Dynamic Vulnerability in Oscillatory Networks and Power Grids

Cheng Ma; Marc Timme; Xiaozhu Zhang

REVIEW 3 major objections 4 minor 35 references

Dynamic Vulnerability in Oscillatory Networks and Power Grids

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a Dynamic Vulnerability Index (DVI), computed from the network's linear response eigenmodes and the signal's power spectral density, to rank which nodes in an oscillator network respond with the largest amplitude to…

desk verdict A genuinely new, cheap vulnerability index with strong apparent rank correlation in one simulated grid, but the paper's central phase-coherence assumption is asserted rather than derived and is likely false for random-phase noise; the validation is too thin to support the robustness claim. read the letter →

arxiv 1908.00957 v1 pith:ZV5DWIKY submitted 2019-08-02 nlin.AO

Xiaozhu Zhang , Cheng Ma , Marc Timme This is my paper · ORCID

classification nlin.AO

keywords DynamicVulnerabilityIndexlinearresponsetheoryoscillatornetworkspowergridsKuramotomodelcolorednoisenoderankingresonance

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

The paper sets out to identify, cheaply and in advance, which nodes in an oscillatory network — in particular a model AC power grid — will show the largest frequency-response amplitudes when the network is driven by a stochastic signal with a known power spectrum. It proposes the Dynamic Vulnerability Index (DVI): for each node, the sum over the resonance band of the amplitude of its linear response to every Fourier component, weighted by the square root of the signal's power spectral density. The authors claim this index's ranking predicts the ranking of the actual maximum resonant responses seen in direct simulation of the nonlinear swing equations, with Pearson correlation above 0.985 and prediction error falling to roughly 15% of a random ranking after 100 seconds. If true, the index gives grid operators a fast way to spot units that may be particularly disturbed by fluctuating renewable infeed without running long simulations.

What carries the argument

The working object is the linear response theory of the second-order Kuramoto model, Eq. (3), which turns the network's dynamical response into a sum over Laplacian eigenmodes. Each eigenmode contributes a frequency-dependent resonance peak whose spatial pattern is set by the overlap of the eigenvector component at the driven node and at the observed node; the DVI is the band-integrated magnitude of this sum, using the signal's PSD to set the amplitude of each Fourier component. The resonance regime, the interval between the lowest and highest eigenmode resonance frequencies, determines which frequencies contribute. The assumption that random-phase Fourier components add like magnitudes after long times is what turns a phase-sensitive superposition into a simple index.

What would settle it

Drive a small oscillator network at a node with a carefully constructed noise whose power spectrum concentrates on two adjacent resonance frequencies whose eigenmode contributions at a particular node have opposite phases, then compare the node's measured maximum response over long $T$ with its DVI; if a low-DVI node systematically attains higher measured maxima than a high-DVI node, the magnitude-sum approximation behind Eq. (6) fails and the ranking claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that the node-specific maximum response amplitude to a noisy driving signal can be ranked a priori by a weighted accumulation of single-frequency linear responses. For the second-order Kuramoto-type model \ddot{\$\theta$}_i = P_i - \$\alpha$ \dot{\$\theta$}_i + \sum_j K_{ij} \sin(\theta_j-\theta_i) + \delta_{ik}D(t), the linearized dynamics decouple into eigenmodes of the weighted graph Laplacian $L$ with eigenvalues $\lambda^{\ell}$ and eigenvectors $v^{\ell}$; each sinusoid of frequency $\omega$ excites mode $\ell$ with amplitude proportional to $i\omega v_k^{\ell} v_i^{\ell} /(-\omega^2 + i\alpha\omega + \lambda^{\ell})$, and resonances occur at $\omega^{\ell}_{\mathrm{res}} = \sqrt{\lambda^{\ell} - \alpha^2/4}$. The DVI of node $i$ is the sum of these resonance amplitudes across the resonance band with the square root of the power spectral density $S(\omega)$ acting as weight. The paper reports that in a sample power-grid network, the ranking of this index matches the ranking of $\max_t |\dot{\Theta}_i^{(k)}(t)|$ from direct simulation, with a normalized Spearman footrule error that drops by about 80% in the first 10 seconds and reaches about 15% of the random-guess level at $T=100$ s, for colored noise exponents $b \in \{0,1,2\}$.

Load-bearing premise

The paper assumes, without proof or error bound, that for sufficiently long noise time series the maximum of the real summed response equals the sum of the magnitudes of its Fourier components, Eq. (3) to Eq. (6); if phases do not align, DVI overestimates the true maximum and the ranking can change.

Editorial extensions

If this is right

Operators can rank all nodes by susceptibility to a fluctuating source at a known node using only the Laplacian spectrum and the PSD exponent, without integrating the nonlinear dynamics.
The same ranking remains highly predictive for white to brown noise ($b \in [0,2]$), so the method needs no re-fit per noise color.
DVI can flag vulnerable nodes located far from the driving node, nodes that topology-only heuristics such as dead-end degree would not single out.
Because the construction is built on linear response, it transfers to other oscillator-network models, including networks with more variables per node.
The fast convergence of the ranking suggests that even short recordings, tens of seconds, suffice for a stable vulnerability ordering.

Reading between the lines

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

Inference: the DVI is best read as an upper-envelope ranking; by replacing complex phase sums with magnitude sums, it estimates the largest response achievable if phases happen to align, rather than the typical response, and real finite-time maxima should be lower.
Inference: a testable extension would apply DVI to real measured wind-power time series at a known grid location and compare its node ranking against both simulated swing dynamics and measured frequency excursions, where systematic over-ranking would reveal the effect of cancelling eigenmode phases.
Inference: one could optimize network topology, damping, or signal shaping to minimize the DVI of critical nodes, turning the index from a diagnostic into a planning objective.
Inference: the method's linearity suggests a spectral interpretation of DVI as a weighted $\ell^1$-type norm of the transfer function on the resonance band, with nonlinear and out-of-band effects as a possible boundary of validity.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

The DVI is a new index, not present in the cited literature, and the core idea is sensible: rank nodes by their expected maximal resonant response using the network's linear response eigenmodes and the driving signal's power spectrum. The derivation up to Eq. (6) is clean, and the benchmark is direct nonlinear simulation, not the same linear response theory the index is built from, so the test is independent. The reported Spearman footrule errors and Pearson correlations around 0.985 for one sample grid across PSD exponents b in [0,2] are credible evidence that the ranking works for that network and that driving setup. The paper is also honest about DVI being a ranking tool, not an amplitude predictor, and the self-citation to the separately published linear response theory is legitimate.

The real soft spot is exactly where the reader and stress-test put it: the sentence after Eq. (6) claims that the response to a noisy fluctuation approaches the sum of the magnitudes of the complex frequency responses for sufficiently long time series. That is not derived, and as stated it is not generally true. For random Fourier phases, the time-domain response at a node is approximately Gaussian with variance proportional to the L2 sum over frequencies of S|H|^2. Its maximum over a long interval grows roughly as sqrt(log T) times that standard deviation, not as the L1 sum of sqrt(S)|H|. The L1 and L2 sums can rank nodes differently; many small off-resonant contributions can dominate the DVI while the true extremum is controlled by a few resonant peaks. So the paper's central approximation is not just unproven; it is the kind of claim that needs an error bound or a worst-case argument, and none is given.

That said, the situation is not a disaster. The numerical validation on one synthetic power-grid network still shows the index works in a nontrivial test, and the error metric is appropriate. The limitations are real but proportionate: one network topology, apparently one driven node, and no explicit statement about how many noise realizations or how the frequency grid in Eq. (6) was chosen. No code or data are provided, which makes independent reproduction harder. I would not call the paper a noise; I would call it an interesting heuristic with an overstated foundation. The phrase 'enables robust predictions' goes beyond what has been shown.

Who should read this? Network scientists and power-grid researchers who want a fast, screening-level vulnerability ranking and are willing to check it on their own systems. It deserves a serious referee, but not a clean acceptance: the authors should either prove or substantially qualify the phase-coherence assumption, add multi-network and multi-realization validation, and specify the numerical details. I would send it to review with the expectation of major revision.

Referee Report

3 major / 4 minor

Summary. The manuscript studies second-order Kuramoto networks used as coarse-grained AC power grid models, driven by stochastic inputs with a prescribed power spectral density. Building on linear response theory (ref. [12]), it defines a Dynamic Vulnerability Index (DVI) in Eq. (6) as a sum over the resonance frequency interval I_res of S(ω)^(1/2) times the magnitude of the single-frequency transfer function from the driven node k to node i. The authors claim that DVI ranks nodes by their maximum time-domain resonant response, and they support this with direct nonlinear simulations on one synthetic power-grid network: the Spearman footrule error between DVI ranking and simulated maximum-response ranking drops to about 15% of a random ranking by T=100 s, with Pearson correlation r>0.985 for PSD exponents b in {0,1,2}. The key approximation after Eq. (6) is stated but not proved, and the numerical evidence is restricted to a single network, a single driven node, and apparently a single noise realization per PSD exponent.

Significance. If the main claim holds, DVI provides a computationally cheap screening tool for dynamic vulnerability in oscillator networks and power grids, and the linear-response basis makes the approach generic in principle. The numerical benchmark is independent of the proposed index because it compares against direct nonlinear simulation, so the validation is not circular. However, the central contribution rests on an unproven and, in general, questionable approximation about random-phase sums, and the validation is too limited to establish the claimed robustness. The paper is therefore significant conditional on a rigorous or at least systematically tested justification of Eq. (6).

major comments (3)

[Section III, Eq. (6)] The approximation stated immediately after Eq. (6) is load-bearing and is not justified. For a finite-length stationary input, the Fourier phases do not align in general, and the maximum over [0,T] of the real-part response is not the sum of the per-frequency magnitudes. For a stationary Gaussian response with variance σ_i^2 proportional to Σ_{ω∈I_res} S(ω)|H_i(ω)|^2, the maximum over [0,T] grows like σ_i sqrt(2 log T) plus a common stochastic term, whereas DVI in Eq. (6) is an L1 sum over the same frequency set. These two functionals are not order-equivalent in general, so the claimed ranking prediction can fail. The authors should either prove the convergence to the sum of magnitudes with explicit rates and error bounds, or replace Eq. (6) by an L2-based quantity that is justified for extremes of stationary Gaussian processes, and then re-run the validation.
[Section III, Fig. 3] The numerical evidence uses one network realization from the random-growth model, one driven node, and apparently one noise realization per PSD exponent. The reported r>0.985 and the 15% error level are therefore a single-sample demonstration; they do not establish robustness to network topology, drive location, or noise realization. I request error bars over at least tens of independent noise realizations and several network realizations, and a direct comparison with the L2-based alternative suggested above. Without such evidence, the claims of 'robust' and 'generic' prediction in the abstract and conclusion are not supported.
[Section III, Eq. (6)] The sum over ω∈I_res is undefined: it is not specified whether the frequency grid is the DFT grid of the finite-time driving signal, a fixed equidistant grid, or a continuum in the limit. If the DFT grid is used, the spacing 2π/T enters the DVI values and the index must be normalized accordingly. The authors should define DVI as an integral over I_res (or fix a grid convention and show that the node ranking is invariant to grid refinement), so that the reported numbers are reproducible.

minor comments (4)

[Section III, paragraph before Eq. (6)] The sentence 'the influence of future remains unknown' appears to contain a typo; it should be rephrased to 'the influence of future fluctuations remains unknown' or similar.
[Section III, Eq. (6)] The relation ε(ω) ∝ S(ω)^(1/2) should state whether S(ω) is a one-sided or two-sided power spectral density, since this affects the interpretation of the sum.
[Section II, Eqs. (4)-(5)] The units of ω, λ^(l), and α should be stated explicitly; the text refers to frequencies in Hz through ω/π, but the eigenfrequencies in Eq. (4) are given in angular units. Clarifying this would help reproducibility.
[Figure 3] The color-coding of PSD exponent b in panels (a) and (b) is mentioned in the caption but is not visually obvious in the printed version; a legend or explicit labels would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DVI ranking is tested against direct nonlinear simulation, and no fitted parameter or self-cited uniqueness claim enters the target.

full rationale

The paper's derivation chain is self-contained against an external benchmark. DVI in Eq. (6) is defined from the linear response eigenmodes of Eq. (3) and the imposed power spectral density S(ω), and it is then compared with the maximum response max_t |dot Theta_i(t)| obtained from direct numerical simulation of the nonlinear model in Eq. (1). That benchmark does not incorporate DVI, and no parameter is fitted to the ranking being predicted. The only self-citation, Ref. [12], supplies the linear response formula (3); it is a separately published, parameter-free theoretical result and does not assume the DVI ranking. The assumption stated immediately after Eq. (6) — that the real part of the sum of complex responses approaches the sum of magnitudes for sufficiently long time series — is a soundness/accuracy approximation, not a circular step: it may be invalid for random Fourier phases, but that is a correctness risk, not a reduction of the prediction to its input. No load-bearing step reduces to a fit, a definitional tautology, or a self-citation chain, so the circularity score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

DVI is built from the linearized dynamics and the driving PSD. The only truly ad hoc element is the phase-incoherent sum approximation. No parameter is fitted to the target ranking; the frequency-grid resolution used to evaluate Eq. (6) is unspecified and is the primary hidden implementation choice.

free parameters (1)

frequency sampling grid for Eq. (6)
The DVI sum over the resonance band I_res requires a discrete set of frequencies, but the paper never specifies the grid resolution or normalization. The ranking may depend on this hidden implementation choice.

assumptions (5)

domain assumption Linearization around a stable fixed point accurately captures the driven network dynamics (Eq. 2).
DVI is built on the linearized Eq. (2) around theta*. It is valid only for sufficiently small perturbations; the paper does not quantify the amplitude range over which the ranking remains valid.
standard math The graph Laplacian has real eigenvalues and orthogonal eigenvectors used in Eq. (3).
Spectral decomposition of the weighted graph Laplacian is standard linear algebra and is used to express the frequency response as a sum over eigenmodes.
ad hoc to paper The maximum time-domain response to a noisy signal equals the sum of the per-frequency response magnitudes.
Stated explicitly after Eq. (6). This is the load-bearing approximation that converts a random-phase superposition into a coherent sum; no proof or error estimate is given.
domain assumption The driving process is characterized by its power spectral density S(omega), with amplitudes epsilon(omega) proportional to S(omega)^(1/2) and random phases.
Used to construct DVI in Eq. (6) and to generate colored noise in Fig. 3. The PSD exponent b is varied, but phase information is discarded.
domain assumption The random growth network model [31] is representative of power grid topologies.
The full validation rests on a single synthetic network generated by the model of Ref. [31]; generality across real grid topologies is assumed but not demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dynamic Vulnerability in Oscillatory Networks and Power Grids." pith.science (2026). https://pith.science/paper/ZV5DWIKY

@misc{pith2026190800957,
  author       = {Pith},
  title        = {Pith review of: Dynamic Vulnerability in Oscillatory Networks and Power Grids},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZV5DWIKY}},
  note         = {Machine review of arXiv:1908.00957}
}

read the original abstract

Recent work found distributed resonances in driven oscillator networks and AC power grids. The emerging dynamic resonance patterns are highly heterogeneous and nontrivial, depending jointly on the driving frequency, the interaction topology of the network and the node or nodes driven. Identifying which nodes are most susceptible to dynamic driving and may thus make the system as a whole vulnerable to external input signals, however, remains a challenge. Here we propose an easy-to-compute Dynamic Vulnerability Index (DVI) for identifying those nodes that exhibit largest amplitude responses to dynamic driving signals with given power spectra and thus are most vulnerable. The DVI is based on linear response theory, as such generic, and enables robust predictions. It thus shows potential for a wide range of applications across dynamically driven networks, for instance for identifying the vulnerable nodes in power grids driven by fluctuating inputs from renewable energy sources and fluctuating power output to households.

Figures

Figures reproduced from arXiv: 1908.00957 by the authors.

**Figure 2.** FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 30 canonical work pages

[12]

Zhang, S

X. Zhang, S. Hallerberg, M. Matthiae, D. Witthaut, and M. Timme, Science Advances 5, eaav1027 (2019)

work page 2019
[1]

Hussain, C

F. Hussain, C. Gupta, A. J. Hirning, W. Ott, K. S. Matthews, K. Josic, and M. R. Bennett, Proc. Nat. Acad. Sci. 111, 972 (2014)

work page 2014
[2]

Jahnke, R.-M

S. Jahnke, R.-M. Memmesheimer, and M. Timme, PLoS Comput. Biol. 10, e1003940 (2014)

work page 2014
[3]

Kundur, N

P. Kundur, N. J. Balu, and M. G. Lauby, Power System Stability and Control, ser. The EPRI Power System En- gineering Series, Vol. 7 (McGraw-Hill New York, 1994)

work page 1994
[4]

Filatrella, A

G. Filatrella, A. H. Nielsen, and N. F. Pedersen, Eur. Phys. J. B 61, 485 (2008)

work page 2008
[5]

Rohden, A

M. Rohden, A. Sorge, M. Timme, and D. Witthaut, Phys. Rev. Lett. 109, 064101 (2012)

2012
[6]

A. E. Motter, S. A. Myers, M. Anghel, and T. Nishikawa, Nature Physics 9, 191 (2013)

2013
[7]

D¨ orﬂer, M

F. D¨ orﬂer, M. Chertkov, and F. Bullo, Proceedings of the National Academy of Sciences 110, 2005 (2013)

work page 2013

Show all 35 references

[8]

Coletta and P

T. Coletta and P. Jacquod, Physical Review E93, 032222 (2016)

2016
[9]

D. H. Zanette, Europhys. Lett. 68, 356 (2004)

2004
[10]

D. H. Zanette, Eur. Phys. J. B 43, 97 (2005)

2005
[11]

Tyloo, T

M. Tyloo, T. Coletta, and P. Jacquod, Phys. Rev. Lett. 120, 084101 (2018)

2018
[13]

Haehne, K

H. Haehne, K. Schmietendorf, S. Tamrakar, J. Peinke, and S. Kettemann, Physical Review E 99, 050301 (2019)

2019
[14]

Sch¨ afer, M

B. Sch¨ afer, M. Matthiae, X. Zhang, M. Rohden, M. Timme, and D. Witthaut, Phys. Rev. E 95, 060203 (2017)

2017
[15]

Kettemann, Phys

S. Kettemann, Phys. Rev. E 94, 062311 (2016)

2016
[16]

Tamrakar, M

S. Tamrakar, M. Conrath, and S. Kettemann, Sci. Rep. 8 (2018)

2018
[17]

Manik, M

D. Manik, M. Rohden, H. Ronellenﬁtsch, X. Zhang, S. Hallerberg, D. Witthaut, and M. Timme, Physical Review E 95, 012319 (2017)

2017
[18]

Sch¨ afer, C

B. Sch¨ afer, C. Beck, K. Aihara, D. Witthaut, and M. Timme, Nature Energy 3, 119 (2018)

2018
[19]

Tyloo, R

M. Tyloo, R. Delabays, and P. Jacquod, Phys. Rev. E 99, 062213 (2019)

2019
[20]

Hindes, P

J. Hindes, P. Jacquod, and I. B. Schwartz, arXiv:1904.12174 (2019)

2019 arXiv
[21]

Tyloo, L

M. Tyloo, L. Pagnier, and P. Jacquod, arXiv:1810.09694 (2018)

2018 arXiv
[22]

(null)">(null)</latexit><latexit sha1_base64=

and estimated via the nodal variance [23]. Yet, it is still unclear which stochastic signals may cause network- wide response patterns and how to quickly and precisely identify those nodes that potentially exhibit most severe responses and thus are most vulnerable to such pert...

1908 arXiv
[23]

S. Auer, F. Hellmann, M. Krause, and J. Kurths, Chaos: An Interdisciplinary Journal of Nonlinear Science 27, 127003 (2017)

2017
[24]

Plietzsch, S

A. Plietzsch, S. Auer, J. Kurths, and F. Hellmann, arXiv:1903.09585 (2019)

2019 arXiv
[25]

Arianos, E

S. Arianos, E. Bompard, A. Carbone, and F. Xue, Chaos: An Interdisciplinary Journal of Nonlinear Science 19, 013119 (2009)

2009
[26]

Simonsen, L

I. Simonsen, L. Buzna, K. Peters, S. Bornholdt, and D. Helbing, Physical Review Letters 100, 218701 (2008)

2008
[27]

Albert, I

R. Albert, I. Albert, and G. L. Nakarado, Physical re- view E 69, 025103 (2004)

2004
[28]

Song and M

H. Song and M. Kezunovic, in 2005 IEEE/PES Trans- mission & Distribution Conference & Exposition: Asia and Paciﬁc (IEEE, 2005) pp. 1–7

2005
[29]

G. Qu, J. Rudraraju, R. Modukuri, S. Hariri, and C. S. Raghavendra, in IASTED International Conference on Communications, Internet, and Information Technology, November 18-20, 2002, St. Thomas, US Virgin Islands, USA (2002) pp. 289–294

2002
[30]

Guti´ errez, F

R. Guti´ errez, F. Del-Pozo, and S. Boccaletti, PloS one 6, e20236 (2011)

2011
[31]

Tyloo and P

M. Tyloo and P. Jacquod, arXiv:1905.03582 (2019)

2019 arXiv
[32]

Schultz, J

P. Schultz, J. Heitzig, and J. Kurths, Eur. Phys. J. Spec. Top. 223, 2593 (2014)

2014
[33]

Anvari, G

M. Anvari, G. Lohmann, M. W¨ achter, P. Milan, E. Lorenz, D. Heinemann, M. R. R. Tabar, and J. Peinke, New J. Phys. 18, 063027 (2016)

2016
[34]

Diaconis and R

P. Diaconis and R. L. Graham, J. Roy. Stat. Soc. Ser. B , 262 (1977)

1977
[35]

P. J. Menck, J. Heitzig, J. Kurths, and H. J. Schellnhu- ber, Nature Comm. 5, 3969 (2014)

2014

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [12]

Zhang, S

X. Zhang, S. Hallerberg, M. Matthiae, D. Witthaut, and M. Timme, Science Advances 5, eaav1027 (2019)

work page 2019

[2] [1]

Hussain, C

F. Hussain, C. Gupta, A. J. Hirning, W. Ott, K. S. Matthews, K. Josic, and M. R. Bennett, Proc. Nat. Acad. Sci. 111, 972 (2014)

work page 2014

[3] [2]

Jahnke, R.-M

S. Jahnke, R.-M. Memmesheimer, and M. Timme, PLoS Comput. Biol. 10, e1003940 (2014)

work page 2014

[4] [3]

Kundur, N

P. Kundur, N. J. Balu, and M. G. Lauby, Power System Stability and Control, ser. The EPRI Power System En- gineering Series, Vol. 7 (McGraw-Hill New York, 1994)

work page 1994

[5] [4]

Filatrella, A

G. Filatrella, A. H. Nielsen, and N. F. Pedersen, Eur. Phys. J. B 61, 485 (2008)

work page 2008

[6] [5]

Rohden, A

M. Rohden, A. Sorge, M. Timme, and D. Witthaut, Phys. Rev. Lett. 109, 064101 (2012)

2012

[7] [6]

A. E. Motter, S. A. Myers, M. Anghel, and T. Nishikawa, Nature Physics 9, 191 (2013)

2013

[8] [7]

D¨ orﬂer, M

F. D¨ orﬂer, M. Chertkov, and F. Bullo, Proceedings of the National Academy of Sciences 110, 2005 (2013)

work page 2013