Let me give a second, more technical answer.
[Edit May 22, 2016: A polished, improved version of this answer is available in the following two articles: The Physics of Virtual Particles and Misconceptions about Virtual Particles.]
Observable particles. In QFT, observable (hence real) particles of mass $m$ are conventionally defined as being associated with poles of the S-matrix at energy $E=mc^2$ in the rest frame of the system (Peskin/Schroeder, An introduction to QFT, p.236). If the pole is at a real energy, the mass is real and the particle is stable; if the pole is at a complex energy (in the analytic continuation of the S-matrix to the second sheet), the mass is complex and the particle is unstable. At energies larger than the real part of the mass, the imaginary part determines its decay rate and hence its lifetime (Peskin/Schroeder, p.237); at smaller energies, the unstable particle cannot form for lack of energy, but the existence of the pole is revealed by a Breit-Wigner resonance in certain cross sections. From its position and width, one can estimate the mass and the lifetime of such a particle before it has ever been observed. Indeed, many particles listed in the tables http://pdg.lbl.gov/2011/reviews/contents_sports.html by the Particle Data Group (PDG) are only resonances.
Stable and unstable particles. A stable particle can be created and annihilated, as there are associated creation and annihilation operators that add or remove particles to the state. According to the QFT formalism, these particles must be on-shell. This means that their momentum $p$ is related to the real rest mass $m$ by the relation $p^2=m^2$.
More precisely, it means that the 4-dimensional Fourier transform of the time-dependent single-particle wave function associated with it has a support that satisfies the on-shell relation $p^2=m^2$. There is no need for this wave function to be a plane wave, though these are taken as the basis functions between the scattering matrix elements are taken.
An unstable particle is represented quantitatively by a so-called Gamov state (see, e.g., http://arxiv.org/pdf/quant-ph/0201091.pdf), also called a Siegert state (see, e.g., http://www.cchem.berkeley.edu/millergrp/pdf/235.pdf) in a complex deformation of the Hilbert space of a QFT, obtained by analytic continuation of the formulas for stable particles. In this case, as $m$ is complex, the mass shell consists of all complex momentum vectors $p$ with $p^2=m^2$ and $v=p/m$ real, and states are composed exclusively of such momentaum vectors. This is the representation in which one can take the limit of zero decay, in which the particle becomes stable (such as the neutron in the limit of negligible electromagnetic interaction), and hence the representation appropriate in the regime where the unstable particle can be observed (i.e., resolved in time).
A second representation in terms of normalizable states of real mass is given by a superposition of scattering states of their decay products, involving all energies in the range of the Breit-Wigner resonance. In this standard Hilbert space representation, the unstable particle is never formed; so this is the representation appropriate in the regime where the unstable particle reveals itself only as a resonance.
The 2010 PDG description of the Z boson, http://pdg.lbl.gov/2011/reviews/rpp2011-rev-z-boson.pdf discusses both descriptions in quantitative detail (p.2: Breit-Wigner approach; p.4: S-matrix approach).
(added March 18, 2012): All observable particles are on-shell, though the mass shell is real only for stable particles.
Virtual (or off-shell) particles. On the other hand, virtual particles are defined as internal lines in a Feynman diagram (Peskin/Schroeder, p.5, or Zeidler, QFT I Basics in mathematics and physiics, p.844). and this is their only mode of being. In diagram-free approaches to QFT such as lattice gauge theory, it is even impossible to make sense of the notion of a virtual particle. Even in orthodox QFT one can dispense completely with the notion of a virtual particle, as Vol. 1 of the QFT book of Weinberg demonstrates. He represents the full empirical content of QFT, carefully avoiding mentioning the notion of virtual particles.
As virtual particles have real mass but off-shell momenta, and multiparticle states are always composed of on-shell particles only, it is impossible to represent a virtual particle by means of states. States involving virtual particles cannot be created for lack of corresponding creation operators in the theory.
A description of decay requires an associated S-matrix, but the in- and out- states of the S-matrix formalism are composed of on-shell states only, not involving any virtual particle. (Indeed, this is the reason for the name ''virtual''.)
For lack of a state, virtual particles cannot have any of the usual physical characteristics such as dynamics, detection probabilities, or decay channels. How then can one talk about their probability of decay, their life-time, their creation, or their decay? One cannot, except figuratively!
Virtual states. (added on March 19, 2012): In nonrelativistic scattering theory, one also meets the concept of virtual states, denoting states of real particles on the second sheet of the analytic continuation, having a well-defined but purely inmaginary energy, defined as a pole of the S-matrix. See, e.g., Thirring, A course in Mathematical Physics, Vol 3, (3.6.11).
The term virtual state is used with a different meaning in virtual state spectroscopy (see, e.g., http://people.bu.edu/teich/pdfs/PRL-80-3483-1998.pdf), and denotes there an unstable energy level above the dissociation threshold. This is equivalent with the concept of a resonance.
Virtual states have nothing to do with virtual particles, which have real energies but no associated states, though sometimes the name ''virtual state'' is associated to them. See, e.g., https://researchspace.auckland.ac.nz/bitstream/handle/2292/433/02whole.pdf; the author of this thesis explains on p.20 why this is a misleading terminology, but still occasionally uses this terminology in his work.
Why are virtual particles often confused with unstable particles? As we have seen, unstable particles and resonances are observable and can be characterized quantitatively in terms of states. On the other hand, virtual particles lack a state and hence have no meaningful physical properties.
This raises the question why virtual particles are often confused with unstable particles, or even identified.
The reason, I believe, is that in many cases, the dominant contribution to a scattering cross section exhibiting a resonance comes from the exchange of a corresponding virtual particle in a Feynman diagram suggestive of a collection of world lines describing particle creation and annihilation. (Examples can be seen on the Wikipedia page for W and Z bosons, http://en.wikipedia.org/wiki/Z-boson.)
This space-time interpretation of Feynman diagrams is very tempting graphically, and contributes to the popularity of Feynman diagrams both among researchers and especially laypeople, though some authors - notably Weinberg in his QFT book - deliberately resist this temptation.
However, this interpretation has no physical basis. Indeed, a single Feynman diagram usually gives an infinite (and hence physically meaningless) contribution to the scattering cross section. The finite, renormalized values of the cross section are obtained only by summing infinitely many such diagrams. This shows that a Feynman diagram represents just some term in a perturbation calculation, and not a process happening in space-time. Therefore one cannot assign physical meaning to a single diagram but at best to a collection of infinitely many diagrams.
The true meaning of virtual particles. For anyone still tempted to associate a physical meaning to virtual particles as a specific quantum phenomenon, let me note that Feynman-type diagrams arise in any perturbative treatment of statistical multiparticle properties, even classically, as any textbook of statistical mechanics witnesses.
More specifically, the paper http://homepages.physik.uni-muenchen.de/~helling/classical_fields.pdf shows that the perturbation theory for any classical field theory leads to an expansion into Feynman diagrams very similar to those for quantum field theories, except that only tree diagrams occur. If the picture of virtual particles derived from Feynman diagrams had any intrinsic validity, one should conclude that associated to every classical field there are classical virtual particles behaving just like their quantum analogues, except that (due to the lack of loop diagrams) there are no virtual creation/annihilation patterns. But in the literature, one can find not the slightest trace of a suggestion that classical field theory is sensibly interpreted in terms of virtual particles.
The reaon for this similarity in the classical and the quantum case is that Feynman diagrams are nothing else than a graphical notation for writing down products of tensors with many indices summed via the Einstein summation convention. The indices of the results are the external lines aka ''real particles'', while the indices summed over are the internal lines aka ''virtual particles''. As such sums of products occur in any multiparticle expansion of expectations, they arise irrespective of the classical or quantum nature of the system.
(added September 29, 2012)
Interpreting Feynman diagrams.
Informally, especially in the popular literature, virtual paricles are viewed as transmitting the fundamental forces in quantum field theory. The weak force is transmitted by virtual Zs and Ws. The strong force is transmitted by virtual gluons. The electromagnetic force is transmitted by virtual photons. This ''proves'' the existence of virtual particles in the eyes of their afficionados.
The physics underlying this figurative speech are Feynman diagrams, primarily the simplest tree diagrams that encode the low order perturbative contributions of interactions to the classical limit of scattering experiments. (Thus they are really a manifestation of classical perturbative field theory, not of quantum fields. Quantum corrections involve at least one loop.)
Feynman diagrams describe how the terms in a series expansion of the S-matrix elements arise in a perturbative treatment of the interactions as linear combinations of multiple integrals. Each such multiple integral is a product of vertex contributions and propagators, and each propagator depends on a 4-momentum vector that is integrated over. In additon, there is a dependence on the momenta of the ingoing (prepared) and outgoing (in principle detectable) particles.
The structure of each such integral can be represented by a Feynman diagram. This is done by associating with each vertex a node of the diagram and with each momentum a line; for ingoing momenta an external line ending in a node, for outgoing momenta an external line starting in a node, and for propagator momenta an internal line between two nodes.
The resulting diagrams can be given a very vivid but superficial interpretation as the worldlines of particles that undergo a metamorphosis (creation, deflection, or decay) at the vertices. In this interpretation, the in- and outgoing lines are the worldlines of the prepared and detected particles, respectively, and the others are dubbed virtual particles, not being real but required by this interpretation. This interpretation is related to - and indeed historically originated with - Feynman's 1945 intuition that all particles take all possible paths with a probability amplitute given by the path integral density. Unfortunately, such a view is naturally related only to the formal, unrenormalized path integral. But there all contributions of diagrams containing loops are infinite, defying a probability interpretation.
According to the definition in terms of Feynman diagrams, a virtual particle has specific values of 4-momentum, spin, and charges, characterizing the form and variables in its defining propagator. As the 4-momentum is integrated over all of $R^4$, there is no mass shell constraint, hence virtual particles are off-shell.
Beyond this, formal quantum field theory is unable to assign any property or probability to a virtual particle. This would require to assign to them states, for which there is no place in the QFT formalism. However, the interpretation requires them to exist in space and time, hence they are attributed by inmagination with all sorts of miraculous properties that complete the picture to something plausible. (See, for example, the Wikipedia article on virtual particles.) Being dressed with a fuzzy notion of quantum fluctuations, where the Heisenberg uncertainty relation allegedly allows one to borrow for a very short time energy from the quantum bank, these properties have a superficial appearance of being scientific. But they are completely unphysical as there is neither a way to test them experimentally nor one to derive them from formal properties of virtual particles.
The long list of manifestations of virtual particles mentioned in the Wikipedia article cited are in fact manifestations of computed scattering matrix elements. They manifest the correctness of the formulas for the multiple integrals associated with Feynman diagrams, but not the validity of the claims about virtual particles.
Though QFT computations generally use the momentum representation, there is also a (physically useless) Fourier-transformed complementary picture of Feynman diagrams using space-time positions in place of 4-momentaa. In this version, the integration is over all of space-time, so virtual particles now have space-time positions but no dynamics, hence no world lines. (In physics, dynamics is always tied to states and an equation of motion. No such thing exists for virtual particles.)
Can one distinguish real and virtual photons?
There is a widespread view that external legs of Feynman diagrams are in reality just internal legs of larger diagrams. This would mean that, in reality, every leg is internal, and thus blurs the distinction between real and virtual particles.
The basic argument behind this view is the fact that the photons that hit an eye (and this give evidence of something real) were produced by excitation form some distant object. This view is consistent with regarding the creation or destruction of photons as what happens at a vertex containing a photon line. In this view, it follows that the universe is a gigantic Feynman diagram with many loops of which we and our experiments are just a tiny part.
But single Feynman diagrams don't have a technical meaning. Only the sum of all Feynman diagrams (or approximately, the sum of all important ones) has predictive value, and the small ones contribute most - otherwise we couldn't do any perturbative calculations.
Moreover, this view contradicts the way QFT computations are actually used. Scattering matrix elements are always considered between on-shell particles. Without exception, comparisons of QFT results with scattering experiments are based on these on-shell results.
It must necessarily be so, as off-shell matrix elements don't make formal sense: Matrix elements are taken between states, and all physical states are on-shell by the basic structure of QFT. Thus the structure of QFT itself enforces a fundamental distinction between real particles representable by states and virtual particles representable by propagators only.
The basic problem invalidating the above argument is the assumption that creation and desctruction of particles in space and time can be identified with vertices in Feynman diagrams. They cannot. For Feynman diagrams lack any dynamical properties, and their interpretation in space and time is sterile.
Thus the view that in reality there are no external lines is based on a superficial, tempting but invalid identification of theoretical concepts with very different properties.
The conclusion is that, indeed, real particles (represented by external legs) and virtual particles (represented by internal legs) are completely separate conceptual entities, clearly distinguished by their meaning. In particular, it never happens that one turns into the other or one affects the other.
This post imported from StackExchange Physics at 2015-11-22 18:28 (UTC), posted by SE-user Arnold Neumaier