Question

Another developer has given me an algorithm that returns a series of strings that contain stringified doubles. I am building unit tests against these string outputs. About 80% of the time I run my unit tests they all pass. The other 20% of the time slight variations occur in the double portion of the returned strings. For example:

Expected: ((B,D),(C,A)); : 0.05766153477579324
Found:    ((B,D),(C,A)); : 0.05766153477579325

Expected: (B,(C,(A,D))); : 0.0017518688483315935
Found     (B,(C,(A,D))); : 0.001751868848331593

I know that double computations can be imprecise, but I have never heard of them being variant. I am assured by the algorithm author that the algo is deterministic. The way that the double is being toStringed is:

    for(Tree gt: geneTrees){
        double prob = probList.next();
        total += prob;
        result.append("\n" + gt.toString() + " : " + prob);
    }

I am at a bit of a loss explain how this variation is possible. Any ideas?

Answer 1

Just based on the summing that you’re doing, I suspect that this might be caused by the issue that double addition is not quite commutative or associative — you’ll get different rounding errors if you add doubles in a slightly different order.

Just add a small epsilon for your unit tests, basically.