Why is it easier to predict the epidemic curve than to reconstruct the underlying contact network?

D\'aniel Keliger; Ill\'es Horv\'ath

REVIEW 2 major objections 4 minor 13 references

Why is it easier to predict the epidemic curve than to reconstruct the underlying contact network?

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Forecasting an epidemic curve is provably more stable than recovering the contact network that generated it.

desk verdict Solid prediction-robustness and a clean inversion formula, but the non-identifiability theorem's counterexample lives outside the model class, so the abstract's reconstruction claim overreaches. read the letter →

arxiv 2506.00059 v2 pith:BC4LWPXP submitted 2025-05-29 q-bio.PE

Dániel Keliger , Illés Horváth This is my paper

classification q-bio.PE MSC 92D3005C8005C82

keywords SISepidemicmodelnetworkreconstructionpredictionrobustnessgraphonsSzemerédiregularitylemmacutnormGrammatrixidentifiability

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper gives a rigorous explanation of a numerical puzzle in epidemic modeling: why a network estimated from early epidemic data can be badly wrong and yet still produce accurate forecasts of the epidemic curve. Working with the deterministic SIS model on weighted dense graphs, the authors prove two complementary statements. The forward map from network to trajectory is stable: if two networks produce almost the same infection ratios on an observation window, they stay close afterwards, uniformly across all large networks in the parameter class. The inverse map is not: for every large dense network there is a different network, far away in the cut norm, whose trajectories are nearly indistinguishable. A separate formula recovers the network exactly from the trajectory whenever the Gram matrix of the observed infection ratios is invertible.

What carries the argument

The argument runs through the Gram matrix $G(T)_{ij} = \int_0^T z_i(t) z_j(t)\,dt$, which converts equality of observable neighborhood statistics into a positive-definiteness condition, and through Szemerédi's weak regularity lemma, lifted to graphons. The lemma guarantees that any large dense weighted graph can be approximated by a block-homogeneous graph with few blocks, and the block statistics alone determine the trajectory. The authors then perturb the block homogeneous approximation by doubling weights inside two parities of each block and zeroing weights between them; this preserves every block sum, so the epidemic curve is unchanged, while the cut norm moves by at least $m/16$. For prediction, a compactness argument on the graphon parameter space shows that the ratio between the future and past Gram norms is bounded uniformly, yielding the uniform $\delta$ in Theorem 3.12.

What would settle it

Compute, for growing $n$, the supremum over the parameter class of $\Lambda^{(n)}(p)$ in inequality (13); if that supremum diverges, the uniform $\delta$ of Theorem 3.12 cannot exist. Alternatively, on a large dense weighted graph with $G(T_0)$ invertible, minimize the observable error $H_{2,T_0}^2$ over candidate $\widehat{W}$: if every minimizer with error below $\delta$ always stays within cut distance $m/16$ of the true $W$, Theorem 3.9's construction is contradicted.

Watch

Extended reading notes

Core claim

The central discovery is that predictability and reconstructability are not dual: for large dense weighted SIS networks, prediction of the epidemic curve is uniformly robust while reconstruction of the underlying contact matrix is ill-conditioned. Formally, Theorem 3.12 states that for every $\varepsilon > 0$ there is a $\delta > 0$, independent of the network size and of the parameters, such that a small past error $H_{2,T_0} < \delta$ forces the future error $H_{2,T} < \varepsilon$. Theorem 3.9 states the opposite for inversion: for every large enough dense network there exists a competitor $\widehat{W}$ with cut-norm distance at least $m/16$ whose epidemic curve stays within $\varepsilon$ of the true curve up to time $T$. The two sides are tied together by the identity $A G(T_0) = R(T_0)$, where $G(T_0)$ is the Gram matrix of the infection ratios and $R(T_0)$ is an observable cross-moment matrix; when $G(T_0)$ is invertible, the network is explicitly $W = R(T_0) G^{-1}(T_0) \Pi^{-1}$.

Load-bearing premise

For the non-identifiability theorem the network must be dense and the communities must be of equal size: every edge weight at least $m>0$ and $\pi_i = 1/n$ for every vertex; if the graph is sparse or the population is spread unevenly, the construction that separates networks in cut norm while keeping trajectories close no longer works.

Editorial extensions

If this is right

Modelers working with dense approximations of contact networks can trust short-horizon curve forecasts even when the estimated network differs strongly from the true one, since the forecasting error $H_{2,T}$ is controlled by the observable past error $H_{2,T_0}$.
When the Gram matrix $G(T_0)$ is invertible, the network is identifiable in the literal sense, and formula (12) recovers $W$ from perfect observations of $z(t)$, $\pi$, and $\gamma$ on an arbitrarily short positive interval.
For the SI model with uniform positive community weights, almost every network is reconstructible when all degrees are distinct, but the reconstruction becomes ill-conditioned as $n$ grows and the Gram matrix approaches singularity.
Discrete-time observations and measurement noise degrade the continuous-time estimate only by an error of order $\Delta = \Delta_1+\Delta_2+\Delta_3+\Delta_4$, uniformly in $n$, so the robustness story survives finite sampling.
The same block-regularity argument implies that any two networks with the same block sums are observationally equivalent for the SIS dynamics, so the identifiable object is the block-averaged network, not its fine structure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A dense-network caveat: the $m \le w_{ij}$ lower bound means the theorems do not directly cover sparse or power-law contact graphs; extrapolating the prediction-robustness result to such graphs would require a separate mechanism, likely based on low-rank or mean-field structure rather than Szemerédi regularity.
The explicit formula $W = R(T_0) G^{-1}(T_0) \Pi^{-1}$ identifies $W\Pi$ rather than $W$ as the genuinely observable object, so in practice errors in community sizes $\pi$ may be the dominant source of reconstruction failure even when the Gram matrix is invertible.
One testable consequence for simulations: on dense weighted networks, an estimator that deliberately searches over the non-identifiable equivalence class of networks sharing the same block sums should produce forecasts as accurate as one that tries to pin down the true $W$; a comparison of the two forecasting errors would quantify how much of the curve is actually determined by the data.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First thing to know: the core new mathematics is real. Theorem 3.1's Gram-matrix characterization is clean, identity (11) and the explicit formula (12) are correct, and Theorem 3.12's uniform-in-n prediction bound is a genuine graphon-compactness argument. That part deserves a serious referee.

But the stress-test note is right, and it lands on the paper's main advertised claim. In the proof of Theorem 3.9, the constructed \hat W sets same-parity weights to 2w'_ij and opposite-parity weights to 0. Since w'_ij ∈ [m,M], \hat W has entries 0 and 2M, so it is not in P_{m,M}^{(n)} (nor P_{0,M}^{(n)} when 2M > M). The theorem says 'there is a \hat W' without restricting it to the parameter class, but the whole reconstruction problem is set in P_{m,M}; an out-of-class network that mimics trajectories does not show the inverse map is ill-conditioned on the model class. And it's not a cosmetic issue: take W ≡ m with homogeneous q. The trajectory depends only on the common row sum, and admissibility forces every entry ≥ m, so the only in-class network with the same trajectory is W itself. For small ε, any in-class \hat W with trajectory error < ε must have row sums within O(ε) of m, which forces cut-norm distance O(ε), not m/16. So the m/16 separation is genuinely unattainable within the class. The dense lower bound isn't a harmless technicality; it's what lets the perturbation exit the class.

There's a second smaller gap: the proof applies the Lemma 4.2/Remark 4.3 implication to the pair (p'', \hat p) even though \hat W is not bounded by M. That's likely fixable by noting \hat W ∈ P_{0,2M}, but as written the constant isn't justified.

What does this leave? The prediction-robustness half (Theorem 3.12) is unaffected and is the strongest part of the paper. The explicit inversion formula and the Gram characterization are useful and correct. The reconstruction non-identifiability claim needs to be restated honestly: either as a statement about estimators that are allowed to leave the class, or with a different construction that stays in class, which the example above shows is impossible in general. The abstract's sentence 'reconstructing the underlying network is ill-conditioned' overstates what is proven.

This is a paper for network epidemiologists and graphon theorists. I'd send it to review because the prediction result and the inversion formula are worth having, but I'd require the authors to fix Theorem 3.9 and the abstract before publication.

Referee Report

2 major / 4 minor

Summary. The paper studies a deterministic SIS metapopulation model on weighted graphs and gives rigorous statements about two related problems: predicting the future epidemic curve from a finite-time observation, and reconstructing the weighted contact network from the same observation. The main results are (i) an explicit reconstruction formula (12) when the Gram matrix G(T0) is invertible, (ii) a uniform prediction-robustness theorem (Theorem 3.12) stating that for all n and all parameters in P_{0,M}, a small past error H_{2,T0} implies a small future error H_{2,T} with a modulus independent of n, and (iii) a non-identifiability theorem (Theorem 3.9) claiming that for large dense networks there exists a network W_hat with large cut-norm distance from W yet nearly identical trajectories. The proofs use the Gram-matrix observability framework, graphon theory, Szemerédi's weak regularity lemma, and a compactness argument for the graphon parameter space.

Significance. The prediction-robustness result (Theorem 3.12) is a strong and interesting theorem: it provides a uniform, dimension-free bound on error propagation for the SIS dynamics, and its proof via compactness of the graphon quotient space is elegant. The explicit reconstruction formula (12) is also a clean linear-algebra identity that clarifies when exact reconstruction is possible. If the reconstruction ill-conditioning claim (Theorem 3.9) can be correctly established within the stated parameter class, the paper would provide a rigorous explanation of the numerical phenomenon observed by Prasse and Van Mieghem. However, as written, the reconstruction claim is not supported in the form needed for the abstract's conclusion, because the counterexample constructed in Theorem 3.9 leaves the admissible parameter class.

major comments (2)

[Section 5.2, proof of Theorem 3.9] The matrix W_hat constructed in the proof of Theorem 3.9 is not admissible for the reconstruction problem defined in Section 2.2. Specifically, the proof sets W_hat entries to 0 for opposite-parity sub-blocks and to 2w'_ij for same-parity sub-blocks, where w'_ij can be as low as m and as high as M. Hence W_hat can contain entries below m or above M, so it need not lie in P_{m,M}^{(n)}, and it may even violate the upper bound of P_{0,M}^{(n)} when doubling exceeds M. The theorem statement does not explicitly require W_hat to be admissible, but the abstract and introduction conclude that 'reconstructing the underlying network is ill-conditioned' for the parameter class in which the reconstruction problem is set. As written, Theorem 3.9 only shows that if one allows estimators outside the class, then a far-away network can mimic the trajectory; it does not establish non-identifiability within the admissible class.
[Section 3.1, Theorem 3.9; Section 2.2, equation (8)] The universal claim of Theorem 3.9 is false under the natural in-class reading. Take W = mJ (all entries equal to m), π_i = 1/n, and a homogeneous q (z_i(0) = z_0, γ_i = γ). Then z_i(t) = s(t) for all i, with (Az(t))_i = m s(t). If W_hat ∈ P_{m,M}^{(n)} generates the same trajectory, then for each i we must have (1/n) Σ_j W_hat_ij = m, i.e., Σ_j (W_hat_ij − m) = 0. Since W_hat_ij ≥ m for all i,j, this forces W_hat_ij = m for all i,j, so W_hat = W and the cut-norm distance is 0. For small ε, any W_hat with ε-close trajectory must have row sums arbitrarily close to nm, and again the lower bound m forces W_hat to be close to mJ in cut norm. Thus there is no admissible W_hat with cut-norm distance at least m/16 for this p, contradicting the theorem's quantification over all p ∈ P_{m,M}^{(n)}. The statement needs either a restriction on p (e.g., excluding parameters with no slack in the window [m,M]) or an explicit enlargement of the estimator class; the current abstract overstates the scope of the result.

minor comments (4)

[Section 3.1, Theorem 3.1] Condition 2 is misprinted: the phrase '⟨v, z(t)⟩ ⇒ v=0' should read '⟨v,z(t)⟩ = 0 for all t ∈ [0,T0] implies v = 0'.
[Section 4, Lemma 4.4] In condition 3), the use of the same symbol (I_k) for both the coarse Szemerédi partition and the original discrete partition is confusing; please use different letters (e.g., (J_k) for the coarse partition, (I_i) for the discrete partition) to make the refinement statement unambiguous.
[Section 5.3, Lemma 5.5] The subsequence notation in the weak-compactness step is garbled: 'f_{N'_k}^{φ_{N''_k}} ⇀ f' should be written with consistent indices for the subsequence and the measure-preserving maps. The displayed estimate '≤ 2∫ ...' also has a missing absolute value inside the integral; please revise for clarity.
[References] Reference [10] is listed as 'L. L´aszl´o and B. Szegedy' in the text; the first author's name is László Lovász, so the citation should read 'L. Lovász and B. Szegedy'.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the prediction and reconstruction theorems are proven from the SIS equations using exact identities, analytic continuation, and standard external mathematics; the only self-citation is not load-bearing.

full rationale

The derivation chain is self-contained. Theorem 3.12 (uniform prediction robustness) is proven via the exact identity H_{2,T}^2 = ∫ ||η(x)||_{G_T}^2 dx (Eq. 28) and Lemma 5.5, a compactness argument over the graphon parameter space P_{0,M} whose compactness (Lemma 4.8) is imported from the standard martingale proof of Lovász–Szegedy, not from the authors' own work; the bound is uniform over all p and n because χ_N → 0, and the future/past nullspace coincidence rests on analyticity (Lemma 2.1) applied to the polynomial SIS vector field. Theorem 3.9 (reconstruction ill-conditioning) is a constructive counterexample built on Szemerédi's weak regularity lemma (Lemma 4.4, proved from the external references [10] and [1]) that produces a block-regular W' with identical coarse-grained dynamics and a bounded cut-norm perturbation; the lower bound m is a stated dense-class assumption, not a hidden input. The reconstruction formula (12), W = R(T0)G^{-1}(T0)Π^{-1}, is an exact rearrangement of the identity AG(T0) = R(T0) (Eq. 11) under the invertibility criterion of Theorem 3.1, which is proved in Section 5.2 from first principles. No parameter is fitted and then relabeled a prediction: H_{2,T0} and H_{2,T} are independently defined observable and future discrepancy measures, and δ in Theorem 3.12 depends only on ε. The sole self-citation [8] (Keliger–Horváth–Takács) is used in Section 2.1 only to motivate the deterministic SIS limit from stochastic graphon processes; it appears nowhere in the proofs of the main theorems and is therefore not load-bearing. The skeptic's observation that the Ŵ constructed in Theorem 3.9 may violate the bounds m ≤ ŵ_{ij} ≤ M, so that the counterexample leaves the admissible class P_{m,M}^{(n)}, is a legitimate scope caveat about what the abstract's 'ill-conditioned' claim covers, but it is not circularity: the theorem as stated asserts only existence of a Ŵ with matching trajectory and large cut distance, and the proof does not assume its own conclusion. Accordingly, with zero circular steps and one minor non-load-bearing self-citation, the appropriate score is 1.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

No fitted parameters appear in the paper; m and M are fixed model bounds, not estimated from data. The central claim rests on the SIS ODE, dense-graph and observability assumptions, and standard graphon analytic tools. No new physical or mathematical entities are postulated.

free parameters (2)

m
Uniform lower bound on edge weights and initial infection levels in P^(n)_{m,M}. Fixed by hand, not fit to data, but load-bearing for the dense-graph ill-conditioning theorem.
M
Uniform upper bound on edge weights and curing rates in P^(n)_{m,M}. Fixed by hand, not fit to data, and it controls Lipschitz constants and compactness arguments.

assumptions (6)

domain assumption The SIS ODE (1), dz_i/dt = (1-z_i)(Az)_i - gamma_i z_i, accurately represents the epidemic dynamics under study.
All theorems are about this system and its graphon extension (21); the model is not derived from first principles in the paper.
domain assumption The weighted graph is symmetric with nonnegative edge weights, and the population weights pi_i satisfy pi_i > 0 and sum pi_i = 1.
This defines A = W Pi and is needed for the reconstruction formula and the Gram-matrix arguments.
domain assumption Parameters lie in P^(n)_{m,M}: m <= w_ij <= M, 0 <= gamma_i <= M, and m <= z_i(0) <= 1-m; Theorem 3.9 also assumes pi_i = 1/n.
These bounds define the model class. The lower bound m makes the graph sequence dense, which is essential for the Szemeredi-style non-identifiability construction.
domain assumption Full continuous observation of z(t) and gamma on [0,T0], and therefore of Az(t) via equation (6), is available.
The reconstruction formula (12) and the discrete-error analysis require node-level curves and recovery rates; real surveillance typically sees aggregates, so this is an idealization.
standard math Szemeredi's weak regularity lemma for graphons and compactness of the factorized parameter space hold as in Lovasz-Szegedy [10] and graphon theory [9].
Lemmas 4.4 and 4.8 are imported from the graphon literature and are used in Theorems 3.9 and 5.1.
standard math Standard analyticity and functional analysis tools: analytic continuation via Lemma 2.1, Gronwall's inequality, and weak compactness of the unit ball in L^2.
These tools propagate zero-error identities from [0,T0] to all times and pass to limits in compactness arguments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Why is it easier to predict the epidemic curve than to reconstruct the underlying contact network?." pith.science (2026). https://pith.science/paper/BC4LWPXP

@misc{pith2026250600059,
  author       = {Pith},
  title        = {Pith review of: Why is it easier to predict the epidemic curve than to reconstruct the underlying contact network?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BC4LWPXP}},
  note         = {Machine review of arXiv:2506.00059}
}

read the original abstract

We study the deterministic Susceptible-Infected-Susceptible (SIS) epidemic model on weighted graphs. In their numerical study [10] van Mieghem et al. have shown that it is possible to learn an estimated network from a finite time sample of the trajectories of the dynamics that in turn can give an accurate prediction beyond the sample time range, even though the estimated network might be qualitatively far from the ground truth. We give a mathematically rigorous derivation for this phenomenon, notably that for large networks, prediction of the epidemic curves is robust, while reconstructing the underlying network is ill-conditioned. Furthermore, we also provide an explicit formula for the underlying network when reconstruction is possible. At the heart of the explanation, we rely on Szemer\'edi's weak regularity lemma.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[1]

N. Alon, W. de la Vega, R. Kannan, and M. Karpinski. Random sam- pling and approximation of MAX-CSPs.Journal of Computer and System Sciences, 67(2):212–243, 2003. Special Issue on STOC 2002

work page 2003
[2]

Beaufort, P.-Y

L.-B. Beaufort, P.-Y. Mass´ e, A. Reboulet, and L. Oudre. Network recon- struction problem for an epidemic reaction–diffusion system.Journal of Complex Networks, 10(6):cnac047, 10 2022

work page 2022
[3]

Delmas, D

J.-F. Delmas, D. Dronnier, and P.-A. Zitt. An infinite-dimensional SIS model, 2020

work page 2020
[4]

´Odor, D

G. ´Odor, D. Czifra, J. Komj´ athy, L. Lov´ asz, and M. Karsai. Switchover phenomenon induced by epidemic seeding on geometric networks.Proceed- ings of the National Academy of Sciences, 118(41):e2112607118, 2021

work page 2021
[5]

Gozzi, M

N. Gozzi, M. Tizzoni, M. Chinazzi, L. Ferres, A. Vespignani, and N. Perra. Estimating the effect of social inequalities in the mitigation of COVID- 19 across communities in Santiago de Chile.Nature Communications, 12, 2021

work page 2021
[6]

Graphon branching processes and fractional isomorphism

J. Hladk´ y, E. K. Hng, and A. M. Limbach. Graphon branching processes and fractional isomorphism, 2024. https://arxiv.org/abs/2408.02528

work page Pith review arXiv 2024
[7]

M. T. Islam, M. Akon, A. Abdrabou, and X. Shen. Modeling epidemic data diffusion for wireless mobile networks.Wireless Communications and Mobile Computing, 14(7):745–760, 2014

work page 2014
[8]

Keliger, I

D. Keliger, I. Horv´ ath, and B. Tak´ acs. Local-density dependent markov processes on graphons with epidemiological applications.Stochastic Pro- cesses and their Applications, 148:324–352, 2022. 27

work page 2022

Show all 13 references

[9]

Lov´ asz.Large Networks and Graph Limits, volume 60 ofColloquium Publications

L. Lov´ asz.Large Networks and Graph Limits, volume 60 ofColloquium Publications. American Mathematical Society, 2012

2012
[10]

L´ aszl´ o and B

L. L´ aszl´ o and B. Szegedy. Szemer´ edi’s Lemma for the Analyst.Geometric and Functional Analysis, 17:252–270, 04 2007

2007
[11]

Murphy, V

C. Murphy, V. Thibeault, A. Allard, and P. Desrosiers. Duality between predictability and reconstructability in complex systems.Nature Commu- nications, 15, 05 2024

2024
[12]

Prasse and P

B. Prasse and P. V. Mieghem. Predicting network dynamics without re- quiring the knowledge of the interaction graph.Proceedings of the National Academy of Sciences, 119(44):e2205517119, 2022

2022
[13]

L. Ruiz, L. F. O. Chamon, and A. Ribeiro. Reply to ’Com- ments on Graphon Signal Processing’ [arXiv:2310.14683], 2024. https://arxiv.org/abs/2401.05326v1. 28

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

N. Alon, W. de la Vega, R. Kannan, and M. Karpinski. Random sam- pling and approximation of MAX-CSPs.Journal of Computer and System Sciences, 67(2):212–243, 2003. Special Issue on STOC 2002

work page 2003

[2] [2]

Beaufort, P.-Y

L.-B. Beaufort, P.-Y. Mass´ e, A. Reboulet, and L. Oudre. Network recon- struction problem for an epidemic reaction–diffusion system.Journal of Complex Networks, 10(6):cnac047, 10 2022

work page 2022

[3] [3]

Delmas, D

J.-F. Delmas, D. Dronnier, and P.-A. Zitt. An infinite-dimensional SIS model, 2020

work page 2020

[4] [4]

´Odor, D

G. ´Odor, D. Czifra, J. Komj´ athy, L. Lov´ asz, and M. Karsai. Switchover phenomenon induced by epidemic seeding on geometric networks.Proceed- ings of the National Academy of Sciences, 118(41):e2112607118, 2021

work page 2021

[5] [5]

Gozzi, M

N. Gozzi, M. Tizzoni, M. Chinazzi, L. Ferres, A. Vespignani, and N. Perra. Estimating the effect of social inequalities in the mitigation of COVID- 19 across communities in Santiago de Chile.Nature Communications, 12, 2021

work page 2021

[6] [6]

Graphon branching processes and fractional isomorphism

J. Hladk´ y, E. K. Hng, and A. M. Limbach. Graphon branching processes and fractional isomorphism, 2024. https://arxiv.org/abs/2408.02528

work page Pith review arXiv 2024

[7] [7]

M. T. Islam, M. Akon, A. Abdrabou, and X. Shen. Modeling epidemic data diffusion for wireless mobile networks.Wireless Communications and Mobile Computing, 14(7):745–760, 2014

work page 2014

[8] [8]

Keliger, I

D. Keliger, I. Horv´ ath, and B. Tak´ acs. Local-density dependent markov processes on graphons with epidemiological applications.Stochastic Pro- cesses and their Applications, 148:324–352, 2022. 27

work page 2022