THnBase::ProjectionAny does not calculate errors correctly

[jgrosseo]

Check duplicate issues.

• Checked for duplicates

Description

It seems to me that THnBase::ProjectionAny does not respect the user's choice of calculating the error correctly even when "E" is given as an option. Please find a minimal reproducer attached which compares the behaviour of TH2 -> TH1 (correct) to THn -> THn (not correct).

Please let me know if I am overlooking something.

Reproducer

void thn()
{
    auto hist = new TH2F("hist", "", 10, 0.5, 10.5, 10, 0.5, 10.5);
    hist->Sumw2();

    for (int i = 0; i < 9; i++)
        hist->Fill(1, 1, 0.1);

    float relError = hist->GetBinError(1, 1) / hist->GetBinContent(1, 1);
    printf("TH2:  %f +- %f --> %f\n", hist->GetBinContent(1, 1), hist->GetBinError(1, 1), relError);

    auto proj1 = hist->ProjectionX();
    relError = proj1->GetBinError(1, 1) / proj1->GetBinContent(1, 1);
    printf("Proj: %f +- %f --> %f\n", proj1->GetBinContent(1), proj1->GetBinError(1), relError);

    int bins[] = {10, 10, 10, 10};
    double xmin[] = {0, 0, 0, 0};
    double xmax[] = {10, 10, 10, 10};
    auto thn = new THnF("thn", "", 4, bins, xmin, xmax);
    thn->Sumw2();

    for (int i = 0; i < 9; i++)
        thn->Fill(0.5, 0.5, 0.5, 0.5, 0.1);

    double values[]{0.5, 0.5, 0.5, 0.5};
    auto binNo = thn->GetBin(values);
    relError = thn->GetBinError(binNo) / thn->GetBinContent(binNo);
    printf("\nTHn:  %f +- %f --> %f\n", thn->GetBinContent(binNo), thn->GetBinError(binNo), relError);

    Int_t dimensions[] = {0, 1};
    // Despite the option "E", the errors are resetted to sqrt(N) instead of keeping the original ones
    THnBase *proj = thn->ProjectionND(2, dimensions, "E");

    binNo = proj->GetBin(values);
    relError = proj->GetBinError(binNo) / proj->GetBinContent(binNo);
    printf("Proj: %f +- %f --> %f\n", proj->GetBinContent(binNo), proj->GetBinError(binNo), relError);
}

The output is

TH2:  0.900000 +- 0.300000 --> 0.333333
Proj: 0.900000 +- 0.300000 --> 0.333333

THn:  0.900000 +- 0.300000 --> 0.333333
Proj: 0.900000 +- 0.948683 --> 1.054093

while I believe all four lines should have identical numbers.

ROOT version

v6-32-06

Installation method

ALICE cvmfs installation

Operating system

Linux Ubuntu 24

Additional context

No response

[jgrosseo]

I believe the following could do to fix it:

diff --git a/hist/hist/src/THnBase.cxx b/hist/hist/src/THnBase.cxx
index 92b4e5c636..e6fa0525c1 100644
--- a/hist/hist/src/THnBase.cxx
+++ b/hist/hist/src/THnBase.cxx
@@ -661,6 +661,10 @@ TObject* THnBase::ProjectionAny(Int_t ndim, const Int_t* dim,
    Bool_t haveErrors = GetCalculateErrors();
    Bool_t wantErrors = haveErrors || (option && (strchr(option, 'E') || strchr(option, 'e')));

+   if (wantNDim && wantErrors) {
+      hn->Sumw2();
+   }
+
    Int_t* bins  = new Int_t[ndim];
    Long64_t myLinBin = 0;

[dpiparo]

Thanks a lot for the report and the proposed fix. Would you like to propose a PR with that and a stripped version of the reproducer as a unit test to preserve your ownership of the solution?

[jgrosseo]

Hi Danilo,

Thanks for your reply and the suggestion. It is fine for me if you take over - and I guess you also have to check if my solution works for all uses of the function.

In any case, this may potentially have affected a huge number of analysis...

Best