Online Supplementary Materials

Evidence that accent is a dimension of social categorization, not a byproduct of perceptual salience, familiarity, or ease-of-processing. 1) Then she said " Hey, it's good to meet you! I heard you'd be here and we were looking forward to meeting up with you. " They had just finished up a long stay there. As I remember, they were keeping pretty busy while they were there. 2) I realized it a few months ago. The Smiths were having a lot of trouble with their car. They brought groceries, some chairs and tables, and helped with our stuff. The last few weeks, though, they had to spend some of their time other places. They were staying somewhere pretty far away. Somewhere I think that's recently been hit with a ton of snow. 3) They started showing up a few months ago. There was a lot of commotion then. But they've totally stopping showing up now. They got carried away with the whole thing. We figured it wouldn't last. Some of us were happy they left, but a couple people seemed really upset. 4) Well, I saw what was left over from it. It was crazy. They had to move thousands of pounds of rubble and debris before doing anything. I was young but I remember that it sucked. It really doesn't help to have your neighborhood trashed when you just moved. 5) I know the kitchens had to be replaced, too. Cracked walls, sagging roof, everything. They had to do some massive wall-to-wall renovation. They tried to hire every able-bodied person in the area qualified enough to work there, even some people who they ended up firing. Everyone was really friendly, though. It's cool to see it now it's up and running smoothly. It's a rocking place. 6) When they lost the medication, he was pretty upset. Said something about checking priorities and keeping track of things a lot more carefully. They wanted to redo the entire system, so that kind of thing wouldn't happen again. They bought some new computers for inventory tracking. He hasn't had any problems since. They seemed to care and actually wanted to fix problems. Things do seem to be better now. You can imagine, he's totally relieved. 7) I don't know how you pick what you're going to do. You know plans don't always fit with what happens. But, more or less, it's all worked …


Introduction
Hydroxyl (OH) and peroxy (HO 2 , RO 2 ) radicals play an essential role in atmospheric chemistry on local to global scales (e.g., Brasseur et al., 2003;Monks et al., 2009). Reactions with the most important atmospheric oxidant, OH, initiate the chemical breakdown of tropospheric trace gases such as CO, SO 2 , NO 2 , CH 4 and other volatile organic compounds (VOCs). Many of these reactions produce HO 2 and RO 2 , which are key intermediates in the formation of secondary, atmospheric pollutants (Finlayson-Pitts and Pitts Jr., 2000). Reactions Oxy radicals, RO, produced by Reaction (R5), are generally converted to oxygenated VOCs (OVOCs), either by fast reaction with O 2 yielding carbonyl compounds, R O (Reaction R6), or by unimolecular decomposition and isomerization reactions (Atkinson, 1997). OVOCs are currently receiving growing attention in atmospheric chemistry. Firstly, they may contribute significantly, sometimes 50% and more to the organic reac-5 tivity in various tropospheric environments, thereby affecting the atmospheric lifetime of OH (Lou et al., 2010;Mao et al., 2010;Shao et al., 2009;Steiner et al., 2008;Emmerson et al., 2007;Yoshino et al., 2006;Lewis et al., 2005). Secondly, photolysis of OVOCs can be a significant source of radicals at urban conditions (Emmerson et al., 2007;Kanaya et al., 2007b;Volkamer et al., 2007;Dusanter et al., 2009). Thirdly, 10 OVOCs are precursors of secondary organic aerosols (SOA) which play a distinct role for air quality and climate (e.g., Hallquist et al., 2009;Kanakidou et al., 2005). The atmospheric abundance of OH and HO 2 , collectively called HO X , depends critically on primary production processes with contributions by photolysis of ozone, photolysis of nitrous acid and OVOCs, and O 3 -alkene reactions (Fig. 1). Ensuing chain reactions (Reaction R1-R7, indicated by red arrows in Fig. 1) oxidize CO, VOCs and NO, but also interconvert RO X radicals (= OH + HO 2 + RO 2 ). Radical recycling by Reactions (R5)-(R7) is of particular relevance, as it constitutes a secondary source of OH which significantly enhances the oxidation efficiency of the troposphere (e.g., Ehhalt, 1999). Radical recombination reactions (e.g., Reactions R10, R11, R14, R15), however, cause chain termination and suppress the concentrations of OH, HO 2 and RO 2 11315 Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | radicals.
OH + NO 2 + M → HNO 3 + M (R14) Highly sensitive instruments for in-situ measurement of HO X became available in the 1990s (Crosley, 1995) and have been employed in an increasing number of field experiments to test atmospheric chemical mechanisms over a broad range of tropospheric conditions. Early studies were carried out mostly in clean (polar, marine) and moderately polluted (rural) regions (Heard and Pilling, 2003), whereas many recent 5 investigations were focussed on environments with high loads of reactive trace gases (Monks et al., 2009). These include polluted cities and their urban-influenced surroundings, as well as forests with high concentrations of biogenic VOCs. Most urban studies have shown reasonable agreement between modelled and measured OH within a factor of two when VOCs and NO X were abundant (e.g., Kanaya et al., 2007b;Shirley et al., 2006;Mihelcic et al., 2003). However, current models tend to underpredict the observed OH by up to an order of magnitude at low NO X , when isoprene reaches mixing ratios of several ppb at the same time. This trend has been clearly identified in forested regions, such as in North America (Tan et al., 2001;Ren et al., 2008), over the Amazonian rainforest (Lelieveld et al., 2008) and the tropical forest of Borneo (Pugh 15 et al., 2010;Whalley et al., 2011). The observations suggest that as-yet-unknown recyling reactions of peroxy radicals constitute a significant OH source, when NO mixing ratios become small. This hypothesis is consistent with a budget analysis of HO X in isoprene-containing urban air near Nashville by Thornton et al. (2002), who supposed that self reactions of peroxy radicals possibly recycle OH, rather than acting as an HO X 20 sink (Reaction R11).
The above findings from urban and forest studies are in agreement with our investigation in Southern China (Hofzumahaus et al., 2009). In summer 2006, we performed a field campaign in the densely populated Pearl River Delta (PRD), close to the megacity of Guangzhou in Guandong province, where atmospheric conditions were strongly data of OH and HO 2 at PRD, Hofzumahaus et al. (2009) implemented two hypothetical reactions (RO 2 + X → HO 2 and HO 2 + X → OH, both of the same rate as for the corresponding NO reactions) into a box model. An NO equivalent of 0.8 ppb was assumed for X to match the mean diurnal profiles of both HO X species.
Recent model studies have attempted to identify the actual reaction mechanisms 15 underlying the non-classical (without NO) OH recycling in forested regions. OH regenerating reactions of HO 2 with acyl peroxy and β-keto peroxy radicals, known from various laboratory studies (Jenkin et al., , 2007Dillon and Crowley, 2008;Hasson et al., 2004), were found to increase modelled OH by at most 7% above tropical rainforests (Peeters and Müller, 2010;Stavrakou et al., 2010;Archibald et al., 2010;20 Pugh et al., 2010). In addition, the OH-neutral oxidation of isoprene hydroperoxides to dihydroxy epoxides, experimentally studied by Paulot et al. (2009), has the potential to further increase the modelled OH concentration by up to 25% in the tropics, depending on the specific isoprene degradation mechanism being used (Stavrakou et al., 2010;Archibald et al., 2010). However, even together the above processes are not sufficient 25 to explain the factor-of-10 discrepancy between modelled and observed OH above the Amazonian rainforest. Based on ab initio calculations, Peeters et al. (2009) and Peeters and Müller (2010) have postulated a new isoprene degradation mechanism (Leuven Isoprene 11318 to a net formation of one HO 2 and up to three OH radicals. The implementation of the LIM0 mechanism increases the modelled OH concentrations for the conditions over the Amazonian rainforest by up to a factor of four, showing potential to explain the nonclassical OH recycling in regions with high isoprene emissions (Stavrakou et al., 2010;Archibald et al., 2010). 10 In the present work, we investigate whether the newly proposed recycling mechanisms can explain the high OH levels at PRD, where NO concentrations span a broad range (0.02-10 ppb) and where anthropogenic VOCs are present besides isoprene. First, we report experimental details of the measurements of HO X during the PRIDE-PRD2006 campaign and demonstrate the stabilizing effect of the unknown OH source 15 on the observed OH concentration as a function of the ozone photolysis frequency, j (O 1 D), and NO X . We test different recycling mechanisms to investigate their impact on atmospheric OH and HO 2 for the conditions at PRD. The analysis is complicated by a recently discovered interference in our HO 2 measurements, which are significantly biased by the detection of specific organic peroxy radicals, such as isoprene 20 hydroxyperoxy radicals (Nehr et al., 2011;Fuchs et al., 2011). The results reported by Hofzumahaus et al. (2009) are revised, accordingly.

Measurement site
The measurements presented in this paper took place alongside a drinking water 25 reservoir in a recreation area called Guangzhou Backgarden about 60 km northwest 11319 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | of downtown Guangzhou on 3-30 July 2006. Backgarden is located in a slightly mountainous area at 23.487 • N, 113.034 • E and is surrounded by farmland, which is mainly covered by tropical shrubs and economic crops like lichees and peanuts. During the intensive campaign period, due to Asian monsoon, the dominant wind direction was south to south east, making Backgarden a receptor site for the outflow of urban emis-5 sions from Guangzhou. However, local wind speeds were generally low (often less than 2 m s −1 ), which is typical for inland PRD during summer season and favors accumulation of air pollutants (Chan and Yao, 2008). During the campaign, the weather was characterized by high humidity with absolute water-vapor mixing ratios of 2.5-4% and high temperatures of about 28-36 • C. 10 Besides HO X (see next section), a comprehensive set of trace gases was measured during the campaign. Table 1 gives an overview of the measured species which are relevant for this paper and specifies the measurement techniques and their performance. Further instrumental details can be found in Lou et al. (2010). The C 3 -C 12 VOCs, measured by online gas chromatography (GC), are specified in Table 2. They include alkanes, alkenes, aromatics, and isoprene. No in-situ OVOCs measurements were available; but on some days, the averaged HCHO and glyoxal concentrations for certain air masses spanning several kilometers were retrieved by MAX-DOAS (Li et al., 2010). Most trace gases were sampled at 10 m above ground on top of a building. HO X , k OH , HONO and meteorological data were measured nearby (30 m distance) at 7 m 20 height on top of two stacked sea containers. On the containers, the photolysis frequencies j (O 1 D) for Reaction (R12) and j (NO 2 ) for Reaction (R8) were measured by calibrated filterradiometers. On the nearby building, j (O 1 D), j (NO 2 ), j (HONO), j (HCHO) etc. were obtained from solar UV spectra measured by actinic-flux spectroradiometry (Bohn et al., 2008). Both sets of j (O 1 D) and j (NO 2 ) data were in agreement within 5%.

Radical measurements
OH, HO 2 and k OH were measured by a compact laser-induced fluorescence (LIF) system built at Forschungszentrum Jülich, Germany (Fig. 2). The technique was initially developed for measurement of OH (Holland et al., 1995(Holland et al., , 1998Hofzumahaus et al., 1996) and was later extended by an additional measurement capability for HO 2 (Hol-5 land et al., 2003). The LIF instrument used at PRD is a follow-up version of the earlier system. It is designed to be smaller and more light-weight than the previous instrument, making the instrument easier to handle in field applications (e.g., Kleffmann et al., 2005;Schlosser et al., 2009). Furthermore, the newly developed LP-LIF (laser-flash photolysis laser-induced fluorescence) technique for measurement of k OH was implemented 10 (Lou et al., 2010) utilizing the same tunable laser source for OH detection. In the following, the LIF system for HO X detection will be described briefly, explaining the general principle and technical differences between this and the earlier instrument version. Radicals are sampled by expansion of ambient air through an inlet nozzle (0.4 mm orifice, Beam-Dynamics) into a low pressure (3.5 mbar) chamber, where OH is detected 15 by LIF at 308 nm. HO 2 is monitored in a separate parallel detection chamber, in which HO 2 is first chemically converted to OH by reaction with injected NO, followed by LIF detection of OH. It should be noted that the two detection cells are mechanically connected by a set of laser-baffle arms, but are separated by a quartz window preventing any possible contamination of the OH cell by NO from the HO 2 cell. The two de-vibronic band at 308 nm. OH resonance fluorescence emitted between 307 nm and 311 nm is collected by an assembly of large-diameter fused-silica lenses and narrowband optical filters and is detected with a highly sensitive channel photomultiplier tube which is mounted perpendicular to the gas beam and the laser axis. The fluorescence is measured by gated photon counting using a time delay (≈ 50 ns) to discriminate 5 the longer-lived OH fluorescence (150 ns lifetime at 3.5 hPa) from the instantaneous laser stray light (≈ 20 ns duration). The signals are further corrected for solar straylight, which enters the measurement cells through the nozzle orifice and is measured after each laser-pulse in a separate time gate with a delay of 25 µs. Furthermore, the laser is tuned periodically on-and off-resonance to distinguish the OH fluorescence signal 10 from non-resonant laser excited background signals (Hofzumahaus et al., 1996). The amount of detected OH fluorescence integrated over successive laser pulses can be converted into an ambient radical concentration, of which the required sensitivity is determined by calibration (see below).
The major technical differences between the present and previous instrument version 15 is the integration of more compact, light-weight components for the laser system, the vacuum pumps and the gated photon-counting system and a more convenient distribution of the 308 nm laser radiation by optical fibres rather than by beam-steering mirrors.
The laser system used in this work consists of an intracavity frequency-doubled tunable dye-laser (Tintura, New Laser Generation) which is pumped by a frequency-doubled 20 Nd-YAG laser (Navigator-I, Spectra Physics). The dye laser uses an intracavity etalon for line narrowing and provides a stable laser bandwidth (7 GHz at 308 nm). This stability is an advantage compared to the previously used laser system where the bandwidth was very sensitive to laser alignment with corresponding need for frequent recalibration of the HO X measurement. The present laser system provides up to 100 mW of total UV power was reduced to 60 mW in the beginning, decreasing further to 10 mW towards the end of the campaign. The laser radiation was delivered by a 6 m long optical multimode fibre (QMMJ-55HP-UVVIS-200/240 µ, AMS Technologies) with an effective transmission of about 50-70% to the detection cells. In-and outcoupling of the laser beam was achieved by AR-coated plano-convex quartz lenses (f = 25 mm 5 and 50 mm, respectively). For fluorescence detection, gated channel-photomultipliers (C1943P, Perkin Elmer) were connected to a gated photon-counter card (PMS300, Becker und Hickl GmbH) and gating signals were provided by a digital delay generator (DDG, Becker und Hickl GmbH) triggered by the laser system. The dry-vaccum pump (IPX500, BOC-Edwards) is connected to each detection chamber by separate flexible 10 metal-bellow tubes (40 mm diameter) including motorized butterfly valves (MKS153), in order to stabilize the pressure in the cells to better than ± 0.1 mbar. During the PRD campaign, the laser was tuned periodically to on-and off-resonance wavelength positions with integration times of 40 s and 8 s, respectively. Including some overhead time for laser scanning, a typical time resolution of 40-70 s was achieved for 15 the HO X measurements. The instrument was calibrated with known amounts of OH and HO 2 radicals which were generated in a flow of synthetic air by photolysis of water vapor at 185 nm from a low-pressure discharge mercury lamp (for details, see Holland et al., 2003;Fuchs et al., 2011). The calibration source was recently tested in intercomparisons against absolute measurement techniques for OH (Schlosser et al., 20 2006(Schlosser et al., 20 , 2009, confirming the estimated accuracy (±10%, 1σ) of the calibration method. At PRD, successive calibration measurements showed an unusally large 1σ variability by 8.7% and 13.6% for OH and HO 2 , respectively, from day to day. This variability is about a factor of two larger than usual and suggests uncontrolled changes of the instrumental detection sensitivity. Since no trend was observed in the calibration 25 data, an average calibration factor was applied for the campaign. The observed variability adds to the calibration uncertainty which is estimated to be 20% (1σ) in total. OH and HO 2 interferences caused by ambient O 3 were corrected by an amount of (6 ± 2) × 10 3 cm −3 and 2 × 10 4 cm −3 , respectively, per ppb of ozone. The detection limit Introduction of the 5 min averaged OH data is 5 × 10 5 cm −3 before 18 July 2006, afterwards due to a reduced laser power, this number increased to 1 × 10 6 cm −3 ; the detection limit of the 5 min averaged HO 2 data is (1 − 2) × 10 6 cm −3 before 18 July 2006, (2 − 3) × 10 6 cm −3 afterwards.

HO 2 measurement interference by RO 2 5
Recent laboratory studies show that our HO 2 detection system exhibits a significant sensitivity to specific RO 2 species, which are converted to OH by a sequence of NO dependent reactions (Nehr et al., 2011;Fuchs et al., 2011). In general, RO 2 reacts in the gas expansion with the injected NO as fast as HO 2 and forms RO (Reaction R5).
In case of simple alkoxy radicals, RO reacts predominantly with O 2 and produces HO 2 10 (Reaction R6). Because of the short reaction time (few milliseconds) and the strongly reduced O 2 number density in the gas expansion, the RO to HO 2 conversion is slow and the following production of detectable OH is marginal. Experimental tests have shown that the corresponding interference by C 1 -C 4 alkyl peroxy radicals is generally not larger than about 5% (Stevens et al., 1994;Kanaya et al., 2001;Tan et al., 15 2001;Creasey et al., 2002;Holland et al., 2003;Ren et al., 2004), consistent with the new results by Fuchs et al. (2011). However, in case of RO 2 from OH reactions with alkenes and aromatics, the RO radicals formed in Reaction (R5) undergo unimolecular reactions and extremely fast decomposition to HO 2 . In this case, the NO-dependent Reactions (R5) and (R7) control the effective rate of RO 2 → HO 2 → OH conversion.

20
A significant amount of RO 2 is eventually detected as OH in the HO 2 detection cell, resulting in relative detection sensitivities (α RO 2 ) for specific RO 2 compared to HO 2 of larger than 50% (Nehr et al., 2011;Fuchs et al., 2011). Experimental α RO 2 values for peroxy radicals from selected VOCs of the groups of alkanes, alkenes, aromatics and OVOCs are shown in Fig. 3  respectively; but values are larger for peroxy radicals from cyclohexane (48%), simple alkenes (85-95%), isoprene (79%), and benzene (86%). Also the peroxy radicals from the OH reactions with major isoprene degradation products, methyl vinyl ketone (MVK) and methacrolein (MACR), have significant α RO 2 values of about 60%. Model simulations of α RO 2 are also presented in Fig. 3 (blue symbols). They are 5 based on the Master Chemical Mechanism MCM v3.1 Jenkin et al., 2003) and reproduce the trend of the experimental values reasonably well (Fuchs et al., 2011). The simulations are shown for the major VOCs which contributed more than 95% to the VOC reactivity at PRD. The HO 2 concentration measured by LIF, denoted as [HO * 2 ], is then expected to be the sum of the true HO 2 concentration and 10 a systematic bias from the mixture of RO 2 species i which are detected with different relative sensitivities α i We have not attempted to correct the measured HO * 2 values in order to obtain true HO 2 concentrations, since RO 2 concentration measurements and their speciation are 15 not available. However, a chemical box model is used to calculate concentrations for both HO 2 and HO * 2 (see below). The difference between these two concentrations can be considered an estimate of the effective instrumental interference from RO 2 at the conditions found at PRD. 20 A zero-dimensional chemical box model was used to calculate concentrations of OH, HO 2 , HO * 2 and photochemical products of nitrogen and carbon compounds. In this work, we call the applied chemical mechanism RACM-MIM-GK. It is based on the Regional Atmospheric Chemical Mechanism (RACM) (Stockwell et al., 1997) which was upgraded with the isoprene degradation scheme by Karl et al. (2006). The latter Introduction  (Pöschl et al., 2000). The complete mechanism of RACM-MIM-GK, which has been used before by Hofzumahaus et al. (2009) andLou et al. (2010), is reported in Table S1 in the Supplement of this paper. The model calculations were constrained to measurements of O 3 , HONO, NO, NO 2 , 5 CO, CH 4 , C 3 −C 12 VOCs, photolysis frequencies, water vapor, ambient temperature and pressure. OLT, the measured value for propene was used and for the aromatics TOL and XYL, the measured α i RO 2 for benzene was applied. In the following, model runs with the RACM-MIM-GK are used as a reference which represents the framework of established photochemistry with classical, NO-dependent radical recycling. This base-case is denoted as M0. To explore possible reaction mech-5 anisms which recycle OH radicals without NO, additional model runs were performed which incorporate assumed generic reaction pathways (M1-M2) or newly proposed mechanisms (M3-M6) into RACM-MIM-GK. Furthermore, the results from RACM-MIM-GK are compared to model runs based on the detailed MCM v3.1 (M7). An overview of the different model scenarios is given in Table 5 and the corresponding reaction mech-10 anisms are listed in the Supplement of this paper. Details and results of the sensitivity runs are given in Sect. 4.3.

Observations of HO X and other photochemical parameters
During PRIDE-PRD2006, concentrations of OH and HO * 2 were measured from 5 to 15 25 July 2006. An overview of the 5 min-averaged data is shown in Fig. 4. Data gaps were caused by heavy rain during the typhoon BILIS (15-17 July 2006), by electric power failure (22 July 2006), and by instrument calibration or maintenance (i.e. 11 and 18 July 2006). The diurnal variations of the observed radicals followed a regular pattern from day to day, with maximum values around noon. To evaluate the vari-20 ability of the daily peak values, mean values of the upper 0.05 percentiles of OH and HO * 2 were calculated for each day. Daily maximum OH concentrations varied from 15 × 10 6 cm −3 to 26 × 10 6 cm −3 , while daily maximum HO * 2 concentrations varied from 3 × 10 8 to 25 × 10 8 cm −3 . Significant HO X concentrations were measured during nighttime as well. While individual OH data points lie close to the limit of detection, hourly 25 averaged nighttime values (at solar zenith angles larger than 90 • ) were in the range of Introduction Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | 0.2 × 10 6 cm −3 to 5 × 10 6 cm −3 with an error of 0.2 × 10 6 cm −3 . When averaged over the whole campaign, the mean nighttime value was 2 × 10 6 cm −3 . For HO * 2 the hourly averaged nighttime values were in the range of 0.1 × 10 8 cm −3 to 10 × 10 8 cm −3 with an uncertainty of 0.1 × 10 8 cm −3 . Over the whole campaign, the mean nighttime value of HO * 2 was 2 × 10 8 cm −3 . The origin of nighttime HO X will be analyzed in a separate 5 publication.
In order to characterize the encountered air masses from a photochemical point of view, time series of important parameters influencing the radical concentrations are presented in Fig. 5. j (O 1 D) and j (NO 2 ) showed regular diurnal patterns with high noontime values that reached up to 4 × 10 −5 s −1 and 1 × 10 −2 s −1 , respectively. Relatively 10 small daytime values were observed on 15-17 July 2006 when the sky was covered by a dark cloud layer during typhoon BILIS. High concentrations of NO (often more than 10 ppb) were measured in the morning hours, while low concentrations below 1 ppb prevailed in the afternoon. Anthropogenic emissions from heavy-duty cars and combustion activities after midnight might be the cause for the morning peak of NO, 15 whereas a ban of heavy duty cars during daytime, the enlarged boundary layer height and photochemical oxidation of NO x may be the reason for the low NO concentrations in the afternoon. The diurnal characteristic of NO 2 is similar to that of NO, but large concentrations of NO 2 often appear near midnight probably produced by titration of O 3 with freshly emitted NO. Ozone showed an anti-correlated diurnal variation compared 20 to NO 2 . Peak values of O 3 mostly appear around noon or afternoon hours, indicating local photochemical ozone production (Lu et al., 2010). Relatively high concentrations of HONO were observed at the PRD site during early morning hours (about 1 ppb) and noon time (about 200 ppt). Details of the HONO budget will be discussed by Li et al. Since the site is located in a forested region, large isoprene concentrations can appear during daytime. As discussed by Lou et al. (2010), daily peak isoprene concentrations were correlated to the wind direction. For the prevailing southerly wind direction, the air had to travel over a large water reservoir and daily peak isoprene concentration were typically 1-2 ppb. For northerly wind directions (13-14 July and 23-25 July 2006), emissions from nearby plants or agriculture fields could directly influence the site and daily peak isoprene concentrations approached 5-6 ppb then. A biomass burning event on 23-25 July 2006 was identified by analysis of measured optical aeorosol properties (Garland et al., 2008;Rose et al., 2010). Thus, the high isoprene concentration during this event could also have been caused by stress-induced plant emissions.

Base-case model results for HO X
Model calculated results (M0) for OH and HO * 2 are compared to the measured time series in Fig. 4. For OH, the base-case model shows diurnal patterns that are systematically different from the observations. During morning hours, modelled and measured concentrations always agree well, while in the afternoon the model underestimates the 15 observed OH by a factor of 2 to 8. On the other hand, observations of HO * 2 are much better reproduced by the model. This picture using the full time resolution of observed data is consistent with what has been reported previously (Hofzumahaus et al., 2009) using diurnally averaged boundary conditions for modelling (see also Fig. S1 in the Supplement). 20 The model has been used to simulate both ambient HO 2 and measured HO * 2 , which are shown in Fig. 4  The radical budget of HO x and the related RO x is analyzed in Figs. 6 and 7 using chemical turnover rates determined by the base model (M0). Figure 6 shows the mean diurnal profiles of the production and loss rates of RO x . During daytime, the primary OH production, P (OH), and primary HO 2 production, P (HO 2 ), were the dominant part of P (RO X ). O 3 and HONO photolysis reactions constituted the major part of P (OH), 5 while HCHO photolysis (about 3-4 ppb h −1 during noon time) dominated P (HO 2 ). The next important processes were the photolysis of dicarbonyls including gloxyal (GLY), αcarbonyl aldehydes (MGLY) and unsaturated dicarbonyls (DCB). In the early morning, HONO photolysis was the most important primary source of HO X , which contributed 60-70% of P (HO X ) between 07:00 and 08:00 CNST. A recently proposed new primary 10 OH source, the reaction of excited NO 2 with H 2 O (Li et al., 2008), has been examined as well. The photolysis frequency for the production of excited NO 2 was calculated from measured actinic flux spectra, assuming that excited NO 2 is formed beyond the photodissociation threshold (420 nm) up to wavelength of 700 nm with unity quantum yield. The estimated radical production rate from this channel was only about 0.1 ppb h −1 at 15 08:30 CNST (SZA = 60 • ) which is almost negligible at our conditions. As indicated by the overlapping grey dashed and red solid lines in Fig. 6a, the total RO X production is balanced by equally large RO X sinks, L(RO X ), in the model. As shown by Fig. 6b, L(RO X ) is dominated in the morning by OH reactions with odd nitrogen compounds and RO 2 reactions with NO x yielding nitrates and PANs. In the afternoon, self reac-20 tions of HO 2 or RO 2 and cross reactions between HO 2 and RO 2 are dominant. The radical chemistry is VOC limited in the morning and NO X limited in the afternoon with the transition taking place around 10:00-11:00 CNST. In Fig. 7a,b, the rates of OH reactions yielding HO 2 and RO 2 are specified. As discussed for the correspondingly measured k OH (Lou et al., 2010), half of the OH loss 25 can be explained by the measured trace gases given in Table 1 and half by modelled daughter products of isoprene, alkenes and aromatics. OH → HO 2 conversion is mainly caused by reaction of OH with CO and formaldehyde throughout the day. During early morning hours, the total OH → RO 2 conversion rate is dominated by oxidation reactions Introduction of alkenes (propene, butenes, pentenes) and aromatics (styrene, toluene, xylenes, trimethylbenzenes), while it is dominated by isoprene and its degradation products in the afternoon. This is reflected in the RO 2 speciation within different NO X regimes (see Fig. S2 and Table S4, Supplement).
Recycling rates for RO 2 → HO 2 and HO 2 → OH are speciated in Fig. 7c,d, respec-5 tively. RO 2 → HO 2 conversion is dominated by NO reactions with MO2, ISOP and MACP. Here, MO2 is predominantly formed from oxidation of methane, isoprene and alkenes. Figure 7d shows that HO 2 →OH recycling (red line) is mainly caused by the reaction HO 2 + NO (blue line). It is nearly balanced by HO 2 formation through OH → HO 2 plus RO 2 → HO 2 reactions (dark red line), demonstrating that the HO 2 budget is mainly 10 controlled by cycling reactions rather than by primary production and termination reactions. Furthermore, it can be noted that the total peroxy radical formation rate, i.e. the sum of R(OH → HO 2 ) and R(OH → RO 2 ) (violet dashed line) agrees most of the day with the recyling rate HO 2 →OH. A relative large difference is found only in the afternoon, indicating inefficient recycling to OH in the RO X propagation implemented in 15 the base model, consistent with the analysis of L(RO X ) in Fig. 6b.

Discussion
HO X concentrations observed during PRIDE-PRD2006 are the highest so far reported for urban and suburban environments at summer time (Table 6). In Mexico City (located at a similar latitude and also in a developing country) relative small HO X con-20 centrations were observed. Among the other cities, which are located in developed countries like US, Europe and Japan, measurements at Nashville and New York City showed relative high HO X concentrations that approach the levels of PRD. One reason for the high radical concentrations at PRD is the strong subtropical insolation leading to large radical production rates by photolysis (cf. Fig. 6a). In fact, the mean noontime j (O 1 D) value was larger during PRIDE-PRD2006 than during other campaigns listed in to characterize the missing chemical processes that sustain the high experimental OH values, the discrepancies between modelled and measured OH will be investigated in the following discussion as a function of other atmospheric parameters. When comparing the reported HO 2 values in Table 6, it must be kept in mind that the data from different campaigns were measured by similar LIF techniques, which all 10 rely on chemical HO 2 conversion. It is likely that the interference from specific RO 2 species, described in Sect. 2.3, has influenced all LIF measurements of HO 2 compiled in Table 6. Since the abundance and speciation of RO 2 is certainly different for the various campaigns, a relatively large, not well quantified uncertainty has to be attached to the HO 2 comparison among different locations.

j(O 1 D) and NO X dependence of OH
A high correlation between OH and j (O 1 D) was observed over the whole range of atmospheric conditions during the PRD campaign, with a linear correlation coefficient of r 2 = 0.81 (Fig. 8a). In past campaigns, similar high correlations were observed (Rohrer and Berresheim, 2006, and reference therein), but at conditions with much lower VOC 20 reactivities. For limiting cases with low or high NO x concentrations, theoretical explanations were attempted using reaction schemes with a simplified VOC chemistry (e.g., Poppe et al., 1995). But even for the limiting cases in the steady state calculation of OH, the role of other photolysis processes (e.g. of HCHO and NO 2 ) is clearly visible and should in principle disturb the linear correlation between OH and j (O 1 D). Ehhalt 25 and Rohrer (2000)  that the observed correlation of OH and j (O 1 D) is just fortuitously linear, resulting from the combined influence of all photolytic processes on OH, which are highly correlated among themselves. According to this analysis, a power-law function with an exponential parameter close to unity is more suitable. Based on the analysis of a five year OH dataset, a simple empirical model was constructed to describe this relation (Rohrer 5 and Berresheim, 2006): The pre-exponential coefficient, a, incorporates the condensed information of the chemical conditions (e.g. NO x or VOCs) at a certain location. Exponent b reflects the combined effects of all photolytic processes (i.e. j (O 1 D), j (NO 2 ), j (HONO), j (HCHO)). 10 These photolytic processes either play a role in primary production processes of HOx radicals or influence its recycling processes. Finally, the offset parameter c accounts for non-photolytic OH sources. When applied to the scatter plot in Fig. 8a, a Levenberg-Marquard fit yields a = 5.6 × 10 6 cm −3 , b = 0.68 and c = 2.3 × 10 6 cm −3 , providing a parameterized description for the PRD conditions. 15 For comparison to other campaigns, a linear fit is more appropriate, since often only linear coefficients are published. The disadvantage in using linear coefficients is a shortfall in the description of twilight conditions for OH. The slope of a linear fit to the PRIDE-PRD2006 observations is 4.0 × 10 11 s cm −3 and the offset 2.4 × 10 6 cm −3 . In the marine boundary layer, reported slopes are relatively low, ≤ 2 × 10 11 s cm −3 20 (Berresheim et al., 2003;Brauers et al., 2001;Smith et al., 2006), while in continental regions reported values lie in the range of (2-4) × 10 11 s cm −3 (Holland et al., 2003;Rohrer and Berresheim, 2006;Ehhalt and Rohrer, 2000). The slope for PRIDE-PRD2006 is at the upper limit of previous continental field observations, which again indicates an intense photochemical activity in the PRD region. 25 The model simulation (M0) cannot reproduce the high correlation between measured OH and j (O 1 D). Unlike the experimental scatter plot (Fig. 8a), the plot of calculated OH versus measured j (O 1 D) is split into two groups of data forming a "V"-like shape Introduction  (Fig. 8b). The upper branch corresponds to morning hours, when modelled and observed OH are in good agreement (cf., Fig. 4). The lower branch of the "V" represents afternoon data, when the model severely underpredicts OH. One major difference between the two branches is the level of NO x , which was high in the morning and low in the afternoon (Table 3). Interestingly, the empirical relationship between measured 5 OH and j (O 1 D) is not only compact in both NO x regimes, but both subsets of data also overlap completely (Fig. 8a). Apparently, some unknown chemical mechanism which is missing in the base model stabilizes the ambient OH concentration. A similar tendency was noticed for a measured five-year OH record at Hohenpeissenberg, where the observed correlation between OH and j (O 1 D) was more stable and compact than 10 could be explained by a chemical box model (Rohrer and Berresheim, 2006). For inspection of the OH dependence on NO X , we remove the strong influence of j (O 1 D) by normalization as shown in Eq. (3).
Here, j (O 1 D) denotes the mean value of the j (O 1 D) data set. To avoid using conditions 15 during twilight for reasons discussed above, the data in this normalization analysis are restricted to j (O 1 D) larger than 1 × 10 −5 s −1 . The NO X dependency of observed and model calculated OH Jnorm are denoted as small dots in Fig. 8c,d, respectively. In addition, trend lines are shown that were obtained by averaging OH Jnorm concentrations over equal ln([NO X ]/ppb) intervals of 0.5.

20
The mean measured OH Jnorm is almost constant over the displayed NO X range, showing a broad curvature with a relative maximum of about 15 × 10 6 cm −3 at 3-4 ppb NO X (Fig. 8c). The observed NO X dependency differs considerably from the model simulation which predicts a steady increase of OH from low values at less than 1 ppb NO X to a maximum OH concentration at about 10 ppb NO X (Fig. 8d) air at POPCORN, where a nonlinear dependence with a maximum at 1 ppb NO 2 was found in good agreement for measured and modelled OH (Ehhalt, 1999).

NO and isoprene dependence of the measured-to-modelled OH ratio
To gain further insight into the shortcomings of the base model calculation M0, the ratio of measured-to-modelled OH, OH obs /OH mod , has been inspected as a function 5 of NO, CO and VOCs. A notable trend is found only for NO and isoprene (Fig. 9a,b), with correlation coefficients r = −0.77 and r = 0.48, respectively. Figure 9a displays a strong NO dependence of OH obs /OH mod with a smooth transition from ratios of about 8 at 0.02 ppb NO to unity ratio at NO mixing ratios larger than 1 ppb. This indicates that chemical processes are missing in the model which compete with NO dependent 10 reactions and become dominant at low NO. With respect to isoprene, two data clouds around 1 ppb and 3 ppb of isoprene are visible in Fig. 9b. For the combined data set, a weak positive correlation between OH obs /OH mod and isoprene seems to exist. No correlation is found for other investigated parameters. For example, indicators of anthropogenic activities like CO (espe- 15 cially from biomass burning) or benzene clearly show no correlation with OH obs /OH mod (Fig. 9c,d). Figure 10a offers a more detailed inspection of the relationship between OH obs /OH mod and isoprene by color coding the data according to the concurrently measured NO mixing ratios. It can be seen that part of the trend of OH obs /OH mod vs. 20 isoprene can be attributed to an anticorrelation between NO and isoprene. Isoprene was generally lower in the morning when NO was high and reached highest values in the afternoon when the NO mixing ratio was small. When the data in Fig. 10a are selected for low NO (< 0.5 ppb), no significant trend with isoprene is found (red circles). Thus, the weak positive correlation in the whole data set between OH obs /OH mod and 25 isoprene seems to be a consequence of the different NO X levels of the two data clouds around 1 ppb and 3 ppb of isoprene. Given the small dynamic range of isoprene and the influence of NO, a functional dependence on isoprene cannot be postulated on the basis of our PRD data. 11335 A model underprediction of measured OH concentrations by up to an order of magnitude has been reported for forested areas, which are characterized by isoprene emissions and low NO x (Tan et al., 2001;Ren et al., 2008;Lelieveld et al., 2008;Pugh et al., 2010;Whalley et al., 2011). Figure 10b compares our OH obs /OH mod ratios with the results from these other studies. For this purpose, we use the PRD data selected 5 for low NO (< 0.5 ppb) from Fig. 10a (red circles). The magnitude of the measuredto-modelled OH ratios of our study agree well with those from previous investigations, i.e. above deciduous forest in North-America during PROPHET (Tan et al., 2001) and INTEX-A (Ren et al., 2008), above the Amazonian rainforest during GABRIEL (Kubistin et al., 2010) and the Borneo rainforest during OP3 (Whalley et al., 2011). When all data 10 sets are combined, a consistent trend of an OH model underprediction with increasing isoprene seems to emerge, as pointed out previously by Ren et al. (2008) and Kubistin et al. (2010).
While the experimental OH data from the different campaigns were measured by similar LIF techniques, differences exist in the chemical models that were applied for 15 the OH simulations. Four different chemical mechanisms were used: RACM-MIM-GK for our study, RACM supplemented with a detailed isoprene chemistry and explicit ozonolysis of terpene (Tan et al., 2001) for PROPHET, MIM (Pöschl et al., 2000) for GABRIEL, and a lumped mechanism described by Crawford et al. (1999) for INTEX-A. For OP3, OH was calculated by an analytical equation with experimentally determined 20 parameters under photostationary steady-state assumptions. The trend in Fig. 10b is apparently independent of the specific model used for OH prediction. It suggests that an OH source mechanism is missing in current models which is related to biogenic emissions or to their photochemical daughter products. Anthropogenic VOCs made also a significant contribution to the OH reactivity at PRD (Lou et al., 2010), but their variability is too small to allow a positive identification of an influence on the OH model underprediction at PRD. for PRIDE-PRD2006. The OH loss and production rates were found to be balanced in the morning when HO 2 was efficiently recycled to OH, but a significant OH source was missing in the afternoon at low NO. Several generic reaction pathways were tested that may explain the mismatch within the OH budget. In that paper, observed OH and HO 2 concentrations were utilized as target parameters 10 for comparison with the model results. The most simple candidate for a new reaction pathway, which quantitatively explains the observations, was RO 2 + X −→ HO 2 in combination with HO 2 +X −→ OH (Table 5, M1). This is the same type of reactions as those of peroxy-radicals with NO. Assuming rate constants as for the NO reactions, a concentration of 0.8 ppb was needed for X to match the mean diurnal profiles of both HO X 15 species.
Since then, a strong interference in the HO 2 measurements by LIF from RO 2 was discovered (Fuchs et al., 2011, see Sect. 2.3). For this reason, the published results of the PRD campaign concerning the HO X budget have to be reevaluated. The previous major conclusion was that the recycling term k HO 2 +NO [NO][HO 2 ] in the OH budget was 20 too small to explain the observed total OH loss rate. Since the interference of HO 2 by RO 2 has one direction, namely the enlargement of HO * 2 with respect to the true HO 2 , the conclusion that additional OH recycling is missing in the model is still valid. Furthermore, the arguments made for the selection of the generic reaction terms with X still hold. However, a formerly excluded generic reaction type, a single reaction 25 HO 2 + Y −→ OH (  Figure 11 shows the mean diurnal variations of the modelled OH, HO 2 , HO * 2 and RO 2 calculated by the RACM-GK-MIM model without (M0) and with additional radical recycling (M1, M2). The results for OH and HO * 2 are also compared to the measured data.

5
In case of M0, there is agreement between modelled and measured OH in the morning, but a large discrepancy exists in the afternoon as discussed previously (Hofzumahaus et al., 2009). The modelled HO * 2 is about 30% larger than the modelled HO 2 and is in very good agreement with the measurement (HO * 2 ). Yet, even the modelled HO 2 is not significantly different from the measurement, given the combined experimental and 10 model uncertainties (Hofzumahaus et al., 2009).
In model runs M1 and M2, the concentrations of the unknown reactants X and Y were optimized for half-hourly bins to achieve a best fit to the observed OH in the time interval 10:00-18:00 CNST (Fig. 11). Agreement is also achieved for modelled and measured HO * 2 within the combined uncertainties. Thus, both types of additional recy- 15 cling scenarios (M1, M2) provide acceptable generic solutions to describe the observed HO x data.
Regarding the M1 case, the modelled HO * 2 is about 30% larger than the modelled HO 2 like in the base case (M0). In both cases, the RO 2 /HO 2 ratio is close to one and about half of the RO 2 species contribute significantly to the interference in the mea-20 sured HO 2 . In the M2 calculation there is a huge difference of more than an order of magnitude between modelled HO 2 and HO * 2 . The reason are very large calculated RO 2 /HO 2 ratios with values of 10-60 and corresponding ISOP/HO 2 ratios of 2-15 during the afternoon hours. The large difference of the RO 2 /HO 2 ratios between mechanism M1 and M2 would be a tool to distinguish which recycling type provides a more 25 realistic description. However, experimental data of RO 2 /HO 2 ratios are not available for PRD and are also generally scarce in literature. In previous campaigns, Matrix Isolation Electron Spin Resonance (MIESR) has been applied to measure directly peroxy radicals in forested and suburban environments. RO 2 /HO 2 ratios were observed Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | to vary from < 1 to 5 (Mihelcic et al., 2003). The ratio RO 2 /HO 2 has also been investigated by measurement techniques that apply chemical conversion, including LIF, peroxy radical chemical amplifiers (PERCA) and chemical ionization mass spectrometry (CIMS). Here, ratios of 4-15 were reported for a remote mountain site in Colorado (Stevens et al., 1997), < 1-9 for a rural area near Berlin (Platt et al., 2002), about 0.5 for a rural area in Pennsylvania , 1-3 for the free troposphere above Burkina Faso (Andrés-Hernández et al., 2010), about one for a remote mountain site in Italy (Hanke et al., 2002) and < 1-2 for the polluted outflow from Mexico City and Asian countries (Hornbrook et al., 2011). As reported by Fuchs et al. (2011) and Hornbrook et al. (2011), HO 2 measurement techniques like LIF or CIMS are susceptible to interferences from specific RO 2 radicals that are produced by atmospheric oxidation of alkenes, dialkenes like isoprene, and aromatics. Thus, the reported RO 2 /HO 2 ratios determined by these techniques were probably underestimated. Furthermore, to our knowledge, the RO 2 /HO 2 ratio has never been determined experimentally for conditions like in this study with k OH well above 15 s −1 . In conclusion, previously reported 15 values of ambient RO 2 /HO 2 -ratios do not allow to choose between scenarios M1 and M2.

Mechanistic chemistry updates
As outlined in the introduction, a number of new chemical mechanisms have been proposed recently to explain the large model underprediction of OH observed in forested 20 areas. Here, we test how well these OH regenerating mechanisms explain the discrepancy between modelled and measured OH at PRD. An overview of the tested mechanisms M3-M6 is given in Table 5. In addition, we show model results of the detailed MCMv3.1 (M7), which contains a more explicit description of the VOC chemistry compared to RACM-GK-MIM (M0). The corresponding model results for OH, HO 2 , 25 HO * 2 and RO 2 and the measured OH and HO * 2 data are compared for M0 and M3-M7 in Fig. 12 The implementation of additional OH formation from the reaction of HO 2 with acyl peroxy and β-keto peroxy radicals (Hasson et al., 2004;Jenkin et al., 2007;Dillon and Crowley, 2008) (M3) and with epoxide peroxy radicals (Paulot et al., 2009) (M4) has only a marginal impact on the modelled concentrations of OH, HO 2 , HO * 2 and RO 2 . For each species, the modelled curves (Fig. 12) of the different scenarios M0, M3 and M4 5 are virtually indistinguishable. The reason for the small sensitivity to the RO 2 + HO 2 reactions is the dominating influence of the competing peroxy radical reactions with NO given average mixing ratios of more than 0.2 ppb NO at PRD.
The mechanisms M5 and M6 calculate significantly larger OH concentrations than M0. The model M5 includes the isoprene chemistry LIM0 postulated by Peeters and 10 Müller (2010), whereas M6 contains additional OH formation by reaction of ISOP with HO 2 as proposed by Lelieveld et al. (2008) and Butler et al. (2008). Both mechanisms contain an "amplification" factor generating additional radicals within the radical recycling processes. In the LIM0 mechanism, amplification is achieved by the photolysis of HPALDs, which are assumed to have a yield of about one OH and one HO 2 (M5a), or 15 alternatively up to three OH plus one HO 2 (M5b). The M6 scenario has two variants with yields of two (M6a) and four (M6b) OH radicals. Among the different M5 and M6 scenarios, M5b gives the largest increase of modelled OH by about a factor of two in the afternoon, relative to the M0 calculation. However, there is still a significant gap compared to the experimental data. The model explains only 40-50% of the measured 20 OH values during the afternoon at PRD, while it reproduces 70-90% of the observed OH for the GABRIEL and INTEX-A campaigns (Stavrakou et al., 2010). In scenario M6b modelled OH increases on average by a factor of 1.7 in the afternoon, reaching only 30-40% of the observed OH concentration. The same mechanism, however, was able to explain the GABRIEL and OP3-1 results by increasing the modelled OH by up 25 to an order of magnitude (Lelieveld et al., 2008;Kubistin et al., 2010;Pugh et al., 2010). The different OH enhancements for PRIDE-PRD2006 and GABRIEL using M5b or M6b are caused by the different levels of NO, which were an order of magnitude larger at PRD compared to the Amazonian rain forest. At PRD, the competing peroxy radical Introduction reactions with NO decrease the sensitivity to radical amplification that is postulated in the mechanisms M5b and M6b. The M5 and M6 scenarios calculate HO * 2 values that are larger than in the base case and overpredict the observed data by a factor 1.2-1.4. The modelled RO 2 is close to the corresponding HO 2 concentrations at daytime, so that the RO 2 to HO 2 ratio seems 5 to be relatively independent of the chemical mechanisms used (M0, M3-M6). Table S7 in the Supplement shows that the relative contributions of speciated RO 2 radicals is not responding very much to changing OH concentrations. Thus, the difference between modelled HO 2 and HO * 2 is 20-40% for M3-M6, close to the 30% difference in M0. An additional model calculation was performed using the well established MCMv3.1 Jenkin et al., 2003) (M7). The calculated results for OH, HO 2 , HO * 2 and RO 2 agree well with the reference model (M0) especially at afternoon (Fig. 12). This demonstrates that the OH underprediction by RACM-GK-MIM is not specifically caused by the lumped representation of the VOC chemistry, but is a fundamental deficit in our current understanding of tropospheric chemistry.

NO dependence of model predictions
Two possible concepts have been discussed above to explain the difference between modelled (M0) and observed OH at PRIDE-PRD2006. One concept introduces generic reactions with unknown species driving additional radical recycling (M1, M2). The other concept relies on the implementation of peroxy-peroxy radical reactions that produce 20 additional OH rather than acting as a radical sink (M3-M6). The potential of each mechanism to explain the observed OH at PRD is compared in Fig. 13 as a function of NO at daytime conditions (j (O 1 D)> 1 × 10 −5 s −1 ). It is apparent that all mechanisms reproduce the measured OH reasonably well at NO concentrations above 1 ppb, while the spread among the model predictions is largest at the low end of the NO scale. 25 Not surprisingly, the generic mechanisms M1 and M2 show good agreement over the whole NO dynamical range, as a consequence of the numerical fitting of the unknown reactants X and Y (cf. Fig. 11 caused by peroxy radical reactions with NO, which compete with peroxy-peroxy radical reactions and become dominant at NO levels above 1 ppb. Among the mechanisms M3-M6, the LIM0 mechanism (M5b) and the proposed OH recycling by reaction ISOP + HO 2 (M6b) offer the largest potential to brigde the gap between modelled (M0) and measured OH at low NO. Considering an estimated ac-5 curacy for OH obs /OH mod of 45%, calculated by error propagation of the corresponding experimental and model uncertainties, the remaining discrepancy of a factor of two at NO < 0.2 ppb is still significant. The postulated mechanisms M5 and M6 are not yet confirmed by laboratory experiments. Thus, the remaining discrepancies in Fig. 13 may be due to the uncertainty of these mechanisms, or indicate other, so far unidentified processes that increased OH at PRD. Figure 13 also shows the modelled OH reactivity (M0) given by CO and VOCs, and its breakdown into different subgroups of organic components for each of the NO bins. The modelled reactivity has an almost constant value of 20 s −1 over the NO range of two orders of magnitude. Over the whole range of NO, isoprene was never dominant 15 making a contribution of 10-16% to the organic reactivity. However, together with its first generation of daughter products (OISO), isoprene made a significant reactivity contribution that increased towards lower NO values.

Heterogeneous radical loss
HO 2 radicals may be lost by heterogeneous uptake onto aerosol particles when com-20 peting gas-phase reactions of HO 2 are relatively slow. The potential influence of heterogeneous loss has been examined in model sensitivity studies mainly for marine environments (e.g., Carslaw et al., 1999Carslaw et al., , 2002Sommariva et al., 2006;Kanaya et al., 2000Kanaya et al., , 2007a. Up to a factor of two of HO 2 reduction was simulated when the heterogeneous uptake coefficient for marine aerosol was assumed to be maximum, Implementing this heterogeneous loss rate into model M0 causes a reduction of the calculated HO 2 by 50%. This implies that a heterogeneous uptake process has the potential to reduce HO 2 at PRD significantly. However, the estimated strong influence is speculative. Though laboratory studies with Cu-II doped aqueous particles have 5 shown large accomodation coefficients close to one (see overview, Kolb et al., 2010), measurements at salt solutions, soot and wet organic particles showed much smaller effective uptake coefficients in the range 0.01-0.1 (e.g., Thornton and Abbatt, 2005;Bedjanian et al., 2005;Taketani et al., 2009Taketani et al., , 2010. Furthermore, the influence of heterogeneous uptake will be strongly diminished by competing HO 2 reactions which are 10 required as additional OH sources (e.g., HO 2 + X). If HO 2 is heterogeneously lost, it will decrease the OH concentration as well. Alternatively, one could speculate about a heterogeneous reaction of HO 2 on particles that releases OH (or an OH precursor) into the gas-phase. However, the potential HO 2 reactivity towards aerosol is at least a factor of five too small compared the required reactivity to match the observed OH 15 (cf. Fig. 11g, scenario M2). Thus, it seems unlikely that a surface-catalyzed HO 2 to OH recycling process is the missing OH source.

OH radical budget analysis
Experimental and modelled (M0-M7) parameters that characterize the HO x chemistry during the afternoon hours 12:00-16:00 CNST at PRD are compared in and P (OH) ( Table 7), which is caused by minor OH sources. These include ozonol-10 ysis of alkenes, photolysis of peroxides, OH recycling by HO 2 + O 3 and prompt OH regeneration, e.g. from ISHP + OH. In case of the model scenarios M2-M6, the difference between D(OH) and P (OH) is larger than in the base case (M0), owing to the additional OH sources implemented in each different mechanism. The respective source strengths are largest for M1 and M2, because they were fitted to match the 15 observed OH. In the other models, the additional source strengths are not sufficient to match the observed OH. Models M3 and M4 have only marginal effects on the total OH production, whereas the scenarios M5b and M6b provide the most efficient additional OH sources among the newly proposed chemical mechanisms. It is interesting to note that the postulated OH formation from HPALD photolysis (M5b) is a factor of 20 1.6 more efficient than the corresponding photolysis of ozone and HONO. Since the HPALD photolysis is also expected to produce HO 2 (Peeters and Müller, 2010), the contribution from HPALD photolysis to the HO X radical pool would actually be larger than from the sum of photolysis rates of O 3 , HONO, and HCHO. reduced ozone production efficiency is expected from M1 and M2. In case of the scenarios M5b and M6b, a reduction of 30% is calculated. The large variability of the calculated ozone production efficiency between the different mechanisms demonstrates that the insufficient understanding of the radical recycling mechanism also introduces a significant uncertainty in the predictions of secondary pollutants.

Conclusions
Ambient OH and HO 2 concentrations were measured by LIF during the PRIDE-PRD2006 campaign at a rural site downwind of Guangzhou in the Pearl River Delta in summer 2006. The most obvious feature of the HO X photochemistry for this campaign was published by Hofzumahaus et al. (2009). A large imbalance between the experimentally determined total OH loss rate and production of OH from known radical sources was discovered, indicating a missing OH source at conditions of high isoprene (2 ppb) and low NO (0.1-0.2 ppb). A generic reaction pathway, RO 2 + X → HO 2 and HO 2 + X → OH was proposed to enable a chemical model to reproduce both the observed OH and HO 2 concentrations. In this work, we have reevaluated the dataset and 15 extended the model analysis, taking into account a newly discovered artefact in the LIF measurement of HO 2 . The major findings are: 1. Considering the interference from RO 2 in the HO 2 detection channel, the need for an additional HO 2 → OH recycling process persists. Since the true HO 2 concentration is smaller than the measured value HO * 2 , the missing OH source may be 20 up to 85% of the OH loss rate rather than 74% calculated previously. Moreover, the need of an adjacent recycling RO 2 → HO 2 to match the observed HO 2 has diminished, since the measured HO 2 data contain a contribution from RO 2 that is not known quantitatively. The concentrations of individual RO 2 species would be needed for a correction of the interference which is not attainable in retrospect. and likely influenced by unspecified interferences from RO 2 similar to our case.
3. The observed OH showed a consistent high correlation with j (O 1 D) over a broad range of NO x conditions. The correlation cannot be reproduced by model simulations, indicating that OH stabilizing processes are missing in current models (e.g., RACM-MIM-GK). 4. The observed OH exhibited only a weak dependence on NO x in contrast to model predictions. While modelled and measured OH agree well at NO mixing ratios above 1 ppb, a continuously increasing underprediction of the observed OH is found towards lower NO concentrations, reaching a factor of 8 at 0.02 ppb NO.

5.
A dependence of the modelled-to-measured OH ratio on isoprene cannot be de- 15 rived from the PRD data set due to the relatively small isoprene variability. However, the magnitude of the ratio fits into the isoprene dependent trend that was reported from other campaigns in forested regions.
6. Two recently postulated isoprene mechanisms (Lelieveld et al., 2008;Peeters and Müller, 2010) lead to significant enhancements of OH expected for PRD, but an 20 underprediction of the observed OH by a factor of 2 remains at low NO. If the photolysis of HPALDs is as efficient as proposed in the LIM0 mechanism by Peeters and Müller (2010), the corresponding OH formation at PRD would be more important than the primary OH production from ozone and HONO.
The isoprene mechanisms by Lelieveld et al. (2008) and Peeters and Müller (2010) 25 have shown potential to explain the unexpectedly large OH concentrations observed 11346 above forests during the GABRIEL and INTEX-A campaigns. The still significant underprediction of OH at the NO concentration regime of PRD may be explained either by the uncertainties of the postulated mechanisms which are not yet confirmed by laboratory studies, or by other so far unidentified OH sources that may have played a role. Further experimental investigations will be needed to get a full picture of the radical 5 chemistry in VOC rich environments. Improved measurement techniques for HO 2 and (speciated) RO 2 would be extremely helpful in future studies to gain more insight into the cycling of radicals and its impact on tropospheric photochemistry.