SciPost Submission Page
Extreme value statistics of edge currents in Markov jump processes and their use for entropy production estimation
by Izaak Neri, Matteo Polettini
This is not the latest submitted version.
This Submission thread is now published as
Submission summary
Authors (as registered SciPost users):  Izaak Neri · Matteo Polettini 
Submission information  

Preprint Link:  https://arxiv.org/abs/2208.02839v2 (pdf) 
Date submitted:  20220817 08:13 
Submitted by:  Neri, Izaak 
Submitted to:  SciPost Physics 
Ontological classification  

Academic field:  Physics 
Specialties: 

Approach:  Theoretical 
Abstract
The infimum of a current is its extreme value against the direction of its average flow. Using martingale theory, we show that the probability mass function of the infima of edge currents in timehomogeneous Markov jump processes is that of a geometric distribution. The mean value of the geometric distribution is determined by the effective affinity measured by a marginal observer that only sees the edge current. In addition, we show that a marginal observer can estimate the average entropy production rate of the underlying nonequilibrium process from the extreme value statistics in the edge current. The estimated average rate of dissipation obtained in this way equals the above mentioned effective affinity times the average edge current, and it is smaller or equal than the average entropy production rate. Moreover, we show that estimates of dissipation based on extreme value statistics can be significantly more accurate than those based on thermodynamic uncertainty ratios and KullbackLeibler divergences.
Current status:
Reports on this Submission
Anonymous Report 2 on 20221030 (Invited Report)
 Cite as: Anonymous, Report on arXiv:2208.02839v2, delivered 20221030, doi: 10.21468/SciPost.Report.6009
Strengths
See report
Weaknesses
See report
Report
The authors consider stochastic processes in Markov networks and, in particular, the statistics of the fluctuating, timeintegrated current across an individual edge of a network. For a statistical characterisation of these currents, this work focuses on the distribution of the infimum of the current, against an overall positive trend. Using methods from the theory of martingales, it is shown that this distribution is always a geometric one. The strength of this work is that it provides an operationally accessible way to evaluate the effective affinity of a single edge of a network, of which all other edges are invisible. This unlocks the possibility to apply bounds on the overall rate of entropy production that have been derived before (by one of the present authors, in Refs. [15,18]) and where this effective affinity enters as well. Unlike in this previous work, it is now possible to evaluate the effective affinity from only typical fluctuations, circumventing the unpractical procedure involving the limit of exponentially unlikely fluctuations. The distribution of the infimum and the resulting bound on the entropy production (compared to established bounds) are illustrated numerically for a wellknown model for a molecular motor. Remarkably, this bound on the entropy production can be much tighter than the established bounds.
The article is well written, with the right attention to detail, and the results are indeed relevant. I appreciate the splitting of the work into two "companion papers", along with Ref. [1]. This manuscript is selfcontained (and I understand the other one is so as well), and the equivalence of the results is discussed in Sec. 2.2. Attempting to sketch both derivations (which are complementary) in one manuscript would have overwhelmed the reader.
A few points remained unclear to me and there are a number of minor points that need to be fixed. I expect that the authors will be able to address these points and change the manuscript adequately, such that the paper will be ready for acceptance.
a.
In the introduction, I'd find it important to mention what class of physical systems is being considered, to provide some context for the first few paragraphs. The bound mentioned in the first paragraph is fairly universal, yet systems with inertia or coherent quantum dynamics can break it.
b.
It would be good to discuss in more detail the relation between the "effective affinity" and the "true affinity" (whatever that is). When are the two the same? I suspect that a*(x>y) is more related to the overall cycle affinity (in case of a unicyclic system) than to the true edgeaffinity a(x>y), even though the notation would suggest some relation to the latter.
c.
p.7: "the system attains the stalling state, which according to the marginal observer is indistinguishable
from equilibrium [1]". It may be distinguishable if the marginal observer does the right analysis, e.g., detect coherent oscillations in the autocorrelation function of the current. Also, I couldn't find the term "marginal observer" in [1], so I couldn't find what it has to say about this.
d.
In Sec. 3, the concept of a current being "proportional" to the entropy needs to be specified better. Is it exactly proportional or up to finite boundary terms?
e.
I don't see how to get from (44) to (46); how can the current and the factor 1/2 be identified? It seems to me that we need to have l(u>v)=k(v>u) (which is the case in the following applications of this formula).
f.
Sec. 6.4 (and also mentioned elsewhere in this manuscript): I doubt that it is possible to have "all microscopic affinities zero except one". When the affinity ln((p_u k(u>v))/(p_v k(v>u)) is zero, then also the current (p_u k(u>v))(p_v k(v>u)) is zero. But, because of the Kirchhoff rules for stationary currents, it is not possible to have a current through one edge, without a backflow through some other edge(s). In Sec. 7.2.1 it looks like the affinity is only the logratio of transition rates (without the contribution from the stationary distribution). In that case, it may be possible to have only one "affinity" equal to zero. But I'm not sure whether the argument from 6.4 then still applies.
g.
In Sec. 7, I'd welcome some more discussion. Beyond confirming the geometric distribution, what other information can be extracted from this kind of measurement? For example, is it possible to decide whether the observed current stems from a single or from several edges? In the latter case, is there a strong deviation from the geometric distribution?
h.
It is not clear to me how the bound derived from the infimum statistics can be so much stronger than the TUR. In which cases can the bound be saturated? (without the TUR saturating at the same time?) In Ref. [15], it is stated (below Eq. (19) there) that the equivalent bound on the splitting probability follows from the TUR. So how can it be stronger?
In the caption of Fig. 5, it is stated that the TUR is evaluated at 100s, which I assume is long enough to capture the longtime limit. Yet, the finitetime TUR can (in some cases) be stronger. How would the TUR evaluated at the optimal time interval compete with the bound from the infimum statistics? I think this would be a "fair game", because also the latter takes into account fluctuations on finite time scales.
Minor points:
1.
abstract: "current" > "integrated current" (the more familiar concept of a current is that of a rate, where the extreme value would be illdefined)
2.
abstract: "estimate" > "bound from below" (If I'm not mistaken, the bound can still be way off. For an "estimate" I would expect a guess with a lower and upper confidence interval)
3.
"largest excursion against its average flow" sounds like max(J<J>), which would not make sense. I think this can be made more precise with a better word choice (or dropped altogether, the explanation before and the Figure already make the concept clear).
4.
In which sense are (1) and (2) "analogous"? Can (1) be derived from (2)?
5.
"Interestingly, far from equilibrium, s_inf captures ..." What is the meaning of s_inf without the argument l? Is the limit to infinity implied, or is l arbitrary?
6.
The notation s_inf(l) in (9), (10) is somewhat confusing, since this quantity does not depend on l.
7.
Below (16): I'm confused about the difference between "geometric" and "exponential" distribution. According to wikipedia, they (only?) differ by whether the support is discrete or continuous.
8.
Eq. (19): The $\to\infty$ is typeset as subscript (please check this throughout, I've also seen this elsewhere, e.g. (125), (126)).
9.
is the f in Eq. (23) just 1p^esc? If so, say so.
Sec. 3: "Such an observer thinks that the observed current J is proportional to entropy production." Just because (s)he is naive, or is that simplest assumption indeed consistent with all observations?
10.
Eq. (31), typo a>c ?
11.
Typo: Full stop at end of sentence including (33).
12.
Beginning of Sec. 4.1: A few examples for "events" may be helpful ("event" sounds rather timelocal, but I believe sigma can depend on several times)
13.
typo: "devined"
14.
Around (37), (38): Difference between Markov process and Markov chain?
15.
I never understood why it's called RadonNikodym "derivative" (and not "ratio"). For a derivative I'd expect the notion of a small difference. What is the meaning of "dQ" and "dP" in (44)?
16.
Below (52): "than" > "then"
17.
Before (54), "M(t)<c". Does c need to be independent of T?
18.
"which is nota bene different from p_ss^xy": worth pointing out that it is also different from p_ss
19.
Before (58): Typo P_q_ss > P_p_ss ?
20.
In (60): is R_q_ss=Q_ss\circ Theta?
21.
"Lastly, introducing the effective microscopic affinity": is this the same quantity as in Ref. [30]?
22.
In what sense is the meaning of the effective affinity "kinematic" (as opposed to the true efficiency?)
23.
p. 15: Notation switches from J_{x>y}(T) to J(T). Make consistent or at least introduce the abbreviation.
24.
I think "generic" should be replaced by "general" in Sec. 6.2. Surely, the initial condition x, which is not "generic", is included in this general case.
25.
Between (81) and (82), the sentence starting with "Therefore ..." appears out of context with the preceding sentence ("Notice ...").
26.
Below (85): I believe this should read "for a *near* stalled current", and a* should be approximately 0, otherwise p_esc and j would be zero exactly.
27.
Before (95): It is not clear how the parameter dependence "does not contribute to dissipation". The affinity doesn't depend on the parameter, but the kinetics depends on it, hence the dissipation rate (if this can be equated to "dissipation") indirectly depends on the parameter as well.
28.
Caption Fig. 3: "lines">"line" (I just see one line)
29.
Sec. 7.2.3: typo "stalled stated"
30.
p. 24: "However, when the statistics of the current J contain strong nonMarkovian effects and when J is not proportional to the entropy production S, than s_KL provides a poor estimate ..." With strong nonMarkovian effects, can s_KL even be measured? (Before, it is stated that nonMarkovian effects are ignored).
31.
I am not sure whether I understand the distinction between s hat and s doublehat. If (121) "applies" only to a single edge, should't it be a doublehat? I understand that in any case, the estimated entropy production is that of the overall network, not just a single edge.
Anonymous Report 1 on 2022922 (Invited Report)
 Cite as: Anonymous, Report on arXiv:2208.02839v2, delivered 20220922, doi: 10.21468/SciPost.Report.5746
Strengths
1) The paper provides a timely, relevant study of martingale methods
applied to the study of currents along a single visible observed edge
(as introduced in Ref. [30]).
2) The first two chapters provide a clear, intuitive introduction
into the methodology and physical significance of this work.
Weaknesses
1) Chapter 8 compares the method to recent entropy estimators, which
are quoted but not discussed properly. Thus, the conclusions might give
a misleading picture of recent entropy estimation methods.
2) The physical significance of the results beyond the summary might be
difficult to grasp due to the heavy mathematical formalism introduced throughout the work.
Report
The study of partially accessible Markov networks is of major interest
in stochastic thermodynamics. Gaining new insights into statistics
of observables and estimators for entropy production is not only a
timely challenge but also crucial from an operational point of view.
The present work adapts a transitionbased description for partially
accessible Markov networks to apply known results from martingale theory
and extant entropy estimators to currents counting transitions along a
single edge in a Markov network. The idea is not novel, but generalizes
previous results that are based on martingale theory (e.g. Ref. [22]).
In a numerical case study for a molecular motor model, the corresponding
entropy estimator is compared to known estimators like the thermodynamic
uncertainty relation, the first passage ratio and a naive bound based
on the KullbackLeibler divergence. The deduced entropy estimator is
interpreted as an effective affinity, a concept already introduced
in Ref. [30].
The manuscript is well written, both in terms of grammar and
mathematical rigor. It provides an introduction and a summary of the
main results that give the reader a clear overview over the paper and
how it relates to recent literature. While the theoretical sections of
the paper are kept at a quite technical level, the paper gives explicit
demonstrations of the main results in a numerical case study.
Requested changes
Major concerns:
1. It is unclear to me how the paper concludes that the derived
estimator "can be significantly more accurate than those based
on ... KullbackLeibler divergences".
Equation (112) is the only one in the manuscript that is claimed to be
an estimator based on the KullbackLeibler divergence. However, it is
certainly not the only estimator of this sort, since more sophisticated
versions have been proved e.g. in Ref. [49] and particularly in
Ref. [48], which discusses an entropy estimator for the identical
molecular motor model.
In light of these recent advances, it is, in my opinion, not justified
to claim greater accuracy without a comparison to these recent bounds.
The authors should either clarify these misleading conclusions or
include these more recent entropy estimators in the numerical study.
(see also the minor remark below)
Minor remarks:
2. Two novel entropy estimators, the infimum ratio and the modified
infimum ratio, are introduced. From an operational point of view, it is
interesting to know how much statistics is actually needed for these
estimators, i.e. how long should the trajectory be for comparable
error/bias? Is there a connection to the results obtained in [18]?
3. The math in Ch. 5 seems to be very similar to the
derivations leading to the "informed partial" estimator in
Gili Bisker et al J. Stat. Mech. (2017) 093210. Both incorporate the
effective affinity of Ref [30] and the stalling distribution (cf. the
paragraph following eq. (63)). How deep is this connection?
More concrete, it might be more appropriate to numerically compare to
the entropy estimators discussed in Bisker et al. These estimators seem
more closely related to the present work than those that are based on
waiting times.
Author: Izaak Neri on 20230110 [id 3226]
(in reply to Report 1 on 20220922)The attachment contains our Reply to the Referee's comments .
Author: Izaak Neri on 20230110 [id 3225]
(in reply to Report 2 on 20221030)The attachment contains a reply to the Referee's comments.
Attachment:
RefereeReply.pdf