The relation between global palm distribution and climate

Fossil palms provide qualitative evidence of (sub-) tropical conditions and frost-free winters in the geological past, including modern cold climate regions (e.g., boreal, or polar climates). The freeze intolerance of palms varies across different organs and life stages, with seedlings in particular less tolerant of sub-zero temperatures than adult plants, limiting successful establishment of populations while permitting adult palms to survive in cultivation outside their natural ranges. Quantitatively, palms indicate minimum cold month mean temperature (CMMT) at 2–8 °C in palaeoclimate reconstructions. These data have accentuated model-proxy mismatches for high latitudes during Paleogene hyperthermals when palms expanded poleward in both hemispheres. We constructed a manually filtered dataset of >20,000 georeferenced Arecaceae records, by eliminating cultivars. Statistically derived mean annual temperature, mean annual temperature range, and CMMT thresholds for the Arecaceae and lower rank subfamilies and tribes reveal large differences in temperature sensitivity depending on lower taxonomic classification. Cold tolerant tribes such as the Trachycarpeae produce thresholds as low as CMMT ≥ 2.2 °C. However, within the palm family, CMMT < 5 °C is anomalous. Moreover, palm expansion into temperate biomes is likely a post-Palaeogene event. We recognize a CMMT ≥ 5.2 °C threshold for the palm family, unless a lower taxonomic rank can be assigned.


Results
The core distribution of the palm family is in subtropical to tropical climates with MAT of 18-28 °C and MART of 0-10 °C (Fig. 2). Importantly, all palm subfamilies and tribes have predominantly tropical climate distributions ( Fig. 3a; Table 1). However, a small but significant proportion of the Arecaceae extend into temperate climates, focused on key tribes principally in the subfamily Coryphoideae. The lowest significant MAT when analyzing the Arecaceae as a group is 10.4 °C, but the lowest significant MAT when analyzing individual tribes is 6.9 °C in the subf. Arecoideae, tr. Areceae (Table 1). The lowest coldest quarter mean temperature (CQtrMT) of the Arecaceae is 6.2 °C, 1.9 °C for Areceae and 1.2 °C for subf. Coryphoideae, tr. Trachycarpeae (Table 1). Importantly, in Areceae and Trachycarpeae, the CQtrMT increases to 3.4 °C in both tribes, if the level of significance is raised to  Table 2). Blue circles indicate recorded occurrences that were eliminated by manual filtering and red indicate the occurrences used in this study. Map generated using ArcGIS 58 (Table 1). This suggests that CQtrMT of ~3.4 °C is at the extremes of the distribution of these tribes. The maximum MAT at which Arecaceae occur is consistently 30 °C. The lowest maximum MAT is in the subf. Arecoideae tr. Podococceae (MAT ≤ 26.2 °C).
CMMT calculations show that Trachycarpeae and Areceae have significant occurrences when CMMT ≤ 5 °C (Fig. 3a). Sabaleae (Coryphoideae), Phoeniceae (Coryphoideae) and Ceroxyleae (Ceroxyloideae) have significant occurrences at CMMT ≤ 8 °C (Fig. 3a). As noted before, the level of significance assigned to the outliers of the distribution of Trachycarpeae and Areceae plays an important role, as the 2σ of both tribes is CMMT ≈ 2.2 °C (Fig. 3a). This CMMTmin value is very similar to the threshold of Trachycarpus fortunei found by Walther et al. 12 .
When considering the 11 genera with occurrences at CMMT ≤ 8 °C separately, the three genera with the coolest range are all in tr. Trachycarpeae (Fig. 3b); Washingtonia and Rhapis have significant occurrences at CMMT ≤ 5 °C, whereas Trachycarpus has a limit of CMMT ≈ 0.2 °C. Rhopalostylis and Linospadix, the palms belonging to the tr. Areceae with the coldest range, have a limit of CMMT ≈ 4.2 °C and 6.0 °C, respectively. The offset here with the range of the tribe, may be explained by a larger range of MAT and MART in the tribe, providing a more encompassing σ. A small sample size of Jubaea included in subf. Arecoideae tr. Cocoseae was too climatically disparate to be representative (Fig. 3a). However, when considering the range of Jubaea separately, this genus extends to CMMT ≈ 6.9 °C (Fig. 3b).

Discussion
The Palm Cold Threshold. The distribution of the Arecaceae is predominantly tropical 22 . The highest diversity and abundance of palms is found in tropical regions where CMMT ≥ 18 °C and MART ≤ 10 °C (Fig. 2). Additionally, all palm tribes have their centres of distribution in the tropics (Supplementary Figures). This is generally well in line with where the main centres of vascular plant diversity in the world are located 26 . Early palm lineage fossils from the Cretaceous were most likely growing in tropical conditions 27 and palms are interpreted to have diversified and spread out from tropical environments during the late Paleogene and Neogene 20,22 . Despite the preponderance of palms in the tropics, Arecaceae can still occur in fully temperate conditions at MAT ≈ 10 °C, with Areceae and Trachycarpeae even found at MAT = 7-8 °C (Table 1). Several palm tribes are exclusively tropical, with a CMMT > 18 °C (Fig. 2). Notably, the monotypic Nypa (Nypoideae) just exceeds this boundary, occurring at CMMT ≈ 17.0 °C. Nevertheless, Nypoideae has the highest CMMT threshold of all palm subfamilies. A mere six out of 25 tribes (not including Eugeissoneae, Chuniophoeniceae and Pegalodoxeae) are exclusively tropical (Fig. 3a). Our results suggest that palm cold tolerances are highly tribe-specific.
Greenwood and Wing 11 suggested that CMMT ≈ 5 °C is the lower threshold of palm distribution. This agrees well with our statistically derived threshold of the Arecaceae of CMMT ≈ 5.2 °C, when excluding the outliers of the whole family distribution. However, when considering the range of the palm tribes separately, it is evident that some of these outliers of the Arecaceae distribution are within the significant range of the Areceae and the Trachycarpeae. Whereas our data includes records of Areceae down to CMMT ≈ 0.8 °C and Trachycarpeae down to CMMT ≈ 0.1 °C, these occurrences are at the extreme end of the probability distribution (<0.01; ≫-2σ) and we would argue not a true reflection of the limit of viable reproducing palm populations (Fig. 3a). The genus with the coolest range in the Areceae tribe, Rhopalostylis, has a significant range down to CMMT ≈ 4.2 °C, which is above the 2σ of the tribe Areceae range (CMMT = 2.3 °C). It is likely that the large variation within the Areceae distribution, in both MAT and MART, causes σ to be large and inclusive of occurrences that are statistical outliers with the more conservative σ of Rhopalostylis. We should therefore assume that the CMMT threshold of Areceae is at 2.3-4.2 °C.
The lower threshold of Trachycarpus (CMMT ≈ 0.2 °C) agrees with that of the tribe Trachycarpeae (CMMT ≈ 0.1 °C). This is well below the CMMT = 2.2 °C threshold for Trachycarpus found by Walther et al. 12 . Problematic in this case is the coordinate precision of the extreme range of Trachycarpus, which at its extreme northern limits occurs in incised valleys with a strong microclimate. Even small coordinate imprecisions could place Trachycarpus in a highland, rather than a valley. Moreover, it is questionable if the global climate models we used here to simulate climatic range are precise enough to capture microclimates. The CMMT = 0.1 °C threshold of Trachycarpeae is therefore questionable and we recommend the 2σ value of CMMT ≈ 2.2 °C, in agreement with Walther et al. 12 . The -3.2 °C CMMT threshold for Trachycarpus fortunei found by Fang et al. 23 in China indicates that this species can be successful in cultivation at CMMT < 2.2 °C, but available data shows that the species does not naturalize at these temperatures 6,7,12 . Trachycarpus vegetative tissue freezes at relatively low temperatures (>-14.0 °C) and can recover even after frost damage, but ground frost of <-8 °C destroys the roots of Trachycarpus fortunei, with seeds killed at -2.5 °C 7 . We can therefore assert that sustained frost is fatal to Trachycarpeae and that the diurnal temperature range should be large enough to go above ground frost temperatures. Finally, Rhapis and Washingtonia (Trachycarpeae) both have a lower CMMT threshold of 2.5 °C, suggesting that cold-tolerance in Trachycarpeae is not confined to Trachycarpus (Fig. 3b) and supporting the lower CMMT threshold in Trachycarpeae of ~2.2 °C.
Cocoseae is shown to be exclusively subtropical to tropical, with a lower CMMT threshold of 12.5 °C and lower MAT threshold of 16.0 °C (Fig. 3b, Table 1). However, this eliminates the range of Jubaea, which can occur down to CMMT ≈ 6.9 °C (Fig. 3b). Though the monospecific genus Jubaea is considered threatened within its native range 28   should not be excluded from the lower CMMT threshold estimate of the Cocoseae. The lower CMMT threshold of Ceroxyleae (7.7 °C) is offset somewhat from Ceroxylon, which still occurs at CMMT < 5 °C (Fig. 3a,b). This is probably due to preponderance of Ceroxyleae in the tropics, especially Ravenea, which is confined to Madagascar and the Comoros islands. Ceroxylon is present at high altitudes (>3000 m) in the tropical Andes, explaining its low MAT and CMMT, as well as the relatively restricted MART range of Ceroxyleae (≤7.9 °C). Sabaleae, comprised only of Sabal, extends into temperate latitudes (principally S. minor) and accordingly has a relatively low CMMT threshold (5.3 °C). Despite a relatively high CMMT threshold in comparison to Trachycarpus, Sabal minor is one of the most frost-hardy palms, only suffering widespread frost damage at <-13.5 °C 7 . The range of Phoenix (Phoeniceae) is somewhat problematic, predominantly because of the widespread cultivation of Phoenix dactylifera, but also P. carariensis, P. reclinata, P. roebelenii, P. rupicola and P. sylvestris (Supplementary Table 1). When considering the native range of Phoenix, the lower CMMT threshold is at 7.4 °C. However, geospatial data in the native range of the most transplanted species, P. dactylifera (date palm), is poor and the second most transplanted species, P. canariensis (Canary Island date palm) is endemic to the Canary Islands, but widely cultivated around the Mediterranean. It is therefore unclear if our lower threshold may be an underestimate. Phoenix has similar frost tolerance to Washingtonia 7 .
The influence of seasonality. In general, our data show that the MAT range of a palm tribe increases with the MART range (Fig. 4). This is indicative of the latitudinal range of particular tribes, as the MAT decreases and MART increases with increasing latitude. A cluster of palm tribes in Fig. 4 have a relatively high MAT and MART range: Trachycarpeae, Phoeniceae, Sabaleae, Areceae, Calameae and Caryoteae. Notably, Arecoideae appear to tend towards a MAT range that exceeds the MART range, whereas Coryphoideae are generally the converse (Fig. 4). Furthermore, the results presented here suggest that in some palm tribes MART may play a more important role in determining distribution than MAT. In 8 tribes, the MAT range strongly exceeds the MART range (Fig. 4). This is notably the case for Areceae, Ceroxyleae, Euterpeae, Geonomeae, Chamaedoreeae and Cocoseae. In the case of Iriarteeae and Phytelephanteae the MART range is close to 0 °C, with a MAT range of ~13 and ~10 °C, respectively. This suggests that these tribes are restricted to ecosystems in the tropics with very little intra-annual variation. Limited seasonality may be more important in determining the distribution of these tribes than the temperature around which this seasonality occurs 29 . Climates with limited MART are usually at tropical latitudes and/or in strongly ocean-moderated climates. These palm tribes are distributed along an altitudinal gradient, probably occurring in lowland and montane tropical rainforest.
There are less extreme cases of palm tribes with a large MART range, such as Phoeniceae, Sabaleae and Caryoteae, without a correspondingly large MAT range (Fig. 4). This is probably due to a combination of a lack of freezing tolerance in palms 7 and that the highest MART ranges occur at temperate latitudes. Phoeniceae and Sabaleae appear to deviate most toward a higher MART range than MAT range. The native range of Phoeniceae Flowers are relatively rare due to their delicate structure 30,39 . During the early Eocene, palms enjoyed a near-cosmopolitan range, including in the Arctic and the Antarctic 14,18,40 . The presence of palms at polar latitudes is puzzling as the physiology of modern palms suggests that winter dormancy is unlikely 8 . Latent heat loss in an absence of insolation during winter would drive temperatures at the poles down, causing sub-freezing temperatures that palms could not withstand. Increased equator to pole heat transfer 41 as well as increased cloud cover 5,42 in combination with high global CO 2 43,44 probably kept polar regions relatively warm in Eocene winters. The Eocene Arctic at 72-85°N had an inferred CMMT ≥ 8 °C 14,40 based on the presence of palms adjusted for high pCO 2 , whereas Antarctica at ~65°S had a CQtrMT ≥ 10 °C 18 based on the presence of a multitude of cold-intolerant species, including palms.
Royer et al. 16 found using chamber experiments, that the freezing sensitivity of palms increases by 1.5-3 °C in high pCO 2 (~800 ppm). Early Eocene hyperthermals, such as the PETM or EECO, have pCO 2 estimates of >800 ppm 43,44 . The CMMT threshold obtained here for Trachycarpeae would put the early Eocene CMMT at 3.7-5.2 °C. However, since diversification and radiation of Trachycarpeae did not occur until the Miocene 45 , the more conservative CMMT of the entire palm family is more appropriate here, putting early Eocene CMMT at 6.7-8.2 °C. This is >30 °C warmer than the Arctic CMMT is today and therefore in support of the enhanced Eocene poleward heat flux and increased cloud cover necessary to keep the poles warm during winter 5,41,42 . However, Royer et al. 16

asserted a 1.5-3 °C increase in freezing sensitivity of plants based on chamber experiments.
Interpreting chamber effects is problematic 46 , and raised CO 2 levels in growth chambers are known to cause high levels of aborted, or malformed, stomata 47 . Aborted stomata are thought to be indicative of a failure of leaf expansion or cell division to adjust the distance of stomatal initials 48 , suggesting that the leaf is non-competitive, or ill adjusted, to the circumstances 49 . It is important to note that natural selection drives adaptation, including to differing levels of atmospheric CO 2 , and that a leaf generation, or a single plant generation, does not represent a natural experiment 50 . Increased freezing sensitivity in high CO 2 chamber experiments may therefore be indicative of a plant poorly adapted to the high levels of CO 2 . Further experiments would be needed to show if plants grown in a high-CO 2 chamber are experiencing the deleterious effects of frost and not high-pCO 2 . For now, we assume a minimum CMMT of 5.2 °C, derived from the modern distribution of the Arecaceae family, corresponding to within 1σ of terrestrial climate from the early Eocene based on leaf physiognomy 51 . This would imply that early Eocene winter temperatures in the Arctic were 38.6 ± 1.1 °C warmer than today (Fig. 5a).
Nearest living relative analyses in the fossil record rely on the assumptions that the modern-day distribution of a taxon is representative of its full climatic potential and that this range is dependent on evolutionary conservative traits 52 . The confidence of this assumption becomes less robust with increasing fossil age, as phylogenetic relationships become more disparate from modern. However, the Arecaceae are highly speciose with some palm tribes diverse in genera and geographically widespread 25 and the phylogeny and origination/diversification ages well-studied 20,22,25,45 , thus fulfilling multiple criterion for strong nearest living relative climate proxies 53 . Caution is required in using individual palm genera as palaeoclimatic indicators; e.g., the monospecific genus Jubaea -if Jubaea were extinct, we may assume that Cocoseae did not occur at CMMT < 12.6 °C (Fig. 3a). However, Jubaea CMMTmin = 6.9 °C (Fig. 3b). In this study, we attempt to compensate for these problems by including a large modern representative dataset and assuming an inclusive minimum threshold range for minimum estimates. Using palm tribe-specific bioclimatic threshold values we can estimate temperature minimums for the palm fossils that have been successfully assigned to a tribe. For example, Carpenter et al. 38 identified Nypa from early Eocene western Tasmania (palaeolatitude: 56-57°S) and suggested that this was indicative of a tropical mangrove environment, which corresponds to the Nypoideae threshold at CMMT = 17.0-19.4 °C or CQtrMT = 17.6-19.9 °C (Table 1, Fig. 3a). Compared to the modern-day climate at 56-57°S, this would mean that CQtrMT during the early Eocene at this latitude was 15.1 ± 0.8 °C warmer (Fig. 5b). Modern day regions with the same CQtrMT are found predominantly at 20°S, or at 25°S in coastal situations. This would put early Eocene Tasmania just within the range of a subtropical climate (CMMT ≤ 18.0 °C), but very likely tropical (CMMT ≥ 18.0 °C). Thomas and De Fransceschi 34 identified Cryosophileae wood remains in the late Miocene of southern France (palaeolatitude: 41-42°N). The Cryosophileae CQtrMT threshold would suggest that late Miocene CQtrMT = 16.1-16.9 °C, 9.9 ± 3.6 °C warmer than CQtrMT of that region at the same latitude in modern times (Fig. 5c).
Palm fossils are relatively abundant but assigning them to a tribe or genus can be problematic 30,32,33 . Further, the identification of palms from fossil pollen is potentially questionable, with some fossil pollen unambiguously identified as 'palm' , with other palm-like pollen variously assigned to Areaceae or other unrelated plants 14,15,18,30,31,40 . However, the three examples above show the potential of using the presence of palm fossils as threshold indicators. In some cases, where adequate taxonomic information is available, a tribe-or subfamily-specific threshold value can be used. However, in assemblages with non-tribe or -subfamily specific taxonomic assignments a threshold value for Arecaceae can still be used. There is potential with several palm fossil organs for tribe-specific taxonomic assignments. Palm leaf fossils should be identified using cuticle if present, as the epidermis often provides genus-diagnostic characters [54][55][56] ; however, great strides have been made in the taxonomic classification of palm wood 34 and palm pollen 31 .

Methods
To obtain a comprehensive dataset of palms worldwide, we extracted a databank of 729,696 georeferenced Arecaceae records from gbif.org on February 27, 2017, using the most current palm classification 25 . This dataset represents an amalgamation of records from several publishing authorities (Supplementary Table 2) and is unbiased in its recordkeeping. However, we manually filtered the dataset according to several criteria: • Entries with no species data were removed. • Entries with poor or non-specific geospatial data were removed. Poor geospatial data can be identified if the geodetic coordinates are strongly rounded, assumed, or depict a location that is not on dry land. • Redundant entries were removed. Particularly when geodetic data is back-entered, a single location may be chosen to represent a sampled population of multiple individuals. This can lead to several hundreds of individual records with identical geodetic coordinates. The rationale for only having a single record representing a location is that when an already small dataset is composed mostly of multiple records from a single location, the climatic probability function becomes highly slanted towards the climate of that location.
The dataset was then manually filtered for cultivated individuals. This was done by: • Identifying palm species that are (1) popular as ornamentals, (2) cultivated crops and/or (3) an invasive species. Within the geodetic data, distinguishing ornamentals, cultivars and invasive species from each other is often problematic. Therefore, no distinction was made between them, despite the possibility of an invasive or naturalized species occurring within its preferred environment. • Identifying the countries that are included in the natural range of the species. The output file of gbif.org includes information of the country in which the plant was recorded. This allows us to exclude occurrences in countries that are not included in the natural range of the species.
We identified 56 palm species as having a problematic range (Supplementary Table 1). Of the original dataset of 729,696 georeferenced Arecaceae records, this rigorous filtering process resulted in 21,323 occurrences that were used (Fig. 1). For the Chilean wine palm (Jubaea chilensis) there were only two recorded occurrences in its native range. However, because of its importance for climatic range estimates as the most temperate species of the tribe Cocoseae, we supplemented the database with 11 recorded populations from Gonzalez et al. 28 . All geodetic coordinates of these occurrences were queried for MAT, summer and winter mean temperature (WQtrMT and CQtrMT) using the gridded climate model of Hijmans et al. 57  Manual filtering eliminates a part of the potential for climatic outliers (Fig. 1). However, the method that we use here still allows for bias. Primarily because the dataset is strongly dependent on organizations, such as government agencies (herbaria, universities, conservation bodies), that make their records available and the degree of data availability can differ strongly between countries. As an example, New Zealand has two native palm species that we consider here, with 2713 records, whereas Papua New Guinea has 129 species of palm that we consider here, with 556 records. This may result in a bias towards the temperate New Zealand species, even though these are much less diverse and not necessarily more abundant. Secondly, we eliminated outliers manually based on the occurrence latitudinal range of a palm species and which countries the species is native in. Therefore, the possibility exists that occurrences are included that are outside the native range of the plant, within a country that it is native in. For example, the palm could be planted on a mountain, whereas its native range is at sea level in the same country. Finally, manual filtering cannot account for the reliability of the geodetic information; data can be incorrectly entered into the database or not recorded accurately at the time of collection (e.g., for older records made prior to the widespread use of handheld GPS systems). We therefore applied a probability density analysis to calculate the likelihood of each individual record being representative. Probability density analysis assumes a normal distribution in the climatic range of each group considered. Assuming a normal distribution is advantageous in this type of analysis because it reduces the risk of bias by poor geospatial coverage. It also allows for querying individual occurrences for their significance in determining the climatic range of a group.
We performed probability density analysis on three taxonomic ranks: family level, subfamily level and tribe level. This provided a significant sample size in most cases. The Eugeissoneae, Chuniophoeniceae and Pegalodoxeae tribes had three, eight and ten recorded occurrences, respectively, that were left after manual filtering. Therefore, these tribes were not further considered. To determine the probability of a data point (x) being representative of the range we combined probability density of both MAT and MART. Because this calculation would yield different probability densities within each group, the probability for each data point was compared to the maximum probability likelihood within a group to determine the relative probability of this data point in constituting the climatic range of the group. We postulate that a data point with a relative probability of f x(relative) < 0.01 is insignificant in the climate range of the group. We determined the range of one (f x(relative) = 0.157) and two standard deviations (f x(relative) = 0.023) from the occurrence within a group with f max to constrain the core distribution of a group. This method is preferable over assigning percentiles as cutoffs for significant climatic range, as we determine outliers based on the relative deviation from the mean climatic range of the group. This also allows us to determine if a group of data does not have any outliers, or may be prone to being influenced by outliers. In addition, this method can consider the significance of a datapoint along multiple environmental gradients. When referring in the results and discussion to tropical, subtropical or temperate climates, we follow the climatic classification of Belda et al. 59 . Climates are considered tropical if CMMT > 18 °C. Climates are subtropical if ≥ 8 months have average temperatures of >10 °C and a CMMT of <18 °C. Since we do not consider monthly temperatures, we here refer to a climate as subtropical if MAT is >16 °C and CMMT is <18 °C. In Belda et al. 59 , climates with 4-7 months of temperatures >10 °C are considered temperate. Here, we refer to temperate climates as MAT < 16 °C.
Data availability. All data used in this study is made available in the Supplementary Information.