivial, and 4D Yang–Mills existence with a mass gap is a Clay Millennium Problem). Further open issues: rigorous treatment of fermionic Grassmann "integrals," the convergence/Borel-summability of the perturbation series defined by the integral, the proper measure for path integrals on curved configuration spaces and over geometries (quantum gravity), and the sign problem obstructing Monte Carlo for finite-density and real-time lattice computations.

Path Integrals — Mathematical Foundations

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also