A unifying basis of auditory thresholds based on temporal summation

Peter Heil* and Heinrich Neubauer*

*Leibniz Institute of Neurobiology, Brenneckestrasse 6, 39118 Magdeburg, Germany; and Institute of Experimental Audiology and Medical Physics, Otto-von-Guericke University, 39118 Magdeburg, Germany Edited by A. James Hudspeth, The Rockefeller University, New York, NY, and approved March 13, 2003 (received for review December 30, 2002)

Thresholds of auditory-nerve (AN) bers and auditory neurons are commonly specied in terms of sound pressure only, implying that they are independent of time. At the perceptual level, however, the sound pressure required for detection decreases with increasing stimulus duration, suggesting that the auditory system integrates sound over time. The quantity commonly believed to be integrated is sound intensity, implying that the auditory system would have an energy threshold. However, leaky integrators of intensity with time constants of hundreds of milliseconds are required to t the data. Such time constants are unknown in physiology and are also incompatible with the high temporal resolution of the auditory system, creating the resolutionintegration paradox. Here we demonstrate that cortical and perceptual responses are based on integration of the pressure envelope of the sound, as we have previously shown for AN bers, rather than on intensity. The functions relating the pressure envelope integration thresholds and time for AN bers, cortical neurons, and perception in the same species (cat), as well as for perception in many different vertebrate species, are remarkably similar. They are well described by a power law that resolves the resolutionintegration paradox. The data argue for the integrator to be located in the rst synapse in the auditory pathway and we discuss its mode of operation.

he detection of sounds constitutes a prerequisite for their further analysis by the auditory system, and a thorough understanding of the systems operation requires knowledge of how threshold is determined. Surprisingly, confusion still exists as to how auditory thresholds are best defined and by which process they are reached. In physiological and audiological studies of the auditory system, thresholds are routinely specified in terms of the sound pressure level (SPL) of the signal needed to evoke a response. This practice, exemplified by tuning curves and audiograms, implies that thresholds of auditory neurons are sufficiently characterized by that pressure and are independent of stimulus duration (1). However, at the perceptual level, the threshold SPL decreases as stimulus duration increases in every species examined (223). Such a trading relationship between a sounds amplitude, or a related parameter, and its duration is consistent with the notion that the auditory system summates or integrates sound over time. It is commonly believed that the physical quantity of sound ultimately integrated by the auditory system is sound intensity, I(t) (211, 1315, 1719, 2129). I(t) is the sound power transmitted per unit area and, for pure tones, is proportional to the square of the peak pressure or the pressure envelope, P(t) (30). Sound power integrated over time yields acoustic energy (30), and so the common interpretation of the perceptual data implies that the auditory system has a threshold that is best specified in terms of the sounds acoustic energy density, and not its pressure. Notably, the energy of threshold stimuli is also not constant but, in contrast to threshold SPL, generally increases with increasing stimulus duration. Only for very short durations has the reverse been suspected (21). This time dependence of the energy of threshold stimuli has been attributed to leaky integration of I(t), combined with spectral spread of energy and limited spectral integration at short durations. However, the physiological foundations of this interwww.pnas.org cgi doi 10.1073 pnas.1030017100

Methods
Electrophysiology. As detailed (35, 39), spikes generated by single AN fibers and single AI neurons were recorded in a soundattenuating chamber from 10 adult barbiturate-anaesthetized cats. AN fibers and AI neurons respond preferentially at sound onset (3439), where I(t) and P(t) both increase dynamically, but differently, because I(t) P(t)2. To exploit these differences, we stimulated each neuron with tones at its characteristic frequency (CF; the frequency to which a neuron is most sensitive) while varying P(t). The latter was achieved by employing different onset functions and systematically varying both the time to the steady-state SPL (onset time) and the SPL. Onset functions were either linear, i.e., P(t) increased linearly with time, t, during the onset time, tr, to the maximum, Pmax: P(t) Pmax t/tr, with onset times of 1, 5, 10, 50, and 100 ms; or cosine-squared (cos2), i.e., P(t) Pmax sin2( /2 t/tr), with onset times of 1.7, 4.2, 8.5, 17, 42, 85, and 170 ms. The SPL was increased from low to high values (usually 090 dB SPL during steady-state) in 5- or 10-dB steps. For each combination of parameters, the stimulus was repeated

This paper was submitted directly (Track II) to the PNAS ofce. Abbreviations: AI, primary auditory cortex; AN, auditory nerve; CF, characteristic frequency; IHC inner hair cell; SPL, sound pressure level.
To

whom correspondence should be addressed. E-mail: peter.heil@ifn-magdeburg.de.

PNAS

May 13, 2003

vol. 100

no. 10

6151 6156

NEUROSCIENCE

pretation are unclear. Very long time constants, up to hundreds of milliseconds, are needed to describe the increase in threshold energy with increasing duration (for summaries, see refs. 18 and 22). Such long time constants contrast with the short membrane time constants of neurons in the auditory periphery (28, 29, 31, 32) and are also difficult to reconcile with the high temporal resolution of the auditory system. Several ideas have been proposed to resolve this resolutionintegration paradox (6, 7, 18, 32, 33). Here we develop a unifying description of auditory thresholds at the physiological and perceptual levels, disclose the parameter that is really traded for stimulus duration, and aid the identification of the processes by which thresholds are reached. We exploit the stimulus-dependent timing of the first spike of auditory neurons to derive their thresholds. Our approach is based on the rationale that the first spike following the onset of a stimulus is triggered when the stimulus reaches the neurons threshold and occurs with some fixed transmission delay thereafter. In this way, we have previously shown that thresholds of auditory nerve (AN) fibers and primary (AI) and other auditory cortical neurons are not specified in terms of fixed pressures (3438). In fact, the instantaneous values of P(t), and hence also of I(t), at which the first spike was triggered decreased as first-spike latency increased, a trading relationship consistent with temporal summation. Here we first examine whether thresholds of AI neurons, as well as perceptual thresholds, are largely determined by integrating I(t), and thus whether they are specified in terms of acoustic energy or, as is the case for AN fibers (38), by integrating P(t). Second, and equally important, we examine the characteristics of the time dependence of these thresholds.

2050 times. Tones had durations of 200 ms (AN) or 400 ms (AI) and were presented at 2 Hz (AN) or 1 Hz (AI) through sound-delivery tubes sealed into the external meati. The ipsilateral ear was used for AN fibers, and the contralateral, ipsilateral, or both ears were used for AI neurons, depending on which was most effective. The instances at which the recorded spikes, after conventional amplification and filtering, first passed a Schmitt-trigger level were stored with a 10- s precision for offline analysis. The first spike following each stimulus onset was used to calculate mean latency. The analysis window always commenced with stimulus onset and usually terminated shortly after stimulus offset (10 ms for AN fibers) to allow for transmission delays. The responses of most AI neurons consisted of 13 spikes tightly locked to stimulus onset. Additional late discharges, if present, were excluded from the analysis. for 11 human subjects (four tested twice) and five cats were measured as detailed (18, 19) and kindly provided by G. M. Gerken (University of Texas, Dallas). All subjects were tested with a set of single- and multiple-burst stimuli differing in duration and envelope characteristics (Fig. 2a). The carrier frequency was 3.125 kHz for humans and 6.25 kHz for cats. The shortest stimulus (stimulus 1) was composed only of onset and offset portions of 4.16 ms each [P(t) of cos2 function]. Multipleburst stimuli were composed of stimulus 1 repeated to form stimuli that differed either in number of bursts (multiple-burst series: stimuli 15) or in interburst interval (interval series: stimuli 3 and 2124). A single-burst series (stimuli 614) was created by adding a constant amplitude plateau in the middle of the stimulus, and an onsetoffset series (stimuli 1 and 1520) was created by varying onset and offset times. For humans (right ear) and cats (binaural), thresholds were determined for all 24 stimuli and 10 stimuli (15, 6, 8, 10, 12, and 14), respectively, by using a simple updown method with decreasing step size. Threshold was the average of six transitions using 2-dB steps in humans and 2.5-dB steps in cats. Following training (at least four complete sets of 24 thresholds in humans and 10 in cats), mean threshold for a subject was based on 12 such thresholds for each of the 24 stimuli in humans (10 in cats). We obtained a grand mean for all subjects from the individual mean threshold SPLs by minimizing the sum of the total variance across subjects without altering the mean SPL across all subjects and stimuli. Results and Discussion Background. An important step toward the aims of the present study was to derive thresholds of AI neurons from the stimulus dependence of their mean first-spike latency, L. Fig. 1 shows the data for one representative neuron. Fig. 1a illustrates that for each onset function, L generally decreased as the SPL of the stimulus tones increased and their onset times decreased, finally approaching a neuron-specific minimum, Lmin, the constant transmission delay (3437). In our study of AN fibers (38), whose latencies behave similarly (38, 39), we initially explained firstspike timing by assuming that threshold, T, is constant and equal to the temporal integral, from stimulus onset to the trigger time of the first spike, L Lmin, of P(t) raised to some power, q:
L

Psychoacoustics and Perception. Detection thresholds (in dB SPL)

Fig. 1. AI thresholds result from temporal integration of the pressure envelope of a sound rather than its intensity. Data are from a single cat AI neuron (except frequency distribution in d) tested at its CF of 22 kHz. (a) Mean rst spike latency, used to derive threshold, plotted against SPL for different onset times and onset functions [linear or cos2 increase in P(t) with time; see key]. (b) Mean latencies (key as in a) plotted against those predicted by a pressure envelope integration threshold that increases linearly with integration time, L Lmin (Eq. 2 with q 1). Fit parameters were Lmin 12.06 ms; T0 1.51 Pa s, and Pc 27.1 Pa. The mean residuals (solid line, right ordinate) reveal only small, but systematic deviations of the data from that model. (c) Individual thresholds (key as in a) for the AI neuron, plotted against integration time. The scatter of individual thresholds and the systematic differences between mean thresholds obtained with different rise functions are smaller for integration of P(t) (q 1; Upper) than for integration of I(t) (q 2; Lower). Lmin, the only free parameter of the ts, was 12.21 ms for q 1 and 12.57 ms for q 2. (d) Costs (purple symbols) and proportion of data points for which Lmin L (green triangles) as a function of the exponent q of P(t), and distribution of the optimal q obtained from 32 neurons tested with different onset functions (black dots, right ordinate). The geometric mean was 0.914 with an error interval of 0.7911.054.

well, but systematic deviations of the data from that model were evident in some fibers. These deviations could be well explained by time-dependent thresholds and an extension of Eq. 1.
L

T L

L min
0

Lmin

P t qdt

T0

Pc L

Lmin

[2]

Compared with Eq. 1, this approach resulted not only in better fits but also in an even narrower distribution of q around 1 (38). Pc (in Pa) can be viewed as a system-inherent pressure-like force that adds linearly to the external sound pressure. not explain the data of AI neurons, indicating that it is also necessary to assume a time dependence of AI thresholds. Eq. 2 described this time dependence satisfactorily for a fraction of AI neurons. Fig. 1b plots, for the same neuron as in Fig. 1a, the measured latencies against those predicted by Eq. 2 when q 1. Despite the rather close match, the residuals (Fig. 1b, solid line, right ordinate) reveal systematic deviations of the data from the model. These suggest that a slight modification of Eq. 2, namely a power law
Heil and Neubauer

Refinement of Previous Models. A fixed integration threshold could

T
0

Lmin

P t qdt

Tconst,

[1]

where T, Lmin, and q are free parameters. The distribution of q obtained from 89 AN fibers was narrow and centered on 1, and not 2, suggesting that AN fibers reach threshold by temporal integration of P(t), and not of I(t). Fixed pressure envelope integration thresholds (q 1) explained the data of most fibers
6152 www.pnas.org cgi doi 10.1073 pnas.1030017100

Fig. 2. Perceptual thresholds result from temporal integration of the pressure envelope of a sound rather than its intensity. (a) Pressure envelopes of the stimuli for which detection thresholds were obtained. Multiple-burst series (red circles), stimuli 15; single-burst series (blue circles), 6 14; onset offset series (green triangles), 1 and 1520; and interval series (yellow squares), 3 and 2124. The same symbols are used in bd and f. Abscissa duration, 275 ms. [Reproduced and modied with permission from Gerken et al. (18) (Copyright 1990, Acoustical Society of America).] (b) Grand means of threshold SPL plotted against stimulus duration for human subjects (right ordinate) and cats (left ordinate). Stimulus duration excludes interburst intervals. Inset shows the upper half of P(t) for eight consecutive bursts (red line) and for a single burst of the same duration but of a 6-dB-lower amplitude (blue area), which approximates to the observed differences in threshold SPLs for such stimuli in humans and cats (arrows). The temporal integrals of P(t) for such stimuli are nearly identical. (c and d) Thresholds of humans (right ordinates) and cats (left ordinates) expressed as temporal integrals of I(t) (i.e., as energy densities, calculated with a specic impedence of 414 Pa s m 1; c) and as temporal integrals of P(t) (d) plotted against stimulus duration. The variances unexplained (1 r2) by power laws between the respective threshold quantity and duration are given. (e) Geometric mean and range of the ratios of unexplained variances for integration of P(t) to integration of I(t) of individual subjects. ( f) Threshold SPL plotted against stimulus duration for theoretical rectangular stimuli (gray squares). The human data from b for single-burst stimuli, multiple-burst stimuli, and interval series are replotted. Straight red and black lines are power law ts for red and gray symbols, respectively, and the solid curved line was calculated from the latter for stimuli with cos2 onset and offset times of 4.16 ms. Envelopes of multiple-burst (red), single-burst (blue), and rectangular (gray) threshold stimuli with durations of 8.32 and 33.28 ms are depicted.
NEUROSCIENCE

T L

L min

L Lmin

P t qdt

k L

Lmin m,

[3]

would provide an even better description, a fact that will become apparent later (Figs. 2 and 3). The exponent m corresponds to Lmin in the slope of the increase in T(L Lmin) with L double-log scales, and the scaling factor k defines the curves vertical position. When L Lmin 1 ms, T(L Lmin) T1. Eqs. 2 and 3 are intimately related, because the right-hand term of Eq. 2 comprises the first two terms of a Taylor series approxi0. mation of the power law, evaluated at a point L Lmin The presentation of the data analysis will first deal with the value of q in the middle term of Eqs. 13, i.e., with the issue of whether P(t) or I(t) is integrated to reach threshold, both at the AI neuronal and the perceptual level. We will then turn to the time dependence of the thresholds defined in this way, including the value of m in the right-hand term of Eq. 3.
AI Thresholds Are Largely Determined by Temporal Integration of P(t).

The optimal exponent q of P(t) for a given AI neuron was obtained by minimizing a cost function that depended on q and Lmin. The cost function was defined such that its minimum was obtained when the spread of thresholds at any given integration time and systematic differences between the thresholds obtained with linear and cos2 onset functions were minimal (Fig. 5, which is published as supporting information on the PNAS web site,
Heil and Neubauer

www.pnas.org). This novel approach only postulates that thresholds to those CF tones of different onset function, onset time, and SPL that elicit responses with identical L Lmin should be identical. It imposes no a priori constraints on thresholds for different values of L Lmin. Thus, threshold may be constant (as in Eq. 1) or may form any function of L Lmin (e.g., those of Eqs. 23). Fig. 1d shows that the cost function (purple) of the example neuron is W-shaped with its global minimum near q 1. Further, the proportion of stimuli for which the measured latency L was shorter than the estimated Lmin increased steadily from near zero for q 1 to 20% for q 2 (Fig. 1d, green triangles), rendering temporal integration of I(t) rather implausible. Fig. 1c shows that when the thresholds (obtained from best fits) to all tones for this neuron are plotted against integration time, the vertical scatter of the data points is appreciably smaller for q 1 (Upper) than for q 2 (Lower). For both exponents, thresholds vary with integration time, an issue that will be dealt with later. Across the AI sample, the optimal values of the exponent q of P(t) were narrowly distributed around 1, and not 2 (Fig. 1d, black dots, right ordinate). These findings demonstrate that AI neurons, like AN fibers (38), reach threshold by temporal integration of P(t), and not I(t).
Perceptual Thresholds Are also Largely Determined by Temporal Integration of P(t). To also distinguish between integration of P(t)

and of I(t) at the perceptual level, we next reanalyzed perceptual thresholds for both cats and humans tested with a large set of
PNAS May 13, 2003 vol. 100 no. 10 6153

Fig. 3. Comparison of the pressure envelope integration thresholds for neuronal (a and b) and perceptual (c and d) measurements. Shown are plots of thresholds for 5 AN bers (a) and 5 AI neurons (b) with different CFs (different colors). Lines connect thresholds derived from tones of different SPLs but of the same onset function and onset time. (c) Analogous functions for perceptual thresholds for two cats (purple and black circles) and three humans (red, green, and blue triangles). Lines connect thresholds obtained with stimuli of different series (see Fig. 2a). (d) Analogous functions for mean perceptual thresholds for a range of vertebrate species, calculated from published data (8 14, 1719). For the porpoise, parakeet, and eld sparrow, absolute thresholds could not be determined from the original articles and thus the position on the ordinate is arbitrary.

single- and multiple-burst tones that differed in duration and envelope characteristics (refs. 18 and 19; Fig. 2a). The results (Fig. 2 bf ) show clearly that, also at this level, integration of P(t) provides a much better fit to the data than integration of I(t). The SPL corresponding to the peak amplitude of threshold stimuli decreased with an increasing number of bursts (stimuli 15 in Fig. 2a), plateau duration (stimuli 1 and 614), and onsetoffset times for a single burst without plateau (stimuli 1 and 1518) (Fig. 2b). These trading relationships between threshold SPL and stimulus duration are all consistent with thresholds reached by temporal summation. Threshold SPL did not differ for stimuli that were temporal mirror images (stimuli 19 and 20) or for multiple-burst stimuli differing in interburst interval (stimuli 3 and 2124) (the five yellow squares representing the threshold SPLs for these stimuli in Fig. 2 bd and f fall on top of one another). The latter result is inconsistent with the leaky integration hypothesis, which would predict an increase in threshold SPL with increasing interburst intervals because leakage would continue during these intervals and its net effect would increase with time (18). Threshold SPL for stimuli in the multiple-burst, single-burst, and onsetoffset series diverged as duration increased (Fig. 2 b and f ). Thus, threshold SPL is not an invariant function of stimulus duration. The energy density of threshold stimuli, i.e., the temporal integral of I(t) (in W s m 2), increased as the number of bursts, plateau duration, and stimulus onsetoffset times increased, again with pronounced divergence of the thresholds between the series (Fig. 2c). This shows that threshold energy is also not an invariant function of stimulus duration, contrary to what would be expected from a temporal integrator of I(t). In contrast, a much closer alignment of the thresholds for stimuli in the different series was obtained by plotting the temporal integral of P(t) (in Pa s; Fig. 2d), again, as for the calculation of energy densities, without any correction of onset offset versus plateau duration. The closer alignment of thresholds for q 1 compared with q 2 was expected from the 6-dB difference between threshold SPLs for single- and multiple-burst stimuli of longer durations (Fig. 2b Inset). Further, it matches the finding on AI neurons (Fig. 1c). Hence, perceptual, like neuronal, thresholds are much better defined by the temporal integral of P(t) than that of I(t), or by SPL only.
6154 www.pnas.org cgi doi 10.1073 pnas.1030017100

perceptual pressure envelope integration thresholds increase almost linearly with stimulus duration on double-log scales for both cats and humans. Consequently, and as stated above, the power law T ts
ts

The Time Dependence of Pressure Envelope Integration Thresholds Is Well Described by a Power Law. Figs. 2d and 3c reveal that

P t 1dt

m k ts

[4]

provides an excellent descriptor of the data that leaves remarkably little unexplained variance (0.84% and 1.11% of the grand means for humans and for cats, respectively), particularly because m and k are the only two free parameters. The duration Lmin of of the threshold stimulus, ts, replaces the term L Eq. 3 and T(ts) T1 for ts 1 ms. The unexplained variance in Eq. 4 was smaller for each subject than that given by the equivalent power law relating threshold energy and duration by, on average, an order of magnitude (Fig. 2e). This was similarly true for the grand mean data (unexplained variances given in Fig. 2 c and d) and again emphasizes that integration of P(t) provides a much better explanation for threshold than of I(t). We also reanalyzed available perceptual thresholds from other studies of temporal summation in various mammals, birds, and fish (814, 17) and replotted the data (Fig. 3d). It is apparent that in each species the dependence of pressure envelope integration thresholds on stimulus duration can be well described by Eq. 4. The adequacy of this power law may be obscured in the conventional plot of threshold SPL versus log duration, particularly when only single-burst stimuli with fixed onset and offset times are used. For such stimuli, threshold SPLs increase more rapidly the shorter the stimulus duration becomes, as illustrated with high resolution for the grand mean human data in Fig. 2f (blue circles). This curved shape could suggest an exponential function with a long time constant. However, it results from the concomitant increase in the proportion of onset and offset times with respect to the entire duration of the tone. To illustrate this, we calculated the threshold SPLs for theoretical stimuli with rectangular P(t) that had durations and temporal integrals of P(t) identical to those of the single-burst threshold stimuli by dividing the temporal integral of P(t) of each stimulus by its duration (gray squares in Fig. 2f ). These theoretical threshold SPLs fall on
Heil and Neubauer

decreases, as m increases by 32 5 dB per m (Fig. 4b). In cats, perceptual sensitivity is 20 dB better than that of the most sensitive neurons (Fig. 4b), but this is readily explained by the conditions used for neuronal recording (anesthesia; stimulation of mostly one ear only and near the tympanum, thus disenabling pinna gain). The values of m are all 1. This excludes a fixed pressure threshold, for which m 1 and the integral T(ts) ts, and corroborates our previous conclusions (3438). The present analyses thus show that both neuronal and perceptual thresholds are better measured in units of pressure multiplied by time (e.g., Pa s), with additional specification of that time, than in units of pressure only or in units of energy.
Fig. 4. Detection thresholds are likely mediated by the most sensitive neurons. (a) Histograms of values for m (Eq. 4) obtained from AI neurons for integration times 8.32 ms (blue) and perceptual tests of cats (yellow). (b) Sensitivity increases, i.e., T1 decreases, as m increases, and regression (dashed blue line) obtained from individual AI neurons (n 57, r2 0.44); symbols show geometric means in bins of 0.1 width. The yellow ellipse shows the mean 1 SD of T1 and m for perceptual data of cats. m derived from perceptual tests matches that of the most sensitive neurons.

a straight line (black line through gray squares in Fig. 2f ) with a slope of 20(m 1) (here, 5.33 0.09 dB per decade of duration). Threshold SPLs measured for multiple-burst stimuli fall on a line (red line through red circles and yellow squares in Fig. 2f ) of similar slope ( 5.64 0.10 dB per decade of duration), because the proportion of onset and offset times also does not change with duration. The line is elevated by 6.15 0.22 dB, i.e., a factor of 2, relative to that for the rectangular stimuli, because the peak amplitudes of multiple-burst stimuli are twice as high as their mean amplitudes and hence as the peak amplitudes of equivalent rectangular stimuli. The power law (Eq. 4 with L Lmin instead of ts) can also well describe the functions relating the pressure envelope integration thresholds to the integration time of cat AI neurons (Fig. 3b) and AN fibers (Fig. 3a), at least over the time range corresponding to the durations tested for perception in that species (Fig. 3 cd). Deviations from the power law can be observed for short integration times in some AI neurons (purple, black, and blue symbols in Fig. 3b) and for long integration times in some AN fibers (purple symbols in Fig. 3a) and are readily explained as follows. In some neurons of the central auditory pathway, latency decreases with increasing SPL for low SPLs but for higher SPLs remains relatively constant (40) or even increases with further increments in SPL (36, 4144). This paradoxical latency shift, which is likely mediated by fast inhibition (43, 44), converts to the steep increase of pressure envelope integration thresholds for short integration times. In AN fibers, latency increases with decreasing SPL until an upper limit is reached that is caused, and its value determined, by the fibers spontaneous activity (38, 39). This upper limit produces roughly constant latencies to low-SPL stimuli and consequently results in a nearly vertical drop of threshold estimates for such stimuli.
The Slope m of the Time Dependence of Pressure Envelope Integration Thresholds. The exponent m of Eq. 4, i.e., the slope of the increase

The Location of the Integrator. Because the functions relating the pressure envelope integration threshold to integration time or stimulus duration for AN fibers, AI neurons, and perception in the same species, namely the cat, are closely matched (Fig. 3 ad), the integrator is most likely peripheral to the site of spike generation in the AN fibers, and more central processes (6, 7, 18, 32, 33) do not seem necessary to explain the present data. The peripheral boundary for the location of the integrator is marked by the membrane potential of the sensory receptors, the inner hair cells (IHCs). This potential almost instantaneously follows the fine structure of the stimulus at low frequencies and the pressure envelope, P(t), at higher frequencies, without any slow changes of its DC component when P(t) is constant (45). These observations agree well with the short, 1.4-ms (46), IHC membrane time constants. Thus, the integration of P(t) over the observed long time scales cannot have been accomplished at, or peripheral to, the level of the receptor potential. Also, the observation that the sensitivity of AN fibers of similar CF in the same individual can differ by several orders of magnitude (Fig. 6, which is published as supporting information on the PNAS web site), together with the compelling evidence that each IHC is innervated by 1030 AN fibers of different spontaneous rates and thresholds (47, 48), argues against an integrator identical for all afferent fibers of a given IHC (38). This limits the possible location of the integrator to the first synapse in the auditory pathway, between the IHC and the distal dendrite of a single AN fiber, including that distal dendrite with its additional synaptic inputs, just below the IHC, from efferent fibers of the lateral olivocochlear system (49). This synaptic region is structurally diverse, which could readily account for the range of AN fiber sensitivities, because there are systematic relationships between morphological and physiological properties (4750). The Possible Nature of the Processes Leading to Threshold. To identify the parameter really traded for stimulus duration, and so to pinpoint the underlying mechanism of temporal summation, Eq. 4 was reformulated such that this parameter, R, multiplied with duration ts yields a constant

R ts

const.

[5]

of the pressure envelope integration thresholds with stimulus duration, averaged 0.76 0.09 for humans and 0.69 0.06 for cats. For the other species (814, 17), we obtained very similar values, ranging from 0.67 to 0.79, and, after weighting them with the number of individuals, frequencies, and stimuli tested, a mean m of 0.75 (Table 1, which is published as supporting information on the PNAS web site). For cat AI neurons, the values of m were more widely distributed (Fig. 4a, blue bars), but for some neurons the values match those for perception (Fig. 4a, yellow bars). These are also the most sensitive neurons, because sensitivity increases, i.e., T1
Heil and Neubauer

Here R c (P) , (1 m) 1, c const (1/k) , and P [ ts P(t)dt]/ts is the mean amplitude of P(t) during ts. R (in s 1) 0 can be interpreted as a mean rate of individual events, possibly point processes. This interpretation results from the units of R, as well as from the plausible idea that a certain number of individual events may be necessary to reach threshold. Thus, the higher (lower) the mean rate of these events, the shorter (longer) the time needed to accumulate the necessary number. This general interpretation of this trading law does not require any assumptions about the necessary number of events or about their distribution (homogeneous or inhomogeneous) over time. Such an accumulation process of individual events, as described by Eq. 5, is not equivalent to the temporal integration of a continuously changing quantity.
PNAS May 13, 2003 vol. 100 no. 10 6155

NEUROSCIENCE

Eq. 5 reveals that there is no real resolutionintegration paradox. Because R P , the duration, ts, over which the events need to be accumulated is short when P is high, indicating high temporal resolution. Conversely, when P is low, ts increases and the temporal resolution decreases, consistent with available data (51, 52). Our reanalysis of the perceptual data (Fig. 3d) reveals no obvious deviations from Eq. 5, in contrast to proposed relatively short maximum integration times ( 200 ms; ref. 29). Eq. 5 also provides a hint to the nature of the integrator. The probability of the occurrence of the individual events, proportional to P , can be viewed as a conditional probability that results from the interaction of subevents, the probability of each of those occurring being proportional to P. Calculating from m for the perceptual data yields values between three and
1. Narayan, S. S., Temchin, A. N., Recio, A. & Ruggero, M. A. (1998) Science 282, 18821884. 2. Hughes, J. W. (1946) Proc. R. Soc. London Ser. B 133, 486490. 3. Garner, W. R. (1947) J. Acoust. Soc. Am. 19, 808815. 4. Green, D. M., Birdsall, T. G. & Tanner, W. P., Jr. (1957) J. Acoust. Soc. Am. 29, 523531. 5. Plomp, R. & Bouman, M. A. (1959) J. Acoust. Soc. Am. 31, 749758. 6. Zwislocki, J. J. (1960) J. Acoust. Soc. Am. 32, 10461060. 7. Watson, C. S. & Gengel, R. W. (1969) J. Acoust. Soc. Am. 46, 989997. 8. Johnson, C. S. (1968) J. Acoust. Soc. Am. 43, 757763. 9. Ehret, G. (1976) J. Acoust. Soc. Am. 59, 14211427. 10. Dooling, R. J. (1976) J. Acoust. Soc. Am. 65, 10581060. 11. Fay, R. R. & Coombs, S. (1983) Hearing Res. 10, 6992. 12. Costalupes, J. A. (1983) Hearing Res. 9, 4354. 13. Brown, C. H. & Maloney, C. G. (1986) J. Acoust. Soc. Am. 79, 10581064. 14. Clark, W. W. & Bohne, B. A. (1986) in Sensorineural Hearing Loss, eds. Collins, M. J., Glattke, T. J. & Harker, L. A. (Univ. of Iowa, Iowa City), pp. 5982. 15. Au, W. W. L. (1988) in Animal Sonar: Processes and Performance, eds. Nachtigall, P. E. & Moore, P. W. B. (Plenum, New York), pp. 753768. 16. Florentine, M., Fastl, H. & Buus, S. (1988) J. Acoust. Soc. Am. 84, 195203. 17. Klump, G. M. & Maier, E. H. (1990) J. Comp. Psychol. 104, 94100. 18. Gerken, G. M., Bhat, V. K. H. & Hutchison-Clutter, M. (1990) J. Acoust. Soc. Am. 88, 767778. 19. Solecki, J. M. & Gerken, G. M. (1990) J. Acoust. Soc. Am. 88, 779785. 20. Fay, R. R. (1992) in The Evolutionary Biology of Hearing, eds. Webster, D. B., Fay, R. R. & Popper, A. N. (Springer, New York), pp. 229263. 21. Schmidt, S. & Thaller, J. (1994) Hearing Res. 77, 125134. 22. OConnor, K. N., Barruel, P., Hajalilou, R. & Sutter, M. L. (1999) J. Acoust. Soc. Am. 106, 954965. 23. Ronacher, B., Krahe, R. & Hennig, R. M. (2000) J. Comp. Physiol. A 186, 10651072. 24. Gollisch, T., Schutze, H., Benda, J. & Herz, A. V. M. (2002) J. Neurosci. 22, 1043410448. 25. Syrlykke, A., Larsen, O. N. & Michelsen, A. (1988) J. Comp. Physiol. A 162, 367374. 26. Tougaard, J. (1996) J. Comp. Physiol. A 178, 669677. 27. Tougaard, J. (1998) J. Comp. Physiol. A 183, 563572. 28. Clock, A. E., Salvi, R. R., Saunders, S. S. & Powers, N. L. (1993) Hearing Res. 71, 3750. 29. Green, D. M. (1985) in Time Resolution in the Auditory System, ed. Michelsen, A. (Springer, Heidelberg), pp. 122140.

five (Table 1). Thus, it seems meaningful to search for events that are mediated by three, four, or five subevents. One candidate for such events is exocytosis at the IHCAN fiber synapse, for which, in mouse (53), four to five Ca2 -binding steps are necessary. Numbers between three and five have been reported for peripheral and central synapses in squid, frog, and rat (5458).
We thank J. Altman, M. Brosch, T. Gollisch, A. V. M. Herz, R. Meddis, and three anonymous reviewers for helpful comments on earlier versions of the manuscript, D. R. F. Irvine for helpful comments on the manuscript and collaboration in the recording of electrophysiological data, and G. M. Gerken for helpful comments on the manuscript and access to his perceptual data. This work was supported by the Bundesministerium fur Bildung und Forschung, the state of Sachsen Anhalt, and grants from the Deutsche Forschungsgemeinschaft to P.H.
30. Buser, P. & Imbert, M. (1992) Audition (MIT Press, Cambridge, MA). 31. deBoer, E. (1985) in Time Resolution in the Auditory System, ed. Michelsen, A. (Springer, Heidelberg), pp. 141158. 32. Viemeister, N. F. & Wakefield, G. H. (1991) J. Acoust. Soc. Am. 90, 858 865. 33. Dau, T., Puschel, D. & Kohlrausch, A. (1996) J. Acoust. Soc. Am. 99, 36233631. 34. Heil, P. & Irvine, D. R. F. (1996) NeuroReport 7, 30733076. 35. Heil, P. (1997) J. Neurophysiol. 77, 26162641. 36. Heil, P. & Irvine, D. R. F. (1998) Cereb. Cortex 8, 125141. 37. Heil, P. (2001) in Computational Models of the Auditory System, eds. Greenberg, S. & Slaney, M. (NATO Advanced Study Institute, IOS, Amsterdam), pp. 189203. 38. Heil, P. & Neubauer, H. (2001) J. Neurosci. 21, 74047415. 39. Heil, P. & Irvine, D. R. F. (1997) J. Neurophysiol. 78, 24382454. 40. Covey, E. & Casseday, J. H. (1991) J. Neurosci. 11, 34563470. 41. Sullivan, W. E. (1982) J. Neurophysiol. 48, 10331047. 42. Phillips, D. P. & Orman, S. S. (1984) J. Neurophysiol. 51, 147163. 43. Klug, A., Khan, A., Burger, R. B., Bauer E. E., Hurley, L. M., Yang, L., Grothe, B., Halvorsen, M. B. & Park, T. J. (2000) Hearing Res. 148, 107123. 44. Galazyuk, A. V. & Feng, A. S. (2001) J. Neurosci. 21, RC147. 45. Kros, C. J. (1996) in The Cochlea, eds. Dallos, P., Popper, A. N. & Fay, R. R. (Springer, New York), pp. 318385. 46. Raybould, N. P., Jagger, D. J. & Housley, G. D. (2001) J. Assoc. Res. Otolaryngol. 2, 362376. 47. Liberman, M. C. (1982) Science 216, 12391241. 48. Merchan-Perez, A. & Liberman, M. C. (1996) J. Comp. Neurol. 371, 208221. 49. Liberman, M. C., Dodds, L. W. & Pierce, S. (1990) J. Comp. Neurol. 301, 443460. 50. Ruel, J., Nouvian, R., Gervais dAldin, C., Pujol, R., Eybalin, M. & Puel, J.-L. (2001) Eur. J. Neurosci. 14, 977986. 51. Giraudi, D., Salvi, R., Henderson, D. & Hamernik, R. (1980) J. Acoust. Soc. Am. 68, 802806. 52. Irwin, R. J., Hinchcliffe, L. K. & Kemp, S. (1981) Audiology 20, 234243. 53. Beutner, D., Voets, T., Neher, E. & Moser, T. (2001) Neuron 29, 681690. 54. Dodge, F. A. & Rahamimoff, R. (1967) J. Physiol. (London) 193, 419432. 55. Augustine, G. A. & Charlton, M. P. (1986) J. Physiol. (London) 381, 619640. 56. Zucker, R. S. (1991) in Presynaptic Regulation of Neurotransmitter Release: A Handbook, eds. Feigenbaum, J. & Hanani, M. (Freund, Tel Aviv), pp. 153195. 57. Schneggenburger, R. & Neher, E. (2000) Nature 406, 889893. 58. Bollmann, J. H., Sakmann, B. & Borst, J. G. G. (2000) Science 289, 953957.

6156

www.pnas.org cgi doi 10.1073 pnas.1030017100

Heil and Neubauer