Functional languages require infrastructure that inevitably adds overheads over what can theoretically be attained using assembler by hand. In particular, first-class lexical closures only work well with garbage collection because they allow values to be carried out of scope.
Beware of self selection. C acts as a lowest common denominator in benchmark suites, limiting what can be accomplished. If you have a benchmark comparing C with a functional language then it is almost certainly an extremely simple program. Arguably so simple that it is of little practical relevance today. It is not practically feasible to solve more complicated problems using C for a mere benchmark.
The most obvious example of this is parallelism. Today, we all have multicores. Even my phone is a multicore. Multicore parallelism is notoriously difficult in C but can be easy in functional languages (I like F#). Other examples include anything that benefits from persistent data structures, e.g. undo buffers are trivial with purely functional data structures but can be a huge amount of work in imperative languages like C.
Functional languages will seem slower than C because you'll only ever see benchmarks comparing code that is easy enough to write well in C and you'll never see benchmarks comparing meatier tasks where functional languages start to excel.
However, you've correctly identified what is probably the single biggest bottleneck in functional languages today: their excessive allocation rates. Nice work!
The reasons why functional languages allocate so heavily can be split into historical and inherent reasons.
Historically, Lisp implementations have been doing a lot of boxing for 50 years now. This characteristic spread to many other languages which use Lisp-like intermediate representations. Over the years, language implementers have continually resorted to boxing as a quick fix for complications in language implementation. In object oriented languages, the default has been to always heap allocate every object even when it can obviously be stack allocated. The burden of efficiency was then pushed onto the garbage collector and a huge amount of effort has been put into building garbage collectors that can attain performance close to that of stack allocation, typically by using a bump-allocating nursery generation. I think that a lot more effort should be put into researching functional language designs that minimize boxing and garbage collector designs that are optimized for different requirements.
Generational garbage collectors are great for languages that heap allocate a lot because they can be almost as fast as stack allocation. But they add substantial overheads elsewhere. Today's programs are increasingly using data structures like queues (e.g. for concurrent programming) and these give pathological behaviour for generational garbage collectors. If the items in the queue outlive the first generation then they all get marked, then they all get copied ("evacuated"), then all of the references to their old locations get updated and then they become eligible for collection. This is about 3× slower than it needs to be (e.g. compared to C). Mark region collectors like Beltway (2002) and Immix (2008) have the potential to solve this problem because the nursery is replaced with a region that can either be collected as if it were a nursery or, if it contains mostly reachable values, it can be replaced with another region and left to age until it contains mostly unreachable values.
Despite the pre-existence of C++, the creators of Java made the mistake of adopting type erasure for generics, leading to unnecessary boxing. For example, I benchmarked a simple hash table running 17× faster on .NET than the JVM partly because .NET did not make this mistake (it uses reified generics) and also because .NET has value types. I actually blame Lisp for making Java slow.
All modern functional language implementations continue to box excessively. JVM-based languages like Clojure and Scala have little choice because the VM they target cannot even express value types. OCaml sheds type information early in its compilation process and resorts to tagged integers and boxing at run-time to handle polymorphism. Consequently, OCaml will often box individual floating point numbers and always boxes tuples. For example, a triple of bytes in OCaml is represented by a pointer (with an implicit 1-bit tag embedded in it that gets checked repeatedly at run-time) to a heap-allocated block with a 64 bit header and 192 bit body containing three tagged 63-bit integers (where the 3 tags are, again, repeatedly examined at run time!). This is clearly insane.
Some work has been done on unboxing optimizations in functional languages but it never really gained traction. For example, the MLton compiler for Standard ML was a whole-program optimizing compiler that did sophisticated unboxing optimizations. Sadly, it was before its time and the "long" compilation times (probably under 1s on a modern machine!) deterred people from using it.
The only major platform to have broken this trend is .NET but, amazingly, it appears to have been an accident. Despite having a Dictionary implementation very heavily optimized for keys and values that are of value types (because they are unboxed) Microsoft employees like Eric Lippert continue to claim that the important thing about value types is their pass-by-value semantics and not the performance characteristics that stem from their unboxed internal representation. Eric seems to have been proven wrong: more .NET developers seem to care about unboxing than pass-by-value. Indeed, most structs are immutable and, therefore, referentially transparent so there is no semantic difference between pass-by-value and pass-by-reference. Performance is visible and structs can offer massive performance improvements. The performance of structs even saved Stack Overflow and structs are used to avoid GC latency in commercial software like Rapid Addition's!
The other reason for heavy allocation by functional languages is inherent. Imperative data structures like hash tables use huge monolithic arrays internally. If these were persistent then the huge internal arrays would need to be copied every time an update was made. So purely functional data structures like balanced binary trees are fragmented into many little heap-allocated blocks in order to facilitate reuse from one version of the collection to the next.
Clojure uses a neat trick to alleviate this problem when collections like dictionaries are only written to during initialization and are then read from a lot. In this case, the initialization can use mutation to build the structure "behind the scenes". However, this does not help with incremental updates and the resulting collections are still substantially slower to read than their imperative equivalents. On the up-side, purely functional data structures offer persistence whereas imperative ones do not. However, few practical applications benefit from persistence in practice so this is often not advantageous. Hence the desire for impure functional languages where you can drop to imperative style effortlessly and reap the benefits.