Methods of noise reduction

● White noise reduction: Impedance /

Temperature / Bandwidth reduction
● Shot noise reduction: Current reduction
● 1/f noise reduction: Band shifting
● Quantization noise reduction: Dithering
● High-frequency techniques

1
 Noise reduction
● For each type of noise:
● Thermal noise (Johnson) Vn2 = 4k BT R B
● Current noise (shot) I n2 = 2eIB
● Amplifier (flicker) noise (1/f)
● Quantization noise

● We can "attack" each part of these equations:

● For thermal noise: Reduce T, R or B

● For current noise: reduce I or B

● For 1/f noise: increase f

● For quantization noise: dithering 2

 White noise reduction: Impedance reduction (if possible)
Reducing the source impedance (if possible) gives two benefits:

● Reduction in thermal noise voltage

● Reduction in capacitive interference coupling (see later)

● Example: Using 20 kHz bandwidth, a 1 MΩ sample at room temperature generates

thermal noise of 20 µV. Reducing the resistance to 10 kΩ would reduce it to 2 µV.

● The best case: superconductors (Rs = 0),

metallic samples (low Rs)

● Sometimes this can not be done:

● Chemical electrodes (high Z)

● Biological probes (High Z)

● Semiconductors

● Photodetectors (High R) 3
 White noise reduction: Impedance matching (if possible)
In case of very low source impedance, and
only for AC, it's worth using an Impedance-
matching Transformer. It increases the signal
by the turns ratio N2/N1, and increases the
equivalent source resistance by the square of
the turns ratio N22/N12.

● The transformer doesn't change the thermal

SNR at the input: Thermal noise voltage
increases by the same ratio as the signal.

● However, the amplifier works at a higher

input resistance, so its NF is lower.

● For high frequencies, impedance matching is essential, to avoid reflections!

● 50Ω is the standard RF impedance (75Ω for video) 4

White noise reduction: Temperature reduction (if possible)
Reducing the source temperature (if possible) also reduces the thermal noise voltage

● Example: Using 20 kHz bandwidth, a 1 MΩ sample at room temperature generates

thermal noise of 20 µV. Reducing the temperature to 4.2 would reduce it to 2.5 µV.

● Can be used in laboratories: when studying samples at low temperatures

● Industrial uses (thermoelectric coolers): cooled photodetectors and image sensors,

cooled RF preamps, space electronics

● In most cases, temperature reduction

is not possible!

● Even worse: high-temperature electronics

5
 White noise reduction: Bandwidth reduction
As reducing the source resistance and temperature is not always possible, the main
tool for noise reduction is bandwidth reduction.

● This usually means: reduce noise by measuring more slowly

● This is equivalent to taking more measurements and averaging the results: for
random noise, we can reduce the fluctuations by a factor N if we average over N2
samples (measurement time will increase by N2). The same will happen if we
reduce the bandwidth by N2, as noise is proportional to N.

● If the sampling (or another factor) is limiting the signal bandwidth, there is no point
in having higher bandwidth at the amplifier. Therefore filtering should be used

● In some cases this is not practical, as in image taking:

● Video frame rate: 20 Hz. Picture size: 640x480 pixels. BW=20*640*480=6.1 MHz!

● CL image: 1s per spectrum, 100x100 pixels. Total time: 2.8 hr! Sample must be very stable…
6
 Bandwidth reduction: Passive, active and digital filters
The best way of bandwidth reduction is by low-pass filtering. This can be
done by different methods:
● Passive filter: made of RC or LC network, it's simple to install and can be added
almost anywhere in the circuit. However, usually it's limited to first-order (single
pole), so high-frequency reduction is limited (slope of 20 dB/decade)

● Active filter: usually part of the amplifier (sometimes the preamp) circuit, it can
provide better performance and more flexibility (even programmability) than the
passive one, at the cost of extra complexity.

● Digital filter: can be implemented in software, the signal from the A/D converter is
processed by algorithms which provide signal filtering (averaging) of any type. It's
easy to implement, but depends on good pre- and amplifier, ADC (should have
enough dynamic range). Analog filtering might be required for correct operation of
the ADC (Nyquist limit).
7
 Low-pass filter examples 1
Simple RC filter:
● To limit the bandwidth to 1 kHz, we can use a 100kΩ resistor with 1.6 nF capacitor.

The filter should be placed between a pre-amp and amp, to buffer the input and output

Ideal
Preamp Amp Cascade RC filters:
R not a great idea…
C

6 db/oct
0
1.E+00
-5 1.E+01 1.E+02 1.E+03 1.E+04 1.E+05

-10 f (Hz)
-15 12 db/oct
G (dB)

-20
-25
Slope: -20 dB/decade) log
-30
gain 18 db/oct
-35
frequency
-40 8
 Low-pass filter examples 2
Active filter: Same bandwidth - 1
kHz - but:
Slope: -60
● Has built-in input/output buffering
dB/decade)
● Has higher slopes

● Can have tailored frequency

response (slope, phase/delay
response, etc.)

● Same can be done in software

(digital filter)

Preamp Filter Amp

9
 Shot noise reduction
Shot noise is proportional to the square root
of DC (bias) current.

● In DC measurements, the relative S/N is

proportional to the DC current.

● For low frequency AC measurements, low

bias current reduces noise

● Bandwidth reduction is useful (same as in

white noise reduction)

● When individual electrons or photons need

to be counted, gated integration or multi-
channel scaling can be used (see below)
10
 1/f noise reduction: Band shifting
In many systems, the signal is DC. Noise reduction by bandwidth limiting is thus an
excellent solution: just apply a low-pass filter! However, there are two problems:

● Amplifier noise is usually highest at DC (1/f noise)

● DC interference (thermal emf, etc.) and drift can be important

The solution is then band shifting. The most widely used instrument to accomplish this
is the lock-in amplifier. The method consists of these steps:

● Transform the input signal to AC at a reference frequency (far from line freq.)

● Use a good AC preamplifier at this frequency (with extra line filtering)

● Filter the amplified signal at the reference frequency with the wanted bandwidth

● Transform the filtered amplitude again to DC

There are even commercial chopper-stabilized DC amplifiers, where an incoming DC

signal is internally switched on/off electronically, amplified and converted back to DC. 11
 Band shifting: How to get AC signals
● In many systems, passive (resistive) transducers provide the signal: strain gauges,
potentiometers, resistance thermometers, etc. In this case, it's enough to replace
the DC drive by an AC drive at the reference frequency.

● In other cases, the system is stimulated by an external excitation source: voltage or

current source in electrical measurements, light source in optical measurements,
etc. The source can be transformed to AC, or light can be chopped.

● There are even commercial chopper-stabilized DC amplifiers, where an incoming

DC signal is internally switched on/off electronically and then amplified and
converted back to DC.

12
 The Phase-sensitive detector (PSD)
● The PSD multiplies the signal (+noise) with a reference signal, either sinusoidal
(e.g. analog multiplier) or square (e.g switch). Here are the waveforms:

13
 How the PSD works
● To analyze the operation of the PSD we use the Fourier transform = spectrum

● Suppose we have a signal: Vssin(ωst) and a reference: Vrsin(ωrt+φ). Their

multiplication in the PSD creates sum and difference frequencies:
VsVr #
Vpsd = Vs sin(ω s t)⋅Vr sin(ω r t + φ ) = $cos ((ω s − ω r )t − φ ) − cos ((ω s + ω r )t + φ )%&
2
● A low-pass filter eliminates the sum frequency, as well as most of the noise.
VsVr
● In the special case where ωs = ωr , the output is DC: Vpsd = cos(φ )
2
In this case the output depends on V
● LPF
the relative phase between the signal
Noise
and reference (can be tuned to zero)

● The noise bandwidth is reduced ω

ωs-ωrENBW ωs ωr ωs+ωr
from ωs to the filter’s bandwidth: NEBW=1/4T.
14
 The benefits of using PSD
● Noise reduction: The low-pass filter
reduces the noise bandwidth from ωs
(typically kHz) to ENBW (usually <1Hz).

● More noise reduction: preamp works above

1/f noise.

● No drift: pure AC measurement

VsVr
● Phase measurements: Vpsd = cos(φ ) Vs
2 V’psd
● If we change the reference phase by π/2,
we get: Vpsd
' VsVr
= sin(φ )
2 Vpsd Vr
● Combining these, we can know the phase
relationship between signal and reference:
Vector measurement
15
 The Lock-in Amplifier
● Combines pre-amplifier, filters, PSD, reference generator … in one
programmable instrument.

Gain TC Display: X/Y, R/θ, etc. Ref. display

Inputs Filters Outputs: X/Y, R/θ, etc. Reference

(differential, In/Out
current)
16
 Two ways to measure…
● Large modulation: Large modulation Small modulation

excitation goes from 0 to 0.85

full value at each cycle.
0.65
Example: Light chopper
0.45
● Small modulation: small
0.25
excitation amplitude Signal
riding on slowly varying 0.05

"DC" signal sweep. 0 1 2 3 4

-0.15
Example: AC I/V curve
-0.35
measurements, showing Derivative
-0.55
dI/dV vs. V curves
-0.75
17
 Example of LIA use: photodiode
● We measure light with a Si photodiode. Main specifications: NEP=5.10-14 W.Hz-1/2,
rise time= 1 ns (BW=350 MHz), S = 0.5 A/W. Noise current: In=2.5.10-14 A.Hz-1/2.

● The DC resistance is: Vb/Id=8GΩ, giving

Johnson noise: Ijn=1.4.10-15 A.Hz-1/2

● The shot noise for bias current of 2.5nA:

Isn=2.8.10-14 A.Hz-1/2.

● The photodiode noise current is

dominated by shot noise! 18
 Example of LIA use
BW#(Hz) I(Johnson) I(shot) I#(1/f) In#(total)
● If we keep the nominal BW, total
1.0E+00 1.4E915 2.8E914 3.0E913 3.0E913
1.0E+02 1.4E914 2.8E913 1.0E915 2.8E913 diode noise: In=0.5nA, equiv.
2.0E+04 2.0E913 4.0E912 5.0E918 4.0E912 NEP=1nW, It’s a lot!
1.0E+06 1.4E912 2.8E911 1.0E919 2.8E911
3.5E+08 2.6E911 5.2E910 2.9E922 5.2E910 ● Using a low-pass filter with BW=1
1.0E909#
1.0E+00# 1.0E+03# 1.0E+06# 1.0E+09#
Hz, will reduce shot noise to
f"(Hz)"
In=2.8.10-14 A or NEP=5.6.10-14 W.
1.0E911# However, typical amplifier’s 1/f
noise could be 2-10 times higher!
I"(A)"

1.0E913#
I(Johnson)# ● With a lock-in amp, TC=250 ms,
I(shot)#
we can measure a signal at
I#(1/f)#
f=100Hz-100kHz, with 1Hz
1.0E915# In#(total)#
bandwidth, and keep the reduced
noise of the diode (2.8.10-14 A)
19
 Advanced lock-in techniques
● Sometimes, the signal is a Source
harmonic of the reference. 1
Vdefl
● Example: particle beam chopper,
2
fed by an AC voltage (reference).
1. Unipolar voltage: beam is blanked at
the reference frequency

2. Bipolar voltage: beam is blanked at Deflection

double the reference frequency plates
● In rare cases, two lock-in or other
noise reduction techniques are
Output slit
combined

20
 More noise is less: Dithering reduces quantization noise
To reduce quantization noise, Dithering the signal means adding some random (white)
noise, equivalent to 0.5 LSB of the DAC. The digitized signal is then averaged to
remove the noise. This removes the deterministic nature of the quantization.

Signal is smaller Signal+noise go

than 1 LSB: it's above the 1 LSB
always measured barrier: The
as 0 ! average value
(deterministic) corresponds to
the signal level
(fraction of 1
LSB!
21
 High-frequency (pulse) techniques
● In many experiments, measurements provide short pulses where the frequency,
height or area represent the data.

● Examples: Photomultipliers, particle detectors.

● Noise reduction is important because of the high frequencies involved.

● The methods are called multi-channel scaling, gated integration, boxcar averaging

22
Pulse example: lifetime measurements
● In this experiments, we measure the lifetime by measuring the time lapse between
a start pulse (from the excitation source) and a stop pulse (the signal from the
experiment). These are repeated many times to get a curve.

● First step: convert time interval to pulse height: Time-Amplitude Converter

● Second step: Scale the pulses into Multiple Channels according to their height,
producing a histogram of no. of pulses vs. height (which corresponds to time delay)

Start
(excitation)
TAC analog MSC
Stop
(signal) pulse

23
 Pulse example: noise calculations
● Noise: random pulses mixed with the signal, giving random contribution to all
channels. If they were constant – no problem: subtract a fixed number! However,
noise has fluctuations, which can blur the signal. For N2 pulses, fluctuations are ≈N

● Noise reduction: noise pulses are divided equally to all channels, signal pulses
accumulate in a specific channel.

● Result: with time, we accumulate more data than noise fluctuations.

● Example: suppose we have 1 signal pulse and 100 noise pulses /sec

● After 1 sec: 101 pulses, fluctuations = 10 pulses, S/N=1/10 (cannot see the signal)

● After 10,000 sec: 1010000 pulses, fluctuations=1000 pulses, S/N=10000/1000 (can

see signal).

● This is actually a bandwidth-narrowing technique!

24
 Gated integration (boxcar averaging)
● Another method to reduce noise: the gated integrator

● First step: a gate allows the signal to pass only during the excitation time

● Second step: an integrator accumulates the signal during a long measurement time

● Dual benefit for noise reduction:

● Noise is acquired only during the signal period (gate on): big improvement for low duty cycles

● Integration reduces noise.

● Schematic diagram of the system:

25
 Gated integration: how it works
● In this example, 1µs excitation pulse width and 100µs period.

● Gating the signal at 1µs reduces the noise bandwidth by 100 and noise by 10

● Further integration of many repetitions can reduce the noise bandwidth even more

● The final (DC) output measures the signal at the specific time point (e.g. excitation)

● Final noise reduction is:

Tint
tg

26
 Gated integration: variable delay
● To trace the full waveform, the gate delay is sweeped (slowly!) so that at each point
the gate samples one point of the signal.

27
 Why “Boxcar” ?
● The American name of merchandise wagons, because of their box-like shape
● Can imagine a “train” of pulses of the same shape…

● Or the gate pulse as the door!

The Boxcar and the LIA

● The lock-in amplifier can be regarded as a
special case of the boxcar integrator, with:

● 50% pulse width

● Phase = delay/period

28