WorksPhysics and Astronomy
Stationary Quantum Time Series: Spectral Theory of Process Memory and Measurement Records
Abstract
Classical time series analysis converts temporal dependence into structure that can be learned from a single observed trajectory, through stationarity, ergodicity, spectral representation, and finite-memory models. Quantum experiments are naturally time series: every run is a sequence of interventions and outcomes, and the memory of an open system is written in its temporal statistics. However, the classical foundation cannot be translated directly: measuring a process and forgetting the outcome differs from not measuring at all, so no single family of joint probability distributions describes a quantum process. We develop a process-tensor framework for stationary quantum time series and establish its first structural theorems. For stationary combs of finite quantum Markov order l relative to a fixed informationally complete causal break, we construct a finite-dimensional causal transfer map of dimension at most d^{4l}, where d bounds the system Hilbert-space dimension at each step. ... These results supply the stationary foundation on which quantum spectral analysis, model identification, and single-trajectory inference can be built.
Introduction
Time series analysis is the statistical theory of ordered data, and it provides one of the most successful examples of structure learned from a single observed trajectory. A stationary process is one whose statistical law is invariant under time translation. The Kolmogorov extension theorem guarantees that a consistent family of finite-dimensional distributions defines a single process [1]; Birkhoff’s ergodic theorem equates time averages along one trajectory with ensemble averages [2], making single-trajectory inference possible; the Wiener–Khinchin spectral representation turns temporal correlation into a frequency-domain object [3, 4]; and finite-memory parametrized families, such as Markov chains and ARMA models, make the theory estimable in practice [5]. Together, these results converted temporal dependence from a nuisance into structure: forecasting, system identification, signal processing, and control all rest on them. The conceptual lesson is that the way a process remembers and forgets is itself a learnable, classifiable object.
Quantum information theory is built on dynamical primitives. A quantum state is the quantum analogue of a random variable, and a quantum channel that of a stochastic map [6]. The temporal evolution of a quantum system, especially an open one, is generically non-Markovian: the environment remembers. Continuous monitoring of quantum systems produces measurement records whose temporal correlations carry information about the dynamics [7, 8, 9, 10], and every quantum experiment, from benchmarking to sensing, is literally a time series of interventions and outcomes. Several pieces of a quantum theory of temporal structure are already known: the process tensor characterizes non-Markovian multi-time processes [11, 12, 13], quantum trajectory theory analyzes continuously measured systems [8, 9, 10], quantum Markov chains have their own statistical theory [14, 15], and the algebraic quantum probability tradition defines stationary quantum processes abstractly [32, 33].
However, none of these pieces supplies the operational time-series level of organization: a theory whose objects are stationary processes, whose questions concern ergodicity, spectra, and finite-memory models, and in which all of the above pieces sit as special cases. Such a theory would unify the description of open-system memory, monitored dynamics, and device-level temporal correlations in the way classical time series analysis unifies forecasting and identification.
The obstruction to such a theory is foundational. The Kolmogorov consistency condition, marginalization, reads “measure and forget the outcome”; quantum-mechanically, this differs from “not measuring at all,” because measurement disturbs the system. Consequently, no single family of joint probability distributions describes a quantum process: outcome statistics depend on which instruments were applied, and incompatible choices cannot be embedded in one distribution. Two correlates follow. First, there is no canonical two-time correlation: what happens between the two times, idle evolution or measurement, changes the answer, so the classical autocorrelation function has no unique quantum descendant. Second, ergodicity becomes subtle: if observation backacts on the process, the sense in which a single trajectory represents the ensemble needs re-examination. Any quantum time series theory must solve these problems before it can ask the classical questions.
The framework we build on has two ancestries. On the operational side, quantum combs were introduced by Chiribella, D’Ariano, and Perinotti [20, 21] and were developed into a complete account of non-Markovian multi-time processes by Pollock et al. [11, 12] (reviewed in [13, 22]); the existence of infinite processes rests on the generalized Kolmogorov extension theorem (GET) of Milz et al. [16], which plays the role of the Kolmogorov extension theorem in our construction. The notion of memory we use is the quantum Markov order of Taranto et al. [17, 18], whose recoverability aspects were studied in [23], and combs as higher-order operations are treated in [24]. Time-translation-invariant process tensors have appeared in the numerical literature [25, 26], where translation invariance provides an ansatz that accelerates computations; our work is complementary to these studies, in that stationarity is imposed as an operational symmetry on the comb family itself.
On the mathematical side, our transfer map parallels the finitely correlated states of Fannes, Nachtergaele, and Werner [19], constructed here for interventions rather than states, and memory channels [27] provide a complementary input-output picture. Hidden quantum Markov models [28, 29] and matrix-product approaches to stochastic modelling [30] correspond to the single-instrument, order-one sector of what we call the record layer. The difference is one of domain: these constructions attach a finite-dimensional memory to a state or to a single fixed instrument, whereas the transfer map built here resolves every intervention element on a stationary comb, which is what the spectral criteria and the record ergodic theorem require.
On the ergodic-theory side, the study of quantum measurement records was pioneered by Belavkin [7] and by Kümmerer and Maassen [8, 9] and continues through [10, 14, 15, 31], all in settings where the environment carries no process-tensor memory; our record ergodic theorem extends this tradition to arbitrary finite Markov order, with the memory spectrum setting the rates. For collision models specifically, the repeated-interaction literature established approach-to-stationarity and return-to-equilibrium results [54] on the same physical objects; those results concern state-level relaxation in concrete Hamiltonian models, while the spectral criteria here concern the operational comb. The algebraic quantum probability tradition (Accardi, Frigerio, and Lewis [32], Kümmerer [33]) defines stationary quantum processes non-operationally; our approach is complementary in that every object is defined directly in terms of interventions. Closest in spirit to our two-level structure, Taranto and Milz showed that records Markovian under every probing can still require memory in the process [34], and Milz et al. studied when the process layer admits a classical description at all [35]; both separations concern Markovianity, while ours concern mixing rates and spectra.
In this paper, we build that theory on the process-tensor foundation and prove its first structural theorems. The basic object is a family of quantum combs, marginalization is replaced by identity insertion (do nothing), the causally correct neutral operation [11, 16], and the GET then provides the existence of infinite processes. On this foundation, the classical questions can be posed operationally; we answer them for the simplest nontrivial model class, stationary processes with finite quantum Markov order relative to a fixed informationally complete (IC) causal break [17, 18]. Concretely, we address the following questions:
(Q1) What is a stationary quantum process, operationally? (Sec. IV)
(Q2) Does a stationary quantum process with finite quantum Markov order admit a finite-dimensional transfer map representation, analogous to the finitely correlated states of Fannes, Nachtergaele, and Werner [19]? (Sec. V)
(Q3) What are the spectral criteria for ergodicity and mixing, and what replaces the autocorrelation function? (Sec. VI)
(Q4) When can a single measurement trajectory be used to infer the process; in other words, is there a quantum ergodic theorem with rates? (Sec. VI, Theorem VI.6)
We answer all four questions, and the answers carry genuinely quantum phenomena: an algebraic signature of measurement backaction, a two-level splitting of the autocorrelation function, and an independence of the two levels of mixing with no classical analogue. The answers are summarized in prose in Sec. II, with pointers to the precise statements.
This paper is organized as follows. In Sec. II, we summarize the main results in prose, with pointers to the precise statements and proofs. In Sec. III, we fix notation and recall quantum combs and Markov order. In Sec. IV, we define stationary quantum time series and prove their unique covariant extension. In Sec. V, we construct the causal transfer map. In Sec. VI, we develop the spectral theory: the two autocorrelations, the ergodicity and mixing criteria, and the record ergodic theorem. Section VII collects physical readings of the theory, Sec. VIII reports the collision-model study, and Sec. IX collects open problems. Appendices A–E give the proofs and the numerical details, each headed by the main-text statement it supplies.
Scope and assumptions. Four levels of restriction must be kept distinct. (i) The finite-dimensional representation assumes finite quantum Markov order . (ii) That Markov order is relative to one fixed, full-rank IC causal break; it is not a statement about every possible probing instrument. (iii) The matrix order, order-unit property, and complete positivity are established on the operational cone generated by measure-and-prepare interventions and identity insertions, although the functional-space dimension bound and the generating formula are defined for arbitrary intervention elements. (iv) The extension of the operational order to coherent interventions is an open problem, formulated below as Conjecture IX.1, not an established conclusion. The spectral criteria for intrinsic process memory use the idle transfer map; the record limit theorems additionally concern a fixed instrument and its instrument-dressed map.
Summary of results
In this section, we provide an informal summary of our main results for the model class studied in this paper: stationary quantum time series of finite quantum Markov order relative to a fixed IC causal break, the simplest nontrivial model class of the theory. The mathematically rigorous formulation of all statements, together with their proofs, will be given in the subsequent sections.
a. Main result 1 (stationary quantum time series exist; Definition IV.1 and Proposition IV.2). The basic object is a family of quantum combs, one per finite time window, consistent under identity insertion and covariant under the time shift (Definition IV.1). Every such family extends uniquely to a two-sided infinite process (Proposition IV.2, proved in Appendix A). This is the operational replacement for the Kolmogorov extension theorem, whose classical consistency condition fails quantumly; in other words, it is what makes the classical questions about stationary processes well posed for quantum experiments.
b. Main result 2 (finite Markov order means finite-dimensional memory; Proposition V.3). Every stationary quantum time series of Markov order relative to the fixed informationally complete causal break admits a generating triple with (Proposition V.3), where is the system Hilbert-space dimension, or an upper bound on it if the local spaces vary: all multi-time statistics factor through a finite-dimensional memory, , with one update map per intervention element. The memory space is intrinsic, the span of all conditional future functionals. Its order-unit and complete-positivity structure is established on the operational cone generated by measure-and-prepare interventions and identity insertions (Lemma V.5); extension to coherent interventions is conjectural (Conjecture IX.1). Within this structure, the identity intervention is the unique intervention whose transfer map is unital for every process in the class (Remark V.2); the normalization axiom is exactly the causality of the comb.
c. Main result 3 (the peripheral spectrum is the intrinsic process memory; Theorems VI.3–VI.5). The autocorrelation splits into an intrinsic process-memory correlation across idle gaps, governed by the idle transfer map , and an instrument-specific record correlation, governed by the instrument-dressed map (Definition VI.1). For the idle dynamics, ergodicity holds if and only if the eigenvalue of is simple (Theorem VI.3), and exponential mixing holds if and only if the peripheral spectrum is trivial. The intrinsic process-memory rate is set by the subleading eigenvalue modulus of , up to polynomial factors from Jordan blocks (Theorems VI.4 and VI.5, proved in Appendix C).
d. Main result 4 (single-trajectory inference at the record level; Theorem VI.6). Fix a finite-outcome measure-and-prepare instrument and a bounded real observable of its record. If the instrument-dressed map has a spectral gap, the measurement record is exponentially -mixing and satisfies separately stated law-of-large-numbers and central-limit conclusions, with asymptotic variance and the degenerate case treated explicitly (Theorem VI.6). The resulting single-trajectory inference rate is instrument-specific: it is controlled by the spectrum of , not by the intrinsic idle-memory spectrum of .
e. Main result 5 (finite-dimensional consistency checks; Sec. VIII). Closed-form finite-dimensional computations and double-precision diagonalization on qubit collision models of Markov order 2 check the implementation of the spectral formulas for instrument-specific record decay rates to five to six decimal places (Sec. VIII, with formulas and sanity checks in Appendix E); they are not independent validations of the general theorems. The examples also display the two-level asymmetry in both directions: a record can be i.i.d. while the idle process fails to mix, and a coherent instrument can sustain a record correlation indefinitely in a process whose idle dynamics mixes.
The table below situates the framework relative to classical time-series concepts and collects the scope qualifications stated above.
Preliminaries
In this section, we fix notation and recall the two ingredients on which the theory rests: quantum combs, which play the role of joint distributions, and quantum Markov order, which plays the role of finite memory. Readers already fluent in the process-tensor formalism may proceed to Sec. IV.
| Classical time series | Quantum time series (this work) | Classical counterpart? |
| Joint distributions + Kolmogorov consistency | Comb family + identity-insertion consistency (GET) | Fails quantumly |
| Autocorrelation function | Two levels: idle-gap (process level, ) and record level () | Splits quantumly |
| Transition/transfer structure | Causal transfer map; identity insertion is the unique intervention unital for every process in the class | Partial |
| Markov order | Instrument-relative quantum Markov order ; memory dimension | Generalizes |
| Ergodicity and mixing | Spectral criteria on ; mixing at the two levels is independent (both implications fail) | New asymmetry |
| — | Coherent-intervention obstruction (temporal interference, Conjecture IX.1) | No counterpart |
Table 2.
Process tensors and combs
At each time the experimenter may apply an instrument : a collection of trace-nonincreasing CP maps summing to a channel [36, 37]. We work in the Choi picture [38, 39]: an intervention element is an operator , and iff it is the Choi operator of a CP map. A -step process is a comb , a multilinear functional on tuples of intervention elements, normalized to 1 on all tuples of deterministic operations and nonnegative on tester elements [16, 20, 21]; equivalently, a positive Choi operator satisfying the causality hierarchy
Outcome probabilities are given by the link product, . The identity intervention (the identity channel, “do nothing”) plays a distinguished role: marginalization is defined by inserting , not by measuring and forgetting. It is the neutral marginalizing operation in the comb, not a suspension of physical evolution: the system continues to evolve between slots, and inserting merely leaves the slot unprobed. This distinction is the operational root of the failure of the classical Kolmogorov consistency condition [16].
The link product admits two readings worth keeping in mind. For a memoryless (Markov) process, factorizes into a product of channel Choi operators, and the link product reduces to the Born rule chained through the channels; memory is exactly the failure of this factorization. The causality hierarchy, in turn, is the statement that later interventions cannot signal to earlier statistics: it is what makes a process rather than an arbitrary multilinear form, and it is the axiom that will reappear as the normalization condition of the transfer map in Sec. V.
We write for finite intervals , for the quasilocal algebra of intervention operators, and for the shift automorphism.
Quantum Markov order
Classical finite memory is counted in steps: a Markov chain of order forgets everything beyond the last symbols. The quantum analogue, introduced by Taranto et al. [17, 18], is formulated through instruments rather than states, because “what the process remembers” depends on how the intervening slots are probed. The precise definition is the following:
Definition III.1 (Quantum Markov order [17, 18]). A process has Markov order relative to an instrument sequence acting on a memory block of consecutive steps if, for every outcome with operation ,
i.e., conditioning on each break outcome factorizes the conditional future from the conditional history.
The physical content is best read through the causal break: an agent who measures the middle block with an informationally complete (IC) POVM and reprepares it from an IC set erases the bridge between past and future. Equation (2) says that after such an erasure, the future depends on the history only through the break outcome. Two caveats matter here. First, Markov order is instrument-relative by construction; Taranto et al. showed that no non-Markovian process has finite order relative to all instruments [17], so “finite memory” is always relative to a probing class, and we fix an IC causal break throughout. Second, the factorization (2) is an identity of operators, not only of probabilities: it constrains the response to arbitrary further interventions on either side.
Two facts from [18] will be used. First, for an informationally complete (IC) history-blocking instrument one has the block expansion
where is a dual set to the instrument Choi operators (not individually positive in general). The index labels a joint measure-and-prepare outcome of the causal break on ; if a basis subfamily is selected below, the instrument operators, the dual operators, and the sum all use that same label set. All contractions use the input/output-leg ordering and Choi transpose of the link product above, while and carry the remaining future and history legs. Second, for a fixed system Hilbert-space dimension , the operator space of the -step memory block is
A causal break [12] (measure with an IC POVM and independently reprepare from an IC set at every step) is IC, so (3) applies to it. We take the IC POVM elements to be full rank [40, 41] (e.g., a depolarized tetrahedral POVM, with elements for , which is full rank and IC), which will provide a uniform lower bound on record probabilities (Lemma V.5).
If the input/output Hilbert spaces vary with time, denotes a uniform upper bound on the system Hilbert-space dimension at every step; (4) is then read with .
Notation and terminology
Table 1 collects the symbols used in the main statements. Throughout, identity insertion names the neutral operation (informally, “do nothing”), idle dynamics names the transfer dynamics generated by , and idle-gap correlation names the corresponding process-level correlator. We use these terms consistently rather than treating measurement followed by forgetting as identity insertion.
| Symbol/term | Meaning |
| Operational comb of the process on the finite time window . | |
| Linear causal transfer map associated with an intervention element . | |
| Idle transfer map for identity insertions; its spectrum controls intrinsic process-memory ergodicity and decay rates. | |
| Instrument-dressed record map for a fixed instrument ; its spectrum controls instrument-specific record rates. | |
| Boundary state/readout functional of the generating triple, normalized by for channels . | |
| Abstract finite-dimensional vector space of intervention functionals in a generating triple; in the construction it is realized by , not by a physical memory Hilbert space. | |
| Concrete span of conditional future functionals. Its operational positive cone is defined by tests on tuples of measure-and-prepare interventions and identity insertions; the cone is a subset of this vector space, not a Hilbert space. | |
| Physical input/output Hilbert spaces at time , distinct from the abstract functional spaces and . | |
| Quantum Markov order relative to the fixed full-rank IC causal break. | |
| System Hilbert-space dimension, or a uniform upper bound on local system dimensions when they vary. | |
| General idle-gap correlation and its collision-model specialization, respectively; both are intrinsic process-memory correlators governed by (or its model realization ). | |
| Record ACF, | Autocorrelation of a classical outcome record for a fixed instrument; an instrument-specific correlator governed by (or its model realization ). |
Table 1. Notation and the two rate levels.
Stationary quantum time series
In this section, we define the central object of the paper and establish its well-definedness on the two-sided infinite time axis. The definition translates the classical notion of a stationary process, replacing joint distributions by combs and marginalization by the identity insertion:
Definition IV.1 (Stationary quantum time series). A family of combs on finite subsets (intervals and finite unions of intervals alike) is stationary if
(consistency) for , , where marginalization inserts the identity intervention at the dropped slots, whether or not they lie at the boundary of an interval;
(translation covariance) for all .
Causal transfer maps
Two remarks bear on this definition. Condition (i) is the only causally available notion of forgetting: any operation that reads the dropped slots, even with its outcome discarded, would in general change the statistics of the remaining slots, so the identity insertion is the unique neutral marginalization. Condition (ii) is the classical shift invariance, now required on every instrument choice at once. Some readers might worry that consistency under identity insertion is too weak to pin down a process, since it never tests the dropped slots. The GET guarantees this is not so: consistency on the operational domain is exactly what is needed for a unique extension, as we now record:
Proposition IV.2 (Unique covariant extension). A stationary comb family extends uniquely to a comb on the quasilocal domain of the GET, and the extension is translation-covariant: .
Proof sketch. Existence and uniqueness of follow from the GET [16]. The shifted functional is well defined because the shift preserves finite-interval liftings and their causal compatibility, and its finite marginals coincide with those of by (ii); uniqueness forces .
This is the quantum form of the Kolmogorov extension theorem: the bi-infinite stationary process exists and is unique as an operational object. Note that uniqueness holds for the comb, not for any underlying system–environment dilation, of which there are many; every representation we build below is a representation of the comb itself.
We calibrate the definition on two anchor examples. A classical stationary process embeds diagonally: its combs are the classical joint distributions, identity insertion is marginalization, and the two consistency notions coincide. At the opposite extreme, the process that applies a fixed unitary to the system at every step and nothing else is stationary and remembers everything; its correlations never decay, and it will reappear as the boundary case of the spectral criteria in Sec. VI.
A fixed instrument repeated at every step produces a classical outcome record. Stationarity of the process implies stationarity of the record; the converse holds if the instrument family is tomographically complete [42].
Causal transfer maps
A classical order- Markov chain is summarized by finitely many numbers: the transition rule on the last symbols. We now construct the quantum analogue: a finite-dimensional operator space that plays the role of the memory, a family of maps on it that play the role of transition matrices, one per intervention, and a state that reads the statistics out. The construction is intrinsic: the memory is not a guessed ancillary system but the space of operationally distinguishable futures.
The generating triple
The representation is packaged in the following definition:
Definition V.1 (Causal transfer map and generating triple). A generating triple for a stationary comb family consists of: a finite-dimensional matrix-ordered space with order unit (the matrix order used throughout this paper is the operational one constructed in Lemma V.5); for every a linear map with
(complete positivity) is CP whenever lies in the operational cone (positive combinations of causal-break elements and identity insertions); whether this extends to coherent intervention elements is part of Conjecture IX.1;
(identity-insertion unitality) ;
(linearity) is linear;
and a state on satisfying the normalization condition for every channel . The triple generates the family if, for all finite windows,
Remark V.2 (Universal unitality of identity insertion). The identity intervention is the unique intervention whose transfer map is unital for every process in the class. A general deterministic intervention does not give : a nontrivial channel changes future statistics. The normalization condition is instead the abstract form of “instrument outcomes sum to one”, and its proof is precisely the causality of the comb (a deterministic operation at the latest slot cannot change earlier statistics).
For a fixed process, holds iff leaves the conditional future statistics unchanged. The universal quantifier in Remark V.2 is essential: degenerate fixed processes can admit other unital interventions.[55]
The generating formula (5) is read from right to left: start with the memory in its distinguished state , apply one update per intervention in time order, and evaluate the resulting functional at the boundary state . Two features matter for what follows. First, there is one map per intervention element, so the transfer structure resolves not only which instrument was applied but also which outcome occurred; summing over outcomes recovers the channel-level update. Second, the family is not a group or even a semigroup of physical channels on a fixed system: it is an operator-valued measure on the space of interventions, and its composition law encodes the process’s memory. When varies over channels, the maps describe how the future is steered by what is done now; this is the object that classical theory reaches only through conditioning, and which here exists without any quantum conditional state.
Construction for finite Markov order
We now show that finite Markov order forces the memory to be finite-dimensional, with an explicit bound:
Proposition V.3 (Finite-dimensional representation). Let be a stationary quantum time series with Markov order relative to the fixed full-rank IC causal break of Sec. III. Then there exists a generating triple with , with the matrix order and complete positivity understood on the operational cone of Definition V.1.
a. Proof roadmap. Stationarity and identity-insertion consistency enter first through Proposition IV.2, supplying the stationary infinite past and hence the order unit and invariant boundary state. Finite Markov order relative to the fixed IC break enters through the block expansion (3), which bounds the span of conditional future functionals by (Lemma V.4). The definition of these functionals on all future windows makes the update map independent of the chosen past and yields the time-ordered generating formula (Lemmas B.1 and B.2). Finally, the factorization (2), full rank of the IC POVM, and restriction of the order tests to the measure-and-prepare/identity operational cone enter the ratio estimate that proves the order-unit, positivity, and state properties (Lemma V.5); no coherent-intervention extension is assumed.
The bound has a transparent meaning: after an -step causal break, everything the past can influence about the future passes through the -step memory block, whose operator space has dimension ; the memory of the process is therefore no larger than the block that mediates it. The construction makes this precise by identifying the memory space with the span of all conditional future functionals: the past matters only through the future statistics it still controls.
The proof occupies the remainder of this subsection together with Appendix B. The main text retains finite dimensionality of the functional space (Lemma V.4) and the operational order structure (Lemma V.5). Well-definedness of the update map requires equality of functionals on every future window, while the boundary-state step requires the stationary infinite past; these hypotheses and conclusions are stated explicitly in Lemmas B.1 and B.2 and summarized below. For a finite past intervention sequence on , define the conditional future functional
a single consistent object on all future windows, and set . The first step of the proof bounds the dimension of this space:
Lemma V.4 (Finite dimensionality). .
Proof. Split a window () into and , and apply the block expansion (3) at the cut :
so with at most terms (for an overcomplete IC instrument, first pass to a basis subfamily of the dual set; biorthogonality forces linear independence, so the count is at most ). More explicitly, the independent subfamily is chosen from the instrument Choi operators spanning , and is then its dual basis. Testing both sides of (3) against these spanning instrument operators shows that this rewriting preserves the decomposition; the symbols thereafter refer to the common independent subfamily. □
This argument has three features worth making explicit. First, the bound concerns functionals on all future intervention tuples, coherent or not: each is defined by the displayed contraction for arbitrary , so the dimension bound is not restricted to any causal-break subspace. Second, the instrument-relativity of the Markov condition enters only through the choice of the cut and its dual set; the block expansion (3) is an identity of operators, so it constrains the response to arbitrary histories and arbitrary future interventions at once. Third, the dual set , which is not individually positive, enters only this counting argument; the positivity arguments below (Lemma V.5) are instead based on the factorization (2), whose factors are, up to positive scalars, partial inner products of positive operators and hence positive.
Two further steps complete the construction; Appendix B gives their full statements and proofs, including the normalization identity. First, prepending one intervention, , defines a linear map that depends only on the functional and not on any past generating it or on any instrument choice; it satisfies , is linear in , and obeys the time-ordered composition law (Lemma B.1). Second, with the boundary state and the order unit , the conditional future functional of the stationary infinite past, which exists by Proposition IV.2, the generating formula (5) holds and (Lemma B.2). It remains to equip the memory space with the order structure that makes the triple physical:
Lemma V.5 (Order structure). Define the matrix order on by testing on the operational test set (positive combinations of intervention tuples whose slots are causal-break elements or identity insertions), with matrix levels entrywise. With respect to this order, is an order unit, is CP for in the operational cone, and is a state. Whether this operational order coincides with the pointwise order on all coherent intervention tuples is open (Remark V.7). The identity-intervention transfer map is moreover a contraction in the evaluation norm
which is a genuine norm on since CPTP Choi operators span .
Proof essentials; full version in Appendix B. Positivity of is tested on the operational test set : positive combinations of intervention sequences whose slots are causal-break elements or identity insertions. The key estimate is the ratio lemma: for a break element on the first slots and arbitrary ,
where is independent of the past by (2), , and because the IC POVM is full rank. The history may be any CP intervention tuple; neither trace preservation nor any other normalization is required. For each fixed history, is finite, and gives for . Once is fixed, this constant is uniform over future tests and does not grow with the past length; appending identity slots leaves the corresponding functional and its domination constant unchanged. The order-unit property needs only this elementwise domination, on the test order, not a bound uniform over all histories. Extension to positive combinations is by linearity, while identity insertions are absorbed by marginalization and identity absorption. The finitely many partial-break patterns also have finite, history-dependent ratios by the same full-rank lower bound, so enlarging the constant if necessary covers the whole of . Finally, because is a channel, prepending it maps every CPTP test tuple to another CPTP test tuple; the supremum defining (8) therefore cannot increase under . This proves contractivity, the only norm property used by the spectral arguments below.
□
Remark V.6 (Operational cone, matrix order, and boundedness). Write for the cone defined by . It is a proper cone: convexity and closedness are immediate from the defining tests, and pointedness follows because tensor products of causal-break and identity Choi operators span every finite-window intervention space. The causal-break elements or identity insertions), with matrix levels entrywise. With respect to this order, is an order unit, is CP for in the operational cone, and is a state. Whether this operational order coincides with the pointwise order on all coherent intervention tuples is open (Remark V.7). The identity-intervention transfer map is moreover a contraction in the evaluation norm
which is a genuine norm on since CPTP Choi operators span .
The ratio lemma gives two-sided domination of every self-adjoint element by a finite multiple of . Thus is generating, the order is Archimedean, and lies in the interior of ; equivalently, the order-unit norm is finite on every self-adjoint element. At matrix level, means entrywise for every . The test set therefore plays the dual role that defines the operational matrix order; it is not identified with the full pointwise order on coherent interventions. Correspondingly, complete positivity of is asserted exactly for elements satisfying within the intervention operational cone. Finite dimensionality makes all norms on equivalent, and evaluation-norm contractivity makes power-bounded, which is the boundedness input used throughout Sec. VI.
Remark V.7 (Coherent interventions). The operational cone is generated by measure-and-prepare interventions, which is also what process tensor tomography implements. For coherent interventions entangled across the memory block, the ratio argument can fail: the memory-block factors are signed, and destructive interference among memory sectors can make the stationary denominator anomalously small. Whether the order-unit domination extends to coherent test tuples (equivalently, whether the operational order coincides with the pointwise order on the full cone) is an open problem; see Conjecture IX.1. The generating formula, the evaluation norm, the order structure on the operational cone, and the spectral conclusions of Sec. VI do not depend on this extension. The failure boundary, if nonempty, is itself a form of temporal nonclassicality, a subject for separate investigation beyond the scope of this work.
Anchor cases
Two degenerate cases calibrate the construction, and we encourage the reader to keep them in mind in what follows. For a classical order- Markov chain with transition matrix , the memory space is the function space on the last symbols, is the shift-by-one-step map, for a measure-and-prepare instrument is the classical “observe, then write” update, is the stationary measure, and the ratio bound (9) is the classical estimate ; every theorem below reduces in this case to a standard fact about Markov chains. For a memoryless quantum process (, IC preparations), is dual to the system state space and is the (pre-)dual action of the channel : the transfer map is the channel itself, iterated. The genuinely new content of the construction lies between these anchors, where the memory is quantum and the instruments are arbitrary.
Spectral theory: ergodicity and mixing
With the transfer map in hand, the classical spectral questions can be posed in earnest, and this section answers them: we determine when the process forgets, and at what rate. The answer comes in two layers, because quantum mechanics supplies two inequivalent readings of “forgetting”.
Two autocorrelations
Classically there is one autocorrelation function. Quantumly there are two levels, governed by two different operators: the record-level ACF of a fixed instrument (governed by ), and the following instrument-independent notion:
Definition VI.1 (Idle-gap correlation). For finite intervention blocks , the idle-gap correlation is
The blocks may be arbitrary intervention elements; the informative case is outcome blocks (trace-nonincreasing CP elements of an instrument). We fix the meaning of the centering by two calibrations. For an outcome block one has , the stationary occurrence probability of the block, so the subtracted term is the product of the stationary one-block marginals: is the covariance of two occurrence events separated by an idle gap, and no separate centering convention is needed. For deterministic (trace-preserving) blocks and the correlation vanishes identically, as there is no randomness to correlate, so the definition is used with outcome or general sub-normalized blocks; arbitrary elements are covered by linearity.
blocks and the correlation vanishes identically, as there is no randomness to correlate, so the definition is used with outcome or general sub-normalized blocks; arbitrary elements are covered by linearity. Definition VI.2 below quantifies over all such blocks. The record-level ACF is the same construction with replaced by the instrument-dressed map , and its events are classical record events throughout.
The idle-gap correlation asks what the process remembers when nothing is done to it: two interventions are separated by slots of identity insertions, so the decay of measures the intrinsic forgetting rate of the dynamics. The record-level ACF instead measures what survives when every slot is actively probed by the same instrument. The difference between and is precisely the difference between “not measuring” and “measuring and forgetting”; their spectral difference is a quantitative expression of measurement backaction. Classically the two readings coincide, because observation is free; the splitting is the first place where the quantum spectral theory departs from the classical one. A related separation at the level of Markovianity, rather than mixing rates, was established by Taranto and Milz [34]: records can be Markovian while the process fundamentally requires memory. With the two correlation levels in hand, the classical notions of ergodicity and mixing translate directly:
Definition VI.2 (Ergodicity and exponential mixing). A stationary quantum time series is ergodic if for all intervention blocks, and exponentially mixing if for some .
As in the classical theory, ergodicity is the minimal demand (time-averaged correlations wash out), while mixing is the stronger demand that they decay at an exponential rate. The two are genuinely different quantumly as well as classically, and the difference will be visible in the spectrum.
The ergodicity criterion
Finite dimensionality plus contractivity (Lemma V.5) make a power-bounded operator on a finite-dimensional space, so all peripheral Jordan blocks are trivial (peripheral semisimplicity) and the mean ergodic projection exists [43, 44]. Throughout this section, a simple eigenvalue means an algebraically simple one; under power-boundedness this is equivalent to a one-dimensional fixed-point space. The correlation quantifier is over arbitrary finite intervention elements. By linearity this is exactly the probe class used in the two-sided spanning argument below; outcome blocks supply the event-covariance interpretation, but do not restrict the spectral quantifier. The first spectral criterion characterizes forgetting on average:
Theorem VI.3 (Ergodicity criterion). For a stationary process with a finite-dimensional generating triple as in Proposition V.3, the following are equivalent: (i) the eigenvalue 1 of is algebraically simple (equivalently here, the fixed-point space is ); (ii) the process is ergodic; (iii) .
a. Proof roadmap. Finite dimensionality and evaluation-norm contractivity of give power-boundedness, exclude nontrivial peripheral Jordan blocks, and make the mean ergodic projection well defined. Stationarity enters as , which identifies the invariant readout in the rank-one projection. The quantification over all intervention blocks in Definition VI.2 enters in both directions: right spanning converts a second fixed point into a nonvanishing Cesàro correlation, while the rank-one form of makes every such averaged correlation vanish (Appendix C).
Operationally, the criterion says that a process forgets on average exactly when its idle dynamics carries no conserved content: any nontrivial fixed point of is a memory direction that never decays, and ergodicity is the absence of all such directions. The projection plays the role of the stationary expectation, mapping every memory element to its stationary shadow.
A trivial peripheral spectrum implies exponential mixing
Ergodicity concerns forgetting on average: only the time-averaged correlations are required to wash out. The operationally decisive notion is stronger, that the correlations themselves decay at a computable rate. The next theorem says that the intrinsic process-memory rate is read off the idle spectrum: once the eigenvalue 1 is simple and no other eigenvalue lies on the unit circle, every idle-gap correlation decays exponentially, and the subleading eigenvalue modulus sets the intrinsic process-memory time. The precise statement is the following:
Theorem VI.4 (Gap implies mixing). For a process with a finite-dimensional generating triple as in Proposition V.3, if the eigenvalue 1 of is simple and no other eigenvalue lies on the unit circle (we call the peripheral spectrum trivial in this case), the process is exponentially mixing: for every strictly larger than the subleading modulus , one has (with itself when the subleading eigenvalues are semisimple; Jordan blocks contribute polynomial factors, absorbed by the strict inequality).
a. Proof roadmap. Simplicity of the eigenvalue 1 and triviality of the remaining peripheral spectrum give the decomposition , with supplied by Theorem VI.3. Finite dimensionality converts the spectral radius of into the stated exponential operator bound, including the Jordan-block qualification. Substitution into Definition VI.1 then proves decay for every pair of intervention blocks, with the block-dependent constant coming only from the readout and excitation factors (Appendix C).
Thus a gapped idle spectrum gives an intrinsic process-memory rate: every idle-gap correlation is bounded using , independently of the measurement scheme. The corresponding intrinsic process-memory time is a property of , not of an instrument-dressed record map.
Exponential mixing implies a trivial peripheral spectrum
The converse direction is what makes the spectral criterion a criterion rather than a sufficient condition: if every idle-gap correlation decays, then no peripheral mode can hide from all probes. Passing from correlation decay to the operator spectrum requires a two-sided spanning property of the construction, which we make precise after the statement:
Theorem VI.5 (Mixing implies gap). For a process with a finite-dimensional generating triple as in Proposition V.3, if for all intervention blocks, then the peripheral spectrum of is trivial (in the sense of Theorem VI.4: the eigenvalue 1 is simple and no other eigenvalue lies on the unit circle).
a. Proof roadmap. The hypothesis is decay for all intervention blocks, not merely one selected pair. The right-spanning property makes these blocks excite every memory direction, and the left-spanning property makes their readouts detect every functional direction; together they upgrade correlation decay to strong convergence . Power-boundedness then rules out nontrivial peripheral Jordan blocks, and any remaining nonunit peripheral eigenvector would produce a non-decaying oscillation detected by a suitable pair of blocks (Appendix C).
Together, Theorems VI.4 and VI.5 close the loop at the process level: triviality of the peripheral spectrum of the idle transfer map is equivalent to exponential forgetting, so this spectrum determines the intrinsic long-time process memory and its intrinsic decay rate. Combined with Theorem VI.3, which detects the weaker property of forgetting on average through the eigenvalue 1 alone, this gives a complete spectral classification of the idle dynamics. One qualification matters in practice: a given intervention pair excites and reads only the modes it couples to, so an observed decay rate can be faster than the slowest mode; Sec. VIII exhibits an instance where this distinction is essential.
b. Two-sided spanning: from all correlations to the full operator. Theorem VI.5 rests on a two-sided spanning property. This is the explicit bridge from decay of every observable correlation to a statement about the full operator spectrum, rather than a merely technical detail of the proof. The ambient space is the finite-dimensional (dim , Lemma V.4), equipped with the evaluation norm (8); finite-dimensionality makes the norm choice immaterial. The construction satisfies two-sided spanning over all finite intervention tuples: , because by Lemma B.1 and these functionals span by definition; and , because these are the point evaluations at intervention tuples and CPTP tuples alone span (Lemma V.5). Definition VI.2 quantifies over all intervention blocks, so correlation decay determines the strong operator limit of : the spectral criteria are intrinsic to the process, not relative to any restricted cone of probes. These three statements parallel Propositions 3.1 and 3.3 and Lemma 5.2 of [19]. As in [19 Example 4(2)], ergodicity does not imply mixing: a peripheral eigenvalue produces nondecaying period-2 oscillations (see Sec. VIII for a physical instance).
The ergodic theorem for measurement records
We now turn from what the process remembers to what an observer can learn. The practical question behind ergodic theory is whether a single experimental trajectory suffices for inference; the following theorem gives a sufficient condition with rates:
Theorem VI.6 (Exponential mixing of records; LLN and CLT). Fix a finite-outcome instrument whose outcome elements lie in the operational cone (a measure-and-prepare instrument, for instance), and let . Let be the stationary record on this finite outcome alphabet, let for a bounded real observable , and set , , , and . Suppose has a simple eigenvalue (with right fixed point , ; note that in general, since is not unital for a non-identity instrument) and no other peripheral eigenvalue. Then the following conclusions hold.
(i) Record mixing. The record is exponentially -mixing: for every strictly larger than the subleading modulus of the dressed operator there is with
with itself when the subleading eigenvalues are semisimple.
(ii) Law of large numbers. almost surely.
(iii) Central limit theorem. The series defining converges absolutely and [45, 46, 47]. If , this is the usual nondegenerate CLT. If , the conclusion is degenerate: in probability, so the centered fluctuations vanish at the scale.
a. Proof roadmap. Process stationarity and comm normalization make the finite-alphabet record stationary and consistent. Positivity on the operational cone gives for positive , and base-norm contractivity of the outcome maps, together with , makes the event bound uniform over events and interval lengths. The assumed peripheral gap of then gives and , hence and . Stationarity, bounded moments, and these summability bounds are precisely the inputs to the cited -mixing LLN and CLT. When , the conclusion is in probability (Appendix C).
Thus, whenever the instrument-dressed map is gapped, a single experimental run suffices to estimate bounded observables of that instrument’s record. The subleading eigenvalue of controls an instrument-specific record rate and the associated confidence intervals; it is not the intrinsic process-memory rate governed by .
Proof sketch. The record is a stationary sequence over the finite outcome alphabet of (record stationarity follows from process stationarity, Sec. IV), and events are cylinder events. For record events on and on , , where is the indicator-weighted composite of outcome maps. The bound holds in the base norm of the operational order, and the normalization identity makes it uniform over all events and all interval lengths (Appendix C). The spectral decomposition with then yields the mixing bound with constants uniform in the events. Since is bounded on the finite outcome alphabet, all moment conditions are automatic, and the same estimate applied to the record autocovariances gives , so the long-run variance converges absolutely; the classical limit theorems for stationary -mixing sequences [45, 46, 47] complete the proof (if the CLT is degenerate: the fluctuations vanish at the scale).
As a direct consequence, single-trajectory estimation is available for the stated record-observable class:
Corollary VI.7 (Single-trajectory record inference). On the class of stationary processes with finite Markov order relative to the fixed IC causal break, for any finite-outcome operational-cone instrument with a gapped record operator and any bounded real observable of its record, the sample mean is a strongly consistent estimator of from a single measurement trajectory, with the CLT of Theorem VI.6 providing its asymptotic fluctuations. The record mixing rate is instrument-specific and controlled by the spectrum of .
Remark VI.8 (Relation to earlier limit theorems). Theorem VI.6 complements the Sanov/CLT results for quantum Markov chain outputs [14] and the limit theorems for quantum trajectories [9, 10, 31], which correspond to memoryless environments; here the memory is an arbitrary finite-order process tensor. Conversely, the trajectory literature covers conclusion types we do not address (large deviations, Berry–Esseen rates [31]) and does not require finite Markov order; the two settings cover different regimes.
Physical scenarios
The objects above are abstract, but each has a direct physical reading. We collect three scenarios; they differ in status: the first identifies physical realizations of the model class, the second is a direct application of Theorem VI.6, and the third describes an illustrative phenomenon whose full development is deferred to the companion work [49].
a. Collision models as the canonical platform. A collision model, in which the system interacts sequentially with fresh ancillas, is both a standard model of open-system dynamics [48] and a standard physical realization of our finite-Markov-order class: the ancillas carry the memory, in the instance of Sec. VIII the Markov order is the number of consecutive collisions each ancilla participates in, and the idle transfer map is the memory channel of the collision chain. Cavity QED with crossing atoms and superconducting qubits coupled to resettable ancilla resonators implement such models directly. Section VIII uses a two-collision model; the ancilla stream is the environment, and its partial retention is the non-Markovianity.
b. Continuous monitoring and single-run inference. Continuously monitoring an open quantum system (homodyne or photodetection records of a driven qubit, for example) produces a measurement record of the kind Theorem VI.6 governs. The theorem’s content for this setting is practical: when the instrument-dressed operator has a gap, a single experimental run suffices to estimate bounded observables of that record, with the long-run variance and instrument-specific inference rate read off the record’s autocorrelation and the spectrum of . The idle-gap correlation of Definition VI.1 is the process’s memory kernel measured without touching it: it answers the question of what the dynamics remembers when left alone, which is exactly the quantity that environment characterization tries to access.
c. Memory-assisted control and dissipation dodging. The two-level asymmetry has a control reading. An instrument that counter-rotates against the process’s internal rotation can keep a memory direction alive indefinitely (Sec. IX), so a dissipative process can nonetheless carry permanent, record-accessible coherence. This is a dynamical-decoupling-type phenomenon seen from the process side: the process’s own dissipation sets the baseline, and the instrument’s coherence against it determines what survives. The companion work uses this phenomenon as motivation for a further study of memory that records can never see [49]; that reference is illustrative and is not needed for any theorem proved here.
Numerical illustration: a collision model with Markov order 2
We now illustrate the spectral criteria of Sec. VI, in particular the mixing criterion of Theorem VI.4 and the record ergodic theorem, Theorem VI.6, on an exactly solvable model. We consider a qubit collision model in which each ancilla, fresh in , collides with the system at two consecutive steps before being discarded, so that the Markov order is 2 and the memory is carried by a single ancilla qubit. We test two collision gates, and the extended gate . The underlying formulas are exact finite-dimensional expressions, evaluated with dense double-precision linear algebra and with no Monte Carlo element; trace preservation, probability normalization, validity of the stationary memory state, and the i.i.d. limit were checked to machine precision (Appendix § 1.6.5). These exact formula evaluations and numerical diagonalizations are consistency checks of the implementation of the stated formulas, not independent validations of the general mathematical theorems.
a. Record level (Theorem VI.6). With the instrument “measure , reprepare ”, the record operator on the ancilla memory was computed as a superoperator. For the base gate, is provably rank one for any measure-and-reprepare instrument (; the record is at most one-step correlated): the ancilla’s written components and read-out direction are misaligned for single-Pauli gates. For the extended gate, the record ACF decays geometrically over 4–10 orders of magnitude, and the fitted instrument-specific record rate agrees with to five to six decimal places:
| fitted ACF rate | ||
| 0.2 | 0.005985 | 0.005985 |
| 0.4 | 0.078037 | 0.078038 |
| 0.6 | 0.276959 | 0.276959 |
| 0.8 | 0.514161 | 0.514161 |
Table 1 (PDF p. 12).
b. Idle level (Theorem VI.4). The idle step map on the joint system–ancilla memory has the closed-form nonzero spectrum (numerically matched to ), and the idle-gap correlation of a -preparation/-measurement pair is exactly . The twofold degenerate eigenvalue 1 (a conserved functional) means the idle process is not mixing; the observed intrinsic process-memory decay rate for this probe pair is that of the fastest nontrivial mode coupled to the chosen preparation/measurement pair, an instance in which the hypotheses of Theorem VI.4 fail and the conclusion must be read mode by mode. At the peripheral eigenvalue appears (period-2 oscillation), a physical analogue of [19 Example 4(2)].
c. Cross-level comparison. At all tested points, ; equivalently, the tested measure-and-prepare instruments increase contractivity at those points. This is an observation at all tested points, not a general theorem. Idle mixing implied record mixing at every tested point (including the nontrivial extended gate), while the converse is false for the base gate: the record is i.i.d. while the idle process fails to mix, because the instrument destroys the conserved quantity. The point spectra are not nested; any general relation must concern the peripheral spectrum or mixing, not the full spectrum.
d. Reproducibility. All quantities above are finite-dimensional and assembled in closed form. With the gate, the reprepare state, and the instrument fixed as stated, the record operator (, acting on the one-ancilla memory) and the idle step map (, acting on the joint system–ancilla memory) are given by the explicit formulas of Appendix E; the stationary memory state is the fixed point of , with Bloch vector ; and record probabilities and autocorrelations follow from the chain rule stated there. There is no Monte Carlo element: the only numerical steps are dense matrix algebra and diagonalization in double precision, with the sanity checks of Appendix E passed to . Fitted rates are extracted from the log-linear decay of over the window between the onset of asymptotic decay and the floating-point floor near (visible in Figure 1); the agreement quoted in the table is to five to six decimal places. At and near the period-2 point , where the peripheral eigenvalue makes any finite-window fit meaningless, the spectra are read from the closed forms of Appendix E and the numerical diagonalization is a cross-check.
Figure 1. Record-level consistency check of the spectral formula in Theorem VI.6 for the extended gate at (see text for details): exact record autocorrelation (markers) versus the spectral prediction (dashed curves), where is the record operator on the one-ancilla memory and its subleading eigenvalue modulus. Bottom right: the base gate , for which the record is memoryless (, rank-one record operator). The plateau at is the floating-point noise floor. All quantities are evaluated from finite-dimensional formulas using double-precision linear algebra.
Figure 2. Intrinsic process-memory consistency check of the spectral formula in Theorem VI.4 for the base gate: the idle-gap correlation of a -preparation/-measurement pair (markers) versus the closed form (curves), for the couplings shown. The sign-alternating cases are visible; is close to the period-2 point , where the eigenvalue enters the peripheral spectrum of the idle step map . All quantities are evaluated from finite-dimensional formulas using double-precision linear algebra.
Taken together, these computations check the implementation of the spectral formulas associated with Theorems VI.4 and VI.6 at both correlation levels, with agreement to five to six decimal places at the tested points; they do not independently verify the general theorems. The model exhibits the two-level asymmetry and illustrates that an observed decay rate reflects the modes coupled to the chosen probes: an intrinsic process-memory rate is read from (the realization of ), whereas an instrument-specific record rate is read from (the realization of ).
Discussion and open problems
We have built the stationary theory of quantum time series on the process-tensor foundation and proved its first structural theorems: the definition (Definition IV.1) with its existence theorem (Proposition IV.2), the finite-dimensional causal transfer map (Proposition V.3), the two-level autocorrelation structure (Definition VI.1), and the spectral criteria with the record ergodic theorem (Theorems VI.3–VI.6). However, our results are established under two restrictions: finite Markov order relative to an IC causal break, and measure-and-prepare operational cones. Whether the spectral criteria persist when either restriction is relaxed, for infinite-order memory on the one hand and for coherent interventions on the other, is open. Beyond these, several further questions are already visible, and we here list five candidates:
Coherent interventions and temporal interference. The order-unit property was established on the operational cone (measure-and-prepare plus identity). Whether it extends to the full cone of coherent in-

Figure 3. (iii) Spectra of record operator and idle operator (unit circle dotted). The spectra behind both levels: eigenvalues of the idle step map and of the record operator in the complex plane, for the two gates and the tested couplings (see text for details). The degenerate eigenvalue 1 of (the conserved functional) and the peripheral approach at are visible; the record operator’s spectrum is strictly inside the unit disk for the extended gate and degenerates to for the base gate.
terventions is open:
Conjecture IX.1. There is a dichotomy between processes for which the operational order extends to coherent interventions and processes for which it fails; the failure class is nonempty and exhibits an interference phenomenon among memory sectors with no classical counterpart.
We do not currently have a decidable criterion for the failure class; finding one (for instance, a spectral or entropic quantity of the transfer map that detects it) is itself part of the problem. This boundary has no classical counterpart: it marks where temporal interference (interference among memory sectors) becomes visible in asymptotic statistics. We leave its systematic study to a separate investigation of temporal nonclassicality.
(ii) Cross-instrument spectra: the two levels are independent.
dependent. Record mixing does not imply process mixing (Sec. VIII: instruments can destroy conserved quantities of the idle dynamics). Process mixing does not imply record mixing either. Consider the single-qubit Markov channel , where (): the Bloch matrix of has eigenvalues , so is mixing, but composing with the unitary instrument gives two fixed states and . With a weak readout of strength , the record autocorrelation is exactly constant, for all , while the idle process mixes at the intrinsic process-memory rate (a direct Bloch-sphere computation). The instrument rotates the memory content away from the contracting direction at every step, dodging the dissipation that drives idle mixing. The forward implication fails even for causal breaks: in the companion work [49] we construct a classical permu[49] tation gate in which the reprepared state acts as a key that keeps a classical memory channel open (record non-mixing) while the idle evolution forgets in two steps (process mixing); a self-contained version of the construction is given in Appendix D. A no-broadcasting argument [50] shows the constraint in the quantum direction: a perfectly preserved quantum memory mode forces the system’s residual state to be memory-independent, making the mode invisible to the record; only classical (copyable) peripheral modes can be record-visible, and the key-locked counterexample is exactly such a classical visible mode. The counterexamples exhibited here are fine-tuned structures. As illustrative motivation only, the companion work reports large random sweeps that found no additional counterexamples [49]; no result in the present paper depends on those sweeps. Classifying, for a given process, the instruments under which records mix is an open problem; already at the channel level, mixing channels are not closed under right composition.
(iii) Peripheral structure. The full analogue of [19 Prop. 3.3] (cyclic peripheral spectrum and periodic decomposition) requires Perron–Frobenius theory on operator systems [51, 52] and is left to future work.
(iv) Uniqueness. Whether two-sided minimal generating triples are unique up to linear isomorphism remains open, as does its finitely-correlated-state counterpart for general -FCS [19].
(v) Outlook. The representation and spectral theory built here are the scaffolding for a quantum time series analysis: spectral (Wiener–Khinchin-type) theory, ARMA-type parametrized model families, estimation and order selection, and nonclassicality measures of temporal statistics are the natural next steps, and the criteria proved here are what those developments will stand on.
Acknowledgments
This work is supported by the National Natural Science Foundation of China [grant number 12401158] and the Fundamental Research Funds for the Central Universities [grant number 30925010422].
Details for Sec. IV: the covariant extension
This appendix expands the proof of Proposition IV.2 (unique covariant extension of a stationary comb family to the two-sided infinite time axis), whose proof sketch is given in Sec. IV. We use the finite-dimensional, lift-based GET of Milz et al. [16]. The hypotheses used here are: (i) every local intervention space is finite-dimensional; (ii) for each finite , is the causally compatible intervention domain and is its lifted domain, with identity insertions outside the finite support; and (iii) whenever , the combs agree under restriction to the lifted compatible domain, equivalently under identity insertion at the slots of . Definition IV.1(i) supplies exactly this consistency, for intervals and finite unions of intervals and for dropped slots in the interior as well as at the boundary; the finite-dimensional and causal-compatibility requirements are already part of the comb setting. The GET therefore associates to a unique comb on the closure of the lifted finitely-supported interventions. For with its shift , define . The proof is divided into three steps.
Step 1: preserves the domain. An element of has the form with . A supported window is mapped to , and a finite union of intervals is mapped in the same way; thus the shifted tuple is . Since the whole family is defined with respect to one fixed temporal causal order, and is defined by compatibility with that order, . Hence finite-interval lifting and causal compatibility are both preserved, , and by continuity .
Step 2: finite marginals agree. For any finite interval , the shifted lifted tuple above has finite marginal by Definition IV.1(ii). The same identity holds on finite unions of intervals by lifting and identity-insertion consistency.
Step 3: uniqueness. The GET states that a comb on is uniquely determined by its finite marginals; applied to and this gives equality.
Remark A.1. The GET’s uniqueness statement concerns the functional on ; the underlying process on the larger container space need not be unique, but the ambiguity is experimentally undetectable [16]. All representations in this paper are representations of .
Details for Sec. V: the causal transfer map
This appendix completes the construction behind Proposition V.3: it verifies the consistency of the conditional future functionals, proves Lemma V.4 in detail, states and proves Lemmas B.1 and B.2 (which the main text summarizes in prose), gives the full proof of Lemma V.5, and works out the anchor cases of the generating triple.
Consistency of the conditional future functionals
This subsection verifies the consistency property invoked in Sec. V when each is treated as a single object on all window lengths. That property follows from consistency (Definition IV.1(i)): for , , since inserting identity interventions marginalizes. Linearity of in each argument follows from multilinearity of .
Proof of Lemma V.4 (finite dimensionality)
Fix the IC causal break and its dual set , . For an overcomplete IC instrument, choose a subfamily of its Choi operators that forms a basis of and use the corresponding dual basis; biorthogonality forces linear independence, and testing against the spanning instrument subfamily shows that the block expansion is preserved. The common label set is thereby reduced to at most elements. For , split , , and apply (3) at the tripartition . Contracting the past slots with and the future slots with gives the displayed expansion, which exhibits as an element of , a space of dimension at most the number of dual-set elements, [[18 Sec. IV]]. Windows shorter than are covered by marginalization.
The update map and the boundary state
This subsection states and proves the two lemmas that the main text summarizes in prose: well-definedness of the transfer map (Lemma B.1) and the boundary state with the generating formula (Lemma B.2).
Lemma B.1 (Transfer map). For , defines a linear map which depends only on (not on any past generating it, nor on any instrument choice); ; is linear; and the composition law respects time order.
Proof. The definition uses only the values of as a functional: if as functionals (equal on all tuples of all lengths), their images are equal. This is the point at which the construction is intrinsic: no choice of past, instrument, or recovery map enters. The update rule follows by evaluating . Linearity in follows from multilinearity of the comb, and the composition law follows by applying the same evaluation identity twice, with the latest intervention placed to the right in the past block.
Lemma B.2 (Boundary state and generating formula). With and (the conditional future functional of the stationary infinite past, which exists by Proposition IV.2), the generating formula (5) holds and .
Proof. Iterating the update rule of Lemma B.1 from the stationary infinite past gives , the conditional future functional of the stationary past followed by the window . Evaluating at the empty future,
where the last equality is the definition of the stationary functional: the statistics of a window preceded by the stationary past are the stationary window statistics. This is (5). The identity is the special case of the normalization identity proved in the identity-insertion unitality paragraph below.
Proof of Lemma V.5 (order structure)
This subsection gives the full proof of Lemma V.5; the main text presents only the key estimate, the ratio lemma of Eq. (9).
a. The operational cone. Let consist of positive combinations of intervention tuples whose every slot is either a causal-break element (with a full-rank IC POVM and an IC set of states) or the identity intervention . Define iff for all ; matrix levels are defined entrywise. Since the Choi operator of the process is positive, every with is positive, and .
b. Positivity for general positive : what holds and what does not. For a functional generated by a single positive past () one has for every tuple of positive Choi elements, because both factors are positive operators; so positive reachable functionals stay positive under for arbitrary positive . However, a general element of the operational cone is a linear combination of such functionals, and the conclusion does not propagate to combinations: comb positivity alone does not control for . The CP statements of this paper are therefore asserted only for in the operational cone; the general extension is Conjecture IX.1. What the ratio lemma adds is not positivity but domination by the order unit.
c. Support alignment. If satisfies , then is supported outside ; a positive operator is supported within the tensor product of the supports of its marginals (its range is contained in theirs), so the joint comb (whose -marginal is the stationary one) gives for every past.
d. The evaluation norm. CPTP Choi operators linearly span (unitary channels and depolarizing combinations suffice). Hence a functional vanishing on all CPTP tuples of all lengths vanishes identically, so (8) is a norm. by comb normalization. Since is a channel, prepending it maps CPTP tuples to CPTP tuples, so is a contraction in this norm.
e. The ratio lemma (order unit property). For a break element on the first slots, the factorization (2) gives
with and (partial inner products of positive operators). Full rank of the POVM gives the state-independent lower bound : at each of the steps, the measurement outcome probability is bounded below by regardless of the state being measured.
The history is allowed to be an arbitrary CP intervention tuple, with no trace-preserving, subnormalization, or probability normalization assumption. For each fixed , every numerator is finite and every stationary denominator is strictly positive. Therefore
This is an elementwise constant: gives for . For fixed it is uniform over all future break tests and does not depend on the past length; appending identity slots to the past leaves the functional and the ratio unchanged by identity-insertion consistency. No tester-probability interpretation of an unnormalized history, and no single bound uniform over all , is used. The lower side of the two-sided domination is simply positivity of .
Extension to positive combinations is termwise by linearity. Identity insertions in a test tuple are absorbed by marginalization and identity absorption. A tuple containing break elements is bounded directly: by positivity, while . Each partial-break pattern therefore has a finite ratio, with the same scaling in . Taking to be the maximum of and these finitely many partial-break ratios covers the whole operational test set. Hence for ,
which is precisely the elementwise two-sided domination making an order unit for the operational cone.
f. Identity-insertion unitality and normalization. because identity insertion marginalizes and the stationary marginal of the stationary process is itself. Positivity of follows from with the empty-window value equal to the value at any single identity insertion; . For any channel ,
where the third equality is comb causality: a deterministic operation at the latest slot does not change earlier statistics (the Choi-level identity combined with the causality hierarchy).
Anchor cases
This subsection works out the two anchor cases that calibrate the generating triple of Proposition V.3 in the main text.
a. Classical diagonal processes. If all intervention operators are diagonal in a fixed basis, the process is a classical stationary process; finite Markov order relative to the break is a classical order- Markov chain. Then is the function space on the last symbols (), is the shift, the stationary measure, and the ratio lemma is the classical bound .
b. Markov processes. For with IC preparations, is (dual to) the system state space, is the predual action of , , and normalization is trace preservation.
Details for Sec. VI: proofs of the spectral criteria
This appendix proves the spectral criteria and the record ergodic theorem of Sec. VI: Theorems VI.3, VI.4, VI.5, and VI.6, together with the base-norm criterion (Lemma C.1) and the classical limit theorems (Lemma C.2) that the record theorem uses.
a. Power-boundedness and the mean ergodic projection. is a contraction in the evaluation norm (Lemma V.5), hence power-bounded; on a finite-dimensional space, power-boundedness implies that all Jordan blocks at peripheral eigenvalues are trivial and that the Cesàro averages converge to the projection onto the fixed-point space along .
b. Proof of Theorem VI.3 (ergodicity criterion). (i)(iii): if the fixed-point space is one-dimensional, for an invariant functional ; invariance and the one-dimensionality of the invariant functionals (which coincides with the one-dimensionality of the fixed points, since power-boundedness makes the algebraic and geometric multiplicities of the eigenvalue coincide) give . (iii)(ii): the Cesàro average of (10) inserts between and , and the rank-one form makes it vanish. (ii)(i): a second fixed point would, by right spanning, be represented exactly as a linear combination of over finitely many blocks (the span is all of , not merely dense), producing a nondecaying Cesàro correlation.
c. Proof of Theorem VI.4 (gap implies mixing). With trivial peripheral spectrum, , , , and for every one has (Jordan blocks at the subleading eigenvalues contribute polynomial factors, absorbed by the strict inequality); substituting into (10) gives the bound with .
d. Proof of Theorem VI.5 (mixing implies gap). By two-sided spanning, pointwise decay of (10) for all blocks implies strongly (finite-dimensional). Any peripheral eigenvalue with eigenvector would contribute a nondecaying oscillation to some correlation (choose with overlapping and with nonvanishing on ), contradiction.
e. Proof of Theorem VI.6 (exponential mixing of records). Event algebra and stationarity. The record is a sequence over the finite outcome alphabet of ; events are cylinder events (finite unions of outcome strings on finite intervals). The record family is stationary: shift-invariance of is the translation covariance of the comb family, and tail consistency follows from the normalization condition (summing the latest outcome applies a channel at the latest slot, which changes no earlier statistics); one-sided Kolmogorov consistency requires no more. It suffices to control for on and on with constants independent of , .
The uniform event bound. For an outcome string write , and for an event set . We work in the base norm of the operational order on the self-adjoint part of ,
a genuine norm because the operational cone is proper and generating [53], and because is strictly positive on it. For the latter, work on the cone generated by reachable positive functionals. If a generator satisfies , then on every joint past/future window the positive operator has trace and hence vanishes. Consequently for every future intervention, so the spanning property of the operational tests gives . A general cone element is a positive combination of generators, and with all terms nonnegative forces each summand to vanish. Positivity of the outcome maps gives pointwise on the test set for positive . The normalization identity (the instrument sums to a channel) is what makes the bound uniform: for positive ,
so for every event of every length. The readout side obeys the same bound, , and . (The order-unit norm would not give this uniformity directly, since is not unital; the base norm, with , is the right tool.)
Mixing estimate. With , the peripheral hypothesis gives (finite-dimensionality; possible Jordan-block polynomial factors are absorbed into at the cost of an arbitrarily small increase of ). Then
with independent of the events, their interval lengths, and their positions, where is the stationary one-sided marginal: indeed the marginal of at position is , which is independent of by stationarity of the record and converges to by the spectral hypothesis, so . This is exponential -mixing, .
Limit theorems. The probability model is a stationary process on a finite outcome alphabet, and is bounded, so every required moment is finite. Exponential mixing gives . Applied to the centered record observable, the covariance estimate gives , hence and the long-run variance converges absolutely. These are, respectively, the stationarity, moment, mixing-summability, and covariance-summability inputs to the classical LLN and CLT for stationary -mixing sequences, see Lemma C.2 below. If the CLT is degenerate in the precise sense that in probability.
The base-norm criterion
The base norm used in the proof of Theorem VI.6 above is controlled by the following criterion, which is the content that makes the event bound uniform.
Lemma C.1 (Base-norm criterion). Let be a proper generating cone in a finite-dimensional space, , and define on the base seminorm . Then for , and is a norm if and only if is strictly positive on . In the setting of this paper, strict positivity holds on the cone generated by reachable positive functionals: if and , the zero-marginal argument above gives on every future window, hence vanishes on every future intervention; the conclusion extends to positive combinations since is additive and nonnegative on each generator.
Proof. For , the decomposition gives , while any decomposition satisfies , so ; equality follows. The seminorm axioms (homogeneity and the triangle inequality) hold by taking infima over decompositions. If for some , then and the seminorm is not definite. Conversely, if on , then with compact (any Euclidean norm ), by continuity, so and every admissible decomposition gives ; hence and the seminorm is definite. The final claim is the zero-marginal argument recalled in the statement. (That proper and generating alone do not suffice can be seen on , , .)
The classical limit theorems used in Theorem VI.6
For completeness we record the classical input in the form in which it is used, with proof.
Lemma C.2 (LLN and CLT for the record). Let be a stationary process on a finite alphabet whose cylinder events satisfy for all and on those . Then for the strong mixing coefficients. If is a bounded real observable, , , then almost surely, the series converges absolutely, and , with a degenerate limit if ; the standardized form requires .
Proof. Write , , and let bound .
Step 1: extension to the tail -algebras. Fix and write , , . For a past cylinder event , the class contains the future cylinders and is a monotone class by continuity of probability, hence equals by the monotone class theorem; the symmetric argument in gives the estimate on , which is .
Step 2: a covariance bound. For measurable on the past and future respectively with , , one has (write , estimate via the layer-cake representation and the mixing definition). Hence for , where ; the covariance series converges absolutely, with .
Step 3: variance asymptotics and the LLN. Stationarity gives , so (the difference is bounded by ). Chebyshev along the subsequence gives a summable deviation bound, hence a.s.; the sandwich extends this to all .
Step 4: the CLT. Use the big-block/small-block decomposition with , : the big blocks are identically distributed by stationarity, their number satisfies , the small blocks contribute in by Step 2, and the block dependence is controlled by (comparing the characteristic function of the block sum with the product of the marginal characteristic functions via repeated use of Step 2). The marginal triangular array satisfies the Lindeberg condition trivially (the normalized blocks are bounded by ), so its characteristic function converges to ; pointwise convergence of characteristic functions to a function continuous at gives the weak convergence. If , Step 3 gives in probability directly.
Details for Sec. IX: a classical key-locked counterexample (self-contained)
This appendix supplies a self-contained version of the classical permutation-gate counterexample cited in Sec. IX, in which the process mixes while the record does not. The construction is short, and we include it for self-containedness. Let be the permutation gate on two qubits , , , (a 3-cycle plus a fixed point), inserted in the two-collision step of the collision model of Sec. VIII. On the idle side, one checks on the computational basis that the joint step map obeys for every operator , so : the idle process mixes in two steps. On the driven side, with the break instrument “measure , reprepare ”, the classical ancilla bit is preserved forever: writes the memory bit into the system and writes it into the fresh ancilla, so the driven memory channel has spectrum , and the record is a constant sequence determined by the initial memory bit. In the symmetric mixture of the two regimes the record autocorrelation is identically : the process mixes while the record is non-ergodic. The preceding computation is self-contained and is the only part of this counterexample used in R1; the companion reference is solely for optional further details and perturbative robustness [49].
Details for Sec. VIII: the collision-model study
This appendix supplies the closed-form formulas, sanity checks, and rate-versus-spectrum comparisons behind the collision-model study of Sec. VIII.
Exact formulas
The step structure is: instrument on ; collision on , discard ; collision on , . With reprepare state , define
and the outcome-selective maps , satisfying exactly. The record probabilities obey the chain rule , and the ACF is
with the fixed point of . The idle step map on the joint memory is
For the base gate, factors through a four-dimensional image, and its nonzero spectrum equals that of the Pauli-diagonal channel , namely ; the second eigenvalue 1 reflects the conserved functional . The -preparation/-measurement idle-gap correlation has the closed form . For the extended gate, the nonzero spectrum of is (numerically matched to at all tested ).
Sanity checks
Trace preservation of and to ; exactly; a valid density matrix with Bloch vector ; record probabilities normalized to ; eigenvalue 1 present in both spectra to ; the i.i.d. limit recovered as ( at , ). All matrix algebra is dense and in IEEE double precision. At the period-2 point , and wherever eigenvalue branches cross, the spectra are evaluated from the closed forms of this section; numerical diagonalization is a cross-check, which is the content of the matching bounds quoted above.
Rates versus spectra
Record level (extended gate): fitted ACF decay rates equal to six significant digits (; see the table in Sec. VIII). Idle level (base gate): the closed form matches numerics to . Cross-level: at all tested points; idle mixing implies record mixing at all tested points; the converse fails (base gate: record i.i.d., idle non-mixing). The rank-one phenomenon of the base gate ( for every measure-and-prepare instrument) is proved by writing the write map and read map and observing that reads only while is independent of the variable part of .
References
- [1]A. N. Kolmogorov, Foundations of the Theory of Probability (Chelsea, New York, 1950) [English translation of Grundbegriffe der Wahrscheinlichkeitsrechnung (Springer, Berlin, 1933)].
- [2]G. D. Birkhoff, Proof of the ergodic theorem, Proc. Natl. Acad. Sci. USA 17, 656–660 (1931).
- [3]N. Wiener, Extrapolation, Interpolation, and Smoothing of Stationary Time Series (MIT Press, Cambridge, MA, 1949).DOI
- [4]A. Ya. Khinchin, Korrelationstheorie der stationären stochastischen Prozesse, Math. Ann. 109, 604–615 (1934).
- [5]G. E. P. Box and G. M. Jenkins, Time Series Analysis: Forecasting and Control (Holden-Day, San Francisco, 1970).
- [6]M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information (Cambridge University Press, Cambridge, 2010), 10th anniversary ed.
- [7]V. P. Belavkin, Quantum stochastic calculus and quantum nonlinear filtering, J. Multivariate Anal. 42, 171–201 (1992).DOI
- [8]B. Kümmerer and H. Maassen, An ergodic theorem for quantum counting processes, J. Phys. A 36, 2155–2161 (2003).
- [9]B. Kümmerer and H. Maassen, A pathwise ergodic theorem for quantum trajectories, J. Phys. A 37, 11889–11896 (2004).DOI
- [10]T. Benoist, J.-L. Fatras, and C. Pellegrini, Limit theorems for quantum trajectories, Stochastic Process. Appl. 164, 288–310 (2023).arxiv.org/abs/2302.06191
- [11]F. A. Pollock, C. Rodríguez-Rosario, T. Frauenheim, M. Paternostro, and K. Modi, Operational Markov condition for quantum processes, Phys. Rev. Lett. 120, 040405 (2018).DOI
- [12]F. A. Pollock, C. Rodríguez-Rosario, T. Frauenheim, M. Paternostro, and K. Modi, Non-Markovian quantum processes: Complete framework and efficient characterization, Phys. Rev. A 97, 012127 (2018).DOI
- [13]S. Milz and K. Modi, Quantum stochastic processes and quantum non-Markovian phenomena, PRX Quantum 2, 030201 (2021).DOI
- [14]M. van Horssen and M. Guţă, Sanov and central limit theorems for output statistics of quantum Markov chains, J. Math. Phys. 56, 020109 (2015).DOI
- [15]M. Guţă and J. Kiukas, Equivalence classes and local asymptotic normality in system identification for quantum Markov chains, Commun. Math. Phys. 335, 1397–1428 (2015).
- [16]S. Milz, F. Sakuldee, F. A. Pollock, and K. Modi, Kolmogorov extension theorem for (quantum) causal modelling and general probabilistic theories, Quantum 4, 255 (2020).DOI
- [17]P. Taranto, F. A. Pollock, S. Milz, M. Tomaničel, and K. Modi, Quantum Markov order, Phys. Rev. Lett. 122, 140401 (2019).DOI
- [18]P. Taranto, S. Milz, F. A. Pollock, and K. Modi, Structure of quantum stochastic processes with finite Markov order, Phys. Rev. A 99, 042108 (2019).DOI
- [19]M. Fannes, B. Nachtergaele, and R. F. Werner, Finitely correlated states on quantum spin chains, Commun. Math. Phys. 144, 443–490 (1992).DOI
- [20]G. Chiribella, G. M. D’Ariano, and P. Perinotti, Quantum circuit architecture, Phys. Rev. Lett. 101, 060401 (2008).pubmed.ncbi.nlm.nih.gov/18764438
- [21]G. Chiribella, G. M. D’Ariano, and P. Perinotti, Theoretical framework for quantum networks, Phys. Rev. A 80, 022339 (2009).DOI
- [22]J. Keeling, E. Stoudenmire, M. Baňuls, and D. Reichman, Process tensor approaches to non-Markovian quantum dynamics, Phys. Rev. X 16, 020502 (2026); arXiv:2509.07661.DOI
- [23]P. Taranto, F. A. Pollock, and K. Modi, Non-Markovian memory strength bounds quantum process recoverability, npj Quantum Inf. 7, 149 (2021).DOI
- [24]P. Taranto, S. Milz, M. Murao, M. T. Quintino, and K. Modi, Higher-order quantum operations, arXiv:2503.09693 (2025).arxiv.org/abs/2503.09693
- [25]V. Link, H.-H. Tu, and W. T. Strunz, Open quantum system dynamics from infinite tensor network contraction, Phys. Rev. Lett. 132, 200403 (2024).DOI
- [26]M. Garbellini, K. Mickiewicz, V. Link, A. Eisfeld, and W. T. Strunz, Uniform process tensor approach for the calculation of multi-time correlation functions of non-Markovian open systems, J. Chem. Phys. 164, 184104 (2026).DOI
- [27]D. Kretschmann and R. F. Werner, Quantum channels with memory, Phys. Rev. A 72, 062323 (2005).DOI
- [28]A. Monras, A. Beige, and K. Wiesner, Hidden quantum Markov models and non-adaptive read-out of many-body states, Appl. Math. Comput. Sci. 3, 93 (2011).arxiv.org/abs/1002.2337
- [29]A. Monras and A. Winter, Quantum learning of classical stochastic processes: The completely positive realization problem, J. Math. Phys. 57, 015219 (2016).
- [30]C. Yang, F. C. Binder, V. Narasimhachar, and M. Gu, Matrix product states for quantum stochastic modelling, Phys. Rev. Lett. 121, 260602 (2018).DOI
- [31]T. Benoist, A. Hautecoeur, and C. Pellegrini, Quantum trajectories. Spectral gap, quasi-compactness and limit theorems, J. Funct. Anal. 289, 110932 (2025).arxiv.org/abs/2402.03879
- [32]L. Accardi, A. Frigerio, and J. T. Lewis, Quantum stochastic processes, Publ. RIMS Kyoto 18, 97–133 (1982).
- [33]B. Kümmerer, Markov dilations on W*-algebras, J. Funct. Anal. 63, 139–177 (1985).DOI
- [34]P. Taranto, T. J. Elliott, and S. Milz, Hidden quantum memory: Is memory there when somebody looks?, Quantum 7, 991 (2023); arXiv:2204.08298.DOI
- [35]S. Milz, D. Egloff, P. Taranto, T. Theurer, M. B. Plenio, A. Smirne, and S. F. Huelga, When is a non-Markovian quantum process classical?, Phys. Rev. X 10, 041049 (2020).
- [36]E. B. Davies and J. T. Lewis, An operational approach to quantum probability, Comm. Math. Phys. 17, 239–260 (1970).DOI
- [37]M. Ozawa, Quantum measuring processes of continuous observables, J. Math. Phys. 25, 79–87 (1984).DOI
- [38]M.-D. Choi, Completely positive linear maps on complex matrices, Linear Algebra Appl. 10, 285–290 (1975).DOI
- [39]A. Jamiołkowski, Linear transformations which preserve trace and positive semi-definiteness of operators, Rep. Math. Phys. 3, 275–278 (1972).
- [40]E. Prugovečki, Information-theoretical aspects of quantum measurement, Int. J. Theor. Phys. 16, 321–331 (1977).DOI
- [41]J. M. Renes, R. Blume-Kohout, A. J. Scott, and C. M. Caves, Symmetric informationally complete quantum measurements, J. Math. Phys. 45, 2171–2180 (2004).DOI
- [42]G. A. L. White, F. A. Pollock, L. C. L. Hollenberg, C. D. Hill, and K. Modi, Non-Markovian quantum process tomography, PRX Quantum 3, 020344 (2022).DOI
- [43]J. v. Neumann, Proof of the quasi-ergodic hypothesis, Proc. Natl. Acad. Sci. USA 18, 70–82 (1932).pubmed.ncbi.nlm.nih.gov/16577432
- [44]U. Krengel, Ergodic Theorems (Walter de Gruyter, Berlin, 1985).
- [45]M. Rosenblatt, A central limit theorem and a strong mixing condition, Proc. Natl. Acad. Sci. USA 42, 43–47 (1956).pubmed.ncbi.nlm.nih.gov/16589813
- [46]I. A. Ibragimov and Yu. V. Linnik, Independent and Stationary Sequences of Random Variables (Wolters-Noordhoff, Groningen, 1971).
- [47]R. C. Bradley, Basic properties of strong mixing conditions. A survey and some open questions, Probab. Surv. 2, 107–144 (2005).DOI
- [48]F. Ciccarello, S. Lorenzo, V. Giovannetti, and G. M. Palma, Quantum collision models: Open system dynamics from repeated interactions, Phys. Rep. 954, 1–70 (2022).DOI
- [49]H. Wang, Two levels of mixing in quantum time series: dissipation dodging, key-locked memories, and a no-broadcasting lemma, in preparation (2026).
- [50]H. Barnum, C. M. Caves, C. A. Fuchs, R. Jozsa, and B. Schumacher, Noncommuting mixed states cannot be broadcast, Phys. Rev. Lett. 76, 2818–2821 (1996).pubmed.ncbi.nlm.nih.gov/10060796
- [51]D. E. Evans and R. Höegh-Krohn, Spectral properties of positive maps on C*-algebras, J. London Math. Soc. (2) 17, 345–355 (1978).DOI
- [52]M. M. Wolf, Quantum channels and operations: guided tour, lecture notes, TU Munich (2012).
- [53]V. S. Paulsen, Completely Bounded Maps and Operator Algebras (Cambridge University Press, Cambridge, 2002).DOI
- [54]L. Bruneau, A. Joye, and M. Merkli, Asymptotics of repeated interaction quantum systems, J. Funct. Anal. 239, 310–344 (2006).
- [55]For an i.i.d. process, every deterministic channel has unital transfer map; this does not affect uniqueness under the universal, all-processes quantifier.

