What is the meaning of orthogonal in validation testing?

I have heard the term "orthogonal * validation" used recently. It was used in the context of experimental platform testing. What does this mean? I cannot find anything on it in literature or Wikipedia.

• Could you provide a reference to what you heard or elaborate on the context a little? – whuber Jun 17 '12 at 13:26
• I believe it was in the context of microarray platform comparison or analytical tool comparison. – user1447630 Jun 17 '12 at 18:09
• Google searching reveals that "orthogonal validation" in an experimental context can mean different things in different fields. "Orthogonal" is often applied loosely to anything that metaphorically can be thought of as "at right angles." The metaphor extends to statistics (uncorrelated variables are at "right angles" to each other), design of experiments (orthogonal designs and polynomials), and to physical realms with orthogonal arrays of cells (in proteomics) and orthogonal plowing patterns (in agriculture). The technical meanings differ radically, which is why some context is crucial. – whuber Jun 18 '12 at 14:32

Basically, it would seem that people use orthogonal as a synonym for independent. So, for orthogonal validation read independent validation The validity of equating orthogonality with independence is discussed here such that "if X and Y are independent then they are Orthogonal" but "the converse is not true".

• That seems to correspond with what I heard. It's not immediately obvious that independence and orthoganality are interchangeable. – user1447630 Jun 17 '12 at 18:29
• The converse is true when the variables are jointly normally distributed. Speaking of variables, it's not totally clear to me what's orthogonal to what in this context but that may just be because I'm not familiar with the application. – Macro Jun 17 '12 at 18:34
• @Macro presumably the different methods of validating some experimental platform are orthogonal? That's how I interpret it anyway. Perhaps OP can confirm. – user1202664 Jun 17 '12 at 18:37
• @user1447630 The answer you checked is wrong. Orthogonality in statistics has nothing to do with orthogonal validation. Check the links out that I gave you for the correct description. A few are abstracts that don't give complete nswers but others do. Also the statistical term orthogonal is not the same as independent. It means that variables are uncorrelated and uncorrelated does not always mean indpendent. – Michael R. Chernick Jun 17 '12 at 18:41
• @MichaelChernick, then can you solve the mystery and tell us all what it is? I'm still not sure, even after looking at both of the answers here. – Macro Jun 17 '12 at 18:47

An orthogonal method is an additional method that provides very different selectivity to the primary method. The orthogonal method can be used to evaluate the primary method. For example, two methods can be used to investigate protein aggregation 1) size-exclusion chromatograph or an orthogonal method such as 2) analytical ultracentrifugation. Both methods are independent approaches that can answer a question such as "is my protein aggregated?"

This "orthogonal" word is rather fashionable part of slang in recent EMA/FDA guidelines. Literary means that something crossed at right angle, so more intuitively understood substitute is "cross-over methodology" i.e. two essentially different methods used to measure the same value ("crossing-point"), so the measurement is reliable. Latin rules again :-)

Orthogonal means independent in loose sense. It sounds different in different contexts and technical fields. like in signal processing orthogonal signals means the signals with zero correlation.

But in the contest of testing (validation and verification of embedded software or hardware), preparing and testing more than one functionality at the same time which are independent. say for example we have a system with 3 subsystems in it with set of inputs a1,a2 for subsystem A and out put A1, b1,b2,b3 as inputs to subsystem B with outputs B1 and B2, and c1,c2,and B2 as inputs to Subsystem C with output C1. In this case subsystem A and B are independent and C is dependent on the output of Subsystem B. So, we can test the functionality of the Subsystem A and B in parallel that will save time instead of testing functionality of the subsystem A and then B. Here part of the subsystem C also be tested with the same test sequence done for the B. So very rational part of the subsystem C is has to be tested. With clear understanding of the functionality of the subsystem C, we can test the entire system with minimal number of test cases this is called orthogonal testing in general.

In-case of hardware IO calibration, It is more quite easy, instead of testing the hardware IO's one by one, we can calibrate the independent IOs at the same time. for example if an ECU has 20 sensor inputs, if there is no dependency among these IOs, then all can be calibrated with a single test ( test duration might be different, but sure it will save 90 to 98 of total time required to calibrate independently)

From my understanding this is more efficient for unit testing, system testing with more independent subsystems/functions. As the number of dependent functions in the system increases then the efficiency of time spent to create orthogonal arrays increases than the robustness testing.