{"id":"f837c8a8-55b7-4e89-a6b4-a2ad49463a86","arxiv_id":"2508.00150","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"For two-state environments and two-state phenotypes, a normalized fitness measure and mutual information are shown to be monotonically equivalent when the marginal distributions are fixed, with diminishing returns per bit.","lead":"This paper derives an exact relationship between mutual information and a normalized fitness measure in two-state populations. With environment and phenotype frequencies held fixed, more information always means more fitness, but the gain per bit shrinks as information grows.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fixed-marginal equivalence is a 2x2-table identity; the paper does not show that general population dynamics can realize it, so the biological claim is narrower than the abstract suggests.","rationale":"The central theorem is mathematically correct: for fixed marginals, Γ is linear in Px1y1, and I is strictly increasing over the paper's stated interval, so the one-to-one relation and the growth-rate independence are not in doubt. The reader's weakest assumption correctly identified that the distribution-based reduction is what makes the equivalence clean, and I agree that the persister example shows the framework's limits when marginals move. My concern is complementary and more operational: even within the reduced-distribution setting, the paper does not demonstrate that the 'fixed marginal distributions' condition can be achieved by varying rates in the general model. This matters because the abstract advertises strategies that adjust proliferation and switching rates to improve information while keeping marginals fixed; only the symmetric model actually enforces that condition. The dormancy example shows the danger of dropping it. This is a scope gap, not an internal inconsistency, so it does not overturn the conditional acceptance; it strengthens the reason for conditioning. The proposed test would settle whether any general strategy can realize the fixed-marginal regime, which is the load-bearing bridge from the mathematical identity to the biological claim.","tokens_in":19427,"tokens_out":16181,"duration_ms":167971,"concrete_test":"In the general asymmetric model of Eq. (1), fix Px1 = 0.3 through the environmental rates and numerically solve for a rate set (growth rates and switching rates ka, kb) that yields Py1 = 0.7 at stationarity. Then vary one switching rate, say kb, over a range while co-varying another rate to hold Py1 constant, and record whether Px1y1 (hence I and Γ) moves over a nontrivial interval with Py1 staying within numerical tolerance. Repeat with a small optimizer to search for continuous paths in rate space with Py1 exactly fixed. If no such path exists outside the symmetric locus, the fixed-marginal regime is not a realizable strategy space, and the central claim lacks dynamical content in the general case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V (Eqs. 16-18) proves that for fixed marginals Px1 and Py1, both Γ = (Px1y1 − Px1Py1)/(Px1(1−Px1)) and I(Px1y1) strictly increase with Px1y1 on the admissible interval Px1Py1 ≤ Px1y1 ≤ min(Px1, Py1). I verified this: I'(p) = log2[p(1−Px1−Py1+p)/((Px1−p)(Py1−p))] is positive for p > Px1Py1, so the algebra is sound. The load-bearing step for the paper's biological message is the implicit assertion that the rate-changing 'strategies' discussed in the abstract can sweep Px1y1 while keeping both marginals fixed. No such mechanism is exhibited outside the symmetric model, where Px1 = Py1 = 0.5 is fixed by symmetry. In the general asymmetric model, every growth and switching rate enters Eq. (1) and generically moves Py1 and Px1y1 together; the dormancy example (Fig. 6E) demonstrates that when Py1 is free, I and Γ decouple and even move oppositely. Thus the headline 'an increase in information implies an increase in population fitness' is established only as a conditional property of a reduced 2x2 distribution, not as a property of any demonstrated control strategy except the symmetric special case.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes a two-state population model in a two-state randomly switching environment. It defines time-averaged joint probabilities for environment and phenotype, the mutual information I, and a normalized fitness Gamma defined as the growth rate relative to the independent case, normalized by the maximum possible relative growth rate. With fixed marginal probabilities Px1 and Py1, the paper shows that Gamma and I are both strictly increasing functions of the joint probability Px1y1, and it derives a growth-rate-independent expression for Gamma, the maximum achievable Gamma and I for fixed marginals, and a marginal information value dGamma/dI. The formalism is illustrated with a symmetric bet-hedging model and a dormant-proliferating phenotype model. The central algebraic derivations are sound, but the biological scope of the claimed information-fitness equivalence is narrower than the abstract suggests.","tokens_in":43,"tokens_out":15314,"duration_ms":273130,"significance":"As a mathematical characterization of a two-by-two probability table, the paper is correct and provides closed-form results: Eq. (18) shows that Gamma is independent of the growth rates, and Eq. (21) correctly evaluates the marginal information value, which I verified to be positive and divergent at I=0. The paper also ships reproducible simulation code and includes a detailed algorithmic description of its numerics. The biological significance, however, is limited by the fact that the equivalence is a property of the 2x2 parametrization rather than a demonstrated property of the population dynamics: changing biological rates generically changes the marginals as well as Px1y1, and the dormant-proliferating example in Fig. 6E shows that Gamma and I can move in opposite directions when marginals are free. The contribution is therefore a moderate unification of information and normalized fitness for two-state bet-hedging, not a general proof that information acquisition increases population fitness.","major_comments":[{"comment":"The headline claim 'with fixed marginal distributions, an increase in information implies an increase in population fitness' is a theorem about a 2x2 probability table, not about the dynamics of Eqs. (1)-(2). For fixed Px1 and Py1, both Gamma and I are strictly increasing functions of Px1y1, and the algebra is correct. However, the paper does not exhibit any control strategy in the general asymmetric model that varies Px1y1 while keeping both Px1 and Py1 fixed. The only mechanism explicitly shown is the symmetric model, where Py1 = Px1 = 1/2 is fixed by symmetry. In the general model, changing growth or switching rates generically moves Py1 and Px1y1 together, as the dormancy example in Fig. 6E demonstrates: there I and Gamma are non-monotonically related precisely because the marginals are not fixed. Since the abstract promises 'strategies' with fixed marginal distributions, the authors should either construct such a mechanism or explicitly restrict the biological claim to the conditional statement and to the symmetric special case.","section":"Section V, Eqs. (16)-(18), and abstract"},{"comment":"The normalized fitness Gamma is defined as (⟨γ⟩−⟨γ⟩ind)/(⟨γ⟩max−⟨γ⟩ind), which is a normalized linear function of Px1y1−Px1Py1. For fixed marginals, every such function is monotone in Px1y1, and the mutual information I is also monotone in Px1y1 on the admissible interval of Eq. (16). Thus the 'equivalence' between information and fitness is largely installed by the choice of Gamma rather than derived from the population dynamics. I verified Eq. (18) and Eq. (21), but the paper's language—'we propose that information and fitness can be treated as equivalent measures' and 'we prove... any increase in population fitness through an arbitrary strategy corresponds to a simultaneous increase in information'—overstates the novelty. The authors should state explicitly that the equivalence is a property of the 2x2 parametrization, not a dynamical law.","section":"Section IV, Eq. (13)"}],"minor_comments":[{"comment":"In the x=x2 branch of Eq. (2), the selection term should be (gx2_y1 − gx2_y2) fy1 (1 − fy1), not (gx1_y2 − gx2_y2) fy1 (1 − fy1); as written, the expression mixes growth rates from the two different environments.","section":"Section II.B, Eq. (2)"},{"comment":"The text states 'gx1_y2 = gx2_y2 = 0, gx1_y1 = −µ1 and gx2_y1 = µ2', which is inconsistent with Table V and with Algorithm 1; those sources use the opposite assignment, with y1 dormant (zero growth) and y2 proliferating (−µ1 in stress, µ2 in normal conditions). Please reconcile the text, the table, and the algorithm.","section":"Section II.D"},{"comment":"The second branch of the piecewise expression for Imax is incomplete: it lacks the condition 'if Py1 < Px1' and the third logarithmic term analogous to the first branch. Please supply the full expression.","section":"Section V, Eq. (20)"},{"comment":"The sentence 'This maximum occurs when Py1 ≈ 0.5' is inconsistent with the Fig. 6C caption ('reaching a maximal value when Py1 = Px1') and with Eq. (19), which gives a global maximum at Py1 = Px1. Please correct the text.","section":"Section V.B"},{"comment":"There are several typographical errors, including 'bet-heging' in the Introduction, 'a a proliferation rate' in Section II.A, and an index mismatch in Eq. (5) where Pyi is defined with fyj. A careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core of the paper is sound and the code availability is a strength, but the novelty is modest because the central equivalence is essentially a 2x2-table identity built into the definition of Gamma. The main risk is overclaiming the biological generality: the abstract's 'strategies with fixed marginal distributions' are not demonstrated outside the symmetric model. This is fixable by reframing and by adding a constructive example or explicit caveats, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this paper expecting another inflated information-fitness claim; it's actually a careful, narrowly scoped piece. The genuinely new results are Eqs. 18-21: a growth-rate-independent normalized fitness Gamma, its parametrization by the joint probability Px1y1 given fixed marginals, the maximum curves, and the marginal information value with diminishing returns. I checked the algebra; Eq. 18 follows cleanly from the 2x2 table, Eq. 19 matches Frechet bounds, and Eq. 21 is positive and diverges as I goes to zero. The symmetric case reproduces earlier results (Rivoire-Leibler, Taylor et al.), which the paper acknowledges.\n\nThe soft spots are real but mostly disclosed. The stress-test note is right that the fixed-marginal equivalence is a 2x2-table identity, not a demonstrated property of any general rate-changing strategy. But the paper explicitly says \"with fixed marginal distributions\" in the abstract and shows in Fig. 6E that when Py1 drifts, Gamma and I decouple and move oppositely. So the limitation is stated, and the reach is honestly bounded. A reader who wants a mechanistic \"how to hold marginals fixed\" won't find it outside the symmetric model; that's a real gap but a minor-to-moderate one for a theory paper. Presentational issues: the sentence before Eq. 19 contradicts the formula, Eq. 15's set braces should be an interval, and the \"never reaches zero\" claim is literally wrong at the boundary (the limit is zero). These need fixing but don't touch the math. The code is on Zenodo, the derivations are in the SI, and the simulations are reproducible.\n\nWho is this for? People working on bet-hedging and information-theoretic fitness, and anyone who wants a concrete design target for two-state phenotypic switching. It deserves a serious referee; the central derivation is correct and the paper is honest about its own scope. I'd send it to review.","headline":"A sound, honestly scoped extension of information-fitness equivalence to a population-level two-state model; the fixed-marginals caveat is in the paper itself, so referee it.","tokens_in":20320,"tokens_out":1359,"would_cite":true,"duration_ms":13609,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92D25","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"With phenotype and environment marginals fixed, mutual information and normalized population fitness are one-to-one increasing functions of the phenotype-environment correlation, independent of the growth rates.","keywords":["bet-hedging","mutual information","phenotypic switching","fluctuating environment","normalized fitness","two-state phenotype","bacterial persistence","information value"],"falsifier":"Run the dormant-proliferating model while actively clamping $P_{y_1}$ constant as $\\mu_1$ varies; if $\\Gamma$ and $I$ still separate, with one falling while the other rises, the fixed-marginal claim fails in exactly the regime intended. Alternatively, scan the full parameter space of the general model and search for any pair of parameter sets with identical $P_{x_1}$ and $P_{y_1}$ but with a larger $P_{x_1y_1}$ and a smaller $\\Gamma$; none should exist if Eq. (18) is correct.","tokens_in":19087,"feed_emoji":"🧬","tokens_out":8348,"duration_ms":79032,"temperature":0.7,"pith_summary":"The paper asks when population-level bet-hedging pays in a fluctuating two-state environment, and proposes that the payoff is fully captured by the correlation between phenotype and environment. For a population of self-replicating individuals with two phenotypes, it defines a normalized fitness $\\Gamma$ as the time-averaged growth rate relative to the hypothetical independent population, divided by the maximum possible relative gain. The main result is that, when the marginal frequencies of the two environments and the two phenotypes are held fixed, $\\Gamma$ and the mutual information $I$ between environment and phenotype are one-to-one increasing functions of the same quantity $P_{x_1y_1}$, the probability that the fittest phenotype appears in its matching environment. The relation does not depend on the numerical values of the growth or switching rates. If true, it means that in this whole class of models the fitness benefit of bet-hedging is governed by a single correlation measure, and that extra information about the environment is worth progressively less as the population already has more of it.","feed_headline":"More phenotype-environment information means higher normalized fitness","feed_subtitle":"With phenotype and environment frequencies fixed, mutual information and relative growth rise together, regardless of growth rates.","key_machinery":"The load-bearing object is the stationary joint probability $P_{x_i y_j}$, obtained from the time-averaged phenotype fraction $\\langle f_{y_j}\\rangle$ and the environmental indicator function. This turns the dynamical system into a time-independent information channel whose input is the environment and whose output is the phenotype. The key identity is Eq. (18): with fixed $P_{x_1}$ and $P_{y_1}$, $\\Gamma = (P_{x_1y_1} - P_{x_1}P_{y_1})/(P_{x_1}(1-P_{x_1}))$, which is independent of the growth rates. The mutual information is likewise parameterized by the single correlation $P_{x_1y_1}$ through the four entries of the joint distribution. The argument that carries the paper is that both $\\Gamma$ and $I$ are monotone functions of this one parameter, so a one-to-one curve between them exists for each fixed pair of marginals; the normalization by the maximum relative growth rate is what removes the growth rates from the formula.","core_discovery":"On the paper's own terms, the central discovery is a parameter-free equivalence between information and fitness. Starting from the deterministic fraction dynamics for two phenotypes in a two-state telegraph environment, the authors construct the stationary joint distribution $P_{x_i y_j}$ by time-averaging the phenotype fraction inside each environmental state, and define the normalized fitness $\\Gamma = (\\langle\\gamma\\rangle - \\langle\\gamma\\rangle_{\\mathrm{ind}})/(\\langle\\gamma\\rangle_{\\max} - \\langle\\gamma\\rangle_{\\mathrm{ind}})$. For fixed marginals $P_{x_1}$ and $P_{y_1}$, this reduces to $\\Gamma = (P_{x_1y_1} - P_{x_1}P_{y_1})/(P_{x_1}(1-P_{x_1}))$, and the mutual information $I=\\sum P_{x_i y_j}\\log_2(P_{x_i y_j}/(P_{x_i}P_{y_j}))$ also depends on $P_{x_1y_1}$ alone; consequently both quantities increase together and each one determines the other. Because the growth rates cancel out of $\\Gamma$, the relation is universal across all parameter choices. The maximum achievable $\\Gamma$ and $I$ occur when the phenotype marginal equals the environmental marginal, $P_{y_1}=P_{x_1}$, and the maximum information is then the environmental entropy. The paper stresses that this master-curve behavior holds only while the marginals are fixed; when the marginals vary, as they do in a dormant-proliferating (persister) population under increasing antibiotic stress, $I$ can rise while $\\Gamma$ peaks and falls, so the two are not equivalent as raw numbers in that broader setting.","pith_inferences":["Because the mutual information here is computed from the stationary joint distribution rather than from trajectories, the paper's equivalence does not by itself capture how quickly the phenotype population tracks the environment; a natural extension would be to define a rate-dependent mutual information and test whether the one-to-one relation survives.","A testable prediction follows: in a synthetic two-state bet-hedging system with the marginal phenotype fraction clamped by a controller, any perturbation that increases phenotype-environment correlation should increase the normalized growth rate, even if the absolute growth rate falls, paralleling persister cells under stress.","The persister example suggests that the quantity evolutionary biology should equate with information is $\\Gamma$, the relative advantage over independence, not the absolute long-term growth rate; if absolute growth is the fitness that matters, the paper's 'information implies fitness' statement applies to a different objective.","The decreasing marginal value curve suggests an optimization problem: if acquiring and storing information has a per-bit cost, an organism's optimal information level would sit below the achievable maximum, which the paper notes but does not model."],"forward_implications":["In the symmetric bet-hedging model with equally likely environments, all growth and switching parameters collapse onto a single master curve $\\Gamma_{\\mathrm{sym}}=4P_{x_1y_1}-1$ against $I_{\\mathrm{sym}}$, so measuring one quantity determines the other.","Any strategy that raises the phenotype-environment correlation, such as responsive switching that favors the fittest phenotype or a slower environment that lets the fittest phenotype dominate, raises both mutual information and normalized fitness for fixed marginals.","The marginal value of information, $\\partial\\Gamma/\\partial I$, is always positive but decreases with $I$; near zero information it diverges, meaning the first bits of correlation yield the largest normalized-fitness gain per bit.","To maximize both information and normalized fitness, a population should match its phenotype marginal to the environmental marginal ($P_{y_1}=P_{x_1}$), reaching an information level equal to the environmental entropy.","When the marginals are free to vary, as in the dormant-proliferating model under rising antibiotic death rate, normalized fitness and mutual information can move in opposite directions, so the equivalence is specifically a fixed-marginal property rather than a universal proportionality."],"supporting_citations":[{"why":"Supplies the symmetric bet-hedging model whose growth and switching matrices are used for the master-curve example.","marker":"[9]"},{"why":"Provides the time-averaged fraction definition used to form the joint phenotype-environment probabilities.","marker":"[47]"},{"why":"Established information-fitness equivalence in simpler stochastic models that this paper generalizes.","marker":"[43]"},{"why":"Offers the earlier argument that mutual information and growth rate can be equivalent, used as a comparative baseline.","marker":"[44]"},{"why":"Introduces the cost-of-information perspective that motivates measuring fitness relative to the independence baseline.","marker":"[39]"},{"why":"Gives the single-individual limit for fast environmental fluctuations, used to delimit when the population-level view applies.","marker":"[42]"},{"why":"Supplies the stationary probability formula for a two-state telegraph process used for $P_{x_i}$.","marker":"[57]"},{"why":"Supports the deterministic fraction dynamics equation with a recent modeling treatment.","marker":"[48]"}],"fun_headline_variants":["Fixed marginals make information and fitness equivalent","Info and fitness rise together when marginals stay fixed","Normalized fitness and mutual information are interchangeable","No growth-rate dependence: info and fitness share one curve","Phenotype info and fitness move together under fixed marginals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The equivalence rests on defining fitness as the normalized, time-averaged growth advantage over the independence baseline, and on keeping the marginal phenotype and environment frequencies fixed; if fitness means the actual long-term growth rate, or if the marginals move, the one-to-one relation can fail.","fun_headline_variants_meta":{"raw":{"variants":["Fixed marginals make information and fitness equivalent","Info and fitness rise together when marginals stay fixed","Normalized fitness and mutual information are interchangeable","No growth-rate dependence: info and fitness share one curve","Phenotype info and fitness move together under fixed marginals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000771,"raw_usage":{"total_tokens":3506,"prompt_tokens":1126,"completion_tokens":2380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":742,"completion_tokens_details":{"reasoning_tokens":2305}},"tokens_in":742,"tokens_out":2380,"duration_ms":17610,"temperature":1.0,"reasoning_tokens":2305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:24:48.858943+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the dormant-proliferating model while actively clamping $P_{y_1}$ constant as $\\mu_1$ varies; if $\\Gamma$ and $I$ still separate, with one falling while the other rises, the fixed-marginal claim fails in exactly the regime intended. Alternatively, scan the full parameter space of the general model and search for any pair of parameter sets with identical $P_{x_1}$ and $P_{y_1}$ but with a larger $P_{x_1y_1}$ and a smaller $\\Gamma$; none should exist if Eq. (18) is correct.","supporting_citations":[{"cited_title":"Thattai and A","cited_arxiv_id":null,"evidence_quote":"Supplies the symmetric bet-hedging model whose growth and switching matrices are used for the master-curve example."},{"cited_title":"Skanata and E","cited_arxiv_id":null,"evidence_quote":"Provides the time-averaged fraction definition used to form the joint phenotype-environment probabilities."},{"cited_title":"Rivoire and S","cited_arxiv_id":null,"evidence_quote":"Established information-fitness equivalence in simpler stochastic models that this paper generalizes."},{"cited_title":"Information and fitness","cited_arxiv_id":"0712.4382","evidence_quote":"Offers the earlier argument that mutual information and growth rate can be equivalent, used as a comparative baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the cost-of-information perspective that motivates measuring fitness relative to the independence baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the single-individual limit for fast environmental fluctuations, used to delimit when the population-level view applies."},{"cited_title":"Gardiner, Stochastic Methods, Vol","cited_arxiv_id":null,"evidence_quote":"Supplies the stationary probability formula for a two-state telegraph process used for $P_{x_i}$."},{"cited_title":"Extinction in agent-based and collective models of bet-hedging","cited_arxiv_id":"2406.11482","evidence_quote":"Supports the deterministic fraction dynamics equation with a recent modeling treatment."}],"review_version":1}