Finite-Time Thermodynamic Uncertainty Relations

• Finite-Time TURs are non-asymptotic bounds that relate the relative fluctuations of currents to entropy production in finite observation windows, revealing a precision–dissipation tradeoff in stochastic systems.
• They employ continuous-time Markov jump processes, many-copy ensembles, and path-space large deviation principles to quantify and constrain fluctuations and thermodynamic costs.
• Extensions to transient, quantum, and discrete-time models demonstrate the limits of classical TURs and motivate more generalized bounds under non-steady and non-Markovian conditions.

Finite-time thermodynamic uncertainty relations (TURs) are non-asymptotic bounds that connect the relative fluctuations of a current or other odd observable measured over a finite interval to entropy production or, in broader formulations, to a path-space asymmetry functional. In the canonical continuous-time nonequilibrium steady-state setting, if $X(t)$ is an integrated current, $J=\langle X(t)\rangle/t$ its stationary mean rate, and $\sigma$ the steady-state entropy production rate, the finite-time TUR takes the form

$\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$

equivalently

$$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$

This finite-time extension was first proposed for continuous-time nonequilibrium steady states and then proved for steady-state currents in continuous-time Markov jump processes (Pietzonka et al., 2017, Horowitz et al., 2017).

1. Canonical formulation and precision–dissipation tradeoff

In its standard form, the finite-time TUR concerns a continuous-time Markov jump process on a finite network with time-homogeneous rates and a stationary distribution. An integrated current $X(t)$ is defined by assigning an antisymmetric increment $d_{ij}=-d_{ji}$ to each jump $i\to j$, so that $\langle X(t)\rangle=Jt$ and $\mathrm{Var}[X(t)]$ is the finite-time variance of the accumulated current. The bound

$J=\langle X(t)\rangle/t$0

states that a current cannot be simultaneously precise and cheap: suppressing relative fluctuations requires dissipation (Pietzonka et al., 2017).

The same relation may be read as an inference inequality for entropy production,

$J=\langle X(t)\rangle/t$1

which is why finite-time TURs became attractive for systems where long trajectories are inaccessible. The finite-time formulation differs from the long-time large-deviation TUR not by changing the structure of the inequality, but by replacing asymptotic diffusion coefficients with directly measurable mean and variance on a finite observation window. In the original steady-state discussion, this made the bound relevant to finite-window work statistics in a stochastically switching optical trap, where non-Gaussian finite-time distributions still respected the inequality (Pietzonka et al., 2017).

A common later notation rewrites the left-hand side as an inverse precision or scaled variance. For an odd current $J=\langle X(t)\rangle/t$2, one defines

$J=\langle X(t)\rangle/t$3

In that language, the classical Barato–Seifert form becomes $J=\langle X(t)\rangle/t$4, making explicit that finite-time TURs are precision–dissipation tradeoffs rather than purely asymptotic fluctuation statements (Ray et al., 2022).

2. Proofs, exact finite-time results, and all-time reformulations

The rigorous proof of the steady-state finite-time TUR for continuous-time Markov jump processes proceeds by replacing the long-time limit with a large ensemble of $J=\langle X(t)\rangle/t$5 independent copies observed for fixed time $J=\langle X(t)\rangle/t$6. The empirical density and empirical current of the $J=\langle X(t)\rangle/t$7-copy ensemble satisfy a path-space large deviation principle, and a quadratic upper bound on the corresponding rate functional yields, after contraction to a scalar generalized current $J=\langle X(t)\rangle/t$8,

$J=\langle X(t)\rangle/t$9

The proof is explicitly steady-state, continuous-time, finite-state, and Markovian; its significance is that the finite-time inequality is not merely conjectural in that setting (Horowitz et al., 2017).

This many-copy construction also clarifies why the result is genuinely finite-time. Large deviations enter through the number of copies rather than through $\sigma$0, so the theorem constrains fluctuations on every fixed observation window $\sigma$1. For the entropy production current itself, the same framework yields $\sigma$2, which is the direct entropy-production specialization of the general finite-time TUR (Horowitz et al., 2017).

A distinct non-asymptotic reformulation was later obtained from the Euclidean geometry of the space of observables. For observables antisymmetric under an involution, especially time reversal, the maximal normalized precision is exactly

$\sigma$3

and an entropy-only consequence is

$\sigma$4

These bounds are explicitly valid for arbitrary dynamics, all time intervals $\sigma$5, and all antisymmetric observables, and they recast finite-time TURs as consequences of symmetry breaking rather than exclusively as large-deviation statements about steady currents (Falasco et al., 2019).

3. Transients, arbitrary initial states, and time-dependent driving

Once the steady-state assumption is dropped, the signal entering the TUR is no longer necessarily the time-averaged integrated current. For continuous-time Markov jump processes with arbitrary initial states, Liu, Gong, and Ueda derived the finite-time bound

$\sigma$6

where $\sigma$7 is the ensemble-averaged instantaneous current at the final time. In this formulation, “the boundary is constrained by the bulk”: the current at the temporal boundary $\sigma$8 is bounded by fluctuations and dissipation accumulated over the whole interval $\sigma$9. When the initial state is stationary, $\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$0 becomes time-independent and the bound reduces to the conventional steady-state TUR (Liu et al., 2019).

The same work also gives a discrete-time analogue for Markov chains with arbitrary non-steady initial states,

$\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$1

with $\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$2 a modified entropy production and $\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$3 the minimal staying probability. This is linear in total entropy production and is explicitly sharper than earlier discrete-time bounds based on exponentiated entropy production (Liu et al., 2019).

For arbitrarily time-dependent driving, the finite-time TUR acquires a response term to protocol-speed rescaling. If the protocol is $\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$4, one defines

$\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$5

Then for current observables,

$\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$6

and for endpoint or time-integrated state observables $\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$7,

$\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$8

These relations are finite-time, non-steady-state, valid for arbitrary initial conditions, and extend TURs beyond currents to state variables in overdamped Langevin systems and continuous-time Markov jump processes (Koyuk et al., 2020).

4. Failure of naive universality and explicit violations

Finite-time TURs do not constitute a single universal inequality valid under arbitrary changes of dynamics, time structure, or observable class. The sharpest illustration is Shiraishi’s counterexample to the natural discrete-time analogue of the continuous-time finite-time TUR. For a stationary two-step Markov chain with two states and two distinct channels, the conjectured discrete-time bound

$\frac{\mathrm{Var}[X(t)]\,\sigma}{J^2 t}\ge 2k_B,$9

fails. In the explicit construction,

$$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$0

which is less than $$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$1 for $$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$2; at $$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$3 one gets $$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$4. The result does not disprove finite-time TURs in continuous time, but it shows that any correct proof must use structure genuinely absent in discrete-time chains (Shiraishi, 2017).

Violations of classical finite-time TUR benchmarks also appear in finite-time cyclic and quantum settings outside the original steady-state Markov domain. In an exactly solvable harmonic-oscillator Otto cycle, both quantum and classical finite-time cycles violate the conventional bound $$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$5 in the small-entropy-production regime, especially near resonance conditions, even though a tighter quasistatic bound holds at larger dissipation (Lee et al., 2020). In a transient two-qubit heat-exchange experiment, generalized fluctuation-theorem-based TURs were always obeyed, while the specialized tighter TUR

$$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$6

was violated in certain finite-time regimes, in agreement with analytic results (Pal et al., 2019). A quantum collisional model exhibits a related pattern: the classical finite-time TUR parameter

$$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$7

can drop below $$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$8 after a threshold number of collisions, with the strongest violation near the nonequilibrium steady state in the Markovian regime and highly mechanism-dependent behavior in non-Markovian regimes (Maity et al., 2024).

These examples do not show that all finite-time TURs fail. They instead delimit the scope of the classical steady-state formula and motivate broader finite-time inequalities adapted to discrete time, transient driving, non-Markovianity, or quantum dynamics.

5. Fluctuation-theorem, higher-moment, and optimal-current formulations

A major branch of the subject derives finite-time TURs directly from fluctuation theorems. For exchange processes satisfying an exchange fluctuation theorem, the tight scalar bound is

$$\frac{\mathrm{Var}[X(t)]}{\langle X(t)\rangle^2}\ge \frac{2k_B}{\sigma t}.$$9

where $X(t)$0 is the inverse function of $X(t)$1. The same framework yields the matrix inequality

$X(t)$2

which constrains not only variances but also covariances of exchanged heat, work, and particle currents in finite-time classical and quantum exchange processes (Timpanaro et al., 2019).

This exchange-TUR structure can be derived even more abstractly from involutions. If $X(t)$3 is an involution on a set of processes and $X(t)$4 is odd under $X(t)$5, then

$X(t)$6

with the same $X(t)$7. In this formulation, fluctuation theorems are no longer the starting point; they enter only when the involution-based KL asymmetry is identified with entropy production (Salazar, 2022).

A more restrictive but exact finite-time theorem holds for time-symmetrically controlled computations (TSCCs), defined by a time-symmetric control protocol and conjugate initial/final distributions. For any current odd under time reversal,

$X(t)$8

and the minimizing current is explicitly

$X(t)$9

This “Thermodynamic Uncertainty Theorem” is finite-time, saturable, and depends on the full stochastic entropy-production distribution rather than only on $d_{ij}=-d_{ji}$0 (Ray et al., 2022).

The role of higher moments is made fully explicit in later fluctuation-theorem families. For a generic detailed fluctuation theorem with possibly different forward and backward processes,

$d_{ij}=-d_{ji}$1

where

$d_{ij}=-d_{ji}$2

The bound is finite-time, valid in classical and quantum regimes, and saturable; its dependence on $d_{ij}=-d_{ji}$3 encodes higher-order moments of entropy production rather than only its mean (Timpanaro, 2024). A related arbitrary-time derivation based on an exchange fluctuation theorem and an ensemble of copies recovers the standard form

$d_{ij}=-d_{ji}$4

at arbitrary finite time, subject to a regularity condition on the transfer distribution (Monnai, 2022).

6. Non-Markovian dynamics, generalized costs, and adjacent developments

For non-Markovian delayed Langevin systems, finite-time TURs can be formulated directly in path space. If $d_{ij}=-d_{ji}$5 is the log-ratio between forward and conjugate path probabilities, then $d_{ij}=-d_{ji}$6, and for any observable $d_{ij}=-d_{ji}$7 odd under the chosen conjugation,

$d_{ij}=-d_{ji}$8

With time reversal this yields a finite-time bound in terms of a generalized entropy production $d_{ij}=-d_{ji}$9; with position reversal it yields an analogous bound in terms of a symmetry-breaking quantity $i\to j$0, which can remain positive even at equilibrium. The derivation requires only path probabilities on a finite interval and not an explicit Markovian embedding (Vu et al., 2019).

A different generalization is needed when the network itself contains irreversible links. For continuous-time Markov jump processes with unidirectional transitions, the finite-time TUR becomes

$i\to j$1

The denominator is no longer purely thermodynamic: it mixes reversible-edge entropy production with the total flux of irreversible edges. This extension covers transient processes with arbitrary initial conditions and includes resetting and Michaelis–Menten catalysis as worked examples (Pal et al., 2020).

Closely related, but not identical to standard current TURs, are finite-time information-geometric speed limits. For a time-dependent distribution $i\to j$2, the Fisher information

$i\to j$3

bounds the intrinsic timescale $i\to j$4 of any observable $i\to j$5 through

$i\to j$6

These results constrain transient rates of change of heat, entropy, entropy production, entropy flow, and dissipated work, and are best regarded as adjacent finite-time uncertainty relations rather than standard TURs for integrated currents (Nicholson et al., 2020).

Taken together, these developments show that finite-time thermodynamic uncertainty relations form a stratified theory rather than a single formula. In continuous-time steady-state Markov jump processes, the classical bound $i\to j$7 is rigorous. Outside that domain, the correct finite-time relation may involve final-time currents, protocol-speed response terms, full entropy-production distributions, path-space KL divergences, exchange-fluctuation symmetries, dynamical activity, or quantum corrections. The modern literature therefore treats “finite-time TUR” as a family of non-asymptotic precision–cost bounds whose exact form depends on dynamical structure, symmetry class, and observable class (Horowitz et al., 2017, Liu et al., 2019, Ray et al., 2022).