The Ship of Theseus question is usually posed in the language of wood and nails: if we replace the planks of a ship one by one, is it still the same ship? Physics poses a sharper version because it forces us to be explicit about what counts as a "part." In an open system, "parts" are not only atoms, but also correlations. If a subsystem \(S\) steadily exports correlations into degrees of freedom we cannot access, then the subsystem’s state may become statistically disconnected from its own past. Yet the global evolution remains perfectly unitary.
This note takes that replacement process seriously and proposes an operational surrogate for “continuity of substance.” The surrogate is not mass, not entropy, and not a metaphysical essence, but a quantity that can be defined for any quantum system and any time: how much information the present subsystem retains about its past. When that quantity drops below a chosen threshold, the subsystem has become a new informational entity in a precise sense.
Beta decay is a clean motivating example. The nucleus + decay products evolve unitarily. But if the neutrino is effectively inaccessible, the nuclear degrees of freedom evolve as if they underwent irreversible amplitude damping. The nucleus, viewed as a subsystem, "forgets" being the parent. The question is not whether the universe forgets; global unitarity forbids that. The question is whether \(S\) remains informationally continuous with the earlier \(S\).
Let a composite Hilbert space factorize as \(\mathcal H = \mathcal H_S \otimes \mathcal H_E\). The global state evolves unitarily:
\[ \rho_{SE}(t) = U(t,t_0)\,\rho_{SE}(t_0)\,U^{\dagger}(t,t_0). \]The reduced subsystem state is
\[ \rho_S(t) = \mathrm{Tr}_E[\rho_{SE}(t)]. \]Even if the global map is reversible, the reduced map \(\rho_S(t)=\mathcal E_t(\rho_S(t_0))\) need not be. In the Markovian regime, \(\{\mathcal E_t\}_{t\ge 0}\) forms a CPTP semigroup generated by a Lindblad operator \(\mathcal L\):
\[ \rho_S(t)=\mathcal E_t(\rho_S(t_0)), \\ \frac{d}{dt}\rho_S(t)=\mathcal L(\rho_S(t)). \]To ask whether \(S\) “remains itself,” we must compare present and past in a way that respects quantum mechanics. Introduce an abstract reference system \(R\) that is initially correlated with \(S\). Let \(\rho_{RS}(t_0)\) be any joint state, often a purification of \(\rho_S(t_0)\). The reference does not evolve, and the joint state evolves as
\[ \rho_{RS}(t) = (\mathrm{Id}_R \otimes \mathcal E_t)(\rho_{RS}(t_0)). \]Define the past-present mutual information (logs base 2 unless otherwise noted)
\[ \boxed{ \mathcal I_S(t) = I(R:S)_t = S(\rho_R) + S(\rho_S(t)) - S(\rho_{RS}(t)), } \]where \(S(\rho)=-\mathrm{Tr}(\rho\log\rho)\) is the von Neumann entropy. Conceptually, \(\mathcal I_S(t)\) measures how much of the subsystem’s earlier identity, as encoded in \(R\), is still present in \(S\) at time \(t\). It is the natural quantum analogue of “memory of the past,” but expressed without narratives: it is a functional of the joint state.
This monotonicity is the structural core of the note. It says: an open subsystem cannot increase its correlations with its past reference. Whatever is lost has been exported into degrees of freedom outside the subsystem’s boundary (and often outside our practical reach). The result is not thermodynamic. It does not rely on ensembles, coarse-graining, or typicality. It is simply what CPTP maps do to correlations.
In the semigroup setting, one may define an instantaneous identity loss rate:
\[ \Phi(t) := -\frac{d}{dt}\mathcal I_S(t) \ge 0. \]On this view, "irreversibility" is not a mysterious physical principle; it is the monotone decay of a specific correlation functional. The universe remains reversible; the subsystem does not. This is the subsystem version of the Theseus replacement story: as correlations are swapped out of the subsystem and into \(E\), the subsystem's present state becomes less and less informative about its earlier instantiation.
For Markovian dynamics, the Lindblad generator can be written
\[ \mathcal L(\rho) = -i[H,\rho] + \sum_k \left(L_k \rho L_k^\dagger - \tfrac12\{L_k^\dagger L_k,\rho\}\right). \]A direct differentiation of \(\mathcal I_S(t)\) yields an exact identity for the rate of mutual-information change:
\[ \boxed{ \frac{d}{dt}\mathcal I_S(t) = \mathrm{Tr}\!\left[(\mathrm{Id}_R\otimes\mathcal L)(\rho_{RS}(t))\,\Big(\log\rho_{RS}(t)-\log\rho_R-\log\rho_S(t)\Big)\right]. } \]This equation is useful not because it always simplifies, but because it localizes where irreversibility lives: the dissipators \(L_k\). The Hamiltonian part of the generator is reversible and (in the appropriate sense) does not produce entropy; it rotates information rather than exporting it. The dissipators are the mechanism by which correlations with \(R\) are shunted into \(E\). When one wants a more “microscopic” picture of identity loss, this is it.
To keep the discussion concrete, model the nucleus as a two-level system \(S\) with basis \(\{|e\rangle,|g\rangle\}\) representing parent and daughter states. Treat beta decay as amplitude damping with rate \(\Gamma\). The decay probability by time \(t\) is
\[ p(t)=1-e^{-\Gamma(t-t_0)}. \]A Kraus representation of the reduced dynamics is
\[ \begin{align} E_0 &= |g\rangle\!\langle g| + \sqrt{1-p}\,|e\rangle\!\langle e|, \\ E_1 &= \sqrt{p}\,|g\rangle\!\langle e|, \\ \rho_S(t)&=\sum_{a\in\{0,1\}} E_a\,\rho_S(t_0)\,E_a^\dagger. \end{align} \]Now correlate the nucleus with a reference system. A natural “maximally informative” choice is a maximally entangled initial state
\[ |\Phi^+\rangle_{RS}=\frac{1}{\sqrt2}(|gg\rangle+|ee\rangle). \]Then \(\rho_{RS}(t)=(\mathrm{Id}_R\otimes\mathcal E_t)(|\Phi^+\rangle\langle\Phi^+|)\). One can compute the relevant entropies in closed form (logs base 2):
\[ \begin{align} S(\rho_R)&=1, \\ S(\rho_S(t))&=h_2\!\left(\frac{1+p}{2}\right), \\ S(\rho_{RS}(t))&=h_2\!\left(\frac{p}{2}\right), \end{align} \]where \(h_2(x)=-x\log_2 x-(1-x)\log_2(1-x)\) is the binary entropy. Therefore
\[ \boxed{ \mathcal I_S(t)=1+h_2\!\left(\frac{1+p(t)}{2}\right)-h_2\!\left(\frac{p(t)}{2}\right). } \]This formula has the correct qualitative behavior: \(\mathcal I_S(t_0)=2\) bits and \(\lim_{t\to\infty}\mathcal I_S(t)=0\). The nucleus, treated as a subsystem, ends up with no mutual information with its past. The global unitary evolution still contains all information, but it has moved into degrees of freedom we have excluded (in practice: a neutrino plus other correlations radiating away). In Theseus language: the “planks” replaced are the correlations; the subsystem has gradually become a different informational object.
Thus the usual population half-life coincides with an identity half-life defined purely in terms of mutual information with the past. The point is not that we have discovered a new parameter of beta decay; rather, we have reframed what the half-life is measuring: it is the characteristic timescale on which the subsystem loses half of its past-present correlations.
Decoherence is often described as "loss of phase information" in a preferred basis. In the present framework, that description becomes quantitative. Decoherence is the mechanism that drives \(\rho_{RS}(t)\) toward a state where all correlations are classical in the pointer basis. Importantly, decoherence alone need not force \(\mathcal I_S(t)\to 0\); it forces the coherent contribution to the mutual information to die.
Fix a pointer basis \(\{|i\rangle\}_{i=1}^{d_S}\) for \(S\) and consider a pure-dephasing Lindblad generator with diagonal Lindblad operators
\[ L_k=\sqrt{\gamma_k}\sum_i \ell_{k,i}\,|i\rangle\langle i|. \]Then the reduced matrix elements satisfy
\[ \begin{align} \rho_{ii}(t)&=\rho_{ii}(0), \\ \rho_{ij}(t)&=e^{-\Gamma_{ij}t}\,\rho_{ij}(0) \ (i\neq j), \\ \Gamma_{ij}&=\frac12\sum_k\gamma_k\,|\ell_{k,i}-\ell_{k,j}|^2. \end{align} \]Let \(\Gamma_{\min}=\min_{i\neq j}\Gamma_{ij}\). Define the dephasing (pinching) map
\[ \Delta(\rho)=\sum_i (I_R\otimes |i\rangle\langle i|)\,\rho\,(I_R\otimes |i\rangle\langle i|). \]A convenient measure of “how far \(\rho_{RS}\) is from being classical in the pointer basis” is the trace distance \(\|\rho_{RS}-\Delta(\rho_{RS})\|_1\). One can bound this distance by a suitable off-diagonal block coherence, and under dephasing it decays exponentially:
\[ \varepsilon(t):=\frac12\|\rho_{RS}(t)-\Delta(\rho_{RS}(t))\|_1 \le \frac12 e^{-\Gamma_{\min}t}\,C_0, \]where \(C_0\) depends only on the initial off-diagonal blocks of \(\rho_{RS}(t_0)\). Now let
\[ \begin{align} I(t)&:=I(R:S)_{\rho_{RS}(t)}, \\ I_{\mathrm{cl}}(t)&:=I(R:S)_{\Delta(\rho_{RS}(t))}. \end{align} \]Since \(\Delta(\rho_{RS}(t))\) is classical-quantum (CQ) in the pointer basis, \(I_{\mathrm{cl}}(t)\) is exactly the mutual information between \(R\) and the pointer outcome stored in \(S\). Mutual information continuity bounds (of Alicki-Fannes type) yield
\[ \boxed{ |I(t)-I_{\mathrm{cl}}(t)| \le 4\varepsilon(t)\log d_S + 2h_2(\varepsilon(t)). } \]Combining with the exponential envelope for \(\varepsilon(t)\) shows that decoherence rates control how fast the mutual information is driven toward its classical part. In intuitive terms, dephasing replaces "quantum identity continuity" with "classical record continuity" on the timescale \(1/\Gamma_{\min}\). If one wants the subsystem to become a genuinely new entity by the stronger criterion \(\mathcal I_S(t)\approx 0\), then decoherence must be accompanied by mixing / relaxation mechanisms that erase even the classical record.
A clean collapse-time bound falls out immediately. If one defines "coherence-mediated identity collapse" by \(|I(t)-I_{\mathrm{cl}}(t)|\le\delta\), then one may choose \(\varepsilon_\delta\) solving \(4\varepsilon_\delta\log d_S+2h_2(\varepsilon_\delta)=\delta\). Any time satisfying
\[ t \ge \frac{1}{\Gamma_{\min}}\log\!\left(\frac{C_0}{2\varepsilon_\delta}\right) \]is sufficient. Up to a logarithmic factor, collapse time scales like \(1/\Gamma_{\min}\).
To obtain full informational collapse, one typically needs a mechanism that contracts all correlations, not merely coherences. This includes amplitude damping, thermalization, and more general mixing semigroups. A common sufficient condition is that the Lindblad generator has a spectral gap \(\lambda>0\) toward a unique stationary state and satisfies a strong data-processing inequality. In that case one can bound mutual information by an exponential contraction:
\[ \boxed{ \mathcal I_S(t) \le \mathcal I_S(t_0)\,e^{-c\lambda t}, } \]where \(c>0\) depends on the functional inequality (e.g., logarithmic Sobolev constants) satisfied by the semigroup. The specific constant is less important than the shape: a single gap parameter controls the forgetting time-scale. In these models, identity does not merely become classical; it disappears altogether. The subsystem becomes independent of its past reference.
This is the cleanest mathematical analogue of the Theseus replacement story. Replace the planks slowly enough and at some stage the ship is still "the same" for practical purposes, but if the replacement process continues and all original structure is washed out, the continuity claim becomes empty. Here the statement is quantitative: \(\mathcal I_S(t)\) measures the remaining continuity.
To turn the functional \(\mathcal I_S(t)\) into an entity criterion, choose an operational threshold \(\varepsilon>0\) representing “indistinguishable from disconnected” at the precision of interest. Define the identity-loss time
\[ \boxed{ \tau_{\varepsilon} := \inf\{t\ge t_0:\ \mathcal I_S(t)\le \varepsilon\}. } \]The subsystem persists as the same informational entity over the interval \([t_0,t)\) whenever \(\mathcal I_S(t)>\varepsilon\). When \(\mathcal I_S(t)\le\varepsilon\), the subsystem is operationally disconnected from its past. This criterion is deliberately modest: it does not declare a metaphysical discontinuity; it declares that the subsystem has lost the correlations required to justify treating it as “the same object” in an information-theoretic sense.
The threshold framing is also what makes the Ship of Theseus analogy work: "sameness" is not a binary predicate but a tolerance-dependent decision rule. The function \(\mathcal I_S(t)\) supplies the quantitative backbone; the threshold supplies the pragmatic boundary.
The reference system \(R\) can be interpreted abstractly (a purification register) or physically (an observer or apparatus record). If one replaces \(R\) by an observer system \(O\), relational descriptions emerge as conditional states. Let \(\{\Pi_o\}\) be a POVM on \(O\). Then the conditional state of \(S\) given outcome \(o\) is
\[ \boxed{ \rho_{S|o}(t) = \frac{\mathrm{Tr}_O\big[(I\otimes\Pi_o)\rho_{SO}(t)\big]}{\mathrm{Tr}\big[(I\otimes\Pi_o)\rho_{SO}(t)\big]}. } \]In this view, "relational quantum states" require no new dynamics: they are conditionalizations inside correlated open-system evolution. The same mutual-information functional applies: one may track \(I(O:S_t)\) as a measure of how robustly the observer's record remains correlated with the system. Decoherence stabilizes the record in a pointer basis; mixing erases it.
The same mathematical skeleton appears across a long list of physical effects: decoherence (collapse of coherent contributions), spontaneous emission and amplitude damping (loss to radiation modes), beta decay (loss to weakly interacting channels), thermalization (mixing to a stationary state), measurement (record formation plus subsequent stabilization), and horizon-like situations (correlations exported beyond causal access).
\[ \boxed{ \text{Open-system CPTP dynamics + monotone decay of } I(R:S_t). } \]What is unified is not a new force or law. It is a shared structure: correlations with a reference encoding the past are monotonically contracted by local CPTP evolution. The “arrow of time” for subsystems is simply the data-processing inequality applied dynamically.
The Ship of Theseus is often treated as a problem about “substance.” In open quantum systems, the closest analogue of substance is not matter but correlation structure. A subsystem is defined by a boundary; open dynamics constantly trades correlation across that boundary. When the subsystem retains substantial mutual information with a reference encoding its past, it is justified, operationally speaking, to treat it as the same entity. When that mutual information decays below threshold, the justification evaporates.
Beta decay supplies an exactly solvable illustration: an informational half-life coincides with the conventional half-life, but the interpretation changes. The half-life is the timescale on which the subsystem loses half of its past-present correlations. Decoherence supplies the complementary lesson: dephasing destroys coherent identity rapidly, but may leave a classical residue. Complete identity collapse requires mixing.
The content of the framework can be compressed to a single statement:
\[ \boxed{ \text{Entity persistence in quantum physics is measured by } I(R:S_t), \text{ and irreversible processes correspond to its monotone decay.} } \]What remains open, philosophically and technically, is how far this viewpoint can be pushed before it becomes genuinely predictive rather than merely organizing. For horizons, cosmology, or field-theoretic settings, “inaccessibility” becomes a structural constraint rather than a practical inconvenience. In such regimes, mutual-information decay may serve as a sharp bridge between operational irreversibility and the geometry of what can be known.