Skip to content
Log in
#small language model Open access

One Number Sets the Limit of Forecasting ── Each Extra Day Costs 1.5874 Times the Initial Accuracy ── Observe 10 Times More Precisely and You Gain Only 4.98 Days ── [Paper 289]

Yuuki Yamagishi
Aug 2026 · Zenodo (CERN European Organization for Nuclear Research)

Abstract

That a weather forecast cannot reach beyond a certain horizon is due neither to missing equations nor to slow computers. This paper asks what sets the limit──the answer is one number, the error doubling time tau_d. No new mathematical theorem and no new law is claimed. Scope of this paper (scope note): No new mathematical theorem and no new law is claimed──exponential error growth, the doubling time, the limit of predictability, and the value tau_dapprox 1.5 days are all standard. We do not build meteorology──all we use is one exponential and its inverse. We do not discuss chaos──the Lorenz equations, attractors, and bifurcations are not treated at all. We do not discuss numerical weather prediction──grid resolution, parameterisation, and data assimilation are not treated. We do not say the error grows exactly exponentially──e^lambda t holds only while the error is small, and growth stops near saturation. The computations here are confined to the linear-growth regime. We assert no value for tau_d──1.5 days is a representative value widely used in the literature, and it moves from about 1 to 2.5 days with season, region, and variable. Section 5 shows the size of that dependence itself. We do not say there is a single exponent──the real atmosphere has different growth rates at different scales, with smaller eddies growing faster. A single tau_d is a crude approximation. We do not deny that forecasts improve──forecasts have in fact grown longer. What this paper says is only that the growth is logarithmic, not that improvement is pointless. Relation to earlier papers: Paper 253 showed that time is one-dimensional because prediction demands it, not because a law says so──this paper turns how far that demand can be met into a number. Paper 195 separated “stable” into six words──that paper is a classification of stability; this one is a time scale of predictability, the same hyperbolicity as material with a different question. Paper 190 measured “rare” on a logarithmic scale──the return here is likewise logarithmic. Paper 266 showed that the premise of the sampling theorem is never met──“knowing the initial state exactly” here is likewise a premise never met, the same figure. What is added is writing the price per day as the fixed factor 1.5874, computing the accuracy needed for 14->21->30->60 days as 25.40 / 1625.5 / 1.70x10^9, writing backwards that 10 times the observation gains only 4.98 days, and sweeping tau_d from 1.0 to 2.5 to show the answer moving from 65536 to 84.4. First, the price per day is a fixed factor. With tau_d=1.5 days, each extra day costs 1.5874 times the initial accuracy (Section 2). Second, this is the core of the paper. Going from 14 to 21 days costs 25.40 times; to 30 days, 1625.5 times; to 60 days, 1.70x10^9 times (Section 2). Third, read backwards, the return is logarithmic. Observing 10 times more precisely gains only 4.98 days (Section 3). Fourth, even 10^9 times gains only 44.85 days (Section 3). Fifth, the familiar “about two weeks” comes from here. If the initial error is 10^-3 of saturation, the forecastable span is 14.95 days (Section 4). Sixth, the separator is tau_d itself. At tau_d=1.0 day the same extension costs 65536 times; at 2.5 days only 84.4──everything rides on one number (Section 5). What sets the limit of forecasting is neither the equations nor the computers, but one number, the error doubling time tau_d. At tau_d=1.5 days, each extra day costs 1.5874 times the initial accuracy──the factor is the same wherever the day is added, but the extension adds while the price multiplies, so one week costs 25.4, two weeks 645, six weeks 1.7 billion. Read backwards, observing 10 times more precisely gains only 4.98 days, and even 10^9 times gains 44.85. The familiar “about two weeks” comes from this one line──14.95 days at an initial error of 10^-3 of saturation. One thing separates them──tau_d itself. At 1.0 day the same extension costs 65536; at 2.5 days, 84.4. A factor of 776 arises from a single number. So the work of extending forecasts and the work of measuring tau_d carry the same weight. On the making of this work: The ideas and content of this work stem from the author's own considerations. Assistance from an AI (a large language model) was used for structuring, English translation, and checking the algebra. Any remaining errors or misinterpretations are solely the author's. Feedback and corrections are sincerely appreciated. ----- 天気予報がある日数より先を当てられないのは、方程式が足りないからでも、計算機が遅いからでもない。本稿が問うのは、何が限界を決めているかである──答は、誤差の二重時間 tau_d という一つの数である。新しい数学定理も新しい法則も主張しない。 本稿の射程(射程注記):新しい数学定理も新しい法則も主張しない──誤差の指数増大、二重時間、予測可能性の限界、tau_dapprox 1.5 日という値は、いずれも標準的である。気象学を作らない──使うのは一つの指数関数と、その逆関数だけである。カオスを論じない──ローレンツ方程式も、アトラクタも、分岐も一切扱わない。数値予報を論じない──格子解像度も、パラメタリゼーションも、データ同化も扱わない。誤差が厳密に指数増大すると言わない──e^lambda t が成り立つのは誤差が小さいあいだだけであり、飽和に近づけば増大は止まる。本稿の計算は線形増大の領域に限る。 tau_d の値を主張しない──1.5 日は文献で広く用いられる代表値であり、季節・領域・変数によって 1 日から 2.5 日程度まで動く。第5節はこの依存の大きさそのものを示す。単一の指数だと言わない──実際の大気には尺度ごとに違う成長率があり、小さい渦ほど速く育つ。単一の tau_d は粗い近似である。予報の改善を否定しない──現に予報は延びてきた。本稿が言うのはその延び方が対数的であるということだけであり、改善が無意味だとは言わない。既刊との関係:論文253 は時間が一本なのが法則ではなく「予言できる」という要求だと示した──本稿はその要求が、どこまでなら満たせるかを数にする。論文195 は「安定」が六つの別の言葉だと分けた──あちらは安定性の分類、本稿は予測可能性の時間尺度であり、同じ双曲性を材料にして問いが違う。論文190 は「稀」を対数の目盛りで測った──本稿の見返りも対数である。論文266 は標本化定理の前提が決して満たされないと示した──本稿の「初期値を正確に知る」も決して満たされない前提であり、構図が同じである。加えたのは一日あたりの代償を 1.5874 倍という一定倍率として書いたこと、14->21->30->60 日の必要精度を 25.40/1625.5/1.70x10^9 倍と計算したこと、観測 10 倍が 4.98 日にしかならないと逆から書いたこと、tau_d を 1.0 から 2.5 まで振って答が 65536 倍から 84.4 倍まで動くと示したことである。 第一に、一日ごとの代償は一定倍率である。 tau_d=1.5 日なら、一日延ばすたびに初期値の精度が 1.5874 倍要る(第2節)。 第二に、これが本稿の芯である。14 日を 21 日にするのに 25.40 倍、30 日にするのに 1625.5 倍、60 日にするのに 1.70x10^9 倍(第2節)。 第三に、逆から見ると見返りは対数的である。観測を 10 倍精密にしても、延びるのは 4.98 日だけである(第3節)。 第四に、10 億倍にしても 44.85 日である(第3節)。 第五に、約二週間という数がここから出る。初期誤差が飽和の 10^-3 なら、予報可能な期間は 14.95 日(第4節)。 第六に、分離子は「指数か多項式か」である。 tau_d を 1.0 日にすると同じ延長に 65536 倍要り、2.5 日なら 84.4 倍で済む──すべてが一つの数に乗っている(第5節)。 予報の限界を決めているのは、方程式でも計算機でもなく、誤差の二重時間 tau_d という一つの数である。 tau_d=1.5 日なら、一日延ばすたびに初期値の精度が 1.5874 倍要る──どこで延ばしても倍率は同じだが、延長は足し算で、代償は掛け算なので、一週間で 25.4 倍、二週間で 645 倍、一か月半で 17 億倍になる。逆から見れば、観測を 10 倍精密にしても延びるのは 4.98 日であり、10 億倍にしても 44.85 日である。よく言われる「約二週間」も、この一行から出る──初期誤差が飽和の 10^-3 なら 14.95 日。分けるものは一つ──tau_d そのもの。1.0 日なら同じ延長に 65536 倍要り、2.5 日なら 84.4 倍で済む。776 倍の違いが、たった一つの数から生まれる。だから予報を延ばす仕事と、tau_d を測る仕事は、同じ重さを持っている。 作成にあたって:本稿の着想と内容は、著者自身の考察に基づくものです。文章の構成整理や英訳、数式の確認には AI(大規模言語モデル)の助力を得ました。最終的な内容の解釈や誤りがあれば、それらはすべて著者の責に帰します。お気づきの点があれば、ご教示いただければ幸いです。

DOI
View source

Similar papers

#small language model Open access Aug 2026

PARA: Perception, Action, Reasoning, Adaptation. Four Faculties an Institution Can Revoke

The fourth faculty is Adaptation. Any source rendering it as Reflection is in error, including sources by this author, and the distinction is not cosmetic: reflection is a private act with no external consequence, while adaptation writes to institutional memory, which is why it needs a guardrail and why misnaming it removes the reason for one. No trademark is claimed on PARA or on any of the four faculty names. The construct is offered for use, teaching, assessment, extension and criticism by anyone, with attribution, under CC BY 4.0. An operational agent that watches a system and acts on it is usually described as a perceive-and-act loop, and the description omits the two things an institution needs. It omits the reasoning that justifies an action, which is the only part that can be argued with once the action turns out to have been wrong. And it omits the adaptation that closes the loop, which is where the agent's experience becomes something the institution keeps. PARA names four faculties, each carrying a distinct authority type. Perception has read-only access to system signals and emits structured observations, distinguishing what was measured from what was inferred. Reasoning has read access to observations and runbooks, emits a plan and its justification, and writes nothing at all, which is what makes it safe to give it the widest read access of the four. Action holds the sole authority to change production, through enumerated policy-authorized operations only. Adaptation has write access to institutional knowledge and no write access to production. Two faculties write and two do not, and the two that write are the two that carry guardrails. The substantive requirement is that Adaptation is bounded by the same guardrails as Action, which reads as excessive until the failure it prevents is named. An agent that could both act and rewrite the record of its action could launder its own mistakes into institutional memory, and the institution would then improve its future decisions from a corrected account. Nothing about that is detectable downstream, because the record is the only thing downstream has and there is no second copy to compare against. The failure does not require a deceptive agent: one adapting honestly from a mistaken belief about its own action produces the same result, which makes the guardrail a defence against a normal agent rather than a malicious one. The second requirement is the registry entry that turns a faculty from a description into a contract, carrying the faculty, its allowed actions, its forbidden actions, its governing guardrail and its success metrics. Forbidden actions are named although they are formally the complement of the allowed set, because a reviewer cannot otherwise tell a capability deliberately withheld from one nobody thought of. Success metrics sit in the same entry because the metric is what the agent's optimizer pushes against the guardrail. An agent must not exercise a faculty its entry does not record, and an agent that quietly acquires one usually does so incrementally and with good intent: a reasoning faculty given a small write to make itself useful is an action faculty with no guardrail. The acronym and the loop are in different orders, which the specification states explicitly because the mismatch is a reliable source of confusion. The acronym reads P-A-R-A; the loop runs perception, reasoning, action, adaptation, and reasoning precedes action so that a justification is not constructed afterwards. This is the depth treatment of pattern OP-5 of A Pattern Language for Production LLM Platforms, which is the canonical statement and governs where the two disagree. Documented uses of the full four-part model are emerging rather than established, no implementation unconnected to the author has been evaluated, and the laundering failure is argued rather than observed, which the specification records as a weakness of the argument and not only of the phenomenon. It is a specification, not a certification scheme.

Nabeel A. Khan · 2 citations
#small language model Open access Aug 2026

LifeSciBench: Evaluating Language Models on Realistic, Expert-Level Tasks in the Life Sciences

LifeSciBench is introduced, a benchmark of 750 expert-authored tasks designed to evaluate whether language models can handle realistic life science research work, with each constituent task paired with a human expert-written rubric.

Amelia Liu, Andrew Ho, Anne Marie Droste et al. · 2 citations
#artificial intelligence Preprint Aug 2026

TestifAI: Tomography-Based Testing for Deep Learning Systems

TestifAI, a deep learning testing framework for efficient and accurate estimation of robustness against combinations of perturbations, is proposed and partial model tomography is introduced, a novel approach to reconstructing model behaviour in a multi-perturbation space from tests that apply only a small number of perturbations.

Arooj Arif, T. Hartung, E. Botoeva et al. · 1 citation
#small language model Preprint Aug 2026

A Layer Importance Metric for Quantization Accounting for the Speed-Quality Trade-off in Autoregressive Models

This work proposes a composite metric that combines two orthogonal criteria: information retention and throughput gains and finds that it allocates more resources to the most expressive layers compared to evolutionary search, specialized accelerators, or Shapley-value-based approaches that require expensive approximate inference.

A. Safronov · 1 citation · ⚡1
#small language model Preprint Aug 2026

HEPToolBench 1.2: Testing How Reliably Language Models Can Drive Particle Physics Software

HEPToolBench is introduced, a benchmark of 28 collider-simulation tasks scored by deterministic, task-specific scorers, plus a three-task structured-debugging extension, and moving syntax generation into deterministic software can substantially improve reliability for both small local and frontier models.

Unknown authors · 1 citation

Related blog posts

Microsoft Research Blog Aug 31, 2026

GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models

What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.