Mathematical Font Art

by Joris van der Hoeven

Laboratoire d'informatique, UMR 7161 CNRS

Campus de l'École polytechnique

1, rue Honoré d'Estienne d'Orves

Bâtiment Alan Turing, CS35003

91120 Palaiseau

August 14, 2016


Currently, only a limited number of fonts are available for high quality mathematical typesetting, such as Knuth's computer modern font, the Stix font, and several fonts from the TeX Gyre family. An interesting challenge is to develop tools which allow users to pick any existing favorite font and to use it for writing mathematical texts. We will present progress on this problem as part of recent developments in the GNU TeXmacs scientific text editor.


For a long period, most documents with mathematical formulas were typeset using Knuth's Computer Modern font [5]. Recently, a few alternative fonts were designed, such as the Stix font [7] and the TeX Gyre fonts [3]. These handcrafted fonts all admit a high quality, but they required an important development effort. Now there exists thousands of fonts for non mathematical purposes. To what extent is it possible to use such fonts for mathematical texts or presentations, or on the web?

In this paper we describe recent developments inside the GNU TeXmacs scientific text editor [1] which aim at a better support of general purpose fonts, thereby making life a bit more colorful. The focus is on fully automatic techniques for using existing fonts inside structured documents with mathematical formulas. Further fine tuning for specific characters in particular fonts is another interesting topic which will not be discussed here.

There are obvious limitations of what we can do with a font if bold and italic declinations or glyphs for various important characters are missing. Nevertheless we will see that quite a lot is often possible even though the resulting quality may be inferior to what can be achieved via manual design. Since various special characters or font effects are often only used at a reduced number of places inside actual documents, the occasional loss of quality may remain within acceptable bounds, even for professional purposes.

Our general strategy for turning existing fonts into full fledged mathematical font families is to remedy each of the font's insufficiencies. The most common problems are the following:

The main countermeasures are font substitution and font emulation. The first technique (see Section 2) consists of borrowing missing glyphs from other fonts. This can either be done on the level of an entire font (e.g. for obtaining bold or italic declinations) or for individual characters (e.g. a missing symbol, or lacking Greek characters). Font emulation consists of combining and altering the glyphs of symbols in a font in order to generate new ones. This can again be done for entire fonts (Section 3) or individual glyphs (Sections 4 and 5).

All techniques described in this paper have been implemented in TeXmacs, version 1.99.5 and beyond. The software can freely be downloaded from our website The virtual character definitions described in Section 4 below can be found in the TeXmacs/fonts/virtual directory; interested users may play with these definitions. Longer examples of what can be obtained using the techniques described in this paper are available here:

In the TeX/LaTeX universe, there have also been several efforts towards better support for modern OpenType fonts, most notably XeTeX [4] and LuaTeX [2]. The first system also contains features that are similar to those described in Section 3. However, these systems do not support full mathematical font emulation as presented in this paper. XeTeX and LuaTeX also tend to diverge from standard LaTeX through the introduction of incompatibilities.

2Font analysis and font substitution

In order to borrow missing characters from other fonts, it is important to be able to determine fonts with a similar design, so that the alien glyphs fit nicely into the main text:

The symbols , , are acceptable inside .

The symbols do not look very well inside .

Usually, rules for font substitution are specified manually for each individual font. Although this often yields the most precise and predictable results, it can be tedious to write such rules. For this reason, we also implemented a more automatic mechanism in order to determine good substitutes.

A prerequisite for our algorithm for automatic font substitutions is a detailed analysis of the main characteristics of all supported fonts. The results of this analysis are stored in a database. Using this database, we may then compute the distance between two fonts. In the case when a symbol is missing in a font , it then suffices to find the closest font that supports this symbol . Notice that the best substitution font may depend on the fonts which are installed on your system.

In our database we both use discrete font characteristics (e.g. sans serif, small capitals, handwritten, ancient, gothic, etc.) and continuous ones (e.g. italic slant, height of an “x” symbol, etc.). Most characteristics are determined automatically by analyzing the name of the font (for some of the discrete characteristics) or individual glyphs (for the continuous ones). Some “font categories” (such as handwritten, gothic, etc.) can be specified manually.

One of the most important font characteristics is the height of the “x” symbol (with respect to the design size). When the font borrows a symbol from the font we first scale it by the quotient of these x-heights inside and . In the example (1) this was done correctly, contrary to (2).

Other common font characteristics are also taken into account into our database, such as the italic slant, the width of the “M” symbol, the ascent and descent (above and above the “x” symbol), etc. In addition, we carefully analyze the glyphs themselves in order to determine the horizontal and vertical stroke widths for the “o” and “O” symbols, the average aspect-ratios of uppercase and lowercase letters, and the average area of glyphs that is filled (how much ink will be used).

Our current implementation manages to find reasonably good font substitutions. Notice that this may even be a problem on certain occasions. For instance, in the example (3) below, the sans serif font is such a good match that it can barely be distinguished from the serif font, thereby defeating its purpose:

This sample text is a bit too good.

3Poor man's font emulation

Various font alterations such as Bold, Italic and Small Capitals can be emulated in rather obvious ways, although with significant loss of quality:

With more work, the above “poor man's” strategies might be further enhanced. For instance, the italic a might be better approximated using a shortened version of d instead of a. In order to improve bold font emulation, we might also replace pixels by small lines of cleverly adjusted lengths.

More elaborate emulation strategies might greatly benefit from a toolkit for “retro-engineering” the design of existing fonts. For instance, given an outline, we might want to determine the curve(s) followed by a “pen” and the size (or shape) of the pen at each point of the curve. This would then make it easy to produce high quality narrowed and widened versions of a font, as well as better emboldened fonts, or variants in which the pen's size is uniform (as needed for sans serif and typewriter fonts). Another interesting question is whether it is possible to automatically detect serifs and to add or remove them.

We have started to experiment with more elaborate emulation algorithms for the generation of “blackboard bold” variants of glyphs. The easiest strategy is to produce an outlined version of the possibly emboldened input glyph. The standard AMS blackboard bold font uses this method (, , , , ), but we consider the result suboptimal with respect to adding a single stroke (, , , , ). We implemented an algorithm for the detection of the part of contour to be “double stroked”. We next embolden this part and hollow it out.

Regular Bold Italic Small Caps Blackboard Bold Mathematics
Optima Bold* Italic Small Caps , , , ,
Cochin Bold* Italic* Small Caps , , , ,
Chartrand Bold Italic Small Caps , , , ,
Essays1743 Bold* Italic* Small Caps , , , , *
Meyne Textur Bold Italic Small Caps , , , ,
Chalkduster Bold Italic Small Caps , , , ,
Comic Sans Bold Italic Small Caps , , , ,
Papyrus Bold Italic Small Caps , , , ,

Figure 1. Emulation of bold, italic, small capitals and blackboard bold.

* These declinations are already supported by the original font.

4Virtual characters

Missing glyphs can be generated automatically from existing ones using a combination of the following main techniques, listed by increasing complexity:

In a similar vein, we need various querying mechanisms: all glyphs come with logical and physical bounding boxes, but we sometimes may want to compute the exact width of some stroke or obtain other kinds of information.

We developed a small language that can be used for defining new “virtual” characters in terms of existing ones. The design of every new virtual glyph can be regarded as a puzzle: finding a clever way to combine existing glyphs into the desired one using the primitives from the language. Of course, we are looking for robust solutions in the sense that they should work for any reasonable font in which the required basic glyphs are available.

Let us consider a few examples. For the construction of arrows, it turns out that the single guillemets ‹ and › are often well suited for the heads (the rescaled symbols < and > are acceptable fallbacks). The arrow bars are obtained from the minus sign , but the determination of an appropriate minus is non trivial. For instance, the width of the dash - is usually too large, so we should avoid using this symbol. The underscore is a better candidate; one may also cut the plus sign into several pieces (avoiding the vertical stroke) and recombine them.

Assuming that we have an appropriate arrow bar and head, we may use the following code for producing an actual arrow:

(rightarrow (right-fit arrowbar (align righthead arrowbar * 0.5)))

The align primitive is used to vertically align the arrow head at the center of the arrow bar. The right-fit primitive is less basic and corresponds to sliding the arrowhead from the right to the left until the arrow bar goes past the head on its right. More direct ways to produce arrows turn out to be less robust. Left and left-right arrows can be defined using

(leftarrow (left-flip rightarrow))


(leftrightarrow (join (part leftarrow * 0.5) (part rightarrow 0.5 *)))

These definitions potentially take advantage of an existing rightarrow in the base font. The part primitive performs two horizontal clippings between the middle and the extremities, whereas join is used for superposition.

An interesting challenge is the emulation of Greek characters. This seems intractable for the lowercase symbols, but is less hopeless for the capitals. For instance, <Gamma*> can be obtained by flipping the Roman L upside down and we already mentioned how to obtain a reasonable <Phi*>. More interesting is the case of <Pi*>, which can be obtained from H by moving the horizontal bar to the top. However, extracting this bar is not so easy in some fonts: consider H. For a robust method, we therefore cut the H into pieces: we first extract <H1> <H2> <H4> and recombine them into <H5>. We next take a tiny piece of the central bar, extend it to the desired length, and move it to the top.

Optima <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Cochin <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Didot <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Cuprum <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Essays1743 <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Am. Typewr. <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Chalkboard <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Chalkduster <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Papyrus <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>
Paper Cuts <Gamma*> <Delta*> <Theta*> <Lambda*> <Xi*> <Pi*> <Sigma*> <Upsilon*> <Phi*> <Psi*> <Omega*>

Figure 2. Emulation of various mathematical symbols in various fonts.

5Rubber characters

One specific problem with mathematical fonts is the need for rubber characters. There are essentially four types of them: big operators (, , ), large delimiters , wide accents (, ), and long arrows (, ).

We produce these rubber characters using essentially the same techniques as in the previous section. Especially horizontal and vertical scaling are very useful, as well as cutting symbols into several parts and reassembling them appropriately.

For instance, moderately large versions of the bracket ( are obtained through magnification, typically with a higher stretch factor in the vertical direction. For larger sizes, this method produces results that are unacceptably thick. In that case, we rather cut the bracket into a top, a bottom, and a tiny middle part. We next repeat the middle part as many times as necessary in order to obtain a bracket of the desired size.

The use of scaling is a very delicate matter. For instance, in the case of square brackets [ (and their potential derivatives and ), the point where horizontally magnified versions get too fat is usually reached much earlier than for ordinary or curly brackets. In the case of wide accents, we typically need very large horizontal stretch factors, which yield unacceptable results. Magnified versions of and typically look allright, but this is much less so for .

We are still in the process of fine tuning our implementation. For better results, one major challenge is to develop magnification methods with a finer control over the stroke widths. In particular, we need a reliable magnification method that preserves all relevant widths.

Figure 3. Assorted rubber symbols from various fonts.

6Conclusion and perspectives

After a moderate development investment, we are now able to use a lot of existing fonts for mathematical typesetting. The quality of the obtained results ranges from “better than nothing” to “professional typesetting quality”. Our virtual font implementation can be regarded as a genuine “metafont”. Paradoxically, and in comparison, Knuth's Metafont initiative [6] has essentially resulted in the creation of a single mathematical font of extremely high quality.

One interesting question that occurred during our development of a virtual mathematical metafont concerns the “essence of a font”: which font features essentially contain all necessary information to reproduce the entire font, and how? For instance, most mathematical symbols can be reconstructed from a few basic glyphs: , , , , , , › (or ), , , , (, [ and {. Similarly, the Greek capitals can essentially be reconstructed from E, H, O, X and Z. So what is the real “fingerprint” of a font?

The development of more and better glyph emulation tools might be valuable for font designers. On the one hand, such tools may be used to automatically generate lots of glyphs. On the other hand, they allow designers to compare their own handcrafted glyphs with automatically generated alternatives. This may help to spot errors or increase consciousness about the distinctive features of a personal design.

For the moment, we developed all our font substitution and emulation tools inside TeXmacs. It might be worthwhile to conceive a separate library with even more systematic tools for font analysis, retro-engineering and glyph emulation. Such a library might come with command line tools for generating mathematically enriched fonts, emboldened or narrowed versions, etc. For the moment, several of our algorithms are also limitated to operating on bitmaps. In the future, it would be nice to systematically work with vector graphics only.

One final issue concerns the purpose of alternative fonts. For instance, certain fonts such as Chalkboard, Chalkduster, Essays1743, Yiggivoo 3D are mainly used in order to produce specific graphical effects: emulate text on a chalkboard or on a blackboard, imitating a degraded retro-font, or producing a 3D sensation. It can be questioned whether these purposes are always best served through the use of a special font. For instance, handwriting might be imitated better by dynamically generating many different versions of the same letter. Better retro and 3D effects might be obtained by applying a suitable graphical filter to an entire portion of text. This might even more be true in the presence of fractions, square roots or geometric pictures.


[1] Massimiliano Gubinelli, Joris van der Hoeven, François Poulain, and Denis Raux. GNU TeXmacs: towards a scientific office suite. In Mathematical Software - ICMS 2014 - 4th International Congress, Seoul, South Korea, August 5-9, 2014. Proceedings, pages 562–569. 2014.

[2] T. Hoekwater, H. Henkel, and H. Hagen. Luatex., 2007.

[3] B. Jackowski, J. Nowacki, and J. Ludwichowski. The TeX Gyre collection of fonts.

[4] J. Kew. Xetex., 2005.

[5] Donald E. Knuth. Computer Modern Typefaces, volume E of Computers and Typesetting. Addison-Wesley, 1986.

[6] Donald E. Knuth. The METAFONTbook, volume C of Computers and Typesetting. Addison-Wesley, 1986.

[7] STI Pub companies. STIX fonts project., 2010.