Terminology • Environment • Local setup • Remote setup • Cloning repos • New repos • Synchronizing • Details • Tracking changes • TeX issues • Collaboration
This no-frills tutorial aims to teach how to set up and use a git repository to collaboratively edit a mathematical paper. It is self-contained and has been kept to a virtually absolute minimum to ensure that you can quickly start using git for your own project. Several of my collaborators have successfully used this tutorial to setup and use git daily. I welcome feedback and questions on this tutorial, which you can email to me.
The basic idea is that all of your TeX files and any auxiliary files (e.g., macro files, source code for pictures, bibliography files, a Makefile, etc.) are stored in a git repository, where they can be edited simultaneously by all collaborators. If two people edit the same file, git automatically merges the changes as explained below. This last aspect clearly distinguishes git from file sharing services such as Dropbox, Google Drive, or Apple iCloud. It often happens that two coauthors edit the same paper and one of them accidentally edits an older version of the manuscript. Recovering from such a situation is a nightmare in the classical setup, but git tracks all versions automatically and this simply cannot happen.
A typical git collaboration involves a remote host, i.e., a server where a copy of the repository will be stored for all your collaborators to access, and multiple local hosts, which are the machines on which you and your collaborators store your own local copies of the repository.
Theoretically, you could use various public servers as your remote host, the most (in)famous example being GitHub. However, GitHub repos are public, so everyone will be able to see your draft. For this reason I strongly recommend to set up your own private remote host. For $15 per year or less you can maintain a virtual private server that will be perfectly sufficient for your git setup.
Some of the commands below have italic text in them, which must be substituted as follows:
This tutorial assumes that you have access to a Unix-like environment on your local and remote hosts. In particular, diff and ssh must be installed, as well as a pager such as less. This does not mean that you have to install Linux or any other Unix-like operating system, all necessary software can be installed in macOS or Microsoft Windows.
For Darwin-based systems such as macOS you will probably need Homebrew, MacPorts, Rudix, or Fink.
For beginners, the easiest to use is probably Homebrew.
Follow the installation instructions for Homebrew.
Once you are finished, run the following command at a Terminal prompt:
brew install less diff git ssh
For Microsoft Windows you will probably need MSYS2 (git for Windows is just a bundle of several MSYS2 packages), Cygwin, or midipix when it is ready.
For beginners, the easiest to use is probably MSYS2.
Follow the installation instructions for MSYS2.
Once you are finished, run the following command at an MSYS2 prompt:
pacman -Sy less diff git openssh
This section should be applied once to every local host. The directory ~/ is the home directory of your Unix installation. You can find it out by running pwd after opening a terminal window. However, on the command line you may use ~/ as written, since the shell is programmed to substitute it accordingly.
To set up the encryption keys for ssh and configure ssh and git, run the following commands:
mkdir -p ~/.ssh
ssh-keygen -N "" -f key
curl https://dmitripavlov.org/config -o ~/.ssh/config
curl https://dmitripavlov.org/.gitconfig -o ~/.gitconfig
Edit the files ~/.ssh/config and ~/.gitconfig, replacing the text surrounded by >>>…<<< with appropriate data.
Furthermore, whoever is maintaining the central git server should append the contents of the public key file ~/.ssh/key.pub
to the file ~/.ssh/authorized_keys on the remote host, e.g.,
cat key.pub >>~git/.ssh/authorized_keys
This section only applies if you want to setup your own git server, as opposed to using a git server set up by somebody else.
On the remote host, run
useradd -c git -e "" -f -1 -k "" -m -r -s /usr/bin/git-shell -U git
passwd -d git
mkdir git-shell-commands .ssh
chown -R git:git git-shell-commands .ssh
You also need to make sure that the sshd daemon is running
and adjust its configuration file (typically /etc/sshd/sshd_config),
ensuring that the following lines are present:
Another recommended option is
for some random port number below 65536 (the standard SSH port is often scanned with malicious intents).
This section only applies if you want to clone an existing repository and not create a new one.
On the local host, change into the directory that contains your projects and run
git clone git@remote-host-name:repo-name
This section only applies if you want to create a new repository and not clone an existing one. Copy the public encryption keys of the people who need to access the project to the remote host; the code below assumes they are stored in files with names of the form *.pub.
On the remote host, run
cat *.pub >>.ssh/authorized_keys
git init --bare
chown -R git:git .
On the local host, change into the directory that contains your projects and run
git add paper-name.tex
git commit -m "Initial commit"
git remote add origin git@remote-host-name:repo-name
git push -u origin master
Git was developed for software projects and has many features that bear little relevance for collaborative editing of mathematical papers. Furthermore, the desired workflow of a mathematical collaboration is very different from a software project: whereas a software engineer would typically develop some new feature in a separate branch, and then merge it once it is ready, a mathematician wants to synchronize his version of the text with the other versions as soon as he makes an edit. For this reason we explain here a simplified setup that should be sufficient for the overwhelming majority of mathematicians.
Below it will be important to understand that at all times at least three potentially different copies of your project are stored:
For us, the entire range of operations is covered by a single command (actually, an alias set up in .gitconfig as given above),
which must be run in project's directory that was created by git clone:
This command should be run every time before you start editing as well as when you finish your editing sessions. It performs the following actions.
If you are offline, the parts responsible for exchanging information with a remote repository will fail, but this does not affect the rest of the command. When you go online later, simply run git sync again and it will synchronize everything.
This section is optional and should be skipped on the first reading. It explains the individual commands behind git sync.
First, recall the definition of git sync from the .gitconfig file above:
git add -u
if ! git rebase --continue; then
if [[ $# == 0 ]]; then
git commit -m \"$*\"
git pull -r
The first command git add -u tells git to store all changes made in the working tree to the index, an intermediate stage between the working tree and the local repository. The -u instructs git not to add any newly created files. The reason for this is that more often than not newly created files are temporary files of some sort that should not be added permanently to the repository. If you do need to add a new file, you can do it using the command git add file-name.
The next command git rebase --continue instructs git to continue the rebasing process after you resolved a merge conflict (if there was one in the first place). The rebasing process amounts to transplanting the edits that you made to some older version of the document on “top” of changes made by somebody else to the same old version of the document. This allows git to combine two versions of the document into one. The base process “replays” your commits one-by-one, and merge conflicts are resolved for each commit individually, so you may encounter new merge conflicts after this command runs. They should be resolved in the same manner, after which you run git sync again. The net effect of this stage is that it looks like you edited a newer version of the document from the start.
The git commit commands instruct git to store all the changes recorded in the index by the git add -u command into the local repository.
The git fetch command retrieves new commits from a remote repository and stores them in your local repository, whereas git rebase replays your changes on “top” of these new changes, as described above.
The git push command sends the changes stored in your local repository to the remote repository.
Finally, git pull -r instructs git to retrieve any changes for submodules (something we will consider below).
To see the list of all commits with the latest ones on top, use git log. If you also want to see the changed lines, use git log -p. The alias git wlog will show the changed words as opposed to mere lines, which is very convenient when you edit TeX files.
To see the differences between your current files and the last commit, use git diff. Use git wdiff to highlight individual words instead of lines.
To add a new file to the repository use git add file-name, then run git sync.
The majority of git's use cases concern the editing of computer code. For the editing of TeX documents it is desirable to make some minor adjustments to the typical workflow and configuration.
Git, like many Unix tools, operates on the level of individual lines. Some people these days do not wrap long lines, so in their TeX files every paragraph is a single line of text. This means that whenever two people edit two different sentences in the same paragraph, one gets a merge conflict. Thus it is advisable to keep the lines relatively short, in order to reduce the number of merge conflicts.
The easiest way to ensure that lines stay short is to set up your editor for automatic word wrapping after a specified column, e.g., 80. It is important, though, to set up the editor in such a way that it does not reflow the entire paragraph whenever a single line is wrapped. Otherwise git thinks that the entire paragraph changed, which makes it difficult to inspect history.
I find myself easy to hit Enter instead of Space at the end of each sentence, and also at the end of major clauses in long sentences. This is the best approach: you don't have to spend any additional time, git always shows correct changes (e.g., if you edit a single clause in a sentence, only this clause is shown by git diff), and it's easier to locate a specific sentence in a paragraph by scanning the left column. See Semantic Linefeeds by Brandon Rhodes for more details.
If you want to move around or remove a large chunk of text, it makes sense to form a separate commit with just this change. This makes the inspection of history much easier. Otherwise, if you move a large chunk of text and edit it at the same time, git shows the entire paragraph as changed, which makes it impossible to see which parts of the paragraph were actually edited, and which ones were just moved around.
Thus the best procedure for moving text around is to commit whatever changes you have accumulated to this point, then move the text and immediately commit the changes, and then continue to edit the document.
It's best to comment text out like this:
instead of like this:
In other words, one places \iffalse and \fi on separate lines before and after the given block of text. The reason for this is the same as above: the git history is not polluted by massive blocks of commented text.
The plain git commands show changes on a line-by-line basis, i.e., you can see which lines were changed, but there is now way to tell which parts of a line were changed. It's common to change just one or two characters in a line when editing a TeX document, which makes it desirable to have a diff file that shows changes on the level of individual words. This is accomplished by git's --color-words option supplied to git commands. The wordRegex option in the configuration file controls how lines are split into words; for TeX documents it is desirable to have a more refined regular expression, so that, for example, displayed formulas don't get treated like single words.
The configuration file given above provides convenient aliases wdiff, wlog, and wshow to git's commands diff, log, and show that enable the option --color-words. I find myself using these variants almost exclusively when editing TeX documents.
Instead of communicating with coauthors by email, one can also leave comments directly in the TeX file. This has the following advantages.
However, git's history tools are very good at tracking small changes. This allows your coauthors to fix minor issues in the text without any comments, which is much more efficient. One can then easily track them using git log -p (or git wlog, as explained above). If a change raises concerns, one can always leave a comment later.