New blog post by Galin Bistrev

.github/actions/spelling/allow/terms.txt

@@ -5,6 +5,7 @@ Backpropagation
 CINT
 CMSSW
 Caa
+Codegen
 Cppyy
 Debian
 EPC
@@ -30,6 +31,7 @@ Ohridski
 OMP
 OpenMP
 PTX
+RAII
 Resugaring
 SBO
 Slib
@@ -44,10 +46,12 @@ biodynamo
 bioinformatics
 blogs
 cms
+codegen
 consteval
 cppyy
 cytokine
 cytokines
+doxygen
 gitlab
 gpu
 gridlay

_posts/2025-23-09-Supporting-AD-in-CMS-Combine-likelihood-scans.md

@@ -0,0 +1,144 @@
+title: "Results from CERN Summer School 2025: Supporting Automatic 
+Differentiation in CMS Combine profile likelihood scans"
+layout: post
+excerpt: "A CERN Summer Student 2025 project aiming at the support of 
+automatic differentiation (AD) for likelihood scans in the CMS Combine 
+tool to accelerate statistical inference by leveraging RooFit's AD 
+support and LLVM-based gradient generation."
+sitemap: false
+author: Galin Bistrev
+permalink: blogs/2025_galin_bistrev_results_blog/
+banner_image: /images/blog/banner-cern.jpg
+date: 2025-09-25
+tags: cern cms root combine c++ RooFit automatic-differentiation	
+---
+
+### **Introduction**
+Greetings! I’m Galin Bistrev, a fourth-year student specializing in
+ Nuclear and Particle Physics at the University of Sofia "St. Kliment Ohridski."  
+As part of the CERN Summer Student Programme 2025, I was working on a 
+project that aimed to provide support for Automatic Differentiation 
+(AD) into the CMS Combine tool profile likelihood scans.
+
+Mentors: Jonas Rembser, Vassil Vasilev, David Lange
+
+### **Description of the Project**
+
+This project aims to enhance support for Automatic Differentiation (AD) 
+in likelihood scans within the CMS Combine framework, the primary 
+statistical analysis tool of the CMS experiment at CERN. Combine is 
+built on top of RooFit, which has recently introduced AD to improve 
+minimization techniques. By providing computationally efficient 
+gradients through AD, RooFit achieves substantial performance 
+improvements. In RooFit, Clad converts internal likelihood 
+representations into standalone C++ code, from which gradient 
+routines for AD are generated. This strategy not only speeds up the 
+fitting process but also increases the portability and shareability 
+of likelihood models, making them usable even by those without 
+detailed knowledge of RooFit or Combine internals.
+
+### **Brief overview of the CMS Combine engine**
+Combine is a statistical analysis framework that compares models of 
+expected observations with real data. It is widely used for tasks such 
+as searching for new particles or processes, setting limits on 
+potential new physics, and measuring physical quantities like cross-sections. 
+Although developed with High Energy Physics (HEP) 
+applications in mind, Combine contains no intrinsic physics assumptions, 
+making it fully general and independent of any specific analysis. 
+This flexibility allows it to be applied across a broad range of 
+statistical problems.
+
+Roughly, Combine performs three main functions:
+
+- Builds a statistical model of expected observations.
+- Runs statistical tests comparing the model with observed data.
+- Provides tools for validating, inspecting, and understanding both the 
+model and the results of the statistical tests.
+
+### **Project goals**
+
+In order for AD to be supported in Combine likelihood scans, a number of goals needed to be achieved:
+
+- Refactoring some of Combine's logic into RooFit, so that Combine can 
+reuse the AD-enabled minimization algorithm already present there. 
+- Integrate gradient computation into likelihood scans, ensuring that 
+derivatives are correctly propagated for efficient and accurate minimization.  
+- Validate correctness and performance, confirming that the AD-based 
+scans produce results consistent with traditional methods while 
+offering improved performance.
+
+## **Overview of Completed Work**
+Over the course of the project, several major tasks were completed to achieve the stated objectives:
+
+- Imported the `RooMultiPdf` class in RooFit from Combine, enabling 
+switching between multiple PDF-s, applying statistical penalties, 
+and supporting code generation for AD.
+
+- The implementation of the new class was made to be supported by 
+`codegen` in RooFit by adding a new function in `MathFunc.h` and 
+extending `CodegenImpl.cxx` to generate code for models making use of it.
+
+- Imported three pieces of code from Combine that handle the 
+minimization procedures within the framework in RooFit's `RooMinimizer.cxx`. 
+The first is a class imported by Jonas Rembser 
+called `FreezeDisconnectedParametersRAII`, which automatically 
+freezes and unfreezes parameters disconnected from the likelihood graph. 
+The second is the function `generateOrthogonalCombinations`, which 
+generates a list of index combinations by initializing a base 
+configuration with all indices set to zero and then varying one category at a time. 
+The third and final piece of code is a function called `reorderCombinations`, 
+which takes the set of indices produced by `generateOrthogonalCombinations` 
+and adjusts each combination by adding the corresponding base values 
+modulo the maximum allowed index, effectively shifting the combinations 
+relative to the current best indices.
+
+- Using the above-stated functions, the discrete profiling algorithm, 
+which is the main minimization algorithm in Combine, was imported 
+into `RooMinimizer.cxx`.
+- A [tutorial](https://root.cern/doc/master/rf619__discrete__profiling_8py.html) 
+was created along with a [benchmark](https://github.com/vgvassilev/clad/issues/1521), 
+made by Jonas Rembser, demonstrating discrete profiling with RooMultiPdf objects 
+and evaluating the performance of AD in the likelihood scans.  
+
+## **Results**
+With those objectives accomplished, RooFit now provides AD support for 
+discrete profiling. However, the developed benchmark indicates that AD 
+does not currently improve efficiency, as the gradient code generated by 
+Clad introduces overhead. Further optimization in Clad is needed to achieve 
+the potential performance gains for RooFit likelihood scans. More information 
+regarding the issue can be found at [#1521](https://github.com/vgvassilev/clad/issues/1521).
+
+## **Conclusions**
+Thanks to this project, RooFit now enables AD support for discrete profiling in Combine, 
+which, after addressing the current overhead in Clad, would allow for 
+significantly faster and more efficient likelihood scans while maintaining 
+accurate optimization of both discrete and continuous parameters. 
+
+## **Future work**
+- Further benchmarking is required to quantify the potential performance 
+gains from automatic differentiation.
+- Additional optimization of Clad is needed to eliminate unnecessary 
+overhead in gradient generation.
+- The discrete profiling logic implemented in RooMinimizer should be 
+tested across different models to evaluate the minimizer’s behavior and 
+robustness.
+- Extend doxygen documentation of RooMinimizer to describe treatment of discrete 
+parameters.
+- Test if the implementation of discrete profiling works also inside CMS Combine , 
+replacing their implementation in `CascadeMinimizer.cxx`. 
+
+## **Acknowledgements**
+I would like to express my sincere gratitude to the CERN Summer School 
+for the opportunity to participate in such an inspiring project. 
+I extend special thanks to Jonas Rembser, Vassil Vassilev, and David Lange for 
+their invaluable guidance and for providing continuous learning opportunities throughout this journey. 
+I am also grateful to the ROOT team for welcoming me and supporting me throughout my stay at CERN. 
+
+## **Related Links**
+- [CMS Combine GitHub page](https://cms-analysis.github.io/HiggsAnalysis-CombinedLimit/latest/)  
+- [ROOT official repository](https://github.com/root-project/root)  
+- [My GitHub profile](https://github.com/GalinBistrev2)
+- [Presentation](/assets/presentations/CaaS_Weekly_25_09_2025_Galin_Bistrev_AD_in_CMS_Combine.pdf)
+
+
+