REVIEW 1 major objections 5 minor 24 references
Inferring Cooperativity From Pooled Measurements
T0 review · 1 major / 5 minor · reviewed 2026-07-12 · grok-4.5
Pith's one-line read Latent interactions among Markov components can be identified and tested from pooled measurements alone.
desk verdict Solid continuous-time fix for an open identifiability problem in pooled multi-channel Markov systems; theorems hold and the package is usable. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Sum-dependent Markov chains (SDMCs): continuous-time finite-state processes whose transition rates for each coordinate depend only on the current sum; their sum process is a birth-death chain whose rate matrix recovers every individual rate, enabling a cooperativity index and a stepdown test.
What would settle it
If two distinct SDMC rate vectors produce statistically indistinguishable discrete-time transition matrices of the sum process (or if a known independent multi-channel recording is declared cooperative by the stepdown test at the claimed asymptotic level), the identifiability and testing claims fail.
Extended reading notes
Core claim
Under permutation invariance and the sparse-transition property, the entire parameter vector of a binary-state sum-dependent Markov chain is uniquely determined by the birth-death rates of its sum process; the same parameters remain identifiable, and therefore estimable and testable, from discrete noisy observations of that sum via a uniquely embeddable hidden Markov model.
Load-bearing premise
The model assumes no simultaneous multi-coordinate jumps on the infinitesimal time scale and that the number of latent units is known or correctly selected.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces sum-dependent Markov chains (SDMCs), continuous-time multivariate Markov processes on {0,1}^L whose rates depend on the current sum of coordinates. Under permutation invariance and sparse transitions (Assumptions 1–2), Theorem 2.5 shows that the full rate vector (λ0…λL−1, μ1…μL) is uniquely recoverable from the birth–death rate matrix of the sum process. A cooperativity index Λ(θ) ∈ [−1,1] is defined from pairwise orderings of these rates, with consistent plug-in and thresholded estimators (Theorem 3.5). For discrete noisy observations of the sum, the authors embed the continuous-time rates into an HMM, prove unique embedding via reversibility (Lemma 4.1), and establish consistency and asymptotic normality of the MLE (Theorems 4.2–4.3). A stepdown multiple-testing procedure (SCoT) for pairwise rate equalities is shown to control FWER asymptotically and to yield a consistent test-based estimator of Λ (Theorem 4.4, Corollary 4.5). Simulations confirm the asymptotics and compare favourably with discrete-time competitors; applications to RyR2 and gramicidin voltage-clamp data recover negative cooperativity and independence, respectively.
Significance. The work gives a clean positive answer to a practically important question: when can latent coordinate interactions be recovered from pooled continuous-time measurements? The continuous-time formulation removes the even-L non-identifiability of the closest discrete-time predecessor (Vanegas et al., 2024) and yields sampling-rate-invariant parameters. Full proofs of identifiability, embedding, MLE asymptotics and FWER control, together with a public R package, make the contribution immediately usable. The multi-state extension via L-informativity (Supplement A) further broadens the scope. If the modelling assumptions hold, the methodology supplies a statistically rigorous tool for quantifying ion-channel cooperativity and, more generally, for structure learning from aggregate Markovian signals.
major comments (1)
- Section 5.3.2: for the gramicidin data both AIC and BIC select L=10, yet the authors override this choice and set L=5 on the basis of a histogram of idealized states. Because the entire parameter vector and the cooperativity index are defined for a fixed L, and because misspecification of L invalidates the algebraic map of Theorem 2.5(iii), the paper should either (i) report the SCoT results for the AIC/BIC-selected L as a sensitivity check, or (ii) supply a formal justification (e.g., a consistent model-selection theorem under the SD-HMM) that legitimates the override. Without this, the real-data claim of “no significant cooperativity” rests on an ad-hoc modelling decision.
minor comments (5)
- Abstract and Introduction: the phrase “simulations and real-data analyses, demonstrate” contains a superfluous comma.
- Definition 4 / Remark 1: the link to Kendall’s τ is elegant; a one-line display of the two data sets D1={(s,λs)}, D2={(s,μs)} would make the identity immediate for readers unfamiliar with rank correlation.
- Figure 1 caption: the truncation of sample quantiles outside [−4,6] is noted, but the number of omitted points per panel would help the reader judge the quality of the normal approximation in the tails.
- Supplement A.2: the Diophantine characterisation of L-informativity is useful; a short numerical example for a rational alphabet with κ=3 would illustrate when the condition fails.
- Package availability is welcome; a brief statement of the computational complexity of the modified Baum–Welch step (Algorithm 2) would aid practitioners with long traces.
Circularity Check
No significant circularity: identifiability, cooperativity index, MLE asymptotics and stepdown test are derived from explicit algebraic maps and standard HMM theory, not from self-referential fits or load-bearing self-citations.
full rationale
The core derivation chain is self-contained. Definition 1 parameterizes the SDMC rate matrix Q by the sum-dependent rates λ_s, μ_s. Theorem 2.2 shows equivalence to Assumptions 1–2 (permutation invariance + sparse transitions). Theorem 2.5 then gives the sum-process rate matrix R explicitly in terms of those rates (birth rates (L-s)λ_s, death rates sμ_s) and inverts the map algebraically: λ_{i-1}=r_{i-1,i}/(L-i+1), μ_i=r_{i,i-1}/i. The cooperativity index Λ(θ) (Definition 4) is a fixed functional of the same rates (pairwise sign comparisons, equivalently Kendall τ); its plug-in and thresholded estimators (Proposition 3.4, Theorem 3.5) inherit consistency from any consistent θ̂_n and do not presuppose the value of Λ. Under the SD-HMM, unique embedding (Lemma 4.1) follows from reversibility of R (Proposition 2.6) plus the external result of Jia (2016); consistency and asymptotic normality of the MLE (Theorems 4.2–4.3) invoke the external HMM theory of Leroux (1992) and Bickel et al. (1998) under standard emission regularity (Assumptions 3–4). The stepdown test (Algorithm 1, Theorem 4.4) applies the external Romano–Wolf (2005) procedure to the asymptotic normal limit of θ̂_n. Self-citations to Vanegas et al. (2024) and Requadt et al. (2025) appear only for comparison of discrete-time analogues, data sources, and robust extensions; they are not used as premises for any uniqueness, embedding or asymptotic claim. No fitted quantity is renamed a prediction, no ansatz is smuggled via self-citation, and no definition is circular. Minor self-citation for context yields score 1 rather than 0.
Assumptions & free parameters
free parameters (5)
- SDMC rate vector θ = (λ0…λL−1, μ1…μL)
- Emission parameters ϕ (baseline b, single-channel conductance a, state-dependent σs)
- Number of channels L
- Truncation sequence an for empirical cooperativity index
- Significance level α for SCoT
assumptions (5)
- domain assumption Permutation invariance of coordinates (Assumption 1)
- domain assumption Conditional independence at infinitesimal times / sparse transitions (Assumption 2)
- standard math Emission densities satisfy identifiability, continuity and moment conditions (Assumptions 3–4)
- ad hoc to paper L-informativity of the alphabet (Definition 5) for multi-state extension
- standard math Reversibility of the sum process under positive rates
invented entities (3)
-
Sum-dependent Markov chain (SDMC)
-
Cooperativity index Λ(θ)
-
Stepdown Cooperativity Test (SCoT)
Cite this review
Pith. "Pith review of Inferring Cooperativity From Pooled Measurements." pith.science (2026). https://pith.science/paper/X5CJL2RY
@misc{pith2026260703088,
author = {Pith},
title = {Pith review of: Inferring Cooperativity From Pooled Measurements},
year = {2026},
howpublished = {\url{https://pith.science/paper/X5CJL2RY}},
note = {Machine review of arXiv:2607.03088}
}
read the original abstract
In many modern experiments, latent interactions drive multicomponent stochastic systems, yet the data are available only as pooled measurements that obscure these dependencies. Whether such interactions can be identified and inferred from aggregate signals remains largely unexplored. Motivated by multi-channel electrophysiological recordings, we address this problem by introducing sum-dependent Markov chains, a class of finite-state continuous-time multivariate Markov processes whose transition rates encode interactions through the aggregate state. Under natural structural conditions, we establish identifiability of the latent dynamic parameters from the aggregate process. We define a cooperativity index that distinguishes positive cooperativity, negative cooperativity and independence, and construct its consistent estimators. For discretely and noisily observed pooled data, we develop likelihood-based inference through a hidden Markov model, address the associated embedding problem, and prove consistency and asymptotic normality. We further propose a stepdown test for cooperativity with asymptotic size control and power guarantees. Simulations and real-data analyses, demonstrate the scope and effectiveness of the methodology.
Figures
Figures from the paper (7 more)
Reference graph
Works this paper leans on
-
[1]
Ball, F., Milne, R. K., Tame, I. D., and Yeo, G. F. (1997). Superposition of interacting aggregated continuous-time Markov chains.Adv. Appl. Probab., 29(1):56–91. 19 0 1 2 3 4 0 1 2 3 Current [a.u.] Time [s] −0.5 0.5 1.0 1.5 2.0 Time [ms] Current [a.u.] 0 200 400 600 800 Figure 6: Dataset 1 of RyR2 channels: current trace, consisting of20000observations, ...
arXiv 1997
-
[2]
Moreover, the model parameters of theκ-state SDMC satisfy θj,k;s = lim δ↘0 P(Xi,t+δ=c k|Xi,t =c j,SXt =s) δ which is independent of bothi∈{1,...,L}and t≥0
in exactly the same manner. Moreover, the model parameters of theκ-state SDMC satisfy θj,k;s = lim δ↘0 P(Xi,t+δ=c k|Xi,t =c j,SXt =s) δ which is independent of bothi∈{1,...,L}and t≥0. Consequently, Theorem 2.2 extends directly to generalκ-state SDMCs. There is also a one-to-one correspondence between the rate matrix (or equivalently the model parametersθj...
1969
-
[3]
Note that ˜z0 = x and ˜zm+n = y
of(Xt)t≥0, we have m+n∏ k=1 q˜zk−1,˜zk >0. Note that ˜z0 = x and ˜zm+n = y. Thus,( Xt)t≥0is irreducible. If both chains(Xt)t≥0and (St)t≥0are irreducible, they are positive recurrent (as the state space is finite), and thus have unique invariant distributions (see e.g. Norris, 1998, Theorem 3.5.2). Letπ= (πx)x∈ALκbe the invariant distribution of(Xt)t≥0. Th...
1998
-
[4]
Let1 ≤i1 <···< im≤L coordinates wherex and y differ, i.e.xj̸= yj for j∈{i1,...,im}and m =∥x−y∥0
Part iii.Consider arbitrarily two distinct statesx,y∈AL κ. Let1 ≤i1 <···< im≤L coordinates wherex and y differ, i.e.xj̸= yj for j∈{i1,...,im}and m =∥x−y∥0. We define a sequence of states as follows:z0 =xandz k = (zk,1,...,zk,L),k∈{1,...,m}, with zk,j = { yj,ifj≤i k, xj,ifj >i k. Then zm =y, and∥zk−zk−1∥0 = 1for k∈{1,...,m}. Thus,qzk−1,zk > 0, by positivit...
1969
-
[5]
We define thediscernibilityofA κover the configuration spaceMκ L by dL(Aκ) := min (m1,...,mκ)̸=(m′ 1,...,m′ κ) ∈Mκ L ⏐⏐⏐⏐⏐ κ∑ i=1 cimi− κ∑ i=1 cim′ i ⏐⏐⏐⏐⏐
reduces toL-informativity (Definition 5). We define thediscernibilityofA κover the configuration spaceMκ L by dL(Aκ) := min (m1,...,mκ)̸=(m′ 1,...,m′ κ) ∈Mκ L ⏐⏐⏐⏐⏐ κ∑ i=1 cimi− κ∑ i=1 cim′ i ⏐⏐⏐⏐⏐. The quantitydL(Aκ)measures the minimal separation between distinct configurations in the aggregated domain and provides an equivalent characterization ofL-δ-i...
1998
-
[6]
A continuous-time Markov chainXt isirreducible, if for allx,y∈X P(Xt+δ=y|Xt =x)>0 for allt≥0,δ >0. By Norris (1998, Theorem 3.2.1), a continuous-time Markov chainXt is irreducible, if and only if for anyx̸=y there exist statess0,...,sn∈X,n∈Nwith s0 =x ands n =ysuch that n−1∏ i=0 qsi,si+1(θ)>0. 33 B.1 Lumpability of Markov chains through sums The lumping p...
1998
-
[7]
C.3 Proof of Lemma 2.3 Proof
It then follows that the parameters λs,µs satisfy (2) and the limits therein are are independent oft≥0andj∈{1,...,L}. C.3 Proof of Lemma 2.3 Proof. By Norris (1998, Theorem 3.2.1), the irreducibility of(Xt)t≥0holds if and only if for any distinctx,y∈Xthere exist a sequence of statesx(0),x (1),...,x (n) with x(0) = x and x(n) = y such thatqx(0),x(1)qx(1),x...
1998
-
[8]
C.4 Proof of Proposition 2.4 Proof
if and only if all parametersλ0,...,λL−1,µ1,...,µL are strictly positive. C.4 Proof of Proposition 2.4 Proof. By Norris (1998, Theorem 3.7.3),(Xt)t≥0is reversible if and only if its rate matrixQ and some initial distributionπsatisfy detailed balance equations, i.e. πxqx,y =πyqy,x for allx,y∈AL 2.(26) To find a possibleπof such, we only need to consider th...
1998
Show all 24 references
-
[9]
By the structure ofR in (3) of Theorem 2.5ii., this is clearly the case if and only if all parameters λ0,...,λL−1,µ1,...,µL are strictly positive, which by Lemma 2.3 is equivalent to the reversibility of(Xt)t≥0. Part ii.Similar to the proof of Proposition 2.4, we employ the sp...
1998
-
[10]
We extend Lemma A.6 from Vanegas et al
P(B 1∪B2) =P(A|B1) =P(A|B2). We extend Lemma A.6 from Vanegas et al. (2024) to the continuous-time setting. Lemma C.2.Let( Xt)t≥0be a continuous-time Markov chain that satisfies Assumption
2024
-
[11]
Then, for anyi∈{1,...,L},x∈AL 2,a∈A2,t≥0andδ >0, it holds P(Xi,t+δ=a|Xt =x) =P(X i,t+δ=a|Xi,t =x i,SXt =Sx). Proof. Let F (x,i ) := {z∈ AL 2 :Sz =Sx,zi = xi}.For anyz∈F(x,i )there exists a permutation matrixP∈AL×L 2 such thatz =Px andzi =xi. By Assumption 1 (permutation invari...
1993
-
[12]
Conversely, if(28) holds, then, by(29), λS(0,x),λS(1,x),µS(0,x),µS(1,x) remain constant for x∈AL−1 2
] I(b= 0), and is thus independent ofx∈AL−1 2 , i.e., (28) is satisfied. Conversely, if(28) holds, then, by(29), λS(0,x),λS(1,x),µS(0,x),µS(1,x) remain constant for x∈AL−1 2 . Then, by the permutation invariance of(Xt)t≥0, we have λ0 =···=λL−1andµ 1 =···=µL, that is,(Xt)t≥0is ...
1993
-
[13]
= (x,y) ) = 1,(30) for all(x,y)∈E. AsΓsatisfies the lumping property with respect to{Jx : x∈{0, 1}L}and{Gy : y∈ {0,1}L}, Theorem 2.11 in Tian and Kannan (2006) implies P(Xδ=z|X0 =x) = ∑ w∈{0,1}L P ( Xδ=z,X ′ δ=w|X0 =x,X ′ 0 =y ) , andP ( X′ δ=w|X′ 0 =y ) = ∑ z∈{0,1}L P ( Xδ=z,...
2006
-
[14]
The above inequality is in fact strict, because the corresponding transition rate Γ((x,y),(x,y+e i)) =λs+1−λs>0, see Norris (1998, Theorem 3.2.1)
Then by (30) po s(δ) =P(Xi,δ= 1|X0 =x) =P ( Xi,δ= 1|X0 =x,X ′ 0 =y ) ≤P ( X′ i,δ= 1|X0 =x,X ′ 0 =y ) =P ( X′ i,δ= 1|X′ 0 =y ) =p o s+1(δ). The above inequality is in fact strict, because the corresponding transition rate Γ((x,y),(x,y+e i)) =λs+1−λs>0, see Norris (1998, Theorem...
1998
-
[15]
Then, because of (30) andΓ((x,y),(x−ei,y)) =µs−µs+1>0, we have pc s(δ) =P(Xi,δ= 0|X0 =x) =P ( Xi,δ= 0|X0 =x,X ′ 0 =y ) >P ( X′ i,δ= 0|X0 =x,X ′ 0 =y ) =P ( X′ i,δ= 0|X′ 0 =y ) =p c s+1(δ), which concludes the proof for the fully positively cooperative case. ii. We next conside...
1993
-
[16]
ii.Caseθ̸= 0.Define the events An := { |ˆθn|≥an,sign( ˆθn) = sign(θ),and|ˆθn−θ|≤an }
Since ˆθn−θ=oP(an), we have P ( |ˆθn|≤an ) →1, which impliesP ( tsfan(ˆθn) = 0 ) →1.Hence,tsfan(ˆθn) P− →0 = sign(θ). ii.Caseθ̸= 0.Define the events An := { |ˆθn|≥an,sign( ˆθn) = sign(θ),and|ˆθn−θ|≤an } . OnA n, we obtain ⏐⏐tsfan(ˆθn)−sign(θ) ⏐⏐ = ⏐⏐|ˆθn|−|θ|−an ⏐⏐≤an→0. Moreo...
2016
-
[17]
Thus, by Leroux (1992, Theorem
1992
-
[18]
E.3 Proof of Theorem 4.3 Proof
and Proposition E.1, the assertion holds. E.3 Proof of Theorem 4.3 Proof. We invoke the general result of Bickel et al. (1998) and verify conditions (A1)–(A6) therein. The assumed stationarity of the latent sum process(St)t≥0, together with its irreducibility (see Lemma 2.3 an...
1998
-
[19]
We next show thatF is strictly increasing on[0,∞)
The continuity ofFimplies thatF(c α(R)) = 1−α. We next show thatF is strictly increasing on[0,∞). There exists a matrixB∈Rd×r with full column rankr = rank(R)such that R =BB⊤.Then Z D=BX with X a standard r-dimensional normal random vector, andMD=N(X)forN(x) = max 1≤j≤d|(Bx)j|...
1998
-
[20]
Therefore, each(i,j )∈H1 belongs to the first-step rejection set
Indeed, at the first step the active set isK=H, and for every(i,j)∈H1, |ˆTn,(i,j)|≥Bn>E n≥ˆcn,α(H). Therefore, each(i,j )∈H1 belongs to the first-step rejection set. Since the final rejection set ˆH1,n contains all hypotheses rejected in the first step, it follows that P ( H1⊆...
2005
-
[21]
By Lemma E.3 and the continuous mapping theorem, max (i,j)∈H0 |ˆUn,(i,j)|D− →max (i,j)∈H0 |Z(i,j)|, Z∼N|H0| ( 0,R◦,H0 )
First, we verify their Assumption A1. By Lemma E.3 and the continuous mapping theorem, max (i,j)∈H0 |ˆUn,(i,j)|D− →max (i,j)∈H0 |Z(i,j)|, Z∼N|H0| ( 0,R◦,H0 ) . By Lemma E.2, the limiting distribution function is continuous and strictly increasing at cα ( R◦,H0 ) . Thus, Assump...
2005
-
[22]
As a direct consequence, 1≥P (ˆΛn,α= Λ(θ◦) ) ≥P(Wn∩Cn)≥1−P(Wc n)−P(Cc n)→1. 57 F Implementation details and additional results F.1 Maximum likelihood estimation by a modified Baum–Welch algorithm Let Y1,...,Yn denote observations generated by an SD-HMM (Model 1), and let(St)t≥...
1970
-
[23]
It may instead converge to a local maximum
is not guaranteed to converge to a global maximum. It may instead converge to a local maximum. Consequently, the choice of the initial value η(1) = (θ(1),ϕ(1))is important. In our numerical experiments and real-data applications, for sufficiently small sampling intervalsδ >0, ...
2024
-
[24]
59 In Algorithm 2, the initial distribution is estimated freely and updated by the usual Baum–Welch update
via functionoptim, and the optimization over parameterϕin (42), involving a constraint that levels are equally spaced, is solved by functionconstrOptim. 59 In Algorithm 2, the initial distribution is estimated freely and updated by the usual Baum–Welch update. As an alternativ...
2018
Reviewed July 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.