{"id":"406dd3c3-fee0-4bba-8bca-854dddc9bbc2","arxiv_id":"1908.08085","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"First photonic experiment demonstrating that mixing two Markovian channels can produce a non-Markovian channel, and that mixing two non-Markovian channels can appear Markovian under the distinguishability criterion.","lead":"This paper reports photonic experiments that mix pairs of quantum channels to test whether two memoryless (Markovian) processes can be added to create a process with memory, and whether the reverse is also possible. The data support the first effect and partially support the second, while showing that experimental noise makes Markovianity very hard to certify.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'nM + nM = M' claim is not established: a decreasing trace distance for one fixed state pair does not certify BLP-Markovianity, and the M+M summands in Fig. 3(b) fail the divisibility test.","rationale":"The reader's verdict is already CONDITIONAL, and I do not recommend moving it. The concern is not about fabrication or alignment but about the logical strength of the evidence relative to the paper's own strongest sentence. A decreasing trace distance for one state pair is necessary but not sufficient for BLP Markovianity, so the sentence in Section III.C overstates what Fig. 5(f) proves. Similarly, Fig. 3(b) shows the assumed Markovian summands failing the divisibility test; the paper candidly discusses this, but it means the experiment demonstrates fragility rather than a clean M+M=nM example. Both gaps are fixable with the existing data (all-pair trace-distance check) or with new measurements with larger count statistics or a channel with strictly positive intermediate eigenvalues. Because these gaps match the paper's own stated limitations and do not affect the theoretical side or the qualitative observations, the conditional acceptance remains appropriate.","tokens_in":16000,"tokens_out":10982,"duration_ms":116369,"concrete_test":"Use the stored 100 reconstructed χ matrices for the total channel (or the raw counts) to compute the trace distance D(t) between a dense grid of initial pure-state pairs, or at least the pair that maximizes D(t) for the reconstructed dephasing map, over t∈[π/4,π/2]. If any pair shows an increase, the BLP-Markovianity claim for the summed channel is falsified; additionally, compute the intermediate Choi eigenvalues for the total channel and test whether they are positive within propagated Poisson error, which would certify Markovianity by divisibility rather than by one decreasing curve.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim, 'nM + nM = M, in terms of distinguishability,' is inferred from Fig. 5(f), where the trace distance between |H⟩ and |V⟩ decreases for the total channel. Under the BLP criterion, a channel is Markovian only if the trace distance is non-increasing for all initial state pairs; a single pair gives only a lower bound on the BLP measure and cannot rule out backflow for another pair. Since full QPT was performed on the total channel, the authors have the data to check all pairs but do not report this check, so the quoted confirmation does not follow. The M+M=nM case has a complementary gap: the individual channels are assumed Markovian, yet their intermediate Choi eigenvalues λmin are negative for every time pair in Fig. 3(b); the paper attributes this to ideal λmin=0 plus noise. That is plausible, but it means the experiment never directly verifies that the summands are Markovian; it only verifies that two noisy near-Markovian channels add to a more strongly non-Markovian channel. The central non-convexity phenomenon is therefore asserted rather than demonstrated at the level claimed.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a photonic experiment that aims to demonstrate the non-convexity of the sets of Markovian and non-Markovian quantum channels. Using a two-path linear-optics setup, it realizes (i) an equal mixture of two dephasing channels, Eqs. (1)-(3), predicted to produce a non-Markovian channel (M + M = nM), and (ii) a weighted 2:1 mixture of two non-Markovian dephasing channels, Eqs. (5)-(7), predicted to produce a Markovian channel (nM + nM = M). The authors perform quantum process tomography at multiple times, reconstruct intermediate Choi eigenvalues, compute the RHP divisibility-based measure, and for the second case also the BLP trace distance for the |H>, |V> pair. They report high process fidelities, find negative minimum Choi eigenvalues for the total channel in case (i), and a decreasing trace distance for the total channel in case (ii). The paper candidly documents places where the data do not match ideal theory, and it uses Monte Carlo error propagation (100 chi matrices per time, 10,000 intermediate maps). The central conclusions, however, are currently stronger than the evidence supports.","tokens_in":16177,"tokens_out":7960,"duration_ms":79644,"significance":"If fully established, this would be the first experimental test of a well-known but counterintuitive theoretical fact: the convex combination of Markovian channels can be non-Markovian, and vice versa. The paper is careful in its error treatment and in reporting disagreements with ideal theory, such as the negative lambda_min of the individual channels in Fig. 3(b). Its strengths include a clear experimental realization of channel addition and the use of two independent Markovianity criteria. Nevertheless, the two case studies have load-bearing gaps: the summands in the first case are not shown to be Markovian, and the 'confirmation' in the second case rests on a single state pair, which is insufficient for the BLP criterion. The manuscript is worth publishing after the claims are corrected and the missing checks are either performed or explicitly identified as open.","major_comments":[{"comment":"The M + M = nM case is not experimentally established because the realized individual channels fail the same Markovianity test: lambda_min for Lambda_1 and Lambda_2 is negative for every time pair shown in Fig. 3(b). The paper attributes this to ideal lambda_min = 0 plus experimental noise, but that is an assumption rather than a measurement. The data therefore demonstrate that two weakly non-Markovian channels add to a more strongly non-Markovian channel; they do not directly demonstrate that two Markovian channels can produce a non-Markovian one. To support the central claim, the authors would need either improved precision that resolves lambda_min >= 0 for the summands or an example whose Markovian summands have strictly positive intermediate Choi eigenvalues, which the paper itself notes is not currently known.","section":"III.B, Fig. 3(b)"},{"comment":"The statement that two non-Markovian channels have been added to make a Markovian channel 'in terms of distinguishability' does not follow from the data shown. The BLP criterion requires the trace distance between any two initial states to be non-increasing for all times; a decreasing trace distance for a single pair, |H> and |V>, provides only a lower bound on the BLP measure and cannot rule out backflow for another pair such as |+> and |->. Because full chi matrices were reconstructed, the authors should compute D(t) for the complete set of initial pairs, or at least report the maximum over a sufficiently dense set, and show that no derivative is positive before claiming Markovianity in the distinguishability sense.","section":"III.C, Fig. 5(f)"},{"comment":"The two criteria give conflicting verdicts for the total channel in the second case: by CP divisibility the total channel is non-Markovian (lambda_min < 0 in the inset of Fig. 4(b), gbar(t) > 0 in Fig. 5(c)), while by the single-pair trace distance it appears Markovian. The paper acknowledges this in the body, but the summary in Sec. IV states only that the total channel was 'Markovian in terms of distinguishability'. Given the mismatch, the safe conclusion is that the experiment realizes a channel that is non-Markovian under CP divisibility and exhibits decreasing trace distance for one state pair; it does not demonstrate that two non-Markovian channels sum to a Markovian channel under either standard criterion.","section":"III.C, Figs. 4(b) and 5(c)"},{"comment":"The central quantitative claim relies on the assumption that the realized map is exactly the convex combination in Eqs. (3) and (7): the two paths must be added incoherently with the specified weights and no path-dependent loss. The paper states that the paths are not temporally aligned and that the 50% loss at the second beamsplitter is unbiased, but it provides no control measurement supporting these assumptions, such as a characterization of the effective mixture weights with a known input state or a check that blocking each path yields the expected relative transmission. Without such a check, deviations in Figs. 3 and 5 could also be explained by a mismatch between the realized and intended channel weights.","section":"II, 'Experimental setup'"}],"minor_comments":[{"comment":"The definition of p1(t) should be written with explicit parentheses, e.g., p1(t) = (3/2)[(1 + e^{-t})/2 - (1/3) cos^2 t], and it would be helpful to state that 0 <= p_i(t) <= 1 for the considered time range.","section":"II.B, Eq. (5)"},{"comment":"The transfer matrix F(s) is inverted to obtain F(t,s) = F(t) F(s)^{-1}; the paper should state how the inversion is regularized when the reconstructed F(s) is not exactly invertible due to noise.","section":"III.A, Eq. (8)"},{"comment":"The data points for channels 1 and 2 are shifted horizontally 'for clarity'; this should be noted in both captions so that readers do not interpret the shift as a physical time offset.","section":"Figs. 3(b), 4(b)"},{"comment":"Because any state pair gives only a lower bound on the BLP measure, the phrase 'confirm experimentally ... nM + nM = M' should be replaced by a statement such as 'consistent with Markovian behavior for the tested pair'.","section":"III.C, trace distance discussion"},{"comment":"The phrase 'not interferometrically aligned temporally' should be quantified, for example by giving the path-length difference relative to the photon coherence length, to strengthen the incoherent-addition argument.","section":"II, 'Experimental setup'"}],"recommendation":"major_revision","confidential_remarks":"The predictions tested in this paper come from Refs. [39] and [43], on which co-author Wudarski appears. The experimental data are new and the prior work is properly cited, so I do not regard this as circular; nonetheless, the authors may wish to note explicitly in the revised text that the predictions being tested originate from their own earlier theoretical work, as this is relevant context for readers. The manuscript fits the journal's scope and the experimental effort is substantial, but the two main claims need to be reworded or supplemented with the additional checks described in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look. This is the first experiment I know of that actually realizes the M+M=nM and nM+nM=M channel combinations in a photonic setup, and that alone is a meaningful step. The setup is simple, the QPT is careful, and the first case is well supported: the total channel's Choi eigenvalue goes clearly negative for all tested time pairs, so the experiment genuinely shows two roughly-Markovian dephasing channels adding to a non-Markovian one.\n\nThe soft spot is the second case. The authors claim to confirm nM+nM=M 'in terms of distinguishability' from a decreasing trace distance for the |H>,|V> pair. That does not follow. BLP Markovianity requires the trace distance to be non-increasing for every initial state pair; one pair can only give a lower bound on the non-Markovianity measure, not a certification of Markovianity. They performed full QPT on the total channel and could check all pairs, but they don't report that check. The divisibility criterion, which they also measured, fails to show the total channel is Markovian. They concede this, and even note that the total channel's g-bar(t) is far from zero. So the strong statement in the summary is not supported.\n\nThere's a second, smaller gap in the first case: the individual Markovian summands have negative intermediate Choi eigenvalues for every time pair. The authors attribute this to ideal lambda_min=0 plus noise, which is plausible, but it means the 'Markovian' label for the summands is asserted, not verified. What is verified is that two noisy near-Markovian channels add to a more strongly non-Markovian one.\n\nTo be fair, the paper is honest about most of this, and the data analysis looks sound. The errors are handled properly and the process fidelities are good. The problem is not sloppy experiment but an overreaching interpretation of the distinguishability result.\n\nWho gets value: people working on non-Markovianity measures, and experimentalists simulating open quantum systems. It deserves a serious referee, but the referee should push for either all-pair BLP check or a softened claim. I'd recommend major revision with a request for the missing analysis.","headline":"First experimental test of non-convex channel addition, but the nM + nM = M claim rests on one state pair and does not survive the divisibility criterion.","tokens_in":16761,"tokens_out":2236,"would_cite":true,"duration_ms":22801,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A photonic experiment shows that mixing two Markovian dephasing channels can create non-Markovian dynamics, and mixing two non-Markovian channels can restore Markovian behavior.","keywords":["quantum channels","Markovianity","non-Markovianity","channel addition","non-convex geometry","dephasing channels","quantum process tomography","open quantum systems"],"falsifier":"Temporally align the two paths, scan the relative phase, and measure the interference visibility: any nonzero visibility would indicate residual coherence between the paths, meaning the realized map is not the claimed incoherent convex sum. Separately, block each path in turn and measure the transmitted count rate at the second beamsplitter to verify that the loss is truly path-independent; unequal loss would alter the effective mixing weights and invalidate the quantitative comparison to theory.","tokens_in":15779,"feed_emoji":"🔀","tokens_out":5132,"duration_ms":52091,"temperature":0.7,"pith_summary":"The paper tries to establish experimentally that Markovian and non-Markovian quantum channels do not form convex sets: taking a weighted average of two Markovian channels can produce a non-Markovian channel, and taking a weighted average of two non-Markovian channels can produce a Markovian channel. The authors build a photonic single-qubit setup in which a polarization-encoded qubit passes through one of two paths, each implementing a different dephasing channel, and the paths are recombined incoherently so the net effect is exactly a convex combination. They classify the resulting channels using two standard criteria: divisibility of the dynamics (via positivity of the intermediate Choi matrices) and distinguishability of evolved states (via trace distance). The experiments confirm the M + M = nM addition clearly with the divisibility criterion, and confirm the nM + nM = M addition with the distinguishability criterion, while also exposing that the divisibility criterion is fragile when the idealized channel sits exactly on the boundary of Markovian behavior.","feed_headline":"Two Markovian channels can add up to a non-Markovian one","feed_subtitle":"A photonic experiment confirms the non-convex geometry of quantum channels, with consequences for classifying memory effects.","key_machinery":"The central object is the convex sum of two quantum channels realized by a two-path interferometer without temporal alignment. A single photon is split at a beamsplitter, each path applies a dephasing channel implemented by probabilistically switching half-wave-plate angles between identity and a Pauli operation, and the paths are recombined at a second beamsplitter whose 50% loss is unbiased and whose paths are not interferometrically aligned, so the resulting output map is the incoherent weighted sum of the two path channels. Classification is carried out by quantum process tomography followed by construction of transfer matrices and the Choi matrix of the intermediate map, whose lowest eigenvalue being negative signals non-divisibility and hence non-Markovianity.","core_discovery":"The central claim is that the non-convex geometry of Markovian and non-Markovian channels is physically real in the laboratory. Specifically, for a qubit subject to two independent phase-damping channels, the equal mixture $\\Lambda_t^{(T)}(\\rho)=\\frac{1}{2}(\\Lambda_t^{(1)}(\\rho)+\\Lambda_t^{(2)}(\\rho))$ of two Markovian dephasing channels along the $x$ and $y$ axes is a non-Markovian channel, while the weighted mixture $\\Lambda_t^{(T)}(\\rho)=\\frac{2}{3}\\Lambda_t^{(1)}(\\rho)+\\frac{1}{3}\\Lambda_t^{(2)}(\\rho)$ of two non-Markovian dephasing channels along the $x$ axis is a Markovian phase-damping channel. The authors verify the first case by observing negative lowest eigenvalues of the intermediate Choi matrices, and verify the second case by observing a monotonically decreasing trace distance between two probe states in the interval $t\\in[\\pi/4,\\pi/2]$, concluding that two non-Markovian channels have been added to make a Markovian channel in terms of distinguishability.","pith_inferences":["A direct extension would be to scan the mixing weight continuously and map the boundary where the sum of two non-Markovian dephasing channels becomes Markovian, giving a quantitative phase diagram for channel addition.","The same two-path addition method could be applied to mixtures involving depolarizing or amplitude-damping channels to test whether the non-convexity observed here is a generic feature of channel sets rather than specific to dephasing.","The authors' identified fragility suggests that future experiments should design Markovian test channels with strictly positive Choi eigenvalues, so that divisibility can certify Markovianity without relying on zero-valued boundaries.","One could test the distinguishability-based Markovianity witness with more than two initial states to see whether the nM + nM = M conclusion is robust for all probe pairs, not just the chosen $|H\\rangle,|V\\rangle$ pair."],"forward_implications":["If correct, the results show that proving a channel is Markovian by decomposing it into a mixture of Markovian building blocks is invalid: the mixture can be non-Markovian.","The results imply that non-Markovian noise can be cancelled or weakened by mixing, so a combination of two memory-bearing channels can behave like a memoryless channel in terms of distinguishability.","The fragility of the Choi-positivity criterion for boundary cases means that experimental claims of Markovianity should not rest solely on divisibility when the ideal channel has a zero lowest eigenvalue.","For practical quantum error correction, the work implies that noise models built from convex combinations of simple Markovian dephasings may hide non-Markovian features that affect memory-enhanced correction protocols."],"supporting_citations":[{"why":"Supplies the theoretical example of two Markovian dephasing channels whose equal convex combination is non-Markovian.","marker":"[39]"},{"why":"Supplies the theoretical example of two non-Markovian dephasing channels whose weighted convex combination is a Markovian semigroup.","marker":"[43]"},{"why":"Provides the divisibility-based non-Markovianity criterion and measure used to test the intermediate maps.","marker":"[9]"},{"why":"Provides the trace-distance distinguishability measure used to witness non-Markovian behavior in the second case.","marker":"[10]"},{"why":"Supplies the quantum process tomography protocol used to reconstruct the channel matrices from the experiment.","marker":"[64]"}],"fun_headline_variants":["Two memoryless channels can make a memoryful one","Adding Markovian channels yields non-Markovian — experiment","Photon experiment shows channel addition breaks memory rules","Memoryless + memoryless = memory: quantum channel surprise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiment's conclusion rests on the assumption that the two optical paths are combined incoherently without residual interference, and that the 50% loss at the recombining beamsplitter treats both paths equally, so the realized output map is exactly the weighted average of the two individual channel maps.","fun_headline_variants_meta":{"raw":{"variants":["Two memoryless channels can make a memoryful one","Adding Markovian channels yields non-Markovian — experiment","Photon experiment shows channel addition breaks memory rules","Memoryless + memoryless = memory: quantum channel surprise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000445,"raw_usage":{"total_tokens":2247,"prompt_tokens":939,"completion_tokens":1308,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1243}},"tokens_in":555,"tokens_out":1308,"duration_ms":10107,"temperature":1.0,"reasoning_tokens":1243,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:49:39.453548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Temporally align the two paths, scan the relative phase, and measure the interference visibility: any nonzero visibility would indicate residual coherence between the paths, meaning the realized map is not the claimed incoherent convex sum. Separately, block each path in turn and measure the transmitted count rate at the second beamsplitter to verify that the loss is truly path-independent; unequal loss would alter the effective mixing weights and invalidate the quantitative comparison to theory.","supporting_citations":[{"cited_title":"Megier, D","cited_arxiv_id":null,"evidence_quote":"Supplies the theoretical example of two Markovian dephasing channels whose equal convex combination is non-Markovian."},{"cited_title":"Chru ´sci´nski, F","cited_arxiv_id":null,"evidence_quote":"Supplies the theoretical example of two non-Markovian dephasing channels whose weighted convex combination is a Markovian semigroup."},{"cited_title":"Rivas, S","cited_arxiv_id":null,"evidence_quote":"Provides the trace-distance distinguishability measure used to witness non-Markovian behavior in the second case."},{"cited_title":"Lindblad, On the generators of quantum dynamical semi- groups, Comm","cited_arxiv_id":null,"evidence_quote":"Supplies the quantum process tomography protocol used to reconstruct the channel matrices from the experiment."}],"review_version":1}