Imaginary-Time and Euclidean Path Integrals

The real-time path integral is a beautiful object and a computational nightmare. Its weight $e^{iS/\hbar}$ is a pure phase of unit modulus — every path contributes with the same magnitude and differs only in its angle, so the sum converges only through delicate cancellation between wildly oscillating terms. You cannot sample it, you cannot bound it, and away from the exactly solvable Gaussians you cannot compute it. The fix is one of the most consequential tricks in theoretical physics: rotate time into the imaginary axis, $t=-i\tau$ , and the phase $e^{iS/\hbar}$ becomes a real, positive, exponentially damped weight $e^{-S_E/\hbar}$ . Oscillation becomes decay; interference becomes probability. What makes this more than a computational convenience is the object the rotation produces. The Euclidean action $S_E=\int(\tfrac12 m\dot x^2+V)\,\mathrm d\tau$ has the potential entering with a plus sign, so it is bounded below and the weight $e^{-S_E/\hbar}$ is a genuine measure on paths — the analytic continuation of Wiener's 1923 Brownian measure. And when you close the imaginary-time interval into a circle of circumference $\beta\hbar$ , the path integral over periodic trajectories computes exactly the thermal partition function $Z=\mathrm{Tr}\,e^{-\beta\hat H}$ . Quantum mechanics in imaginary time is statistical mechanics, with the dictionary $\beta=it/\hbar$ . You come to this already holding the real-time construction (The Feynman Path Integral Formulation) and the exact Gaussian propagators (Free-Particle and Harmonic-Oscillator Path Integrals). Here we perform the Wick rotation carefully, watch the free kernel turn from a Fresnel oscillation into the heat-diffusion Green's function, and compute the harmonic-oscillator partition function $Z=1/[2\sinh(\beta\hbar\omega/2)]$ directly from periodic Euclidean paths, reading its zero-point energy $\tfrac12\hbar\omega$ off the low-temperature limit. This machinery — real weights, periodic paths, the QM↔stat-mech bridge — is the foundation for lattice field theory and, in the very next article, for instanton tunnelling.

Key takeaways

Wick rotation $t=-i\tau$ turns the oscillating weight $e^{iS/\hbar}$ into a real, positive, damped weight $e^{-S_E/\hbar}$ — the reason Euclidean path integrals can actually be computed and sampled.
The Euclidean action $S_E=\int(\tfrac12 m\dot x^2+V)\,\mathrm d\tau$ has $V$ with a plus sign (bounded below): $e^{-S_E/\hbar}$ is a genuine probability measure on paths.
The free Euclidean propagator $K_E=\sqrt{m/2\pi\hbar\tau}\,e^{-mx^2/2\hbar\tau}$ is the heat/diffusion kernel with diffusion constant $D=\hbar/2m$ — no factor of $i$ .
Closing imaginary time into a circle of length $\beta\hbar$ and summing over periodic paths gives the partition function $Z=\mathrm{Tr}\,e^{-\beta\hat H}$ .
For the oscillator this yields $Z=1/[2\sinh(\beta\hbar\omega/2)]$ ; the low- $T$ limit $e^{-\beta\hbar\omega/2}$ exhibits the zero-point energy $\tfrac12\hbar\omega$ .
The dictionary is $\beta=it/\hbar$ : quantum dynamics in imaginary time is statistical mechanics at temperature $T=1/k_B\beta$ .
Key idea: rotating time to $t=-i\tau$ replaces the intractable oscillating phase by a real Boltzmann-like weight, turns the wave kernel into the diffusion kernel, and makes the periodic path integral compute the thermal partition function.

The core insight

Wick-rotating time $t=-i\tau$ converts the oscillating path-integral weight $e^{iS/\hbar}$ into the real, positive weight $e^{-S_E/\hbar}$ ; summed over periodic paths of period $\beta\hbar$ it computes the partition function $Z=\mathrm{Tr}\,e^{-\beta\hat H}$ , making quantum mechanics in imaginary time identical to statistical mechanics.

Where this lives today

Every large-scale computation in nonperturbative quantum field theory is Euclidean. Lattice QCD discretizes imaginary-time spacetime, replaces $e^{iS/\hbar}$ by the real Boltzmann weight $e^{-S_E/\hbar}$ , and samples field configurations by Monte Carlo — impossible with the oscillating Minkowski weight — to compute hadron masses, the proton's internal structure, and the finite-temperature phase transition of quark matter. The identical machinery runs in condensed matter: quantum Monte Carlo simulations of magnets and superconductors are path integrals over periodic imaginary-time worldlines, with the temperature set by the circle's circumference $\beta\hbar$ . When a physicist speaks of the "sign problem" that plagues finite-density QCD or frustrated magnets, they mean a residual complex phase that survives the Wick rotation and spoils the very positivity this article's rotation is designed to deliver.

The physical picture

The single most important picture in this article is what happens to the weight. In real time, every path carries $e^{iS/\hbar}$ : a unit-modulus phasor spinning at a rate set by its action. Paths near the classical trajectory spin slowly and reinforce; paths far away spin so fast that they cancel in pairs. The sum is a mirage of cancellations. Rotate time by ninety degrees into the complex plane and the phasor stops spinning and starts shrinking: $e^{iS/\hbar}\to e^{-S_E/\hbar}$ , a real number between 0 and 1. The classical path is now the path of largest weight, and everything else is exponentially suppressed. This is why Euclidean integrals converge and Minkowski ones fight you.

Figure

The Wick rotation: an oscillating phasor becomes a damped Boltzmann weight

The second picture is thermal. Take the trace of the evolution operator in imaginary time: $\mathrm{Tr}\,e^{-\tau\hat H/\hbar}$ with $\tau=\beta\hbar$ . Writing the trace as $\int\mathrm dx\,\langle x|e^{-\beta\hat H}|x\rangle$ means the path must return to where it started — it is a closed loop in imaginary time, a circle of circumference $\beta\hbar$ . Hot systems ( $\beta$ small) live on short circles where the path barely moves; cold systems ( $\beta$ large) live on long circles where the path can explore. The partition function is a sum over these closed imaginary-time worldlines, and every thermal average in statistical mechanics becomes a path-integral expectation value.

Figure

Periodic imaginary-time paths compute the partition function Z

The core equations

We begin with the rotation itself, applied to the real-time weight $e^{iS/\hbar}$ . Substituting $t=-i\tau$ carries a Minkowski action into a Euclidean one; the crucial bookkeeping is that two factors of the time derivative flip the sign of the kinetic term while the measure $\mathrm dt=-i\,\mathrm d\tau$ supplies the overall factor. The result is worth pinning as the master statement of the article.

i\frac{S}{\hbar}\;\xrightarrow{\;t=-i\tau\;}\;-\frac{S_E}{\hbar},\qquad S_E=\int\!\Big(\tfrac12 m\dot x^2+V(x)\Big)\mathrm d\tau \qquad (1)

where: $\dot x=\mathrm dx/\mathrm d\tau$ now denotes the imaginary-time derivative; $S_E$ — the Euclidean action. Reading it physically: the potential $V$ has flipped sign relative to the Lagrangian, so $S_E$ is a sum of two positive-definite pieces and is bounded below. That is exactly why $e^{-S_E/\hbar}$ is a legitimate statistical weight — the analytic continuation of the Brownian-motion measure — whereas $e^{iS/\hbar}$ is not. The free-particle kernel makes the transformation concrete. Rotating the Minkowski $\sqrt{m/2\pi i\hbar t}\,e^{imx^2/2\hbar t}$ removes both factors of $i$ and leaves a real Gaussian.

K_E(x,\tau;0,0)=\sqrt{\frac{m}{2\pi\hbar\tau}}\;\exp\!\Big[-\frac{m x^2}{2\hbar\tau}\Big] \qquad (2)

where: $\tau$ — the imaginary-time interval; the prefactor and exponent are now purely real. Reading it physically: this is the heat-kernel — the Green's function of the diffusion equation $\partial_\tau K_E=D\,\partial_x^2 K_E$ with diffusion constant $D=\hbar/2m$ . The single factor of $i$ that made the real-time kernel spread with interference is gone; imaginary-time "evolution" is genuine diffusion, and $K_E$ is a normalized probability density, $\int K_E\,\mathrm dx=1$ . Closing imaginary time into a circle produces the partition function. Setting $\tau=\beta\hbar$ and taking the trace forces periodic boundary conditions.

Z=\mathrm{Tr}\,e^{-\beta\hat H}=\oint_{x(0)=x(\beta\hbar)}\!\mathcal D x\;e^{-S_E[x]/\hbar} \qquad (3)

where: the loop symbol denotes the sum over paths that return to their start after imaginary time $\beta\hbar$ ; $\beta=1/k_BT$ . Reading it physically: every thermal quantity is now a path-integral average with the Boltzmann weight $e^{-S_E/\hbar}$ . Expectation values of operators become correlation functions of the periodic path — the working definition used in every Monte-Carlo lattice simulation.

Derivation AWick rotation of the action

Our goal is equation (1): show that the real-time phase $iS/\hbar$ becomes the real Euclidean damping $-S_E/\hbar$ , with the potential switching sign. The one subtlety is that the time derivative appears squared in the kinetic term, so the rotation acts on it twice, while the measure $\mathrm dt$ contributes once — we must track both. We work at the level of the Lagrangian density (per unit imaginary time) and reassemble the integral at the end.

Step 1. Substitute $t=-i\tau$ , so $\mathrm dt=-i\,\mathrm d\tau$ and $\dfrac{\mathrm dx}{\mathrm dt}=\dfrac{\mathrm dx}{-i\,\mathrm d\tau}=i\,\dfrac{\mathrm dx}{\mathrm d\tau}$ . Why: the chain rule is the whole content; the imaginary-time velocity is $i$ times the real-time velocity, so its square picks up a minus sign.

Step 2. The kinetic term: $\big(\tfrac{\mathrm dx}{\mathrm dt}\big)^2=\big(i\,\dot x\big)^2=-\dot x^2$ , where now $\dot x=\mathrm dx/\mathrm d\tau$ . Why: the sign flip here is the origin of the entire effect — the kinetic energy in $L=\tfrac12 m\dot x^2-V$ becomes $-\tfrac12 m\dot x^2$ when re-expressed in imaginary-time derivatives.

Step 3. Assemble the integrand of $iS/\hbar$ per unit $\mathrm d\tau$ : $\dfrac{i}{\hbar}\Big[-\tfrac12 m\dot x^2-V\Big]\,(\mathrm dt/\mathrm d\tau)=\dfrac{i}{\hbar}\Big[-\tfrac12 m\dot x^2-V\Big](-i)=-\dfrac{1}{\hbar}\Big[\tfrac12 m\dot x^2+V\Big]$ . Why: the two factors of $i$ (one from the measure, one already present) combine to $i\cdot(-i)=1$ , turning the whole expression real; the bracket is now $\tfrac12 m\dot x^2+V$ , the Euclidean Lagrangian.

Step 4. Integrate: $\displaystyle i\frac{S}{\hbar}=-\frac{1}{\hbar}\int\!\Big(\tfrac12 m\dot x^2+V\Big)\mathrm d\tau=-\frac{S_E}{\hbar}$ . Why: the potential now enters with a plus sign, so $S_E\ge0$ and the weight $e^{-S_E/\hbar}\in(0,1]$ — a bona fide measure. The classical Euclidean path minimizes $S_E$ (motion in the inverted potential $-V$ , the fact exploited by instantons). Verified: `verify_aq-qm-euclidean-path-integral.py` confirms the integrand identity $iS/\hbar=-(1/\hbar)S_E$ symbolically and the plus-sign of $V$ in $S_E$ .

Go deeper · Derivation B · 1 minThe oscillator partition function from periodic paths

Derivation BThe oscillator partition function from periodic paths

Here we compute $Z=\mathrm{Tr}\,e^{-\beta\hat H}$ for the harmonic oscillator and recover the standard result $Z=1/[2\sinh(\beta\hbar\omega/2)]$ , then read the zero-point energy from its low-temperature limit. The strategy is to use the known spectrum $E_n=(n+\tfrac12)\hbar\omega$ and sum the Boltzmann factors — the same number the periodic Euclidean path integral produces — because it makes the geometric series and the zero-point offset fully transparent.

Step 1. Write the trace in the energy eigenbasis: $Z=\sum_{n=0}^{\infty}e^{-\beta E_n}$ with $E_n=(n+\tfrac12)\hbar\omega$ . Set $b\equiv\beta\hbar\omega$ . Why: the trace is basis-independent; the eigenbasis makes it a geometric series in $q=e^{-b}$ .

Step 2. Factor the zero-point piece: $Z=e^{-b/2}\sum_{n=0}^{\infty}e^{-bn}=e^{-b/2}\dfrac{1}{1-e^{-b}}$ . Why: the $\tfrac12\hbar\omega$ in every level factors out as an overall $e^{-b/2}$ ; the remaining sum is the ordinary geometric series $\sum q^n=1/(1-q)$ for $|q|<1$ .

Step 3. Convert to a hyperbolic form: $Z=\dfrac{e^{-b/2}}{1-e^{-b}}=\dfrac{1}{e^{b/2}-e^{-b/2}}=\dfrac{1}{2\sinh(b/2)}$ . Why: multiplying numerator and denominator by $e^{b/2}$ produces $e^{b/2}-e^{-b/2}=2\sinh(b/2)$ — the compact closed form.

Step 4. Low-temperature limit $b\to\infty$ : $Z\to e^{-b/2}=e^{-\beta\hbar\omega/2}$ , so $Z\,e^{+b/2}\to1$ . Why: at low $T$ only the ground state survives, and $Z\approx e^{-\beta E_0}$ with $E_0=\tfrac12\hbar\omega$ — the zero-point energy read directly off the partition function. Verified: the verifier confirms $Z=1/[2\sinh(\beta\hbar\omega/2)]$ , the low- $T$ limit $Z e^{+b/2}\to1$ , and numerically $Z(\beta\hbar\omega=1)=0.95952$ against the direct Boltzmann sum.

Limiting cases

case	result	meaning
$\tau$ real (undo rotation)	$K_E\to$ Minkowski Fresnel kernel	recovers real-time quantum mechanics
free particle	$K_E=$ heat kernel, $D=\hbar/2m$	imaginary-time evolution is diffusion, no interference
$\beta\to\infty$ (low $T$ )	$Z\to e^{-\beta\hbar\omega/2}$	only the ground state contributes; zero-point energy $\tfrac12\hbar\omega$
$\beta\to0$ (high $T$ )	$Z\to k_BT/\hbar\omega$	classical equipartition; short circle, frozen path
$V\to$ double well	classical Euclidean solution = the kink	instanton tunnelling (Instantons and Quantum Tunneling)

Common trapWick rotation is not just "put an i in front of t" — the potential changes sign

The most common error is to rotate the kinetic term but forget that the potential's sign flips in the Euclidean action. In real time $L=\tfrac12 m\dot x^2-V$ ; in imaginary time $S_E=\int(\tfrac12 m\dot x^2+V)\,\mathrm d\tau$ — the $V$ now adds. Equivalently, the classical Euclidean equation of motion $m\ddot x=+V'(x)$ describes motion in the inverted potential $-V$ . Getting this sign wrong makes $S_E$ unbounded below (weights blow up) and destroys the entire point of the rotation. It is also precisely the sign that lets a particle "roll over" a barrier in imaginary time — the seed of instantons.

Worked examples

Example 1 — the free Euclidean propagator is a normalized diffusion kernel. Verify that $K_E=\sqrt{m/2\pi\hbar\tau}\,e^{-mx^2/2\hbar\tau}$ is a probability density and identify its diffusion constant.

Step 1. Normalization. This is a Gaussian with variance $\sigma^2=\hbar\tau/m$ , so $\int_{-\infty}^{\infty}K_E\,\mathrm dx=\sqrt{m/2\pi\hbar\tau}\cdot\sqrt{2\pi\hbar\tau/m}=\mathbf{1}$ .

Step 2. Diffusion equation. Compute $\partial_\tau K_E$ and $\partial_x^2 K_E$ : direct differentiation gives $\partial_\tau K_E=\dfrac{\hbar}{2m}\,\partial_x^2 K_E$ , so $K_E$ solves the heat equation with $D=\hbar/2m$ .

Step 3. Read the physics. The mean-square spread grows as $\langle x^2\rangle=\sigma^2=\hbar\tau/m=2D\tau$ — ordinary linear-in-time diffusion, not the interfering spread of real-time quantum mechanics. The imaginary-time free particle is literally a Brownian walker with $D=\hbar/2m$ .

Example 2 — oscillator partition function and its zero-point energy at $\beta\hbar\omega=1$ . Compute $Z$ and extract the ground-state energy.

Step 1. Use the closed form: $Z=1/[2\sinh(\beta\hbar\omega/2)]$ . With $\beta\hbar\omega=1$ , $Z=1/[2\sinh(0.5)]$ .

Step 2. Evaluate: $\sinh(0.5)=0.521095$ , so $Z=1/(2\times0.521095)=\mathbf{0.95952}$ .

Step 3. Cross-check against the direct Boltzmann sum $Z=\sum_n e^{-(n+1/2)}$ : $e^{-0.5}+e^{-1.5}+e^{-2.5}+\cdots=0.60653+0.22313+0.08208+\cdots=0.95952$ . The two agree to all digits, confirming the geometric resummation.

Step 4. Zero-point energy. As $\beta\hbar\omega\to\infty$ , $Z\to e^{-\beta\hbar\omega/2}$ , i.e. $Z\approx e^{-\beta E_0}$ with $E_0=\tfrac12\hbar\omega$ . The half-quantum zero-point energy is visible directly in the low-temperature partition function — no separate calculation needed.

Origins · 1 mingian-carlo wick and how this was discovered

gian-carlo wick

1909–1992 · Italian-American theorist; introduced the Wick rotation and Wick's theorem, foundations of perturbative and Euclidean field theory

Wick introduced the rotation $t\to-i\tau$ in 1954 (Phys. Rev. 96, 1124) as a way to make the divergent, oscillatory integrals of relativistic scattering theory convergent by continuing to Euclidean signature, where the metric becomes positive-definite. The same continuation underlies his earlier "Wick's theorem" for reducing products of field operators. He worked at Berkeley, Carnegie Tech, and Columbia; his rotation is now the standard bridge between quantum field theory and statistical mechanics, and the reason lattice QCD is a Euclidean, not Minkowski, computation.

History

1923

— Wiener constructs the rigorous measure on Brownian paths, later recognized as the Euclidean path-integral measure

1949

— Kac proves the Feynman-Kac formula, expressing e^{−τH} as an average over Wiener paths

1954

— Wick introduces the rotation t → −iτ to render field-theory integrals convergent

1955

— Matsubara formulates imaginary-time (thermal) field theory with period βℏ

1979

— lattice gauge theory (Creutz, following Wilson 1974) computes Euclidean path integrals by Monte Carlo

Going further

This article Wick-rotates the real-time construction of The Feynman Path Integral Formulation and the exact Gaussians of Free-Particle and Harmonic-Oscillator Path Integrals, turning the free propagator into the diffusion kernel. The QM↔stat-mech bridge it builds is the setting for the next article, Instantons and Quantum Tunneling, where the classical Euclidean solution in a double well — motion in the inverted potential flagged in the warning — computes the tunnelling splitting. Readers wanting the rigorous measure-theory foundation should follow the Wiener-measure and Feynman-Kac references.

Provenance. Base article QUANT.RQM.EUCL.1096 (AllQFT "quantum-mechanics" track, Sakurai §2.6 / Feynman-Hibbs). Upgraded to v3 canon; the Wick-rotation identity, the heat-kernel identification, and the oscillator $Z$ are verified in `verify_aq-qm-euclidean-path-integral.py`.

Prerequisites. The Feynman Path Integral Formulation · Free-Particle and Harmonic-Oscillator Path Integrals · Canonical Ensemble and Partition Function · The Time-Evolution Operator. Leads to. Instantons and Quantum Tunneling · lattice field theory · finite-temperature field theory.

Why does Wick rotation make the integral converge when the original oscillates?

Because it replaces a unit-modulus phasor $e^{iS/\hbar}$ with a decaying real weight $e^{-S_E/\hbar}\le1$ . The Minkowski integral converges only through cancellation between rapidly oscillating contributions — mathematically it is a conditionally convergent, distribution-valued object that you cannot sample or bound. After rotation, paths far from the classical one are exponentially suppressed, so the integral converges absolutely and can be estimated by Monte Carlo or saddle point. This is the entire reason lattice field theory and quantum Monte Carlo are formulated in Euclidean time.

Is imaginary time "real" or just a computational trick?

Both, depending on what you ask. As a computational device, you rotate, compute a real Euclidean quantity, and rotate the answer back — imaginary time never appears in a measured number. But the Euclidean theory is also physically meaningful in its own right: with $\tau=\beta\hbar$ it is thermal equilibrium at temperature $T=1/k_B\beta$ . So imaginary-time periodicity is not a fiction; it is temperature. The KMS condition in quantum statistical mechanics is exactly the statement that thermal correlation functions are periodic in imaginary time with period $\beta\hbar$ .

What is the diffusion constant ℏ/2m physically?

It is the rate at which a free quantum particle's imaginary-time "position uncertainty" grows: $\langle x^2\rangle=2D\tau=\hbar\tau/m$ . In real time the same spreading is the wavepacket dispersion, but with the crucial factor of $i$ that turns spreading into interference. Rotating time removes the $i$ and leaves honest Brownian diffusion with $D=\hbar/2m$ . Larger mass diffuses slower, exactly as heavier Brownian particles do — the classical limit $\hbar\to0$ freezes the diffusion entirely, collapsing the path integral onto the classical trajectory.

Why must the paths be periodic to compute the partition function?

Because $Z=\mathrm{Tr}\,e^{-\beta\hat H}=\int\mathrm dx\,\langle x|e^{-\beta\hat H}|x\rangle$ — the same $x$ appears in the bra and the ket. The path integral for $\langle x_f|e^{-\beta\hat H}|x_i\rangle$ runs from $x_i$ to $x_f$ in imaginary time $\beta\hbar$ ; setting $x_f=x_i$ and integrating over that common value forces the path to close into a loop. So the trace geometrically becomes a sum over closed imaginary-time worldlines of circumference $\beta\hbar$ . Fermionic fields, by contrast, get antiperiodic boundary conditions — the origin of the difference between Bose and Fermi thermal distributions.

Does the oscillator Z include or exclude the zero-point energy?

It includes it. $Z=1/[2\sinh(\beta\hbar\omega/2)]=e^{-\beta\hbar\omega/2}/(1-e^{-\beta\hbar\omega})$ — the numerator is exactly the zero-point Boltzmann factor $e^{-\beta E_0}$ with $E_0=\tfrac12\hbar\omega$ . Some textbooks quote $Z=1/(1-e^{-\beta\hbar\omega})$ by choosing to measure energies from the ground state (dropping the $\tfrac12\hbar\omega$ ); that convention shifts the free energy by a constant $\tfrac12\hbar\omega$ but leaves all thermal derivatives (energy, heat capacity) unchanged. The path-integral computation naturally produces the sinh form with the zero-point energy intact.

How does the sign problem relate to this rotation?

The Wick rotation is supposed to make the weight real and positive so it can be sampled. For many systems it does. But for others — QCD at finite baryon density, real-time dynamics, or fermions in frustrated geometries — a residual complex or negative weight survives the rotation, so $e^{-S_E/\hbar}$ is no longer a probability. Then Monte Carlo importance sampling fails catastrophically (the "sign problem"), and you are back to fighting cancellations. Solving or circumventing the sign problem is one of the central open challenges in computational many-body and high-energy physics.

Is the Euclidean action always bounded below?

For a standard nonrelativistic Lagrangian with a potential bounded below, yes: $S_E=\int(\tfrac12 m\dot x^2+V)\,\mathrm d\tau$ is a sum of a nonnegative kinetic term and a bounded-below potential, so $e^{-S_E/\hbar}$ is well-behaved. This positivity is what the whole construction relies on. In gauge theories the Euclidean action is also bounded below and, for instanton sectors, is bounded below by a topological charge (the BPS bound). The pathological cases where $S_E$ is unbounded — certain gravitational actions, for instance — are exactly where the Euclidean path integral becomes subtle or ill-defined.