1
Technical Note
The Principle of Time-Correlated
Single Photon Counting
Time-resolved fluorescence spectroscopy is a powerful
analysis tool in fundamental physics as well as
in the life sciences. Implementing it in the time domain
requires recording the time dependent intensity
profile of the emitted light upon excitation by a
short flash of light, typically a laser pulse. While in
principle, one could attempt to record the time decay
profile of the signal from a single excitation-emission
cycle, there are practical problems preventing such
a simple solution in most cases. First of all, the decay
to be recorded is very fast. Typical fluorescence
from commonly used organic fluorophores lasts only
some hundred picoseconds to some tens of nanoseconds.
In order to recover not only fluorescence
lifetimes but also the decay shape, typically in order
to resolve multi-exponential decays, one must
be able to temporally resolve the recorded signal at
least to such an extent, that the decay is represented
by some tens of samples. For a decay lasting, e.g.,
500 ps the signal would have to be sampled at time
steps of say 10 ps.
At first glance it might seem reasonable to do this
with a photo-diode and a fast oscilloscope or some
similar electronic transient recorder. However, the required
temporal resolution is hard to achieve with ordinary
electronic transient recorders. Moreover, the
emitted light may be simply too weak to create an analog
voltage representing the optical flux. Indeed, the
optical signal may consist of just a few photons per
excitation/emission cycle. Then the discrete nature
of the signal itself prohibits analog sampling. Even
if one has some reserve to increase the excitation
power to obtain more fluorescence light, there will
be limits, e.g., due to collection optic losses, spectral
limits of detector sensitivity or photo-bleaching at
higher excitation power. Ultimately, problems would
arise when the observed sample consists of just a
few or even single molecules, a situation that is absolutely
real in confocal microscopy applications.
The solution for these problems is Time-Correlated
Single Photon Counting (TCSPC). With periodic
excitation, e.g., from a laser, it is possible to extend
the data collection over multiple cycles of excitation
and emission. One can then accept the sparseness
of the collected photons and reconstruct the fluorescence
decay profile from the multitude of single photon
events collected over many cycles.
The method is based on the repetitive, precisely
timed registration of single photons of, e.g., a fluorescence
signal[1],[2]
. The reference for the timing is
the corresponding excitation pulse. As a single photon
sensitive detector a Photomultiplier Tube (PMT),
Micro Channel Plate (MCP), a Single Photon Avalanche
Diode (SPAD) or Hybrid PMT can be used.
Provided that the probability of registering more than
one photon per cycle is low, the histogram of photon
arrivals per time bin represents the time decay one
would have obtained from a “single shot” time-resolved
analog recording. The precondition of single
photon probability can (and must) be met by simply
attenuating the light level at the sample if necessary.
Figure 1 illustrates how the histogram is formed
over multiple cycles. In the example, fluorescence is
excited repetitively by short laser pulses. The time
difference between excitation and emission is measured
by electronics that act like a stopwatch. If the
Time-Correlated Single Photon Counting
Michael Wahl
PicoQuant GmbH, Rudower Chaussee 29, 12489 Berlin, Germany, info@picoquant.com
2
single photon probability condition is met, there will
actually be no photons at all in many cycles. In the
example this situation is shown after the second laser
pulse. It should be noted that (by the laws of quantum
physics) the occurrence of a photon or an empty
cycle is entirely random and can only be described in
terms of probabilities. Consequently, the same holds
true for the individual stopwatch readings.
The stopwatch readings are sorted into a histogram
consisting of a range of “time bins”. The width
of the time bins typically corresponds to the resolution
of the stopwatch (some picoseconds) but may
be chosen wider in order to cover a longer overall
time span. The typical result in time-resolved fluorescence
experiments is a histogram with an exponential
drop of counts towards later times (Figure 2).
The reason why there typically is an exponential
drop is very similar to that of nuclear decay. As quantum
mechanics predict, we have no means of knowing
exactly when a nuclear decay will occur. All we
can predict is the likelihood of an atomic species to
decay in a given period of time. Similarly, all we can
predict about the lifetime of an excited state is its statistical
expectation. The exponential drop of fluorescence
intensity in a single shot experiment with many
molecules may be explained as follows: Suppose we
begin with a population of 1000 excited molecules.
Let the probability of each molecule returning to the
ground state be 50 % in the first nanosecond. Then
we have 50 % of the exited population after the first
nanosecond. In the next nanosecond of observation
we lose another 50 %, and so on. Since the intensity
of light is determined by the number of photons emitted
in any period of time, it is directly proportional to
the surviving population of excited molecules. When
the experiment is done with single molecules it is of
course no longer meaningful to speak of populations.
Nevertheless, by virtue of ergodicity, the likelihood of
observing a photon, i.e., a molecule’s return to the
ground state as a function of time follows the same
exponential drop over time.
It is important to note that we can but need not
actually do ­TCSPC with single molecules. Sufficiently
attenuating the light, so that the detector receives
only single photons, has the same effect. Indeed,
in order to use ­TCSPC we must attenuate the light
to this level. One may ask now, if we do have light
from many molecules, why waste it by attenuation
and use ­TCSPC? The good reason to do so is that a
single photon detector can be built with much better
time resolution than an analog optical receiver.
In practice, the registration of a photon in time-resolved
fluorescence measurements with ­TCSPC involves
the following steps: First, the time difference
between the photon event and the corresponding
excitation pulse must be measured. For this purpose
both optical events are converted into electrical pulses.
For the fluorescence photon this is done via the
single photon detector mentioned before. For the excitation
pulse it may be done via another detector
(typically called trigger diode) if there is no electrical
synchronization signal (sync) supplied directly by the
laser. Obviously, all conversion to electrical pulses
Figure 1: Measurement of start-stop times in time-resolved fluorescence measurement with TCSPC.
Figure 2: Histogram of start-stop times in time-resolved fluorescence
measurement with TCSPC.
3
must preserve the precise timing of the signals as
accurately as possible.
The actual time difference measurement is done
by means of fast electronics which provide a digital
timing result. This digital timing result is then used to
address the histogram memory so that each possible
timing value corresponds to one memory cell or
histogram bin. Finally, the addressed histogram cell
is incremented. All steps are carried out by fast electronics
so that the processing time required for each
photon event is as short as possible. When sufficient
counts have been collected, the histogram memory
can be read out. The histogram data can then be
used for display and, e.g., fluorescence lifetime calculation.
In the following sections we will expand on
the various steps involved in the method and associated
issues of importance.
Count Rates and Single Photon
Statistics
It was mentioned that it is necessary to maintain a
low probability of registering more than one photon
per cycle. This is to guarantee that the histogram of
photon arrivals represents the time decay one would
have obtained from a single shot time-resolved analog
recording. The reason for this is briefly the following:
Detector and electronics have a “dead” time
for at least some nanoseconds after a photon event.
During this time they cannot process another event.
Because of these dead times ­TCSPC systems are
usually designed to register only one photon per excitation
cycle. If now the number of photons occurring
in one excitation cycle were typically > 1, the system
would very often register the first photon but miss
the following ones. This would lead to an over-representation
of early photons in the histogram, an effect
called ‘pile-up’. It is therefore crucial to keep the
probability of cycles with more than one photon low
(Figure 3).
To quantify this demand one has to set acceptable
error limits for the lifetime measurement and apply
some mathematical statistics. For practical purposes
one may use the following rule of thumb: In order
to maintain single photon statistics, on average only
one in 20 to 100 excitation pulses should generate
a count at the detector. In other words: the average
count rate at the detector should be at most 1 to
5 % of the excitation rate. E.g., with a pulsed diode
laser driver of PicoQuant’s ­PDL ­Series running at
80 MHz repetition rate, the average detector count
rate should not exceed 4 MHz. This leads to another
issue: the maximum count rate the system (of both
detector and electronics) can handle.
Indeed, 4 MHz are stretching the limits of some
detectors and certainly are way beyond the capabilities
of old NIM based ­TCSPC systems. On the other
hand, one wants high count rates in order to acquire
fluorescence decay histograms quickly. This may
be of particular importance where dynamic lifetime
changes or fast molecule transitions are to be studied
or where large numbers of lifetime samples must
be collected (e.g., in 2D scanning configurations).
PMTs (dependent on the design) can handle count
rates of up to 1 to 20 millions of counts per second
(cps), passively quenched SPADs saturate at a few
hundred kcps. Old-fashioned NIM based ­TCSPC
electronics are able to handle a maximum of 50,000
to 500,000 cps. With modern integrated ­TCSPC designs,
e.g. TimeHarp 260, count rates up to 40 Mcps
can be achieved.
It is also worth noting that by virtue of quantum
mechanics the actual count arrival times are random,
so that there can be bursts of high count rate and
periods of low count rates. Bursts of photons may
well exceed the average rate. This should be kept in
mind when an experiment is planned. Even if an instrument
can accommodate the average rate, it may
drop photons in bursts. This is why the length of the
dead time is of interest too. This quantity describes
the time the system cannot register photons while it
is processing a previous photon event. The term is
applicable to both detectors and electronics. Through
Figure 3: Distortion of the TCSPC measurement by pile-up effect and dead time.
4
pile-up effects dead time can distort the decay histogram
and thereby affect the lifetime to be extracted
from the latter. Furthermore, the photon losses prolong
the acquisition time or deteriorate the SNR if the
acquisition time remains fixed. In applications where
the photon burst density must be evaluated, e.g., for
molecule transition detection or imaging, long dead
times might be a problem.
Timing Resolution
The characteristic of a complete ­TCSPC system that
summarizes its overall timing precision is its Instrument
Response Function (IRF). The basic idea is
that if the system is ideal, i.e., has an infinitely sharp
excitation pulse and infinitely accurate detectors and
electronics, it should have an infinitely narrow IRF.
Any deviation from this ideal results in a broadening
of the IRF. Before looking into how the individual
error contributions add up, the most critical sources
shall be introduced here.
The weakest component in terms of timing resolution
in ­TCSPC measurements will usually be the
detector. However, as opposed to analog transient
recording, the time resolution of ­TCSPC is not limited
by the pulse response of the detector. Only the
timing accuracy of registering a photon determines
the ­TCSPC resolution. The timing accuracy is limited
by the timing uncertainty the detector introduces
in the conversion from a photon to an electrical
pulse. This timing error (or uncertainty) can be as
much as 10 times smaller than the detector’s pulse
response. The timing uncertainties are usually quantified
by specifying the r.m.s. error or the Full Width
Half Maximum (FWHM) of the timing error distribution.
Note that these two notations are related but not
identical. In the case of a Gaussian error distribution
the FWHM value is about 2.3 times as large as the
corresponding r.m.s. value. Good but also expensive
detectors, notably MCPs, can achieve timing uncertainties
as small as 25 ps FWHM. Low cost PMTs or
SPADs may introduce uncertainties of 200 to 400 ps
FWHM. Modern SPADs show timing uncertainties as
small as 30 ps FWHM.
The second most critical source of IRF broadening
usually is the excitation source. While most laser
sources can provide sufficiently short pulses, it
is also necessary to obtain an electrical timing reference
signal (sync) to compare the fluorescence
photon signal with. Where this signal is derived
from depends on the excitation source. With gain
switched diode lasers, e.g., the LDH Series along
with the laser drivers of the PDL Series, a low jitter
electrical sync signal is readily available. The signal
type used here is commonly a narrow negative pulse
of -800 mV into 50 Ohms (NIM standard). The very
sharp falling edge is synchronous with the laser pulse
(typically < 3 ps r.m.s. jitter for the PDL Series). With
other lasers, e.g., many Ti:Sa lasers, a second detector
must be used to derive a sync signal from the
optical pulse train. This is commonly done with a fast
photo diode (APD or PIN diode, e.g., the TDA 200
from PicoQuant). The light for this reference detector
must be coupled out from the excitation laser beam,
e.g., by means of a semi-transparent mirror. The reference
detector must be chosen and set up carefully
as it also contributes to the overall timing error.
Another source of timing error is the timing jitter of
the electronic components used for ­TCSPC. This is
caused by the finite rise/fall-time of the electric signals
used for the time measurement. At the trigger
point of compactors, logic gates, etc., the amplitude
noise, e.g., thermal noise and interference, always
present on these signals is transformed to a corresponding
timing error (phase noise). However, the
contribution of the electronics to the total timing error
usually is relatively small. Modern ­TCSPC electronics
cause an r.m.s. jitter of < 10 ps. Nevertheless, it
is always a good idea to keep the RF noise low. For
this reason signal leads should be properly shielded
coax cables, and strong sources of RF interference
should be kept away from the ­TCSPC detector and
electronics.
The contribution of the time spread introduced by
the individual components of a ­TCSPC system to
the total IRF width strongly depends on their relative
magnitude. Strictly, the total IRF is the convolution of
all component IRFs. An estimate of the overall IRF
width system can be obtained from the geometric
sum of the individual components, e.g., in the form of
r.m.s. or FWHM values according to statistical error
propagation laws:
lu39c1s
size 12{e rSub { size 8{ ital "IRF"} rSub { size 8
"system"} } } » sqrt { Sum {e rSub { size 8{ ital
"component"} } rSup { size 8{2} } } } } {}
eIRF
system
»√∑ecomponent
2
(1)
Obviously, due to the squares in the sum, the total
will be more than proportionally dominated by the
largest component. It is therefore of little value to improve
a component that is already relatively good. If,
e.g., the detector has an IRF width of 200 ps, shortening
of the laser pulse from 50 to 40 ps is practically
of no effect.
Apart from predicting the approximate IRF width
according to Eqn. 1 one can of course measure it.
The typical approach is to place a scattering medium
in the sample compartment so that there is no
fluorescence but only some scattered excitation light
reaching the detector. The IRF measurement is not
only a means of optimizing and characterizing the
instrument. It also serves as an input to data analysis
with “deconvolution” and is therefore a frequent
measurement task.
It was mentioned earlier that the total IRF is the
convolution of all component IRFs. Similarly, the
measured fluorescence decay is the convolution of
the “true” physical process of exponential decay with
5
the IRF. With this theoretical model it is possible to
extract the parameters of the “true” decay process
from the convoluted results in the collected histo-
grams[3]
. This is often referred to as “deconvolution”
although it should be noted that the term is not mathematically
precise in this context. The procedure that
most data analysis programs actually perform is an
iterative reconvolution. If the detector’s temporal response
shows a spectral dependency then it is not
recommendable to measure the IRF at the excitation
wavelength while the fluorescence will be measured
at a different wavelength. The deconvolution
would then not be meaningful. In such cases the IRF
should be measured at the target sample’s emission
wavelength using strongly quenched fluorophores
showing very short lifetimes.
Having established the role of the IRF and possibly
having determined it for a given instrument
leads to the question what the actual lifetime measurement
resolution of the instrument will be. Unfortunately,
it is difficult to specify a general lower limit
on the fluorescence lifetime that can be measured
by a given ­TCSPC instrument. Apart from the instrument
response function and noise, factors such as
quantum yield, fluorophore concentration, and decay
kinetics all affect the measurement. However, as a
rule of thumb, one can assume that under favorable
conditions, most importantly sufficient counts in the
histogram, lifetimes down to 1/10 of the IRF width
(FWHM) can still be recovered via iterative recon-
volution.
A final time-resolution related issue worth noting
here is the bin width of the ­TCSPC histogram. As
outlined above, the analog electronic processing of
the timing signals (detector, amplifiers, etc.) creates
a continuous, e.g., Gaussian, distribution around the
true value. In order to form a histogram, at some
point the timing results must be quantized into discrete
time bins. This quantization introduces another
random error that can be detrimental if chosen too
coarse. The time quantization step width, i.e., the bin
width, must therefore be small compared to the width
of the analog error distribution. As a minimum from
the information theoretical point of view one would
assume the Nyquist frequency. I.e., an analog signal
should be sampled at least at twice the highest frequency
contained in it. The high frequency content
depends on the shape of the distribution. In a basic
approximation this can reasonably be assumed
to be Gaussian and therefore having very little high
frequency content. For practical purposes there is
usually no point in collecting the histogram at time
resolutions much higher than 1/10 of the width of
the analog error distribution. Nevertheless, a good
histogram resolution is helpful in data analysis with
iterative reconvolution.
Photon Counting Detectors
General Characteristics
What we expect of an ideal photon detector for
­TCSPC is an electrical output pulse upon each photon
arriving at the detector. In the real world there are
various deviations from this ideal. As outlined in the
section on timing resolution, one imperfection of real
world detectors manifests itself in an uncertainty of
time delay between photon arrival and electrical output.
We describe it by the IRF width of the detector.
Furthermore, due to finite sensitivity we do not get
an output pulse for each input photon. The most important
characteristic in this context is the quantum
efficiency of the detector. It critically depends on material
properties and incident wavelength. The peak
quantum efficiency is between 5 and 90 % for typical
detectors. There are other factors beyond quantum
efficiency that may affect the overall conversion efficiency
of the detector but they are less dominant.
Because of noise from various sources in the detector,
the output may contain pulses that are not
related to the light input. These are referred to as
dark counts. In terms of a characteristic of the detector
it is common to use the dark count rate (DCR) in
counts per second. It can be close to zero for some
specialized detectors but as high as some thousands
of cps in other cases. Since it is mostly driven by
thermal effects, it is typically higher in detectors with
sensitivity in the near-infrared. For the same reason
it can typically be reduced by cooling.
Another type of “false” output from a photon detector
is due to so-called afterpulsing. The term describes
the observation that some time after a true
photon event the detector emits another pulse. The
internal reasons for such effects are different in the
various types of detectors. As a common problem
such false counts are temporally correlated with true
photon events. The typical delay time is in the order
of microseconds. The correlation is also weak, so
that it does not show in many applications. In fact,
in classical ­TCSPC with nanosecond time spans afterpulses
very often appear just like a background
noise that increases with illumination. However, afterpulsing
can induce spurious effects when the typical
delay of the afterpulses is in the same time range
as the photon correlation times to be observed in
the experiment. This is typically the case in Fluorescence
Correlation Spectroscopy (FCS). Fortunately,
there exist detectors with very little or even zero af-
terpulsing.
Photomultiplier Tube (PMT)
A PMT consists of a light-sensitive photocathode
that generates electrons when exposed to light.
These electrons are directed onto a charged electrode
called a dynode. The collision of the electrons
with the dynode produces additional electrons. Since
each electron that strikes the dynode causes several
electrons to be emitted, there is a multiplication ef-
6
fect. After further amplification by multiple dynodes,
the electrons are collected at the anode of the PMT
and output as a current. The current is directly proportional
to the intensity of light striking the photocathode.
Because of the multiplicative effect of the
dynode chain, the PMT is a photoelectron amplifier
of high sensitivity and remarkably low noise. PMTs
have a wide dynamic range, i.e., they can also measure
relatively high levels of light. They furthermore
are very fast, so rapid successive events can be reliably
monitored. PMTs are also quite robust. The high
voltage driving the tube may be varied to change the
sensitivity of the PMT.
When the light levels are as low as in ­TCSPC, the
PMT “sees” only individual photons. One photon on
the photocathode produces a short output pulse containing
millions of electrons. PMTs can therefore be
used as single photon detectors. In photon counting
mode, individual photons that strike the photocathode
of the PMT are registered. Each photon event
gives rise to an electrical pulse at the output. The
number of pulses, or counts per second, is proportional
to the light impinging upon the PMT. As the
number of photon events increases at higher light
levels, it will become difficult to differentiate between
individual pulses and the photon counting detector
will become non-linear. Dependent on the PMT design
this usually occurs at 1 to 10 millions of counts
per second. The timing uncertainty between photon
arrival and electrical output is small enough to permit
time-resolved photon counting at a sub-nanosecond
scale. In single photon counting mode the tube is
typically operated at a constant high voltage where
the PMT is most sensitive.
PMTs usually operate between the blue and red
regions of the visible spectrum, with greater quantum
efficiency in the blue-green region, depending
upon photo-cathode materials. Typical peak quantum
efficiencies are about 25 %. For spectroscopy
experiments in the ultraviolet and visible region
of the spectrum, a photomultiplier tube is very well
suited. In the near infrared the sensitivity drops off
rapidly. Optimized cathode materials can be used to
push this limit, which may on the other hand lead to
increased noise. The latter can to some extent be
reduced by cooling.
Because of noise from various sources in the
tube, the output of the PMT may contain pulses that
are not related to the light input. These are referred
to as dark counts, as outlined previously. The detection
system can to some extent reject these spurious
pulses by means of electronic discriminator circuitry.
This discrimination is based on the probability that
some of the noise generated pulses (those from the
dynodes) exhibit lower signal levels than pulses from
a photon event.
Microchannel Plate PMT (MCP)
A microchannel plate PMT is also a sensitive photon
detector. It consists of an array of glass capillaries
(10 to 25 µm inner diameter) that are coated on the
inside with an electron-emissive material. The capillaries
are biased at a high voltage applied across
their length. Like in the PMT, an electron that strikes
the inside wall of one of the capillaries creates an
avalanche of secondary electrons. This cascading
effect creates a gain of 103 to 106 and produces a
current pulse at the output. Due to the confined paths
the timing jitter of MCPs is sufficiently small to perform
time-resolved photon counting on a sub-nanosecond-scale,
usually outperforming PMTs. Good
but also expensive MCPs can achieve timing uncertainties
as low as 25 ps. Microchannel plates are also
used as an intensifier for low-intensity light detection
with array detectors.
Avalanche Photo Diode (APD)
APDs are the semiconductor equivalent of PMTs.
Generally, APDs may be used for ultra-low light detection
(optical powers < 1 pW), and can be used in
either “linear” mode (bias voltage slightly less than
the breakdown voltage) at gains up to about 500, or
as photon-counters in the so-called “Geiger” mode
(biased slightly above the breakdown voltage). In the
case of the latter, the term gain is meaningless. A
single photon may trigger an avalanche of about 108
carriers but one is not interested in the output current
or voltage because it carries no information other
than “there was a photon”. Instead, in this mode the
device can be used as a detector for photon counting
with very accurate timing of the photon arrival. In
this context APDs are referred to as Single Photon
Avalanche Diodes (SPAD). Widespread commercial
products attain timing uncertainties on the order of
400 ps FWHM. Single photon detection probabilities
of up to approximately 50 % are possible. Maximum
quantum efficiencies reported are about 80 %. More
recent SPAD designs focus on timing resolution and
can achieve timing accuracies down to 30 ps but are
less sensitive at the red end of the spectrum. The
dark count rate (noise) of SPADs strongly depends
on the active area. In SPADs it is much smaller than
in PMTs, which can make optical interfacing difficult.
Hybrid PMT detectors
By combining a PMT front end with an APD amplification
stage it is possible to design a hybrid detector
that provides a very clean instrument response and
virtually zero afterpulsing. The timing uncertainty
is on the order of 50 to 100 ps and the distribution
of timing error is nearly Gaussian. The PMT front
end requires a very high voltage but the detectors
are available as compact modules readily including
the high voltage supply (PicoQuant’s PMA Hybrid).
Due to the PMT front end their sensitivity follows the
same fundamental dependency on cathode material
and wavelength as in ordinary PMTs.
7
Other detectors
The field of photon detectors is still evolving. Recent
developments that are beginning to emerge
as usable products include so called silicon PMTs,
superconducting nanowire detectors and APDs with
sufficient gain for single photon detection in analog
mode. Each of these detectors has its specific benefits
and shortcomings. Only a very brief overview
can be given here. Silicon PMTs are essentially arrays
of SPADs, all coupled to a common output. This
has the benefit of creating a large area detector that
can even resolve photon numbers. The drawback
is increased dark count rate and relatively poor timing
accuracy. Superconducting nanowires, typically
made from NbN, can be used to create photon detectors
with excellent timing performance and high
sensitivity reaching into the infrared spectrum. The
shortcomings for practical purposes are the extreme
cooling requirements and the low fill factor of the wire
structures, making it difficult to achieve good collection
efficiencies. Another class of potentially interesting
detectors are recently emerging APDs with very
high gain. In combination with a matched electronic
amplifier they have been shown to detect single
photons. As opposed to Geiger mode, this avoids
afterpulsing and allows very fast counting rates. The
disadvantage is a high dark count rate, currently way
too high for any practical ­TCSPC application.
Basic Principles Behind the TCSPC
Electronics
PicoQuant’s ­TCSPC systems are quite advanced. To
begin explanation, we start with traditional ­TCSPC
systems. They typically consist of the building blocks
shown in Figure 4.
The Constant Fraction Discriminator (CFD) is
used to extract precise timing information from the
electrical detector pulses that may vary in amplitude,
typically those from a PMT or MCP detector. This way
the overall system IRF may be tuned to become narrower.
The same could not be achieved with a simple
level trigger (comparator). Particularly with PMTs
and MCPs, constant fraction discrimination is very
important as their pulse amplitudes vary significantly.
Figure 5 shows a comparison between level trigger
and CFD operation. The pulses are shown with negative
voltages as typically delivered from PMTs.
The most common way of implementing a CFD
is the comparison of the original detector signal with
an amplified and delayed version of itself. The signal
derived from this comparison changes its polarity exactly
when a constant fraction of the detector pulse
height is reached. The zero crossing point of this
signal is therefore suitable to derive a timing signal
independent from the amplitude of the input pulse.
This is done by a subsequent comparison of this signal
with a settable zero level, the so called zero cross
trigger. Making this level fixable allows to adapt to
the noise levels in the given signal, since in principle
an infinitely small signal could trigger the zero cross
comparator.
It should be noted that modern CFDs, like in most
of PicoQuant’s systems, work a little differently. They
basically detect the vertex of each pulse and trigger
on that point. Effectively this is like applying a constant
fraction of 1. For practical purposes it makes no
difference (we get a timing signal independent from
the amplitude of the input pulse) and in terms of implementation
it makes things easier.
Typical CFDs furthermore permit the setting of a
so called discriminator level, determining the lower
limit the detector pulse amplitude must pass. Random
background noise pulses can thereby be suppressed.
Particularly pulses originating from random
electrons generated at the dynodes of the PMT can
be suppressed as they had less time to amplify, so
that their output pulses are small (Figure 6).
Similar as for the detector signal, the sync signal
must be made available to the timing circuitry. Since
the sync pulses are usually of well-defined amplitude
and shape, a simple settable comparator (level trigger)
is sufficient to adapt to different sync sources.
In the classical design the signals from the CFD and
SYNC trigger are fed to a Time to Amplitude Converter
(TAC). This circuit is essentially a highly linear
ramp generator that is started by one signal and
stopped by the other. The result is a voltage proportional
to the time difference between the two signals
(Figure 7).
The voltage obtained from the TAC is then fed
Figure 4: Block diagram of a traditional TCSPC system. CFD = Constant Fraction Discrimonator, TAC = Time to Amplitude Converter,
ADC = Analog to Digital Converter
8
to an Analog to Digital Converter (ADC) which provides
the digital timing value used to address the
histogrammer. The ADC must be very fast in order
to keep the dead time of the system short. Furthermore,
it must guarantee a very good linearity, over
the full range as well as differentially. These are criteria
difficult to meet simultaneously, particularly with
ADCs of high resolution, typically 12 bits, as desirable
for ­TCSPC over many histogram channels. Furthermore,
the TAC range is limited.
The histogrammer has to increment each histogram
memory cell whose digital address in the histogram
memory it receives from the ADC. This is
commonly done by fast digital logic, e.g., in the form
of Field Programmable Gate Arrays (FPGA) or a microprocessor.
Since the histogram memory at some
point also must be available for data readout, the
histogrammer must stop processing incoming data.
This prevents continuous data collection. Sophisticated
­TCSPC systems solve this problem by switching
between two or more memory blocks, so that one
is always available for incoming data.
While this section so far outlined the typical structure
of conventional ­TCSPC systems, it is now time
to note that the tasks performed by TAC and ADC
can be carried out by a single fully digital circuit, a so
called Time to Digital Converter (TDC). These circuits
can measure time differences based on the delay
times of signals in semiconductor logic gates or the
conductor strips between them[4]
. The relative delay
times in different gate chains can be used to determine
time differences well below the actual gate delay.
Other TDC designs use interpolation techniques
between the pulses of a coarser clock. This permits
exceptionally small, compact and affordable ­TCSPC
solutions, as the circuits can be implemented as Application
Specific Integrated Circuits (ASICs) at low
cost and high reliability. All PicoQuant ­TCSPC systems
make use of such a state-of-the-art design.
In the simplest form of a TDC based ­TCSPC system
the TAC and ADC of the classical approach are
replaced by a TDC. However, this simple solution is
no longer used in current systems. Modern ­TCSPC
devices are fundamentally different in design. Instead
of operating like a stopwatch, they have independent
TDCs for each input channel. The important
detail is that the individual TDCs are running off the
same crystal clock. If a timing difference is needed,
like in classical histogramming, it can be obtained by
simple arithmetics in hardware. Figure 8 shows this
in a block diagram.
Observe the symmetry of the input channels,
now both having a CFD. The symmetry as well as
the separation of the input channels allow many adFigure
5: Comparison between level trigger (left) and CFD operation (right).
Figure 6: Typical train of PMT pulses and the CFD’s discriminator
level (red line).
Figure 7: Operation principle of a TAC (Time to Amplitude
Converter) level (red line).
9
vanced ­TCSPC concepts and new applications that
will be discussed further below. For the time being
we will first take a look at a ­TCSPC set-up in the
lab. This is fairly independent from the design of the
­TCSPC electronics used.
Experimental Set-up for Fluorescence
Decay Measurements with
TCSPC
Figure 9 shows a simple set-up for fluorescence lifetime
measurements with ­TCSPC. The picosecond
diode laser is running on its internal clock. The driver
box (PicoQuant’s PDL Series) is physically separate
from the actual laser head (PicoQuant’s LDH Series),
which is attached via a flexible lead. This permits to
conveniently place the small laser head anywhere in
the optical set-up.
The light pulses of typically 50 ps FWHM are directed
at the sample cuvette, possibly via some
appropriate optics. A neutral density filter is used to
attenuate the light levels to maintain single photon
statistics at the detector. Upon excitation, the fluorescent
sample will emit light at a longer wavelength
than that of the excitation light. The fluorescence
light is filtered out against scattered excitation light
by means of an optical cut-off filter. Then it is directed
to the photon detector, again possibly via some
appropriate collection optics, e.g., a microscope objective
or just a lens. For timing accuracies of 200 ps
FWHM (permitting lifetime measurements even
shorter than this via reconvolution), a economic PMT
is sufficient. The electrical signal obtained from the
detector, e.g., a small negative pulse of -20 mV, is fed
to a pre-amplifier, and then to the ­TCSPC electronics
via a standard 50 Ohms coax cable. In this example
the complete ­TCSPC electronics are contained on a
single PC board (TimeHarp 260). Other models are
designed as separate boxes connected via USB.
The laser driver also provides the electric sync
signal needed for the photon arrival time measureFigure
8: Block diagram of a modern TDC based TCSPC system running in classical histogramming mode.
Figure 9: Simple experimental set-up for fluorescence decay measurements with TCSPC.
10
ment. This signal (NIM standard, a narrow pulse of
-800 mV) is fed to the ­TCSPC electronics via a standard
50 Ohms coax cable.
Figure 10 shows two ­TCSPC histograms obtained
with this kind of setup including the TimeHarp 260
­TCSPC electronics. Excitation source was a
PDL  800-B with a 470 nm laser head running at
10 MHz repetition rate. A PDM SPAD from MPD was
used for detection. The narrower peak (blue curve)
represents the system IRF, here dominated by laser
and detector. The other curve (green) corresponds
to the fluorescence decay from a solution of Atto 488
in water, a fluorescent dye with fairly short fluorescence
lifetime (~ 3.8 ns). The count rate was adjusted
to < 1 % of the laser rate to prevent pile-up. The plot
in logarithmic scale shows the perfect exponential
nature of the decay curve, as one would expect it.
Note that this is obtained even without a deconvolution
of the IRF.
The approximate mono-exponential fluorescence
lifetime can be obtained from a simple comparison of
two points in the exponential display with count rates
in the ratio of e:1, e.g., 27,180:10,000. In this particular
experiment this results in a lifetime estimate of
3.8 ns as expected for Atto 488.
Of course it is easy to measure long lifetimes with
or without reconvolution, since the IRF is of less influence.
However, for a precise measurement of short
lifetimes one would perform an iterative reconvolution
fit taking into account the IRF. Nevertheless, one
can measure lifetimes significantly smaller than the
IRF with this method. Additionally, the r.m.s. residue
from the fit can be used to assess the quality of the
fit and thereby the reliability of the lifetime measurement.
The FluoFit decay analysis software package
from PicoQuant provides this functionality and even
allows an advanced error analysis in order to assess
the confidence boundaries of the numerical results.
The high-end software package SymPhoTime 64
provides the same functionality, in addition to many
other features, in the context of microscopy and imaging
applications.
Reverse Start-Stop Mode No Longer
Required
For logic and understanding it is obvious to assume
that the time delay measurement is directly causal,
i.e., the laser pulse causes a photon event and one
measures the time delay between laser pulse and the
subsequent photon event. However, old style TAC
based ­TCSPC systems need to give up this logical
concept because of the typically high repetition rates
of the excitation lasers: Since the time measurement
circuit cannot know in advance whether there will
be a fluorescence photon, it would have to start a
time measurement upon each laser pulse. This is
something the old-fashioned TAC cannot deal with.
By reversing the start and stop signals in the time
Figure 10: TCSPC histograms of fluorescence and instrument response in log scale (direct screenshot from the data acquisition software).
11
measurement, the conversion rates are only as high
as the actual photon rates generated by the fluorescent
sample. These can be handled by the TAC. The
consequence of this approach, however, is that the
times measured are not those between laser pulse
and corresponding photon event, but those between
photon event and the next laser pulse, unless a long
cable delay is inserted. This still works (by software
data reversing) but is inconvenient in various ways.
PicoQuant’s recent ­TCSPC electronics are very
different in this respect, as they work in forward
start stop mode, even with ultrafast lasers. This is
facilitated by independent operation of the TDCs of
all channels and a programmable divider in front of
the sync input. The latter allows to reduce the input
rate so that the period is at least as long as the dead
time. Internal logic determines the sync period and
re-calculates the sync signals that were divided out.
It must be noted that this only works with stable sync
sources that provide a constant pulse-to-pulse period,
but all fast laser sources known today meet this
requirement within an error band of a few picosec-
onds.
Advanced TCSPC
Historically, the primary goal of ­TCSPC was the determination
of fluorescence lifetimes upon optical
excitation by a short light pulse[1],[2]
. This goal is still
important today and therefore has a strong influence
on instrument design. However, modifications and
extensions of the early designs allow for the recovery
of much more information from the detected photons
and enable entirely new applications.
Classical ­TCSPC for fluorescence lifetime measurements
only uses the short term difference between
excitation and emission. It was soon realized
that other aspects of the photon arrival times were
of equally great value in the context of single molecule
fluorescence detection and spectroscopy. For
instance, in single molecule experiments in flow capillaries;
an important option is to identify the molecules
passing through the detection volume based
on their fluorescence lifetime. Each molecule transit
is detected as a burst of fluorescence photons. Each
time such a transit is detected its fluorescence decay
time has to be determined.
Also in the area of single molecule detection and
spectroscopy, photon coincidence correlation techniques
were adopted to observe antibunching effects
that can be used to determine the number of observed
emitters as well as the fluorescence lifetime.
Figure 11 shows an example for such a coincidence
correlation experiment which can be performed with
a stopwatch type ­TCSPC instrument and either
pulsed or CW excitation since the laser sync is not
used in photon timing.
Another important method that makes use of temporal
photon density fluctuations over a wider time
range is Fluorescence Correlation Spectroscopy
(FCS). From the fluorescence intensity fluctuations
of molecules diffusing through a confocal volume,
one can obtain information about the diffusion constant
and the number of molecules in the observed
volume and calculate the molecular concentration.
This allows sensitive fluorescence assays based on
molecule mobility and interaction. Due to the small
numbers of molecules, the photon count rates in
FCS are fairly small. Therefore, the only practical
way of collecting the data is by means of single photon
counting. In order to obtain the time resolution
of interest for the diffusion processes, counting with
microsecond resolution is required. Hardware correlators
for FCS can be implemented very efficiently
and recent designs are widely used. However, these
instruments are dedicated to correlation with nanosecond
resolution at best, and cannot perform picosecond
­TCSPC.
The requirements of all these analytical techniques
based on single photon timing data have
much in common. Indeed, all of them can be implemented
with the same experimental set-up and are
based on photon arrival times. A first step towards
unified instrumentation for this purpose was a modification
of classical ­TCSPC electronics. The start-stop
timing circuitry is used as previously, providing the
required picosecond resolution for ­TCSPC. In order
to maintain the information embedded in the temporal
patterns of photon arrivals the events are stored
Figure 11: Coincidence correlation for investigation of bunching/antibunching effects.
12
as separate records. In addition to that, a coarser
timing (time tagging) is performed on each photon
event with respect to the start of the experiment. This
is called Time-Tagged Time-Resolved (TTTR) data
collection[5]
. Figure 12 shows a scheme of this data
acquisition mode where the start-stop events (the
so-called ­TCSPC times, t) are not sorted into a histogram
but stored directly along with an additional
time information – the time tag (T). This time tag represents
the macroscopic arrival time of the photon
with respect to the beginning of the experiment. Furthermore,
channel information (CH) can be recorded,
that depending on the set-up, determines the
wavelength or polarization of the detected photon.
To synchronize the data acquisition with the movement
of the scanner for, e.g., FLIM imaging, external
marker signals (M) from the scanner are incorporated
into the file format. This enables to reconstruct 2D
or even 3D images from the stream of TTTR records,
since the relevant XY position of the scanner can always
be determined during the data analysis. The
photon records are collected continuously and the
data stream can be processed immediately for display
and analysis. Thus, within the generated TTTR
file the complete photon dynamics are conserved
and no information is lost.
In classical TTTR the different time scales are processed
and used rather independently. However, it is
of great interest to obtain high resolution timing on
the overall scale, i.e., combining coarse and fine timing
into one global arrival time figure per event, with
picosecond resolution. In a most generic approach,
without implicit assumptions on start and stop events,
one would ideally just collect precise time stamps of
all events of interest (excitation, emission, or others)
with the highest possible throughput and temporal
resolution, and then perform the desired analysis on
the original event times (Figure 13).
Ideally this is done on independent channels, so
that between channels even dead time effects can
be eliminated. These requirements are met by the
TimeHarp  260, PicoHarp  300, and HydraHarp  400
­TCSPC systems from PicoQuant. Their radically new
design enables temporal analysis from picosecond
to second time scale, thereby covering almost all dynamic
effects of the photophysics of fluorescing molecules.
This is achieved by means of independent
TDC timing channels allowing picosecond cross-correlations
and very high throughput[6],[7],[8]
. In addition
to this enhanced functionality, the new designs eliminate
the need for operating in reverse start-stop
mode. TTTR mode is a fairly powerful method with
many aspects to consider, it is therefore covered in a
separate technical note[5]
.
One last feature of advanced TDC based ­TCSPC
systems to be covered here briefly is their multi-stop
capability. While TAC based systems can process
only one photon per excitation cycle, modern systems
with TDCs and independent channels can record
multiple photons per cycle, provided the cycle
Figure 12: TTTR (Time-Tagged Time-Resolved) data acquisition mode.
13
is longer than the dead time of the detector and the
TDC. This is quite commonly the case in luminescence
and phosphorescence measurements with
lifetimes in the range of micro- or milliseconds. The
TDC based system can then collect data much more
efficiently.All recent ­TCSPC systems from ­PicoQuant
support this multi-stop collection[6],[7],[8]
.
PicoQuant TCSPC Electronics and
System Integration
In addition to the fast timing electronics for acquisition
of, e.g., time-resolved fluorescence decays,
­PicoQuant provides pulsed diode lasers and other
light sources, bringing the technology for such measurements
to a degree of compactness and ease
of use unseen before. This permits the transfer of
revolutionary methods from the lab to real life industry
applications, e.g., in quality control or high
throughput screening. PicoQuant ­TCSPC systems
outperform conventional systems in many parameters.
Due to a versatile design they support many
useful measurement modes such as oscilloscope
mode for on-the-fly optical alignment or continuous
and time-tagging modes. Hardware synchronization
pins permit real-time scanning set-ups with sub-millisecond
stepping. DLL libraries as well as demo code
are available for custom programming or system integration.
A powerful data analysis software for timetagged
data is also available and is being improved
constantly to accommodate new methods and algo-
rithms.
Figure 13: Ideal instrument for TCSPC.
14
Copyright of this document belongs to PicoQuant GmbH. No parts of it may be reproduced, translated or transferred to third parties
without written permission of PicoQuant GmbH. All Information given here is reliable to our best knowledge. However, no responsibility
is assumed for possible inaccuracies or ommisions. Specifications and external appearances are subject to change without notice.
PicoQuant GmbH
Rudower Chaussee 29 (IGZ)
12489 Berlin
Germany
Phone	+49-(0)30-6392-6929
Fax	 +49-(0)30-6392-6561
Email	info@picoquant.com
WWW	http://www.picoquant.com
© PicoQuant GmbH, 2014
References and Further Reading
[1]	 Lakowicz, J. R. , “Principles of Fluorescence Spectroscopy”, 3rd Edition, Springer Science+Business
Media, New York, 2006
[2]	 O’Connor, D.V.O., Phillips, D., “Time-correlated Single Photon Counting”, Academic Press, London, 1984
[3]	 O’Connor, D.V.O., Ware, W.R., Andre, J.C.,“Deconvolution of fluorescence decay curves. A critical comparison
of techniques”, J. Phys. Chem. 83, 1333-1343, 1979
[4]	 Kalisz, J., “Review of methods for time interval measurements with picosecond resolution”, Metrologia,
41(1), 17-32, 2004
[5]	 Wahl, M., Orthaus, S., Technical Note on TTTR, PicoQuant, 2014
[6]	 Wahl M., Rahn H.-J., Gregor I., Erdmann R., Enderlein J., “Dead-time optimized time-correlated photon
counting instrument with synchronized, independent timing channels”, Review of Scientific Instruments, 78,
033106, 2007
[7]	 Wahl M., Rahn H.-J., Röhlicke T., Kell G., Nettels D., Hillger F., Schuler B., Erdmann R., “Scalable
time-correlated photon counting system with multiple independent input channels”, Review of Scientific Instruments,
79, 123113, 2008
[8]	 Wahl M., Röhlicke T., Rahn H.-J., Erdmann , R., Kell G., Ahlrichs, A., Kernbach, M., Schell, A.W., Benson,
O.: “Integrated Multichannel Photon Timing Instrument with Very Short Dead Time and High Throughput”,
Review of Scientific Instruments, 84, 043102, 2013