The parameter sheet: a defensible 2026 model of a LEO network參數表：一份站得住腳的 2026 年 LEO 網路模型
Build a dated parameter sheet for a 2026 LEO networking study. Identify filed, flown, measured and claimed values, together with parameters awaiting public evidence. Use these inputs to calculate an explicit end-to-end latency and capacity budget for one user.本章建立一份供 2026 年 LEO 網路研究使用的附日期參數表，區分申報值、實飛值、量測值與宣稱值，並列出仍待公開證據的參數。這些輸入用來計算單一使用者的端到端延遲與容量預算，讓讀者能追溯每項假設。
This chapter collects the numbers of Chapters 1, 2 and 3 into one table, so it uses their vocabulary throughout. These earlier terms appear most often below. Each chip links to its definition.
本章把第 1、2、3 章的數字收進同一張表，因此全章都使用它們的詞彙。下列較早定義的名詞在本章最常出現。點選即可回到它的定義。
New terms in this chapter (3)本章新名詞（3）
Each term is defined in the text where it first appears. The link jumps there.
每個名詞都在它第一次出現的段落中定義，點連結可跳過去。
- parameter sheet參數表 The single dated table of every constant an evaluation rests on, each with its value, provenance and date. 一份附日期的表，列出評估所依賴的每個常數及其數值、出處與日期。
- evidence tag證據標籤 A label attached to a number stating what kind of document it comes from and therefore how a reader can check it. 附在數字上的標記，說明它來自哪一類文件，因此讀者能用什麼方式查證。
- filing snapshot申報快照 The constellation geometry stated in one dated regulatory document, which later filings replace and which the papers of its year freeze. 某一份附日期監管文件中陳述的星座幾何，之後的申報會取代它，而當年的論文則把它凍結下來。
4.1Why a parameter sheet, and how to read the tags為什麼需要參數表，以及如何閱讀標籤
The first three chapters introduce each quantity alongside its physical or measurement basis. Simulation, emulation and analysis require these values together. A parameter sheet is a dated table of the constants supporting an evaluation. It includes shell geometry, laser rates, node delays, scheduler period and queue size.
Each value carries its source document and date. Terrestrial networking papers rarely need such a table because the hardware they model has data sheets. Satellite papers use three source classes written for separate purposes. Their definitions of Starlink can differ.
Thus, the sheet has to carry provenance as a column rather than as a footnote.
前三章一次產生一個數字，每個數字都出現在解釋其物理或量測的那個句子裡，而現在想要模擬、仿真或分析 LEO 網路的讀者需要同時拿到全部。參數表就是這樣的集合：一份附日期的表，列出評估所依賴的每個常數，從殼層的高度與軌道面數、到單雷射速率、節點延遲、排程器週期與佇列大小，每個數值都附上它所出自的文件以及該文件的日期。地面網路論文很少需要這種表，因為它們所模擬的硬體有規格書，而衛星網路論文的常數繼承自三類為各異目的而寫、甚至對「Starlink 是什麼」看法各異的來源，這正是出處必須成為表中一欄並隨每項數值清楚列示的原因。
The first source is the regulatory record. A filing snapshot records the constellation geometry in one dated FCC document. Later modifications replace it, while contemporary papers preserve it in their simulators. 第一類來源是監管紀錄。每一份 FCC 文件固定了一個申報快照，也就是那份附日期文件中所陳述的星座幾何，之後的修正會取代它，而當年的論文則把它凍結進模擬器裡。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Constellation configuration | The 2016 application placed 1,584 satellites at 1,150 km in 32 planes of 50. The November 2018 modification moved them to 550 km in 24 planes of 66 (SAT-MOD-20181108-00083). August 2019 redistributed the same fleet across 72 planes of 22 (SAT-MOD-20190830-00087, McDowell’s transcription secondary. Fcc.report record regulatory). Handley’s HotNets 2018 paper, CoNEXT 2019 and Hypatia [6] therefore model separate Starlink snapshots. Intra-plane neighbor distance is 659 km for 24×66 and 1,969 km for 72×22 [33, §I, III] [34, §3]. |
| Evaluation conditions | Measurement literature began when service became publicly reachable. Examples include Michel et al. and Kassem et al. at IMC 2022. Tanveer et al. followed at CoNEXT 2023, Mohan et al. at WWW 2024, and Izhikevich et al. in Proc. ACM MACS 2024 and Garcia et al. at IMC 2026. These studies observe the deployed network. |
| Link or data rate | Company reports form the third source class. They give 100 Gbps sustained, 200 Gbps peak, 9,000 lasers and 42 PB daily. Journalists relayed these figures from a January 2024 SPIE talk (TechSpot, Hackaday company via secondary). Its latency figures come from its March 2024 latency document company. |
| 星座配置 | 2016 年的申請把 1,584 顆衛星放在 1,150 km、32 個軌道面各 50 顆，2018 年 11 月的修正（SAT-MOD-20181108-00083）把它們移到 550 km、24 個軌道面各 66 顆，2019 年 8 月的修正（SAT-MOD-20190830-00087）又把同樣的 1,584 顆重排成 72 個軌道面各 22 顆（McDowell 對申報的轉錄 二手。 |
| 軌道面配置 | fcc.report 紀錄 監管文件），於是 Handley 的 HotNets 2018 論文、Bhattacherjee 與 Singla 的 CoNEXT 2019 論文以及 Hypatia[6]各自描述了一個相異的「Starlink」，同一軌道面相鄰衛星的距離在 24×66 佈局中是 659 km、在 72×22 佈局中則是 1,969 km[33, §I, III][34, §3]。 |
| 評估條件 | 第二類來源是服務對外開放後才開始的量測文獻，包括 IMC 2022 的 Michel 等人與 Kassem 等人、CoNEXT 2023 的 Tanveer 等人、WWW 2024 的 Mohan 等人、Proc. ACM MACS 2024 的 Izhikevich 等人以及 IMC 2026 的 Garcia 等人，它們觀察的是實際飛行的網路，與申報的網路各有適用範圍。 |
| 評估條件 | 第三類來源是公司本身，其單雷射 100 Gbps 持續、200 Gbps 峰值、9,000 具雷射與每日 42 PB 的數字，是透過記者對 2024 年 1 月一場 SPIE 演講的報導進入文獻的（TechSpot。 |
| 延遲比較 | Hackaday 公司（二手）），而其延遲數字則來自公司自己 2024 年 3 月的延遲文件 公司。 |
An evidence tag identifies a value’s source class. It tells readers where to check the claim. Peer-reviewed values point to a refereed paper’s section, table or figure. Regulatory values point to filings or standards, including SDA’s optical specification.
Preprint values point to a specific arXiv version with review status requiring confirmation. Company values point to operator statements. Textbook values point to established references such as Vallado. The tag identifies provenance.
Accuracy requires examination of the underlying claim and evidence. The tags let the two cases be told apart at a glance.
證據標籤是本教材附在數字上的標記，用來說明它屬於上述哪一類，因此讀者能用什麼方式查證。同儕審查的數字可以對照審查過的論文的某一節、表或圖來檢查，監管文件的數字可以對照申報或像 SDA 光學終端規格這類標準，預印本的數字對照的是仍待確認審查的 arXiv 版本，公司的數字只能對照營運商自己的說法，而教科書的數字則對照 Vallado 這類標準參考書。標籤標示出處，數值的準確度則須另行檢查原始主張及證據。SpaceX 的單雷射 100 Gbps 很可能是準確的，且仍待獨立驗證。標籤讓這兩種情況一眼就能分辨。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Path length and hops | A sixth label, computed , marks arithmetic performed in this primer from tagged inputs, and a computed value inherits the weakest tag among its inputs. The 1,969 km hop is computed from a regulatory filing. The 3.9 Tbps mean implied by 42 PB daily is computed from a company claim. |
| Path length and hops | Chaudhry and Yanikomeroglu’s peer-reviewed OJCOMS 2022 paper explicitly assumes 10 ms per-hop node delay in Sec. IV. SpaceX’s 100 Gbps laser rate may be accurate and still requires independent verification. |
| 路徑長度與跳數 | 第六種標記計算值標示本教材由已標籤的輸入算出的結果，而計算值會繼承其輸入中最弱的標籤，所以 1,969 km 的跳段長度是由監管申報算出的，每日 42 PB 所隱含的 3.9 Tbps 平均速率則是由公司宣稱算出的。 |
| 路徑長度與跳數 | Chaudhry 與 Yanikomeroglu 在 OJCOMS 2022 論文中的每跳 10 ms 節點延遲通過了同儕審查，卻仍是作者在 Sec. IV 中聲明的假設。 |
The three sources also define two eras of the research literature, and the parameter sheet exists largely to keep them apart.這三類來源也劃分出研究文獻的兩個時代，而參數表的存在很大程度上就是為了把它們分開。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Model configuration | Modeling-era papers span Handley’s HotNets 2018 work, CoNEXT 2019, SIGCOMM CCR 2020 and Hypatia. They froze filings, assumed four or five laser terminals and used +Grid links with instant reconfiguration. Scheduling remained outside their models because service observation was still unavailable. |
| Queue behavior | Measurements from 2022 onward reveal fifteen-second reconfiguration at :12, :27, :42 and :57 [13, Fig. 2]. They identify a shared 1,400–1,600-packet drop-front queue [45, §7]. PoP location strongly shapes idle latency [49, §6.1]. Current satellites carry three lasers, with active topology proprietary [39, §2.1, §6.1]. |
| Round-trip time | A four-ISL +Grid simulation represents a model-era design. Using measured RTT to validate it requires evidence linking that design to the measured deployment. This occurs because the +Grid rows carry regulatory and modeling tags while the RTT rows carry measurement tags from a different network. |
| 建模假設 | 建模時代從 Handley 的 HotNets 2018 論文經 CoNEXT 2019、SIGCOMM CCR 2020 到 Hypatia，凍結了一份申報，把每顆衛星的四或五具雷射終端接成 +Grid，把雷射鏈路當成可瞬間重組，排程器則仍待辨識，相關部署行為須待服務開放後才能觀察。 |
| 佇列行為 | 量測時代從 2022 年起發現終端是透過每分鐘第 :12、:27、:42、:57 秒的 15 秒全球重組來連線的[13, Fig. 2]，下行瓶頸是單一個約 1,400–1,600 封包的前端丟棄佇列[45, §7]，閒置延遲主要由 PoP 所在位置主導，傳播則提供延遲下限[49, §6.1]，而衛星如今載有三具雷射終端，其拓樸 SpaceX 仍以專有資訊管理[39, §2.1, §6.1]。 |
| 往返時間 | 一篇 2026 年的論文如果混用兩個時代，例如跑一個 4-ISL 的 +Grid 模擬器再引用量測到的 RTT 作為驗證，就犯了一個參數表會讓它顯形的範疇錯誤，因為 +Grid 那幾列帶的是監管與建模標籤，而 RTT 那幾列帶的是來自另一個網路的量測標籤。 |
Original figure note原始圖說出處
Drawn for this primer from fact base Part A Chapters 1–3: McDowell's filing transcription, fcc.report SAT-MOD-20190830-00087, FCC DA 26-36, Michel et al. IMC 2022, Tanveer et al. CoNEXT 2023, Mohan et al. WWW 2024, Izhikevich et al. 2024, Garcia et al. IMC 2026, TechSpot/Hackaday Feb 2024, Starlink latency PDF Mar 2024, ISPreview Jan 2025, Converge Digest Jul 2025.4.2The sheet參數表
The parameter sheet gives 2026 study inputs with their values, evidence tags and exact source locations. Rows follow evaluation structure: constellation geometry, ISLs, ground segment, scheduler and satellite bus. They end with Part III’s terrestrial references and Chapter 11’s Suncatcher cluster. Values from separate sources remain separate entries. A contested or unknown value needs an explicit assumption, such as "we assume 1 ms per hop, pending measurement." Readers can then assess that choice.
下表彙整 2026 年研究可用的參數，每列均附上數值、證據標籤與可查證的來源位置。各列按照評估章節通常的組織方式分組，從星座幾何經星間鏈路、地面段與排程器到衛星平台，最後是第三部將要對照的地面參考數字以及第 11 章詳細處理的 Suncatcher 叢集。同一格中出現兩個數值時，它們來自相異來源，因此各自列示並保留依據。數值仍有爭議或待確認時，研究須明列假設，例如「量測資料仍待補充，因此假設每跳 1 ms」，讓讀者能直接檢查這項選擇。
| Parameter參數 | Recommended 2026 assumption建議的 2026 年假設 | Evidence證據 | Source and location來源與位置 |
|---|---|---|---|
| Constellation星座 | |||
| Canonical research shell標準研究殼層 | 550 km, 53°, 72 planes × 22 (1,584). State the filing date550 km、53°、72 個軌道面 × 22（1,584 顆）。註明申報日期 | regulatory via secondary監管（二手） peer-reviewed同儕審查 | SAT-MOD-20190830-00087. Hypatia [6] Table 1SAT-MOD-20190830-00087。Hypatia[6]Table 1 |
| Real Starlink altitude rangeStarlink 實際高度範圍 | ~340–570 km. 525–535 km shells may be lowered to 475–485 km. Only ~575 satellites remain at 552 km約 340–570 km。525–535 km 殼層可能降至 475–485 km。552 km 僅剩約 575 顆 | regulatory監管文件 preprint預印本 | FCC DA 26-36 ¶6, ¶13(c). [16, Tab. 1]FCC DA 26-36 ¶6、¶13(c)。[16, Tab. 1] |
| Period / speed at 550 km550 km 的週期／速度 | 95.65 min. 7.585 km/s95.65 分。7.585 km/s | computed計算值 | Kepler relations, Vallado App. D constantsKepler 關係式，Vallado 附錄 D 常數 |
| Minimum elevation最低仰角 | Starlink 25° (user and gateway). Kuiper 35° user / 20° gateway. Telesat 10°. OneWeb ≥55° guaranteed visibilityStarlink 25°（使用者與閘道）。Kuiper 使用者 35°／閘道 20°。Telesat 10°。OneWeb 保證可見 ≥55° | regulatory監管文件 peer-reviewed同儕審查 | ISPreview Apr 2021. Kuiper Tech. App. Table 7. Telesat filing p. 15. OneWeb Att. B p. 12ISPreview 2021 年 4 月。Kuiper 技術附錄 Table 7。Telesat 申報 p. 15。OneWeb 附件 B p. 12 |
| Maximum pass at 25° / 40°25°／40° 的最長過頂時間 | 4.49 / 2.74 min (550 km). 9.29 / 5.99 min (1,200 km)4.49／2.74 分（550 km）。9.29／5.99 分（1,200 km） | computed計算值 | λ(ε)/n, Chapter 1 Panel 1.2λ(ε)/n，第 1 章面板 1.2 |
| Slant range and one-way delay at 25°25° 的斜距與單向延遲 | 1,123 km, 3.75 ms (550 km). SpaceX says 1.8–3.6 ms per leg1,123 km、3.75 ms（550 km）。SpaceX 稱每段 1.8–3.6 ms | computed計算值 company公司 | d(ε) formula. Starlink latency PDF (Mar 2024)d(ε) 公式。Starlink 延遲 PDF（2024 年 3 月） |
| Inter-satellite links星間鏈路 | |||
| ISL propagationISL 傳播 | 3.336 µs/km vacuum. Fiber 4.895 µs/km (n = 1.4675). 47% advantage真空 3.336 µs/km。光纖 4.895 µs/km（n = 1.4675）。優勢 47% | computed計算值 peer-reviewed同儕審查 | [3, §4][3, §4] |
| Intra-plane / inter-plane ISL length (72×22)同面／鄰面 ISL 長度（72×22） | 1,969 km. 604 km (equator) → 363 km (53° latitude)1,969 km。604 km（赤道）→ 363 km（緯度 53°） | computed計算值 preprint預印本 | chord 2R sin(π/N). Ali & Zhang, arXiv 2602.07756, Sec. 3弦長 2R sin(π/N)。Ali 與 Zhang，arXiv 2602.07756，Sec. 3 |
| ISL line-of-sight limitISL 視線極限 | 5,016 km (80 km atmosphere). SpaceX claims links beyond 5,400 km5,016 km（80 km 大氣層）。SpaceX 宣稱鏈路超過 5,400 km | peer-reviewed同儕審查 company via secondary公司（二手） | [17, Eq. 1] . TechSpot Feb 2024[17, Eq. 1] 。 TechSpot 2024 年 2 月 |
| Per-laser rate單雷射速率 | 100 Gbps sustained, 200 Gbps peak (SpaceX claim). Model at 100 Gbps per direction (Ali & Zhang. StarCDN) or 200 Gbps (TinyLEO)持續 100 Gbps、峰值 200 Gbps（SpaceX 宣稱）。建模時每方向 100 Gbps[34][62]或 200 Gbps（TinyLEO） | company via secondary公司（二手） preprint預印本 peer-reviewed同儕審查 | TechSpot. TinyLEO [39] §6.1. StarCDN [62] Table 1TechSpot。TinyLEO[39]§6.1。StarCDN[62]Table 1 |
| Laser terminals per satellite每顆衛星的雷射終端數 | 3 (current Starlink page). 4 for legacy +Grid modeling. 5 reserved for historical designs3（目前 Starlink 網頁）。傳統 +Grid 建模用 4。5 限於歷史設計 | company via secondary公司（二手） | NewSpace Economy May 2026. Ali & Zhang Sec. 2NewSpace Economy 2026 年 5 月。Ali 與 Zhang Sec. 2 |
| Acquisition / setup delay擷取／建立延遲 | 2–30 s today (CONDOR 30 s first, 2 s re-acquire). SDA ≤100 s cold. "milliseconds" is hypothetical今日 2–30 s（CONDOR 首次 30 s、重擷取 2 s）。SDA 冷啟動 ≤100 s。「毫秒級」仍屬假想 | preprint預印本 standard標準 | [32, §II] . [41][32, §II] 。 [41] |
| Per-hop node delay每跳節點延遲 | unknown. Papers assume 1 ms or 10 ms待確認。論文假設 1 ms 或 10 ms | preprint預印本 peer-reviewed同儕審查 | [32] (1 ms) . [33, §IV] (10 ms)[32] （1 ms） 。 [33, §IV] （10 ms） |
| Ground segment and measured performance地面段與量測效能 | |||
| Idle RTT閒置 RTT | 40–50 ms (academic 2022–23). 25–35 ms [63]. ~100 ms at ISL islands40–50 ms（學術界 2022–23）。25–35 ms[63]。ISL 孤島約 100 ms | peer-reviewed同儕審查 company公司 | [49, Figs. 5–7, 12] . Starlink latency PDF[49, Figs. 5–7, 12] 。 Starlink 延遲 PDF |
| Bufferbloat / queue緩衝膨脹／佇列 | 2–4× to 400–500 ms under load. Single ~1,400–1,600-packet drop-front queue負載下 2–4× 至 400–500 ms。單一約 1,400–1,600 封包的前端丟棄佇列 | peer-reviewed同儕審查 preprint預印本 | [49, Fig. 8] . [45, §7][49, Fig. 8] 。 [45, §7] |
| Throughput and asymmetry吞吐量與上下行速率差距 | 100–250 Mbps down, 10–30 Mbps up, ~10:1下行 100–250 Mbps，上行 10–30 Mbps，約 10:1 | peer-reviewed同儕審查 | Michel et al. §3.3. Kassem et al. Table 3. Ma et al. Table IIMichel 等人 §3.3。Kassem 等人 Table 3。Ma 等人 Table II |
| Loss封包遺失 | 0.4–2% typical, bursts to 50%典型 0.4–2%，叢發可達 50% | peer-reviewed同儕審查 | Michel et al. Table 2. Kassem et al. Fig. 6cMichel 等人 Table 2。Kassem 等人 Fig. 6c |
| User downlink waveform使用者下行波形 | 8 × 240 MHz OFDM, 1,024 subcarriers, 1.333 ms frame, 4QAM/16QAM8 × 240 MHz OFDM，1,024 個子載波，1.333 ms 幀，4QAM/16QAM | peer-reviewed同儕審查 | [42, Tab. II][42, Tab. II] |
| Per-satellite capacity單顆衛星容量 | contested: 6.5 Gbps (model), 24 / 96 Gbps (V1.5 / V2 Mini, preprint), >1 Tbps V3 (claim)有爭議：6.5 Gbps（模型）、24／96 Gbps（V1.5／V2 Mini，預印本）、V3 >1 Tbps（宣稱） | peer-reviewed同儕審查 preprint預印本 company公司 | [31] . [26, §IV-D] . Converge Digest Jul 2025[31] 。 [26, §IV-D] 。 Converge Digest 2025 年 7 月 |
| PoPs / gatewaysPoP／閘道 | 49 PoPs, 98 backbone links (Dec 2025). 9 dishes per gateway. >1,500 US antennas49 個 PoP、98 條骨幹鏈路（2025 年 12 月）。每閘道 9 面天線。美國 >1,500 面天線 | preprint預印本 peer-reviewed同儕審查 company公司 | Wang et al. Sec. III . [44, §III-B] . Converge DigestWang 等人 Sec. III 。 [44, §III-B] 。 Converge Digest |
| Scheduler排程器 | |||
| Scheduler period排程器週期 | 15 s at :12/:27/:42/:57. ~74 ms spike in the first 140 ms of each period15 秒，於 :12/:27/:42/:57。每週期前 140 ms 內約 74 ms 的尖峰 | peer-reviewed同儕審查 preprint預印本 | [13, Fig. 2] . [57, Fig. 3][13, Fig. 2] 。 [57, Fig. 3] |
| Satellite bus衛星平台 | |||
| Satellite bus power衛星平台功率 | ~28 kW from ~105 m² at 22% for a V2 Mini analogue (Google estimate). 1.5 kW (2020 estimate)V2 Mini 類比體約 105 m²、效率 22% 得約 28 kW（Google 估計）。1.5 kW（2020 年估計） | peer-reviewed同儕審查 | [36, p. 10] . [2, §4][36, p. 10] 。 [2, §4] |
| Space solar / battery太空太陽能／電池 | 380–430 W/m² BOL, ~30 W/kg typical (200 max). Li-ion 150–270 Wh/kg壽命初期 380–430 W/m²，典型約 30 W/kg（最高 200）。鋰離子 150–270 Wh/kg | agency report機構報告 | [66, Tab. 3-1][66, Tab. 3-1] |
| Radiator散熱板 | ~413 W/m² per side at 300 K, ε = 0.9, before backloads300 K、ε = 0.9 時每面約 413 W/m²，計算範圍排除回載 | computed計算值 | Stefan–Boltzmann, Chapter 8 Panel 8.2Stefan–Boltzmann，第 8 章面板 8.2 |
| Eclipse at 550 km550 km 的地影 | up to 35.6 min of 95.65 min (37%). Dawn–dusk SSO is eclipse-free95.65 分中最多 35.6 分（37%）。晨昏太陽同步軌道全程受日照 | computed計算值 | cylindrical shadow, Chapter 1 Section 1.3圓柱陰影，第 1 章第 1.3 節 |
| Radiation at 650 km650 km 的輻射 | ~150 rad(Si)/yr behind 10 mm Al. 750 rad over 5 yr10 mm 鋁屏蔽後約 150 rad(Si)/年。5 年 750 rad | peer-reviewed同儕審查 | [36, p. 5][36, p. 5] |
| Design life設計壽命 | 5 years5 年 | preprint預印本 company公司 | [16, §3.3] . SpaceX filing via DCD[16, §3.3] 。 SpaceX 申報（經 DCD） |
| Launch cost發射成本 | $3,600/kg (Falcon 9 reusable, 2025). $200/kg by the mid-2030s is a projection$3,600/kg（可重複使用獵鷹 9，2025）。2030 年代中期的 $200/kg 是預測 | peer-reviewed同儕審查 | [36, pp. 6, 9–10][36, pp. 6, 9–10] |
| Terrestrial reference地面參考 | |||
| Terrestrial AI reference地面 AI 參考 | 400 Gbps NIC per GPU. 3.2 Tbps per 8-GPU host. NVLink 400–900 GB/s. 25.6–51.2 Tbps switches. Tens of µs multi-hop latency. 1–2 job crashes/month at 15K GPUs每 GPU 400 Gbps NIC。每 8-GPU 主機 3.2 Tbps。NVLink 400–900 GB/s。交換器 25.6–51.2 Tbps。多跳延遲數十 µs。15K GPU 下每月 1–2 次工作崩潰 | preprint預印本 peer-reviewed同儕審查 company公司 | Llama 3 §3.3. HPN [65] §2–3. MegaScale [59] §3.6. TPU v6e docsLlama 3 §3.3。HPN[65]§2–3。MegaScale[59]§3.6。TPU v6e 文件 |
| Suncatcher clusterSuncatcher 叢集 | |||
| Suncatcher clusterSuncatcher 叢集 | 81 satellites, R = 1 km, 100–200 m spacing, 650 km dawn–dusk SSO, 9.6–12.8 Tbps per aperture, 800 Gbps bench81 顆衛星，R = 1 km，間距 100–200 m，650 km 晨昏太陽同步軌道，每孔徑 9.6–12.8 Tbps，桌面測試 800 Gbps | peer-reviewed同儕審查 | [36, pp. 3–4, 8][36, pp. 3–4, 8] |
Three habits make the sheet usable rather than decorative. The first is to read the tag column before the value column. This occurs because a row whose tags are all peer-reviewed, such as the downlink waveform or the idle RTT, can be quoted as a fact. A company via secondary row requires the claim and its date. Laser terminal count is one example.使用參數表時，可依以下三項原則核對證據。第一是先讀標籤欄再讀數值欄，因為標籤全為同儕審查的列（例如下行波形或閒置 RTT）可以當作事實引用，而唯一標籤是「公司（二手）」的列（例如雷射終端數）則必須連同日期一起當作宣稱引用。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Path length and hops | Computed rows are reproducible from their formulas and tagged inputs. Disputing the 1,969 km hop therefore requires examining the underlying 72×22 filing. |
| PoP coverage | Treat the source date as part of each value. Write "49 PoPs in December 2025" and "25.7 ms in SpaceX’s July 2025 report". Both quantities change over time. |
| 路徑長度與跳數 | 第二是留意哪些列是計算值，因為這些列任何人都能從公式與已標籤的輸入重現，一位質疑 1,969 km 跳段長度的審稿人，實際上質疑的是背後的 72×22 申報。 |
| PoP 覆蓋 | 第三是把來源欄裡的日期視為數值的一部分，於是「49 個 PoP」永遠是「2025 年 12 月的 49 個 PoP」，「25.7 ms」永遠是「SpaceX 2025 年 7 月報告中的 25.7 ms」，因為這兩個數字隨年份持續變化。 |
4.3Parameters awaiting public evidence仍待公開證據的參數
Seven simulator inputs still require public evidence from SpaceX. A 2026 paper must state an assumption for each. The first is the topology itself. Since Hypatia, +Grid tools connect two intra-plane neighbors and one same-index satellite in each adjacent plane. Pan et al. and TinyLEO describe SpaceX’s actual topology as undisclosed [39, §6.1].
The current satellites carry three terminals rather than the four the +Grid requires. Thus, the honest statement is "three terminals per satellite, topology undisclosed, +Grid or 3-ISL grid used as a baseline". The second is the primary per-laser rate, whose SpaceX primary document still requires retrieval.
七項模擬器輸入仍待 SpaceX 的公開證據，因此 2026 年的論文須逐項明列假設。第一個是拓樸本身。自 Hypatia 以來的每一個 +Grid 工具都把每顆衛星接到同面的兩個鄰居與相鄰軌道面各一顆同索引衛星，然而 SpaceX 的啟用鏈路配置仍屬專有資訊，Pan 等人與 TinyLEO 都指出真實拓樸仍待公開[39, §6.1]，而目前的衛星載有三具終端，+Grid 則需要四具，因此誠實的寫法是「每顆衛星三具終端，拓樸仍待公開，以 +Grid 或 3-ISL 網格作為基準」。第二個是原始的單雷射速率，SpaceX 提供數值的原始文件仍待取得。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Capacity comparison | Recent 100 Gbps sustained and 200 Gbps peak assumptions trace to journalists covering one conference talk. Chaudhry and Yanikomeroglu expected "up to 10 Gbps" in 2021 [17]. Choosing 100 or 200 Gbps doubles Part II capacity results. |
| Path length and hops | Per-hop node delay needs public internal measurements. Literature assumptions of 1 ms [32] and 10 ms [33, §IV] differ tenfold. Across ten hops, this difference can determine competitiveness. |
| 容量比較 | 近期每篇論文使用的持續 100 Gbps 與峰值 200 Gbps 都源自記者對同一場會議演講的轉述，Chaudhry 與 Yanikomeroglu 在 2021 年的預期是「最高 10 Gbps」[17]，而在 100 與 200 Gbps 之間的選擇會讓第二部的每一個容量結果差上兩倍。 |
| 路徑長度與跳數 | 第三個是每跳節點延遲，Starlink 衛星內部的公開量測資料仍待補充，文獻中的 1 ms[32]與 10 ms[33, §IV]相差一個數量級，而在十跳的路徑上，這就是延遲競爭力的顯著差異。 |
The fourth missing quantity is the scheduler's interior. Global ISL-traffic share requires a published estimate. Belgian traceroutes showed groundward routing in 2022 [48, §3.1]. Nigerian traces show ISL-sized spikes during roughly one third of sampled time [35, §7.2]. Thus, a paper evaluating an ISL routing scheme must leave its share of current traffic open. Every orbital-computing paper that scales a payload from "a Starlink bus" is building on this estimate.第四個缺失的量是排程器的內部。第五個是流量究竟有多少比例走 ISL，全球數字仍待論文量化，只有 2022 年比利時的 traceroute 顯示採用對地路徑[48, §3.1]，以及奈及利亞的追蹤中 ISL 尺度的尖峰佔了約三分之一的時間[35, §7.2]，因此一篇評估 ISL 路由方案的論文須將方案涵蓋的現有流量比例列為待確認項目。每一篇從「一個 Starlink 平台」放大酬載的在軌運算論文，都建立在這個估計之上。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Scheduling period and phase | Scheduler period and phase are measured. Tanveer et al. observed :12, :27, :42 and :57 shifts at every site. Casparsen et al. resolved initial spikes with 500 Hz probes. Allocation rules remain inferred from preference for 45–90° satellites [13, §5.1, Fig. 4]. Parallel latency bands, spaced a few milliseconds apart, suggest round-robin frame allocation (Secs. 3–4). |
| Capacity comparison | Gen1 capacity remains contested. A peer-reviewed model gives 6.5 Gbps [31], while a preprint gives 24 Gbps for V1.5 [26, §IV-D]. Talk figures in the high teens require source evidence. |
| Power budget | Secondary sources place V2 Mini mass from under 600 kg to about 800 kg. Google’s 28 kW array figure estimates an analogue [36, p. 10]. Its scope differs from an operator specification. |
| 排程週期與相位 | 它的週期與相位是量測到的事實，因為 Tanveer 等人在每個觀測點都發現了 :12、:27、:42、:57 的狀態切換，Casparsen 等人也用 500 Hz 探測量到了週期起點的尖峰，但排程器分配什麼、依據什麼規則，則屬於推論，須與直接觀測區分，推論來自對 45–90° 仰角衛星的偏好[13, §5.1, Fig. 4]以及作者歸因於輪詢式幀分配的、相隔數毫秒的延遲帶（Secs. 3–4）。 |
| 單星容量來源 | 第六個是 Gen1 的單顆衛星容量，同儕審查的物理模型給出 6.5 Gbps[31]，一份預印本給出 V1.5 的 24 Gbps[26, §IV-D]，而演講中流傳的十幾 Gbps 高段數字則仍待來源證據。 |
| 功率預算 | 第七個是 V2 Mini 的質量與功率，二手來源給出由低於 600 kg 至約 800 kg 的各種數字，而 28 kW 的太陽能板數字是 Google 對一個類比體的估計[36, p. 10]，其依據為類比估計，SpaceX 規格則仍待取得。 |
These gaps are the residue of the two eras introduced in Section 4.1. The fact base's register of disagreements lists six assumptions that the modeling era made and the measurement era has since overtaken. The table lists six assumptions that recur as Limits items in Part II. Recognizing them helps readers date a paper’s model.
這些缺口是第 4.1 節所介紹的兩個時代留下的殘餘，而事實庫的分歧登錄表列出了六項建模時代曾採用、如今已被量測時代超越的假設。下表重製它們，因為每一項都會在第二部的某篇摘要中以「限制」條目重新出現，也因為知道這六項的讀者光從假設就能判定任何一篇 LEO 論文的年代。
| Modelling-era assumption建模時代的假設 | Who assumed it誰採用了它 | What the measurement era found量測時代的發現 | What a 2026 paper should do2026 年論文該怎麼做 |
|---|---|---|---|
| Four-ISL +Grid is "Starlink's topology"四 ISL 的 +Grid 就是「Starlink 的拓樸」 | HotNets 2018, CoNEXT 2019, CCR 2020, Hypatia, StarryNet, xeoverse, LeoEM, SkyLink, SaTE, StarCDN, SpaceCDNHotNets 2018、CoNEXT 2019、CCR 2020、Hypatia、StarryNet、xeoverse、LeoEM、SkyLink、SaTE、StarCDN、SpaceCDN | Three terminals per satellite. Topology undisclosed (TinyLEO §6.1). 3-ISL grid disconnected on real TLEs (Ali & Zhang Sec. 5.6)每顆衛星三具終端。拓樸仍待公開（TinyLEO §6.1）。3-ISL 網格在真實 TLE 上分成多個連通分量[34, §5.6] | Present +Grid as a design-space baseline把 +Grid 當作設計空間的基準 |
| Laser links reconfigure instantly雷射鏈路可瞬間重組 | Every topology and routing paper except Chaudhry & Yanikomeroglu and Bhattacharjee et al.除 Chaudhry 與 Yanikomeroglu 及 Bhattacharjee 等人以外的每篇拓樸與路由論文 | Setup takes 2–30 s on current terminals. SDA allows ≤100 s cold start現有終端建立需 2–30 s。SDA 允許冷啟動 ≤100 s | State the setup delay and whether links are static over the evaluation列出建立延遲及評估期間鏈路的靜態或動態設定 |
| Scheduler excluded排程器排除 | Hypatia, StarryNet, Celestial, xeoverse, SaTE, StarCDN (15 s only as a time step)Hypatia、StarryNet、Celestial、xeoverse、SaTE、StarCDN（15 秒只作為時間步長） | 15-s global reconfiguration with a ~74 ms spike. Modeled only by LeoEM/[67], LeoCC and OrbCC15 秒全球重組並伴隨約 74 ms 尖峰。只有 LeoEM/SaTCP、HotNets 2024、LeoCC 與 OrbCC 建模 | Say whether the 15-s schedule is modeled明列 15 秒排程的建模設定 |
| Propagation-only "half the RTT of fiber"只算傳播的「光纖 RTT 的一半」 | [3], Gearing Up 2018, CCR 2020 at 1,150 km with 4–5 ISLs[3]、Gearing Up 2018、CCR 2020，1,150 km、4–5 ISL | Measured bent-pipe floor ~40 ms. Idle medians 25–50 ms量測到的彎管底線約 40 ms。閒置中位數 25–50 ms | Separate theoretical path latency from end-to-end latency把理論路徑延遲與端到端延遲分開 |
| claims federated learning requires ground-only connectivity宣稱聯邦學習僅能使用對地鏈路 | FedSpace (2022), Razmi et al. (2022). OrbitalBrain (2026) assumes ~100 MbpsFedSpace（2022）、Razmi 等人（2022）。OrbitalBrain（2026）假設約 100 Mbps | Starlink has flown lasers since v1.5. Claimed 100–200 GbpsStarlink 自 v1.5 起即配備雷射。宣稱 100–200 Gbps | Date the assumption為假設標註日期 |
| Simulator ISL rates of 10 Mbit/s–10 Gbit/s模擬器 ISL 速率 10 Mbit/s–10 Gbit/s | Hypatia (10 Mbit/s experiments), Celestial (10 Gb/s or 100 Mb/s), SaTE (200 Mbps scaled)Hypatia（10 Mbit/s 實驗）、Celestial（10 Gb/s 或 100 Mb/s）、SaTE（200 Mbps 縮放） | 100–200 Gbps claimed per laser宣稱每雷射 100–200 Gbps | Treat simulator throughput results as illustrations of dynamics把模擬器的吞吐量結果當作動態的示意 |
Reproduced from fact base Part C, items 55–60.
重製自事實庫第 C 部第 55–60 項。
4.4Worked example: an end-to-end latency and capacity budget for one user完整範例：單一使用者的端到端延遲與容量預算
The sheet earns its keep when it is used. Thus, This section builds the latency and capacity budget of a single user from its rows and checks the result against measurement. The first path is the bent pipe of Chapter 3: up to a satellite, down to the gateway, then over ground fiber to the PoP. The paper leaves its one-way or round-trip definition open.參數表要用了才有價值，所以本節從它的各列建立單一使用者的延遲與容量預算，並拿結果對照量測。第一條路徑是第 3 章的彎管：上行到衛星、下行到閘道，再經地面光纖到 PoP。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| PoP coverage | The user is a terminal in the Seattle area, a region that appears in every measurement campaign [46, Tab. 3] [49, §4] . This primer places gateway and PoP 300 km away so terminal and gateway share a satellite footprint. Mohan et al.’s typical terminal–gateway distances below 1,200 km permit this example (Fig. 13). |
| Node delay | Both radio legs use the longest permitted geometry, the 25° mask [23]. Satellite forwarding explicitly assumes 1 ms node delay following Bhattacharjee et al. [32]. Public measurements are needed to establish this delay. The ground segment uses Mohan et al.’s roughly 5 ms gateway–PoP value [49, App. D], treated here as one-way. |
| PoP 覆蓋 | 這位使用者是西雅圖地區的一台終端，這個地區出現在每一次量測活動裡[46, Tab. 3][49, §4]，而服務它的閘道與 PoP 設定在 300 km 之外，這個距離由本教材選定，使終端與閘道共享同一個衛星覆蓋區，正如 Mohan 等人 Fig. 13 中典型小於 1,200 km 的終端到閘道距離所允許的。 |
| 星座配置 | 為了讓每個輸入都可追溯，兩段無線電鏈路都取 25° 仰角遮罩，這是申報允許的最長幾何（ISPreview，2021 年 4 月 監管（二手）），衛星的轉送計入一個仍待公開量測的節點延遲，因此依 Bhattacharjee 等人[32]明確假設為 1 ms，地面段則計入 Mohan 等人約 5 ms 的閘道到 PoP 數字[49]，該論文報告這個數字時將其單向或往返定義保留為待確認，這裡取為單向。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Slant range at 25°, 550 km550 km、25° 的斜距 | \(R_E\big[\sqrt{(a/R_E)^2-\cos^2\varepsilon}-\sin\varepsilon\big]\) | 1,123 kmcomputed. R_E = 6,378.137 km, a = 6,928.1 km [5, App. D]計算值。R_E = 6,378.137 km，a = 6,928.1 km[5] |
| One radio leg at 25°25° 的單段無線電鏈路 | \(1{,}123\ \text{km}\times 3.336\ \mu\text{s/km}\) | 3.75 mscomputed. The zenith leg of 550 km would be 1.83 ms計算值。天頂方向 550 km 的一段為 1.83 ms |
| Bent pipe, both radio legs彎管的兩段無線電鏈路 | \(2\times 3.75\ \text{ms}\) | 7.49 mscomputed. Worst case with both ends at 25°計算值。兩端都在 25° 的最壞情況 |
| Bent pipe, one-way to the PoP彎管到 PoP 的單向延遲 | \(7.49 + 1 + 5\ \text{ms}\) | 13.5 mscomputed. Node delay 1 ms assumed (Bhattacharjee et al.). GS→PoP ~5 ms (Mohan et al. App. D)計算值。節點延遲假設 1 ms（Bhattacharjee 等人）。GS→PoP 約 5 ms（Mohan 等人附錄 D） |
| Bent pipe, round trip彎管的往返延遲 | \(2\times 13.5\ \text{ms}\) | 27.0 mscomputed. 45.0 ms if the node delay is 10 ms [33, §IV]計算值。若節點延遲為 10 ms 則為 45.0 ms[33, §IV] |
| One intra-plane ISL hop (72×22)一個同面 ISL 跳段（72×22） | \(1{,}969\ \text{km}\times 3.336\ \mu\text{s/km}\) | 6.57 mscomputed. Chord 2R sin(π/22), R = 6,921 km計算值。弦長 2R sin(π/22)，R = 6,921 km |
| Five ISL hops, propagation only五個 ISL 跳段，僅傳播 | \(5\times 6.57\ \text{ms}\) | 32.8 mscomputed. 9,845 km of path計算值。路徑長 9,845 km |
| ISL chain, one-way to the PoPISL 鏈到 PoP 的單向延遲 | \(3.75 + 32.8 + 6\times 1 + 3.75 + 5\ \text{ms}\) | 51.3 mscomputed. Six satellites each charged 1 ms. 105 ms at 10 ms per satellite計算值。六顆衛星各計 1 ms。每顆 10 ms 時為 105 ms |
| ISL chain, round tripISL 鏈的往返延遲 | \(2\times 51.3\ \text{ms}\) | 103 mscomputed. Compare ~100 ms at Mohan et al.'s ISL islands and 181 ms minimum for Izhikevich et al.'s Seychelles yacht計算值。對照 Mohan 等人 ISL 孤島的約 100 ms 與 Izhikevich 等人塞席爾遊艇的最小 181 ms |
| Fibre, 300 km one-way光纖 300 km 單向 | \(300\ \text{km}\times 4.895\ \mu\text{s/km}\) | 1.47 mscomputed. N = 1.4675 (Bhattacharjee et al.). 2.94 ms round trip計算值。n = 1.4675（Bhattacharjee 等人）。往返 2.94 ms |
| Fibre over the ISL chain's 9,845 km光纖走 ISL 鏈的 9,845 km | \(9{,}845\ \text{km}\times 4.895\ \mu\text{s/km}\) | 48.2 mscomputed. Against 32.8 + 6 = 38.8 ms by ISL with 1 ms nodes, Handley's "beyond about 3,000 km" [3]計算值。對照 ISL 在 1 ms 節點下的 32.8 + 6 = 38.8 ms，即 Handley 的「約 3,000 km 以上」[3] |
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | Geometry, one assumed millisecond per satellite and a measured ground segment give 27 ms bent-pipe RTT. This is 9–21 ms below Mohan et al.’s 36–48 ms medians [49, Figs. 12–15]. It also falls below Michel et al.’s Belgian 46–52 ms medians [48, §3.1]. The gap identifies latency requiring further explanation. |
| Latency comparison | Chapter 3 identifies scheduler polling, frame allocation, downlink queues and satellite forwarding as contributors beyond geometry. Pan et al. give roughly 20 ms minimum access RTT from propagation plus polling [44, §III-A]. |
| Node delay | The calculation shows why end-to-end data alone leave attribution open. Raising node delay from 1 to 10 ms raises modeled RTT from 27 to 45 ms, within the measured band. Thus, the data are consistent with a slow satellite and a fast scheduler or with a fast satellite and a slow scheduler. This is precisely why the sheet lists node delay as unknown. |
| Path length and hops | A five-hop ISL chain costs 103 ms RTT with 1 ms nodes and roughly twice that with 10 ms nodes. These values bracket Mohan et al.’s roughly 100 ms island latency and Izhikevich et al.’s Seychelles minimum of 181 ms. The 300 km fiber path to the same PoP needs 2.9 ms, a factor of nine below the bent pipe. |
| Propagation delay | Over 9,845 km, fiber needs 48 ms and the ISL chain 39 ms. This restates Handley’s argument: long distances let propagation savings dominate, provided node delay stays small. |
| 星座配置 | 由純幾何、每顆衛星假設 1 ms 與一段量測到的地面區段得出的 27 ms 彎管往返延遲，比 Mohan 等人量到的 36–48 ms 彎管中位數[49, Figs. 12–15]低 9–21 ms，也低於 Michel 等人比利時的 46–52 ms 中位數[48, §3.1]，而這個差距是本章最有教育意義的數字。 |
| 延遲比較 | 它是幾何估計範圍之外的延遲，第 3 章已點出它的來源：排程器的輪詢與幀分配，Pan 等人把它描述為由傳播加輪詢延遲構成的約 20 ms 最小接取 RTT[44, §III-A]，下行前方的佇列，以及衛星本身加上的任何轉送延遲。 |
| 節點延遲 | 計算表顯示了最後這一項為何需要端到端量測以外的證據：把節點延遲從 1 ms 提高到 10 ms，會讓幾何往返延遲從 27 ms 變成 45 ms，正好落在量測區間內，所以資料既與「慢衛星加快排程器」相容，也與「快衛星加慢排程器」相容，這正是參數表把節點延遲列為待確認的原因。 |
| 路徑長度與跳數 | 五跳的 ISL 鏈在 1 ms 節點下往返需 103 ms，在 10 ms 節點下約為兩倍，恰好夾住 Mohan 等人在 ISL 孤島看到的約 100 ms 與 Izhikevich 等人從塞席爾看到的最小 181 ms，而通往同一個 PoP 的 300 km 光纖只需 2.9 ms，比彎管低九倍。 |
| 傳播延遲 | 最後一列用參數表自己的數字重述了 Handley 的開創性論證：走完整 9,845 km，光纖需要 48 ms 而 ISL 鏈需要 39 ms，所以只有當目的地遠到傳播佔主導時軌道才會贏，而且只有在節點延遲很小時才成立。 |
Original figure note原始圖說出處
Drawn for this primer from the calc table above. Measured band from Mohan et al., WWW 2024, Figs. 12–15.| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Capacity comparison | The capacity budget combines satellite capacity and service-cell assumptions. Capacity ranges from a 6.5 Gbps physical model [31] to 24 Gbps V1.5 and 96 Gbps V2 Mini estimates [26, §IV-D]. Pan et al. relay roughly 24 km cells and "100 users in every 300 km²" [44, §III-A]. |
| Capacity comparison | The sanity check gives each user TinyLEO’s 100 Mbps demand. It asks how many cells one satellite can serve [39, §6.1] . Depending on assumed capacity, one satellite supports demand from roughly one to six fully active cells. The density rule therefore describes subscribers, with statistical multiplexing across roughly 2.8 million km² above 25°. |
| 容量比較 | 預算的容量半邊用到參數表中兩列很少被並排的數字：單顆衛星容量，在 6.5 Gbps 的物理模型[31]與 V1.5 衛星 24 Gbps 或 V2 Mini 96 Gbps[26, §IV-D]之間有爭議，以及營運商自己的經驗法則，也就是 Pan 等人轉述的直徑約 24 km 的服務格與「每 300 km² 100 位使用者」[44, §III-A]。 |
| 服務格容量 | 這個合理性檢查問的是：如果每位使用者都用掉 TinyLEO 分配給每人的 100 Mbps[39, §6.1]，一顆衛星能服務多少個這樣的服務格，而答案依所相信的容量數字介於一到六格之間，顯示這條經驗法則是訂戶密度，同時需求則須另行定義，而且服務依賴的是對一顆衛星在 25° 以上所見約 280 萬 km² 範圍的統計多工。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Area of a 24 km cell直徑 24 km 服務格的面積 | \(\pi\,(12\ \text{km})^2\) | 452 km²computed. Diameter from [44, §III-A]計算值。直徑來自 [44, §III-A] |
| Users per cell at the rule of thumb依經驗法則每格的使用者數 | \(452 \times 100/300\) | 151computed. "100 users in every 300 km²" (Pan et al.)計算值。「每 300 km² 100 位使用者」（Pan 等人） |
| Demand per cell if all users draw 100 Mbps若所有使用者都用 100 Mbps 的每格需求 | \(151\times 100\ \text{Mbps}\) | 15.1 Gbpscomputed. Per-user rate from TinyLEO §6.1計算值。每人速率來自 TinyLEO §6.1 |
| Fully active cells one satellite can carry一顆衛星能承載的全活躍服務格數 | \(C_{\text{sat}} / 15.1\ \text{Gbps}\) | 0.4 / 1.6 / 6.4computed. At 6.5 (Pachler), 24 (Wang, V1.5), 96 Gbps (Wang, V2 Mini)計算值。各自在 6.5（Pachler）、24（Wang，V1.5）、96 Gbps（Wang，V2 Mini） |
| Users at 100 Mbps per satellite每顆衛星可供 100 Mbps 的使用者數 | \(C_{\text{sat}} / 100\ \text{Mbps}\) | 65 / 240 / 960computed. The 960 at 96 Gbps is TinyLEO's own figure (§6.1), a consistency check計算值。96 Gbps 下的 960 正是 TinyLEO 自己的數字（§6.1），可作一致性檢查 |
| Ground area one satellite sees above 25°一顆衛星在 25° 以上所見的地面面積 | \(2\pi R_E^2\,(1-\cos 8.45^\circ)\) | 2.77 × 10⁶ km²computed. λ(25°) = 8.45° from Chapter 1計算值。λ(25°) = 8.45° 來自第 1 章 |
| Subscribers in that area at the rule of thumb該面積依經驗法則的訂戶數 | \(2.77\times 10^6 \times 100/300\) | ~925,000computed. Against 65–960 simultaneous users at 100 Mbps, so the rule is a subscriber-density estimate, with simultaneous demand defined separately計算值。對照 100 Mbps 下 65–960 位同時使用者，所以這條法則是訂戶密度，同時需求須另行定義 |
The propagation part of the 90° delay barely moves, yet the total changes a great deal.跨 90° 延遲中的傳播部分變化很小，總延遲卻變化很大。是參數表中的哪一列在起作用？
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Node delay | Set the shell to 24×66 and the node delay to 10 ms, then switch to 72×22. |
| Latency comparison | Which parameter-sheet row causes the difference? Explain why 2019 and 2020 ISL estimates can differ even when both assume 10 ms per hop. |
| 節點延遲 | 把殼層設為 24×66、節點延遲設為 10 ms，再切換到 72×22。 |
| 延遲比較 | 為什麼這使得一篇 2019 年論文的 ISL 延遲結果與一篇 2020 年論文的結果需要各自對照其幾何條件，即使兩者都假設每跳 10 ms？ |
4.5How to cite a satellite number in a paper如何在論文中引用一個衛星數字
The habits of Section 4.2 become a procedure when a number leaves the sheet and enters a manuscript. Reviewers ask five questions about each LEO constant. Answering them with the value makes its meaning and scope easier to assess. The fourth is the evidence tag, stated in the sentence rather than left to a reference list that readers rarely open. The fifth is the location in the source, down to the section, table, figure or paragraph number. This occurs because a citation to a forty-page FCC order or a twelve-page measurement paper requires an exact location for timely review.第 4.2 節的習慣在數字離開參數表、進入稿件時就成為一套程序。熟悉這批文獻的審稿人會對任何 LEO 常數提出五個問題，而一個事先回答了這五個問題的句子，既更容易辯護，也並能減少誤讀。第四個是證據標籤，須直接寫在句子裡，讓讀者就地查閱。第五個是在來源中的位置，精確到節、表、圖或段落編號，因為對一份四十頁的 FCC 命令或一篇十二頁的量測論文須同時附上具體位置，讓審稿人能在有限時間內查證。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Plane configuration | First identify the dated geometry filing. November 2018 gives 24 planes of 66, while August 2019 gives 72 planes of 22 for the 550 km shell. Intra-plane hop length differs threefold. |
| Filed and operational altitudes | Then distinguish filed geometry from flown geometry. FCC DA 26-36 authorizes shells operated at other altitudes in Ali et al.’s observations [16, Tab. 1]. Roughly 575 satellites remain in the canonical 552 km shell. |
| Model configuration | Identify measurement or modeling provenance. Tanveer et al.’s fifteen-second period is measured, while Hypatia latency counts modeled propagation. Mohan et al.’s 36–48 ms measurements and Handley’s 57–66 ms New York–London estimates also have separate scopes. |
| 軌道面配置 | 第一個問題是幾何來自哪一份申報、何時申報，因為「Starlink 的 550 km 殼層」在引用 2018 年 11 月修正的論文裡意指 24 個軌道面各 66 顆，在引用 2019 年 8 月的論文裡則是 72 個軌道面各 22 顆，兩者的同面跳段長度相差三倍。 |
| 軌道高度 | 第二個是這個數字描述的是申報中的星座還是實際飛行的星座，因為 FCC DA 26-36 授權的殼層，Ali 等人觀察到實際運作在各異高度[16, Tab. 1]，而標準的 552 km 殼層只剩約 575 顆衛星。 |
| 量測與模型範圍 | 第三個是數字是量測的還是建模的，這把 Tanveer 等人的 15 秒週期與 Hypatia 只算傳播的延遲分開，也把 Mohan 等人的 36–48 ms 與 Handley 的紐約–倫敦 57–66 ms 估計分開。 |
| Check檢查項 | Question to answer in the sentence句中要回答的問題 | Example that passes合格的範例 | Example that fails待修訂的範例 |
|---|---|---|---|
| Filing date申報日期 | Which dated FCC document fixes the geometry?哪一份附日期的 FCC 文件固定了幾何？ | "1,584 satellites at 550 km in 72 planes of 22 [18]"「550 km 的 1,584 顆衛星，72 個軌道面各 22 顆[18]」 | "Starlink's 550 km shell"「Starlink 的 550 km 殼層」 |
| Filed vs flown申報 vs 實飛 | Is this the authorized constellation or the one TLEs show?這是被授權的星座，還是 TLE 顯示的星座？ | "authorized at 530 km [27]. Observed at 559 km ([16, Tab. 1])"「授權高度 530 km[27]。觀測高度 559 km（[16, Tab. 1]）」 | "the 43° shell at 530 km"「530 km 的 43° 殼層」 |
| Measured vs modeled量測 vs 建模 | Did an instrument see it, or did a simulator produce it?是儀器看到的，還是模擬器產生的？ | "bent-pipe median RTT 36–48 ms measured from 98 RIPE Atlas probes [49, Fig. 12]"「由 98 個 RIPE Atlas 探針量得的彎管中位 RTT 36–48 ms[49, Fig. 12]」 | "Starlink latency is about 40 ms"「Starlink 延遲約 40 ms」 |
| Evidence tag證據標籤 | What kind of document, and can a reader check it?是哪一類文件，讀者可採用的查證方法為何？ | "100 Gbps per laser, a SpaceX figure reported from a Jan 2024 SPIE talk (TechSpot) [company via secondary]"「單雷射 100 Gbps，為 2024 年 1 月 SPIE 演講中 SpaceX 的數字（TechSpot 報導）[公司（二手）]」 | "Starlink lasers run at 100 Gbps [12]"「Starlink 雷射以 100 Gbps 運作 [12]」 |
| Location in source來源中的位置 | Which section, table, figure or paragraph?哪一節、表、圖或段落？ | "≤100 s cold-start acquisition (SDA OCT Standard v3.0.1, req. OCT-016)"「冷啟動擷取 ≤100 s（SDA OCT 標準 v3.0.1，需求 OCT-016）」 | "acquisition takes about a minute [SDA]"「擷取約需一分鐘 [SDA]」 |
| Assumption declared聲明假設 | When a source is absent, is the assumption explicit?來源仍待補充時，這項選擇的假設性質有清楚標示嗎？ | "we assume 1 ms per hop, as Bhattacharjee et al. do, pending measurement"「我們依 Bhattacharjee 等人假設每跳 1 ms，量測資料仍待補充」 | "the per-hop delay is 1 ms"「每跳延遲為 1 ms」 |
The ladder in Figure 4.3 is the picture to keep in mind while filling in the checklist. This occurs because it shows where each class of parameter actually sits. The scheduler period, the idle RTT medians, the downlink waveform and the line-of-sight limit stand on the top rung. Thus, a paper can state them as facts with a figure reference.
The shell geometry, the elevation mask and the acquisition requirement stand on the regulatory rung, where they are checkable but describe permissions rather than operations. Laser rates, terminal count, company latency medians and V3 capacity occupy the company-claim rung. Report them with explicit attribution: "SpaceX states".
填寫檢查表時腦中要留著圖 4.3 的梯子，因為它顯示了每一類參數實際上位於哪裡。排程器週期、閒置 RTT 中位數、下行波形與視線極限站在最上層，所以論文可以附上圖號把它們當作事實陳述。殼層幾何、仰角遮罩與擷取需求站在監管那一層，可以查證，但描述的是許可範圍，實際運作則須另行量測。雷射速率、終端數、公司自己的延遲中位數與 V3 容量站在最底層，唯一誠實的寫法是「SpaceX 表示」。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Orbital altitude | Flown altitudes, the 74 ms spike, 24 and 96 Gbps capacities, and PoP count occupy the preprint rung. These current estimates can change during review. |
| Evidence for computed values | Computed numbers sit beside the ladder rather than on it. Computed values inherit their weakest input’s evidence class. The 550 km period uses textbook physics. The 1,969 km hop uses regulatory geometry. The 3.9 Tbps average derived from 42 PB daily retains company-claim status. |
| 軌道高度 | 實飛高度、74 ms 尖峰、24 與 96 Gbps 容量以及 PoP 數量站在預印本那一層，它們是目前最好的數字，但在審查中仍可能改變。 |
| 計算值的證據層級 | 計算值依其輸入的證據層級定位，因為每一個都繼承其最弱輸入所在的那一層：550 km 的週期是教科書物理，1,969 km 的跳段是監管層級，而每日 42 PB 所隱含的 3.9 Tbps 平均速率則是偽裝過的公司宣稱。 |
Seven quantities require further public evidence: active topology, primary laser rate, node delay and scheduler allocation rules. The remaining quantities are ISL traffic share, Gen1 capacity, and V2 Mini mass and power. Each must be declared as an assumption. Citing a satellite number means stating its filing date, filed-versus-flown status, measured-versus-modeled status, evidence tag and location in the source.模擬器需要的七個量仍待 SpaceX 以外的公開證據，即拓樸、原始單雷射速率、每跳節點延遲、排程器的分配規則、ISL 流量比例、Gen1 單顆容量以及 V2 Mini 的質量與功率，每一個都必須聲明為假設。若要將餘下部分拆分為節點延遲與排程器延遲，還須更多證據。五跳 ISL 鏈約需 103 ms，只有在幾千公里以上才勝過光纖。引用一個衛星數字，意指寫明它的申報日期、申報或實飛、量測或建模、證據標籤以及在來源中的位置。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Model configuration | Three source classes supply LEO constants. Regulatory filings support frozen 2018–2021 models, while campaigns from 2022 onward measure the deployed network. Company statements await independent verification. Section 4.2’s thirty-row sheet carries evidence tags and source locations, with disputed values identified. |
| Node delay | Geometry, assumed 1 ms nodes and a measured ground segment explain about 27 ms of 36–48 ms bent-pipe RTT. Attributing the remainder needs further evidence. A five-hop chain costs about 103 ms, with fiber crossover beyond a few thousand kilometers. Statistical multiplexing reconciles 100 users per 300 km² with 6.5–96 Gbps per satellite. |
| 評估條件 | 一篇 LEO 網路論文的常數繼承自三類來源：會被後續修正取代、且被 2018–2021 年建模論文凍結的監管申報，2022 年起描述真正在飛的網路的量測活動，以及仍待獨立驗證的公司說法，因此第 4.2 節的參數表在三十列的每一列都帶有證據標籤與來源位置，而且將各異的數值分開列示。 |
| 星座配置 | 完整範例顯示幾何、每顆衛星假設的 1 ms 與一段量測到的地面區段，解釋了量測到的 36–48 ms 彎管往返中的約 27 ms。 |
| 訂戶密度與容量 | 而每 300 km² 100 位使用者的法則是訂戶密度，只有統計多工才能把它與每顆衛星 6.5–96 Gbps 的容量調和。 |
Chapter 5 begins Part II with Handley’s fiber crossover, +Grid, motifs, ground relays and real-ephemeris topologies from 2026. The parameter sheet identifies each result’s filing, ISL count and node-delay assumption. The sheet itself returns in every digest's "Limits" section and in Chapter 10 , where its terrestrial reference row is set against the constellation rows.
第 5 章以回到建模時代本身來開啟第二部：Handley 的「3,000 km 以上勝過光纖」論證、+Grid 與 motif 拓樸、地面中繼，以及 2026 年建立在真實星曆上的拓樸，每一項都對照參數表來閱讀，使它所凍結的申報、所假設的 ISL 數與所選擇的節點延遲都清楚可見。參數表本身會在每篇摘要的「限制」一節以及第 10 章再次出現，在那裡它的地面參考列會與星座各列並置比較。