STATEMENT AS TO RIGHTS TO INVENTIONS MADE UNDER FEDERALLY SPONSORED RESEARCH AND DEVELOPMENT The U. S. Government has rights in this invention pursuant to Contract No.

DE-FG03-97ER82510 awarded by the Department of Energy.

BACKGROUND OF THE INVENTION The present invention relates generally to signal processing systems, and more particularly to processing the step-like output signals generated by non-ideal, nominally single-pole ("N-1P") devices responding to possibly time-varying, pulse-like input signals of finite duration, wherein the goal is to recover the integrated areas of the input signals.

The specific embodiments described relate to processing step-like signals generated by detector systems in response to absorbed radiation or particles and, more particularly, to digitally processing such step-like signals in high resolution, high rate gamma ray (y-ray) spectrometers with resistive feedback preamplifiers connected to large volume germanium detectors. The application of measuring the step-like output signals from y-ray detector preamplifiers to measure the y-rays'energies is just a specific example, and is described because this was the area in which the method was first developed.

The techniques that we have developed solve this problem generally, and therefore should not be construed as being limited to this specific application. Any detection system, for example, that produces output current signals that are integrated by charge sensitive preamplifiers could be treated by these techniques, whether the detected quantities are light pulses, x-rays, nuclear particles, chemical reactions, or otherwise. The techniques, in fact, is not limited to"detector systems"per se, but are, in fact, general purpose signal processing techniques which may be broadly applied, once understood.

The outputs from superconducting bolometers, for example, produce step-like signals that

are readily treated by the invention. The field of gamma spectroscopy, where 0.1% or less makes the difference between a bad and a good detector, however, provides particularly stringent tests of our techniques.

The term"step-like signal"also requires some discussion. The output of an ideal single-pole ("1P") device to an ideal impulse (delta) function input, is an infinitely fast rise time followed by an exponential decay whose time constant Td is characteristic of the pole. Viewed on a time scale short compared tord, this output will look like a pure step, while, when viewed on a time scale long compared to Td, it will look like a pulse. A real 1P device output, however, will have a finite risetime, Tr, whose duration will be determined both by the nature of the device and, particularly, by the duration of its real input signal. Provided that Tr is significantly shorter than Td, a real 1P device output signal, viewed on a time scale comparable to Td, will then show a risetime region, whose shape may be difficult to describe mathematically, followed, after a period comparable to ir, by an exponential decay with time constant Td. The output of a N-1P device will be similar, with additional distortions. We will refer to such signals, viewed on this time scale, as"step-like".

Gamma-ray (y-rav) detection requirements The detection and measurement of y-ray energies is a well-established discipline whose primary goal is to accurately determine both the number and energies of y-rays emitted from some target source. The requirements of good energy resolution and high count rate capability usually conflict, however, since count rates are enhanced by increasing detector volume, which increases output signal distortion and so degrades energy resolution. High count rates also degrade energy resolution directly due to practical problems in preamplifier design.

Description of the problems The field of y-ray detection is highly developed. A fairly comprehensive introduction to the state of the art may be found in the volume"Radiation Detection and Measurement, 2nd Ed."by Glenn F. Knoll [KNOLL-1989]. Below we note only the issues relevant to the present invention. In the first section, we discuss how pole/zero cancellation errors introduce a second pole, spoiling the preamplifier's single pole response. In the second section, we examine how the finite input signal duration, in this case due to charge collection, distorts the preamplifier's output from the ideal.

Pole/zero cancellation errors Fig. 1A shows a typical solid state y-ray spectrometer comprising a semiconductor detector diode 7 biased by a voltage supply 8 and connected to a preamplifier 10 comprising an amplifier 13 with a feedback capacitor C 15 and resistor R

17. As drawn, preamplifier 10 is a single pole circuit whose response to an impulse (delta function) input is A exp (-t/T2), where z2 = RC and A is the area under the impulse.

Because T2 is typically of order 1 ms, which is too long for the following circuits, a pole/zero (P/Z) network 20 cancels the pole at 1ITN and replaces it with a pole at 1/il, where T1 typically is 50 J. s. Gain stage 22 then amplifies and buffers the preamplifier's output signal for shaping amplifier 23 which feeds multichannel analyzer (MCA) 24.

If the time duration of the current pulse arising from the charge deposited in detector 7 by a y-ray absorption is very short compared tory, the output of stage 22 will be an exponentially decaying step whose amplitude is the pulse integral and proportional to the deposited charge. y-ray spectrometers are therefore designed to measure these step amplitudes to measure the charge deposited by the absorbed y-ray.

Other forms of radiation, including neutrons, alpha and beta particles, and x-rays behave similarly and their energies are measured the same way.

Commonly, however, both the input's finite duration and the pole-zero circuit's imperfections distort the preamplifier's response, destroying the proportionality between the output step's amplitude and the deposited charge and so degrading the system's energy resolution. Imperfections in P/Z network 20 arise from difficulties in precisely canceling the T2 component, leaving a small residual fraction, of order 1-2%, in the output signal. Fig. 1B shows a 5% residual T2 component for ease of viewing: an exponential decay signal 25 with time constant T2, input to the P/Z network 20, produces either output signal 27 or 29, depending upon whether the residual T2 term is positive or negative.

These T2 residuals are particularly bothersome at high counting rates, where each signal step rides upon a if background from all preceding steps. As these arrive randomly, the resulting baseline bias also fluctuates randomly in time, which the spectroscopy amplifier's baseline restoration circuit cannot track well. These terms, which may only be a few tenths of 1%, become a significant resolution degradation at 1 MeV where 0.05% energy resolution is desired.

Signal risetime fluctuations and ballistic deficit Fig. 2 shows a preamplifier 10 front end with a cross sectional view of the detector 7 of Fig. 1, for the common coaxial geometry, The dashed lines show electric field line within the detector body 30, which vary considerably with local geometry. Two factors cause charge collection time variations within the detector and thus risetime variations in the preamplifier's 10 signal output: 1) the difference between carrier velocities; and 2) the existence of different path lengths within the detector.

RAUDORPH-1982 describes these issues. These risetime variations produce ballistic deficit by two paths, one direct, one indirect. The direct effect is well understood, per

GOULDING-1988: the output filter's response varies with the time dependent shape of the charge arrival, being the convolution of the two. A trapezoidal filter greatly reduces this effect in the absence of exponential decay.

The indirect effect source of ballistic deficit is due to fluctuations in charge loss through the feedback resistor with differing risetime, as seen in Fig. 3A with two risetimes, 40 and 42, where Fig. 3B enlarges their peak regions. The slower risetime signal loses less charge and thus is larger once charge collection is complete. Even filters which ignore the charge collection region are still sensitive to this lost charge effect, and relatively small errors of this size can substantially degrade resolution. For a trapezoidal filter, the collection time difference shown Fig. 3B produces a 0.2% amplitude difference (2,000 eV at 1 MeV) which will degrade ideal 1.7 keV resolution to 2.6 keV. Ballistic deficit errors must therefore be reduced to less than 0.05% to obtain ideal spectrometer resolutions at 1 MeV.

Charge Trapping Losses Charge trapping also produces errors in y-ray energy measurements since trapped charges are lost to the measurement. The present invention does not seek to address this problem.

Generalizations It is important to note that the pole-zero cancellation errors described above do not arise from the preamplifier's use in y-ray spectroscopy, but are a generic problem in low noise, charge sensitive preamplifiers. Similarly, while the described risetime fluctuations described above were attributed to the geometry and construction of large volume Ge detectors, it is clear that such problems fundamentally arise from interactions between the finite charge collection time and the electrical characteristics of the preamplifier and not from the physics of the collection processes. Geometric heat flow variations in the photon absorbing mass produce similar effects in the superconducting bolometers mentioned earlier. Risetime issues are therefore a potentially general problem as well, and may need to be corrected for in other, non-y-ray, detectors whenever the highest measurement accuracies are required. The methods we describe offer just that capability. Further, the terminology"single-pole"or"multiple-pole" comes from the LaPlace Transform treatment of differential equations describing time variant phenomena. Any device which shows"single pole"behavior, for example, will display exponential time decay in response to an impulse input and, therefore, may be, for example, mechanical, thermal, chemical, or magnetic in nature in addition to the electronic case presented here. Our method can be directly applied to these devices as well, as will be apparent from the teachings herein.

Existing Correction Schemes The prior art deals with pole/zero errors in two ways: first, by canceling T2 as accurately as possible; and, second, with baseline restoration schemes which try to track the shaping amplifier's"no signal"output as closely as possible, an approach which degrades as rates becomes high. We have not found any approaches which measure and/or correct for the effect directly.

Over the years, various heuristic schemes have been developed which attempt to compensate for ballistic deficit. RADEKA-1982 introduced trapezoidal filtering and developed a time-variant implementation, using a gated filter following a semi-Gaussian shaper, that provided significant resolution improvements. WHITE-1988 proposed a different gated integration approach using a series switch to excise the charge collection region out of the preamplifier signal entirely. The final circuit was complex and had enhanced deadtime problems. GOULDING-1988, RAUDORF-1982, and SIMPSON-1990 disclose schemes that depend on directly measuring the signal's risetime tr and correcting the energy filter output by a term like trn. These approaches are complex to implement and require precise expert adjustments to operate. The underlying assumptions are not particularly valid and improvements in energy resolution have been modest in practice.

HINSHAW-1991 and KUMAZAWA-1998 describe attempts to correct for ballistic deficit by capturing peak amplitudes from two filters which respond to the ballistic deficit differently, one an energy measuring filter and one a differentiating (or bipolar shaping) filter. Typically a significant fraction of their difference in peak heights is added to the energy filter's peak to correct it.

Related Art There is some related art wherein the details of the shapes of the preamplifier output signals are sampled digitally and used either to distinguish between different types of particles absorbed in the detector (e. g., MILLER-1994) or to distinguish between single and multiple interaction events in large germanium detectors. See, for example, TAKAHASHI-1994 and AALSETH-1998.

WARBURTON-1997, WARBURTON-1998, and WARBURTON-1999 describe methods for implementing digital filtering and x-ray spectroscopy. While they do not address the issues under consideration, some of their filtering techniques will be employed in the present invention and are referenced in the specification.

SUMMARY OF THE INVENTION The present invention provides techniques for measuring a step-like output signal from a nominally single-pole (N-1P) device in response to a pulse-like input signal to determine the integrated area of said input signal. The invention addresses the possibility that the device deviates from ideality due to the presence of additional poles, zeros and/or a DC offset and that the input pulses may have finite time durations and variable amplitudes. In a specific example, the measurement determines collected detector charges from step-like preamplifier output signals in the presence of both risetime fluctuations and imprecise pole/zero cancellations.

In brief, the present invention contemplates processing the N-1P device's (e. g., preamplifier's) output signal using a set of one or more shaping filters. Where plural filters are used, they typically have different time constants. A set of samples of the outputs of this filter set is captured in such a manner that the multiple sample values bear prescribed time relationships to one other. We refer to this set of sample values as a correlated multiple output sample (or"cMOS"for short). The relationships among the individual sample values may be determined by the times of their capture, by delay elements inserted in the signal paths, or some combination of the two. Further, the different individual sample values can be obtained from associated different filters, or plural sample values can be obtained from the same filter, but captured at different times or with different delays. The term cMOS is intended to cover these multiple possibilities and is discussed further in § 5.1 below.

The input pulse's integrated area is determined by capturing a cMOS in response to detecting a step-like signal (sometimes called an event). A weighted sum of the individual sample values in the cMOS (sometimes referred to as cMOS values) is then formed as a measure of the input pulse's integrated area (e. g., total charge). The weighting factors can be computed directly from information about the N-1P device's decay constants, the filter set, and the prescribed time intervals in the cMOS.

While embodiments of the invention use triangular and trapezoidal filters, the invention does not require specific filter shapes. Further, the invention does not require that the filter's capture times be precisely located relative to the step-like signal's leading edge. Rather, the method derives its accuracy from repeatably reproducing the set of prescribed time relationships between the sample values in the cMOS. The underlying capture method is therefore time-based rather than amplitude-based, which differentiates the invention from prior art methods which capture filter samples based on their maximum amplitudes.

While the preferred implementation uses digital signal processing, implementations using solely analog processing or hybrid approaches are also feasible.

Embodiments of the invention require neither a measurement of the step-like signal's

risetime nor more precise information about its arrival time than is found by the pileup inspection circuits of commonly available commercial shaping amplifiers.

When using a single cMOS, however, the measurement's noise is increased by the number N of additional filter samples required to compensate for the N-1P device's non-ideal terms. For those cases where this increased noise is significant, we also show how N may be reduced by creating a parameterized model of the non-ideal terms; making baseline measurements to determine the parameters; and then using them to correct area (e. g., charge) measurements made using simpler sets of filters which have less noise. In some implementations, these baseline measurements are cMOS values captured at times when the filters are not processing the step-like signals. In the detector- preamplifier case described in detail, the resultant implementation achieves the low electronic noise levels of conventional trapezoidal filters while also eliminating resolution loss due to both ballistic deficit and high count rates.

A further understanding of the nature and advantages of the present invention may be realized by reference to the remaining portions of the specification and the drawings.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1A shows a schematic of a semiconductor y-ray detector front end, including a resistive feedback integrator, pole/zero network and output gain stage; Fig. 1B shows voltages generated by the preamplifier portion of the circuit in Fig. 1A when the pole/zero circuit is not ideally adjusted; Fig. 2 expands the detector structure of Fig. lA to demonstrate the geometric and other factors which contribute to signal risetime variations; Fig. 3A shows the variation in risetime between pure electron and pure hole collection for a simple collection path in the detector structure of Fig. 2; Fig. 3B shows an expanded view of the signal peaks in Fig. 3A; Fig. 4 shows a step-like feature in a general preamplifier output, broken into its component signals, and establishes a nomenclature system for describing these signals mathematically; Fig. 5A and 5B show how to compute the appropriate intervals between capture times of the multiple filters; Fig. 6A and 6B show a pair of purely digital circuits which implement the described method; Fig. 7 shows a purely analog circuit which implements the described method; and

Fig. 8 shows a hybrid analog/digital circuit which implements the described method.

DESCRIPTION OF SPECIFIC EMBODIMENTS 1. Introduction The invention has three basic steps: applying a set of filters to the preamplifier output; capturing a correlated multiple output sample ("cMOS") from the filter set in response to a detected event; and forming a weighted sum of these sample values to accurately recover the energy of the detected event. The cMOS is a set of samples bearing a prescribed set of time relationships between one another. The appropriate time relationships are achieved by setting the times of capture, inserting delays into the signal paths, or a combination thereof, and are independent of the specific risetime shape of any particular event. Our overall goal, therefore, is to describe how to select an appropriate set of filters; explain what it means to capture a cMOS and how to achieve it; and show how to compute the weighting functions.

The theory underlying the invention method is based on modeling the preamplifier's (or N-1P device's) step-like output response (the"output") to the input charge (or impulse: the"input") generated by the event as the sum of a small number of analytical functions which are represented by the same number of amplitudes. In a specific embodiment, five amplitudes are used to describe: the Tel and T2 exponential components prior to the step-like signal, the ideal amplitude of the step itself, an error term due to the ballistic deficit, and a DC level. We first reduce the number of amplitudes to four by making differential filter measurements which are DC level insensitive. To find these four amplitudes, we make four independent measurements (from three filters) in the step's vicinity, express each measurement as a linear combination of the four amplitudes, and solve the resultant set of four linear equations to recover the ideal step amplitude, which is proportional to the input charge (or impulse area). Obtaining a set of equations which is well conditioned and insensitive to preamplifier noise, ballistic deficit, and T2 amplitude is then primarily a matter of selecting an appropriate set of filters and time correlated capture points, which then define the cMOS.

Step amplitudes obtained using the four-filter measurement method have almost a factor of 3 higher electronic noise than simple trapezoidal filters. In cases where this is important, we show how simpler filters can be corrected for the T2 and DC components by developing a parameterized model of these terms and making one or more baseline measurements to determine their values. When multiple measurements are made, the parameters must be compensated for their evolution in time, creating a time- compensated model.

The description below is organized as follows In § 2, using a static, captured step-like signal, we locate several filters in the vicinity of its leading edge and mathematically describe their outputs in terms of the base amplitudes. In § 3, we use these results to set up and solve the set of linear equations to obtain the input charge.

This solution has the form of a weighted sum of the filter values, showing the genesis of the weighting coefficients. In § 4, we evaluate the method for two filter sets which compensate simultaneously for inexact pole/zero cancellation and ballistic deficit. In § 5, we introduce the concept of the cMOS to show how to implement the invention in real time processing situations. In § 6, we will describe various circuits which implement this method: a purely digital pair, a purely analog circuit and a hybrid analog-digital circuit.

In § 7, we analyze our method's electronic noise. Finally, in § 8, we show how a simpler, three filter measurement can be used to achieve similar accuracy with lower noise, provided time recorded baseline measurements are made to correct for the 12 and DC preamplifier components 2. Mathematical description of generalized filter outputs 2.1. Decav components of the step-like signal Fig. 4 shows a typical preamplifier step-like output signal as it might look if it were digitized by an analog to digital converter (ADC) and saved in a computer memory. The shown time scale is arbitrary, with zero set a little before the onset of charge collection for convenience. To simplify the image, only a few members of the set of discrete values have been identified explicitly: subscripted capital Q's denote values at specific instants since the signal really represents the charge integrated on feedback capacitor 15 by amplifier 13. The T 1 component has three regions: an exponential decay region 50 prior to the step (i. e., times less than 0), a charge collection region 52, and an exponential decay region 53 following the step (i. e., times greater than 1). The curve 55, showing how curve 50 would have continued to decay had the step not arrived, represents the continued time decay of all previous steps, that is, the preamplifier's first pole residual response to those previous steps. The signal's T2 component has the equivalent four regions 57,58,60, and 62. Finally, there is the preamplifier's DC offset value Bo 65, which is constant. The Tl and T2 charges injected by the y-ray pulse decay independently, since the preamplifier is a linear device. The total collected charge QgT generating this event thus divides into Q1gT = (1-ß) QgT for il and Q2gT (3QgT for i2, (3 being the division ratio.

It is important to note that, in the following, we will be describing a signal processing technique which relies on the mathematical form of the signals and not on their physical source. While we develop the method in the context of preamplifier signals from y-ray detector preamplifiers, the method itself can be usefully applied to any signals

which have the appropriate form. Similar outputs would generally be obtained from nominally single pole devices (N-1P devices) in response to impulse inputs of extended duration and variable amplitude. In the present case these signals result from a single pole electronic circuit responding to an input pulse from a detector, in other cases the detector itself may produce signals of the appropriate form. The temperature response of a superconducting bolometer, for example, shows this behavior [STAHLE-1999] and could be treated by the method.

2.2. Running-sum filter output If r values are sampled at uniform intervals At from an exponential decay with time constant #j, then the ith value Qj, i is given by: Qj, i=Qj,0 exp(-i #t/#j)=Qj,0 bji, for (0 i r-1) (1) where bj = exp (-At/Tj) is the decay per At. Therefore, a kth digital filter, defined by weighting constants {wi k, 1 i r-1}, acting on the values in Eqn. 1 gives: where the * marks this as a general filter result. This shows that the filter output is linear in the amplitude Qj, o, multiplied A*j kr, which is a constant for a particular filter and decay constant T1. For a simple running-sum filter (see WARBURTON-1999), w equals unity, so Eqn. 3 also defines the division factor rj, which is equal to the sum to infinity.

2.3. General trapezoidal filter output with risetime in its gap Next, consider a running-sum cv+ of m samples starting at time 1 over the signal in Fig. 4. Decomposing the signal into 5 components with amplitudes {Qg13, Q13, Qg23, Q23, and Bo}, Eqns. 2 and 3 give: +=QgnA)+Q,3A)+QgA+(A+mBo(4) where the index j in Aj equals 1 or 2 for T 1 and T2. The equivalent sum (TE-, starting at time-4, which only sees amplitudes {Q13, Q23, and Bo} is then: Q20A2,m+mB0.(5)##=Q10A1,m+ The difference (6E+-Cz) is a trapezoidal filter se of risetime m and gap g. Since Q1o = bi--SQi3andQ20=b2--8 Q23: ##=##+-##=Qg13A1,m + Q13α(1,m,m-m-g) + Qg23A2, m + Q23a (2m~m~g), (6)

where a (j, r, s) = Aj, r (l-bjs) and the DC components cancel. In Eqn. 6, we stress that Q13 represents, at time instant 3, the preamplifier's dominant pole (rj) response to all preceding step-like signal events; Q23 similarly represents the secondary pole (X2); while Qgl3 and Qq23 are their amplitude responses to the present event. Eqn. 6 takes its simple form because the signal rise occurs entirely within the filter gap region g, so only decaying exponential terms are measured outside. Thus cry is relatively insensitive to the precise location of the step within the gap region because its weighting constants are negligible (zero, in fact) there.

2.4. Triangula filter over the charge collection region To explicitly probe the step's risetime region, we will use a triangular filter (gap = 0) cFg = (ag+-ag_) of risetime g, whose subtractive leg su sites precisely within the gap of sz. (This relationship will also be easy to achieve when we implement a digital real time system.) We will show that the precise location of the step within the gap is not important. Then, as above: where the qjg, are the sampled charge values at each interval in the subtracted running- sum region, each comprising the new charge increment collected in that interval plus the exponentially decaying remnants of the charges collected in the previous intervals.

Since each charge increment decays to zero independently, while their sum equals the total collected charge, we can write Splitting the infinite summation, however, also gives: so that The correction charge Qc defined in Eqn. 10 is the difference between the total T1 collected charge and the amount remaining at the end of gap g. Qc will typically be of order 2% of QgT. As a small correction term, Qc will be relatively uncorrelated to Qgl3,

so the set of linear equations we solve for Qgl3 and Qc be well conditioned under inversion.

Similarly, for the T2 component: where Qc2 will be smaller than Qc by approximately the ratio Tl/T2. Eqns. 10 and 11 allow us to express vu entirely in terms of QgT, the total charge, and values from outside of the charge collection region, where the signals behave strictly as decaying exponentials. Equivalent integral expressions may similarly be derived in the continuous signal case or for filters with other weighting constants.

2.5. Eliminatini When the gap region is much smaller than both il and T2, Qc2 becomes proportional to Qc to a high degree of accuracy. This may be seen by defining Qg ; as the total y-ray charge collected at time interval i in the collection region. Since each Qgi decays independently, we can write for the two components: while, by definition, Thus, for Qe: Expanding b1g-i to good accuracy as (l-(g-i) tt/T1) gives (1-b1g-i) equal to (g-i) #t/#1 so: Similarly: where p2 is equal to At/T2. Comparing Eqns. 15 and 16:

Qc2 is therefore quite small, being about 0.1% of Qc or 0.002% of QgT. We can therefore express any of our filters in terms of the four quantities QgT, Qc, Qgl3 and Qg23 using: 3. Solving for collected charge QgT from the captured waveform 3.1. Filter expressions in terms of the base amplitudes § 2.5 shows how all difference filters in the vicinity of the step's leading edge can be expressed as linear equations in the four amplitudes QgT, Qc, Qg13 and Qg23 (non-difference filters will also require Bo). Therefore we must make four (or five) independent filter measurements to solve for QgT. While a wide variety of filters could be used, certain sets will be better conditioned than others and will also require less measurement precision to accommodate step risetime variations. Research and experimentation have shown the following set to function well: I) the trapezoidal "energy"filter of risetime m from Eqn. 6; 2) the triangular"risetime probe"filter of Eqn.

7; and, 3) a pair of triangular"T T and T2 background"filters of risetime m/2, situated at times-4 and +1 in Fig. 4, respectively. It is important to notice that, in this selection, only the risetime probe filter processes the step's risetime, which lies in the energy filter's gap and is entirely outside the range of the two other triangular filters. This placement, where only the risetime probe filter has non-negligible weighting constants in the risetime region, assures that their outputs, and hence QgT estimate, will be insensitive to the precise location of the signal step within the gap, which, in real time implementations, will remove any need for great timing precision. That the filters share some common starting and stopping points is primarily a matter of mathematical and implementational convenience, but is not fundamental to the technique.

Substituting Eqns. 18 into Eqns. 4 through 7 gives four linear equations for cyz+, Cy£, and vg in terms of QgT, Qc, Q13 and Q23 :

We divided On by rl so that the coefficient of Qc will approximately equal unity for inversion stability.

Using the same methods for the outputs of three filters of length m/2 (two triangular filters C743 and #21 whose running-sum first points are: o43 = [time +3: time +1] and cy21 = [time-2: time-4] and a trapezoidal filter cs32 = [time +2: time-2]) gives: These filters also exclude the step's rising edge for location insensitivity.

While this is not a requirement of our invention, it significantly simplifies both the coefficient derivations and trigger system timing requirements.

3.2. The four-filter solution for QgT Collecting results for the four filters listed at the start of § 3.1, we have:

as four linear equations in four unknowns. The unknown parameters are of three types: 1) (Pi) the unknown charge QgT (the area of the input impulse in the general case); 2) (P2) charges describing the step's finite arrival time (Qg) and the residual charges in the il and T2 decay modes (Q13 and Q23) from previous events; and 3) (P3) a set of fixed parameters {ri, T2, P, m and g} which describe the preamplifier's (or N-1P device's) transfer function and the applied filters. Since the coefficients in []'s are functions only of the P3 parameters, they are constants which need be computed only once for a given preamplifier and filter set.

Eqns. 21 can be solved by several methods, including casting them into matrix notation and then inverting the coefficient matrix. Thus:

becomes J#Q,= <BR> (23a)<BR> <BR> <BR> <BR> <BR> <BR> so =J-1##,Q <BR> (23b) As a solution for QgT, Eqn. 23c has the advertized form: a weighted sum of the four filter outputs. The weighting coefficients Jij-1 are computed only once for a particular set of filters and P3 parameters {#1, T2, P, m and g} and then multiplied by the four filter outputs ##, 6g/r1,043 and 621 to obtain the total pulse charge QgT (and hence the energy of the absorbed y-ray) for each captured event. We note that if more filters are measured than there are parameters Q, the matrix can be inverted using least squares methods to obtain Eqn. 23b.

4. Evaluation for two filter sets 4.1. A four-filter set with typical parameters We now evaluate a typical case of the filter set leading to Eqns. 22 and 23: we set m equal to 4 µs, g to 1 µs and assume typical values of T1 = 50 ps and T2 = 1 ms with a division ratio ß=0. 04. In this case, for 40 MHz sampling, m = 160, g = 40 and we can compute and <BR> <BR> #g-1.593231E-2#43-1.549595E-2#21(24c)QgT=6.309486E-3##-4.811 789E-4

6g is typically of order 0.256z, so the 6g term is typically less than 2% of QgT. In spite of its small size, however, 6g carries the burden of the ballistic deficit correction, since it is the only filter that probes the step's risetime directly.

The coefficients in Eqn. 24c also depend upon the value ß. When the pole-zero is perfectly balanced, (ß = 0) Eqn. 24c becomes: QgT = 6.309486E-3c-4.811789E-Cg-l.59323lE-2c-1.549595E-2) where u.'s coefficient has changed by 4%, cyy_'s by 0.04%, and the C743's and #21's by about 0.008%. The need for P accuracy is therefore a function of how accurately one wishes to correct for risetime fluctuations and pole-zero inaccuracies. The J-1 weighting coefficients are found to vary linearly with P, which allows ß to be determined experimentally if it is not precisely known. To do this, first select a set of P values bracketing its expected value and compute Eqns. 24 to produce a set of weights V"/ (the coefficients in Eqn. 24C) for each ß value. Next collect data for each weight set and measure the standard deviation 6QgT (ß) in QgT (i. e., the energy resolution) which is obtained. Then, by plotting or interpolating the 6QgT (ß) results versus ß, select a P value which minimizes #QgT(ß). If the preamplifier is stable, (3 should only need to be determined once.

We tested the accuracy of Eqn. 24c's in the presence of both risetime variations and a stable-12 decay term (equal to the T1 step amplitude) using a spreadsheet to compute the filter responses for modeled waveforms. Table 1 presents these results.

As may be seen, the Eqn. 24c is insensitive to both risetime fluctuations and T2 components at the few ppm level and is three orders of magnitude better than a simple trapezoidal filter.

Table 1: Modeled charge QgT for 5 filters in the presence of varying risetimes and #2 amplitudes for the case of ß equals 0.02.

Risetime = 450 ns Risetime = 950 ns Model <-c2> = 0 <#2>=1 <#2>=0 <#2> = 1 Triangular 0.861966 0.859787 0.831149 0.827170 (100.00%) (99.75%) (96.43%) (95.96%) Trapezoidal 0.867397 0.862421 0.863901 0.858928 (100.00%) (99.43%) (99.59%) (99.02%)

Section 4.1 0.999997 0.999997 1.000001 1.000001 (Eqn. 24c) (100.000%) (100.000%) (100.000%) (100.000%) Section 4.2 1.000119 1.000113 1.000028 1.000022 (Eqn. 33) (100.000%) (99.999%) (99. 991%) (99.990%) 4.2. A three-filter simplification which still has good performance For cases which do not demand the precision of Eqn. 24c, or have small 12 components, we show a simpler, three-filter set which still has good performance. We use trapezoidal/triangular filters for their insensitivity to the"DC-like"slow variation in the T2 component: namely the same"energy"and"risetime probe"filters as in § 4.1 and a second trapezoidal filter 632 (Eqn. 20c), aligned to exclude the risetime region, to estimate Q13. Thus: In Eqn. 26 we reversed the sign of 6r in order to create a filter that is differential on both sides of the step's rising edge and we weighted 632 by W to determine the value that gives the best results. We find the lowest noise is for W = 1, so we can repeat Eqns. 24a and 24b to obtain: QgT = 9. 338925E-3 a-5. OOOOOOE-4 6g + 3. 106015E-2 a32.

Table 1 shows that this result is sensitive to the 12 component at the few ppm level and to risetime variations at the 100 ppm level. For many practical cases, this is sufficiently accurate. Assuming a random distribution of risetimes, for example, this would add less than 100 eV in quadrature to the energy resolution at 1 MeV, which would be scarcely perceptible.

5. Correlated multiple output sampling in real time implementations 5.1. Defining"correlated multiple output sampling" The filter weighting Eqns. 25 and 27 derivations were developed as if the signals had been captured into memory, so the running-sums could simply be taken over the desired sample values. This implementation is feasible at lower data rates, but in the

preferred implementation, as taught by WARBURTON-1997, a gate array is used to realize the difference of running-sum filters, which run in real time, each outputting a new value its own fixed number of clock cycles after a new ADC sample is supplied. A priori, then, there may be no single instant when a set of filter values can be captured which will reproduce the inter-filter time relationships required by Eqns. 25 and 27. Instead, in general, the outputs of the different filters will have to be captured at different times and these times will have to be carefully correlated if Eqns. 25 and 27 are to produce accurate values of QgT. Because multiple filter values must be captured, we call the resultant set of captured values a"correlated multiple output sample"or"cMOS"and refer to these values either as"cMOS values"or as"individual sample values in the cMOS".

We also observe that the number of cMOS values required by Eqns. 25 and 27 and number of realized filters needed to produce them need not be the same. For example, in Eqn. 25, the 621 and 643 filter values both come from triangular filters of length m/2 with their starting points separated by the interval m+g. Thus two outputs from the same realized filter can be captured to supply the two required values.

While we have noted that the filter ranges do not have to be precisely positioned relative to the step's rising edge, the accuracy of their positions relative to one other in the cMOS capture determines the accuracy that the method can attain. This is because the running-sum ranges, which translate directly into filter lengths and time intervals between starting locations, are encoded into the J-1 weighting coefficients through terms like a (j, r, s) = Aj, r (l-bjs) (see, e. g., Eqn. 6), where r is the filter length and s the interval between the sum start points. For simplicity, we aligned the various filters so their starting points coincided, but other cases are trivially treated by scaling Eqns. 4-7 by terms like bit, where t is the offset between starting locations. Therefore, there is an intimate relationship between the weighting coefficients and the filter ranges which, if violated, will introduce inaccuracies. The J-I's are computed for a particular set of filter sum ranges. If another set of ranges is captured, then expected cancellations between filter terms will not occur exactly and the differences will appear as errors in the detected charge. In digital implementations, the fact that the signal samples are discrete and enumerated makes is relatively easy to maintain accurate time relationships by measuring time in"ticks"of the digitization clock and then simply counting out the required time intervals.

The requirement of appropriate time correlations between the cMOS values clearly differentiates the present method from prior art as represented by the approaches of HINSHAW-1991 and KUMAZAWA-1988. While two filter outputs were captured and weighted in those methods, the capturing steps used peak detecting means.

The methods are therefore amplitude based rather than time based, as in the present method, and the captured values carry different information about the step shapes.

Further, the method of capturing based on peak detection assures that there will be no fixed time relationship between the captured samples since the two filters'times of maximum amplitude depend differently on step shape, which will vary from step to step.

In light of the above, it is clear that there will be a wide variety of possible approaches to capturing filter cMOS values with the prescribed time relationships. The outputs of single filters can be captured multiple times. Filter outputs can be captured as they appear from the gate array or can be delayed for convenience. We will show several examples in § 6. Which method is best will depend upon implementation details and will vary from case to case. We therefore intend the phrase"cMOS" (for"correlated multiple output sample") to cover the multiplicity of possibilities for capturing, in a real-time implementation, filter values whose time relationships are correct for the set Of J-1 weighting coefficients that will be applied to them to recover the input charge from the detected event. When the cMOS capture is made in response to the detection of a step in the preamplifier's output signal, with the intention of using it to determine the charge deposited in the detector (in the general case the impulse area) and hence the gamma- ray's energy, we will often refer to it as an"energy cMOS capture" (or"charge cMOS capture"or"area cMOS capture"as appropriate).

5.2. Timing considerations Filters, in real time implementations, require finite times to compute. For example, consider a trapezoidal filter of length (risetime) m and gap g over a set of values {qi}, where qn is the last sample value ("end sample") to enter the filter: If implemented in gate arrays according to the method of WARBURTON-1997, the output a (n) will appear 3 filter clock cycles after sample qn is presented. This is its propagation delay. For a 4 ps peaking time (filter clock equals 200 ns), the propagation delay is 600 ns. These delays must be included in the design to assure that the cMOS capture constraints are satisfied.

Figs. 5A and 5B show how to compute sampling times for the Eqn. 24c filter set. The calculation has two steps: finding the time delays between their end samples of the different filters and then correcting for their different propagation delays.

For the first step we use Fig. 5A, which shows the filter regions relative to the preamplifier output step 72. os 74,643 76,621 77, and Og 79 were all described earlier.

6t 81 is a fast channel timing filter, as described by WARBURTON-1997, which is used to detect step-like signals in the preamplifier's output. The timing filter's output triggers a digital discriminator when it is located approximately as shown with respect to the leading edge of the signal step 72. We can now measure end sample delays relative to 6t

81 as shown, with the filters 6z 74,643 76, and Cg 79 appearing at times Atz, At43, and Atz after at 81 and 621 77 appearing At21 earlier. From Fig. 5A, if at has a peaking time of T, then At equals (m+g-2T).

It is important to note that, while the firing of #t may jitter somewhat depending upon the step's risetime shape, this only causes the step location to jitter within the gap g. The relative timings of all the filters to 6t, and thus to each other, are fixed by construction. Thus, provided #g is wide enough to catch the step for any reasonable discriminator triggering, the filter will work as designed. This relative insensitivity to the timing accuracy with which the step-like signals are detected is a significant advantage of the present invention.

Fig. 5B shows the second step: correcting for propagation delays. We note that these delays may exceed a microsecond in long time constant filters and typically increase with filter length. Fig. 5B shows propagation delays with respect to the ADC sample time ts 83 for a 40 MHz implementation with g = 1 µs and m = 4 µs, so 8tt = 75 ns, §tg = bt21 = bt43 = 300 ns, and #t# = 1,200 ns. Combining Figs. 5A and 5B, we should sample the output of a given filter ax, at =+5t,-5t, (29) following the discriminator firing on 6t's output to assure the cMOS temporal relationships required to extract QgT. If tx is positive, then a counter can be used to measure tx and capture the filter value. If tx is negative, however, then its propagation delay btx will have to be increased by at least |tx| to meet the demands of causality (capturing must follow step detection in time). In digital implementations this is a simple operation using a FIFO memory. In our present example, no additional delay is required since a typical timing filter peaking time is 100 ns, so At21 =-200 ns and t21 = At2l + set21 + btt =-200 + 300-75 = 25 ns, (30) which is positive.

6. Circuits implementing the cMOS method 6.1. Fully digital filter implementations Two general approaches are possible: 1) asynchronously capturing values as they appear from the sub-filters, per the discussion of Sect. 5.2, and 2) extending the filters'propagation delays so that they can be synchronously captured. In the latter case all pileup inspection is carried out before the capture signal is given, while in the former case the values are captured as they appear and the pileup inspection is used to decide whether anything should be done with them afterwards. We will demonstrate one implementation of each to show the principles.

6.1.1. Asynchronous capture implementation Fig. 6A shows a preferred implementation of Eqn. 24c. The digital processor 90 comprises an ADC 92, a real-time digital processing unit (RT-DPU) 93, and a digital signal processor (DSP) 95. Per the method taught by WARBURTON-1997, the RT-DPU processes data at the ADC output rate while the DSP processes data at the event rate (impulse signal rate). The RT-DPU produces values 6z, 643, 621, and cFg, while the DSP multiplies them by the coefficients J-I to obtain QgT.

The RT-DPU 93 also has two parts, a fast channel 97 and a number of slow channels 98. The fast channel contains the fast timing filter at 100, a digital discriminator 102, some pileup test and timer logic 103, a clock 104, and an output register 105. Each slow channel comprises a triangular or trapezoidal digital filter and an output register with either a digital delay line (if Eqn. 29 is negative) or a timer (if Eqn.

29 is positive). Notice that 643 and 621, both being of length m/2, are being implemented using a single triangular filter C74321 107.

In operation, at 100 processes ADC 92 data until a step-like signal arrives.

Discriminator 102 detects this step and signals the pileup inspector 103, which strobes a trigger line 106 and initiates pileup inspection. If the step's arrival time is important, the output of clock 104 can be captured to output register 105 by the same strobe signal 106.

Filter 64321's 107 output, delayed by t21 using 21-delay 108 if required by Eqn. 29 causality, is captured immediately in the 21-register 109 by this trigger strobe, which also starts the 43-timer 110, E-timer 111, and g-timer 112 to time and t I shortest, the g-timer 112 times out first, capturing filter cy.'s 115 output in the g-register 116. Similarly, the 43-timer 110 and E-timer 111 time out at t43 and tE to capture filter (74321 107 and 6 117 outputs in the 43-register 118 and E-register 119. If this step is not piled up, then all four filter values are ready to be read into the DSP 95 via data bus 121.

The pileup inspector can be implemented as taught by WARBURTON-1999 relative to the length of the c77 filter and will not be discussed further.

As per the cMOS discussion in § 5.1, the invention method does specifically require that the number of filters equal the number of filter values to be captured. Here we have captured 64321 twice.

6.1.2. Synchronous capture implementation Fig. 6B shows an implementation of Eqn. 27 wherein the three filter values are captured just as the value emerges from the slowest (6E) filter. The primary physical difference between this implementation and that of Fig. 6A is that the timers 108,109, and 110 have been eliminated and a second delay (g-delay 125) has been added to compensate for the differences between t and tn and t32 from Eqn. 29. All the remaining parts are the same and carry the same reference numbers as is Fig. 6A. Notice that the trigger line 106 now connects directly to the output registers, rather than through

timers as before. For this topology, the functional operation of the pileup test and timer #1 103 is essentially identical to that presented by WARBURTON-1999.

A disadvantage of this circuit, compared to Fig. 6A, is that relatively long delays may be required, which may consume excessive gate array resources. Recent programmable gates, however, are appearing with blocks of dedicated memory, so this restriction is becoming less significant. One significant advantage of this design is that, because of the synchronous capture, the pileup inspection logic design is much simpler.

6.2. Fully analog implementation While our preferred implementation is fully digital, the method is not inherently digital per se. In certain cases, particularly for very fast step risetimes and very short filter time constants, or to allow a high degree of integration, it may become preferable to use an analog implementation. Fig. 7 shows an analog version of Fig. 6A.

The topologies are identical up to the point where sample and hold circuits replace digital registers to capture the filter values, so the operation of the circuit up to this point will be clear to one skilled in the art, given the teaching presented herein. The derivation of the J matrix, of course, will have to proceed using weighting functions appropriate for the analog filters used. Then analog multipliers (160, etc.) and an analog adder 185 are used to implement the J-1 coefficients and addition of Eqn. 24C. We note that, for an application with fixed filters, the multiply-and-add function is just an op-amp circuit with appropriate resistor inputs to its summing node. If the pileup tester 143 deems the value good, the gate and shape circuit 145 passes out the adder 185 output in a form suitable for multichannel analysis.

6.3. Hybrid implementations Shaping filter characteristics which are difficult to implement digitally may be required in some situations where the other benefits of digital processing are still desired. In these situations hybrid implementations may be considered. Fig. 8 shows one such implementation 190. The topology is nearly identical to that of Fig. 6A. The major differences are: 1) that the fast channel 191 is implemented in analog circuitry identical to that of Fig. 7 (although the digital circuit of Fig. 6A could as easily be used); and 2) that, in the slow channel section 192, the digital filters nu4321 107, z 117, and su 115 have been replaced by analog filters cl4321 193,6E 194, and cg 195 followed by ADCs 196, 197, and 198. The time behavior of this circuit is the same as was described for the asynchronous digital circuit in Fig. 6A.

7. Noise considerations The Eqns. 24C and 27 filters have higher noise than simple trapezoidal filters since they weight the sampled set of data points with values larger than unity. For example, by breaking the 2m+g filtered region of Eqn. 27 into its 4 major sub-regions of length m/2 (ignoring the gap g for this argument) and calling running-sums over these regions cyl through v4, we can recompose Eqn. 27 as: QgT =-9.338925E-3 C4 + 2.172123E-2 ou-2.172123E-2 c2 + 9.338925E-3 a,.

If the noise standard deviation per sample is a, then the total noise sig (QgT) will be which can be compared to a trapezoidal filter of the same length: Thus the Eqn. 27 filter will have approximately 2.7 times as much electronic noise as a simple trapezoidal filter. The result for the Eqn. 24c filter is similar. For a 4 ps filter, whose electronic noise can be as low as 150 eV, this effect is significant at energies like 100 keV (Fano noise 415 eV) but not at 1 MeV (Fano noise 1,300 eV). And, as peaking times shorten to obtain higher count rates, the problem will worsen at higher energies.

We therefore consider how to obtain the benefits of correcting for both ballistic deficit and pole-zero errors without paying the additional filter noise price.

8. Treating non-ideal terms with baseline measurements Our approach is to extend the baseline averaging concept of WARBURTON-1997 to treat time varying terms as well. In the particular case at hand, we would like a way to measure Q23 using baseline measurements in spite of the fact that it both exhibits exponential decay and is increased each time an event occurs.

8.1. Derivation of an example filter We begin by developing a simple cMOS filter, composed of three running averages, of lengths m, g, and m, which we make contiguous for the practical reasons of simplifying the math and minimizing deadtime, but not for any fundamental reason. As in previous sections, the length m filters will primarily serve to determine QgT, while the length g filter will be used to eliminate ballistic deficit effects by probing the signal step's risetime region. Starting with Eqns. 19A, 19B, and 19D, these filters are: E+=QgTAn, QcAm+Ql3Alm +Q23A2, m + mB0 (34a)

=+Q.3'A,+Q'A,g+gB<BR> <BR> <BR> <BR> d"l-=QnbA)+Q23bA+mB where <BR> <BR> <BR> <BR> #m = ß)A1,m+- =A1,m+FA2,m,#=r1+Fr2#m <BR> (34d) In Eqns. 34 we now have 4 classes Pi of parameters: Classes 1 and 2 (P1, P2), as before, which include QgT, Qc, and Q13; Class 3 (P3), Q23 and Bo, which describe the preamplifier's"non-ideal"behavior, Q23 being the amplitude of its second pole's response to any previous steps and Bo being the DC offset; and Class 4 (P4): all the remaining fixed parameters (e. g., il, T2, (3, m, g) which describe both the filters'and the preamplifier's transfer functions. We will treat QgT, Qc, and Q13 as unknowns to be recovered from a particular cMOS filter measurement, with the intention of independently determining Q23 and Bo from baseline measurements. Eqns. 34 are simple enough to be solved by direct elimination of unknowns to get: The constants K, L, and M depend only on Class 4 parameters, which are assumed to be time independent for a particular detector-preamplifier pair and set of filter lengths. Eqn.

35A is therefore a transform function (the first bracketed term) between the Eqn. 34 cMOS filter values o+, cyy-, and Og and the desired detector charge QgT, minus an error term [LQgT + MB0]#m-1, that is the transform's response to the non-ideal (NI) preamplifier terms Q23 and Bo. Using Eqn. 35a thus requires that Q23 and Bo to be determined at a time appropriately correlated to the cMOS capture (e. g., time 1 in Fig. 4) and substituted into the error term to produce an accurate value of QgT. The cMOS

capture itself can be physically implemented using any of the approaches described above. The subscript q on the bracketed weighted sum of cMOS values indicates that this was an"area"cMOS capture of the area filter set to determine deposited charge.

8.2. Single measurement error term correction A simple, although not the lowest noise, method to estimate the error term is to measure it locally for each cMOS energy capture by making the same Eqn. 34 cMOS filter capture in close time proximity but without a step-like signal present in the filters, so that QgT is zero. Then, from Eqn. 35a, with QgT equal to zero, we have: where the subscript b is for"baseline", called so since, as shown, this"baseline"cMOS filter capture produces the transform's baseline response to the preamplifier's non-ideal terms. We may also call such a measurement a"null charge"or"null energy" ("null area"in the general case) capture since it is made when QgT is zero. Substituting Eqn.

34 c into Eqn. 33A gives: where the subscript q labels the cMOS charge measurement to be corrected.

This correction is moderately simple and has been found to work acceptably in some cases. However, if Q23 is large, as may be true at high output step rates, then time decay effects in [LQ23 + MBo] b may still be significant. This can be approximately corrected by measuring the time interval 6t = te-tb (in clock ticks At) between the energy and baseline captures and correcting by the decay term b2bt, to obtain: In this method, if Bo is truly a constant and the time interval 8t between the energy and baseline captures is fixed, then the error term MBob2sT, will also be a constant and merely produce an offset in the spectrum.

Eqn. 38 has two significant problems: 1) increased noise, due to the noise in the single baseline measurement, which is the same size as in the measurement itself and degrades resolution by sqrt (2); and 2) it increases deadtime by at least a factor of two if the two measurements are independent and so separated by at least 2m+g in time. It is

therefore useful to be able to make multiple, independent measurements of Bo and Q23 and average them in order to reduce their variance. We present a method for doing so below.

8.3. The time compensated baseline model In making filter derivations throughout this work, we have assumed that we have an accurate preamplifier model in terms of its several decay constants and DC offset. We therefore take the logical step and use precisely the same model to correlate measurements made at different times so they may be averaged. Thus, if we measure Q13, Q23, or Bo at one time, then their values At clock cycles later will be given by b1htQl3 b2ttQ23, and Bo, respectively, unless a step arrives within the interval At. In that case, at the end of the step arrival time (e. g., time 1 in Fig. 4), we increment Q13 by (1- P) QgT Q23 by ßQgT, and Bo by zero, after which their exponential decays will continue as before. For the purposes of baseline averaging, then, Q13, Q23, or Bo become"time- compensated"model (or"TC-model") parameters, whose averaged values are <Q13>, <Q23>, and <Bo>, respectively.

If we make measurements of Bo and Q23 from time to time, we can now update our estimate of Bo simply by using the method of WARBURTON-1997 or some other averaging scheme. For Q23 our TC-model now gives: Q23(I)=Q23(I-I\(39) so, <Q23>i-I represents an"averaged"estimate of Q23 (i-1), last measured at time ti 1, we update <Q23>i, using a new baseline measurement of Q23 (i) at time ti, by the following extension of Warburton and Hubbard's Eqn. 7, where Ati = tu-tri: Equivalently, we update <Q23> for the arrival of an impulse signal of charge QgT (i) at time ti using: <BR> <BR> <BR> b2#ti#Q23#(i-1+ßQgT(i)#Q23#i= <BR> <BR> <BR> (41) If required, Q13 can be similarly extended from its last time of measurement by replacing b2inEqn.39.b1 <Q23> and <Bo> are used with an charge cMOS capture to determine QgT using Eqn. 35a as follows. First, <Bo> simply replaces Bo in the equation. <Q23>, however, has to be extended from its time of last update ti-1 to the present capture time ti using Eqn. 39, according to our TC-model, where Ati = ti-ti 1:

Because <Q23> and <Bo> can be determined by averaging over many baseline measurements, the noise contribution of the correction term in Eqn. 42 can be made much smaller than in the single baseline correction procedure of § 8.2. Since the weighting function ofog. is small compared to unity, the measurement noise in Eqn. 42 will then be determined primarily by 6S+ and a-, and will be approximately the same as for a trapezoidal filter of the same peaking time and about a factor of sqrt (2) better than by using Eqn. 38. Furthermore, because a baseline cMOS filter capture can now be made at any time the interval between successive steps is adequately long, rather than requiring one in conjunction with each charge cMOS filter capture, the method's deadtime is cut in half : to that of a single cMOS filter capture.

8.4. Measuring the non-ideal terms O_ and Bo An effective way to measure Q23 at time ti is by making a baseline cMOS capture, per Eqn. 36, with the difference now that the times ti need no longer be correlated with charge cMOS captures. Using an averaged value <Bo> for Bo, plus computed values for L and M, we can then solve [LQ23 + MBO] B for Q23 : and use it to update <Q23>i using Eqn. 40.

We normally expect Bo to be stable and nominally rate independent, so that <Bo> can be measured with no input counts to the detector and used thereafter to obtain Q23 values. However, to compensate for any temperature and/or rate related changes in Bo, it is prudent to also measure Bo from time to time and update our estimate of <Bo>. One approach is to use pairs of baseline measurements, separated in time by At, where, if we denote the error term E as: then, measuring Ei and Ei l separated by At = ti-ti 1, we have, from Eqn. 39:

Bo, i measurements can be averaged, as noted above, to obtain <Bo>, a low variance estimation of Bo, for use in Eqn. 35a. From Eqn. 45, it is clear that the longer the time interval At, the more accurate the measurement of Bo will be, since both the numerator and denominator are small differences of large numbers. In any case, enough values Bo, i can be averaged so that errors in <Bo> do not contribute substantially to the error in QgT.

8.5. Dealing with model parameter drifts For this compensation scheme to work well, the P4 parameter (3 needs to be precisely known and stable over time. Else both the extrapolations of Q13 and increments in Q23 will incorporate systematic errors which will propagate into the QgT calculations. The proposed Q23 baseline measurements offer a method to monitor P and either refine its value or track it in time.

In Eqn. 40 we update our average Q23>i = b2t<Q23>i l at time ti with the measured value Q23 (i). However, we can also record the differences between these two values and compute their mean <AQ23> over time. If <AQ23> is not zero, then a systematic error exists and we can adjust ß using: §P=<AQ23>/Sp, (46) where the slope Sp = d<AQ23>/d is previously obtained from calibration runs, made at system setup time, wherein P is systematically adjusted and the resultant <AQ23> values recorded, so that Sp can be calculated. Similar methods can be used to monitor and correct other P4 parameters such as T1 and T2, if appropriate.

9. References The following are incorporated by reference: AALSETH-1998:"Using pulse shape discrimination to sort individual energy deposition events in a germanium crystal", C. E. Aalseth, F. T. Avignone III, R. L. Brodzinski, H. S.

Miley, and J. H. Reeves in Journal of Radioanalytical and Nuclear Chemistry, Vol. 233, Nos 1-2 (1998) 119-123.

GOULDING-1988:"Ballistic Deficit Correction in Semiconductor Detector Spectrometers", F. S. Goulding and D. A. Landis, I. E. E. E. Trans. Nuclear Science, Vol.

35, No. 1, (1988) 119-124.

HINSHAW-1991: U. S. Patent No. 5,021,664, issued Jun. 4,1991 to S. M. Hinshaw for "Method and apparatus for correcting the energy resolution of ionizing radiation spectrometers".

KNOLL-1989:"Radiation Detection and Measurement, 2nd Ed."by Glenn F. Knoll (J.

Wiley, New York, 1989) KUMAZAWA-1988: U. S. Patent No. 4,727,256, issued Feb. 23,1988 to Y. Kumazawa for"Semiconductor radiation detector".

MILLER-1994: U. S. Patent No. 5,347,129, issued Sept. 13,1994 to W. H. Miller and R. R.

Berliner for"System for determining type of nuclear radiation from detector output pulse shape".

RADEKA-1982:"Trapezoidal Filtering of Signals from Large Germanium Detectors at High Rates", V. Radeka, I. E. E. E. Trans. Nuclear Science, Vol. NS-19 (1982) 412-428.

RAUDORF-1982:"Pulse Shape and Risetime Distribution Calculations for HPGe Coaxial Detectors", T. W. Raudorf, M. O. Bedwell and T. J. Paulus, I. E. E. E. Trans.

Nuclear Science, Vol. NS-29, No. 1 (1982) 764-768.

SIMPSON-1990: U. S. Patent No. 4,937,452, issued Jun. 26,1990 to M. L. Simpson and T. W. Raudorf for"Charge trapping correction in photon detector systems".

STAHLE-1999:"Design and performance oftheAstro-E/XRF microcalorimeter array and anti-coincidence detector", C. K. Stahle et al., in Proc. of the SPIE No. 3765, "EUV, X-ray and Gamma-ray Instrumentation for Astronomy X" (Denver, CO, July 21-23,1999), in press.

TAKAHASHI-1994:"A Multiparametric Waveform Analysis of Ge Detector Signal Based on Fast ADC Digitizing Technique", by H. Takahashi, S. Kinjoh, J.

Kawarabayashi, T. Iguchi, and M. Nakazawa, I. E. E. E. Trans. Nuclear Science, Vol.

41, No. 4, (1994) 1246-1249.

WARBURTON-1997: U. S. Patent No. 5,684,850, issued Nov. 4,1997 to W. K.

Warburton and B. Hubbard for"Method and apparatus for digitally based high speed x-ray spectrometer".

WARBURTON-1998: U. S. Patent No. 5,774,522, issued Jun. 30,1998 to W. K.

Warburton for"Method and apparatus for digitally based high speed x-ray spectrometer for direct coupled use with continuous discharge preamplifiers".

WARBURTON-1999: U. S. Patent No. 5,873,054, issued Feb. 16,1999 to W. K.

Warburton and Z. Zhou for"Method and apparatus for combinatorial logic signal processor in a digitally based high speed x-ray spectrometer".

WHITE-1988:"Pulse Processing for Gamma Ray Spectrometry: a Novel Method and its Implementation", G. White, I. E. E. E. Trans. Nuclear Science, Vol. 35, No. 1, (1988) 125-130.

10. Conclusion The foregoing description of specific embodiments has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise forms described, and obviously, many modifications and variations are possible in light of the above teaching. These embodiments were chosen and described in order to best explain the principles of the invention and its practical application to thereby enable others in the art to best utilize the invention in various embodiments and with such modifications as best suit the invention to the particular uses contemplated.

As a first example, while all the described embodiments used trapezoidal or triangular filters, there is no such restriction in the technique. As shown, any filter describable by a set of weighting constants applied to successive signal samples, or its analog equivalent, can be used. The method will therefore work with essentially any realizable filter. Second, while the specific embodiments employed two, three, or four filters, there is no specific limit on the number of filters which can be used. Even a single filter can be used if its output is correctly sampled at several times. If the number of filters N exceeds the number of base amplitudes M, then matrix inversion methods can be used to find least squares best estimates of the M values, rather than exact values. Third, while all the filters used in the specific embodiments were symmetric, this is not a requirement of the method, as may be seen from the derivations. Fourth, while specific configurations of electronic circuits were shown in the implementations, other configurations can readily be devised which accomplish the same functions. The number of ADC bits or value lengths captured to the registers could be varied, for example. Fifth, even the use of electronic circuits is not required to implement the method. As per the initial derivations, for example, if successive ADC values are instead read into computer memory, all the subsequent processing steps can be carried out in software. Sixth, the method can also be applied to signals which are formally equivalent to outputs from nominally single pole filters, whether the signal source is actually a single pole filter or not. Superconducting bolometers produce such output signals, for example.

Therefore, the above description should not be taken as limiting the scope of the invention, as defined by the appended claims.