Extreme-range precipitation probability across global weather systems Authors: Suqin Q. Duan, Wei-Ming Tsai, Fiaz Ahmed, J. David Neelin Published: 2026 ABSTRACT Precipitating weather systems are major contributors to hazardous extreme rain, yet weather and climate models struggle to capture their dynamics and relative impacts. Using high-resolution observations, we examine precipitation probability distributions at hourly timescale linked to major system types across the globe. We show that the high-intensity tails of these distributions share a typical shape, which can be characterized by an intensity scale. Dynamic processes largely govern this universal shape, while thermodynamic factors modulate its magnitude. We find that intensity scales of major contributing systems largely match those of total precipitation due to shared thermodynamic environments, dominance of intense systems, and modest variations in dynamic scales. This universality of various system types’ precipitation probability tail supports risk assessment without overly depending on weather-type identification for most regions. We identify exceptions where intense tropical low-pressure systems or mesoscale convective system co-occurrences disproportionately affect extremes, yielding gray-swan-prone distributions. INTRODUCTION Precipitation is a fundamental component of the Earth system and a key element of the hydrological cycle. Its global and long-term mean behavior is relatively well constrained by energetic considerations (1–4). However, precipitation varies across a wide range of spatial and temporal scales (5–8). This variability renders its full probability distribution—and particularly its extremes—relevant for both scientific inquiry and societal impact. Extreme precipitation events are among the most damaging weather-related hazards, responsible for widespread flooding, power outages, infrastructure damage, and economic disruption (9–12). Accurately characterizing the statistical distribution and physical controls of extreme precipitation is therefore essential for climate diagnostics, model evaluation, and risk management. Much attention has been paid to how the intensity of extreme precipitation scales with warming, often guided by the thermodynamic arguments, i.e., the Clausius-Clapeyron relation (13–17). However, considerable uncertainties remain, in part because precipitation extremes do not arise from thermodynamics alone (18–21). In reality, many such events are generated by specific weather systems, such as atmospheric rivers (ARs), mesoscale convective systems (MCSs), low-pressure systems (LPSs), and fronts (Fronts), each involving complex interactions between moisture availability and dynamical processes such as convergence and uplift (22–27). Standard climate models (28, 29) often lack the resolution or physical detail needed to adequately resolve these extreme-producing weather systems, leading to biases in the simulation of precipitation intensity (30–32). Moreover, tracking these weather features in observations or model output is computationally intensive and often inconsistent across algorithms (33–35). Bridging weather scale processes and climate-scale diagnostics across diverse weather systems, and connecting the statistical characteristics of extreme precipitation to large-scale thermodynamic and dynamic conditions, remains an unmet challenge. Here, we address this challenge by using high-resolution satellite observations and reanalyses to systematically characterize precipitation extremes associated with major weather phenomena. We examine how the high-intensity tails of precipitation probability distributions vary across weather systems, seasons, and locations, and how they reflect underlying thermodynamic and dynamic factors. First, we show that the tails of precipitation probability distributions can be effectively captured by a precipitation scale associated with the exponentially decaying tail. We then analyze how this precipitation scale, denoted PL, varies across weather systems and environments. We examine how a mixture distribution (36) of weather-system components shapes the total precipitation distribution, and identify cases where a less-common weather system type can contribute disproportionately to extreme events, yielding distributions prone to “gray-swan” events (37–39). Last, we disentangle the thermodynamic and dynamic controls on PL, including the dependence on weather type, and discuss how these relationships can inform future projections, even in the absence of explicit weather-feature identification. RESULTS Precipitation scale for the probability distribution high-intensity range Probability distributions and extremes in Earth science are commonly described using summary statistics such as the mean, standard deviation, percentiles, or return levels estimated from extreme value theory (40–48). The first two measures can well depict quantities whose probability distribution is closer to Gaussian, such as temperature. For precipitation (P), its probability density function [PDF; denoted as f(P)] has a form closer to a gamma distribution, with a scale-free regime at low intensities and a scale-dependent regime at high intensities (49–51). It has been shown that using a functional form f(P)∼cP−τG(−P/PL), where G is often modeled as exponential and c is a normalizing factor, precipitation PDFs can be characterized by two parameters: a power-law exponent τ governing the low to moderate precipitation range, and a characteristic probability cutoff scale PL governing the high-intensity precipitation range (50–53), providing a more tailored representation than mean and standard deviation. For G approximated as exponential, f(P)∼cP−τe−P/PL. When P increases by an amount PL, the probability density approximately drops by a factor of e−1. In log-PDF–linear-P coordinates, PL is inversely proportional to the slope of the PDF decrease in the high-intensity rangelog f(P)=c0−τlogP−1PLP (1)where c0=log c. The larger the PL value, the shallower the slope of the probability density decrease, and the more likely extreme values of P occur. Figure 1 (A and D) shows examples of observed precipitation PDFs from Integrated Multi-satellitE Retrievals for Global Precipitation Measurement mission (GPM IMERG) data (54) at locations with a variety of geographical and meteorological conditions—including warm tropical oceans (blue), moist continental regions (green/yellow), and drier land areas (brown)—for June–August (Fig. 1A) and December–February (Fig. 1D). The warm tropical oceans exhibit the longest tails (i.e., shallowest slopes) and the largest PL values around 15 to 20 mm hour−1. This indicates that the probability of high-intensity precipitation decreases most slowly in these regions, making extreme events more likely. In contrast, moist land regions typically have intermediate PL values around 10 mm hour−1, while drier land areas exhibit the steepest tails and smaller PL values near 5 mm hour−1, signifying a rapid drop-off in the likelihood of extreme rain. Precipitation probability distribution across example locations and seasons, illustrating the collapse obtained using the intensity scale. (A) Observed precipitation PDFs (solid lines) during June, July, August (JJA) from IMERG at eight selected locations: tropical/subtropical oceans [South Pacific (SPaci), West Pacific (WPaci), and Indian Ocean (IndO); blues], moist land regions [Amazon rainforest (Amazon), southeast US (SEUS), and northeast US (NEUS); greens/yellow], and dry continental regions [western continental US (WConUS) and northern China (NChina); browns]. The eight locations are labeled using pink boxes in fig. S1. For three of them—West Pacific, Amazon rainforest, and western continental US—we show the statistical fits (dashed lines) used to estimate the high-intensity range decay scale PL, and use symbols to mark a few key statistics: nonraining conditions (P<0.02 mm hour−1, circle), PL (triangle), and the 99th (star), 99.9th (oblique cross), and 99.99th (upright cross) percentiles of raining conditions (i.e., P≥0.02 mm hour−1). (B) Same as (A), but PDFs are rescaled by their corresponding PL values (annotated at top right), and aligned at P=2PL (marked by the vertical gray line). (C) Same as (A), but PDFs are rescaled by their raining 99th percentile (P¨99th) and aligned at P=P¨99th (marked by the vertical gray line). (D to F) Same as (A) to (C) but for December, January, and February (DJF). To aid interpretation, we annotate the PDFs (Fig. 1, A and D) for three representative locations—tropical ocean (West Pacific, blue), moist land (Amazon rainforest, green), and dry land (western continental US, brown)—with a few key statistics: the frequency of zero precipitation (solid circles), the 99th (star), 99.9th (oblique cross), and 99.99th (upright cross) percentiles (calculated from nonzero values), and the fitted PL value (triangle). These markers highlight that the “tail” evident in the log-linear space corresponds to extreme events exceeding the 99th and extending beyond the 99.99th percentile. We refer to this exponential decrease range [linear in the log(f(P))−P coordinates] as the “extreme-range” or “high-intensity range.” Dashed lines show the fits used to estimate PL (see the “Precipitation PDFs and PL” section in Materials and Methods), which align with the observed PDFs. When these PDFs are rescaled by their local PL values, their upper tails collapse onto a similar curve (Fig. 1, B and E), demonstrating that PL serves as a robust, characteristic scale for describing the tail behavior of precipitation distributions across regions and seasons. For comparison, we also rescale the PDFs by the local raining 99th percentile of precipitation (P¨99th; Fig. 1, C and F). Unlike being rescaled by PL, this does not yield a universal collapse of the tails, indicating that P¨99th is less effective than PL at capturing the full shape of the upper tail, especially over a broad intensity range. Weather feature extreme-range precipitation PDFs track the total distribution Globally, more than 50% of total precipitation and 75% of extreme precipitation are associated with weather systems and their co-occurrences (25, 27). Examining the precipitation probability distributions (PDFs) and extremes linked to different weather systems is essential for understanding the physical processes shaping precipitation distributions and patterns, and for improving the simulation and prediction of future extreme precipitation risks. At a given location, precipitation from different systems can be thought of as components of a mixture distribution, where the total PDF is the weighted sum of the component PDFs, with weights reflecting their relative frequencies. Figure 2 illustrates three representative scenarios in which contributions from two weather types combine to shape the total distribution, focusing on the behavior of their high-intensity tails: 1) Similar-tail mixture: Both systems have similar tails, so the total inherits the same tail shape regardless of their relative contributions. 2) Single-tail dominance: One system (weather type 2) has a much longer tail that dominates the extremes, while the other (weather type 1) drops off well before reaching high intensities and thus contributes little to the extreme range of total precipitation. 3) Gray-swan-prone mixture: Both system types have sufficiently long tails to influence the total extremes, but the less frequent system (weather type 2) has a longer tail and emerges as dominant at the very highest intensities. We describe this as a gray-swan-prone distribution, since such events can be underestimated when extrapolating from the less extreme portions of the total distribution or from limited samples (38), even though their probability is, in principle, predictable with sufficient care. Schematics illustrating three relevant scenarios for how weather system types combine to shape the total precipitation distribution tails. (A) Similar-tail mixture: Both system types have the same PL=20 mm hour−1, with system type 2 (wtype2; orange-red dashed) occurring 20% times as frequently as system type 1 (wtype1; teal dashed), yielding a total distribution (black solid) with the same tail form. (B) Single-type tail dominance: The total distribution (black solid line) is set primarily by system type 2 (PL=20 mm hour−1), while system type 1 (PL=5 mm hour−1) contributes little to the tail despite equal overall frequency. (C) Gray-swan-prone mixture: The total distribution is shaped by system type 1 (PL=15 mm hour−1) and a less frequent (20% of system type 1) but more intense system type 2 (PL=20 mm hour−1). The rarer high-intensity system emerges to disproportionately affect the tail at the very extreme end. In all three scenarios, we use τ=1 for both system types. Fitted PL values of the constructed total distributions are shown in the legends. In between the second and the third scenarios, there could be the situation where both weather types contribute to influence the total extremes, and the type with a longer tail is also more frequent. For this situation, extrapolating from the total distribution tends not to underestimate the risk brought by a weather system component. For all three scenarios illustrated in Fig. 2, we set τ=1 for both system types; the more extreme system type 2 is assigned PL=20 mm hour−1, akin to conditions of a location in the West Pacific. System type 1 is given different PL to illustrate the cases. Frequency distributions are normalized by bin size and by the total count of the constructed total distribution (equal to the sum of counts of the two system types). Under this normalization, the total distribution corresponds to a probability density, while the component distributions indicate their relative contributions to the total. In the single-tail dominance case (Fig. 2B), the tail of the total distribution is, not unexpectedly, almost entirely set by the more extreme system type 2, and the fitted PL closely matches that of this system. Although the mixture distribution for the total is not exactly of the form as Eq. 1, the fit works well (see caveats in the “Precipitation PDFs and PL” section in Materials and Methods). In the gray-swan-prone mixture case (Fig. 2C), systems with PL=15 and PL=20 combine to yield a fitted total PL of 16.5. Since system type 2 is less frequent, and its dominance does not show up until the very extreme end of the total PDF, the distribution used for risk assessment may miss this effect if not carefully sampled. For instance, conducting risk assessment based solely on the total fit (i.e., using PL=16.5 to estimate the extreme probability) could underestimate the actual extremes (the contributing cases may follow a distribution with PL=20). This problem may be exacerbated if model simulations, whose outputs feed risk estimates, fail to sufficiently capture the dynamics of such systems. In the similar-tail mixture case (Fig. 2A), both system types have identical tails, producing a total with the same decay shape. While one might expect single-tail dominance to be the prevailing scenario, similar-tail mixtures turn out to be unexpectedly relevant, which we highlight in the observational results below. What do the PDFs of actual weather-system components look like? To answer this, we turn to observations. Using an existing set of assembled weather feature identifications (27, 55) based on state-of-the-art satellite observations GPM-IMERG (54) and the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis ERA5 (56), we examine the precipitation PDFs of four major weather systems: ARs, Fronts, MCSs, LPS (including tropical cyclones), and four common types of co-occurrences between them: Front-AR, Front-AR-MCS, LPS-MCS, and Front-AR-MCS-LPS (All). Figure 3 (A to D) shows normalized frequency distributions for the total precipitation and the weather feature contributions at four example locations, encompassing tropical ocean (WPaci), midlatitude coastal land (NEUS), subtropical moist land (SEUS), and dry land (WConUS) during June–August. The distribution for total precipitation is normalized as probability density, while normalizations for weather features reflect their contribution to the total (see the “Precipitation PDFs and PL” section in Materials and Methods). For example, during summertime, MCSs and LPS-MCSs, followed by LPSs, are the primary contributors to extreme precipitation in the tropical West Pacific. In contrast, Front-AR-MCSs and Front-ARs dominate in the northeastern US. In the southeastern US, extreme rainfall is primarily associated with MCSs, while both MCSs and Fronts are major contributors in the western continental US. Precipitation probability distributions associated with major contributing weather systems. (A to D) Normalized frequency distribution during June–August for total precipitation (black) and for selected weather systems: AR (yellow), Fronts (pink), MCSs (orange), LPSs (blue), occurring alone, and their co-occurrences—Front-AR (purple), Front-AR-MCS (magenta), LPS-MCS (crimson), and all four combined Front-AR-MCS-LPS (All; indigo). The frequency distributions are normalized by bin width and the total count of the total precipitation so that the total precipitation curve is equivalent to the probability density, and the weather system curves reflect their frequency contributions in each intensity bin (see the “Precipitation PDFs and PL” section in Materials and Methods for details). Thick lines denote weather systems contributing more than 10% to the local total extreme precipitation (i.e., for P>PL). Uncertainty in each bin is shown as error bars representing the standard deviation across 50 bootstrap groups (see the “Bootstrapping” section in Materials and Methods for details). (E to H) Precipitation PDFs from weather systems that contribute more than 10% to the local total extreme precipitation (contributions are annotated in the legend), rescaled by PL of the total precipitation. The rescaled PDFs are aligned at P=2PL to highlight the tail similarity. The four demonstrated locations are labeled in Fig. 8. Despite the difference in the contribution at each location, weather features, especially those that make major contributions to the total extreme precipitation, have PDF tails highly parallel with each other and with the total, indicating that they have a similar probability decay rate in the high-intensity range. As a reference, the raining 99th percentile is marked on each PDF. Notably, while the tail slopes are similar among major contributors, their respective 99th percentiles differ, suggesting that the percentile-based thresholds alone may not holistically capture the common structure in the extreme precipitation tails. Can a single parameter characterize the precipitation tail across all major contributing weather systems? In Fig. 3, (E to H), we retain weather systems that contribute more than 10% to the high-intensity range (P>PL) of total precipitation and show their PDFs alongside the total, rescaled by the decay scale of total precipitation PL_tot (see the “Precipitation PDFs and PL” section in Materials and Methods for details). The rescaled PDFs are aligned at P=2PL to illustrate tail similarity: When scaled by the local parameter PL_tot, the tails of the PDFs from major contributing weather systems collapse onto one another. This indicates that a single parameter can effectively characterize the leading-order tail behavior across different weather systems, potentially simplifying the assessment of extremes and their associated risks. In Fig. 4, we evaluate how the decay scale of precipitation associated with individual weather features compares to that of the total precipitation more generally across regions and seasons. Locations where a given feature contributes more than 10% to the total extreme precipitation are highlighted with solid markers—circles for the tropics and triangles for the extratropics—colored by their contribution to the total extreme rain. The spatial distribution of PL for precipitation associated with different weather features, along with regions where they make substantial contributions to total extreme rainfall, is shown in fig. S1. Different features dominate in different regions and seasons: in the tropics, MCSs primarily drive extreme precipitation, while in the extratropics, the dominant contributors are Front-AR-MCSs, Front-ARs, and, in some regions, Fronts alone. The average PL values for feature-specific precipitation—computed over locations and seasons where each feature contributes more than 10% to total extremes—decrease as follows: around 16 mm hour−1 for LPSs, 14 mm hour−1 for LPS-MCS co-occurrences, 11 mm hour−1 for both Front-AR-MCS-LPS and MCSs, 7 mm hour−1 for Front-AR-MCS, and around 5 mm hour−1 for Fronts, Front-ARs, and ARs. The mean, median, and interquartile ranges for each category are shown in the box plots on the right side of each panel. This ordering is broadly consistent with expectations based on the typical intensity of these weather systems. Comparison of the extreme-range precipitation scale (PL) from weather system-associated extreme rain against the total extreme precipitation. (A to H) Different weather systems or their co-occurrences: (A) AR only, (B) Front only, (C) MCS only, (D) LPS only, (E) Front-AR co-occurrence, (F) Front-AR-MCS, (G) LPS-MCS, and (H) co-occurrence of all four (Front-AR-MCS-LPS). Each point represents one location in one season. Circles denote tropical locations (30°S to 30°N), triangles denote extratropical locations (30° to 60°N/S). Solid-colored symbols indicate locations/seasons where the weather system contributes ≥10% to total extreme precipitation (P>PL); transparent symbols indicate <10% contribution. Contribution strength is further indicated by the yellow-red color scale. The thin gray line marks the 1:1 reference. Oblique cross symbols highlight four example locations (as in Fig. 3 and labeled in Fig. 8), with error bars showing one standard deviation from bootstrapping; in some cases, error bars are smaller than the symbol size. Data points with fractional bootstrapping error >25% in either the feature or total PL are excluded. At the right of each panel, a box plot (blue) summarizes the distribution of feature-specific PL values for the retained data points (solid-colored symbols), showing the mean (black triangle), median (black line), interquartile range (box), and full range contained by Q1−1.5(Q3−Q1) and Q3+1.5(Q3−Q1) where Q1 and Q3 are the 25th and 75th percentiles (whiskers; points beyond the whiskers are considered as outliers). To further interpret Fig. 4, we highlight the four example locations—marked with oblique crosses—whose PDFs are shown in Fig. 3. The tropical West Pacific lies at the high end of the PL spectrum, where MCSs and MCS-LPS co-occurrences contribute substantially to extreme precipitation. In contrast, the western and southeastern US lie at the lower end of PL, with MCSs and Fronts (occurring alone) being the dominant contributors. The estimation errors for PL from both total and feature-specific precipitation are small, with bootstrap-derived standard deviations represented by error bars that are often smaller than the plotting symbols. In the northeastern US, major contributors include Front-AR and Front-AR-MCS co-occurrences; both these features and the total precipitation exhibit intermediate PL values, with slightly larger estimation uncertainty compared to the other locations. For the remaining locations, we exclude those with fractional estimation errors exceeding 25% in either feature-specific or total PL. Nonetheless, despite differences in the roles, structures, and spatial distribution of various weather features, their precipitation PDFs exhibit intensity scales that tend to follow that of the total precipitation (the correlation between the weather-specific and the total precipitation PL is above 0.8 for most weather types; see fig. S3). In other words, while the extreme probability decay scale varies by location—and different weather features cluster in different locations—it is not strongly dependent on the type of precipitating weather system at each location. Beyond this first-order alignment, a second-order difference is observed: Features containing MCSs tend to lie closer to the 1:1 line, whereas Fronts and ARs tend to fall below it. This suggests that convective systems are more likely to produce higher-intensity extremes and to dominate total extreme precipitation. In addition, LPSs with larger PL values tend to align with or exceed the 1:1 line—these are likely associated with tropical cyclones or severe tropical storms. Returning to the three conceptual scenarios of tail mixtures, the results in Figs. 3 and 4 show that, in most cases, weather-system components shape the total as either similar-tail mixtures (e.g., Fig. 3, A, B, and D, and points along the 1:1 lines in Fig. 4) or single-tail dominance (e.g., Fig. 3C). For risk assessment, this means that the intensity scale (or related extreme statistics) estimated from the total precipitation would be generally reliable. However, exceptions arise when a system’s intensity scale exceeds that of the total, leading to potential underestimation of “gray swan” risks. Figure 4 indicates that most such cases occur with LPSs (points above the 1:1 line). We seek these PL>PL_tot events, and in Fig. 5, we show the geographical distribution of these cases for LPSs and Front-AR-MCS co-occurrences and illustrate their PDFs. Illustration of the “gray-swan-prone” distributions and their geographic/seasonal locations. (A and B) Locations where LPS-driven PL>PL_tot events occur during June–August and December–February. Yellow dots mark cases where LPSs contribute <10% to total extreme precipitation (P>PL), and red triangles mark cases where LPSs contribute ≥10%. The four pink boxes indicate the four example regions shown in Figs. 3 and 4; the two green boxes highlight the regions further examined in (C) to (F): the West Pacific (WPaci2) and the southern Indian Ocean (SIndO). (C to F) Probability distributions for the two cases marked by green boxes in (A) and (B), shown in the same format as Fig. 3. In (E) and (F), the PDFs for the main weather types are shown for precipitation normalized by the total PL. (G to L) Similar to (A) to (F), but for PL>PL_tot events associated with Front-AR-MCS co-occurrences, with example locations shown for the Great Lakes (GrtLakes) and the Argentina coast (ArgenCoa). For LPSs, exceedances are concentrated in regions and seasons with high tropical storm activity (Fig. 5, A and B), and the associated PL values are particularly large (deep blue shading). We illustrate two cases, one in the West Pacific and another in the southern Indian Ocean (green boxes in Fig. 5, A and B), where LPSs contribute more than 10% to extreme rain. In both cases (Fig. 5, C to F), LPSs are the third-largest contributor, yet their PL values exceed those of the top two contributors and also surpass PL_tot. Visually, the difference of these PDFs may be small, but using PL_tot in such cases would not adequately capture the probability of these high-intensity LPS-driven gray swans, echoing the conceptual prototype shown in Fig. 2C. For Front-AR-MCSs, most PL>PL_tot cases are found in regions with moderate PL. For example, in the Great Lakes region (Fig. 5, I and K), Front-AR-MCS is the second-largest contributor, with PL=6.8 mm hour−1, higher than that of Fronts, the largest contributor, and slightly above PL_tot. A few regions show more severe cases. Along the Argentina-Atlantic coast, some Front-AR-MCS systems produce exceptionally large PL values that far exceed PL_tot (Fig. 5, J and L). The importance of Front-AR-MCS in the extreme tail implies that accounting in detail for changes in this weather type would be important for risk assessment in these regions. Thermodynamic and dynamic determinants of precipitation PDF extremes Why do precipitation from major contributing weather systems exhibit probability decay scales that commonly follow that of the total? We investigate this by first analyzing the key factors that potentially determine the intensity scale PL. Precipitation results from the condensation of water vapor as it rises, and the macroscopic control of precipitation can be understood in terms of two primary factors: atmospheric water vapor and the strength of vertical motion, or, equivalently, via mass continuity, horizontal mass convergence, especially at low altitudes where water vapor content is high (see the “The C distribution, CL, and mass convergence” section in Materials and Methods for further details). Water vapor reflects the thermodynamic environment, while mass convergence constitutes the dynamic component. In Fig. 6A, we show the precipitation PDF at the West Pacific location and indicate the average column-integrated water vapor (CWV) in each precipitation intensity bin. Across the tail of the distribution, CWV values remain fairly homogeneous, centered around 65 mm. In fact, the CWV distribution during heavily precipitating conditions (shown by the lime line in the inset) spans a narrow range, with the 10th and 90th percentiles around 60 and 71 mm, respectively. Even without conditioning on heavy precipitation (cyan lines in the inset), the temporal variability in CWV is moderate—note that values corresponding to densities below 10−2 are uncommon. This homogeneity in CWV during extreme precipitation at a given location contrasts with the much broader range observed across different regions and seasons, from near zero to over 70 mm (Fig. 6, A and D, and fig. S4). Therefore, we use the average CWV during extreme events P>PL, denoted as CWVP>PL¯, as a proxy for the thermodynamic condition of extreme precipitation at each location. Thermodynamic and dynamic controls on the high-intensity range precipitation scale PL. (A) Precipitation PDF for the West Pacific grid box during June–August (JJA), with bin-averaged CWV overlaid as colored dots. Inset: PDFs of CWV for all conditions (dashed cyan), raining (P>0.02 mm hour−1, solid cyan), and heavily raining (P>PL, lime); red dots mark the 10th and 90th percentiles of the heavily raining CWV distribution. (B) PDF of the dynamic proxy C=P/CWV (purple; defined instantaneously; conditions with CWV<15 mm excluded) and the 850 hPa mass convergence under raining conditions (C850,P>0, magenta) for West Pacific JJA. Only the converging (positive) side of C850,P>0 is shown. The probability decay scale CL is estimated from the C distribution. A rescaled C850,P>0 distribution—by its own probability decay scale and CL, and anchored at C=2CL—is shown as the yellow line. (C) PDF of C [purple, repeated from (B)] and of P scaled by the average CWV during heavy rain, CWVP>PL¯ (blue). (B) and (C) are shown for more locations in figs. S5 and S6. (D to F) Scatter plots showing the relationship between PL and (D) CWVP>PL¯, (E) CL, and (F) their product CL×CWVP>PL¯, across seasons and locations. Tropical locations are shown as green circles; extratropical locations as magenta triangles. On the dynamic side, we explore a proxy for atmospheric convergence in Fig. 6B. A simple approach to isolating the dynamic contribution is to approximately remove leading thermodynamic effects from precipitation (1, 57). We define the ratio C=P/CWV, and treat C as a proxy for convergence, as well as other factors not accounted for by the simple measure of thermodynamics (see the “The C distribution, CL, and mass convergence” section in Materials and Methods). The distribution of C, calculated with instantaneous P and CWV, is shown as the purple line in Fig. 6B. For comparison, we also plot the distribution of mass convergence at 850 hPa from ERA5 (C850, magenta line). Both distributions exhibit exponential decay in the extreme (tail) range. Notably, the PDF of C shows a tail behavior similar to that of precipitation, while the PDF of CWV (inset of Fig. 6A) does not. This suggests that dynamics primarily shape the form of the precipitation PDF tail—i.e., the rate at which the probability decreases with increasing intensity—while thermodynamics modulate its magnitude. When the precipitation PDF is normalized by the thermodynamic constant CWVP>PL¯ (Fig. 6C, blue line), it exhibits a rate of decay at the tail closely resembling that of the C distribution (Fig. 6C, purple line), reinforcing the interpretation that dynamics govern the shape of the extreme tail. We estimate a dynamical decay scale, CL, from the C distribution using a similar fitting approach as for PL. This CL serves as a proxy for the dynamic contribution to the probability decay of extreme precipitation. When the ERA5 mass convergence at 850 hPa (C850) is normalized by its own decay scale and by CL, its PDF tail (Fig. 6B, yellow line) aligns with that of the C distribution, supporting the use of CL as an estimator for the dynamic decay scale. To reconstruct PL, we combine the thermodynamic and dynamic components by multiplying CWVP>PL¯ with CL. For instance, at the West Pacific location during June–August, CWVP>PL¯≈65 and CL≈0.3, yielding a value that approximately matches the PL value, which is about 20.1. This reconstruction is evaluated more broadly in Fig. 6 (D to F), across various locations and all four seasons. Each point represents one location (magenta triangles for extratropics and green circles for tropics). Both CWVP>PL¯ and CL show positive correlations with PL, and their product effectively recovers the values of PL. We present the relationships between each component with PL in log-log space, which linearizes the multiplicative relationship, to facilitate gauging the relative contributions. Variations in CWVP>PL¯ and in CL contribute comparably to the overall variations in PL. A multiple linear regression of log10PL on log10CWVP>PL¯ and log10CL yields coefficients of 0.69 and 0.67, respectively, indicating that each predictor accounts for ~40% of the relative explanatory importance in determining the variability of PL across locations and seasons. Results above combine CWV from ERA5, to have consistent values between land and ocean, with precipitation from IMERG retrievals, a choice that avoids imperfections in reanalysis precipitation location and intensity (56, 58). However, mismatches between the two datasets can occur. To assess the sensitivity of our findings to this, we repeated the analysis in Fig. 6 using both CWV and P from ERA5 (see fig. S7). Key results remain similar. ERA5 exhibits a tighter linear relationship between CWVP>PL¯ and PL (fig. S7D), and a more distinct separation among the CWV PDFs for all-sky, raining, and heavily raining conditions (fig. S7A, inset). These differences are likely due to the improved consistency between CWV and P when sourced from the same dataset. However, the intensity of heavy precipitation in ERA5 is substantially lower than that in IMERG, resulting in significantly smaller PL values. To intuitively illustrate the differences, we provide a visual comparison of ERA5 and IMERG precipitation snapshots overlaid on CWV fields at a corresponding moment in fig. S8. When using P from ERA5, the relative importance of CWV and CL in explaining variations in PL are ~50 and 30%, respectively. Having examined the determinants of PL, we now return to why PL for major contributing weather systems closely track that of the total. In Fig. 7 (A to H), we compare the thermodynamic factor CWVP>PL¯ and the dynamic scale CL between feature-specific precipitation and the total precipitation. To streamline the presentation, we show four representative weather feature categories: the combined set of AR-only, Front-only, and AR-Front co-occurrences—grouped due to their lower convective influence (A and E); MCSs (B and F); Front-AR-MCS co-occurrences (C and G); and the combined set of LPSs and LPS-MCS co-occurrences (D and H). Thermodynamic and dynamic relationships between representative weather feature precipitation and total precipitation. (A to D) Comparison of thermodynamic conditions, measured by the average CWV during heavy precipitation (CWVP>PL¯), between feature-specific and total precipitation. (E to H) Comparison of the probability decay scale associated with dynamics (CL) between feature-specific and total precipitation. (I to L) Thermodynamic scaling of feature-specific CWVP>PL¯ using surface air temperature (T2m) and pressure during total extreme precipitation (P>PL). To streamline visualization while retaining representativeness, we focus on four groups of weather systems: [(A), (E), and (I)] combined AR-only, Front-only, and AR-Front co-occurrences; [(B), (F), and (J)] MCSs without co-occurrence (MCS-only); [(C), (G), and (K)] co-occurring Fronts, ARs, and MCSs (Front × AR × MCS); [(D), (H), and (L)] combined LPSs and their co-occurrence with MCSs (LPS and LPS × MCS). Symbols and colors follow the convention used in Fig. 4. In (E) to (H), samples with bootstrapping error >25% in either the feature or total CL are excluded; bootstrap error bars shown for the example locations are often smaller than the symbol. In (I) to (L), samples where surface elevation exceeds the 850 hPa geopotential height are excluded. Also as in Fig. 4, on the right edge of each panel in (A) to (H), a box plot summarizes the distribution of feature-specific CWVP>PL¯ and CL values for the solid-colored data points, showing the mean (black triangle), median (black line), interquartile range (box), and full range contained by Q1−1.5(Q3−Q1) and Q3+1.5(Q3−Q1), where Q1 and Q3 are the 25th and 75th percentiles (whiskers; points beyond the whiskers are considered as outliers). For the thermodynamic environment, weather features exhibit strong alignment with the total, lying closely along the 1:1 line. This indicates that CWVP>PL¯ for feature-specific precipitation generally mirrors that of the total precipitation. Using a low-level specific humidity yields similar alignment (fig. S9). A portion of the contributing LPSs and LPS-MCSs (Fig. 7D) and Front-AR-MCSs (Fig. 7C) has slightly higher CWVP>PL¯ than the total. For the dynamical decay scale, overall, the CL values of weather features generally track those of the total precipitation, although some notable differences exist. For example, the contributing LPSs and LPS-MCSs (Fig. 7H)—some of which include tropical cyclones—tend to have higher CL values than the total (the regression slope is the largest among those in Fig. 7, E to H, and is greater than 1 if considering LPSs only). This likely reflects the rarity and intensity of such storms: While tropical cyclones do not occur frequently, their extreme intensity yields a large CL, whereas the total CL is influenced by a mix of systems. This is consistent with the >1:1 slope in Fig. 4D for PL. MCSs (Fig. 7F) exhibit CL values that match more closely to the total. Given their relatively high frequency and dominant contribution to extreme precipitation in many regions, MCSs can strongly influence the total CL. ARs, Fronts, and their co-occurrences (Fig. 7E), as well as Front-AR-MCSs (Fig. 7G), tend to have smaller CL values than the total. Beyond the comparison between the feature-specific versus the total precipitation, as in the discussion for PL, another notable aspect is the spatial distribution. LPSs and LPS-MCSs that contribute more than 10% to total extreme precipitation tend to occur in regions and seasons with particularly high values of CWVP>PL¯, averaging around 63 mm (Fig. 7D); whereas ARs, Fronts, Front-AR, and Front-AR-MCS co-occurrences, which concentrate more in the midlatitudes, show lower CWVP>PL¯ values, averaging around 30 to 35 mm (Fig. 7, A and C). MCSs, which occur in both the tropics and midlatitudes, average to a value in between ~50 mm (Fig. 7B). When the thermodynamic contribution is removed, the dynamical decay scale CL exhibits less regional contrast and more comparable values among features. ARs and Fronts alone, AR-Front co-occurrences, MCSs, and Front-AR-MCS co-occurrences all show similar mean values around 0.23 hour−1 (Fig. 7, E to G). LPSs and LPS-MCSs show a slightly higher mean value around 0.29 hour−1 (Fig. 7H). Note that CL is the probability decay scale of the C distribution. It is a proxy for convergence but is essentially a statistical measure and may not directly reflect the local dynamics of individual storms. By construction, CL contains all factors beyond large-scale CWV. These factors could include the vertical structure of temperatures and winds (59), low-level vorticity, baroclinic instability, and other environmental and microphysical influences (18, 19, 60). Taking a step further, we assess the extent to which the thermodynamic factor CWVP>PL¯ can be approximated using surface air temperature. Establishing such a relationship can be valuable for projecting PL under climate change scenarios: Assuming relatively modest changes in dynamics, warming-driven increases in temperature can drive increases in CWVP>PL¯ and, hence, in PL. Temperature scaling for precipitation extremes has been widely discussed in the context of climate change (13–15, 17, 61). Here, we examine the temperature scaling for the current-climate CWVP>PL¯ associated with different weather features. To this end, we apply a simple scaling formulation (see the “Temperature scaling of CWVP>PL¯” section in Materials and Methods), XWV(T2m,P>PL¯)=0.8⋅q2m,P>PL∗¯⋅Δp/(ρg), where q2m,P>PL∗ is the near-surface saturated specific humidity conditioned on P>PL, calculated as a function of near-surface temperature T2m and pressure, Δp is the effective vertical depth over which water vapor is concentrated (in pressure units), ρ is the density of air, g is the gravitational acceleration, and XWV stands for scaled CWV. In Fig. 7 (I to L), we compare CWVP>PL¯ during weather-feature precipitation with the estimate based on temperature during times with total precipitation P>PL. Since some of the assumptions in the scaling formula do not hold well for high-topography regions (see the “Temperature scaling of CWVP>PL¯” section in Materials and Methods and fig. S10), we filter out locations where the orographic height exceeds the geopotential height at 850 hPa in the comparison. Despite some variability, the observed and temperature-scaled CWV values lie close to the 1:1 line. This result indicates that, the (spatial and seasonal) variability due to the thermodynamics can largely be attributed to the variability in surface air temperature, regardless of the feature types. Regional and seasonal distribution of precipitation extremes with the weather-system lens Last, we revisit the precipitation extreme decay scale PL and discuss its regional and seasonal distribution with the lens of its dominant weather systems. Figure 8 (A and D) shows maps of total precipitation PL during June–August (JJA) and December–February (DJF). The pattern features tropics exhibiting higher PL values. Total PL is shaped by the major contributing weather systems, and here, we highlight two representative weather features with distinct seasonal cycles: Front-AR-MCS co-occurrences (midlatitude) and LPSs combined with LPS-MCS co-occurrences (tropical). Spatial distribution of the precipitation scale PL and its thermodynamically normalized values. (A to F) Maps of PL for total precipitation, Front-AR-MCS co-occurring systems, and LPSs (LPS only and LPS × MCS combined) during June–August [(A) to (C)] and December–February [(D) to (F)]. (G to L) Corresponding thermodynamically normalized PL values, i.e., PL/XWV(T2m,P>PL¯), for the same seasons and system types. Locations with insufficient data are left blank (see the “Precipitation PDFs and PL” section in Materials and Methods), and those with fractional bootstrap error in PL greater than 25% are masked (dark gray; see the “Bootstrapping” section in Materials and Methods). In (G) to (L), regions with low CWV (i.e., CWVP>PL¯<15 mm) are also masked, such as the midlatitudes in the Northern Hemisphere during December–February. Locations where these features contribute more than 10% to total precipitation are marked with yellow/red dots. For Front-AR-MCS, contributions are strongest during DJF in the Northern Hemisphere midlatitudes, corresponding with higher PL values relative to JJA or the Southern Hemisphere. In contrast, LPSs and LPS-MCS co-occurrences shift with the seasonal migration of the Intertropical Convergence Zone, peaking in the tropical West Pacific during JJA and in the tropical South Pacific during DJF—regions also characterized by large PL values. While the seasonal migration of these systems is well known, it is revealing to see how this migration manifests in the geographic and seasonal variability of PL. A large portion of the spatial and seasonal variability in PL is thermodynamically controlled—warmer and more humid tropical regions tend to exhibit higher PL values. When we remove the thermodynamic component using the temperature-scaled proxy XWV(T2m,P>PL¯), much of the spatial variability is diminished (compare Fig. 8, G and J, with Fig. 8, A and D). Nonetheless, the dynamic component also plays a role and is positively correlated with PL. Figure 8K reveals features resembling Northern Hemisphere storm tracks, suggesting a dynamic influence associated with baroclinic instability. Similarly, Fig. 8 (I and L) show that regions with more active LPS activities tend to have larger CL values, reflecting dynamic influences that may stem from factors such as stronger low-level vorticity. These variations contribute to the spatial pattern observed in Fig. 8 (C and F). DISCUSSION Implications for risk assessment An essential element of risk assessment is estimating the probability that precipitation exceeds a critical threshold Pc. This can be quantified using the exceedance (or survival) function F¯(Pc)≡Prob(P>Pc)=∫Pc∞f(P)dP, where f(P) is the precipitation PDF. Using the functional form introduced above, f(P)∼cP−τe−P/PL, the exceedance function becomes F¯(Pc)=c∫Pc∞P−τe−P/PLdP. For sufficiently large Pc, this integral is asymptotically proportional to cPLPc−τe−Pc/PL. Thus, while the power-law factor Pc−τ provides a slower modulation, the exponential term governed by PL controls the leading-order decay of the exceedance probability. In this sense, PL directly sets the dominant scale governing the likelihood of very extreme precipitation. A commonly used metric in risk assessment is the return level (38, 62) ZT, defined through the relation Prob(P>ZT)=1/T, where T is the return period. In the generalized extreme value (GEV) framework, return levels are typically estimated by fitting block maxima to a GEV distribution (47, 48), F(z)=exp−[1+ξz‐μσ]‐1/ξ, from which ZT can be derived. Within our framework, an analogous connection can be made by equating the exceedance probability to 1/T: F¯(ZT)∼cPLZT−τe−ZT/PL≈1/T. Both τ and PL influence the return level; however, for large ZT, the exponential term dominates, yielding the leading-order approximation ZT≈PLlogT. This relationship highlights that PL plays a central role in controlling how rapidly return levels increase with return period, while not replacing the full information contained in extreme value theory-based approaches. In practice, robust risk assessment still requires explicit estimation of return levels and their associated uncertainties, including potential nonstationarity arising from changes in absolute precipitating frequency, precipitation rates, and shifts in the relative occurrence of different weather system types. Our analysis focuses on the decay scale PL, for which we identify common behavior across major contributing weather systems; the power-law exponent τ may still vary among systems (see fig. S11). Even for the primary tail-controlling parameter PL, second-order differences of order 20% exist across contributing weather systems (Fig. 4). These differences become particularly relevant in gray-swan-prone cases, where a relatively infrequent weather system exhibits a larger PL than the total precipitation distribution. In such cases, its influence emerges primarily at very large thresholds Pc (or long return periods T), and changes in the frequency of these systems could disproportionately affect exceedance probabilities and associated return levels. We performed a thermodynamic-dynamic decomposition to explain the physical controls on PL and to understand why different weather system types tend to exhibit similar PL values in the present climate. Taking this framework a step further, one may ask whether such a decomposition could also inform future risk projections. The motivation for considering extensions to future climate lies in the commonly taken role of thermodynamics as a first-order control on changes in extreme precipitation intensity. Whether variability of thermodynamic conditions in the current climate could be taken as an analog for climate-change scaling is an ergodicity-like transferability assumption (63–67). However, while extreme precipitation variability within the current climate may inform future projections (64), the specific scaling rates are often not comparable (68). This is because the deviations from the Clausius-Clapeyron scaling are not directly transferable. Equally important, dynamic processes may not remain fixed (13, 20, 21). Here, the residual decay scale CL encapsulates the effects of multiple dynamic processes—including vertical velocity structure, wind shear, entrainment, and land-atmosphere interactions—that vary across weather systems and regions and may respond to warming in distinct ways (18, 69, 70). As shown in Fig. 7 (E to G), CL still exhibits appreciable spread across weather types for a given value of the total CL, indicating the intrinsic variability of the dynamic contribution. These results underline that understanding this variability and quantifying how it evolves with warming would be important for assessing the uncertainty associated with thermodynamics-based estimates of extreme precipitation risk. Our results imply that, in the present climate, the thermodynamic control on extreme precipitation is primarily set by the local temperature environment rather than by weather-system identity. If this similarity in thermodynamic environments across weather types persists under warming, then—for the thermodynamic contribution—changes in extreme precipitation risk could be approximated using information from the total precipitation distribution, without requiring explicit tracking of individual weather systems. In this restricted sense, future changes in the relative frequency of weather types would not strongly alter the thermodynamic scaling inferred from total precipitation. An exception is precipitation associated with the gray-swan-prone cases, such as strong LPSs, likely tropical cyclones, which show a slightly larger thermodynamic contribution (Fig. 7, D and L) and potentially higher sensitivity to mean temperature than the total. Summary and outlook Using a cutoff power-law distribution to describe the precipitation PDF (51, 53), we show that the intensity scale for the extreme-range probability decay, PL, serves as a characteristic parameter for the high-intensity tail of the distribution. It condenses key information and allows the upper tails across locations and seasons to be compared on a common scale. Unlike percentile-based metrics—which, while intuitive, can be skewed by variability in the low-intensity range—PL offers a robust alternative. By examining the thermodynamic (i.e., water vapor) and dynamic (i.e., mass convergence) controls on precipitation, we demonstrate that PL is shaped by both components. We use the average CWV during heavily precipitating conditions, CWVP>PL¯, as a proxy for the thermodynamic factor, and define a dynamic proxy, C, as instantaneous precipitation normalized by CWV. We show that this dynamic factor C yields the shape of the precipitation PDF’s tail and can be characterized by probability decay scale CL. The thermodynamic factor, which remains relatively stable under intense precipitation, modulates the magnitude of the precipitation intensity. While dynamic and thermodynamic scaling has long been discussed in the context of mean and extreme precipitation, there is no settled theory for the probability distribution of precipitation. These results provide observationally grounded evidence for distinct thermodynamic and dynamic influences on precipitation distribution tails. In particular, the decay of the dynamics is closely linked to the drop-off in the probability of extreme precipitation. This diagnostic separation, though approximate, provides a useful basis for future theoretical and modeling advances. Analyzing the PDFs of precipitation associated with major contributing weather systems, we find that their PL closely tracks that of the total precipitation. This suggests that while the extreme-range probability decay scale varies by location—and different weather features cluster in different locations—it is not strongly dependent on the system type at each location. This alignment arises as thermodynamic conditions are similar across different weather systems at a given location, intensely precipitating systems tend to dominate the total extremes, and the dynamic decay scale CL varies only moderately across weather types. We interpret the relationship between the weather system–specific and total precipitation PL using three conceptual tail-mixture scenarios: similar-tail mixture, single-tail dominance, and gray-swan-prone mixture. In most situations, using the total PL for risk assessment is sufficient, although we highlight “gray-swan-prone” cases: regions/seasons where a system type with a longer tail can disproportionately affect the most extreme end of the total distribution while occurring infrequently enough that it is difficult to detect from the less extreme portion of the total precipitation tail. Looking ahead, future work could explore how PL and its relations with the thermodynamic and dynamic factors evolve under warming scenarios, and how they might be integrated into practical risk assessment frameworks for precipitation extremes. Developing ways to apply these relationships in climate model outputs and impact assessments could bridge the gap between physical process understanding and actionable climate adaptation strategies, including long-term infrastructure resilience planning, utility operations, and water resource management. MATERIALS AND METHODS Data Precipitation observations are obtained from the IMERG Final Run product, version 06B (54). IMERG provides global precipitation estimates at 0.1° × 0.1° spatial resolution and 30-min intervals by combining measurements from a constellation of passive microwave sensors, calibrated with data from the GPM Dual-Frequency Precipitation Radar and adjusted with gauge analyses. The IMERG data for this study are regridded to 0.25° spatial resolution using bilinear interpolation and resampled to 6-hourly temporal resolution to align with the weather feature identifier datasets (Fronts are identified at 6-hourly intervals). Note that although the data are resampled in time, the precipitation values remain samples of the original instantaneous 30-min observations; no averaging is performed, and thus, high-intensity precipitation rates are preserved. Analyses are conducted for the 19-year period from 2001 to 2019. Atmospheric variables are taken from the ERA5 reanalysis (56), produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). ERA5 provides hourly fields at 0.25° spatial resolution. We use data at the original spatial resolution and 6-hourly temporal resolution to match with the precipitation and weather feature datasets. Variables used include total CWV, mass divergence (convergence) at 850 hPa, and 2-m air temperature and surface pressure. These variables are used to characterize the thermodynamic and dynamic states associated with the precipitation extremes. We also used total precipitation from ERA5 to compare with IMERG. Weather system identification here relies on the Integrated Global Feature-Precipitation Dataset (55), which systematically assembles data identification of major precipitating weather systems tracked by a set of objective detection algorithms (27, 33, 34, 71–73). The weather feature identifiers include ARs (34), Fronts (71), MCSs (72), and LPSs (33, 73), and the various combinations of co-occurrences between them. In the Feature-Precipitation Dataset (55), precipitation not associated with the four primary weather system types and their co-occurrences is further classified into nonfeature deep convection, nonfeature nondeep convection, nonfeature stratiform clouds, and nonfeature drizzle. In this study, we focus exclusively on precipitation categorized as feature-related, which accounts for the majority of extreme rain (25, 27). ARs, Fronts, and LPSs are identified based on ERA5 data at its native resolution; MCSs are identified based on IMERG data at 0.1° resolution and are regridded to the 0.25° ERA5 grid. Together, this integrated dataset provides feature masks at 0.25° × 0.25° resolution and 6-hourly intervals. We use the feature masks to extract and analyze the probability distribution of precipitation associated with each weather system and their co-occurrences. Precipitation PDFs and PL To calculate the PDF of precipitation, we first aggregate the GPM-IMERG precipitation data, which were preprocessed to be at 6-hourly intervals and 0.25° spatial resolution, into 10° × 10° grid boxes. Note that the precipitation data are observed at 30-min temporal and 0.1° spatial resolutions, with no temporal or spatial averaging applied. We then create the bins for PDF calculation as 99 logarithmically spaced bins over the range [0.02, 150] mm hour−1, and added a first bin at 0.001 mm hour−1 to host samples with precipitation less than 0.02 mm hour−1. Next, at each grid box and for each season (December–February, March–May, June–August, and September–November), we calculate the frequency and density with this set of bins. For each 10° × 10° grid box and one season, the total raw sample count is on the order of 107. The large sample size provides a robust basis for the extreme range of the precipitation PDF. The algorithm for calculating the probability density is equivalent to normalizing the frequency in each bin by the bin width and the total count of precipitating samples, producing a unit-integral PDF for each location and season. In Fig. 3 (A to D), we show the normalized frequency distribution for the weather features. It is calculated by normalizing the weather feature precipitation frequency in each bin by the bin width and the total count of the total precipitation. Since we use the same set of bins for both feature-specific and total precipitation, this calculation makes the curve for total precipitation equivalent to its density function and those for features reflect their contribution to the total in each bin. In Fig. 3 (E to H), we retain weather systems that contribute more than 10% to the total extreme precipitation and rescaled their PDFs by PL of the total precipitation. The rescaling process is equivalent to transforming the precipitation intensity P into the dimensionless variable (53) P∗=P/PL_tot. Under this transformation, the rescaled distribution is H(P∗)=f(P∗⋅PL_tot), which preserves the relative shape of the distribution but does not integrate to unity. A properly normalized density would be h(P∗)=PL_totf(P∗⋅PL_tot); however, since we focus on the similarity of the tail slope, we do not apply this additional normalization. To highlight tail similarity, we align the rescaled PDFs of the major contributing weather systems to that of the total precipitation at a common reference point, P∗=2 (i.e., P/PL_tot=2). This anchoring point lies well within the high-intensity range where exponential decay dominates. Specifically, for each weather system i, we define an alignment factor Ai=H0(P∗=2)Hi(P∗=2)=f0(P=2PL_tot)fi(P=2PL_tot), where the subscript 0 denotes the total precipitation. Applying this factor yields the aligned curve Hi~(P∗)=AiHi(P∗). Because the PDFs are plotted on a logarithmic vertical axis, this alignment is equivalent to a vertical shift logHi~(P∗)=logAi+log[Hi(P∗)]. This procedure allows us to directly assess whether the high-intensity tails of the weather system–specific PDFs collapse onto that of the total precipitation when scaled by PL_tot. While we anchor the curves at P=2PL_tot, any reference point sufficiently distanced from the low-intensity range where the power-law behavior dominates would yield the same qualitative conclusion. We classify a weather system as a major contributor when its contribution to the total extreme precipitation exceeds 10% and retain these systems for the analysis of PDF tail similarity. The contribution is defined as the ratio of occurrences summed over the high-intensity range (P>PL) for a given weather system to those of the total precipitation. Our results are not sensitive to the exact choice of this threshold. As shown in fig. S3, the similarity of PDF tails persists across a wide range of contribution thresholds. In particular, the correlation between the weather system–specific and the total PL, which exceeds 0.8 for most weather types, does not decrease as the threshold is increased. We therefore adopt the 10% threshold as a balance between retaining a sufficient number of samples and focusing on systems that materially contribute to extreme precipitation. To estimate PL, we fit the precipitation probability density using Eq. 1: log f(P)=c0−τlogP−1PLP, where log f(P) is treated as the response variable and logP and P are the predictors. The parameters c, τ, and 1PL are determined by minimizing the sum of squared residuals in log f(P). Because the model is linear in the parameters after the transformation in Eq. 1, the estimation is effectively equivalent to multiple linear regression in the transformed variable space. The fitting is conducted only when more than 10 bins contain data. In the spatial maps of PL (Fig. 8 and fig. S1), blank areas indicate locations where PL is not computed due to insufficient data (i.e., fewer than 10 bins with valid values; in some cases, this reflects a complete lack of data—for example, if a particular weather feature does not pass through those locations). In prior studies (51–53) that model precipitation probability densities using f(P)∼cP−τe−P/PL, the two key controlling parameters τ and PL have often been estimated using moment-based methods under the assumption that precipitation follows a gamma distribution. Under this assumption, the parameters can be estimated from mean and variance mom=σ2(P¨)P¨¯, <τ>mom=1−P¨¯2σ2(P¨), where the subscript “mom” denotes a moment estimate, and P¨ represents precipitation rates under raining conditions. However, this moment estimate relies on the gamma-distribution constraint τ<1, which directly affects both parameter estimates. As seen in fig. S11, τ commonly exceeds 1 for subhourly precipitation when estimated by the fitting method, as expected from prior studies (50) for short timescale precipitation intensity. Higher moments would thus be required for an accurate moment-based approach at these timescales [for daily timescales, the moment estimator above is typically fine; (51)]. In applying Eq. 1 to characterize precipitation distributions associated with individual weather systems, we note the following caveats. It is not guaranteed a priori that PDFs of individual feature types follow the same functional form as Eq. 1. When they do, a mixture of two components each following Eq. 1 may produce a total distribution that in general is not exactly the same form. Furthermore, the fitting algorithm we use to estimate PL tends to slightly underestimate the extremity of the underlying PDFs. This can be seen by comparing the dashed (fit) and solid (observed) curves in Fig. 1 (A and D): At higher P values, the observed PDFs often maintain higher probabilities than implied by the fit. The precipitation intensity bins are defined in log space (to accommodate the low frequency of extreme events), which causes the regression to have substantial weighting in the light-to-moderate range. Nevertheless, the key test of the adequacy of the fit for the intended purpose is whether PDFs collapse in the high-intensity range when rescaled by PL, as demonstrated in Figs. 1 (B and E) and 3 (E to H). To assess bias, we perturbed PL and τ in selected cases: Slightly decreasing τ and increasing PL tend to slightly improve the fit quality. Uncertainties in the PL estimate are further quantified through bootstrapping, as described in the “Bootstrapping” section in Materials and Methods. These imperfections do not affect our main conclusions. From the perspective of the three mixture scenarios, in single-tail dominance, a more precise fit would bring the total PL even closer to that of the dominant system. In similar-tail mixtures, both system-specific and total PL values would be slightly larger, but their alignment would remain. In gray-swan-prone mixtures, the total PL still resembles that of the dominant system, while the gray-swan distribution retains a larger PL. Thus, the overall consistency of system-specific and total tails and the associated risk-assessment implications remain robust. The C distribution, CL, and mass convergence To characterize the dynamic contribution to the precipitation distribution, we define a convergence proxy C=P/CWV, where P is the instantaneous precipitation rate in units of millimeters per hour, CWV is the column water vapor in units of millimeters, and the resulting C is in units of per hour—same as the mass convergence. This proxy can be understood as the governing factor of precipitation apart from the thermodynamic influence captured by CWV. Therefore, in addition to the mass convergence at low levels where water vapor concentrates, it also includes effects from factors such as moisture vertical structure and precipitation efficiency (74–76). This ratio reflects the portion of precipitation not directly explained by the thermodynamic influence of water vapor. In separating thermodynamic and dynamic contributions to extreme precipitation, our use of CWVP>PL¯ and a residual dynamic component represents one of several possible decompositions. A large body of previous work (13–16, 20) has examined thermodynamic and dynamical controls on precipitation extremes, exploring their effectiveness across temporal and spatial scales and using moisture measures defined either as column integrals or at specific vertical levels. Physically, extreme precipitation is closely related to the vertical integral of condensation, modified by precipitation efficiency (13, 20, 21). Within clouds, condensation is commonly expressed in terms of vertical moisture advection, for example, −ωdq∗dpθe∗ (13) or −wdq∗dzθe∗ (20), which explicitly involves both vertical velocity and the moisture lapse rate (vertical gradient of saturation-specific humidity). Vertical velocity represents a dynamical contribution, while the moisture lapse rate contains both a thermodynamic component (linked to Clausius-Clapeyron scaling of q∗) and a stability-related component (e.g., moist-adiabatic lapse rate changes with temperature). The decomposition we adopt here is based on the moisture convergence term (77, 78) 〈−q∇h⋅v〉, with 〈⋅〉 denoting a mass-weighted vertical integral and ∇h⋅v representing horizontal divergence. During heavily precipitating conditions, this term forms the leading-order balance to precipitation (7, 78) (in climate models, grid-cell saturation is not required). This convergence-based formulation is equivalent to vertical-advection-based expressions through mass continuity 〈−ωdqdp〉≈〈qdωdp〉≈〈−q(∇h⋅v)〉∼CWV⋅C, linking vertical motion to low-level horizontal convergence. While a more complete scaling would require layer-by-layer integration of thermodynamic and dynamic quantities, previous studies have shown that precipitation-extreme scaling can often be reasonably captured using boundary-layer saturated specific humidity together with midtropospheric vertical velocity (21, 79). In the main text, we use CWV as a leading indicator of lower-tropospheric moisture; using specific humidity at 850 hPa (fig. S9) retains the similarity of the thermodynamic environment across weather features. The temperature dependence of the thermodynamic component can be reasonably captured with surface air temperature (see Fig. 7, I to L). C is computed for all grid points and time steps where CWV≥15 mm. This threshold filters out conditions where heavy precipitation is physically unlikely due to insufficient moisture, and also mitigates spurious values that may arise from mismatches between the CWV and precipitation fields. The PDF of C is calculated at each 10° × 10° region and season, and with 99 logarithmically spaced bins over the range [0.002, 20] hour−1 and a first bin at 0.001 hour−1 added to host samples less than 0.002 hour−1. We estimate its probability decay scale CL of the upper tail using the same regression method used for PL. To validate that the probability decay in C reflects atmospheric dynamics, we compare it with the actual horizontal mass convergence at 850 hPa (C850) from ERA5. We compute PDFs of C850, after converting ERA5 divergence to convergence, and restrict the analysis to the converging (positive) branch. Figure 6B shows the PDF of C850 under precipitating conditions (P>0.02 mm hour−1), exhibiting an exponential decay in the high-intensity range. The PDF of C850 for all-sky conditions also shows an exponential decay—although with lower density—at high values. To further compare the proxy C and C850, we rescale the PDF of C850,P>0 by its own decay scale and then with CL, and then align them at 2CL. The alignment of the rescaled C850,P>0 PDF tail with C (fig. S5 shows more locations) suggests that both the proxy and the actual dynamic variable exhibit similar exponential decay behavior in their tails—supporting the interpretation of CL as a dynamic scale. Bootstrapping To quantify the uncertainty of the estimated distributional tail parameters (PL and CL), we use bootstrapping. For each grid box and season, we randomly resample the time series of precipitation (or C) with replacement to generate 50 bootstrap sample groups (the raw sample size for each group is on the order of 107). For each sample group, we keep the diurnal cycle and the monthly seasonal cycle, and shuffle across years and days within each month. We then recalculate the PDF and perform the same fitting procedure as described in the “Precipitation PDFs and PL” section in Materials and Methods. We define the fractional bootstrap error as the ratio of the standard deviation to the mean estimate of the parameter across the bootstrap groups. In Figs. 3 and 5, data points with fractional bootstrap error greater than 25% are excluded, and in fig. S1 and Fig. 8, grid boxes with fractional bootstrap error greater than 25% are masked in gray. As expected, fractional bootstrap error is larger for locations where a given type of weather feature does not occur often. Temperature scaling of CWVP>PL¯ To relate CWVP>PL¯ to surface air temperature T2m, we first calculate the near surface saturation specific humidity q2m∗ using T2m and surface pressure (ps) from ERA5. Matching with the precipitation and weather feature mask, we then calculate the average q2m∗ conditioned on P>PL, q2m,P>PL∗¯. We then apply the following scaling formula XWV(T2m,P>PL¯)=0.8⋅q2m,P>PL∗¯⋅Δp/(ρg), to get the scaled thermodynamic factor XWV(T2m,P>PL¯). The coefficient 0.8 represents an approximate relative humidity (RH) under heavily precipitating conditions, and Δp/(ρg) is the effective vertical depth over which water vapor is concentrated. We assume the pressure depth Δp corresponds to a 350 hPa layer, capturing the bulk of lower-tropospheric moisture. We take the density of air ρ=1 kg m−3 and the gravitational acceleration g=9.8 m s−2. Although RH and Δp/(ρg) vary regionally, our aim is to establish a first-order scaling with temperature; small deviations in their values are not expected to affect the core conclusion. However, in high-elevation regions, the assumption of a 350 hPa effective depth can be less fair, and large-scale environmental RH may also be lower than 0.8, even during rain. We therefore exclude locations where the orographic height exceeds the geopotential height at 850 hPa. Figure S10 shows spatial maps of the scaling performance for both total and feature-based precipitation.