Code Length Measured in 14 Languages(blog.wolfram.com)
blog.wolfram.com
Code Length Measured in 14 Languages
http://blog.wolfram.com/2012/11/14/code-length-measured-in-14-languages/
12 comments
Judge for yourself:
http://reference.wolfram.com/mathematica/guide/Mathematica.h...
The Mathematica 'ecosystem' doesn't benefit much from OSS, but it does benefit from hundreds of smart engineers with diverse backgrounds who work on it full-time. And SW as benevolent dictator, vetting every design decision for consistency and elegance.
Don't write a web server in Mathematica -- it's not particularly well suited to that. But if you're writing natural language parsing, knowledge base infrastructure, and algorithmic content for something like Wolfram|Alpha, I can't imagine anything better to use.
http://reference.wolfram.com/mathematica/guide/Mathematica.h...
The Mathematica 'ecosystem' doesn't benefit much from OSS, but it does benefit from hundreds of smart engineers with diverse backgrounds who work on it full-time. And SW as benevolent dictator, vetting every design decision for consistency and elegance.
Don't write a web server in Mathematica -- it's not particularly well suited to that. But if you're writing natural language parsing, knowledge base infrastructure, and algorithmic content for something like Wolfram|Alpha, I can't imagine anything better to use.
That's the issue
Mathematica is too different from the rest. Or mathematically speaking, the intersection between the capabilities of Mathematica and Python for example is smaller than Python and C#
You can't have Mathematica serving a web site (for any non trivial amount of users) (WA has a lot of infrastructure to make it happen)
On the other hand try to calculate an integral over the gamma function in the complex domain, that's one line in Mathematica and how many I don't know in Python even using some libraries.
Mathematica is too different from the rest. Or mathematically speaking, the intersection between the capabilities of Mathematica and Python for example is smaller than Python and C#
You can't have Mathematica serving a web site (for any non trivial amount of users) (WA has a lot of infrastructure to make it happen)
On the other hand try to calculate an integral over the gamma function in the complex domain, that's one line in Mathematica and how many I don't know in Python even using some libraries.
No APL? That language always wins on code golf. You wouldn't want to maintain it though. However it absolutely shines if you're on a 110 baud teletype and have to wait for each character to print. /s
This is truly an apt criticism. There are a lot of J solutions for problems which Mathematica also has a solution, but for some reason it wasn't included.
I quickly ran the code on J and got ratios of lines:0.5, characters: 0.74, tokens:0.5, over 432 comparisons IE J is about half the length of Mathematica code.
APL, has ratios of between 1 and 2 (IE a little longer than Mathematica, but over only 32 comparisons which is too few to be reliable).
Julia was also asked about, but there are only 9 comparisons which is certainly too few to be meaningful.
APL, has ratios of between 1 and 2 (IE a little longer than Mathematica, but over only 32 comparisons which is too few to be reliable).
Julia was also asked about, but there are only 9 comparisons which is certainly too few to be meaningful.
The C# AGM example should be retitled "C# as written by an insane Java programmer who read a .NET 1.1 book and called it a day"
http://rosettacode.org/wiki/Arithmetic-geometric_mean#C.23
http://rosettacode.org/wiki/Arithmetic-geometric_mean#C.23
Aaaaand, this is a huge reason why the whole "experiment" is utterly invalid. Without making use of experienced devs for each language, and perhaps a statistically significant number of samples for each language, the amount you can trust any of this data is precisely zero.
Really stunning
This is everything that's wrong with 'Java lovers' (even though it's in C#)
But then again more lines of code and function points to bill your customer
This is everything that's wrong with 'Java lovers' (even though it's in C#)
But then again more lines of code and function points to bill your customer
Participation involves gardening to clean up entries such as this and to add entries that are missing. Comparisons such as these are only as good as the data, and much of this data is...eww.
Upvotes to you if you clean up that example.
Upvotes to you if you clean up that example.
This article does not measure what it professes to measure. When comparing a language to itself and for large programs (>1000 lines), line count is a decent metric for expressibility. However, for a short amount of code (like that found in Rosetta) and for comparing one language to another, you want to compare both line count and character count.
If you space a few lines of very expressive code out, you don't want that to penalize the results. Similarly, you don't want someone to cheat by just having one huge line.
Edit: I stopped reading when he started describing the fetching/parsing process. My objection is covered later on. Thanks to carlob for pointing this out.
I'll also note that being the best may actually be a bad thing. Perl can be quite succinct, but also quite noisy at times.
Similarly, while I'm a fan of K/J (and I pretty much guarantee that K is more succinct than Mathematica), I think that the code can be very difficult to read sometimes.
If you space a few lines of very expressive code out, you don't want that to penalize the results. Similarly, you don't want someone to cheat by just having one huge line.
Edit: I stopped reading when he started describing the fetching/parsing process. My objection is covered later on. Thanks to carlob for pointing this out.
I'll also note that being the best may actually be a bad thing. Perl can be quite succinct, but also quite noisy at times.
Similarly, while I'm a fan of K/J (and I pretty much guarantee that K is more succinct than Mathematica), I think that the code can be very difficult to read sometimes.
Did you read the whole article? Towards the end it says:
While the “Large tasks – Line count ratio” gives the most impressive result for Mathematica, I think that the “Large tasks – Character count ratio” is the really the fairest comparison.
While the “Large tasks – Line count ratio” gives the most impressive result for Mathematica, I think that the “Large tasks – Character count ratio” is the really the fairest comparison.
This is a fairly meaningless comparison. Rosetta code is not written for terseness, there is no fair comparison about code length to be made based on these example solutions to common problems.
You can be as terse or as verbose as you want in many languages (think forth, perl, apl).
If you want to do a meaningful comparison pick a non-toy problem and write a full sized application, something that in the worst language is a few thousand lines with a proper UI, a web app, a desktop application and a large simulation of something or other. You will still have to account for language gifts such as libraries or built-ins somehow.
And you tell the devs they get points for producing short code.
Then you can compare along several axis, performance, code maintainability (closely related to how easy it is to read) and so on.
That's useful data. This isn't, regardless of whether mathematica looks good.
edit: ok. point taken.
You can be as terse or as verbose as you want in many languages (think forth, perl, apl).
If you want to do a meaningful comparison pick a non-toy problem and write a full sized application, something that in the worst language is a few thousand lines with a proper UI, a web app, a desktop application and a large simulation of something or other. You will still have to account for language gifts such as libraries or built-ins somehow.
And you tell the devs they get points for producing short code.
Then you can compare along several axis, performance, code maintainability (closely related to how easy it is to read) and so on.
That's useful data. This isn't, regardless of whether mathematica looks good.
edit: ok. point taken.
No, he was right, I stopped reading when the author started discussing the mechanics of crawling web pages. I assumed that the blog post would be structured sensibly (analysis, then implementation).
Perl's noisiness would probably be penalized by the token count metric from the article. Although you really need to have language specific lexers to do this properly.
Of course Mathematica is more succinct - it is a domain specific language, not general purpose like most of the others compared. The only meaningful comparisons to Mathematic would be R and MATLAB (other DSLs).
Well, it is certainly concise to represent formulae and equations in Mathematica, but that isn't what makes it a concise language.
That I think is owed to Mathematica being an heir of LISP that implemented m-expressions so cleanly that M doesn't even feel like a homoiconic language. Most other LISP-derivates (like Clojure) stuck to s-expressions, which are historical baggage more than anything else. Some day I'm sure we'll get an m-expression version of Clojure.
Of course in practice one defines new DSLs quite often in Mathematica. And it comes with a variety of DSLs "built-in", mainly for user interface construction, vector graphics, mathematics, and statistical computations.
Disclosure: I too work for WRI.
That I think is owed to Mathematica being an heir of LISP that implemented m-expressions so cleanly that M doesn't even feel like a homoiconic language. Most other LISP-derivates (like Clojure) stuck to s-expressions, which are historical baggage more than anything else. Some day I'm sure we'll get an m-expression version of Clojure.
Of course in practice one defines new DSLs quite often in Mathematica. And it comes with a variety of DSLs "built-in", mainly for user interface construction, vector graphics, mathematics, and statistical computations.
Disclosure: I too work for WRI.
Julia ( http://julialang.org/ ) is another homoiconic language that doesn't use S-expressions. They're both mathematical languages, so it's plausible that Julia was inspired by Mathematica in this choice, though it takes other ideas straight from Common Lisp, so maybe not.
Mathematica is not a domain specific language, otherwise many of the examples in Rosetta Code wouldn't be completed…
See this[0] for more info
[0] http://reference.wolfram.com/legacy/v5/Tour/MathematicaAsAPr...
See this[0] for more info
[0] http://reference.wolfram.com/legacy/v5/Tour/MathematicaAsAPr...
Maybe domain-specific is the wrong term. It's certainly a Turing complete language. But it's optimised to make a particular set of operations easier, so those operations are going to come out shorter than the same operations in a general purpose language.
This is only true if you think the test set in Rosette Code is biased towards the domain that Mathematica is optimized for.
There is no reason to think this looking at the website.
Since he uses median ratios, then a single spectacular ratio will not skew the results.
There is no reason to think this looking at the website.
Since he uses median ratios, then a single spectacular ratio will not skew the results.
Not the test set in Rosetta Code.
The part of the test set that people have actually provided Mathematica solutions for. There is every reason to believe that people are more likely to provide Mathematica solutions for problems that Mathematica is likely to be good at.
The part of the test set that people have actually provided Mathematica solutions for. There is every reason to believe that people are more likely to provide Mathematica solutions for problems that Mathematica is likely to be good at.
This is a potential bias that he does call out in the article.
I would say, rather than being a DSL, the Mathematica language was informed by a particular goal. Much as Prolog, R, RPG, Perl, job control shells, etc., were born of a need to 'scratch an itch' and have found themselves pushed beyond that initial itch, Mathematica's Lisp-like syntax, pattern matching, term rewriting, and functional programming capabilities have resulted in a number of applications written by people who each push mma just a little bit further.
Most of these apps come from people in the target market, but it's similar to the number of Excel spreadsheets that end up running surprisingly large companies. A "non-programmer" develops a bunch of macros and VBA code to make their job easier, and those "DSL" type capabilities become critical core aspects of the system.
Most of these apps come from people in the target market, but it's similar to the number of Excel spreadsheets that end up running surprisingly large companies. A "non-programmer" develops a bunch of macros and VBA code to make their job easier, and those "DSL" type capabilities become critical core aspects of the system.
Probably could have done with a better, more visual graph than the pyramid shown throughout the article (especially when he was using Mathematica!). The trends are there but don't pop out at all, and these are very simple metrics.
One takeaway I didn't expect: Clojure is, on average, nearly as terse as Mathematica.
Also, has anyone actually used Mathematica for something aside from data processing? Like, any of the tasks you would normally use Ruby, C++, or Clojure for?
One takeaway I didn't expect: Clojure is, on average, nearly as terse as Mathematica.
Also, has anyone actually used Mathematica for something aside from data processing? Like, any of the tasks you would normally use Ruby, C++, or Clojure for?
Disclosure: I work for Wolfram.
I might not surprise you to know that most of the code in Wolfram|Alpha is Mathematica code. I don't know what kind of examples you have in mind when you talk about "Ruby, C++, or Clojure", but Alpha is probably the largest Mathematica based project in production.
I might not surprise you to know that most of the code in Wolfram|Alpha is Mathematica code. I don't know what kind of examples you have in mind when you talk about "Ruby, C++, or Clojure", but Alpha is probably the largest Mathematica based project in production.
I did know that a lot of Wolfram|Alpha (And Mathematica itself too, right?) is written in Mathematica code, but as mumrah points out, that only proves that it works well in production as a DSL. When I said "Ruby, C++, or Clojure," I was trying to evoke more common tasks like a desktop application, a simple server, a text-processing library, etc.
Edit: Even if Mathematica isn't technically a DSL, it certainly seems to be geared heavily towards symbolic mathematics and data processing/visualization. My question is really, how does it fare outside its element?
Edit: Even if Mathematica isn't technically a DSL, it certainly seems to be geared heavily towards symbolic mathematics and data processing/visualization. My question is really, how does it fare outside its element?
Yes, some parts of Mathematica are written in Mathematica, though most fast build in functions (linear algebra is one example that comes to mind) are written in faster, compiled languages.
As I said in response to mumrah, Mathematica is not a DSL. You might argue that a lot of 'library' counts as language in Mathematica, but not that you can't write general purpose applications.
As I said in response to mumrah, Mathematica is not a DSL. You might argue that a lot of 'library' counts as language in Mathematica, but not that you can't write general purpose applications.
Mathematica is certainly a domain-specific language. Its domain is symbolic mathematics.
You may be unfamiliar with the term "DSL." It refers to a language designed for solving a specific computational task.
This was, in fact, the stated goal of Mathematica. DSL does not mean the language cannot solve other tasks, only that it was meant for solving a specific one.
You may be unfamiliar with the term "DSL." It refers to a language designed for solving a specific computational task.
This was, in fact, the stated goal of Mathematica. DSL does not mean the language cannot solve other tasks, only that it was meant for solving a specific one.
i don't think it's a DSL. its evaluation paradigm is symbolic in the same way you would say Java's paradigm is OOP, yet we wouldn't say Java is an OOP DSL even if that's true in some technical sense
in this respect i think the name "Mathematica" does a disservice to the system. originally it may have been intended for math, but in its current form it's more like a general prototyping/presentation/construction system that makes thorough use of its capacity as a research platform
in this respect i think the name "Mathematica" does a disservice to the system. originally it may have been intended for math, but in its current form it's more like a general prototyping/presentation/construction system that makes thorough use of its capacity as a research platform
Maybe I'm not reading the pyramid properly, but it looks like Clojure is on average ~2x as lengthy as Mathematica? IE: The entries in the horizontal Clojure row are ~1/2 the entries in the horizontal Mathematica row.
Definitely a confusing table.
Definitely a confusing table.
Yep, that's how I got "nearly as terse" - It's 1x for 'All Tasks - Line count tatio' and about 1.6x for every other metric. (In the first row)
That's a very confusing table, but an interesting comparison.
I think you reveal different things comparing line count and character count, and both miss what I'm looking for in a terse language: scanability.
A lot of the verbosity that annoys me with Java isn't really about number of lines -- it's bad, not particularly worse than any of the other languages descended from ALGOL. No, what makes Java worse than its cousins is the propensity to use English words, and often really long ones, in places where symbols would be easier to scan. "public class Foo extends Bar" is a lot of English text, and I have to read it all to separate out the tokens. "@interface Foo : Bar" makes it easier for me to see at a glance what's going on.
And don't get me started on wading through rotation.multiply(other_rotation).multiply(velocity).add(rotation.multiply(other_velocity)). That's technically one line, but that doesn't make it easy to read.
So winning on line count doesn't capture everything.
On the other hand, measuring characters biases you toward symbol heavy languages like perl or APL, and '$\="\n"' takes some picking apart, too.
I did think APL would win an outright terseness competition, though, by either metric. It's a shame it isn't on the list.
I think you reveal different things comparing line count and character count, and both miss what I'm looking for in a terse language: scanability.
A lot of the verbosity that annoys me with Java isn't really about number of lines -- it's bad, not particularly worse than any of the other languages descended from ALGOL. No, what makes Java worse than its cousins is the propensity to use English words, and often really long ones, in places where symbols would be easier to scan. "public class Foo extends Bar" is a lot of English text, and I have to read it all to separate out the tokens. "@interface Foo : Bar" makes it easier for me to see at a glance what's going on.
And don't get me started on wading through rotation.multiply(other_rotation).multiply(velocity).add(rotation.multiply(other_velocity)). That's technically one line, but that doesn't make it easy to read.
So winning on line count doesn't capture everything.
On the other hand, measuring characters biases you toward symbol heavy languages like perl or APL, and '$\="\n"' takes some picking apart, too.
I did think APL would win an outright terseness competition, though, by either metric. It's a shame it isn't on the list.
"public class Foo extends Bar" is a lot of English text, and I have to read it all to separate out the tokens. "@interface Foo : Bar" makes it easier for me to see at a glance what's going on.
Never mind the @interface declaration. If you're going to be writing or maintaining objc, you have to understand that. However, the x : y notation often leads to confusion. Does x extend y? Does y extend x?
It's probably not a big deal for someone who lives in one or two languages, as the knowledge is refreshed frequently. However, when you swing between, say, Haskell, Python, Ruby, C++, Java, or whatever, and maybe only touch objc when you need to update your iOS app to match your web app, then there's a cognitive hurdle because ':' doesn't have a clear interpretation.
Given its Smalltalk heritage and model of message passing, "@interface y subclass: x" (or "@interface x subclasses: y") would carry less of a cognitive load while perhaps being less verbose.
Again, this is only a superficial example, but choice of names or symbols is important in the language UI.
I've seen metrics done on the basis of symbol counts, not characters or lines. In theory this normalize differences in coding styles, such as verbose identifiers, or positioning of braces, or liberal use of whitespace.
I agree with you about Java. I have a line on my screen that has 6 identifiers and is 99 characters long and it's not even a constructor call. One of many problems is long identifiers with similar names.
I agree with you about Java. I have a line on my screen that has 6 identifiers and is 99 characters long and it's not even a constructor call. One of many problems is long identifiers with similar names.
Scroll all the way to the bottom.
The organization of this post is strange -- a result, a long section about how to write a web crawler in Mathematica, and then about a dozen more different results. If you (like me) are inclined to tune out when you hit code, keep scrolling.
The organization of this post is strange -- a result, a long section about how to write a web crawler in Mathematica, and then about a dozen more different results. If you (like me) are inclined to tune out when you hit code, keep scrolling.
I haven't looked at the Rosetta code examples but I didn't expect the difference between Common Lisp and Clojure to be so pronounced.
I'm mainly a Common Lisp user, does anyone who has used both care to share an opinion?
I'm mainly a Common Lisp user, does anyone who has used both care to share an opinion?
Some thoughts from an ex-CL'er that has dabbled in Clojure for 3-4 years:
Clojure's combination of declarative idioms and highly functional idioms (to a greater extent than CL) lead to it being able to by default express certain things in a more terse fashion.
http://rosettacode.org/wiki/Factors_of_an_integer
Compare Common Lisp and Clojure here.
Common Lisp:
Clojure solutions:
Take a look at the other things that work across the seq'able data structures. Pretty powerful base set of idioms.
Okay, but lets provide a second opinion on these idioms. What about Haskell?
And Mathematica?
I give up. These are dumb comparisons. Don't listen to this Wolfram post. At least the Common Lisp/Clojure code wasn't pointless.
Clojure's combination of declarative idioms and highly functional idioms (to a greater extent than CL) lead to it being able to by default express certain things in a more terse fashion.
http://rosettacode.org/wiki/Factors_of_an_integer
Compare Common Lisp and Clojure here.
Common Lisp:
(defun factors (n &aux (lows '()) (highs '()))
(do ((limit (isqrt n)) (factor 1 (1+ factor)))
((= factor limit)
(when (= n (* limit limit))
(push limit highs))
(nreconc lows highs))
(multiple-value-bind (quotient remainder) (floor n factor)
(when (zerop remainder)
(push factor lows)
(push quotient highs)))))
Clojure solutions:
; Trivial version
(defn factors [n]
(filter #(zero? (rem n %)) (range 1 (inc n))))
; Less trivial
(defn factors [n]
(into (sorted-set)
(mapcat (fn [x] [x (/ n x)])
(filter #(zero? (rem n %)) (range 1 (inc (sqrt n)))) )))
; Now with a for-comprehension.
(defn factors [n]
(into (sorted-set)
(reduce concat
(for [x (range 1 (inc (sqrt n))) :when (zero? (rem n x))]
[x (/ n x)]))))
Take a look at the other things that work across the seq'able data structures. Pretty powerful base set of idioms.
Okay, but lets provide a second opinion on these idioms. What about Haskell?
import HFM.Primes(primePowerFactors)
import Data.List
factors = map product.
mapM (uncurry((. enumFromTo 0) . map .(^) )) . primePowerFactors
And Mathematica?
Factorize[n_Integer] := Divisors[n]
I give up. These are dumb comparisons. Don't listen to this Wolfram post. At least the Common Lisp/Clojure code wasn't pointless.
Common Lisp code basically equivalent to that first "trivial" Clojure implementation (WARNING: untested, may consist entirely of bugs):
(defun factors (n)
(loop for r upfrom 1 to n when (zerop (remainder n r)) collect r))
Perhaps what we can take away from this is that Mathmatica (at least in this case, and I would imagine many others) is more terse that other languages because the base language includes features such as factorization that require libraries in other languages. This is obviously beneficial in many cases, but I agree with you that it is disingenuous to suggest that Mathematica's terseness is primarily due to well-designed syntax/semantics as opposed to feature-richness.
This. As I said before: there is a lot that is a library elsewhere that's built-in in Mathematica. And that's why I think Mathematica is very good at rapid prototyping.
I agree with you there.
Another thought from my experiences:
The only way I've seen to scale up the "expressiveness" in a programming language in a way that's generally applicable is to make it more declarative. I have mixed feelings about this.
Two primary ways of doing so are functional idioms for "iteration" and concatenative programming.
Things like juxt, mapcat, et al have their equivalents in languages like J, K, Factor, etc.
The only way to elevate yourself above loops and manually assigning intermediates is to flow the data through high-level descriptions of what you want.
Whether that takes the form of a concatenative or functional langauge depends on the coder.
OOP is more about scaling up procedural idioms so that state management and code reuse aren't a total Vietnam. They aren't really more "expressive" with the exception of some of the idioms Smalltalk tossed about.
All that's left otherwise is "ecosystem leverage" such as Mathematica exhibits here. That's not really about the semantics of the programming language, but it's still worth noting if the leverage is relevant to the problem you're solving.
Another thought from my experiences:
The only way I've seen to scale up the "expressiveness" in a programming language in a way that's generally applicable is to make it more declarative. I have mixed feelings about this.
Two primary ways of doing so are functional idioms for "iteration" and concatenative programming.
Things like juxt, mapcat, et al have their equivalents in languages like J, K, Factor, etc.
The only way to elevate yourself above loops and manually assigning intermediates is to flow the data through high-level descriptions of what you want.
Whether that takes the form of a concatenative or functional langauge depends on the coder.
OOP is more about scaling up procedural idioms so that state management and code reuse aren't a total Vietnam. They aren't really more "expressive" with the exception of some of the idioms Smalltalk tossed about.
All that's left otherwise is "ecosystem leverage" such as Mathematica exhibits here. That's not really about the semantics of the programming language, but it's still worth noting if the leverage is relevant to the problem you're solving.
The Common Lisp code is carefully crafted to minimize the number of divisions.
Of course Mathematica is good at succinctly expressing mathematical functions.
I'd like to see how much Mathematica code it'd take to write a compiler or web server.
I'd like to see how much Mathematica code it'd take to write a compiler or web server.
probably quite a bit less code than the equivalents in other languages. Mathematica is the most powerful language i know of in terms of abstractive power. essentially it's a symbolically-evaluated Lisp
the reason you don't see more general-purpose code written in Mathematica is simply because it's a proprietary environment. you can't share general-purpose programs that are written in Mathematica
the reason you don't see more general-purpose code written in Mathematica is simply because it's a proprietary environment. you can't share general-purpose programs that are written in Mathematica
You should check out Mathics; It's a free and open source implementation of the 'Mathematica language' written in Python.
Admittedly, there are bugs, and the performance is not as good as Mathematica, but a lot of progress has been made (For example, PyPy compatibility is planed for the upcoming release). Give it some time, and I think (hope) it will grow to fill this niche.
Also, The fact that Mathics is open, allows you to extend it (in Python for example) with minimal effort.
Disclaimer: I contribute to Mathics development in my free time. We are hoping to get more people involved!
Admittedly, there are bugs, and the performance is not as good as Mathematica, but a lot of progress has been made (For example, PyPy compatibility is planed for the upcoming release). Give it some time, and I think (hope) it will grow to fill this niche.
Also, The fact that Mathics is open, allows you to extend it (in Python for example) with minimal effort.
Disclaimer: I contribute to Mathics development in my free time. We are hoping to get more people involved!
Ok, now make a chart of "code length per dollar cost to use".
Am I missing something, or the solutions have vastly different approaches, sample code, etc? What will you get by measuring that?
Consider that a few years ago, Reddit switched from Lisp to Python, not because they thought the latter was in any way a better language, but (in their own words) because of Common Lisp's "lack of widely used and tested libraries."[1]
How does Mathematica's ecosystem compare to those of the popular languages in this comparison?
--
[1] http://blog.reddit.com/2005/12/on-lisp.html