I intend to write a function which calculates the result and the error for any formula with any amount of variables using the Gauß Error Propagation.

The error $\mathrm{d}R$ for a function $A(a,b,c)$ and the corresponding errors $\mathrm{d}a, \mathrm{d}b$, and $\mathrm{d}c$ are calculated with


Since there is huge amount of formulas one could wish to use with different amounts of variables, the function has to be adaptive.

My first shot is the following function, which already does the job pretty well.

Gaus[φ_, data_?MatrixQ] := (
  Do[Evaluate[Symbol["x" <> ToString[i]]] = data[[i, 1]], {i, Length[data]}];
  α = Norm[Derivative[##][φ][Sequence @@ 
         Evaluate[Table[Symbol["x" <> ToString[i]], {i, Length[data]}]]] &
            @@@ IdentityMatrix[Length[data]]*data[[All, 2]]];
  Α = φ[Sequence @@ Evaluate[Table[Symbol["x" <> ToString[i]], {i, Length[data]}]]];
  Print[{N[Α], N[α]}]

It accepts a function $\phi$ with a layout like

R[a_, b_, c_] := a*b/c

and a list of {value,error} pairs. Then it creates an corresponding amount of intermediate variables x1,...,xn with n being the number of pairs. These are used to calculate the error $\alpha$ and the value A, which are then printed.

And now to my problem. I really want the intermediate variables xi and $\alpha$ and A to be local. Currently they are global, which isn't a problem for the function, as existing variables are overwritten, but proved to be annoying in other calculations.

My first try was to use Module similar to this

Module[{Evaluate[Table[Symbol["x" <> ToString[i]], {i, 2}]], a, b}, a + b]    

Module::lvsym : "Local variable specification {Evaluate[Table[Symbol["x" <> ToString[i]], {i, 2}]], a, b} contains Evaluate[Table[Symbol["x" <> ToString[i]], {i, 2}]], which is not a symbol or an assignment to a symbol.

which produces only the error message above. I could of course define the variables x1 to x10 directly, as I will rarely have more than ten values per formula, but this would be terribly inelegant.

Several futile tries later I am running out of possibilities to rephrase my search, since I neither am a native speaker nor have significant programming background.

So I would be happy, if someone could provide an idea to create adaptive amount of local variables.

Update: Thanks to the answer of Jens about the use of downvalues, i could modify my code to the following.

Gaus[ϕ_, data_?MatrixQ] := Module[{df, f, x},
  Do[x[i] = data[[i, 1]], {i, Length[data]}];
  df = Norm[Derivative[##][ϕ] [Sequence @@ Array[x, Length[data]]
        ] & @@@ IdentityMatrix[Length[data]]*data[[All, 2]]];
  f = ϕ[Sequence @@ Array[x, Length[data]]];
  Print[{N[f], N[df]}]

For some reason I had to get rid of the Elavaluate[..] Function, although, from what I understand of it, there no reason for an error. This basically answers my introductory question.

The comment of acl and a night of sleep made me realize that the approach is indeed wrong, as the task can be achieved without using intermediate variables just using sequence, which somehow eluded me.

Gauss[ϕ_, data_?MatrixQ] := Module[{df, f},
  df = Norm[Derivative[##][ϕ] [Sequence @@ data[[All, 1]]
        ] & @@@ IdentityMatrix[Length[data]]*data[[All, 2]]];
  f = ϕ[Sequence @@ data[[All, 1]]];
  Print[{N[f], N[df]}]

This seems to me like much cleaner and easier function.


4 Answers 4


Perhaps what you're looking for is something like this:

Module[{x, a, b},
 x[1] = 1;
 x[2] = 10;
 a + b/x[2] + x[1]


Here I defined a single additional local variable x but then refer to "indexed" variables sharing the same name and differing only in the index x[1] and x[2] etc. These indices are actually a way of storing arbitrarily many values under a single name x, as DownValues. All of them are then local to the module in which x was localized.

The fact that variables are local is indicated in the above output line by the appearance of the random-looking dollar suffixes on the variable names a and b (the variable x had received values, so you don't se their internal names). This never has to concern you because the local variables would usually not be output like I did in this example. This is only to show what goes on internally to the Module.

You don't need any table constructs or loops of any kind to get a flexible number of variables this way.

But if you do want to define, say, a list of the x[i] over some range of i, you can use a construct like xList = Array[x, 2] for a 2-element list xList whose entries would be {x[1], x[2]}.

Another possible approach, especially if you do insist on using a table to define your variables, would be to use Unique, but I don't see any advantage for it in your application.


There is an interesting discussion of how local variables are actually realized in Mathematica below the question "Local variables in Module leak into the Global context".

  • $\begingroup$ I'd go for x/@Range[2]. But yes I agree and, as I say in my answer, this is almost certainly the wrong approach. $\endgroup$
    – acl
    Commented Dec 21, 2012 at 22:46
  • $\begingroup$ @acl I just think Array is more readable, so why not use it? $\endgroup$
    – Jens
    Commented Dec 21, 2012 at 22:50
  • $\begingroup$ Certainly; I'm not implying it's better to use Map. I find x/@Range[2] more readable myself, for some reason. $\endgroup$
    – acl
    Commented Dec 21, 2012 at 22:54
  • $\begingroup$ Thank you for your answer, i wasn't aware of the possibility of downvalues. I just assumed that I would need a new symbol for every value. I adjusted my code accordingly The track, which lead to this approach started with the problem, that the derivate function doesn't accept lists as arguments and my first idea was to use variables as intermediate. In addition the comment from acl about the wrong approach got me thinking and I realized, that the sequence solves the problem, avoiding the variables. $\endgroup$
    – Crown42
    Commented Dec 22, 2012 at 7:55

Rather than answering your question as posed, let me instead save you the effort of writing such a function and at the same time demonstrate how it can be done by posting some code that I've already written for this purpose:


Unprotect[var, cov]; ClearAll[var, cov];
SetAttributes[var, HoldAll];
SetAttributes[cov, {HoldAll, Orderless}];

SetAttributes[PropagateCovariance, HoldAll];
Options[PropagateCovariance] = {
   "ExpansionOrder" -> 1, 
   "InitialCovarianceMatrix" -> Automatic


cov[x_Symbol, x_Symbol] := var[x];

   exprs : {Except[_List] ..}, vars : {__Symbol},
  ] :=
     partialTrace = Map[Tr, #, {2, ArrayDepth[#] - 1}] &,
     covarianceMatrix = If[# === Automatic,
       Outer[cov, vars, vars],
       #] & @ OptionValue["InitialCovarianceMatrix"]
    With[{contracted = partialTrace@D[exprs, {vars, #}]},
      1/#! contracted.MatrixPower[covarianceMatrix, #].Transpose[contracted]
    ] & /@ Range@OptionValue["ExpansionOrder"] // Total

(* Deal with univariate case *)
   expr_, var_Symbol,
   opts : OptionsPattern[]
  ] :=
  PropagateCovariance[expr, {var}, opts];

(* If a scalar expression is given, return a scalar rather than
   a 1×1 covariance matrix *)
   expr : Except[_List], vars : {__Symbol},
   opts : OptionsPattern[]
  ] :=
  PropagateCovariance[{expr}, vars, opts][[1, 1]];


Protect[var, cov];


I'll confess that I've been waiting for the right question to arrive so that I can post this, but I think you'll find it helpful. It provides for the propagation of arbitrary covariance matrices, thus dealing properly with correlated errors, and you can easily find the covariance matrix of multiple values calculated from the same set of quantities. Also, I provide for an approximation of arbitrary order--while in practice greater than second-order approximations are seldom useful, one can certainly very often do much better than the conventional first-order approximations.

Let's try it:

err = PropagateCovariance[a*b/c, {a, b, c}] // Simplify
(* -> (1/(c^4))(2 a b c^2 cov[a, b] - 2 a b^2 c cov[a, c] - 
                2 a^2 b c cov[b, c] + b^2 c^2 var[a] + a^2 c^2 var[b] + 
                a^2 b^2 var[c]) *)

This may look somewhat complicated, but is the right answer (to first order) when a, b, and c are correlated. The second-order expression is even more involved, so I won't reproduce it here, although you can obtain it using

PropagateCovariance[a*b/c, {a, b, c}, "ExpansionOrder" -> 2] // Simplify

if you so wish.

Let us now suppose that a and b are not correlated with each other, and nor are b and c:

 b /: cov[_, b] = 0;
 err2 = err // Simplify
(* -> (-2 a b^2 c cov[a, c] + b^2 c^2 var[a] +  a^2 (c^2 var[b] + b^2 var[c]))/c^4 *)

Or, perhaps we wish to assume that none of the variables are correlated, and want to find the standard deviation rather than the variance (which will give the same result as your method):

Block[{a, b, c},
 cov[a, b] ^= 0;
 c /: cov[_, c] = 0;
 Sqrt[err] /. var[x_] :> σ[x]^2 // Simplify
(* -> Sqrt[(b^2 c^2 σ[a]^2 + a^2 (c^2 σ[b]^2 + b^2 σ[c]^2))/c^4] *)

This result can also be obtained by the following approach, which may even save a little effort in the calculation:

 a*b/c, {a, b, c},
 "InitialCovarianceMatrix" -> DiagonalMatrix@Thread@σ[{a, b, c}]^2
] // Simplify
(* -> Sqrt[(b^2 c^2 σ[a]^2 + a^2 (c^2 σ[b]^2 + b^2 σ[c]^2))/c^4] *)

Or, let's say we wish to find the covariance matrix between $a b/c$ and $a/b + c^2$ (again assuming uncorrelated errors and a first-order approximation, so that the result is not too unwieldy):

 {a*b/c, a/b + c^2}, {a, b, c},
 "InitialCovarianceMatrix" -> DiagonalMatrix@Thread@σ[{a, b, c}]^2
] // Simplify

$\left( \begin{array}{cc} \frac{\left(\text{var}(c) b^2+c^2 \text{var}(b)\right) a^2+b^2 c^2 \text{var}(a)}{c^4} & \frac{b^2 \text{var}(a)-a \left(2 \text{var}(c) b^3+a \text{var}(b)\right)}{b^2 c} \\ \frac{b^2 \text{var}(a)-a \left(2 \text{var}(c) b^3+a \text{var}(b)\right)}{b^2 c} & \frac{\text{var}(b) a^2}{b^4}+\frac{\text{var}(a)}{b^2}+4 c^2 \text{var}(c) \\ \end{array} \right)$

To show the effect of using approximations of different orders, let us now consider the function $f(x) = \sin \log x^2$. Accurate errors can be obtained numerically using a Monte Carlo approach, and these will be our benchmark. We can obtain a first-order approximation numerically using Mathematica's built-in significance arithmetic, i.e.:

Sqrt@var[x] 10^(Accuracy[x] - Accuracy@f[x])

Or, analytically:

Sqrt@PropagateCovariance[f[x], x]
(* -> 2*Sqrt[(Cos[Log[x^2]]^2*var[x])/x^2] *)

But, as we can see, it's somewhat less than perfect (Monte Carlo values in blue):

Plot of first-order error versus Monte Carlo result

The second-order expression is much better, though, failing only when the gradient of $f$ is large enough that the approximation that the errors are Gaussian begins to break down:

Plot of second-order analytic error versus Monte Carlo result

The remaining failings won't be improved by taking an expansion to higher order, but could be addressed with appropriate consideration of the higher-order moments. Unfortunately, in physics, we very rarely have access to the complete covariance matrix, let alone the coskewness and cokurtosis tensors, so I think we can be satisfied for most purposes with the current level of approximation.

Finally, a demonstration of the importance of considering correlation. Taking $g(x,y) = \sqrt{x^2-y^2}$ with strongly correlated errors on $x$ and $y$, we obtain the following plot:

Plot of errors calculated using diagonal vs. full-rank correlation matrices

Here, the green curve is given by the full second-order expression for the error, while the red one represents the outcome of the same procedure but with the covariance matrix approximated as diagonal. Again, the points represent the Monte Carlo errors. Obviously, neglecting the off-diagonal terms of the covariance matrix can give rise to an egregiously wrong result.

  • 1
    $\begingroup$ (+1) All that work has to be rewarded, of course. $\endgroup$
    – Jens
    Commented Dec 22, 2012 at 2:41
  • $\begingroup$ A suggestion though: it would be helpful to point out where in the code you solve the problem of defining a set of variables whose number is adjustable. The construct of Block and With is probably worth explaining to make it more useful to the OP or for people searching for the words in the title... $\endgroup$
    – Jens
    Commented Dec 22, 2012 at 2:47
  • $\begingroup$ @Jens because I use a matrix formalism I avoid the need to define any symbols explicitly, so I didn't solve that problem here. I wasn't too keen on either the OP's approach or their implementation and so I decided to address the wider issue using idiomatic Mathematica code, but without actually answering the question directly. I hope that the OP and anyone else looking for error propagation code will find this more useful in the long run. $\endgroup$ Commented Dec 22, 2012 at 3:02
  • $\begingroup$ Thanks for your detailed answer, although I struggling a bit to understand it. I'm not studying physics but am required to do a few chosen experiments to see and learn the basics. So I chose the opportunity to do things, I could probably do by hand or with excel, with mathematica in order to learn it. But I am interested in learning, so I will try to work your solution out. $\endgroup$
    – Crown42
    Commented Dec 22, 2012 at 8:13
  • 1
    $\begingroup$ +1 This is great. Something like this has been on my to-do list for ages, now I can cross it off :-) Thanks for the early Christmas present, Oleksandr! $\endgroup$ Commented Dec 22, 2012 at 13:56

Mathematica 12 introduced a new built-in way to handle error propagation, using the Around function:

Around[a, da]*Around[b, db]/Around[c, dc]


$\frac{a b}{c} \pm \sqrt{a^2 \left(b^2 \left( \frac{\text{dc}}{c^2}\right) ^2+\frac{\text{db}^2}{c^2}\right)+\frac{b^2 \text{da}^2}{c^2}}$

You can use numeric values, or vectors of numeric values as well.


The last code snippet could be made to work like this

   {vars = Evaluate[Table[Symbol["x" <> ToString[i]], {i, 2}]~Join~{a, b}]},
   Module[vars, a + b+x1]

which replaces vars with the right thing in Module[vars,body].

I don't think using x1 and the like is optimal though.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.