{"id":"01457fa1-e28a-476e-a879-6bad44987d59","arxiv_id":"2501.16676","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Flexibility, defined as EI(joint) minus EI(system) minus EI(environment) under uniform interventions, is PID synergistic information and correlates with complex adaptive features like feedback loops and the edge of chaos.","lead":"This paper introduces a new indicator called flexibility, defined as the joint effective information of a system and its environment minus their individual effective informations. The authors prove it equals synergistic information in the partial information decomposition framework and test it on cellular automata, Boolean networks, and gene regulatory networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The formal PID identification in Theorem 1 is sound, but the central semantic claim that Eq. (7) measures real environmental responsiveness is supported only by a weak in-sample GRN correlation (r=0.487); a direct causal test is missing.","rationale":"Theorem 1 is a genuine mathematical result: after the do-intervention, the two source variables are independent, Lemma 1 forces Red=0, and Eq. (C10) follows from the PID lattice. I could not find a gap in that derivation; the proof is parameter-free and the calculation is reproducible from Eqs. (2)-(7). The vulnerability is at the level of interpretation. The abstract and introduction claim that Eq. (7) measures 'the ability of a system to flexibly respond to the environment.' For that to be true, the uniform do-intervention must be the right idealization of responsiveness. The paper's evidence is (i) CA correlations with Langton λ, which show that complex rules have high Syn but do not establish environmental responsiveness; (ii) FBL examples, which are qualitative and well-known; and (iii) the GRN experiment, whose quantitative anchor is a Pearson r=0.487 computed on one apoptosis network (Appendix D), with no held-out validation and no code. A 0.487 in-sample correlation leaves substantial room for confounds—e.g., motifs with high Syn may simply have more states or more output diversity, which would also correlate with any diversity-based response score. The concrete test isolates exactly this: rank-order motifs by Syn and by directly measured response diversity under non-uniform environment distributions. If the rank correlation holds out-of-sample, the semantic claim is supported; if not, the paper should retreat to the narrower formal claim that Syn is the PID synergy of the intervened mechanism, not a measure of real-world flexibility. Until that test is done, conditional acceptance is the right verdict.","tokens_in":13637,"tokens_out":8840,"duration_ms":89150,"concrete_test":"On the Kadelka et al. Boolean GRN corpus, compute Syn from Eq. (7) for every three-node motif in all 63 networks. For each motif, simulate the actual response protocol: initialize at every possible system state, apply each environmental state e drawn from a non-uniform empirical distribution (e.g., the observed state frequencies of the network), and measure the diversity of reachable steady states and the mutual information between environment switches and system-state changes, exactly as in Appendix D but with held-out motifs. Compute the Spearman rank correlation between Syn and this direct responsiveness score, both within the apoptosis network and across the full corpus. Repeat with two alternative system/environment partitions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is not the algebra but the semantics. Theorem 1 is proved correctly: after the intervention do(X^t,E^t∼U(Ω_X,E)), the sources are independent, Lemma 1 gives Red=0, and Eq. (C10) reduces Eq. (7) to the PID synergy of the intervened mechanism. But the paper's central claim—that this quantity measures a system's ability to flexibly respond to the environment—requires that the uniform do-intervention is the right idealization of real responsiveness. The only quantitative evidence for that identification is the GRN experiment in Appendix D: a Pearson r=0.487 between Syn and mean consecutive-moment mutual information under environmental switches, computed on subgraphs of one apoptosis model, with no out-of-sample validation, no code, and no variation of the environmental state distribution. A moderate in-sample correlation on one network cannot establish that the uniform-intervention synergy ranks actual systems by flexibility; it could be tracking generic state-space diversity or motif complexity. If, for non-uniform environmental distributions, systems with high Syn are not the ones that respond more diversely, the central claim fails even though Theorem 1 stands.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new quantity, called flexibility, defined as the difference between the effective information (EI) of a joint system-environment transition mechanism and the EIs of the two marginal mechanisms. The authors prove that, under a uniform do-intervention on the joint system-environment state, this quantity coincides with the synergistic information of the intervened sources with respect to the next system state in the partial information decomposition (PID) framework. They also decompose flexibility into two interpretable components, expansiveness and introversion, and study its behavior in cellular automata, random Boolean networks, and Boolean gene regulatory networks (GRNs). Finally, they show that a trained neural network can approximate the flexibility from simulated data when the underlying dynamics are unknown.","tokens_in":13808,"tokens_out":13665,"duration_ms":124001,"significance":"The formal identification of the proposed indicator with PID synergy is the paper's main strength. The proofs in Appendix C are correct given the stated axioms and the independence condition imposed by the uniform intervention, and the non-negativity of the synergy follows from the independence of the intervened sources. The decomposition into expansiveness and introversion offers a useful interpretative tool, and the machine-learning experiment provides a plausible path to data-driven application. The paper also generates falsifiable predictions, notably that feedback loops in GRNs exhibit high flexibility. However, the central semantic claim that flexibility measures a system's ability to flexibly respond to environmental changes rests on a single moderate in-sample correlation (r=0.487) on one apoptosis network, and the experimental sections omit crucial details about how the system/environment partition and the transition probability matrices are constructed. If the measure is to be adopted as an indicator of adaptive flexibility, the empirical validation and reproducibility must be substantially strengthened.","major_comments":[{"comment":"The GRN experiments do not specify how Boolean functions are converted into the transition probability matrices (TPMs) required to compute EI, nor how the system and environment variables are defined for the three-node subgraphs. The paper states that the GRNs are modeled as Boolean networks and refers to the repository of Kadelka et al., but it never describes the mapping from a deterministic Boolean update rule to a stochastic TPM, the treatment of asynchronous or noisy dynamics, or the choice of environmental variables (e.g., direct regulators outside the subgraph versus all remaining nodes). Without these details, the quantitative results in Figure 4 and the correlation in Figure 8 are not reproducible, and it is unclear whether the flexibility values reflect the intrinsic properties of the subgraphs or the arbitrary choices of embedding and TPM construction. Please provide a precise specification of the TPM generation and the partition rule for the GRN experiments.","section":"Section III.B and Appendix D"},{"comment":"The central claim that flexibility quantifies a system's ability to flexibly respond to environmental changes is supported only by a single in-sample Pearson correlation of r=0.487 (p<0.05) between Syn and a 'mean mutual information' measure computed on subgraphs of one apoptosis network. This is a moderate effect, obtained without out-of-sample validation, without controlling for state-space size, motif complexity, or other confounds, and without varying the distribution of environmental switches. Because this evidence is the main quantitative link between the mathematical quantity and the biological interpretation, I cannot regard the semantic claim as established. The paper should either provide a direct test with synthetic systems of known flexibility (e.g., systems whose responsiveness can be manipulated), randomize over multiple networks, or substantially temper the claims about what the correlation establishes.","section":"Section III.B, Figures 4 and 8"},{"comment":"The identification of flexibility with PID synergy depends entirely on the do-intervention do(X^t, E^t ~ U(Ω_{X,E})), which forces the system and environmental states to be independent and uniformly distributed. The paper motivates this by the desire to define a mechanism-level, state-independent quantity, but it provides no justification for why a uniform distribution of environmental inputs is the correct idealization for 'flexible response' in real environments. If the relevant environmental disturbances are non-uniform or rare, systems ranked by the uniform-intervention synergy need not be those that respond flexibly in practice. The paper should include a discussion of this modeling choice and ideally a sensitivity analysis showing how Syn changes under alternative input distributions.","section":"Section II.A, Eq. (7)"},{"comment":"The cellular automaton experiments do not explicitly define the system and environment variables for which flexibility is computed. The text mentions 'flexibility of a specific rule-based cellular automaton' and mutual information of 'a single cell', but it does not state whether the system X is one cell and the environment E is its two neighbors, or whether another partition is used. Since the value of Syn is partition-dependent, this omission makes the CA results ambiguous and prevents independent replication. Please state the partition explicitly and justify it.","section":"Section III.A, Figure 3"}],"minor_comments":[{"comment":"The base of the logarithm in the EI definition should be stated explicitly. The numerical values and the claimed ranges for Exp and Int (0 to 2 log2 |Ω_X|) suggest base 2, but this is not stated.","section":"Appendix B, Eq. (B1)"},{"comment":"The keyword list contains a typo: 'flexibilty' should be 'flexibility'.","section":"Keywords and text"},{"comment":"The notation for the intervention is inconsistent: the world-level intervention do(U^t ~ U(Ω_U)) is introduced first, but later the intervention is written as do(X^t, E^t ~ U(Ω_{X,E})). The relationship between these two formulations should be clarified.","section":"Section II.A"},{"comment":"The gene names in the caption ('BAG4, BAG4 TNFRSF1A, and TNF BAG4 TNFRSF1A') appear garbled or duplicated. Please verify the intended gene identifiers.","section":"Figure 4(c) caption"},{"comment":"The neural network experiment does not report the number of training samples, the training epochs, or the accuracy of the learned TPM. Adding these details would make the machine-learning validation more convincing.","section":"Section III.C"}],"recommendation":"major_revision","confidential_remarks":"The formal theory is sound, but the empirical validation is thinner than the claims. The major issues are the missing experimental specifications (Boolean-to-TPM conversion, CA partition) and the weak single-correlation validation of the semantic interpretation. If the authors can provide the omitted methodological details and a more convincing validation (or explicitly temper the claims), the paper could become suitable for publication. I would also encourage the authors to make code and data available, as the current manuscript offers no way to reproduce the reported numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives a clean, state-independent measure of system-environment synergy defined directly on the TPM: Syn = EI(X,E->X') - EI(X->X') - EI(E->X'), with both sources forced uniform. The formal result is real. Given the axioms in Appendix A, Lemma 1 is correct: if X and E are independent, redundancy vanishes, and the definition collapses to the PID synergy of the intervened variables. Theorem 1 is not an empty restatement—it makes precise that under the uniform do-intervention, the quantity really is an information atom. The Exp/Int decomposition also gives a useful way to talk about two routes to synergy. Credit where due: the CA experiments are clean, the analytic relation between introversion and the Langton parameter is a nice touch, and the machine-learning demo, while basic, at least shows the quantity can be computed from learned TPMs.\n\nThe soft spots are where the paper asks you to believe the semantics. The central claim is that Syn measured under a uniform do-intervention tracks a system's real ability to flexibly respond to environmental changes. The only quantitative support is the GRN experiment in Appendix D: a Pearson r = 0.487 between Syn and a mutual-information-based response score, on subgraphs of one apoptosis model, with no code, no out-of-sample check, and no variation of the environmental distribution. That correlation is modest, in-sample, and could easily reflect a confound such as state-space diversity or motif complexity. The paper does not test whether systems with high Syn actually adapt better under non-uniform environments. If the uniform intervention is the wrong idealization, the measure may rank systems incorrectly even though Theorem 1 stands. So the formal core is fine, but the semantic validation is not.\n\nTwo smaller issues: Corollary 1 as stated in the main text is inconsistent with the appendix derivation. The appendix formula has Syn = Exp + Int - 2 log|Omega_X|; the main text just says Syn = Exp + Int. That needs fixing. Also, the GRN experiments under-specify the Boolean-to-TPM conversion; without that, the results are not reproducible. The claim in the abstract that the indicator satisfies the PID axiomatic system is too broad: the proof only works under the independent-source intervention, not for arbitrary joint distributions. The authors should hedge that.\n\nWho is this for? Complex-systems people who work on EI and PID, and who want a mechanism-level synergy measure for discrete dynamics. The paper deserves a serious referee—the formal part is worth engaging with—but it needs a major revision, not just acceptance. The semantic claim should be tested properly or dropped to a conjecture.\n\nRecommendation: send to peer review, but the editor should ask for the code/data, the Corollary fix, and either a direct causal validation or a re-framed claim.","headline":"A sound formal identity (flexibility = PID synergy under uniform intervention) in search of a valid semantic interpretation; the GRN validation is too weak to carry the load.","tokens_in":14334,"tokens_out":2037,"would_cite":false,"duration_ms":20057,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","37B15","92C42"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new indicator, flexibility, is the synergistic information between system and environment and tracks adaptive responsiveness across cellular automata, gene regulatory networks, and learned dynamics.","keywords":["synergy","flexibility","effective information","partial information decomposition","gene regulatory networks","cellular automata","random Boolean networks","causal mechanisms"],"falsifier":"Find one Boolean network whose flexibility is high but whose response to a diverse set of environmental shocks is poor, and one with low flexibility that tracks the environment well; if such a pair exists, flexibility is not a reliable indicator of flexible response. A more quantitative version: on the paper's apoptosis gene regulatory network, compute flexibility and response diversity for all three-node subgraphs and test whether the reported positive correlation (r approximately 0.487) holds; a near-zero or negative correlation would refute the claim.","tokens_in":13364,"feed_emoji":"🧬","tokens_out":8192,"duration_ms":76151,"temperature":0.7,"pith_summary":"The paper proposes 'flexibility' as a causal-mechanism-level indicator of how well a system can respond to a changing environment. Flexibility is defined as the difference between the effective information from the joint system-environment state to the next system state and the sum of the two separate effective informations, and the authors prove that under a uniform intervention this difference is exactly the synergistic information of system and environment in the partial information decomposition framework. They then decompose flexibility into expansiveness and introversion, showing that it is controlled by the degree of system-environment entanglement and by noise. Experiments on cellular automata, random Boolean networks, and curated gene regulatory networks support the interpretation: complex cellular automata and feedback loops score high, and the measure can be estimated from learned dynamics.","feed_headline":"Formula singles out flexible responders to environmental change","feed_subtitle":"Proven equivalent to synergistic information, it highlights feedback loops in gene networks.","key_machinery":"The engine of the argument is effective information (EI), a function of a transition probability matrix defined as mutual information between input and output when the input is forced to a uniform distribution. The paper applies EI to three TPMs: the joint system-and-environment mechanism, the individual mechanism averaged over environments, and the external driving mechanism averaged over system states, and defines flexibility as their difference. The proof mechanism is the cited PID axiom system together with the lemma that redundancy vanishes for independent sources: the uniform intervention makes $\\tilde{X}^t$ and $\\tilde{E}^t$ independent, so the joint-minus-unique subtraction isolates synergy. A second object, the EI decomposition into determinism and degeneracy, yields the splitting of flexibility into expansiveness and introversion.","core_discovery":"On the paper's own terms, the central discovery is that a three-variable partial information decomposition can be made computable through effective information. For a Markovian system $X$ and environment $E$ influencing the next state $X^{t+1}$, the quantity $\\mathrm{Syn}(P_{X,E\\to X^{t+1}}) = EI(P_{X,E\\to X^{t+1}}) - EI(P_{X\\to X^{t+1}}) - EI(P_{E\\to X^{t+1}})$, computed after $do(X^t, E^t \\sim U(\\Omega_{X,E}))$, is the synergistic information of $\\tilde{X}^t$ and $\\tilde{E}^t$ with respect to $\\tilde{X}^{t+1}$. The proof uses the independence created by the uniform intervention to make redundant information vanish, leaving synergy as the excess of joint over separate effective information. The authors call this excess flexibility and characterize it as the system's capacity to respond flexibly to environmental changes.","pith_inferences":["Editorial extension: if flexibility is treated as a design target, its decomposition into expansiveness and introversion suggests two routes to higher responsiveness — diversify state transitions across environments or consolidate internal order — and either route could be tuned by adjusting coupling weights or noise.","Editorial extension: because the measure is built on a uniform intervention, it may disagree with responsiveness under strongly non-uniform environmental statistics; comparing flexibility against response diversity in environments with realistic state distributions would test how much this matters.","Editorial extension: the same joint-minus-separate EI subtraction could be applied to systems with more than two source variables, but the proof via vanishing redundancy would need new axioms or a new route once the intervening variables are not jointly independent in the required way."],"forward_implications":["Flexibility can be computed directly from the transition probability matrix, giving a state-distribution-free and initial-condition-free comparison of adaptive capacity across systems.","In cellular automata, flexibility increases from Class I to Class IV and peaks in the Langton-parameter range associated with complex dynamics, while introversion is a fixed function of the Langton parameter in noise-free rules.","Across expert-curated Boolean gene regulatory networks, feedback-loop subgraphs carry the highest flexibility, tying the measure to biological functions such as dynamical compensation and oscillators.","In random Boolean networks, flexibility is maximized at balanced self-versus-environment coupling and at a moderate noise level where expansiveness gains outweigh introversion losses.","When dynamics are unknown, a neural network trained on observed transitions can recover the TPM closely enough for flexibility to match the ground-truth curve."],"supporting_citations":[{"why":"Defines effective information as a causal-mechanism-level quantity on transition probability matrices, the basis of the flexibility formula.","marker":"[11]"},{"why":"Introduces partial information decomposition and the redundant-unique-synergy taxonomy used to interpret flexibility.","marker":"[19]"},{"why":"Supplies the axioms that the proof uses to identify the flexibility difference with synergistic information.","marker":"[21]"},{"why":"Establishes the PID axiom system, including the lemma that redundant information vanishes for independent sources.","marker":"[24]"},{"why":"Provides the do-operator intervention used to define uniform input distributions and isolate mechanism-level quantities.","marker":"[18]"},{"why":"Frames the system-environment information dynamics and the causal diagram that motivates the three mechanisms.","marker":"[10]"},{"why":"Supplies the curated Boolean gene regulatory network models used in the feedback-loop and response experiments.","marker":"[26]"},{"why":"Defines the Langton parameter used to relate flexibility to cellular automaton complexity.","marker":"[2]"}],"fun_headline_variants":["Synergy metric quantifies flexible response to environment","Effective info decomposition links flexibility to synergy","Feedback loops score high in new flexibility measure","Gene networks' flexible responses tied to synergistic info","New indicator measures a system's dynamic flexibility"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that forcing the system and environment to start uniformly across all possible states reveals the same responsiveness the system would show under real, non-uniform conditions.","fun_headline_variants_meta":{"raw":{"variants":["Synergy metric quantifies flexible response to environment","Effective info decomposition links flexibility to synergy","Feedback loops score high in new flexibility measure","Gene networks' flexible responses tied to synergistic info","New indicator measures a system's dynamic flexibility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0004,"raw_usage":{"total_tokens":2096,"prompt_tokens":957,"completion_tokens":1139,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1072}},"tokens_in":573,"tokens_out":1139,"duration_ms":11009,"temperature":1.0,"reasoning_tokens":1072,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:29:29.221344+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find one Boolean network whose flexibility is high but whose response to a diverse set of environmental shocks is poor, and one with low flexibility that tracks the environment well; if such a pair exists, flexibility is not a reliable indicator of flexible response. A more quantitative version: on the paper's apoptosis gene regulatory network, compute flexibility and response diversity for all three-node subgraphs and test whether the reported positive correlation (r approximately 0.487) holds; a near-zero or negative correlation would refute the claim.","supporting_citations":[{"cited_title":"Rosas, P","cited_arxiv_id":null,"evidence_quote":"Defines effective information as a causal-mechanism-level quantity on transition probability matrices, the basis of the flexibility formula."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces partial information decomposition and the redundant-unique-synergy taxonomy used to interpret flexibility."},{"cited_title":"Bertschinger, J","cited_arxiv_id":null,"evidence_quote":"Establishes the PID axiom system, including the lemma that redundant information vanishes for independent sources."},{"cited_title":"Dynamical Reversibility and A New Theory of Causal Emergence based on SVD","cited_arxiv_id":"2402.15054","evidence_quote":"Provides the do-operator intervention used to define uniform input distributions and isolate mechanism-level quantities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Frames the system-environment information dynamics and the causal diagram that motivates the three mechanisms."},{"cited_title":"Albantakis, L","cited_arxiv_id":null,"evidence_quote":"Supplies the curated Boolean gene regulatory network models used in the feedback-loop and response experiments."},{"cited_title":"Proof 3 According to the definition of EI, EI (PX t→X t+1) = I( ˜X t; ˜X t+1), (C11) 22 while still satisfying do(X t, Et ∼ U(ΩX,E))","cited_arxiv_id":null,"evidence_quote":"Defines the Langton parameter used to relate flexibility to cellular automaton complexity."}],"review_version":1}