{"id":"4ceafb15-ea63-4c0d-92d3-7ed5208e51b5","arxiv_id":"2502.04555","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors define Partial Information Rate Decomposition (PIRD), a spectral extension of PID that splits the mutual information rate between random processes into unique, redundant, and synergistic components.","lead":"This paper introduces a way to break down the information flowing between time-varying signals into unique, shared, and synergistic pieces, working frequency by frequency. It matters for physiology and neuroscience, where signals like blood pressure and brain blood flow interact rhythmically and older methods ignore temporal correlations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SMMI redundancy inherits MMI's negative-atom failure for M>2; even a memoryless Gaussian example with an irrelevant third source yields negative coarse-grained synergy, contradicting the claimed non-negative PIRD.","rationale":"The paper is careful in its definitions and appendices, and the M=2 simulations are internally consistent. However, the reader's weakest assumption focused on the Gaussian/VAR practical implementation, whereas the more load-bearing defect is internal to the Gaussian setting itself: the SMMI redundancy rate, as a frequency-wise minimum-MI measure, inherits the known failure of MMI-based redundancy to guarantee non-negative atoms for more than two sources. The counterexample above is exactly the memoryless limit the paper claims to recover, so it attacks the central non-negativity claim without invoking non-stationarity, non-Gaussianity, or estimation error. This does not destroy the whole framework; the decomposition still provides a valid additive accounting of the MIR, and the M=2 case is sound. But the advertised non-negative unique-redundant-synergy interpretation is unsupported for M>=3, and the physiological application uses M=3. The correct response is to require a proof of non-negativity under stated conditions or an explicit restriction of the non-negativity claim to M=2, together with a re-interpretation of the M=3 results if negative atoms occur. This keeps the reader's conditional verdict, but sharpens the condition that must be met before the paper can be trusted as a validated tool.","tokens_in":30572,"tokens_out":23270,"duration_ms":242555,"concrete_test":"Use a VAR(0) Gaussian model with Z=(Y,X1,X2,X3), exact covariance chosen so that X1=X2, X3 is independent of all other variables, and Y=X1+N with small additive noise. Compute the SMMI PIRD from the exact PSD matrices via Eqs. (23)-(25) and Eq. (15), with no finite-sample estimation. If the coarse-grained synergy rate is negative, or if any PI rate atom from Eq. (25) is negative, the non-negativity claim fails. A useful variation is to let the X1-X2 correlation approach 1 from below; the coarse-grained synergy should approach -I_{Y;X2} rather than zero.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the PIRD with the SMMI redundancy rate yields a non-negative decomposition is not established for M>2, and appears to be false. In the memoryless Gaussian limit the authors themselves show that SMMI coincides with the standard MMI redundancy (Appendix A.1), and MMI-based redundancy is known to produce negative partial information atoms when more than two sources are considered. Concretely, let M=3 memoryless Gaussian variables satisfy X1=X2, X3 independent of Y and of X1,X2, and Y=X1+N with small noise. Then all information about Y is carried redundantly by X1 and X2, while X3 contributes nothing. The bottom redundancy rate in Eqs. (23)-(24) is zero because min(I_{Y;X1}, I_{Y;X2}, I_{Y;X3})=0. Equation (15) then gives U_{Y;X1}=I_{Y;X1}, U_{Y;X2}=I_{Y;X2}, and the coarse-grained synergy is S_{Y;X}=I_{Y;X}-U_{Y;X1}-U_{Y;X2}-R_{Y;X}\\approx -I_{Y;X2}<0. The same negativity appears in the spectral PI rate at the atom {1}{23} through Möbius inversion of the SMMI redundancy function. Since the paper explicitly claims a non-negative decomposition and applies the M=3 version to the physiological network, this is a property of the proposed redundancy function rather than a finite-sample or non-Gaussian estimation issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Partial Information Rate Decomposition (PIRD) that extends the Williams-Beer partial information decomposition from random variables to stationary random processes, replacing mutual information by mutual information rate and using a newly defined spectral minimum-MI (SMMI) redundancy rate. The decomposition is implemented for Gaussian processes through the power spectral density and VAR models, and it is applied to simulated VAR networks and to a physiological network with three source processes. The paper claims that PIRD yields non-negative unique, redundant, and synergistic information-rate contributions, captures full temporal and spectral structure, and reduces to static PID in the memoryless case and to PID of transfer entropy in the strictly causal case.","tokens_in":30913,"tokens_out":11968,"duration_ms":128716,"significance":"The conceptual step of decomposing mutual information rate over the same redundancy lattice as PID is natural and potentially useful, and the frequency-domain formulation offers a computationally tractable, spectrally resolved decomposition for Gaussian VAR processes. The appendix proofs (A.1-A.3) give useful support for the basic properties of the SMMI redundancy rate. However, the central non-negativity claim fails for more than two sources, and some of the reduction claims are overstated. The M=3 physiological application therefore rests on an unsupported property of the proposed redundancy function.","major_comments":[{"comment":"The claimed non-negative decomposition is false for M>2. In the memoryless Gaussian limit the SMMI redundancy rate coincides with the MMI redundancy rate (Appendix A.1). Take M=3 with Y=X1+N, X1=X2, X3 independent of Y, X1, X2, and N independent small Gaussian noise. Then I(Y;X3)=0, so R_{Y;X}=min(I_{Y;X1}, I_{Y;X2}, I_{Y;X3})=0, and Eq. (15) gives S_{Y;X}=I_{Y;X1,X2,X3}-U_{Y;X1}-U_{Y;X2}-U_{Y;X3}-R_{Y;X}=I-I-0-0=-I<0. For the full lattice, take X1=X2=X3=Y+N_i with i.i.d. noises so all individual MIs are equal to some finite I; Eq. (25) gives PI({1})=I-PI({1}{2}{3})-PI({1}{2})-PI({1}{3})=I-I-I-I=-2I<0. Since the M=3 physiological analysis uses exactly this redundancy, the non-negativity claim and the interpretation of the M=3 atoms are unsupported. The authors should either restrict non-negativity claims and applications to M=2, impose additional constraints to enforce non-negativity, or adopt a redundancy function known to yield non-negative atoms.","section":"Frequency-domain PIRD, Eqs. (23)-(25), (15)"},{"comment":"The statement that in the strictly causal case 'the PIRD reduces to a PID applied to the TE' is only true at the level of the total quantity. The PIRD is solved with the SMMI redundancy rate (23)-(24), whereas the PID of the transfer entropy as implemented in the paper uses a time-domain MMI redundancy (17). Because the integral of a minimum of spectra is not equal to the minimum of the integrals (Appendix A.1 only proves an inequality), the PIRD atoms and the TE-based MMI atoms will generally differ in the presence of self-dependencies. This should be stated explicitly, or the comparison in Section 'Effects of changes in the network topology' should be framed as comparing two different redundancy schemes rather than as PIRD 'reducing' to TE-PID.","section":"Partial Information Rate Decomposition, Eq. (16)"},{"comment":"The simulations are not an independent validation of the decomposition: the VAR parameters are chosen to produce the behaviors that PIRD then recovers, as the text itself says ('Simulations are designed to induce expected behaviors and thus aim to provide a validation'). This demonstrates internal consistency with the chosen axioms but does not test whether SMMI is a correct or interpretable redundancy function. An independent check for M=3 would require a setting with known, non-negative ground-truth atoms, or at least a direct stress test of the negative-atom issue described above.","section":"Theoretical Examples, first paragraph"}],"minor_comments":[{"comment":"There are several typos that should be corrected: 'th spontaneous variability' should be 'the spontaneous variability'; 'operalization' should be 'operationalization'; 'strightforward' should be 'straightforward'; 'withing' should be 'within'.","section":"Application to Physiological Networks"},{"comment":"The caption refers to 'the parameters c1, c2', but the text and the simulation use a single parameter c; this should be corrected.","section":"Figure 5 caption"},{"comment":"The term 'pointwise' is used for frequency-specific redundancy. This should be carefully distinguished from pointwise (realization-specific) local mutual information in the PID literature to avoid confusion.","section":"Frequency-domain PIRD"},{"comment":"The discussion of fitting a VAR model as 'assuming a linear model rather than a linear process' is useful, but the actual condition for Eqs. (28)-(29) is joint Gaussianity and stationarity; the Wold-decomposition argument alone does not justify the Gaussian assumption, and this should be clarified.","section":"Formulation for Gaussian processes"}],"recommendation":"major_revision","confidential_remarks":"The lattice machinery and the Gaussian/VAR implementation are sound for M=2, and the frequency-domain idea is worth publishing after revision. The main obstacle is the failure of non-negativity for M>2, which directly affects the physiological application. I would ask the authors to either restrict the scope to two sources, find a redundancy function that guarantees non-negative atoms, or substantially rewrite the claims and re-interpret the M=3 results with explicit caveats about possible negative atoms."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"To be candid: the core idea of this paper—swap MI for mutual information rate in the PID lattice, define a spectral minimum-MI redundancy, and recover static PID and TE-PID as limit cases—is genuinely new and should be taken seriously. The M=2 simulations are clean and the spectral decomposition is a nice trick. The appendices prove what they claim: the SMMI redundancy function satisfies the four Williams–Beer axioms.\n\nThe load-bearing problem is that the paper also claims a non-negative decomposition of the MIR. That does not hold for more than two sources. In the memoryless Gaussian limit the SMMI redundancy coincides exactly with the standard MMI redundancy (the paper proves this in A.1). MMI is known to produce negative partial information atoms for M≥3. A concrete example: three memoryless Gaussians with X1=X2, X3 independent of everything, Y=X1+N. Then R=min(I1,I2,I3)=0, U1=I1, U2=I2, and the coarse-grained synergy S=I−U1−U2−R≈−I2<0. Same thing happens with the spectral PI rates. Since the paper applies the M=3 version to the physiological network, this is not a minor technicality; it breaks the stated 'non-negative decomposition' promise.\n\nWhat else is soft: no code or data are released, the Gaussian/VAR assumptions are not validated on the physiological time series (only the model is fitted), and the simulation validation is self-referential—the behaviors are built in and then recovered. The paper honestly notes that redundancy choice is axiomatic, so that alone is not a flaw, but it weakens the claim that SMMI is the right choice.\n\nWhat survives: the lattice formalism and Möbius inversion are correct, the spectral integration equivalence (A.2) is fine, and the M=2 behavior is exactly what you want. The idea is worth publishing, but the non-negativity claim needs to be either proven under restricted conditions or removed. A revision that adds an explicit caveat about M≥3 and provides code would be credible.\n\nI'd send this to a serious referee. The idea is important, the formalism is mostly sound, and the flaw is a specific overclaim, not a collapse. But I would not cite it as a validated tool until the non-negativity issue is resolved.","headline":"PIRD is a genuine extension of PID to time series; the lattice formalism and M=2 simulations are solid, but the claimed non-negative decomposition fails for M>2 because SMMI inherits MMI's negative-atom problem.","tokens_in":31398,"tokens_out":4778,"would_cite":false,"duration_ms":45663,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","94A15","60G10","62H20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the mutual information rate between a target random process and a set of source processes can be decomposed over the same redundancy lattice as PID into non-negative unique, redundant, and synergistic…","keywords":["partial information decomposition","mutual information rate","spectral redundancy","multivariate time series","network physiology","Gaussian processes","transfer entropy","frequency-domain analysis"],"falsifier":"Take a stationary nonlinear vector process with a known mutual information rate, for example a bivariate process generated by a nonlinear coupling whose exact transfer entropy can be computed analytically, and compare the PIRD atoms computed through the Gaussian VAR route against a consistent model-free estimate of the same spectral MIR atoms; if the two sets of atoms disagree substantially or become negative, the Gaussian spectral implementation is not a faithful estimator of the general MIR decomposition.","tokens_in":30418,"feed_emoji":"📈","tokens_out":5542,"duration_ms":56013,"temperature":0.7,"pith_summary":"The paper introduces the partial information rate decomposition (PIRD), a framework that extends partial information decomposition (PID) from random variables to random processes. It replaces the mutual information between a target variable and source variables with the mutual information rate between a target process and source processes, and runs the same redundancy lattice of PID over this rate. The load-bearing new object is a spectral redundancy rate function: at each frequency, redundancy is the minimum of the spectral MIRs between the target and each source group, and the global redundancy rate is its integral over frequency. The paper argues that this yields a non-negative decomposition into unique, redundant, and synergistic information rates, that it equals static PID when the processes are memoryless, and that it equals PID applied to transfer entropy when the coupling is strictly causal with no instantaneous or reverse interactions. If correct, analysts of multivariate time series no longer need to pretend the samples are independent and identically distributed.","feed_headline":"Time series information decomposes without the memoryless assumption","feed_subtitle":"A spectral minimum-MI rule splits dynamic information into unique, redundant, and synergistic rates.","key_machinery":"The carried object is the spectral minimum mutual-information (SMMI) redundancy rate, defined in Eq. (23) as the minimum, at each frequency $\\omega$, of the spectral MIRs between the target process and the source groups composing an atom. For jointly Gaussian processes, each spectral MIR is given by the determinant formula $\\frac{1}{2}\\log\\frac{|P_X(\\omega)|P_Y(\\omega)}{|P_{[YX]}(\\omega)|}$, so every atom is computed from determinants of sub-blocks of the power spectral density matrix; the PSD itself is estimated by fitting a vector autoregressive model and applying spectral factorization. This machinery lets redundancy be defined locally in frequency, where the spectral MIR is non-negative, and then integrated to produce time-domain information-rate atoms, while also permitting band-limited decompositions restricted to specific oscillatory components.","core_discovery":"The central claim is that the information a target process $Y$ shares per unit time with source processes $X_1,\\ldots,X_M$ splits, without loss or double-counting, into atoms indexed by the redundancy lattice, exactly as static PID does for random variables. The split is achieved by defining a redundancy rate as a pointwise minimum of spectral MIRs: for each atom, the redundancy rate is the integral over frequency of the minimum, over the atom's elements, of the spectral MIR between the target and that element. For jointly Gaussian processes, each spectral MIR is computed from determinants of sub-blocks of the power spectral density matrix, and the paper shows that the spectral redundancy rate satisfies the standard PID axioms, that the frequency-domain and time-domain decompositions commute with integration, and that the method collapses to the zero-lag PID and to the PID of the joint transfer entropy in the two limiting dynamic regimes. The framework is therefore claimed to resolve the mismatch between PID's implicit memorylessness and the temporal correlations present in real network data, and to do so while preserving a non-negative, interpretable decomposition.","pith_inferences":["Beyond the paper: because the provided implementation is parametric and Gaussian, a natural testable extension is a model-free estimator of the spectral MIR and of the SMMI redundancy rate, which would show whether the atoms change materially when gaussianity or linearity fails.","Beyond the paper: the authors note that spectral pointwise minimization is less conservative than time-domain MMI redundancy; if that ordering holds more generally, PIRD may reduce the known overestimation of redundancy in MMI-style schemes and make redundancy rates more comparable across different redundancy functions.","Beyond the paper: the decomposition treats the chosen target and the specified sources only, so a hidden common driver outside the source set could inflate the redundant or synergistic atoms; conditioning on exogenous processes or embedding PIRD in a graphical model is a plausible next step that the paper does not develop.","Beyond the paper: the band-limited formulation suggests a direct application to any rhythmic dataset, such as neural oscillations or cardiorespiratory coupling, where whole-band information measures may obscure the coexistence of synergy in one band and redundancy in another."],"forward_implications":["Temporal correlations no longer disqualify a dataset from information decomposition: the MIR-based atoms remain non-negative and are well defined for processes with memory, unlike the zero-lag PID applied to dependent samples.","Frequency-band integration makes the decomposition scale-specific: unique, redundant, and synergistic information rates can be reported within physiologically meaningful bands such as low-frequency and high-frequency ranges, revealing higher-order interactions that may cancel out in whole-band averages.","The framework subsumes two existing practices: it reduces to the static PID in the memoryless case and to the PID of transfer entropy in the strictly causal case, so previous results can be reinterpreted as boundary cases of the same lattice construction.","In the physiological application, the finding that cerebrovascular and cardiovascular interactions are predominantly redundant and that redundancy increases with postural stress provides a spectrally resolved, target-specific readout of high-order coupling in network physiology.","The equivalence between frequency-domain and time-domain PIRD means that researchers can choose either representation freely, or combine them, without changing the resulting information-rate atoms."],"supporting_citations":[{"why":"Introduces the redundancy lattice and the consistency equations for PID that PIRD reuses intact for information rates.","marker":"[16]"},{"why":"Defines the minimum-MI redundancy principle that the spectral redundancy rate adapts frequency-by-frequency.","marker":"[34]"},{"why":"Defines transfer entropy, the quantity whose PID becomes a special case when coupling is strictly causal.","marker":"[44]"},{"why":"Provides the spectral expansion of mutual information rate for Gaussian processes used in Eqs. (27)-(28).","marker":"[45]"},{"why":"Supplies the spectral MIR representation and its multivariate interpretation for jointly Gaussian processes.","marker":"[46]"},{"why":"Defines the coarse-grained PID that PIRD extends to the rate domain for scalable unique, redundant, and synergistic terms.","marker":"[18]"},{"why":"Links Gaussian mutual information to variance ratios from linear regression, underpinning the closed-form spectral MIR.","marker":"[48]"},{"why":"Provides the vector autoregressive model whose least-squares identification and spectral factorization yield the PSD matrix used in the practical implementation.","marker":"[57]"}],"fun_headline_variants":["Spectral PID decomposes dynamic information rates in time series","PID extended to time series via spectral mutual information rates","Dynamic information split into unique, redundant, synergistic rates","Frequency-based redundancy rate solves memoryless PID limitation","Decompose information rates in networks of correlated processes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical version of the framework assumes the analyzed vector process is stationary and jointly Gaussian, so that the spectral MIR equals the determinant formula and the spectrum can be estimated from a finite-order VAR model; real physiological series may violate these conditions, and no model-free estimator of the decomposition is provided.","fun_headline_variants_meta":{"raw":{"variants":["Spectral PID decomposes dynamic information rates in time series","PID extended to time series via spectral mutual information rates","Dynamic information split into unique, redundant, synergistic rates","Frequency-based redundancy rate solves memoryless PID limitation","Decompose information rates in networks of correlated processes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1411,"prompt_tokens":1004,"completion_tokens":407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":620,"tokens_out":407,"duration_ms":4449,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T22:20:27.292553+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a stationary nonlinear vector process with a known mutual information rate, for example a bivariate process generated by a nonlinear coupling whose exact transfer entropy can be computed analytically, and compare the PIRD atoms computed through the Gaussian VAR route against a consistent model-free estimate of the same spectral MIR atoms; if the two sets of atoms disagree substantially or become negative, the Gaussian spectral implementation is not a faithful estimator of the general MIR decomposition.","supporting_citations":[{"cited_title":"Schreiber, Physical review letters85, 461 (2000)","cited_arxiv_id":null,"evidence_quote":"Defines transfer entropy, the quantity whose PID becomes a special case when coupling is strictly causal."},{"cited_title":"Geweke, Journal of the American statistical associa- tion 77, 304 (1982)","cited_arxiv_id":null,"evidence_quote":"Provides the spectral expansion of mutual information rate for Gaussian processes used in Eqs. (27)-(28)."},{"cited_title":"Chicharro, Biological cybernetics105, 331 (2011)","cited_arxiv_id":null,"evidence_quote":"Supplies the spectral MIR representation and its multivariate interpretation for jointly Gaussian processes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the coarse-grained PID that PIRD extends to the rate domain for scalable unique, redundant, and synergistic terms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Links Gaussian mutual information to variance ratios from linear regression, underpinning the closed-form spectral MIR."},{"cited_title":"Lütkepohl, New introduction to multiple time series analysis (Springer Science & Business Media, 2005)","cited_arxiv_id":null,"evidence_quote":"Provides the vector autoregressive model whose least-squares identification and spectral factorization yield the PSD matrix used in the practical implementation."}],"review_version":1}