Internet Backbones in Space太空中的網際網路骨幹
A satellite transit network presents changing attachment points to Internet routing. This paper compares four integration designs and quantifies their connectivity, ground-infrastructure and weather costs. Its conclusions depend on a dated constellation model and propagation-only paths.衛星中轉網路向網際網路呈現持續變化的連接點。研究比較四種整合方式，量化連線能力、地面設施與天氣的成本，其結論適用於指定年代的星座模型及純傳播路徑。
Giuliari and colleagues examine a LEO constellation as a BGP-facing transit autonomous system. Their top-venue study asks how the wider Internet sees its changing connectivity. The paper argues for more than halving long-distance latency through path-aware or CDN-like integration. BGP alone offers a separate integration model. All of it is propagation-only simulation on a filing SpaceX has since replaced.Giuliari 與同事是頂級會議中唯一把 LEO 星座當作面向 BGP 的中轉自治系統、並追問網際網路的其他部分會看到什麼的作者。論文結論是衛星路徑可以把長距離延遲削減超過一半，但只能透過路徑感知或 CDN 式的整合BGP 的直接整合則面臨連線變動問題，而這一切都是對一份 SpaceX 之後已取代的申請文件所做的純傳播模擬。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground infrastructure | The simulation uses then-filed Starlink Phase 2, four ISLs per satellite and 1,833 ground-station cities between 56° S and 56° N. The elevation mask is 40°. A six-minute visibility filter still leaves about 20 daily connectivity events at 40° latitude. Filtering discards 15–45% of available connectivity. [73, §2.4, Fig. 1] |
| Ground infrastructure | A terrestrial backup network raises ground-segment cost to about $27–30 billion, compared with roughly $10 billion for the constellation. Restricting stations to 353 IXPs adds about 10 ms mean latency. Rain rerouting loses a few percent of traffic on average. Loss exceeds 30% in the average worst case and 80% in the worst scenario. [73, §3, Tab. 1] [73, §6.1, Fig. 2] [73, §6.2, Fig. 3] |
| 星座配置 | 他們模擬 SpaceX 當時申請的 Starlink 第二階段，每顆衛星四條雷射鏈路，地面站設於南緯 56° 到北緯 56° 之間的 1,833 個城市，仰角遮罩 40°，發現即使經過丟棄可見時間較短衛星的六分鐘過濾器，位於 40° 緯度的地面站每天仍經歷接近 20 次連線與斷線事件，而過濾器本身浪費了 15–45% 的可用連線能力[73, §2.4, Fig. 1]。 |
| 評估條件 | 把這種變動藏在地面備援網路後面，地面段的成本約 270–300 億美元，對比星座本身約 100 億[73, §3, Tab. 1]。 |
| 延遲比較 | 只在 353 個網際網路交換中心設站平均增加約 10 ms 延遲[73, §6.1, Fig. 2]。 |
| 評估條件 | 而雨衰重新路由平均損失幾個百分點的流量，但平均最壞情況超過 30%、最壞情境超過 80%[73, §6.2, Fig. 3]。 |
1. Where this paper sits1. 這篇論文的位置
HotNets 2018 and CoNEXT 2019 establish low-latency laser paths. This CCR paper examines their Internet integration. Satellite attachment changes every few minutes, while inter-domain routing assumes links lasting months. The paper comes from the group behind the SCION path-aware architecture together with Singla.
Path awareness is one conclusion. Churn, cost and rain-fade results also stand independently. Chapter 5 cites latency in Section 5.1 and modeling assumptions in Section 5.7. This early study prices ground infrastructure and simulates a year of weather.
Later measurements show PoP placement and gateway availability strongly affecting real Starlink latency.
2018 年的兩篇 HotNets 論文與 2019 年的 CoNEXT 論文確立了配備雷射鏈路的星座可以提供低延遲路徑，而這篇 CCR 論文問的是任何想販售這些路徑的人接下來必須面對的問題：一個接入點每隔幾分鐘就改變的網路，要如何接到一個跨域路由假設鏈路能維持數月的網際網路？論文出自 SCION 路徑感知架構背後的團隊加上 Singla，所以它的答案部分是在為路徑感知辯護，但它對變動、成本與雨衰的量測本身站得住腳，這也是第 5 章在第 5.1 節引用它的延遲宣稱、在第 5.7 節引用它的假設的原因。它也是最早為地面段定價並模擬整年天氣的論文，而這兩個主題在量測時代後來被發現透過 PoP 位置與閘道可用性主導了真實的 Starlink 延遲。
2. The problem, measured2. 問題，用量測說話
The measurement at the heart of the paper is a count of events rather than a latency. Stations see satellites for a few minutes. If each visible satellite provides BGP connectivity, each arrival or departure changes that connectivity. Each event triggers at least one BGP update. Thus, a naive "white-box" satellite AS would flap continuously and create transient black holes [73, §2.4] . Figure A below draws the mechanism.論文核心的量測是連線事件計數，延遲則另行分析。地面站只能看到每顆衛星幾分鐘它就落下，如果星座透過該站向 BGP 宣告自己，每一顆可用衛星的到來與離去都是該站連線能力的改變，至少觸發一次 BGP 更新，所以一個樸素的「白箱」衛星 AS 會持續擺動並製造暫態黑洞[73, §2.4]。下方圖 A 繪出這個機制。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Constellation configuration | A visibility filter excludes short contacts. With a six-minute filter on 10% of Phase 2, a station at 40° still experiences about 20 daily events. The filter discards 15–45% of geometrically available connectivity. At higher latitudes, Fig. 1b reaches about 52 events daily, with waste approaching 100%. [73, §2.4, Fig. 1] |
| Ground infrastructure | The second measurement is economic: a ground station is costed at about $7 million of infrastructure plus about $3 million for 30 antennas |
| Terrestrial fiber cost | terrestrial fiber to interconnect the stations at $10,000 per kilometer |
| Wide-area network cost | and a Delaunay-triangulated wide-area network joining 1,833 stations at $8.5–12 billion |
| Total ground-segment cost | for a total ground segment of $27–30 billion against roughly $10 billion for the constellation itself [73, §3, Tab. 1], which the calc table reassembles. |
| Ground infrastructure | The weather experiment uses 2018 NOAA GFS-ANL reanalysis, sampled twice daily. 11 rain thresholds yield 7,744 scenarios requiring traffic movement around faded stations. [73, §6.2] |
| 星座配置 | 自然的修正是忽略可見時間少於某個窗口的衛星，作者在第二階段 10% 部署上模擬六分鐘過濾器：位於 40° 緯度的地面站每天仍看到接近 20 次事件，過濾器浪費了 15–45% 幾何上可用的連線能力，而在更高緯度圖 1b 顯示每天多達約 52 次事件、浪費接近 100%[73, §2.4, Fig. 1]。 |
| 地面設施 | 第二個量測是經濟的：一個地面站的成本估計為約 700 萬美元的基礎設施加上約 300 萬美元的 30 面天線，連接各站的地面光纖每公里 10,000 美元，連接 1,833 個站的 Delaunay 三角化廣域網路 85–120 億美元，地面段合計 270–300 億美元，對比星座本身約 100 億[73, §3, Tab. 1]，計算表重新組合了這些數字。 |
| 地面設施 | 第三個是氣象的：2018 年 NOAA GFS-ANL 再分析資料，每天取樣兩次並以 11 個雨量門檻切分，得到 7,744 個某些地面站被雨衰中斷、流量必須移動的情境[73, §6.2]。 |
| Question | 問題 | Formula | 公式 | Result | 結果 |
|---|---|---|---|---|---|
| Fibre latency in the paper's model relative to the vacuum geodesic論文模型中光纖延遲相對真空大圓的比值 | \( \tfrac{t_{fib}}{t_{c}}=\tfrac{2.3\,L/(2c/3)}{L/c}=1.5\times2.3 \) | 3.45computed from the paper's assumptions: fiber at 2c/3 along a path 2.3× the geodesic [73, App. A.2]. A satellite path with stretch below 3.45 beats fiber, below 1.73 halves it由論文假設計算：光纖以 2c/3 沿 2.3 倍大圓的路徑傳播（附錄 A.2）。拉伸低於 3.45 的衛星路徑勝過光纖，低於 1.73 則將其減半 | |||
| Ground segment cost from the per-station figures由每站數字得到的地面段成本 | \( 1{,}833\times(7+3)\ \text{M\$}+\text{WAN} \) | 18.3 B$ + 8.5–12 B$ ≈ 27–30 B$computed from §3 and Table 1. The paper prints 27–30 B$ and 11.85 B$ for the WAN in the text由 §3 與表 1 計算。論文正文印出 27–30 B$，WAN 為 11.85 B$ | |||
| Ratio of ground segment to constellation cost地面段與星座成本的比值 | \( \tfrac{27\text{–}30}{10} \) | ≈ 2.7–3.0×computed. The ≈ 10 B$ constellation figure is the paper's estimate for Starlink [73, Tab. 1]計算值。約 100 億美元的星座數字是論文對 Starlink 的估計（表 1） |
3. The design3. 設計
Rather than one system the paper compares four ways of integrating a satellite network AS, which Figure A arranges as a ladder. The white-box design runs BGP directly over the churning ground stations and is rejected because of the flapping and black holes just measured. The clean-slate design is path-aware networking in the style of SCION, in which endpoints choose among explicitly advertised paths. Thus, a satellite path can be offered and withdrawn with convergence handled through explicit path choice.
The paper uses it as the performance upper bound [73, §4] . The deployable CDN-inspired design directs traffic to nearby stations with good satellite paths. This resembles CDN steering among edge servers. This keeps BGP unchanged and can be done today [73, §5] . Two experiments then separate the designs.
論文比較四種整合衛星網路 AS 的方式，圖 A 把它們排成一道梯子。白箱設計直接在變動的地面站上執行 BGP，因為剛剛量到的擺動與黑洞而被排除。黑箱設計把星座藏在一個連接所有地面站的穩定地面廣域網路之後，使 BGP 看到單一穩定的 AS，而內部控制器把流量導向任何當下可用的衛星。全新架構設計是 SCION 式的路徑感知網路，端點在明確宣告的路徑中選擇，所以衛星路徑可以透過顯式路徑選擇直接提供與撤回，論文把它當作效能上界[73, §4]。
可部署的設計受 CDN 啟發：流量被導向當下有良好衛星路徑的鄰近地面站，就像 CDN 在邊緣伺服器之間引導客戶端，保留原有 BGP，今天就能做[73, §5]。接著兩個實驗區分這些設計。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground infrastructure | A terrestrial wide-area network interconnects ground stations in the black-box design. BGP sees one stable AS. An internal controller selects available satellite paths. This works but is what costs $27–30 billion [73, §3] . |
| Ground infrastructure | Using only 353 IXP sites instead of all cities adds about 10 ms mean latency. Relative penalties span −50% to +75%, largest for low-latency paths. The authors therefore call satellites "most beneficial for long-distance communication". [73, §6.1, Fig. 2] |
| Ground infrastructure | Rain then tests the two reactive designs, re-routing to another station against path control at the endpoints. Their loss CDFs are "very close", with path control giving more than 10% improvement in the 95th percentile [73, §6.2, §7]. |
| 評估條件 | 這可行，但正是花費 270–300 億美元的地方[73, §3]。 |
| 延遲比較 | 在每個城市設站對比只在 353 個 IXP 地點設站，顯示使用現有互連站點的延遲成本，平均約 10 ms，相對懲罰從 −50% 到 +75%，在端到端延遲最低處最大，由此得出衛星網路「對長距離通訊最有益」的觀察[73, §6.1, Fig. 2]。 |
| 評估條件 | 雨衰接著測試兩種反應式設計，重新路由到另一個站對比端點的路徑控制，兩者的損失 CDF「十分接近」，而路徑控制在第 95 百分位有超過 10% 的改善[73, §6.2, §7]。 |
Original figure note原始圖說出處
Drawn for this primer from Giuliari et al., CCR 2020, §2.4–§6 and Fig. 1.4. Evaluation setup4. 實驗設定
A custom orbital simulator supplies all results. Emulation, packet transport, TCP effects and live-network measurements lie outside the study. The constellation is described as Starlink "phase two", the first complete constellation with inter-satellite links, with altitude and plane count left unspecified. The primer's fact base records it as the Phase 2 filing of the time. Each satellite carries four laser links, two fore and aft in the same plane and two cross-plane, in the +Grid pattern [73, §6, App. A.2] .所有結果都來自自行開發的軌道模擬器，模擬平台、封包傳輸及實際網路量測皆位於範圍之外，TCP 效應位於討論範圍之外。星座被描述為 Starlink「第二階段」，即第一個配備星間鏈路的完整星座，但高度與軌道面數仍待補充，本教材的事實庫將它記錄為當時的第二階段申請文件。每顆衛星有四條雷射鏈路，兩條在同軌道面前後、兩條跨軌道面，為 +Grid 模式[73, §6]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground infrastructure | Ground stations occupy 1,833 cities above 300,000 population between 56° S and 56° N, or 353 IXP sites. The 40° minimum elevation exceeds Starlink's later 25° setting. They are interconnected by a Delaunay-triangulated terrestrial WAN in the black-box design. |
| Latency comparison | Satellite latency divides path length by light speed. Fiber latency uses speed 2c/3 over routes 2.3 times the geodesic length. Thus, both sides are propagation-only models with queueing, node delay, acquisition and link capacity outside the model. |
| Ground infrastructure | Rain tests use twice-daily 2018 NOAA GFS-ANL samples at 11 thresholds, producing 7,744 scenarios. Experiment 2 limits ground-station failures to rain. |
| 地面設施 | 地面站設在南緯 56° 到北緯 56° 之間人口超過 30 萬的 1,833 個城市，或改設在 353 個 IXP 地點，最低仰角 40°，比 Starlink 後來採用的 25° 嚴格，在黑箱設計中由 Delaunay 三角化的地面 WAN 互連。 |
| 延遲比較 | 延遲在衛星段為路徑長度除以光速，在地面光纖為沿 2.3 倍大圓路徑的長度除以 2c/3，所以兩側都是純傳播模型，排隊、節點延遲、對準時間及鏈路容量皆位於模型範圍之外。 |
| 地面設施 | 雨衰使用 2018 年 NOAA GFS-ANL 資料，每天取樣兩次、11 個門檻，得到 7,744 個情境，而第二個實驗假設地面站故障僅限雨衰。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Close to 20 connect/disconnect events per day per ground station at 40° latitude after a six-minute filter, wasting 15–45% of connectivity. Up to ≈ 52 events/day and ≈ 100% waste at high latitudes [73, §2.4, Fig. 1].經六分鐘過濾後，40° 緯度的每個地面站每天接近 20 次連線／斷線事件，浪費 15–45% 連線能力。高緯度每天多達約 52 次事件、浪費接近 100%[73, §2.4, Fig. 1]。 | 10% of Phase 2 deployed. The high-latitude values are read from Fig. 1b. Each event triggers at least one BGP update.第二階段部署 10%。高緯度數值由圖 1b 讀出。每次事件至少觸發一次 BGP 更新。 |
| Ground cost is about 27–30 B$, versus about 10 B$ for the constellation. Each station adds about 7 M$ infrastructure and 3 M$ for 30 antennas. Fiber costs 10 k$/km, with total WAN cost 8.5–12 B$. [73, §3, Tab. 1]地面段約 27–30 B$（每站約 7 M$ 加 30 面天線約 3 M$，WAN 每公里 10 k$，WAN 合計 8.5–12 B$），對比星座約 10 B$[73, §3, Tab. 1]。 | 1,833 stations at GDP-sampled or large-city locations. The authors call the model a lower bound. The text gives the WAN as 11.85 B$.1,833 個站位於 GDP 取樣或大城市地點。作者稱該模型為下限。正文給 WAN 為 11.85 B$。 |
| IXP-only placement adds ≈ 10 ms average latency over city placement. Relative change −50% to +75%, largest for short paths [73, §6.1, Fig. 2].只在 IXP 設站比在城市設站平均增加約 10 ms 延遲。相對變化 −50% 到 +75%，短路徑最大[73, §6.1, Fig. 2]。 | 353 IXP sites versus 1,833 cities. Percentages read from the figure. Basis for "most beneficial for long-distance communication".353 個 IXP 站點對比 1,833 個城市。百分比由圖讀出。「對長距離通訊最有益」的依據。 |
| Rain re-routing loses a few percent of traffic on average, > 30% average-maximum and > 80% worst case. Path control and re-routing CDFs "very close", path control > 10% better at the 95th percentile [73, §6.2, §7, Fig. 3].雨衰重新路由平均損失幾個百分點的流量，平均最壞 > 30%、最壞情況 > 80%。路徑控制與重新路由的 CDF「十分接近」，路徑控制在第 95 百分位好 > 10%[73, §6.2, §7, Fig. 3]。 | 7,744 scenarios from 2018 NOAA data. Station failures limited to rain. "Fluctuations of a few milliseconds are tolerable" is a qualitative remark, with measured jitter outside its evidence scope.來自 2018 年 NOAA 資料的 7,744 個情境。地面站故障限於雨衰。「幾毫秒的波動可容忍」是定性評論，量測抖動則位於證據範圍之外。 |
| Satellite paths "could cut long-range latency by more than half" [73, §9].衛星路徑「可以把長距離延遲削減超過一半」[73, §9]。 | Follows from the fiber model at 3.45× the vacuum geodesic [computed]. City-pair RTT values remain unreported, so the claim is qualitative in this paper.由 3.45 倍真空大圓的光纖模型推出［計算值］。城市對 RTT 數值仍待補充，所以在這篇論文中該宣稱是定性的。 |
6. Limits and open questions6. 限制與待解的問題
- The latency claim follows from the fiber model. The fiber model uses speed 2c/3 and route stretch 2.3×, implying 3.45 times vacuum-geodesic latency. City-pair RTTs are unreported. The "more than half" claim follows from those assumptions. Handley and Bhattacherjee supply modeled latency values. Measured Starlink figures describe separate evidence: about 40 ms floor and 25–50 ms idle medians (fact base Part C, item 58).延遲宣稱由光纖模型推得。論文城市對 RTT 數值仍待補充，光纖比較範圍限於 2c/3 與 2.3 倍模型，所以「超過一半」是假設光纖為真空大圓 3.45 倍的結果其性質為模型推論。Handley 與 Bhattacherjee 的論文提供了數字，而實測的 Starlink 下限約 40 ms、閒置中位數 25–50 ms（事實庫 C 部第 58 項）必須與它們全部分開。
- Phase 2 with four ISLs is a superseded model. The paper leaves altitude and plane count unspecified. Later Starlink shells occupy 340–570 km, including a canonical 72×22 layout (Part C, item 12). The four-terminal +Grid is a baseline. Actual three-terminal topology remains proprietary (Part C, items 13 and 55). Setup delay and 15-second scheduling lie outside the churn model (Part C, items 56–57). This means the real ground-station churn is governed by a scheduler that later measurements identified.四條 ISL 的第二階段是已被取代的模型。星座的高度與軌道面數數值待補，Starlink 的殼層此後已移到 340–570 公里、典型殼層為 72×22 面（C 部第 12 項），四終端的 +Grid 是建模基準而三終端的真實拓樸仍待公開（C 部第 13 與 55 項），變動模型中雷射建鏈延遲與 15 秒排程器皆位於範圍之外（C 部第 56–57 項），這表示真實的地面站變動由一個後續量測才辨識的排程器所主導。
- The 40° mask inflates churn. A 40° minimum elevation shortens every pass relative to the 25° mask Starlink adopted. Thus, the event counts and the 15–45% waste are upper estimates for the geometry that flew. The paper also stops at ±56° latitude, excluding the high-latitude users who in practice depend most on ISL paths.40° 遮罩放大了變動。相對於 Starlink 採用的 25° 遮罩，40° 最低仰角縮短了每次過頂，所以事件計數與 15–45% 的浪費對實際飛行的幾何而言是偏高的估計。論文也止於 ±56° 緯度，排除了實務上最依賴 ISL 路徑的高緯度使用者。
- Cost is a lower bound with heuristic placement. The authors say so themselves, the $10 billion constellation figure is an estimate rather than a disclosed number. The comparison omits the user terminals and the operating cost of the ground segment. Measured Starlink latency depends on its much smaller deployed PoP set. The hypothetical 1,833-city ground segment describes a separate design.成本是採啟發式選址的下限。作者自己這麼說，100 億美元的星座數字是估計值，公開部署數字則須另行確認，比較也省略了使用者終端與地面段的營運成本。量測時代後來發現 Starlink 的延遲由它實際建造的少得多的 PoP 決定，其配置與模型中的 1,833 城市地面段各異。
- Rain is the only failure. Experiment 2 covers weather outages through twice-daily snapshots. Minute-scale fades require finer data. Reported connectivity loss measures a different quantity from packet loss or RTT change. Thus, the "jitter" a reader might expect from the title is connectivity churn, with millisecond variance outside its measurement scope.雨是唯一的故障。實驗 2 假設地面站中斷僅限於天氣，使用每天兩次的再分析快照分鐘尺度的衰落則位於取樣範圍之外，報告的是連線能力的損失封包遺失及 RTT 變化則位於範圍之外，所以讀者可能從標題期待的「抖動」其實是連線變動，毫秒級變異則位於量測範圍之外。
- The SCION comparison is partly a position. The authors are the architects of SCION, the path-aware design is used as the upper bound. At the 95th percentile, black-box rain performance improves by "more than 10%", while the CDFs remain "very close". These modest gains support the deployable option.SCION 比較部分是立場。作者是 SCION 的設計者，路徑感知設計被用作上界，而它在雨衰下相對 CDN 式方案的優勢只是第 95 百分位幅度較小的「超過 10%」、其餘 CDF「十分接近」，所以論文自己的資料比其框架更支持可部署的選項。
7. How it connects7. 與其他內容的關聯
- Section 5.1 uses the "more than half" statement as the third of the founding latency claims and shows why it is qualitative here.第 5.1 節把「超過一半」的陳述當作三個奠基延遲宣稱中的第三個，並說明為什麼它在這裡是定性的。
- Section 5.7 lists the Phase 2 snapshot, four ISLs, 40° mask and absent scheduler among the superseded assumptions.第 5.7 節把第二階段快照、四條 ISL、40° 遮罩與缺席的排程器列入已被取代的假設。
- Section 5.2 treats the four-link +Grid this paper assumes as the baseline topology, and Section 2.5 gives its geometry.第 5.2 節把這篇論文假設的四鏈路 +Grid 當作基準拓樸，第 2.5 節給出其幾何。
- Section 3.3 explains deployed gateway and PoP placement as a determinant of latency. Section 3.4 gives measured rain effects.第 3.3 節描述 Starlink 實際建造的閘道與 PoP 結構，它以這篇論文地面段原本預期的方式決定了實測延遲，第 3.4 節給出實測的雨衰效應。
- Sibling digests: Handley 2018 and Gearing Up 2018 supply the latency numbers this paper's claim rests on. Handley 2019 is the other paper that prices ground infrastructure against lasers. CoNEXT 2019 optimizes the +Grid assumed here. Ali & Zhang 2026 and TinyLEO redo topology with three terminals. TinyLEO's geographic anycast is the modern answer to the stable-namespace problem this paper exposed. HitchHiking and Mohan 2024 show how PoP placement governs the real network's latency.相關導讀：Handley 2018 與 Gearing Up 2018 提供這篇論文宣稱所依賴的延遲數字。Handley 2019 是另一篇把地面基礎設施與雷射對比定價的論文。CoNEXT 2019 最佳化這裡假設的 +Grid。Ali 與 Zhang 2026 和 TinyLEO 以三個終端重做拓樸，而 TinyLEO 的地理任播是對這篇論文揭露的穩定命名空間問題的現代回答。HitchHiking 與 Mohan 2024 顯示 PoP 位置如何主導真實網路的延遲。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| ≈ 20 / day | Connect/disconnect events per ground station at 40° latitude after a 6-minute filter, 10% of Phase 2 deployed [73, §2.4, Fig. 1]. | 經 6 分鐘過濾後，40° 緯度每個地面站每天的連線／斷線事件數，第二階段部署 10%[73, §2.4, Fig. 1]。 | |
| 15–45% | Share of geometrically available connectivity the 6-minute filter discards at 40° latitude. Up to ≈ 100% at high latitudes [73]. | 6 分鐘過濾器在 40° 緯度丟棄的幾何可用連線能力比例。高緯度多達約 100%（圖 1b）。 | |
| 4 ISLs, Phase 2, 40° | Two in-plane and two cross-plane lasers per satellite on the then-filed Phase 2, 40° minimum elevation [73, §6, App. A.2]. Altitude and planes unspecified. | 當時申請的第二階段上每顆衛星兩條面內、兩條跨面雷射，最低仰角 40°[73, §6]。高度與軌道面數數值待補。 | |
| 1,833 / 353 | Ground-station sites: cities over 300,000 inhabitants within ±56° latitude / IXP locations [73, §6]. | 地面站站點：±56° 緯度內人口超過 30 萬的城市／IXP 地點[73, §6]。 | |
| ≈ 27–30 B$ vs ≈ 10 B$ | Ground segment (stations + antennas + WAN) against the constellation. The authors call it a lower bound [73, §3, Tab. 1]. | 地面段（站點＋天線＋WAN）對比星座。作者稱其為下限[73, §3, Tab. 1]。 | |
| ≈ 7 M$ + 3 M$ / 10 k$ per km | Per-station infrastructure plus 30 antennas / terrestrial WAN fiber cost [73, Tab. 1]. | 每站基礎設施加 30 面天線／地面 WAN 光纖每公里成本（表 1）。 | |
| + ≈ 10 ms | Average latency added by IXP-only station placement. Relative −50% to +75% [73, §6.1, Fig. 2]. | 只在 IXP 設站所增加的平均延遲。相對 −50% 到 +75%[73, §6.1, Fig. 2]。 | |
| few % / > 30% / > 80% | Rain re-routing loss: average / average-maximum / worst case over 7,744 NOAA 2018 scenarios [73, §6.2, Fig. 3]. | 雨衰重新路由損失：7,744 個 NOAA 2018 情境中的平均／平均最壞／最壞情況[73, §6.2, Fig. 3]。 | |
| 3.45× | Fibre latency relative to the vacuum geodesic in the paper's model, 1.5 × 2.3 [computed]. The basis of "more than half". | 論文模型中光纖延遲相對真空大圓的倍數，1.5 × 2.3［計算值］。「超過一半」的依據。 | |
| ~ 40 ms | Measured Starlink bent-pipe RTT floor, for contrast. External comparison (fact base Part C, item 58). | 實測的 Starlink 彎管 RTT 下限，供對照。外部對照（事實庫 C 部第 58 項）。 |