Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Fluctuations in email size modeled using a gamma-like distribution

T0 review · 3 major / 5 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 The gamma-like distribution pGL(s) fits observed email size fluctuations better than the earlier log-normal-like model, with lower least-squares errors in both direct fitting and size-generation simulation.

desk verdict A small, honest incremental paper that swaps a normal for a gamma distribution in an email-size model and reports a better fit, but the improvement is not statistically substantiated. read the letter →

arxiv 2506.03500 v1 pith:QH2F7MI6 submitted 2025-06-04 physics.soc-ph

Yoshitsugu Matsubara This is my paper
classification physics.soc-ph PACS 89.20.Hh05.40.-a
keywords emailsizegamma-likedistributionlog-normal-likepower-lawfluctuationsgenerationmodelfrequencyMIMEcontenttypes
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 is trying to establish that replacing the normal distribution for single-word length with a gamma distribution improves the fit of an email-size frequency model. The improved model, called gamma-like, is compared with the earlier log-normal-like model using the same observed email-size data from a university mail system. The gamma-like model yields smaller least-squares fitting errors in both the direct fit to observed frequencies and the fit of sizes generated by the corresponding generation model. If the improvement is real, it offers a modestly more accurate statistical description of email-size fluctuations, which matters for understanding the load that email services must handle.

What carries the argument

The central object is the gamma-like density pGL(s) = 1/a * 1/(s ln s) * 1/$\theta$^k * (ln ln s)^{k-1} * exp(-ln ln s / $\theta$), defined for s > e, which is obtained by substituting a gamma distribution for the normal distribution inside the earlier log-normal-like construction. It is paired with the size generation model s_t = b s_{t-1}^c $e^{{epsilon_t}}$, where epsilon_t, the length of a single word, is now drawn from a gamma distribution (k, $\theta$) instead of a normal distribution N(mu, $sigma^{2}$). The gamma distribution removes the possibility of negative word lengths while preserving the power-law tail of the generated sizes.

What would settle it

Count single-word lengths in bytes from a privacy-preserving corpus of actual email bodies and fit a gamma distribution to them; if the fitted shape and scale are far from k = 24.1803 and theta = 0.0489, or if a gamma fit is poor, then the generative mechanism behind the improved D values is unsupported.

Watch

Extended reading notes

Core claim

The paper's central claim is that the gamma-like probability density pGL(s), formed by combining a gamma distribution with the transformation x = ln ln s, fits the observed frequency distribution of 'no attachment' email sizes better than the log-normal-like density pLNL(s). For the direct fit to observed data, the least-squares error is D = 70.3002 for pGL versus D = 88.8502 for pLNL. For the size generation model, which produces 191,993 simulated emails matching the observed count, the gamma-based model gives D = 13.99222 versus D = 15.19232 for the log-normal-based model. Both models preserve the same asymptotic power-law behavior for large email sizes, so the improvement is concentrated in the small-size range where the earlier model's normal assumption allowed negative word lengths.

Load-bearing premise

The model assumes that the length of a single word in an email follows a gamma distribution, and this assumption is never checked against actual email content, which the paper says it did not analyze.

Editorial extensions

If this is right

  • The gamma-like density pGL(s) is claimed to be a better empirical description of 'no attachment' email-size frequencies than the log-normal-like density, with D = 70.3002 versus 88.8502.
  • The gamma-based generation model is claimed to reproduce observed email-size frequencies more closely, with D = 13.99222 versus 15.19232 over 191,993 simulated emails.
  • Because both models have the same asymptotic power-law behavior, the improved fit does not change the operational implication that very large emails are rare but should still be considered when setting email size limits.
  • The gamma substitution fixes the formal flaw of negative single-word lengths in the earlier generation model, making the generative mechanism more plausible at the level of word structure.

Reading between the lines

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

  • If real single-word lengths are indeed gamma distributed, the nested exponentials in the generation model imply that sentence and compound-sentence lengths inherit heavier tails; this could be checked against sentence-length corpora for Japanese and English.
  • The improvement in D is modest and the paper reports no uncertainty estimates, so a reader should treat the advantage as suggestive until a model-selection criterion such as AIC or BIC, or a holdout-data comparison, confirms it.
  • The same substitution of a gamma for a normal distribution in the innermost exponential could be applied to other log-normal-like size distributions of text or file sizes, offering a direct way to test whether the improvement generalizes beyond email data.
  • Since pGL(s) has an extra shape parameter k compared with pLNL(s), part of the improved fit may simply reflect added flexibility rather than a deeper generative mechanism.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes a gamma-like probability density pGL(s) for email size fluctuations, replacing the log-normal-like distribution pLNL(s) used in earlier work. The model is intended to improve fits in the small-size range by assuming that the length of a single word follows a gamma distribution rather than a normal distribution, thereby avoiding negative word lengths. The author fits both models to the size-frequency distribution of 'no attachment' emails and to a simulated email size generation model, reporting smaller least-squares discrepancy D for pGL in both cases. The central claim is that pGL provides a significantly improved fit over pLNL.

Significance. If the improved fit is real and statistically validated, the paper would provide a modestly better empirical description of email size fluctuations and a more plausible generative mechanism. However, the manuscript's significance is limited by the absence of any uncertainty quantification, model-selection test, or out-of-sample validation, and the generative mechanism is explicitly not tested against email content. The paper is honest in stating these limitations, which is a strength, but the central quantitative claim currently rests on raw point estimates of D with no error bars.

major comments (3)
  1. [Section 4, D values] The claim that pGL fits better than pLNL is based solely on the reported D differences (70.3002 vs. 88.8502 for the direct fit and 13.99222 vs. 15.19232 for the generation model). No standard errors, confidence intervals, or significance tests are provided for D or for the difference between the two models. Because the same binned observed data are used both to estimate parameters by least squares and to compute the winning D, the comparison has no protection against sampling noise or overfitting. Moreover, pGL and pLNL are non-nested functional forms, so a raw D difference cannot be interpreted without a null distribution. The word 'significantly' in the abstract is therefore unsupported.
  2. [Section 4, generation model simulation] The simulation based on the generation model is not an independent test of the gamma-like distribution. The parameters used for pGL (k=24.1803, theta=0.0489) are exactly those obtained from the direct fit to the observed distribution, so the simulation reproduces the fitted distribution rather than validating it. Additionally, the pLNL generation parameters (mu=1.259, sigma=0.235) differ from the direct-fit pLNL parameters (mu=0.0489, sigma=0.2461) without any explanation; if the pLNL simulation is not optimized, the reported D gap of 15.19232 vs. 13.99222 may reflect suboptimal pLNL parameters rather than a genuine advantage of pGL.
  3. [Section 3, Eq. (3) and the gamma word-length assumption] The model's generative mechanism rests on the assumption that the length of a single word epsilon_t follows a gamma distribution, introduced to fix the negative-length problem of the normal distribution. No empirical evidence is provided that actual word lengths in the analyzed emails are gamma-distributed; the conclusion explicitly states that email contents were not analyzed. If the word-length distribution is not gamma, the improved fit of pGL is a curve-fitting artifact rather than support for the proposed mechanism. This should be acknowledged more explicitly and ideally tested on auxiliary data or at least framed as an untested assumption.
minor comments (5)
  1. [Section 3, Eq. (3)] The gamma density in Eq. (3) is written as (1/theta^k)(ln ln s)^(k-1) exp(-ln ln s/theta) without the factor 1/Gamma(k); presumably this normalization is absorbed into the constant a, but the expression is not a normalized density as written and this should be clarified.
  2. [Conclusion] The text refers to 'the size–frequency distribution pLGL' which appears to be a typo for 'pGL’; correct this notation.
  3. [Section 4] The definition of D is typeset awkwardly ("D = ∑_s(ln yO(s) − ln y(s))^2" with a superscript 2 after the sum). Please render the equation clearly and number it explicitly.
  4. [Abstract and Introduction] Several typos interrupt the reading: 'wh ich' in the abstract, 's ending' in the introduction, and 'modeled using a gamma-like distribution' in the title is fine but the abstract contains 'sending mail' which is acceptable. A light proofreading pass is recommended.
  5. [Figure 1 caption] The caption refers to 'sGL(s)' but should be 'pGL(s)'.

Circularity Check

1 steps flagged · score 6.0 of 10

The gamma-like model's 'prediction' is a Monte Carlo redraw of parameters fitted to the same observed distribution; the improved-fit claim is in-sample.

  1. fitted input called prediction [Section 4 (Discussion), Figures 1-2 and the simulation paragraph; direct fit parameters carried into the generation simulation.]
    "The degree of fitting in pLNL was D = 88.8502 (µ = 0.0489, σ = 0.2461), whereas that in pGL was D = 70.3002 (k = 24.1803, θ = 0.0489). ... The parameter values of pLNL were µ = 1.259, σ = 0.235, whereas those of pGL were k = 24.1803, θ = 0.0489. The degrees of fit were D = 15.19232 and 13.99222 for pLNL and pGL, respectively, indicating that the frequency distribution of the gamma-like model, pGL, is a better fit than that of the log-normal-like model, pLNL."

    The pGL generation simulation uses k = 24.1803, θ = 0.0489, which are exactly the least-squares estimates obtained by fitting pGL to the observed 'no attachment' size distribution in the same section (D = 70.3002). The simulated distribution is then scored with D against that same observed histogram. This is not an out-of-sample prediction: the simulation is a Monte Carlo draw from a model already calibrated to minimize D on the target data, so its lower D (13.99222 vs 15.19232) largely inherits the fitted advantage of pGL rather than independently confirming it. The direct fit is also evaluated on the fitting data, and no held-out subset, standard error, or null distribution for the D difference is reported.

full rationale

The central new model pGL(s) is defined as a gamma density on x = ln ln s, while pLNL(s) is a lognormal density on the same x. Replacing the normal law for the per-word innovation ε_t by a gamma law is a legitimate parametric modification, and the direct least-squares comparison (D = 70.3002 vs 88.8502) is an ordinary in-sample goodness-of-fit comparison between two two-parameter, non-nested families. That part is not circular. The circularity burden comes from the 'email size generation model' simulation in Section 4: the pGL parameters fed into st are the same k and θ values obtained by fitting pGL to the observed 'no attachment' distribution, and the simulated output is then compared with that same observed distribution. Thus the simulated D values are in-sample, not predictive; the simulation cannot validate the model beyond the fit. The abstract's word 'significantly' is also unsupported by any uncertainty estimate (no SE/CI, no null distribution for the D difference, non-nested models), and the paper itself concedes that the email contents were not analyzed (Conclusion), so the gamma word-length mechanism is an assumption rather than an empirically established generative cause. These are correctness and limitation concerns, not additional circular steps. The baseline log-normal-like model and data come from the author's prior papers ([13], [15]), but invoking those prior models as a starting point is normal self-citation and is not load-bearing for the new gamma variant. Overall, the main quantitative claim is a fitting result with one re-used in-sample simulation presented as if it were a generated prediction: partial circularity, score 6.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The model rests on a distributional assumption for word lengths that is chosen to avoid negative values, on the transformation x=ln ln s, and on fitted shape parameters. The observed data are from a single organization over three months and are not public; the generation model also depends on unknown coefficients b and c.

free parameters (7)
  • mu_LNL_direct_fit = 0.0489
    Location parameter of the log-normal-like density pLNL(s) fitted to observed email size frequencies in Section 4.
  • sigma_LNL_direct_fit = 0.2461
    Shape parameter of pLNL(s) fitted to observed email size frequencies in Section 4.
  • k_GL = 24.1803
    Shape parameter of the gamma-like density pGL(s) fitted in Section 4.
  • theta_GL = 0.0489
    Scale parameter of pGL(s) fitted in Section 4.
  • mu_LNL_generation = 1.259
    Mean of the normal word-length distribution used in the size generation simulation for pLNL (Section 4).
  • sigma_LNL_generation = 0.235
    Standard deviation of the normal word-length distribution used in the generation simulation for pLNL (Section 4).
  • b_and_c_in_size_generation_model = not reported
    Coefficients in Eq. 2, st = b * s_{t-1}^c * exp(epsilon_t), are needed to simulate email sizes but their fitted values are not stated.
assumptions (4)
  • domain assumption Email body size is generated by a multiplicative cascade st = b * s_{t-1}^c * exp(epsilon_t) where epsilon_t is a single-word length.
    This relation is stated in Eq. 2 and is the mechanism connecting word length to email size, but it is not tested against actual email content in this paper.
  • ad hoc to paper The length of a single word follows a gamma distribution.
    Section 3 changes epsilon_t from normal to gamma solely because normal allows negative lengths; no empirical data on word lengths are presented.
  • standard math The transformation x = ln ln s maps the gamma density into the email-size density pGL(s).
    This is a standard change of variables, used in Eq. 3; it is mathematically valid for s > e.
  • domain assumption The observed email size frequencies from one organization, May to July 2015, are representative of email size statistics.
    The data come from a single university mail server over three months and are not public; generalizing beyond this population is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fluctuations in email size modeled using a gamma-like distribution." pith.science (2026). https://pith.science/paper/QH2F7MI6

@misc{pith2026250603500,
  author       = {Pith},
  title        = {Pith review of: Fluctuations in email size modeled using a gamma-like distribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QH2F7MI6}},
  note         = {Machine review of arXiv:2506.03500}
}
read the original abstract

A previously established frequency distribution model, which integrates a lognormal distribution with a logarithmic equation, effectively characterizes fluctuations in email size during sending requests. In addition, an email size generation model has been developed based on this log-normal-like framework. While the fitting of these models has been deemed satisfactory, they can be further enhanced in the range of small email sizes. This study advances these models by incorporating a gamma distribution alongside a logarithmic equation. The resulting gamma-like model demonstrates a significantly improved fit compared with the log-normal-like model. These results contribute to the knowledge on the statistical properties of sending mail.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 21 canonical work pages

  1. [1]

    Faloutsos, P

    M. Faloutsos, P. Faloutsos, C. Faloutsos, On power–law relation ships of the internet topology. SIGCOMM Comput. Commun. Rev. 29(4), 251–262 (1999). https://doi.org/10.1145/316194.316229

    work page arXiv 1999

  2. [2]

    no attachment

    V. Paxson, S. Floyd, Wide area traffic: the failure of Poisson mode ling. IEEE/ACM Trans. Networking 3, 226–244 (1995). https://doi.org/10.1109/90. 5 10-10 10-8 10-6 10-4 10-2 100 101 102 103 104 105 106 Observed data Log-normal-like model Gamma-like model Relative frequency of number of send requests s: Email size (100 B) Fig. 2 Size-frequency distribution...

    doi:10.1109/90 1995

  3. [3]

    Csabai, 1/f noise in computer network traffic

    I. Csabai, 1/f noise in computer network traffic. J. Phys. A: Mat h. Gen. 27(12), L417 (1994). https://doi.org/10.1088/0305-4470/27/12/004

    work page doi:10.1088/0305-4470/27/12/004 1994

  4. [4]

    Takayasu, H

    M. Takayasu, H. Takayasu, T. Sato, Critical behaviors and 1 /f noise in infor- mation traffic. Physica A 233, 824–834 (1996). https://doi.org/10.1016/S0378- 4371(96)00189-6

    doi:10.1016/s0378- 1996

  5. [5]

    Tadaki, Power-law fluctuation in internet traffic

    S. Tadaki, Power-law fluctuation in internet traffic. J. Phys. Soc . Jpn. 76(3), 044001–044001–5 (2007). https://doi.org/10.1143/jpsj.76.044001

    work page doi:10.1143/jpsj.76.044001 2007

  6. [6]

    Eckmann, E

    J.P. Eckmann, E. Moses, D. Sergi, Entropy of dialogues creates coherent struc- tures in e-mail traffic. Proc. Nattl. Acad. Sci. U.S.A. 101(40), 14333–14337 (2004). https://doi.org/10.1073/pnas.0405728101 6

    work page doi:10.1073/pnas.0405728101 2004

  7. [7]

    Barab´ asi, The origin of bursts and heavy tails in human dynam ics

    A.L. Barab´ asi, The origin of bursts and heavy tails in human dynam ics. Nature 435, 207–211 (2005). https://doi.org/10.1038/nature03459

    doi:10.1038/nature03459 2005

  8. [8]

    Goh, A.L

    K.I. Goh, A.L. Barab´ asi, Burstiness and memory in complex syste ms. EPL (Europhys. Lett.) 81(4), 48002 (2008). https://doi.org/10.1209/0295-5075/81/ 48002

    work page doi:10.1209/0295-5075/81/ 2008

Show all 28 references
  1. [9]

    Malmgren, D.B

    R.D. Malmgren, D.B. Stouffera, A.E. Motter, L.A.N. Amaral, A Poisso nian expla- nation for heavy tails in e-mail communication. Proc. Natl. Acad. Sci. U.S.A. 105(47), 18153–18158 (2008). https://doi.org/10.1073/pnas.0800332105

    2008 doi

  2. [10]

    Anteneodo, R.D

    C. Anteneodo, R.D. Malmgren, D.R. Chialvo, Poissonian bursts in e -mail corre- spondence. Eur. Phys. J. B 75, 389–394 (2010). https://doi.org/10.1140/epjb/ e2010-00139-9

    2010 doi

  3. [11]

    Karsai, K

    M. Karsai, K. Kaski, A.L. Barab´ asi, J. Kert´ esz, Universal fe atures of correlated bursty behaviour. Sci. Rep. 2, 397 (2012). https://doi.org/10.1038/srep00397

    2012 doi

  4. [12]

    Matsubara, Y

    Y. Matsubara, Y. Hieida, S. Tadaki, Fluctuation in e-mail sizes we akens power- law correlations in e-mail flow. Eur. Phys. J. B 86, 209 (2013). https://doi.org/ 10.1140/epjb/e2013-40209-x

    2013 doi

  5. [13]

    Matsubara, Y

    Y. Matsubara, Y. Musashi, Fluctuations in email size. Eur. Phys . J. Plus 132, 507 (2017). https://doi.org/10.1140/epjp/i2017-11767-2

    2017 doi

  6. [14]

    Milojeviˇ c, Modes of collaboration in modern science: Beyond power laws and preferential attachment

    S. Milojeviˇ c, Modes of collaboration in modern science: Beyond power laws and preferential attachment. Journal of the American Society for I nformation Science and Technology 61(7), 1410–1423 (2010). https://doi.org/10.1002/asi.21331

    2010 doi

  7. [15]

    Matsubara

    Y. Matsubara. Fluctuations in the email size modeled by a log-nor mal-like distribution (2025). https://doi.org/10.48550/arXiv.2501.04042

    2025 doi arXiv

  8. [16]

    Freed, D.N.S

    N. Freed, D.N.S. Borenstein. Multipurpose Internet Mail Exten sions (MIME) Part One: Format of Internet Message Bodies. RFC 2045 (1996). https://doi. org/10.17487/RFC2045

    1996 doi

  9. [17]

    Freed, D.N.S

    N. Freed, D.N.S. Borenstein. Multipurpose Internet Mail Exten sions (MIME) Part Two: Media Types. RFC 2046 (1996). https://doi.org/10.17487/RFC2046

    1996 doi

  10. [18]

    K. Moore. MIME (Multipurpose Internet Mail Extensions) Part Three: Message Header Extensions for Non-ASCII Text. RFC 2047 (1996). https://doi.org/10. 17487/RFC2047

    1996

  11. [19]

    Freed, D.N.S

    N. Freed, D.N.S. Borenstein. Multipurpose Internet Mail Exten sions (MIME) Part Five: Conformance Criteria and Examples. RFC 2049 (1996). https://doi. org/10.17487/RFC2049 7

    1996 doi

  12. [20]

    Klensin, N

    D.J.C. Klensin, N. Freed. Multipurpose Internet Mail Extensions (MIME) Part Four: Registration Procedures. RFC 4289 (2005). https://doi.org/10.17487/ RFC4289

    2005

  13. [21]

    Freed, D.J.C

    N. Freed, D.J.C. Klensin, T. Hansen. Media Type Specifications an d Registration Procedures. RFC 6838 (2013). https://doi.org/10.17487/RFC6838

    2013 doi

  14. [22]

    Arai, Sentence length and lognormal distribution : A case stu dy of akutagawa and dazai

    H. Arai, Sentence length and lognormal distribution : A case stu dy of akutagawa and dazai. Hitotsubashi Rev. 125(3), 205–223 (2001). http://hermes-ir.lib.hit- u.ac.jp/rs/handle/10086/10418 (In Japanese)

    2001

  15. [23]

    Furuhashi, Y

    S. Furuhashi, Y. Hayakawa, Lognormality of the distribution of japanese sentence lengths. J. Phys. Soc. Jpn. 81(3), 034004 (2012). https://doi.org/10.1143/JPSJ. 81.034004

    2012 doi

  16. [24]

    Sasaki, Distribution of sentence-length

    K. Sasaki, Distribution of sentence-length. Math. Linguist. 78, 13–22 (1976). (In Japanese)

    1976

  17. [25]

    Ishida, K

    M. Ishida, K. Ishida, On distributions of sentence lengths in Jap anese writing. Glottometrics 15, 28–44 (2007). URL https://api.semanticscholar.org/CorpusID: 12215774

    2007

  18. [26]

    Borb´ ely, A

    G. Borb´ ely, A. Kornai. Sentence length (2019). https://doi.org/10.48550/arXiv. 1905.09139

    2019 doi arXiv

  19. [27]

    J. Liu, N. Yang, H. Liu, Distribution of sentence length of english complex sen- tences. Moderna Spr ˚ ak118(3), 51–69 (2024). https://doi.org/10.58221/mosp. v118i3.15574

    2024 doi

  20. [28]

    Milojevi´ c, Power law distributions in information science: Mak ing the case for logarithmic binning

    S. Milojevi´ c, Power law distributions in information science: Mak ing the case for logarithmic binning. J. Am. Soc. Inf. Sci. Technol. 61(12), 2417–2425 (2010). https://doi.org/10.1002/asi.21426 8

    2010 doi

Pith tools

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