0% found this document useful (0 votes)
42 views

Introduction To Measurement Uncertainty and Error Analysis

This document introduces students to the concepts of measurement uncertainty and error analysis in physics experiments. It explains that the scientific method involves making observations, developing theories, and designing experiments to test those theories. Well-designed experiments are important for advancing scientific understanding. The document also discusses how comparing experimental results to accepted theoretical values or other studies allows researchers to evaluate their results and determine if their measurements agree with existing models or require further refinement. Calculating the percentage error between measured and expected values provides a meaningful way to assess experimental accuracy and determine if results are consistent or inconsistent with accepted physics principles.
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
42 views

Introduction To Measurement Uncertainty and Error Analysis

This document introduces students to the concepts of measurement uncertainty and error analysis in physics experiments. It explains that the scientific method involves making observations, developing theories, and designing experiments to test those theories. Well-designed experiments are important for advancing scientific understanding. The document also discusses how comparing experimental results to accepted theoretical values or other studies allows researchers to evaluate their results and determine if their measurements agree with existing models or require further refinement. Calculating the percentage error between measured and expected values provides a meaningful way to assess experimental accuracy and determine if results are consistent or inconsistent with accepted physics principles.
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

Introduction

 to  Measurement  Uncertainty  and  Error  Analysis  


 
Welcome  to  university  physics.    
 
What  you’ll  discover  in  your  courses  is  that  science  builds  up  a  “model”  describing  how  the  world  
and  the  universe  work.  If  we  understand  the  model  then  we  can  make  predictions  as  to  how  things  
will  behave,  and  we  can  design  complex  machines  and  systems.  If  we’re  clever,  our  understanding  of  
science  may  have  meaningful  impact  on  the  health  and  welfare  of  the  people  and  the  planet  around  
us.  Knowledge  is  important  and  powerful.  
 
The  foundation  of  this  picture  of  how  things  are  related  stems  from  observation,  theory,  and  
experimentation:  the  scientific  method.  When  we  observe  what  we  perceive  to  be  a  relationship  
between  two  properties  (or  variables)  of  a  system  we  might  hypothesize  that  the  relationship  takes  
on  a  particular  mathematical  form.  Ideally  we  can  then  take  measurements,  controlling  one  variable  
while  watching  the  second.  If  our  observations  fit  our  theory  then  we’ve  succeeded  in  coming  to  
some  form  of  deep  understanding.  If  the  results  don’t  match  our  expectations,  then  our  theory  needs  
to  be  modified  and  improved.  By  this  process  of  refinement  we  improve  our  understanding.  Our  
experiments  are  a  critical  part  of  the  scientific  method,  and  as  such,  we  need  to  ensure  we  go  about  
performing  and  reporting  them  in  ways  that  are  thoughtful,  repeatable  and  verifiable,  and  convincing  
to  others.  The  laboratory  component  of  a  science  program  serves  this  important  purpose.    
 
1.  Comparisons  to  Theory  or  Accepted  Values  -­‐  Percentage  Error  
 
Before  we  step  into  the  minutia  of  measurements  and  calculations,  we  should  take  a  moment  to  state  
an  important  reality  of  experimental  physics:  measurements  and  calculations,  in  and  of  themselves,  
are  rarely  interesting.  
 
What  is  interesting  though  is  when,  by  virtue  of  our  measurements  and  calculations,  we  can  compare  
a  result  to  an  accepted  value  or  validate  a  theoretically  predicted  value.    
 
By  way  of  example,  if  we  have  devised  an  experiment  and  measured  the  acceleration  of  a  falling  
object  to  be  9.78±0.05m/s2,  this  is  a  fine  outcome.  We  have  made  our  measurement,  and  we  have  
conscientiously  quantified  the  uncertainties  in  our  measurements  such  that  we’re  confident  our  
result  is  correct  to  ±0.05m/s2,  possibly  due  to  instrumentation  limitations  or  slight  variations  in  our  
results  from  one  trial  run  to  the  next.  However,  our  measurement  lacks  any  meaningful  context  by  
itself.  Our  result  might  apply  to  one  object  only,  and  for  all  we  know,  other  objects  could  fall  at  
completely  different  acceleration  rates.    
 
If  however  we  consider  Newton’s  law  of  gravitation,  we  realize  that  at  the  surface  of  the  earth  we  
expect  that  
 
g  =  -­‐GM/r2  =  9.80665m/s2,  
 
since  G  (the  gravitational  constant),  M  (the  mass  of  Earth),  and  r  (the  radius  of  Earth)  are  already  
known  to  quite  good  accuracy  and  precision.1  We  can  report  (happily)  that  our  result,  within  the  
limitations  of  our  experimental  uncertainty,  is  “consistent”  with  the  accepted  value  of  g.  A  calculation  
of  “Percentage  Error,”  where  
 

                                                                                                               
1  Precision  refers  to  the  number  of  significant  figures  in  a  measurement.  In  general,  the  better  the  

resolution  of  our  tools  or  the  more  samples  we  can  measure,  the  better  our  precision.  Accuracy  
meanwhile  refers  to  how  “realistic”  our  measurements  are.  Systematic  errors  like  poorly  calibrated  
instruments  mean  that  we  might  achieve  results  with  good  precision  but  poor  accuracy.    
Measured − Accepted
%Error = ×100% ,  
Accepted
 
can  be  useful  to  convey  the  accuracy  of  our  result,  showing  how  closely  it  matches  accepted  values  or  
theoretical  predictions.  
 
Additionally,  our  uncertainty  (±0.05m/s2)  is  deemed  “insignificant”  since,  within  the  range  of  our  
uncertainty,  our  result  overlaps  the  accepted  value.  Given  our  apparently  successful  result,  we  might  
therefore  extend  our  study  to  objects  of  other  masses,  or  go  further  and  study  objects  subject  to  
additional  forces,  like  air  resistance.  
 
If,  on  the  other  hand,  we  had  measured  the  acceleration  due  to  gravity  to  be  7.5±0.7m/s2  then  we  
must  report  that  our  result  is  “inconsistent”  with  the  commonly  accepted  value  of  the  acceleration  
due  to  gravity.    
 
In  this  second  case,  the  discrepancy  between  our  result  and  the  accepted  value  in  our  gravity  
example  is  “significant”  because  our  result  is  quite  different  than  the  expected  value,  even  at  the  
extremes  of  our  range  in  uncertainty.  The  percentage  error  when  compared  against  the  accepted  
value  is  approximately  23%.  This  situation  would  probably  lead  us  to  re-­‐evaluate  our  calculations  
(were  mistakes  made?),  our  methodology  (does  our  experiment  have  flaws  we  overlooked?),  and  our  
tools  (are  systematic  problems  affecting  our  results,  e.g.  slow/fast  timers,  or  poorly  calibrated  length  
scales?).  We  may  need  to  rethink  our  experiment.  Or  perhaps  we  find  ourselves  in  the  interesting  
situation  that  our  experiment  is  correct  and  we  have  discovered  a  flaw  in  the  theory  we’re  
investigating  (many  unexplained  observations  and  results  have  eventually  led  to  historic  scientific  
revolutions!).    
 
Finally  we  should  consider  a  case  where  our  experiment  tells  us  that  acceleration  due  to  gravity  is,  
say,  11±2m/s2.  In  this  case  our  result  is  “consistent”  with  the  accepted  value  of  9.80665m/s2,  
however  our  percentage  error  (~12%)  and  uncertainty  (±2m/s2)  are  both  relatively  large.  In  some  
scenarios,  depending  on  the  experimental  conditions,  this  might  be  a  completely  appropriate  and  
acceptable  outcome,  but  it  might  also  indicate  that  improvements  in  methodology  or  instrumentation  
are  required  to  better  our  precision  and  confirm  the  accuracy  of  our  result.  We  would  then  need  to  
consider  whether  efforts  at  improving  our  experiment  would  be  warranted  and/or  successful.  
 
2.  Commonalities  Between  Experiments  
 
While  most  of  our  experiments  will  touch  on  a  variety  of  unique  and  apparently  unrelated  topics,  
they  will  generally  have  in  common  that  we  will  compare  a  result  to  an  accepted  value  or  validate  a  
result  predicted  by  theory,  as  in  the  example  scenarios  above.  This  serves  as  an  effective  way  to  learn  
to  experiment,  and  eventually  students  may  find  themselves  involved  in  research  investigating  
entirely  unknown  and  new  science.  
 
3.  Measurements  and  Uncertainty    
 
Let  us  now  begin  to  formalize  our  approach  to  dealing  with  experimental  uncertainties,2  
mathematically  speaking.  We  have  three  important  tasks  ahead:  
 

                                                                                                               
2  Note,  in  an  experimental  context,  the  term  “uncertainty”  is  often  used  interchangeably  with  the  
term  “error.”  Neither  term  refers  to  an  error  on  the  part  of  the  researcher,  but  rather  to  the  
quantifiable  limitations  of  the  experimental  tools  or  methods.  We  will  usually  use  the  term  
“uncertainty”  in  this  document,  but  you  might  encounter  both  in  your  studies.    
1) We  must  establish  how  uncertainty  arises  in  experiments.  We’ll  see  that  in  most  experiments  
there  are  various  factors  contributing  to  uncertainty,  and  depending  on  the  scenario,  some  
may  be  significant  while  others  might  be  negligible.    
2) We  must  learn  to  quantify  uncertainty  in  our  experiments.  To  some  extent  this  will  require  
common  sense,  both  to  minimize  uncertainties  where  possible  and  to  assign  them  
“reasonable”  values.  We’ll  address  the  three  main  types  of  uncertainties  as  we  proceed.  
3) We  must  finally  determine  how  to  correctly  carry,  or  “propagate,”  these  uncertainties  
through  our  calculations.  Our  experiments  will  often  require  directly  measuring  two  (or  
more)  properties  of  a  system  in  order  to  calculate  a  third  dependent  property  (e.g.  by  
hanging  a  known  mass  from  a  spring,  k=F/x  we  can  “indirectly”  determine  the  spring’s  
constant,  k,  if  we  can  “directly”  measure  the  displacement,  x,  caused  by  the  known  force,  
F=mg).  We  will  need  to  know  how  the  uncertainties  in  the  measured  values  carry  through  to  
affect  the  final  result.  
 
 
Notation  
 
In  your  classes  you  should  notice  that  you  encounter  examples  almost  always  dealing  in  numbers  
that  are  exactly  defined.  As  such,  the  solution  to  a  problem  usually  has  a  single  “correct”  answer:  
a+b=c,  or  xy=z.  This  is  a  convenient  way  to  teach  theory  and  to  practice  calculations.  However,  this  
tendency  to  deal  in  exact  numbers  risks  that  students  develop  an  unrealistically  “clean”  perception  of  
the  science  they’re  studying.    
 
In  the  practical  reality  of  experimental  physics  we  must  deal  with  the  fact  that  measurements  are  
never  known  with  absolute  certainty.  No  matter  the  effort  applied,  even  the  best  experimentalist  
must  contend  with  unavoidable  limitations  in  how  confidently  they  can  express  a  measurement.    
 
Therefore  when  we  measure  a  mass,  we  require  a  notation  to  capture  both  our  measurement  and  
also  this  “uncertainty”  in  our  measurement.  In  real  terms  we  must  express  our  result  not  as  an  exact  
value  “m,”  but  rather  as  a  best  effort  at  measurement  of  m,  with  an  uncertainty  recorded  to  indicate  
our  confidence  in  our  measurement.  That  is,  the  preferred  form  for  recording  m  would  be  
 
      m  =  m(best  measurement)  ±  Δm,  
 
or,  even  more  succinctly,  we  simply  record  the  mass  as  m±Δm.  Similarly,  lengths  would  be  expressed  
as  L±ΔL,  times  expressed  a  t±Δt,  momentum  as  p±Δp,  etc.  
 
This  serves  as  a  useful  notation  when  dealing  with  individual  measurements  and  also  with  larger  sets  
of  data,  and  in  all  labs  this  will  be  our  preferred  way  of  recording  any  measurement.    
 
4.  Sources  of  Uncertainty  
 
Uncertainty  in  experiments  arises  from  three  basic  sources:  reading  uncertainties,  random  
uncertainties,  and  systematic  uncertainties.  As  a  scientist,  our  job  is  to  recognize  where  each  type  
enters  into  an  experiment,  and  to  minimize  their  influence  to  the  best  of  our  ability.  Ideally  we  can  
make  two  out  of  the  three  negligibly  small  compared  against  their  more  influential  counterpart  (that  
is,  we  identify  which  of  the  three  sources  of  uncertainty  most  affects  our  measurement).  We  then  
simply  quantify  the  source  of  uncertainty  and  propagate  it  through  any  calculations  to  arrive  at  our  
final  result.  
 
4.1  Reading  Uncertainty  
 
Frequently  we  encounter  situations  where  our  instrumentation  is  a  limiting  factor  in  our  ability  to  
make  a  measurement.    
 
Because  most  instruments  have  an  incremented  scale,  when  a  measurement  falls  between  the  
smallest  marked  increments  (“ticks”)  it’s  often  appropriate  and  acceptable  to  estimate,  or  
“interpolate,”  this  final  digit.  
 
For  example,  if  a  ruler  has  millimeter  markings,  the  length  of  an  object  might  
be  observed  to  be  slightly  greater  than,  say,  45mm  but  definitely  less  than  
46mm.  In  this  case  we  can  look  closely  and,  by  our  best  effort,  we  decide  that  
the  object’s  length  is  very  close  to  45.4mm.    
 
However,  we’re  also  cognizant  that  because  we’re  essentially  estimating  our  
final  digit  we  may  be  slightly  in  error,  and  so  we  include  in  our  recorded  
measurement  an  estimate  of  our  uncertainty.    
 
By  common  convention,  we  often  choose  to  use  half  the  value  of  the  smallest  
increment  on  our  measuring  instrument  as  our  estimate  of  the  uncertainty.  So  
in  the  case  of  our  ruler  measurement  with  millimetre  markings,  we  would   Figure  1.  
record  our  result  as  45.4±0.5mm.     Measurement  
  interpolation.  
Note  though  while  the  “half  the  smallest  division”  convention  is  generally  a  
good  rule  of  thumb,  we  may  still  encounter  cases  where  it  would  overestimate  our  uncertainty  
significantly.  This  tends  to  occur  if  our  instrument’s  divisions  are  widely  spaced.  If  a  ruler  provided  
only  centimeter  divisions  for  instance,  then  we  might  quite  reasonably  estimate  our  measurement  to  
an  uncertainty  of  ±0.1cm  rather  than  the  more  conventional  ±0.5cm.  To  use  ±0.5cm  simply  by  
convention  would  in  this  case  be  overly  imprecise,  which  would  affect  any  subsequent  calculations.  
We  always  need  to  apply  good  judgment  in  order  to  achieve  best  results.  
 
Parallax  
 
When  reading  measuring  instruments  with  scale  markings  
we  must  also  be  careful  to  read  from  directly  in  front  of  the  
instrument’s  scale  in  order  to  avoid  “parallax.”  Parallax  can  
occur  if  the  experimenter  moves  their  head  to  either  side  of  
the  plane  perpendicular  to  the  scale  and  the  object.  This  can  
cause  a  visual  angular  misalignment  between  the  object  and  
the  measuring  scale,  such  that  the  experimenter  misreads  or  
mis-­‐interpolates  their  measurement.    
 
Digital  Displays  
  Figure  2.  Parallax  adds  reading  error.  
Electronic  measuring  instruments  frequently  come  equipped   Measuring  tools  should  be  viewed  
with  digital  readouts  or  will  display  results  to  a  computer.  As   "straight  on."  
such,  we  need  to  be  able  to  record  our  results  correctly  when  
using  these  tools.    
 
One  needs  to  be  somewhat  careful  in  deciding  on  the  appropriate  uncertainty  to  record  when  using  a  
digital  display.  For  example,  if  an  ammeter  display  reads,  say,  10.51mA  then  it  could  be  reasonable  to  
believe  the  worst  case  uncertainty  would  be  plus  or  minus  the  smallest  digit,  i.e.  10.51±0.01mA.  
However,  this  may  not  always  be  the  case.  Depending  on  the  equipment,  a  more  representative  
recording  of  the  measurement  might  actually  be  10.510±0.005mA  (you  would  need  to  check  the  
instrument  labeling  or  manual  to  confirm  the  applicable  convention).  And  with  some  
instrumentation,  you  might  find  a  label  indicating  that  the  reading  is  “accurate  to  ±1%  of  full  scale,”  
(such  that  in  the  case  of  our  ammeter,  if  the  full  scale  covers  20mA,  then  our  uncertainty  would  be  
1%  of  20mA,  or  0.2mA,  leading  to  a  reading  of  10.5±0.2mA).  Despite  these  variations  in  reading,  with  
a  little  effort  you’ll  be  able  to  identify  the  most  appropriate  way  to  record  your  measurement  and  its  
uncertainty.  
 
Finally  we  mention  that  in  some  cases  the  reading  from  a  digital  display  may  be  “noisy”  and  fluctuate  
to  some  degree.  (This  can  and  does  occur  for  analog  displays  as  well,  as  when  a  needle  on  the  display  
jumps  about  erratically).  This  can  be  a  result  of  random  variations  influencing  the  experiment  or  the  
measuring  instrument  itself,  such  as  electrical  noise,  thermal  fluctuations,  etc.  We’ll  see  next  that  
random  fluctuations,  if  non-­‐negligible,  may  require  us  to  increase  our  estimates  of  uncertainty  to  
values  larger  than  those  justified  by  “reading  uncertainty”  alone  (that  is,  our  “half  the  smallest  
division”  convention  or  “±1  digit”  on  a  digital  display  may  significantly  underestimate  the  realistic  
value  of  our  uncertainty).  We’ll  also  see  that  for  such  cases,  if  we  can  repeat  the  measurement  several  
times,  statistical  analysis  will  provide  a  more  methodical  mathematical  way  of  determining  our  
measurement  and  uncertainty.  
 
4.2  Random  Uncertainty  
 
While  “reading  uncertainty”  is  applicable  and  useful  when  the  object  or  property  we  are  measuring  is  
reasonably  well  defined,  in  many  cases  the  “thing”  we  are  measuring  is  naturally  “fuzzy.”  In  another  
sense,  if  we  have  difficulty  defining  when  or  where  our  measurement  should  start,  repeating  the  
measurement  may  give  slightly  varying  results;  there  is  an  imprecision  in  our  technique  that  is  
difficult  to  eliminate,  so  despite  best  efforts,  some  of  our  measurements  are  slightly  too  large  or  too  
small.  In  such  cases  the  instrument  reading  uncertainty  may  become  negligible  compared  to  larger  
“random”  fluctuations  in  the  object  or  property  being  measured.  As  such,  where  random  
uncertainties  dominate  the  measurement  we  must  determine  how  to  assign  a  more  appropriate  best  
estimate  of  uncertainty.  
 
There  are  a  few  cases  where  such  random  uncertainties  become  dominant  in  a  measurement.    
 
Problem  of  definition  
 
Consider  the  difficulty  in  using  a  lens  to  focus  a  sharp  image  of  an  object  onto  a  screen.  According  to  
the  “thin  lens  equation”  we  expect  that  
 
1 1 1
+ =  
d d' f
 
where  d  is  the  distance  between  object  and  lens,  d’  is  the  distance  between  image  and  lens,  and  f  is  
the  lens  focal  length.  If  d  is  large,  1/d~0  and  therefore  f~d’  so  the  image  location  should  equal  the  
lens’s  focal  length.  However  in  practice,  the  image  formed  by  the  lens  may  look  focused  and  sharp  as  
the  position  of  the  screen  is  adjusted  over  many  millimetres.    
 
This  can  be  a  case  where  the  image  distance  is  inherently  difficult  to  define,  so  we  encounter  an  
unavoidable  randomness  in  our  selection  of  where  we  believe  we  observe  the  sharpest  image  
location.  As  such,  even  though  we  might  measure  the  distance  between  lens  and  image  using  a  ruler  
with  millimeter  markings,  a  larger  uncertainty  of  ±0.5cm  or  even  more  might  be  the  appropriate  and  
more  representative  value.  (You  can  see  the  difficulty  for  yourself  by  using  a  magnifying  lens  to  focus  
the  image  from  your  window  onto  a  paper  screen  in  a  darkened  room.)  
 
Similarly,  consider  the  case  where  we  are  trying  to  measure  the  period  of  a  pendulum.  Using  a  
stopwatch  we  count  out  the  time  required  for  ten  oscillations,  and  then  divide  accordingly.  We  might  
discover  that  the  time  required  for  ten  oscillations  is  12.62s  and  conclude  that  the  period  of  a  single  
oscillation  is  1.262±0.005s,  and  we  might  be  quite  happy  with  our  result.  However,  if  we  then  repeat  
the  measurement  perhaps  we  would  find  that  in  our  second  attempt,  ten  oscillations  take  12.28  s,  so  
that  one  oscillation  takes  1.228±0.005s.  Of  course  we  notice  the  “significant”  discrepancy  between  
our  results  (they’re  quite  close,  but  their  uncertainties  don’t  overlap)  and  become  concerned.    
 
On  closer  consideration  we  probably  realize  that  our  ability  to  start  the  stopwatch  at  exactly  the  
beginning  of  an  oscillation  and  to  stop  at  exactly  the  end  of  an  oscillation  are  somewhat  error-­‐prone.  
There  is  in  fact  a  degree  of  randomness  that  affects  how  well  we  start  and  stop  our  timer.  We  repeat  
the  measurement  several  more  times  and  observe  periods  of  1.262,  1.228,  1.211,  and  1.254  s.  We  
realize  we  can  average  these  readings  such  that  our  most  likely  oscillation  period  is  
 
T  =  [1.262+1.228+1.211+1.254]/4  =  1.239s  
 
and  we  could  reasonably  estimate  the  uncertainty  to  be  ±0.003s  based  on  the  range  we  see  in  our  
value,  giving  a  best  measurement  of  T=1.24±0.03s  per  oscillation.  (Keep  in  mind  that  the  stopwatch  
may  not  represent  the  only  significant  source  of  error  in  such  an  experiment,  and  we  should  remain  
mindful  and  observant  lest  we  overlook  another.  Perhaps  the  way  in  which  we  set  the  pendulum  into  
oscillation,  its  amplitude,  or  even  air  friction,  influence  the  result.  The  possibilities  are  broad  and  
very  experiment-­‐dependent.)  
 
Statistical  Determination  of  Uncertainty  
 
In  such  experiments  or  studies  where  there  are  many  samples  available  to  measure,  or  where  the  
same  measurement  can  be  made  many  times,  statistical  analysis  allows  us  a  more  mathematical  
approach  for  calculating  our  best  measurement  and  uncertainty  value  than  the  interpolation  or  best  
estimates  we’ve  opted  to  use  thus  far.    
 
Luckily,  in  many  cases  we’ll  encounter  in  physics,  we  can  indeed  repeat  our  measurements.  If  we  can  
take  the  same  measurement  several  times,  random  uncertainties  will  cause  some  of  our  
measurements  to  be  slightly  too  large  or  slightly  too  small.  The  set  of  measurements  will  exhibit  a  
“distribution”  about  the  most  probable  “best  value.”  As  a  result,  when  averaged,  we  expect  the  
random  uncertainties  to  cancel  out,  providing  a  very  good  and  representative  final  result.    
 
We  can  calculate  the  average,  or  “mean”  value,   x ,  as  
follows:  
 
n

x1 + x2 +... + xn
∑ xi
x= = i .  
n n
 
We  can  gain  a  measure  of  the  average  variance  of  
our  data  from  the  mean  by  calculating  the  “standard  
deviation,”   σ ,  according  to:   Figure  3:  Probability  distribution  for  a  set  of  
  measurements  subject  to  random  
n uncertainties,  showing  mean  value  and  
standard  deviation.  
∑(x − x )i
2

σ= i
.  
(n −1)
 
This  standard  deviation  indicates  the  average  separation  of  a  data  point  from  the  mean.  Note  that  as  
the  number  of  data  points  recorded  increases,  this  average  value  should  remain  relatively  constant.      
 
Finally,  as  our  measure  of  uncertainty  in  the  mean  value  of  our  data  set,  we  calculate  the  “standard  
deviation  of  the  mean,”   σ m ,  given  by:  
 
σ
σm = .  
n
 
Here  we  note  that  as  we  increase  the  number  of  measurements  recorded,  n,  the  standard  deviation  of  
the  mean  decreases  (thus  the  precision  with  which  we  know   x  increases).  So  by  taking  more  
measurements,  we  are  able  to  increase  our  overall  precision.  Note  though  that  while  useful  to  a  point,  
the  technique  can  become  time  consuming  and  yields  diminishing  returns,  as  for  instance,  to  increase  
σ m  by  a  factor  of  10  we  must  increase  the  number  of  measurements  by  100.    
 
So,  in  summary,  given  a  set  of  data  from  a  repeatable  measurement  subject  to  random  variations,  our  
best  result  is  given  by:  
 
σ
      x ± Δx = x ± σ m = x ± .  
n
 
Given  the  option,  we  should  always  repeat  a  measurement  several  times  and  use  the  mean  and  
standard  deviation  of  the  mean  as  our  values  of  best  measurement  and  uncertainty,  respectively.  
This  provides  the  most  rigorous  method  available  to  confidently  and  precisely  determine  our  result,  
whereas  when  forced  to  interpolate  or  estimate  the  risk  of  significantly  over-­‐  or  under-­‐estimating  
our  measurement  is  much  greater.    
 
4.3  Systematic  Uncertainty  
 
Systematic  uncertainties  are,  unfortunately,  often  the  most  difficult  to  identify  and  account  for  in  an  
experiment.    
 
Typical  examples  would  be  calibration  errors  in  our  instruments:  for  instance,  a  clock  that  runs  too  
fast,  or  a  ruler  whose  markings  are  stretched  or  scaled  incorrectly.  To  acquire  a  set  of  measurements  
using  a  stopwatch  that  consistently  runs  too  slowly  for  instance,  we  would  of  course  still  encounter  
the  random  fluctuations  that  govern  the  precision  of  our  estimate  (or  statistical  calculation)  of  
uncertainty.  But  rather  than  acquiring  a  set  of  data  distributed  about  the  true  value,  all  
measurements  would  skew  slightly  to  smaller  readings  than  should  properly  be  the  case.  The  result  
might  be  a  precise  looking  result,  but  one  that  is  ultimately  inaccurate.    
 
Similarly,  subtle  effects  like  air  pressure  or  temperature  changes  in  the  laboratory  might  have  
significant  influence  on  the  outcome  of  measurements,  unbeknownst  to  the  experimenter.  
 
Even  experimenter  bias  might  contribute  to  introduce  systematic  uncertainty.  Think  of  parallax,  
described  earlier  as  the  tendency  to  misread  an  instrument’s  scale  if  viewed  from  an  angle.  Normally  
the  experimenter  would  take  readings  from  directly  in  front  of  the  scale,  but  could  be  forgiven  for  
introducing  small  random  variations  if  the  head  or  eye  position  shifts  slightly  left  or  right  (with  
relatively  equal  tendency).  To  consistently  read  the  scale  from  an  angle  always  left  of  straight  on  
would  tend  to  skew  the  readings.  This  bias,  or  ones  like  it,  should  be  identified  and  corrected.  
 
Overall,  it  can  be  quite  difficult  to  compensate  for  systematic  uncertainties.  Better-­‐calibrated  tools  
(or  recalibrating  the  tools  available  against  certified  calibration  standards)  can  improve  results,  as  
can  attention  to  details,  like  parallax,  that  might  be  surprisingly  influential.    
 
5.  Propagation  of  Uncertainty  
 
As  you  can  see,  recording  a  good  and  reliable  measurement,  though  not  necessarily  difficult,  does  
require  awareness  of  where  one  risks  over-­‐  or  under-­‐estimating  uncertainty.  Interpolating  or  
estimating  random  variations  are  sometimes  necessary,  but  if  done  badly  can  significantly  mis-­‐
represent  the  actual  uncertainty,  giving  a  result  with  unrealistic  precision.  Statistical  analysis  
provides  the  most  methodical  and  mathematically  rigorous  approach  available  for  identifying  a  
reliable  and  precise  mean  value  (best  measurement)  and  standard  deviation  of  the  mean  
(uncertainty).  However,  systematic  uncertainties,  if  not  identified  and  minimized,  could  still  render  a  
highly  precise  effort  at  a  measurement  completely  inaccurate.  Now  that  we’re  aware  of  these  
possible  pitfalls  though,  we  can  endeavor  to  be  both  precise  and  accurate,  which  is  worthwhile  if  we  
want  our  experiments  to  be  successful.    
 
So  what  we  have  achieved  thus  far  is  the  knowledge  needed  to  acquire  a  set  of  good  measurements.  
We  should  be  able  to  minimize  or  eliminate  avoidable  sources  of  error,  and  we  can  therefore  record  
our  results  with  realistic  uncertainties.    
 
5.1  Uncertainty  Propagation  for  Addition  and  Subtraction  
 
Often  in  experiments  we’ll  be  taking  direct  measurements  and  performing  calculations  to  determine  
(indirectly)  a  property  of  interest.  Whereas  in  our  “clean”  textbook  examples,  a+b=c  or  xy=z,  we  must  
now  develop  a  formalism  for  handling  the  more  realistic  calculations  where,  for  instance,  
 
(a ± Δa) + (b ± Δb) = (c ± Δc)   and   (x ± Δx)(y ± Δy) = (z ± Δz) .  
 
We  begin  by  deriving  Δc  for  the  case  of  addition  of  two  numbers  that  include  uncertainties.  If  we  
consider  the  extremes  a+Δa  and  b+Δb,  then  we  can  regroup  as  follows:  
 
(a + Δa) + (b + Δb) = (a + b) + (Δa + Δb)
      .  
= c + Δc
 
Similar  consideration  applies  to  the  opposite  extreme  where  we  use  a-­‐Δa  and  b-­‐Δb.  
 
(a − Δa) + (b − Δb) = (a + b) − (Δa + Δb)
.  
= c − Δc
 
Thus,  in  the  case  of  addition,    
 
(a ± Δa) + (b ± Δb) = (a + b) ± (Δa + Δb)
 
= (c ± Δc)
 
and  we  see  clearly  that  our  uncertainties  can  be  expressed  as    
 
Δc ~ Δa + Δb .  
 
An  identical  result  can  be  derived  for  subtraction,  where    
 
(a ± Δa) − (b ± Δb) = (c ± Δc) .  
 
We  should  note  that  this  expression  “slightly”  overestimates  Δc  because  statistically  our  average  
uncertainties  Δa  and  Δb  will  be  somewhat  less  than  the  limiting  value  we  used  in  our  illustration.  We  
instead  use  the  expression  
 
Δc = Δa 2 + Δb 2 .  
 
We  can  see  that  the  expression  is  mathematically  equivalent  to  the  
Pythagorean  theorem  governing  the  lengths  of  the  sides  of  right  angle  
triangles.  We  can  therefore  appreciate  how  the  calculated  uncertainty,  Δc,  is  
always  less  than  the  sum  of  the  added  uncertainties  (i.e.  the  hypotenuse  is  
always  less  than  the  sum  of  the  sides  of  a  right  triangle).  
  Figure  4.  When  
  uncertainties  are  
5.2  Uncertainty  Propagation  for  Multiplication  and  Division   added,  the  resulting  
  uncertainty  is  
Whereas  in  the  case  of  addition  and  subtraction,  when  dealing  with   calculated  by  
uncertainties  we  dealt  with  “absolute”  uncertainties,  in  the  case  of   “quadrature.”  
multiplication  and  division,  we  need  to  deal  in  “fractional”  uncertainties.    
 
Fractional  uncertainties  are  expressed  in  the  form  Δa/|a|,  where  |a|  is  the  “absolute  value”  of  a.  
 
We  follow  a  similar  approach  to  our  derivation  of  the  propagation  of  uncertainty  in  the  case  of  
addition.  Consider  the  quotient  
 
(x ± Δx)(y ± Δy) = (z ± Δz) .  
 
As  before,  we  consider  the  uncertainty  extremes,  where  x  is  given  by  x+Δx  and  y  is  given  by  y+Δy.  
 
" Δx %" Δy %
(x + Δx)(y + Δy) = xy $$1+ ''$$1+ ''
# x &# y &
" Δx Δy ΔxΔy %
= xy $$1+ + + '    
# x y x y '&
" " Δx Δy %%
~ xy $$1+ $$ + ''''
# # x y &&
 
(Since  Δx  and  Δy  should  be  small,  ΔxΔy~0.  )  
 
Similarly,  considering  the  limiting  case  of  x-­‐Δx  and  y-­‐Δy,  
 
# Δx &# Δy &
(x − Δx)(y − Δy) = xy %%1− ((%%1− ((
$ x '$ y '
# Δx Δy ΔxΔy &
= xy %%1− − + (  
$ x y x y ('
# # Δx Δy &&
~ xy %%1− %% + ((((
$ $ x y ''
 
Combining  the  two  results  and  substituting  z  for  xy,  then  we  conclude  that,  
 
 
" " Δx Δy %%
(x ± Δx)(y ± Δy) = z $$1± $$ + ''''  
# # x y &&
 
such  that  
 
Δz " Δx Δy %
        = $ + '' .  
z $# x y &
 
 
Finally,  we  mention  that  by  using  the  extremes  of  our  uncertainty  range  (i.e.  x+Δx,  x-­‐Δx)  in  our  
derivation,  the  formula  given  above  again  “slightly”  overestimates  Δz.  A  slightly  more  precise  version  
is  given  by    
 
2 2
Δz " Δx % " Δy %
= $ ' + $ ' .  
z # x & # y &
 
Using  a  similar  derivation,  the  same  result  is  also  found  to  apply  in  the  case  of  division.  
 
5.3  Summary  of  Propagation  of  Uncertainty  
 
Addition  and  Subtraction:  
 
For  
 
  (a ± Δa) + (b ± Δb) = (c ± Δc)     or   (a ± Δa) − (b ± Δb) = (c ± Δc) ,  
 
Δc  is  given  by  
 
        Δc = Δa 2 + Δb 2 .  
 
Multiplication  and  Division:  
 
For  
 
(x ± Δx)
(x ± Δx)(y ± Δy) = (z ± Δz)     or     = z ± Δz ,  
(y ± Δy)
 
Δz  is  given  by  
 
2 2
Δz " Δx % " Δy %
= $ ' + $ ' .  
z # x & # y &
 
We  now  have  the  important  tools  required  to  navigate  uncertainty  propagation  in  most  scenarios  
we’ll  encounter  in  our  experiments.    
 
Before  we  proceed  though  we  should  take  a  moment  to  review  how  we  carry  significant  figures  
through  our  calculations.  
 
 
6.  Significant  Figures  and  Calculation  Methods  Involving  Uncertainties  

6.1  Significant  Figures  


 
Much  as  when  we  described  the  importance  of  correctly  reading  your  measuring  tool,  researchers  
must  also  be  conscientious  in  correctly  carrying  significant  figures  through  calculations  performed  
from  their  measurements.  Several  standardized  conventions  apply.  
 
Definition:  In  any  given  number,  the  significant  figures  include  all  digits  that  are  considered  reliable,  
as  well  as  the  first  digit  considered  doubtful  (i.e.  the  digit  defining  our  uncertainty).  Furthermore,  we  
ignore  zeros  acting  as  placeholders.  
 
Examples  –  Determining  Significant  Figures  
 
480491  kg     à   Six  significant  figures  
480500  kg   à   Four  significant  figures  (last  two  zeros  are  placeholders)  
400000  kg   à   One  significant  figure  (last  five  zeros  considered  placeholders)  
480.491  g   à   Six  significant  figures  
480.500  g   à   Six  significant  figures    
(including  last  two  zero  decimals  implies  they  are  considered  significant)  
0.0003489  kg   à   Four  significant  figures  (all  zeros  before  3  are  placeholders)  
0.00034  kg   à   Two  significant  figures  
0.034  kg   à   Two  significant  figures  
1.23  g       à   Three  significant  figures  
0.12  g     à   Two  significant  figures    
(leading  zero  in  this  case  is  a  placeholder,  included  by  convention)  
0.120  g     à   Three  significant  figures    
(including  the  final  zero  in  this  case  implies  that  it  is  considered  significant)    

6.2  Calculations,  Taking  into  Consideration  Significant  Digits  and  Uncertainty  Propagation    
 
It  is  particularly  important  to  properly  account  for  significant  figures  when  performing  calculations,  
so  that  results  are  not  stated  with  too  great  or  too  little  precision.  
 
Rule  1:  Significant  Figures  when  Adding  or  Subtracting    
 
When  adding  or  subtracting  numbers,  the  position  of  the  least  significant  digit  in  the  result  is  the  
same  as  the  position  of  the  last  significant  digit  in  the  least  precise  number  being  added  or  
subtracted.  
 
Simple  Examples:  
4.983 1.2873
+3.587 −0.484
        3.14159 − 2.718 + 42.0 = 42.4  
8.570 0.803
 
Experimental  Example:  
 
Say  we  want  to  add  the  masses  4.51±0.04  kg,  575.62±0.09  g,  and  2.1±0.5  kg.  First,  convert  to  similar  
units,  and  then  proceed  with  the  calculation  as  follows,  
4.51 Δm = (Δm1 )2 + (Δm2 )2 + (Δm3 )2
0.57562 = (0.04)2 + (0.00009)2 + (0.5)2
+2.1
          = (1.6 ⋅10 −3 ) + (8.1⋅10 −9 ) + (0.25)
7.18562
= 0.2516000081
= 0.501
= 0.5
The  last  significant  digit  in  each  value  in  the  summation  is  indicated  by  the  underscore.  We  see  that  
since  2.1  kg  is  the  least  precise  value  in  our  calculation,  its  least  significant  figure  determines  the  
overall  number  of  significant  figures  in  our  final  result.  As  such,  we  round  up  and  report  that   m  =  
7.2±0.5  kg.  
 
Rule  2:  Significant  Figures  when  Multiplying  and  Dividing  
 
When  multiplying  or  dividing  a  set  of  numbers,  the  final  result  should  include  only  as  many  
significant  digits  as  the  number  in  the  set  with  the  fewest  significant  digits.  
 
Simple  examples:  
 
1.25 × 2.0 = 2.5       (Not  2.50!)  
 
4.882
= 2.441 = 2.4     (Round  off  to  two  significant  figures.)  
2.0
 
1.72
5.286 ⋅ = 9.091738... = 9.09     (Round  off  to  three  significant  figures.)  
1.00002
Experimental  Example:  
 
Say  we  intend  to  divide  force   F  =  8.24±0.05  N  by  mass   m  =  2.251±0.006  kg  in  order  to  determine  
acceleration  in  accordance  with  Newton’s  Second  Law,  where   a = F / m .  Then  
 
2 2
Δa " ΔF % " Δm %
= $ ' +$ '
a # F & # m &
2 2
" 0.05 % " 0.006 %
= $ ' +$ '
F # 8.24 & # 2.251 &
a=
m = (3.68⋅10 −5 ) + (7.10 ⋅10 −6 )
8.24
=           = 0.0066  
2.251
Δa = a(0.0066)
= 3.6605
= 3.6605(0.0066)
= 0.024
 
 
Since  we  have  three  significant  figures  in   a (i.e.   a =3.66),  we  round   Δa  to  0.02.  Therefore
a = 3.66 ± 0.02 .  
 
Note  that  in  our  calculation  of   a ,  the  number  of  significant  digits  retained  in  the  result  is  the  same  as  
the  number  of  significant  digits  of  the  least  precise  value  entered  in  the  calculation  (in  this  case,  8.24  
N,  which  has  only  three  significant  digits  versus  the  four  of   m ).  
 
6.3  Scientific  Notation  (aka  “Standard  Form”)  
 
Scientific  notation  makes  use  of  powers  of  ten  to  adjust  the  position  of  the  decimal  point.  It  has  two  
advantages:    
 
i) Convenience  and  compactness.    
 
The  mass  of  an  electron  is  9.11×10-­‐31  kg.  Compare  this  value  with  
0.00000000000000000000000000000000000911  kg.  Try  entering  the  latter  into  a  
calculator!    
 
ii) Elimination  of  ambiguity  regarding  precision.    
 
Suppose  we  see  the  statement:  “The  average  distance  between  Earth  and  Sun  is  
150,000,000  km”.  Are  the  zeros  significant  digits  or  are  some  of  them  used  only  as  
“spacers”?  By  writing  the  value  as  1.50×108  km,  we  are  clearly  implying  a  precision  of  
three  significant  figures,  if  such  is  actually  warranted.  
 
 
References  
 
J.R. Taylor, An Introduction to Error Analysis (University Science Books, Mill Valley, California, 1982).
 
 
 
 
   
Experiment:  Identification  of  Metal  by  Determination  of  Mass,  Volume,  and  Density    
 
In  this  experiment  you  will  determine  the  mass,  volume,  and  density  of  several  metal  blocks  using  a  
variety  of  measuring  instruments  and  techniques.  You  will  pay  particular  attention  to  recording  
results,  documenting  your  uncertainties,  and  propagating  any  uncertainties  through  calculations.    
 
Theory  
 
Metals  and  metal  alloys  (mixtures  of  pure  metals)  often  have  well  defined  densities.  By  measuring  
the  volume  and  mass  of  a  piece  of  metal,  density,  ρ,  can  be  calculated  according  to    
 
m
            ρ= ,  
V
 
where  m  is  mass  and  V  is  volume.    
 
When  the  shape  of  a  piece  of  metal  has  a  well-­‐defined  geometry,  the  volume  may  be  determined  
using  rulers  or  similar  tools  to  measure  critical  dimensions  necessary  for  volume  calculation.    
 
When  the  shape  of  a  piece  of  metal  does  not  have  a  well-­‐defined  geometry,  it  may  still  be  possible  to  
determine  its  volume  using  alternative  methods.  Submersion  in  water  can  be  used  to  determine  
metal’s  volume  indirectly  by  measuring  the  volume  of  the  water  that  is  displaced  during  the  
submersion  process.      
 
Apparatus  
 
2  metal  blocks           ruler  
triple  beam  balance           Vernier  calipers  
graduated  cylinder           beaker  
 
Procedure  
 
Two  blocks  of  different  metals  will  be  provided.  Ensure  that  you  can  distinguish  which  is  which.    
 
Determination  of  Mass  
 
The  mass  of  each  block  will  be  determined  using  a  triple  beam  balance.  Ensure  that  you  read  and  
correctly  record  the  appropriate  uncertainty  value.    
 
Determination  of  Volume  using  Ruler  and  Vernier  Caliper  
 
Determine  the  volume  of  block  1  using  a  standard  ruler  or  meter  stick  (1mm  divisions),  taking  care  
to  ensure  that  the  uncertainty  for  each  measurement  is  recorded  correctly.    
 
Determine  the  volume  of  block  2  using  a  Vernier  caliper.    If  unsure  as  to  how  to  read  a  Vernier  
caliper,  see  Appendix.  Take  care  to  ensure  that  the  uncertainty  for  each  measurement  is  recorded  
correctly.    
 
Determination  of  Volume  by  Displacement  of  Water  
 
Submerge  block  1  in  a  beaker  of  water,  such  that  the  water  covers  the  top  of  the  block.  Mark  the  
height  of  the  water  level  using  a  piece  of  masking  tape  and  a  pen.  Remove  the  block  from  the  beaker  
of  water.  Fill  a  graduated  cylinder  to  a  level  of  60  or  100  mL.    
 
Record  the  volume  of  water  in  the  graduated  cylinder,  being  careful  to  correctly  record  uncertainty.  
Carefully  pour  water  from  the  graduated  cylinder  back  into  the  beaker  until  the  water  rises  to  the  
original  water  level  when  the  block  was  submerged.  Record  the  amount  of  water  remaining  in  the  
graduated  cylinder,  and  use  your  results  to  calculate  the  volume  of  the  block.    
 
Repeat  for  block  2.    
 
Results  
 
Based  on  your  measurements,  determine  the  density  of  each  block  as  determined  from  each  
measurement  method.  Provide  a  table  showing  your  results.    
 
In  your  report,  provide  examples  for  each  type  of  calculation  required  to  determine  the  block  
densities.  
 
Identify  the  type  of  metal  each  block  is  made  of  by  referencing  the  table  of  accepted  densities  listed  in  
the  Appendix.  Provide  a  calculation  of  Percentage  Error.    
 
In  your  discussion,  comment  on  the  strengths  and  weaknesses  of  each  technique.    
 
 
 
 
 
 

   
Appendix  I  –  Reading  a  Vernier  Caliper:  
 
Vernier  calipers  are  a  clever  modification  on  a  ruler  that  allows  for  higher  precision  readings  than  
the  ruler’s  scale  by  itself  can  provide.  They  are  particularly  useful  for  measuring  the  dimensions  of  
precision  machined  components.    
 

 
 
By  design,  as  the  jaws  of  the  calipers  slide  over  a  distance  of  1  mm,  the  “Vernier  scale’s”  markings  
align  sequentially  with  the  ruler  markings  above,  such  that  if  the  jaws  are  at  0.2  mm  then  the  2nd  
Vernier  scale  marking  will  be  aligned  with  the  ruler  marking  above,  and  if  the  jaws  are  0.5  mm  the  5th  
Vernier  marking  will  be  aligned  with  the  ruler  marking  above,  etc.  
 
We  demonstrate  with  a  series  of  examples:  
 
 

 
(Note  that  the  zeros  on  the  ruler  and  the  Vernier  scale  align,  and  also  that  the  10th  marking  on  the  
Vernier  scale  corresponds  to  the  39th  mm  on  the  ruler.  This  “scaled”  grating  results  in  sequential  shift  
in  marking  alignment  as  the  jaws  are  slid  open  or  closed.)  
 
As  we  slide  the  Vernier  scale  (say  we’re  measuring  the  diameter  of  a  fibre  optic  cable)  we  might  have  
the  following:  
 

 
 
Notice  that  it  would  be  quite  difficult  to  read  the  exact  distance  the  Vernier  scale’s  zero  has  shifted  
right  of  the  ruler’s  zero  position  by  eye  alone.  We  could  estimate  the  reading  to  be  somewhere  
around  0.1  mm  to  perhaps  0.4  mm.  Reasonable,  but  we  can  do  better…  
 
As  we  clamp  the  fibre  in  our  caliper  jaw,  we  see  that  the  Vernier  scale  actually  slid  along  the  ruler  
such  that  its  “3”  aligns  nicely  with  a  marking  on  the  ruler.  None  of  the  other  Vernier  scale  markings  
align  very  well  with  the  ruler  markings  above.  As  such,  this  tells  us  that  the  fibre  diameter  is  0.30  mm  
(that  is,  the  zero  on  the  vernier  scale  has  moved  3/10ths  of  1  mm  from  the  zero  on  the  ruler  scale).  
And  since  we  assign  uncertainty  of  half  of  the  smallest  division,  which  is  0.05  mm  on  our  Vernier  
scale,  the  uncertainty  is  ±0.03  mm  (rounded  up  from  0.025).  So  the  measurement  of  the  fibre  
diameter  is  therefore  0.30±0.03  mm.  
 
Try  another…  
 

 
Again,  you  can  see  that  it  would  be  difficult  to  estimate  directly  by  eye  that  the  Vernier  scale’s  0  is  
0.50  mm  from  the  ruler’s  0  mm  marking,  but  we  see  quite  clearly  that  the  5  on  the  Vernier  scale  
aligns  with  a  marking  on  the  ruler  above  (whereas  the  other  Vernier  markings  aren’t  well  aligned  
with  the  ruler  markings),  so  the  reading  is  0.50±0.03  mm.  
 
Another  example…  
 

 
The  reading  would  be  5.40±0.03  mm  (perhaps  even  5.37±0.03  mm,  since  it’s  a  little  hard  to  say  
whether  3.5  or  4  align  better  with  the  ruler  markings  above).  
 
Try  some  on  your  own…  
 

 
Solution:  _____  ±  _____  mm  
 

 
Solution:  _____  ±  _____  mm  
 
 
Appendix  II:  
 
Useful  conversion  factor:  1  mL  =  10-­‐6  m3  
 
Various  densities1:  
 
Aluminum   2.702x103  kg/m3  
Copper  
  8.92x103  kg/m3  
Iron     7.86x103  kg/m3  
Lead     11.3437x103  kg/m3  
Tin     7.28x103  kg/m3    
 
Brass                  8.47x103  kg/m3  2  
Acrylic                1.19x103  kg/m3  2  
 
 
References  
 
1The  Chemical  Rubber  Company  (1970).  Handbook  of  Chemistry  and  Physics.  51st  Ed.  Cleveland:  The  

Chemical  Rubber  Co.  


2Oberg,  E.,  et  al.,  Machinery’s  Handbook  (2012).  29th  Ed.New  York:  Industrial  Press.  

You might also like